content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{Introduction}
\thispagestyle{empty}
\noindent
The Dirac equation is one of the most beautiful creations of human intellect. It opened a new era in particle physics, providing impressive results among which we can mention the prediction of antiparticles such as the positron. Indeed, the prediction of the positron by Dirac in 1928, --and the posterior discovery by Anderson in 1932--, marks one of the major events in the history of science. However, in spite of their impressive agreement with experiments, it is not clear whether the Dirac equation will \emph{always} prevail as an untouchable scientific truth. We specifically refer to the ultra-high energy regimes, i.e, $p>>m$ , where possible deviations from Lorentz invariance might be present. These possible violations of Lorentz invariance have received attention in the literature and are being extensively studied in the last years, specially in the context of ultra-high energy cosmic-rays.\cite{Macc,Bie}\\
It is not the purpose of this letter the study of a competitor to the Dirac wave equation. There is no doubt that fermions follow the Dirac equation, whose unrivalled success in the understanding of matter is not subjected to discussion. However, fermions such as the electron also satisfy the Schr\"odinger equation, but in the low-energy limit. Indeed, as is well known, the Dirac equation replaces the Schr\"odinger equation in the relativistic regime. Then, it seems natural to wonder: Is there life beyond the Dirac equation? In this work we try to provide an answer to this question. In particular, we want to explore if a theoretical alternative to the Dirac equation at very high energies can be possible. From a theoretical point of view, the study of such a possibility is not only a legitimate research program, but also a quite interesting intellectual exercise.
Indeed, as we will see, in the regime $p>>m$, it seems to be room for the existence of another matter wave equation, which can be derived starting from first principles. This matter wave equation also includes spin in a natural way. In fact, we will show that the wave equation is not only suitable for fermions of spin 1/2, but does also describe massive bosons of spin 1 by means of a substitution of the spin operators $\boldsymbol{\sigma}\rightarrow\boldsymbol{\alpha}$, where $\boldsymbol{\alpha}$ are the Majorana-Oppenheimer matrices \cite{Maj,Opp31}.
The paper is organized as follows. In sec.\ref{Derivation}, we derive the wave equation starting from first principles. In sec.\ref{Canonical formulation}, we establish the Lagrangian formulation and the Hamiltonian formalism. Finally, in sec.\ref{conclusions}, we present the conclusions of this work..
\section{Derivation and physical interpretation}
\label{Derivation}
\thispagestyle{empty}
\noindent
In order to derive a wave equation in Physics, we have to focus on the energy-momentum relation assumed. In the same fashion that Schr\"odinger's equation is derived assuming a classical relation, $E=p^{2}/2m$, the Klein-Gordon equation can be obtained taking the relativistic, $E^{2}=m^{2}+p^{2}$, and the Dirac equation emerges assuming a linear relation, $E=\alpha^{i}p_{i}+\beta m$. Then, it seems natural to wonder what wave equation would correspond to the case, $E=p+\frac{m^{2}}{2p}$, which is the subject of this work.
In the theoretical discussion that follows, we will maintain the constants $\hbar$, $c$, in all the expressions unless otherwise noted. Let us begin the discussion with the energy-momentum relation of the Special Theory of Relativity (STR) for a free particle with positive energy
\be\label{Energy}
E=\sqrt{c^{2}p^{2}+m^{2}c^{4}}
\,
\ee
Where $m$ is the rest mass. As is well known, in the non-relativistic limit, $p<<mc$, we can approximate this equation as
\be\label{Energy2}
E\simeq mc^{2}\Big(1+\frac{p^{2}}{2m^{2}c^{2}}\Big)= mc^{2}+\frac{p^{2}}{2m}
\,
\ee
However, in the ultra-relativistic limit, $mc << p$, equation (\ref{Energy}) provides
\be\label{Energy3}
E\simeq cp\Big(1+\frac{m^{2}c^{2}}{2p^{2}}\Big)=cp+\frac{m^{2}c^{3}}{2p}
\,
\ee
Multiplying by $p$ the last equation and making explicit the operator representation, $\widehat{E}\displaystyle\rightarrow i\hbar\frac{\partial}{\partial t}$, $\widehat{p}\displaystyle\rightarrow-i\hbar\nabla$, we arrive to the following partial differential wave equation
\be\label{Wave_equation}
|\nabla|\frac{\partial\psi}{\partial t}=-c\nabla^{2}\psi+\frac{m^{2}c^{3}}{2\hbar^{2}}\psi
\,
\ee
\\
Where, $\nabla^{2}\equiv\partial^{2}_{x}+\partial^{2}_{y}+\partial^{2}_{z}$ is the Laplacian operator and, $|\nabla|\equiv(\partial^{2}_{x}+\partial^{2}_{y}+\partial^{2}_{z})^{1/2}$ its square root. We can express the square root of the Laplacian as, $|\nabla|=\vec{\sigma}\cdot\nabla$; In cartesian coordinates
\begin{equation}\label{spinors}
|\nabla|=\vec{\sigma}\cdot\nabla=\sigma_{x}\partial_{x}+\sigma_{y}\partial_{y}+\sigma_{z}\partial_{z}
\,
\end{equation}
Where, $\sigma_{x}$, $\sigma_{y}$, $\sigma_{z}$ are certain operators. The condition, $(\vec{\sigma}\cdot\nabla)^{2}=\nabla^{2}$ can be written as
\begin{equation}
\Big(\sigma_{x}\partial_{x}+\sigma_{y}\partial_{y}+\sigma_{z}\partial_{z}\Big)\cdot \Big(\sigma_{x}\partial_{x}+\sigma_{y}\partial_{y}+\sigma_{z}\partial_{z}\Big)=\partial^{2}_{x}+\partial^{2}_{y}+\partial^{2}_{z}
\,
\end{equation}
Then, the operators $\sigma_{i}$ are subjected to the constraints
\begin{eqnarray}
\sigma^{2}_{i}=I
\qquad
\{\sigma_{i},\sigma_{j}\}=2\delta_{ij}I
\end{eqnarray}
This allows to choose a representation of the sigmas given in terms of the Pauli matrices
\begin{eqnarray}\label{Pauli}
\sigma_{x}=\begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} \qquad
\sigma_{y} = \begin{pmatrix} 0 & -i \\ i & 0 \end{pmatrix} \qquad \sigma_{z} = \begin{pmatrix} 1 & 0 \\ 0 & -1
\end{pmatrix}
\end{eqnarray}
Finally, we can write the explicit form of the wave equation (\ref{Wave_equation}) is terms of the Pauli matrices as follows
\begin{equation}\label{wave_equationexplicit}
\sigma_{x}\frac{\partial^{2}\psi}{\partial x\partial t}+\sigma_{y}\frac{\partial^{2}\psi}{\partial y\partial t}+\sigma_{z}\frac{\partial^{2}\psi}{\partial z\partial t}=-c\nabla^{2}\psi+\frac{m^{2}c^{3}}{2\hbar^{2}}\psi
\end{equation}
\\
Or in the compact form
\begin{equation}\label{compact}
(\vec{\sigma}\cdot\nabla)\frac{\partial \psi}{\partial t}=-c\nabla^{2}\psi+\frac{m^{2}c^{3}}{2\hbar^{2}}\psi
\,
\end{equation}
This explicit emergence of the Pauli matrices (\ref{Pauli}) in the wave equation indicates that (\ref{compact}) is describing particles of spin $1/2$ in the ultra-high energy regime, $p>>mc$.
It is important to note that this wave equation is not covariant (with respect to a Lorentz Transformation). This is because the energy and the momentum do not share the same power in the energy-momentum relation, which implies that the spatial and temporal derivatives that appear in the wave equation are not of the same order. Since we are neglecting in the power series expansion of (\ref{Energy3}) all the terms beyond, $m^{2}c^{2}/2p$, the breakdown of Lorentz invariance (LI), given by (\ref{Energy3}) is therefore of order, $\mathcal{O}((mc)^{4}/p^{2})$, this represents a very small violation of LI, almost negligible.
At any rate, experiments have the last word. If a wave equation explains the experimental results but fails to be Lorentz invariant, then the problem is not necessarily in the wave equation. In this sense, it is well known that there is no clue so far of Lorentz invariance violations (LIV) in particle physics experiments. However, a large amount of theoretical effort is being spent trying to study this possibility, specially in the context of the cosmic rays, which are conformed by ultra-relativistic particles.\cite{Macc,Bie}. Although the wave equation is not LI, it is interesting to note that with some manipulations it can be written in a form that resembles a covariant equation:
\\
\begin{equation}\label{covariant}
(\sigma_{\mu}\partial^{\mu})\frac{\partial\psi}{\partial x^{0}}=\Big(-\partial_{\mu}\partial^{\mu}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi
\end{equation}
where $\sigma_{0}=I_{2x2}$. Indeed, since $\sigma_{0}\partial^{0}\partial_{0}\psi=\frac{1}{c^{2}}\frac{\partial^{2}\psi}{\partial t^{2}}I_{2x2}$, this equation leads to
\\
\begin{equation}\label{covariant2}
\frac{1}{c^{2}}\frac{\partial^{2}\psi}{\partial t^{2}}+(\boldsymbol{\sigma}\cdot\nabla)\frac{\partial\psi}{c\partial t}=\Big(-\nabla^{2}+\frac{1}{c^{2}}\frac{\partial^{2}}{\partial t^{2}}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi
\end{equation}
Then, we have a cancellation between the second-order partial derivatives with respect to the time coordinate, recovering (\ref{compact}). From the observation of (\ref{covariant2}), we can see that the term $\partial_{0}\psi$ is preventing the covariance of the wave equation. Indeed, this is the zero component of a four-vector, $\partial_{\mu}\psi$, referred to a particular coordinate system, meanwhile all the other terms of the wave equation can be written in a manifest Lorentz invariant fashion. On the other hand, the Hamiltonian form of the wave equation can easily be provided:
\begin{equation}\label{covariant}
(-i\sigma_{\mu}\partial^{\mu})\mathcal{\hat{H}}\psi=\Big(-\hbar\partial_{\mu}\partial^{\mu}+\frac{m^{2}c^{2}}{2\hbar}\Big)\psi
\end{equation}
\\
where
\begin{equation}\label{standard Hamiltonian}
\mathcal{\hat{H}}\psi=i\hbar\frac{\partial\psi}{\partial t}
\end{equation}
$\mathcal{\hat{H}}$ is the standard Hamiltonian (or time-evolution) operator. In the last part of this work we shall see a peculiar aspect of this theory, namely, that the time evolution operator $\mathcal{\hat{H}}$ and the canonical Hamiltonian $\hat{H_{c}}$ that comes from the Lagrangian formulation are not the same algebraic object.
\subsection{The plane wave solution. Dispersion relation, phase and group velocities of the ultrarelativistic waves.}
Given the wave equation (\ref{wave_equationexplicit}), it seems natural to look for a solution with a plane-wave structure
\be\label{plane_wave}
\psi(r,t)={\chi \choose \phi}e^{i(\vec{k}\cdot\vec{r}-wt)}
\,
\ee
\\
Where, $\vec{k}\cdot\vec{r}=k_{x}x+k_{y}y+k_{z}z$. Substituing this ansatz in (\ref{wave_equationexplicit}) we obtain after a bit of algebra the following matrix equation
\begin{equation}\label{Matrix}
\begin{pmatrix}
k_{z}\omega-ck^{2}-\frac{m^{2}c^{3}}{2\hbar^{2}}& (k_{x}-ik_{y})\omega\\
(k_{x}+ik_{y})\omega&-k_{z}\omega-ck^{2}-\frac{m^{2}c^{3}}{2\hbar^{2}}\\
\end{pmatrix}
{\chi \choose \phi}={0 \choose 0}
\end{equation}
\\
Where, $k^{2}=k_{x}^{2}+k_{y}^{2}+k_{z}^{2}$. In order to obtain non-trivial solutions we must impose the condition, $\det\hat{A}=0$.
It gives the result
\begin{equation}
\det\hat{A}=-k_{z}^{2}\omega^{2}+\Big(ck^{2}+\frac{m^{2}c^{3}}{2\hbar^{2}}\Big)^{2}-\omega^{2}(k_{x}^{2}+k_{y}^{2})=0
\end{equation}
Which implies a dispersion relation, $w(k)$ given by
\begin{equation}\label{dispersion}
w(k)=\pm\Big(ck+\frac{m^{2}c^{3}}{2\hbar^{2}k}\Big)
\end{equation}
The minus sign corresponds to the negative energy solutions.Then, taking into account both possibilities, the most general solution of the wave equation will be a superposition of positive and negative energy modes, namely:
\begin{equation}\label{general solution}
\psi(x)=\sum_{k}\Big(a(k)\boldsymbol{u}(k)e^{i(\vec{k}\cdot\vec{r}-\omega t)}+b^{\dagger}(k)\boldsymbol{v}(k)e^{-i(\vec{k}\cdot\vec{r}-\omega t)}\Big)
\end{equation}
with an spinor of positive energy $\boldsymbol{u}(k)$, and another of negative energy, $\boldsymbol{v}(k)$ satisfying the orthogonality condition $\boldsymbol{u}(k)\cdot\boldsymbol{v}(k)=0$. Indeed, in order to better understand the last result, let us consider the particular case of the one-dimensional propagation along the z-axis. The simplified version of (\ref{wave_equationexplicit}) will be
\be\label{Wave_equation2}
\sigma_{z}\frac{\partial^{2}\psi}{\partial z\partial t}=-c\frac{\partial^{2}\psi}{\partial z^{2}}+\frac{m^{2}c^{3}}{2\hbar^{2}}\psi
\,
\ee
To solve this wave equation we take an ansatz similar to (\ref{plane_wave})
\begin{equation}\label{plane_wave2}
\psi(z,t)={\chi \choose \phi}e^{i(kz-wt)}
\,
\end{equation}
For simplicity we have denoted the $z$ component of the wave vector $\vec{k}$ simply as $k$. The substitution of (\ref{plane_wave2}) in (\ref{Wave_equation2}) provides the following system
\begin{eqnarray}
\frac{kw}{c}\chi=\Big(k^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\chi\,\\
\frac{kw}{c}\phi=-\Big(k^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\phi\,
\,
\end{eqnarray}
These relations imply
\begin{eqnarray}\label{dispersion_relation}
w_{\chi}(k)=ck+\frac{m^{2}c^{3}}{2\hbar^{2}k}\,\\
w_{\phi}(k)=-\Big(ck+\frac{m^{2}c^{3}}{2\hbar^{2}k}\Big)\,
\,
\end{eqnarray}
Of course, they are nothing else than the dispersion relations that we deduced for the 3D case in (\ref{dispersion}). Note that the field, $\phi$ that appears in (\ref{plane_wave}-\ref{plane_wave2}) is the piece of the wave function that corresponds to the negative energy solution. i.e, the field associated to the antiparticle.
On the other hand, it is possible to compute directly from (\ref{dispersion}) the phase and group velocities associated to the positive energy solutions.
\be\label{phase_velocity}
V_{ph}=\frac{w}{k}=c+\frac{m^{2}c^{3}}{2\hbar^{2}k^{2}}=c\Big(1+\Big(\frac{mc}{\sqrt{2}\hbar k}\Big)^{2}\Big)
\,
\ee
\be\label{group_velocity}
V_{g}=\frac{dw}{dk}=c-\frac{m^{2}c^{3}}{2\hbar^{2}k^{2}}=c\Big(1-\Big(\frac{mc}{\sqrt{2}\hbar k}\Big)^{2}\Big)
\,
\ee
Note that in the previous relations $mc<<\hbar k$. Then, although the phase velocity of the ultra-relativistic wave satisfies, $V_{ph}\geq c$, the group velocity (the meaningful concept related with the true energy propagation of the wave), cannot exceed the speed of light. We can therefore conclude that the propagation of these waves is causal and consistent with the STR, the superluminal propagation is not possible in this theory.
Furthermore, given the values of $V_{g}$ and $V_{ph}$, we can assure after a straightforward computation that their product has an upper bound given by $c^{2}$
\begin{equation}
V_{ph}\cdot V_{g}=c^{2}\Big(1-\Big(\frac{mc}{\sqrt{2}\hbar k}\Big)^{4}\Big)
\,
\end{equation}
On the other hand, note that by means of the dispersion relation (\ref{dispersion_relation}), the plane wave solution with positive energy of (\ref{plane_wave2}), can be written in the following manner
\begin{align}\label{neutrino_waves}
\chi(z,t)&=\displaystyle\chi(0)\exp\Big(i(kz-w_{\chi}t)\Big)\nonumber\\
&=\displaystyle\chi(0)\exp\Big(i(kz-kct-\frac{m^{2}c^{3}}{2\hbar^{2}k}t)\Big)\nonumber\\
&\simeq\chi(0)\exp\Big(-i\frac{m^{2}c^{2}z}{2\hbar^{2}k}\Big)
\,
\end{align}
Where we can approximate $z\simeq ct$ if the ultra-relativistic particle travels close to the speed of light.
The wave function (\ref{neutrino_waves}) is a standard result frequently found within the context of the theory of neutrino oscillations\cite{Bil,Bala} . This theory makes the initial assumption that the mass eigenfunctions that describe the propagation of such particles are plane-waves, $|\nu_{i}(z,t)\rangle = \exp(i(k_{i}z-w_{i}t))|\nu_{i}(0)\rangle$, then it is used the approximation (in natural units), $E=p+m^{2}/2p$ to simplify the argument of the exponential and finally obtain
\begin{equation}\label{neutrino_waves2}
|\nu_{i}(z)\rangle = \exp\Big(-i \frac{m_{i}^{2} z}{2p}\Big)|\nu_{i}(0)\rangle
\,
\end{equation}
\\
Where $z$ is the distance between the neutrino production and detection points. Note that both wave functions have the same structure. It is worth noting that we have been able to derive this standard result following a non-standard approach. Indeed, we have proved that the family of plane waves (\ref{neutrino_waves2}), are only particular solutions of the ultra-relativistic wave equation (\ref{Wave_equation}).\\
On the other hand, eigenstates with different masses propagate at different speeds, this is evident following equation (\ref{group_velocity}) which establishes the dependence of the group velocity upon $m^{2}$. This fact is also directly derived from the wave equation.\\
\subsection{ The spin 1 case. Majorana-Oppenheimer matrices}
Let us briefly illustrate how the formalism can be naturally adapted to describe ultrarelativistic bosons of spin 1. This can be achieved by a subtle change of operators without changing the structure itself of the wave equation. Indeed, let us rewrite (\ref{compact}),
in the following form:
\begin{equation}\label{compact2}
(\boldsymbol{\alpha}\cdot\nabla)\frac{\partial \boldsymbol{\psi}}{\partial t}=-c\nabla^{2}\boldsymbol{\psi}+\frac{m^{2}c^{3}}{2\hbar^{2}}\boldsymbol{\psi}
\,
\end{equation}
\\
The difference with respect to the spin 1/2 case lies in the left hand side, but it is worth noting that the structure of the wave equation remains the same with the only modification $\boldsymbol{\sigma}\rightarrow\boldsymbol{\alpha}$. This means that instead of the Pauli matrices, now we have another spin operators. These operators are given by the following matrices:
\begin{eqnarray}\label{Majorana-Oppenheimer1}
\alpha^{1}=\begin{pmatrix} 0 & 0 & 0 \\ 0 & 0 & i \\ 0 & -i & 0 \end{pmatrix} \qquad
\alpha^{2} = \begin{pmatrix} 0 & 0 & -i \\ 0 & 0 & 0\\i & 0 & 0 \end{pmatrix}\qquad
\end{eqnarray}
and
\begin{eqnarray}\label{Majorana-Oppenheimer2}
\alpha^{3} = \begin{pmatrix} 0 & i & 0 \\ -i & 0 &0\\0 & 0 & 0\end{pmatrix}
\end{eqnarray}
These matrices satisfy the angular--momentum commutation rules
\begin{equation}\label{commutation}
[\alpha_i,\alpha_k]=-i\varepsilon_{ikl}\alpha_l\quad
\end{equation}
Matrices (\ref{Majorana-Oppenheimer1},\ref{Majorana-Oppenheimer2}) were introduced by Majorana \cite{Maj} and Oppenheimer \cite{Opp31} in their independent attempt to formulate Maxwell's Electrodynamics as the Field Theory of a massless spin 1 particle within the framework of a Dirac-type equation. Such as the Pauli matrices, these operators are hermitian $\boldsymbol{\alpha}^{\dagger}=\boldsymbol{\alpha}$, and obey the commutation rules of the rotation group SO(3) (\ref{commutation}). Since they are 3x3 matrices, the field $\boldsymbol{\psi}$ must be decomposed as $\boldsymbol{\psi}=(\psi_1,\psi_2,\psi_3)$. In particular, the plane wave solution will be of the form:
\begin{equation}\label{plane_wave_spin1}
\boldsymbol{\psi}(r,t)=\left(\begin{array}{c}\psi_{1}\\\psi_{2}\\\psi_{3}\end{array}\right)e^{i(\vec{k}\cdot\vec{r}-\omega t)}
\,
\end{equation}
\\
The substitution of this ansatz in the wave equation (\ref{compact2}) provides, after some elementary manipulations, a homogeneous matrix system $\hat{A}\boldsymbol{\psi}=\boldsymbol{0}$, similar to that of the spin 1/2 case (\ref{Matrix}), but $\hat{A}$ will be now a 3x3 hermitian matrix. Indeed, the explicit form of this homogeneous system is the following:
\small
\begin{equation}
\begin{pmatrix}
-ck^{2}-\frac{m^{2}c^{3}}{2\hbar^{2}}& ik_{z}\omega & -ik_{y}\omega\\
-ik_{z}\omega & -ck^{2}-\frac{m^{2}c^{3}}{2\hbar^{2}} &ik_{x}\omega\\
ik_{y}\omega& -ik_{x}\omega & -ck^{2}-\frac{m^{2}c^{3}}{2\hbar^{2}}\\
\end{pmatrix}
\left(\begin{array}{c}\psi_{1}\\\psi_{2}\\\psi_{3}\end{array}\right)=\left(\begin{array}{c}0\\0\\0\end{array}\right)
\end{equation}
\normalsize
\\
Note that $\hat{A}^{\dagger}=\hat{A}$. To obtain non-trivial solutions, we must impose again the consistency condition $\det\hat{A}=0$. Through this condition one gets the result:
\begin{equation}
\Big(ck^{2}+\frac{m^{2}c^{3}}{2\hbar^{2}}\Big)\Big(k^{2}\omega^{2}-\Big(ck^{2}+\frac{m^{2}c^{3}}{2\hbar^{2}}\Big)^{2}\Big)=0
\end{equation}
which implies
\begin{equation}
\omega(k)=\pm\Big(ck+\frac{m^{2}c^{3}}{2\hbar^{2}k}\Big)
\end{equation}
This is the dispersion relation $\omega(k)$ expected for a ultrarelativistic particle of energy-momentum relation $E\simeq p+m^{2}/2p$, a result that was already derived for the spin 1/2 case, in (\ref{dispersion}) and (\ref{dispersion_relation}). As in the spin 1/2 case, the negative frequency solution (negative energy), corresponds to the antiparticle.
Finally, the Hamiltonian form of the wave equation for the spin 1 case in a ``covariant" fashion can be written as :
\begin{equation}
(-i\alpha_{\mu}\partial^{\mu})\mathcal{\hat{H}}\boldsymbol{\psi}=\Big(-\hbar\partial_{\mu}\partial^{\mu}+\frac{m^{2}c^{2}}{2\hbar}\Big)\boldsymbol{\psi}
\end{equation}
where $\alpha_{0}=I_{3x3}$
\newpage
\section{Canonical formulation}
\label{Canonical formulation}
\thispagestyle{empty}
\noindent
In this section we proceed to the construction of the Lagrangian and Hamiltonian formulation of the theory. As is well known, a remarkable feature of Field Theory is that all the well defined matter wave equations can be derived from a Lagrangian density, from which a continuity equation, $\frac{\partial\rho}{\partial t}+\nabla\cdot\boldsymbol{J}=0$, follows. The aim of this section is to prove that the wave equation (\ref{compact}) also admits naturally a Lagrangian formalism. The canonical analysis is a powerful tool, not only to study the symmetry transformations of the Lagrangian which allows to apply Noether's theorem to collect the associated conservation laws, but also to build the associated Hamiltonian, a necessary step to carry out the canonical quantization of the field.\footnote{The study of the canonical quantization of the field will be the subject of future work}
\subsection{Lagrangian formulation}
The starting point of the canonical analysis are the following Lagrangians:
\begin{equation}\label{Lagrangian density}
\mathcal{L}_{1/2}=\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}+\nabla\psi\cdot\nabla \psi^{\dagger}+\frac{1}{2c}\Big(\frac{\partial\psi}{\partial t}\boldsymbol{\sigma}\cdot\nabla\psi^{\dagger}+\frac{\partial \psi^{\dagger}}{\partial t}\boldsymbol{\sigma}\cdot\nabla\psi\Big)
\,
\end{equation}
\begin{equation}\label{Lagrangian density_bosons}
\mathcal{L}_{1}=\frac{m^{2}c^{2}}{2\hbar^{2}}\boldsymbol{\psi}\boldsymbol{\psi}^{\dagger}+\nabla\boldsymbol{\psi}\cdot\nabla \boldsymbol{\psi}^{\dagger}+\frac{1}{2c}\Big(\frac{\partial\boldsymbol{\psi}}{\partial t}\boldsymbol{\alpha}\cdot\nabla\boldsymbol{\psi}^{\dagger}+\frac{\partial \boldsymbol{\psi}^{\dagger}}{\partial t}\boldsymbol{\alpha}\cdot\nabla\boldsymbol{\psi}\Big)
\,
\end{equation}
where $c$ is the speed of light. As is well known, the matrices $\boldsymbol{\sigma}$, $\boldsymbol{\alpha}$, are hermitian, $\sigma_{i}^{\dagger}=\sigma_{i}$, $\alpha_{i}^{\dagger}=\alpha_{i}$ which guarantees the hermiticity of both Lagrangians, $\mathcal{L}=\mathcal{L}^{\dagger}$. Since the spin 1 case is identical to the 1/2 case with the replacement $\boldsymbol\sigma\rightarrow\boldsymbol{\alpha}$ in the Lagrangian, we shall restrict ourselves to the study of the 1/2 case, and it should be understood that a similar analysis holds for the case of spin 1. Having made this clarification, the Euler-Lagrange (E-L) equations for the fields $\psi$, $\psi^{\dagger}$, are given by
\begin{equation}\label{Lagrangian density}
\mathcal{L}=\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}+\nabla\psi\cdot\nabla \psi^{\dagger}+\frac{1}{2c}\Big(\frac{\partial\psi}{\partial t}\vec{\sigma}\cdot\nabla\psi^{\dagger}+\frac{\partial \psi^{\dagger}}{\partial t}\vec{\sigma}\cdot\nabla\psi\Big)
\,
\end{equation}
\\
Where $c$ is the speed of light. As is well known, the Pauli matrices are hermitian, $\sigma_{i}^{\dagger}=\sigma_{i}$, which guarantees the hermiticity of the Lagrangian, $\mathcal{L}=\mathcal{L}^{\dagger}$. The Euler-Lagrange (E-L) equations for the fields $\psi$, $\psi^{\dagger}$, are given by
\begin{align}\label{Euler-Lagrange}
\partial_{\mu}\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi)}\Big]-\frac{\partial\mathcal{L}}{\partial\psi}=0,
\qquad
\partial_{\mu}\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi^{\dagger})}\Big]-\frac{\partial\mathcal{L}}{\partial\psi^{\dagger}}=0
\end{align}
\\
Making explicit the summation over the index $\mu$, the E-L equation associated to the hermitian field $\psi^{\dagger}$ will be
\begin{equation}\label{Euler-Lagrange2}
\partial_{0}\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{0}\psi^{\dagger})}\Big]+\nabla\cdot\Big[\frac{\partial\mathcal{L}}{\partial(\nabla\psi^{\dagger})}\Big]-\frac{\partial\mathcal{L}}{\partial\psi^{\dagger}}=0
\,
\end{equation}
\\
Applying the derivatives of (\ref{Euler-Lagrange2}) to the Lagrangian density (\ref{Lagrangian density}), we find
\begin{align}\label{Euler-Lagrange3}
\frac{\partial\mathcal{L}}{\partial(\partial_{0}\psi^{\dagger})}=\frac{1}{2c}\vec{\sigma}\cdot\nabla\psi,
\qquad
\frac{\partial\mathcal{L}}{\partial(\nabla\psi^{\dagger})}=\frac{1}{2c}\frac{\partial\psi}{\partial t}\vec{\sigma}+\nabla\psi
\end{align}
\\
The substitution of these results in (\ref{Euler-Lagrange2}) gives
\begin{align}
0&=\partial_{0}\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{0}\psi^{\dagger})}\Big]+\nabla\cdot\Big[\frac{\partial\mathcal{L}}{\partial(\nabla\psi^{\dagger})}\Big]-\frac{\partial\mathcal{L}}{\partial\psi^{\dagger}}\nonumber\\
&=\frac{\partial}{\partial t}\Big(\frac{1}{2c}\vec{\sigma}\cdot\nabla\psi\Big)+\nabla\cdot\Big(\frac{1}{2c}\frac{\partial\psi}{\partial t}\vec{\sigma}+\nabla\psi\Big)-\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\nonumber\\
&=\frac{1}{c}(\vec{\sigma}\cdot\nabla)\frac{\partial\psi}{\partial t}+\nabla^{2}\psi-\frac{m^{2}c^{2}}{2\hbar^{2}}\psi
\end{align}
\\
We have therefore been able to derive the wave equation (\ref{compact}) from a Lagrangian density by means of the corresponding E-L equations.
\subsection{Global gauge invariance. Noether's theorem and conserved current}
The Lagrangian density (\ref{Lagrangian density}) is invariant under the transformation
\begin{equation}\label{gauge}
\psi\rightarrow\psi^{\prime}=e^{i\theta}\psi
\,
\end{equation}
Then, according to Noether's theorem it must exist a conserved quantity. Indeed, it can be proved that the associated current $J^{\mu}$, satisfies the differential equation $\partial_{\mu}J^{\mu}=0$, where
\begin{equation}\label{current}
J^{\mu}=\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi)}\Big]\delta\psi+\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi^{\dagger})}\Big]\delta\psi^{\dagger}
\,
\end{equation}
\\
This result implies that there exists a certain ``charge" $Q\equiv\int_{V} J^{0}d^{3}x$, which is a constant of motion, i.e, $dQ/dt=0$. For a transformation of the type given by (\ref{gauge}) we have
\begin{align}
\delta\psi=\psi^{\prime}-\psi=\Big(e^{i\theta}-1\Big)\psi\approx i\theta\psi\nonumber\\
\delta\psi^{\dagger}\approx-i\theta\psi^{\dagger}
\,
\end{align}
\\
Substituing the results of equations (\ref{Euler-Lagrange3}) together with these last identities in (\ref{current}), we obtain
\begin{align}
J^{0}&=\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{0}\psi)}\Big]\delta\psi+\Big[\frac{\partial\mathcal{L}}{\partial(\partial_{0}\psi^{\dagger})}\Big]\delta\psi^{\dagger}\nonumber\\
&=\frac{i\theta}{2c}\Big[\Big(\vec{\sigma}\cdot\nabla\psi^{\dagger}\Big)\psi-\Big(\vec{\sigma}\cdot\nabla\psi\Big)\psi^{\dagger}\Big]
\,
\end{align}
\begin{align}
\vec{J}&=\Big[\frac{\partial\mathcal{L}}{\partial(\nabla\psi)}\Big]\delta\psi+\Big[\frac{\partial\mathcal{L}}{\partial(\nabla\psi^{\dagger})}\Big]\delta\psi^{\dagger}\nonumber\\
&=i\theta\Big[\psi\Big(\frac{1}{2c}\frac{\partial\psi^{\dagger}}{\partial t}\vec{\sigma}+\nabla\psi^{\dagger}\Big)-\psi^{\dagger}\Big(\frac{1}{2c}\frac{\partial\psi}{\partial t}\vec{\sigma}+\nabla\psi\Big)\Big]
\,
\end{align}
\\
It is straightforward to show that these functions satisfy the hermiticity condition, $J^{0}=(J^{0})^{\dagger}$, $\vec{J}=\vec{J}^{\dagger}$. On the other hand, since the parameter $\theta$ is an arbitrary constant, we can take $\theta=1$. Finally, the conserved ``charge" will be
\begin{equation}\label{charge}
Q=\displaystyle\int_{V} J^{0}d^{3}x=\frac{i}{2c}\int_{V} \Big[\Big(\vec{\sigma}\cdot\nabla\psi^{\dagger}\Big)\psi-\Big(\vec{\sigma}\cdot\nabla\psi\Big)\psi^{\dagger}\Big]d^{3}x
\end{equation}
\\
In order to verify the robustness of these results, we can check if the divergence $\partial_{\mu}J^{\mu}$ vanishes or not. After a bit of algebra we find
\begin{align}\label{divergence}
\partial_{\mu}J^{\mu}&=\partial_{0}J^{0}+\nabla\cdot\vec{J}\nonumber\\
&=i\psi\Big(\frac{1}{c}\vec{\sigma}\cdot\nabla\frac{\partial\psi^{\dagger}}{\partial t}+\nabla^{2}\psi^{\dagger}\Big)-i\psi^{\dagger}\Big(\frac{1}{c}\vec{\sigma}\cdot\nabla\frac{\partial\psi}{\partial t}+\nabla^{2}\psi\Big)\nonumber\\
&=i\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}-i\frac{m^{2}c^{2}}{2\hbar^{2}}\psi^{\dagger}\psi=0
\,
\end{align}
\\
Therefore, the divergence $\partial_{\mu}J^{\mu}$ vanishes identically as expected. Then, we can conclude that the Lagrangian formulation of the ultra-relativistic wave equation is a consistent theory, and the Lagrangian density (\ref{Lagrangian density}) has a global gauge symmetry compatible with a conserved current.
\subsection{Local gauge invariance}
The generalization to the $U(1)$ case is straightforward. As is well known, for a local phase transformation, i.e, $\psi\rightarrow\psi^{\prime}=e^{i\theta(x)}\psi$, the usual derivative transforms in the following way
\begin{align}
\partial_{\mu}\psi\rightarrow\partial_{\mu}\psi^{\prime}&=\partial_{\mu}(e^{i\theta(x)}\psi)=\partial_{\mu}(e^{i\theta(x)})\psi+e^{i\theta(x)}\partial_{\mu}\psi\nonumber\\
&=e^{i\theta(x)}(i\partial_{\mu}\theta(x))\psi+e^{i\theta(x)}\partial_{\mu}\psi\nonumber\\
&=e^{i\theta(x)}[i\partial_{\mu}\theta(x)+\partial_{\mu}]\psi
\,
\end{align}
Then, the Lagrangian density (\ref{Lagrangian density}) is no longer invariant under this transformation and we must look for a generalization. This generalization is
\begin{equation}\label{Lagrangian_U(1)}
\mathcal{L}=\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}+\mathcal{D}_{i}\psi(\mathcal{D}_{i}\psi)^{\dagger}+\frac{1}{2c}\Big(\mathcal{D}_{0}\psi(\vec{\sigma}\cdot\vec{\mathcal{D}}\psi)^{\dagger}+(\mathcal{D}_{0}\psi)^{\dagger}\vec{\sigma}\cdot\vec{\mathcal{D}}\psi\Big)
\,
\end{equation}
\\
Indeed, the covariant derivative, $\mathcal{D}_{\mu}\equiv(\partial_{\mu}+A_{\mu})$ is subjected to the transformation rule
\begin{equation}
\mathcal{D}_{\mu}\psi\rightarrow\mathcal{D}^{\prime}_{\mu}\psi^{\prime}=(\partial_{\mu}+A_{\mu}^{\prime})e^{i\theta(x)}\psi=e^{i\theta(x)}[i\partial_{\mu}\theta(x)+\partial_{\mu}+A_{\mu}^{\prime}]\psi
\end{equation}
\\
Then, in order to compensate the term $i\partial_{\mu}\theta(x)$, we take the condition $A_{\mu}\rightarrow A^{\prime}_{\mu}=A_{\mu}-i\partial_{\mu}\theta(x)$. which implies, $\mathcal{D}_{\mu}^{\prime}\psi^{\prime}=e^{i\theta(x)}\mathcal{D}_{\mu}\psi$, assuring the invariance of (\ref{Lagrangian_U(1)})
\subsection{The Hamiltonian formalism}
Armed with a consistent Lagrangian theory, the next logical step after the analysis of the internal transformations such as (\ref{gauge}) is the study of the external symmetries and the Hamiltonian formalism. The canonical energy-momentum tensor that comes from the Lagrangian density (\ref{Lagrangian density}), under space-time translational invariance is
\begin{equation}\label{energy-momentum tensor}
T^{\mu}_{\nu}=\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi)}(\partial_{\nu}\psi)+\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\psi^{\dagger})}(\partial_{\nu}\psi^{\dagger})-\delta^{\mu}_{\nu}\mathcal{L}\\
\,
\end{equation}
On the other hand, we can define the Hamiltonian density $\mathcal{H}$, as
\begin{equation}\label{Hamiltonian_density}
\mathcal{H}=T^{0}_{0}=\frac{\partial\mathcal{L}}{\partial\dot{\psi}}\dot{\psi}+\frac{\partial\mathcal{L}}{\partial\dot{\psi^{\dagger}}}\dot{\psi^{\dagger}}-\mathcal{L}=\pi(x)\dot{\psi}+\pi^{\dagger}(x)\dot{\psi^{\dagger}}-\mathcal{L}
\,
\end{equation}\\
The canonical momenta, $\pi(x)$, $\pi^{\dagger}(x)$ are given by the following relations
\begin{align}\label{canonical_momenta}
\pi(x)=\frac{\partial\mathcal{L}}{\partial\dot{\psi}}=\frac{1}{2c}\vec{\sigma}\cdot\nabla\psi^{\dagger},
\qquad
\pi^{\dagger}(x)=\frac{\partial\mathcal{L}}{\partial\dot{\psi^{\dagger}}}=\frac{1}{2c}\vec{\sigma}\cdot\nabla\psi
\end{align}\\
Therefore, substituing the last results in (\ref{Hamiltonian_density}) and using (\ref{Lagrangian density}) we obtain
\begin{align}
\mathcal{H}&=\pi\dot{\psi}+\pi^{\dagger}\dot{\psi^{\dagger}}-\mathcal{L}=\frac{1}{2c}\Big(\dot{\psi}\vec{\sigma}\cdot\nabla\psi^{\dagger}+\dot{\psi}^{\dagger}\vec{\sigma}\cdot\nabla\psi\Big)-\mathcal{L}\nonumber\\
&=\frac{1}{2c}\Big(\dot{\psi}\vec{\sigma}\cdot\nabla\psi^{\dagger}+\dot{\psi}^{\dagger}\vec{\sigma}\cdot\nabla\psi\Big)-\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}-\nabla\psi\cdot\nabla \psi^{\dagger}\nonumber\\
&-\frac{1}{2c}\Big(\frac{\partial\psi}{\partial t}\vec{\sigma}\cdot\nabla\psi^{\dagger}+\frac{\partial \psi^{\dagger}}{\partial t}\vec{\sigma}\cdot\nabla\psi\Big)\nonumber\\
&=-\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}-\nabla\psi\cdot\nabla \psi^{\dagger}
\end{align}\\
The conserved currents and their ``charges", such as $J^{0}$ and $T^{0}_{0}$, are only determined up to a constant. This means that we are free to redefine $\mathcal{H}\equiv-T^{0}_{0}$, in order to have a positive defined Hamiltonian density. Then, we can adopt
\begin{align}
\mathcal{H}&=\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}+\nabla\psi\cdot\nabla \psi^{\dagger}\nonumber\\
&=\frac{m^{2}c^{2}}{2\hbar^{2}}\psi\psi^{\dagger}+\nabla\cdot\Big(\psi^{\dagger}\nabla\psi\Big)-\psi^{\dagger}\nabla^{2}\psi
\,
\end{align}
With this result, the relation between the Hamiltonian $H$ and their density $\mathcal{H}$, is given by
\begin{align}
H\equiv\int_{V}\mathcal{H}d^{3}x&=\int_{V} \psi^{\dagger}\Big(-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi d^{3}x\nonumber\\
&+\nabla\cdot\int_{V} \psi^{\dagger}\nabla\psi d^{3}x
\,
\end{align}
The second term is a divergence which does not change the action, and can be neglected. We collect the final expression
\begin{equation}\label{expectation_value}
H=\int_{V} \psi^{\dagger}\Big(-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi d^{3}x=\int_{V} \psi^{\dagger}\widehat{H}\psi d^{3}x=<\widehat{H}>
\end{equation}
\\
Since, $\widehat{p}=-i\hbar \nabla$, we can write the {\emph{canonical}} Hamiltonian operator $\hat{H_{c}}$ as
\begin{equation}\label{Hamiltonian operator}
\hat{H_{c}}=-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}=\frac{\hat{p}^{2}}{\hbar^{2}}+ \frac{m^{2}c^{2}}{2\hbar^{2}}
\,
\end{equation}
\\
It is easy to see that the operator $\hat{H_{c}}$ is hermitian given their own definition, $\hat{H_{c}}^{\dagger}=\hat{H_{c}}$. On the other hand, the vanishing of the divergence, $\partial_{\mu}T^{\mu}_{\nu}=0$ implies, $d/dt(\int_{V} T^{0}_{0}d^{3}{\bf x})=d/dt(<\hat{H_{c}}>)=0$. Therefore, the expectation value (\ref{expectation_value}) turns out to be a constant of motion.
The Hamiltonian formalism allows us to reinterpret some of the results obtained previously. For instance, the conserved ``charge" (\ref{charge}) associated to the invariance of the Lagrangian under the global phase transformation (\ref{gauge}) can be written as
\begin{align}\label{electric_charge}
Q=\displaystyle\int_{V} J^{0}d^{3}{\bf x}&=\frac{i\theta}{2c}\int_{V} \Big[\Big(\boldsymbol{\sigma}\cdot\nabla\psi^{\dagger}\Big)\psi-\Big(\boldsymbol{\sigma}\cdot\nabla\psi\Big)\psi^{\dagger}\Big]d^{3}{\bf x}\nonumber\\
&=i\theta\int_{V} \Big(\pi(x)\psi(x)-\pi^{\dagger}(x)\psi^{\dagger}(x)\Big)d^{3}{\bf x}
\end{align}
Where we have employed the canonically conjugated momenta that were derived in (\ref{canonical_momenta}).
The physical interpretation of this formula is now transparent. The parameter $\theta$ is the electric charge times a constant. Therefore the conservation law (\ref{divergence}) is expressing nothing but the conservation of the electric charge. If $\psi(x)=\psi^{\dagger}(x)$ then according to (\ref{electric_charge}), $Q=0$ and the field describes a neutral particle. On the other hand, with this explicit expression for the charge given in terms of the fields and their canonical momenta, one is ready to make the next step and promote $Q$ from a classical quantity, to a quantum operator (the charge operator). In other words, the transition $Q\rightarrow\hat{Q}$ is automatic from (\ref{electric_charge}), once the standard anticommutation rules for a fermionic field, $\{\psi(x),\pi(y)\}=i\hbar\delta^{3}(x-y)$, are fixed. This task, however, will not be undertaken in this work.
\subsection{Canonical Hamiltonian Operator Vs Time Evolution Operator}
From the above equations, we can say some important things about the canonical Hamiltonian operator $\hat{H_{c}}$. In the first place it does not depend on time. Secondly, this canonical Hamiltonian involves certain time evolution, but {\emph{stricto sensu}}, it will not be equal to the pure time evolution operator, i.e, $\hat{H_{c}}\neq\mathcal{\hat{H}}$, where $\mathcal{H}$ is the standard Hamiltonian operator defined in (\ref{standard Hamiltonian}). To show this point, let us consider the action of the canonical Hamiltonian $\hat{H_{c}}$ over a plane wave solution of the wave equation
\begin{align}
\hat{H_{c}}\psi_{0}e^{i(\vec{k}\cdot\vec{r}-\omega t)}&=\Big(-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi_{0}e^{i(\vec{k}\cdot\vec{r}-\omega t)}\nonumber\\
&=\Big(k^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi_{0}e^{i(\vec{k}\cdot\vec{r}-\omega t)}=\frac{\omega k}{c}\psi_{0}e^{i(\vec{k}\cdot\vec{r}-\omega t)}
\end{align}
where we have used the dispersion relation $\omega(k)$, associated to the positive energy solution, $\omega(k)=ck+\frac{m^{2}c^{3}}{2\hbar^{2}k}$. We have therefore obtained an eigenvalue equation, $\hat{H_{c}}\psi_{+}=\lambda_{+}\psi_{+}$, with $\lambda_{+}=\omega k/c$. Similarly, for the negative energy solution, $\psi_{-}=\boldsymbol\psi_{0}\exp(i(-\vec{k}\cdot\vec{r}+\omega t))$, one finds:
\small
\begin{align}
\hat{H_{c}}\psi_{0}e^{i(-\vec{k}\cdot\vec{r}+\omega t)}&=\Big(-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi_{0}e^{i(-\vec{k}\cdot\vec{r}+\omega t)}\nonumber\\
&=\Big(k^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi_{0}e^{i(-\vec{k}\cdot\vec{r}+\omega t)}=-\frac{\omega k}{c}\psi_{0}e^{i(-\vec{k}\cdot\vec{r}+\omega t)}
\end{align}
\normalsize
These relations can be condensed in the compact expression, $\hat{H_{c}}\psi_{\pm}=\pm (\omega k/c)\psi_{\pm}$. Since the standard Hamiltonian verifies a different eigenvalue equation, namely, $\mathcal{\hat{H}}\psi_{\pm}=\pm\hbar\omega\psi_{\pm}$, it is obvious the non-equivalence of both operators. However, a close algebraic relation exists among them, and can be found paying attention to the structure of the wave equation. Indeed, since we already know that the wave equation (\ref{compact}) can be written in a Hamiltonian form (\ref{covariant}), we can identify
\begin{equation}
(\boldsymbol{-i\sigma}\cdot\nabla)\mathcal{\hat{H}}\psi=c\hbar \Big(-\nabla^{2}+\frac{m^{2}c^{2}}{2\hbar^{2}}\Big)\psi=c\hbar\hat{H}_{c}\psi
\,
\end{equation}
Then, the exact algebraic relation between $\mathcal{\hat{H}}$ and $\hat{H}_{c}$ is given by the operator equation
\begin{equation}
(\boldsymbol{-i\sigma}\cdot\nabla)\mathcal{\hat{H}}=c\hbar \hat{H}_{c}
\,
\end{equation}
Another interesting consequence that can be extracted from the above relations is that, $[\mathcal{\hat{H}},\hat{H_{c}}]=0$. Of course this implies the conservation of $\hat{H}_{c}$, which is consistent with the result, $d/dt(<\hat{H_{c}}>)=0$, that we derived in the previous section from the canonical formalism. Then, the field theory developed here possesses all the ingredients required to proceed further. In particular, the canonical quantization and the construction of the Hilbert space of physical states are tasks that seem attainable once a Lagrangian formulation is provided. These important questions will be addressed in a forthcoming work.
\section{Discussion}
\label{conclusions}
\thispagestyle{empty}
\noindent
In this work we have presented a wave equation that works for particles whose energy can be approximated by, (ignoring constants), the relation $E\simeq p+\frac{m^{2}}{2p}$.
If such energy-momentum relation encloses a ``hidden" wave equation, then this wave equation can only be the one that we have introduced in this paper, which is a hyperbolic second order linear PDE with well-behaved physical solutions. As we have demonstrated, it can be useful to explain some properties of ultra-relativistic particles. For instance, the family of plane-wave functions usually employed in the theory of neutrino oscillations (\ref{neutrino_waves2}) are only particular solutions of the wave equation discussed here.
Indeed, such as the Dirac equation, the wave equation (\ref{compact}) describes particles of spin $1/2$. In fact, the spin operators are incorporated in a natural way by means of the Pauli matrices, which emerge explicitly in the square root of the Laplacian that appears in the derivation of the wave equation. Interestingly enough, a similar wave equation (\ref{compact2}), can describe massive ultrarelativistic bosons of spin 1, if we replace the Pauli matrices $\boldsymbol{\sigma}$ (\ref{Pauli}), by the Majorana-Oppenheimer $\boldsymbol{\alpha}$ (\ref{Majorana-Oppenheimer}), maintaining the rest of the wave equation unaltered. Therefore the study of the behaviour of this wave equation under different interactions $V(x)$, will allow to enlarge the possible family of solutions, which may be useful to improve the understanding of ultra-relativistic processes, including perhaps the ultra-high energy cosmic-rays.
In addition, a detailed Lagrangian formulation of the wave equation was also provided. In particular, we have proved that this is a consistent theory, where, through the symmetries of the Lagrangian density, some standard and well defined conservation laws are derived in a natural way. On the other hand, by means of the Hamiltonian formalism, we have demostrated that in the free case the expectation value of the operator, $-\nabla^{2}+m^{2}c^{2}/2\hbar^{2}$ is a conserved constant of motion (\ref{expectation_value}).\\
In conclusion, we point out that all the consistent matter wave equations in Physics derive of non-trivial energy-momentum relations. Indeed, if we think of the space of possible non-trivial energy-momentum configurations, we will realize that it is quite constrained: It seems that there are only four consistent possibilites: i). The non-relativistic, $E=p^{2}/2m$. ii). The linear, $E=\alpha^{i} p_{i}+\beta m$. iii). The quadratic, $E^{2}=p^{2}+m^{2}$. iv). The ultra-relativistic, $E=p+m^{2}/2p$. The first three options are all associated with consistent wave equations that describe particles with different properties in their appropriate physical regime. The study of option iv) deserves an analysis, and has been the subject of this work. The hypothetical existence of another matter wave equation is a very interesting possibility that deserves to receive further attention. In this sense, we point out that the field theory presented in this work is a natural alternative to the Dirac wave equation at very high energies. We have demostrated that it reproduces some standard results of the Dirac theory in the limit $p>>mc$, in a quite natural way (\ref{neutrino_waves2}). Besides, it incorporates the possibility of an explicit (and small) Lorentz invariance violation. However, our theory is far from being completely satisfactory. It lacks a canonical quantization, and the construction of a consistent Hilbert space; Nevertheless, with a Lagrangian formulation and a Hamiltonian formalism, the required ingredients to carry out these tasks are available.
|
\section{Introduction}
The last 35 years may be the most rapidly developing 35 years for cosmologists. With the help of quickly-increasing data from observations, peoples now have established the so called standard model of universe. According to this model, the early universe experience a very short time of inflation\cite{2,3}, after which matter and anti-matter begin to form simultaneously through reheatings. If nothing special happens, the amount of matter and anti-matter should be equal. But observations indicate that there are more matter than anti-matter in the universe, the so-called
baryon asymmetry. Quantitatively, this is parameterized by the baryon-to-photon ratio~$\eta$~, whose observation value reads:
\beq{}
\eta_{obs}\thickapprox6\times 10^{-10}.
\label{etaObservation}
\end{equation}
It is unreasonable to explain this asymmetry as the initial condition of universe evolution. Because after inflations, any pre-inflation particle's number density would be diluted to zero so any asymmetries between matter and anti-matter should be wiped out totally. So, to implement the observed the baryon asymmetry, one must invoke some mechanism to generate a net baryon number after the inflation.
In 1967, Sakharov \cite{4} came up with three conditions that processes which can produce the baryon asymmetry should satisfy:
\begin{itemize}
\item The process violate baryon charge conservation.
\item The process violate C and CP invariance.
\item The process should take place in a nonequilibrium thermodynamic state.
\end{itemize}
The first condition comes directly, the second is for the decay of the particles and antiparticles to produce different
numbers of baryons and anti-baryons. The third is mainly to prevent the inverse process from annihilate the baryon asymmetry.
Among large number of theories trying to describe the baryon asymmetry, the most interesting one may be the Affleck-Dine mechanism\cite{5}, which uses scalar field dynamics to get a net baryon number. Their basic idea is, in a matter or a radiation dominated universe, introducing a complex scalar field with $U(1)$-symmetry broken self-interaction and letting the evolution of the scalar field to produce the desired baryon asymmetry. In reference \cite{6}, by associating with inflation scenario, Andrei Linde give a more physical realization for this mechanism. While in reference \cite{7,8,9,10,11,12,13,14,15,16}, more interesting and important questions are discussed.
Both in Affleck-Dine's original work and A. Linde's improvements, the complex scalar field is identified with the classical squark-slepton scalar field or their avatar, neither of which has direct relevance with inflatons. But in a recent work \cite{17}, the Hertzberg and Karouby proposed the idea that the complex scalar field just be the inflaton field $\phi$. So the non-zero net $\phi$-particle number is generated just during latter stage of inflation. While at the end of inflation, by reheating process, the net $\phi$-particles decay into baryons, thus achieve the desired baryon asymmetry.
Hertzberg and Karouby illustrated their idea with the chaotic inflation scenario\cite{17.5}. However, ignited by the recent BICEP2 observation \cite{1}, more focus of the community is attracted to the ``natural inflation''\cite{18} scenario. The natural inflation model is favored by its ``naturalness'' in physical realizations. As is well known \cite{19}, to solve the horizon-, flatness- and other related questions, any successful slow-roll single-field inflation model must satisfy
\beq{}
\chi\equiv\Delta V/(\Delta\phi)^4\leq \mathcal{O}(10^{-6}\sim10^{-8}),
\end{equation}
where $\Delta V$ and $\Delta\phi$ are the change of potential and field respectively during the inflation era. This small ratio of mass scales required is known as the fine-tuning problem in inflation.
It quantifies how flat the inflaton potential should be. In the natural inflation, the flatness of the potential is easily achieved by a shift symmetry under which $\phi\rightarrow\phi + \mathrm{constant}$. Our purpose in this paper is just to adapt Hertzberg and Karouby's idea to the natural inflation model.
This paper is organized as follows: In Section II we introduce the natural inflation model with complex scalar fields and illustrate the process of inflaton asymmetries generation, in Section III
we let the inflaton decay into baryons and give our final results on baryon asymmetries. Section IV is a summary of our work and some discussions.
\section{Net $\phi$-particle generations in natural inflation models}
Let us begin our investigation from the simplest complex scalar field inflaton models, whose action has the form
\beq{}
S=\int d^4 x\sqrt{-g}~\left[\frac{1}{16\pi G}\mathcal{R} +\frac{1}{2}\,|\partial\phi|^2-V(\phi,\phi^\ast)\right],\end{equation}
where $g$ is the determinant of the metric, $\mathcal{R}$ is Ricci scalar and $V(\phi,\phi^\ast)$ is the effective potential of inflaton field $\phi$.
In the current paper we will use the normal flat FRW metric with signatures (+ - - -) and natural units $\hbar=c=1$. Differences between various
inflation models root in their potential function $V$. In the original natural inflation model, the potential, or the lowest order approximation
of the potential is generally of the form:
\beq{}
V(\phi)=\Lambda^4(1\pm\mathrm{cos}(N\phi/f)),
\end{equation}
where $\phi$ is real scalar field; $f$ is the characteristic scale of global symmetry spontaneously breaking; $\Lambda$ is a lower scale associating with some explicit soft symmetry breaking;
the choice of the sign do not affect the physical results and we will choose the negative sign in this paper; the coefficient $N$ is always assumed to be equal to 1. For our purpose in this paper, we will take $\phi$ as a complex scalar field and modify the natural inflation potential and decompose it into two parts:
\beq{}
V(\phi,\phi^\ast)=V_s(|\phi|)+V_b(\phi,\phi^\ast).
\end{equation}
The $V_s$ part conserves the global $U(1)$ symmetry while the $V_b$ part breaks down it. To implement the desired natural inflation, we assume that $V_s$ always dominates and set
\beq{}
V_s(|\phi|)=\Lambda^4(1-\mathrm{cos}(\frac{|\phi|}{f})).
\end{equation}
Obviously $V_s(\phi)$ is invariant under the global $U(1)$ transformation $\phi\rightarrow e^{-i \alpha}\phi$. According to Noether's theorem, any global symmetry
leads to a conserving current. The $U(1)$ symmetry of $V(\phi)$ is related to the net $\phi$-particle number $N_{\phi}-N_{\bar{\phi}}$. So, to obtain a non-zero $\phi$-particle
number which is related to baryon numbers from an initially $\phi\bar{\phi}$ symmetric universe, we must break down this $U(1)$ symmetry. Following Hertzberg and Karouby, we implement this goal by setting the symmetry breaking part in the potential as:
\beq{}
V_b(\phi,\phi^\ast)=\lambda(\phi^n +\phi^{\ast n}),
\end{equation}
where the integer $n\geq 3$ and $\lambda$ is a symmetry breaking parameter. Although the cross terms like $\phi^{n-m}\phi^{\ast m}+\phi^{\ast n-m}\phi^m$ also break the $U(1)$ symmetry, we will not consider them for simplicities.
The smallness of $\lambda$ is natural in physics by 't Hooft's critirial \cite{20}: a small parameter in a theory is natural if, in the limit it is set to zero, the symmetry of the system increases. Obviously, when $\lambda=0$, the $U(1)$ symmetry is recovered, and
the symmetry of the system increases. We also need the smallness of $\lambda$ to preserve the shape of the potential for inflaton, otherwise the character of the natural inflation would be destructed. From the observation aspect, a small value of $\lambda$ is also favored by small baryon-to-photon ratios. Because $\lambda$ is just the measure of $U(1)$-symmetry breaking degree which is responsible for the net particle number's generation.
It's worth mention that despite the $U(1)$ symmetry is broken by $V_b$, the charge conjugation symmetry $\phi\leftrightarrow\phi^{\ast}$ is still respected. We assume that this symmetry is broken in the following process, or the Sakharov's conditions would be violated. In the original Affleck-Dine mechanism, it is spontaneously broken by the interaction with some other light fields. However, the detailed mechanism is
not important to us, and it do not affect our results, so we will not discuss it in this paper.
\subsection{Net $\phi$-particles from $\phi$ and $\bar{\phi}$}
Firstly, noting that the function $V_s$ is a periodic function of period $2\pi f$, we restrict
the value of $|\phi|\in[0,\pi f]$. Secondly, using the fact that $V_s$ takes minimum at $|\phi|=0$, we make Taylor expansion of it at this point as $V_s\approx \frac{1}{2}\Lambda^4(\frac{|\phi|}{f})^2$ when $|\phi|$ is small. Thirdly, since $n\geq3$ in $V_b$, at the late time of inflation during which $|\phi|$
is small, $V_b$ decreases faster than $V_s$ so soon becomes negligible. As results, the effective potential of the inflaton at later times conserves the global $U(1)$ symmetry. According to Noether's theorem, we can derive out the conserving charge as the net particle number:
\beq{}
\Delta N_\phi=N_\phi-N_{\bar{\phi}}=i\int d^3 x \sqrt{g_s} \,(\phi^\ast\, \dot{\phi}-\dot{\phi}^\ast\, \phi),
\end{equation}
here $d^3 x \sqrt{g_s}$ is the spatial volume measure, $N_\phi$ and $N_{\bar{\phi}}$ are the number of $\phi$-and $\bar{\phi}$-particles.
As the roughest approximation, we take $\phi$ as spatial-homogeneous. Substituting the FRW metric into this definition, we can work out the integral and get:
\beq{}
\Delta N_\phi=N_\phi-N_{\bar{\phi}}=iV_{com}a(t)^3\,(\phi^\ast\, \dot{\phi}-\dot{\phi}^\ast\, \phi),
\end{equation}
where $V_{com}$ is the comoving volume and $a(t)$ is the scale factor.
To get the equation of motion for $\phi$, we vary the total action of the system with respect to $\phi^\ast$ and get
\beq{}
\ddot{\phi}+3H\dot{\phi}+\frac{\Lambda^4}{f}\sqrt{\frac{\phi}{\phi^\ast}}\mathrm{sin}(\frac{|\phi|}{f})+2\,\lambda\, n\, \phi^{\ast n-1}=0,
\label{eomPhi}
\end{equation}
where $H=\dot{a}/a$ is the Hubble parameter. Taking the time derivative of $\Delta N_\phi$
\beq{}
\frac{\partial}{\partial t} \Delta N_\phi=i\, V_{com}a^3(3H(\phi^\ast\, \dot{\phi}-\dot{\phi}^\ast\, \phi)+\phi^\ast\, \ddot{\phi}-\ddot{\phi}^\ast\, \phi),
\end{equation}
and substituting the results into an appropriate combination of \eqref{eomPhi} with its complex conjugate, we will get
\bea{}
\Delta N_\phi(t_f)=\Delta N_\phi(t_i)+2i\,\lambda\,V_{com}\,n\,\int^{t_f}_{t_i}\,dt\,a(t)^3\times
\\
(\phi(t)^n-\phi^\ast(t)^n),
\nonumber
\end{eqnarray}
where $\Delta N_\phi(t_i)$ is the initial net particle number at time $t_i$, while $t_f$ denotes the final time. From this equation, we easily see that when the $U(1)$ symmetry is unbroken, i.e. $\lambda\rightarrow0$, the net particle number will indeed be conserved. Since any initial particle number would be diluted by inflation, and the process we feel interested in
happens at the late time of inflation, we will set $\Delta N_\phi(t_i)=0$ from now on.
For conveniences in the latter derivations, we express the scalar field $\phi$ in the polar coordinate as
\beq{}
\phi(t)=\Phi(t)\, e^{i\theta(t)},
\end{equation}
and rewrite the net particle number $\Delta N_\phi$ in the form
\beq{}
\Delta N_\phi(t_f)=-4\lambda\,V_{com}\,n \int^{t_f}_{t_i}\,dt\,a(t)^3\,\Phi(t)^n\,\mathrm{sin}(n\,\theta(t)).
\label{DeltaNphi}
\end{equation}
Like the field $\phi$ and $\phi^*$, the polar field $\Phi$ and angular field $\theta$ also satisfy differential equations similar to \eqref{eomPhi}, which can be solved order by order in $\lambda$.
By equation \eqref{DeltaNphi}, $\Delta N_\phi$ is proportional to $\lambda$. So if we need only calculate $\Delta N_\phi$ to first order approximation which is reasonable when $\lambda$ is small, then we only need to calculate
the integral to the zeroth order in $\lambda$. It can be proved that the evolution of $\theta$ is determined by the symmetry
breaking term. So when we neglect the effect of $\lambda$, $\theta$ doesn't evolve at all, i.e., $\dot{\theta}=0$. For this reason, the factor $\mathrm{sin}(n\,\theta(t))$ in (14) can be extracted out of the integrations. Using $\Phi_0(t)$ and $a_0(t)$ to denote $\Phi(t)$ and $a(t)$ when we
neglect the effect of $\lambda$ in the equations of motion, we can write $\Delta N_\phi$ as the form
\beq{}
\Delta N_\phi(t_f)=-4\lambda\,V_{com}\,n\, \mathrm{sin}(n\,\theta_i)\int^{t_f}_{t_i}\,dt\,a_0(t)^3\,\Phi_0(t)^n ,
\end{equation}
where $\theta_i$ is the initial value of $\theta$.\\
The equation of motion for $\Phi_0$ is easy to derive:
\beq{}
\ddot{\Phi}_0+3H_0\dot{\Phi}_0+\frac{\Lambda^4}{f}\mathrm{sin}(\frac{\Phi_0}{f})=0,
\end{equation}
while the corresponding Friedmann equation for $H_0$ reads
\beq{}
H^2_0=\frac{8\pi}{3 m^2_{Pl}}(\frac{1}{2}\dot{\Phi}_0^2+\Lambda^4(1-\mathrm{cos} (\frac{\Phi_0}{f}))),\end{equation}
where $m_{Pl}\equiv1/\sqrt{G}=1.22\times10^{19}$GeV is the Plank mass. By these two equations of motion, supplemented with appropriate initial conditions, we will be able to do the integral in eq.(15) very fluently.
\subsection{The value of $\Lambda$ and $f$ from observations}
According to reference \cite{24}, to be consistent with known cosmic-microwave background observations such as WMAP\cite{21}, the Planck\cite{22,23} and the BICEP2\cite{1}, the $\Lambda$ and $f$ parameters in the natural inflation should satisfy that $f\gtrsim m_{_{Pl}}$ and $\Lambda\thicksim m_{\mathrm{GUT}}\thicksim 10^{16}$GeV. Although the natural inflation in this paper is implemented with complex scalar fields, the parameter determination logic could be adapted from \cite{24} routinely.
\begin{itemize}
\item Constraints from the density perturbation spectrum index $n_s$. Both in real and complex scalar field, we can derive that
\beq{}
n_s=1-\frac{m^2_{Pl}}{8\pi f^2}.
\end{equation}
While according to reference \cite{23}, the observation value of $n_s\approx0.96$. This means that in the complex scalar field natural inflation model, $f\approx m_{_{Pl}}$
\item Constraints from the tensor-to-scalar(perturbation amplitudes) ratio. Theoretical considerations \cite{24} require
\beq{}
V_H=(2.2\times 10^{16}\mathrm{GeV})^4\,\frac{r}{0.2},
\end{equation}
for natural inflation models $V_H=2\Lambda^4$. According to the observation of BICEP2 $r=0.20^{+0.07}_{-0.05}$. So $\Lambda\approx 10^{16}\mathrm{GeV}$ is very normal choices
\end{itemize}
According to reference \cite{25}, the number of inflation e-foldings to solve the flatness- and horizon-problem of non-inflation cosmologies also implies constraints on the choice of model parameters $\Lambda$ and $f$. Its basic logic is as follows.
Firstly, according to the slow-roll scenario, the number of inflation e-foldings reads
\bea{}
N_e=\ln(\frac{a_2}{a_1})=\int^{t_2}_{t_1}H\,dt=\frac{8\pi}{m_{Pl}^2}\int^{\phi_1}_{\phi_2}\frac{V(\phi)}{V'(\phi)}d\phi
\\
=\frac{8\pi f^2}{m_{Pl}^2}\ln\left[\frac{1+\cos(|\phi_2|/f)}{1+\cos(|\phi_1|/f)}\right],
\nonumber
\end{eqnarray}
where $a_1$ and $\phi_1$ are initial values when the inflation begins, $a_2$ and $\phi_2$ are the values at the end of inflation, and $V'$ denotes $dV/d\phi$.
Secondly, to associate $N_e$ with $f$, define a ``possibility'' $P(f)$ to quantify whether a given $f$ value is likely to generate sufficient inflation that $N_e\approx 60$,
\beq{}
P(f)=\frac{\pi f-\phi_{min}(f)}{\pi f},
\end{equation}
where $\phi_{min}(f)$ is the minimal value of $|\phi_1|$ that getting $N_e\geq60$ for a given $f$. Obviously, as long as $\phi_{min}$ can drive
sufficient inflation, all values of $|\phi_1|\in(\phi_{min},\pi f)$ will yield $N_e>60$, as shown in Fig \ref{figVphiPf}.
\begin{center}
\begin{figure}[ht]
\includegraphics[width=7cm]{fig1potential.eps}
\includegraphics[width=8cm]{fig1Pf.eps}
\caption{The top figure illuminates the shape of $V(\phi)$. It's clear that a higher potential generates a larger e-folding number, so
$|\phi_1|\in(\phi_{min},\pi f)$ would get $N_e>60$. The button figure shows the numeric feature of $P(f)$, from which we easily see that the larger is $f$, the more closer $P(f)\rightarrow1$.}
\label{figVphiPf}
\end{figure}
\end{center}
Thirdly, using the slow-roll parameter definition
\beq{}
\epsilon=\frac{m_{Pl}^2V'(\phi)^2}{16\pi V(\phi)^2}=\frac{m_{Pl}^2}{16\pi f^2}\left[\frac{\sin(\phi/f)}{1-\cos(\phi/f)}\right]^2,
\end{equation}
and the inflation ending condition $\epsilon\approx1$, we can get the $\phi$ field value at the inflation ending:
\beq{}
\phi_2(f)=f\,\arccos\left[\frac{16\pi f^2-m_{Pl}^2}{16\pi f^2+m_{Pl}^2}\right].
\end{equation}
Finally, combining eqs.(20)(21)(23) we can exactly work out $P(f)$ and the result is shown in Fig \ref{figVphiPf}. While from the definition of $P(f)$, we know that to get enough number of inflation e-foldings, $P(f)$ should be as close as possible to $1$. For $f=m_{Pl}$, we get $P(f)=0.194$, which is grudgingly in the desired range. More large values of $f$ will generate more sufficient inflation.
In following calculations, we will set $f=m_{Pl}$ and $\Lambda=10^{16}$GeV when necessary.
\subsection{Dimensionless representation}
Since $\Delta N_\phi$ is proportional to the size of the expanding universe, it is not a good quantity for numerics, even though it is dimensionless. The more appropriate quantity measuring the baryon asymmetry is
\beq{}
\alpha\equiv \frac{\Delta N_\phi}{N_{tot}}=\frac{\Delta n_\phi}{n_\phi+n_{\bar{\phi}}},
\end{equation}
where $N_{tot}$ is the total number of $\phi$ and $\bar{\phi}$ particles and $n=N/V_{com}a^3$ stands for particle number densities. After the inflation finishes, but before the decay of $\phi$ particles into baryons, all energies that fill the universe is stored in the non-relativistic $\phi$ particles. So we have
\beq{}
m_\phi(n_\phi+n_{\bar{\phi}})=\varepsilon_0=\frac{1}{2}\dot{\Phi}_0^2+\Lambda^4(1-\mathrm{cos} (\frac{\Phi_0}{f})),
\end{equation}
where $m_\phi=\frac{\Lambda^2}{f}$ is the mass of $\phi$ particle and $\varepsilon_0$ is the energy density of the universe. From the above two equations, we can derive
\beq{}
\alpha=\frac{m_\phi \Delta n_\phi}{\varepsilon_0} .
\end{equation}
To get further dimensionless representation for $\alpha$, we introduce the following dimensionless quantities:
\beq{}
\tau\equiv m_\phi t=\frac{\Lambda^2 t}{f},~~\tilde{\Phi}\equiv \frac{\Phi_0}{f},~~\tilde{H}\equiv\frac{H_0}{m_\phi}=\frac{fH_0}{\Lambda^2},
\end{equation}
and write
\beq{}
\alpha=-\frac{\lambda\,f^n}{\Lambda^4}\,\mathrm{sin}(n\,\theta_i)\,A_n(\tau_i,\tau_f),
\label{alphabyAn}
\end{equation}
where $\tau$ and $\tilde{\Phi}$ are dimensionless time and field variables respectively, while
\beq{}
A_n(\tau_i,\tau_f)=\frac{4\,n\,\int^{\tau_f}_{\tau_i}d\tau\,a_0(\tau)^3\tilde{\Phi}(\tau)^n}{a_0(\tau_f)^3(\frac{1}{2}\dot{\tilde{\Phi}}(\tau_f)^2+1-\mathrm{cos}(\tilde{\Phi}(\tau_f)))}.
\label{Antitf}
\end{equation}
By numerically solving the dimensionless version of equations (16) and (17), we will obtain the time dependence of $\tilde{\Phi}_0(\tau)$, $a_0(\tau)$ very easily, see FIG \ref{figphita0} for references.
\begin{center}
\begin{figure}[ht]
\includegraphics[width=7.5cm]{fig2phit.eps}
\includegraphics[width=7.5cm]{fig2a0.eps}
\caption{
The evolution of the dimensionless field variable $\tilde{\Phi}$ and scale factor $a_0$.
In this plot, the initial conditions were set to $\tilde{\Phi}=2.5$, $f=m_{_{Pl}}$ and $\Lambda=10^{16}$GeV to implement the e-folding number $\sim$60. Without loss of generality, we set $\dot{\tilde{\Phi}}=0$ and $a_i=1$. The right figure shows that indeed about 60 e-folding numbers are generated.}
\label{figphita0}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[ht]
\includegraphics[width=7.5cm]{fig3aphiCubic.eps}
\includegraphics[width=7.5cm]{fig3IntaphiCubic.eps}
\caption{
The top is the integrand in eq.(29), which is proportional to the production rate of the net particle number. The button is the integrated value, expresses the whole net particle number produced till $\tau_f$. We can easily see that the early and late time contributions do not significantly affect the final net particle number. Almost all of the net particles are produced during a very short time around $\tau=20$.
}\label{figa3phi3related}
\end{figure}
\end{center}
From FIG \ref{figphita0}, we can easily see that at the beginning of inflation, $a_0(\tau_i)$ is very small(relative to $a_0(t_f)$). As results, in the integration (29) contributions from the early time are negligible. While at the matter dominating era marked by $\tau_f$, $\tilde{\Phi}$ is evolving to almost zero. So integrations from that period also contribute little to $A_n$, see FIG \ref{figa3phi3related} for quantitative references. From the figure, it's easy to see that $A_n$ is totally determined by the ``middle'' area of the integrand. Under the limit that $\tau_i\rightarrow 0$ and $\tau_f\rightarrow \infty$, $A_n$ can be looked as a constant which depends only on the lower index $n$;
\beq{}
A_n(\tau_i\rightarrow 0,\tau_f\rightarrow\infty)\equiv c_n.
\end{equation}
As examples, we numerically compute this parameter when $n=3,4,\cdots,10$, the result is as follows
\beq{}
\begin{array}{c}
c_3\approx8.0,~c_4\approx3.7,~c_5\approx1.3,~c_6\approx0.56
\\
c_7\approx0.25,~c_8\approx0.12,~c_9\approx0.060,~c_{10}\approx0.031
\end{array}
\end{equation}
Substituting this results into equations \eqref{alphabyAn}-\eqref{Antitf}, we will get the final expression for the dimensionless baryon asymmetry parameter as following
\beq{}
\alpha=-c_n\,\frac{\lambda\,f^n}{\Lambda^4}\,\mathrm{sin}(n\,\theta_i).
\label{alphaFourth}
\end{equation}
Obviously, for some special values of the initial angle $\theta_i$, for
instance $\theta_i=\pi$, the factor $\sin(n\,\theta_i)$ vanishes. In such cases, no baryon asymmetry is generated. Such special values can generate
large isocurvature fluctuations \cite{17}. But that's not our main goal, we will set $\theta_i$ to be general values so that $|\sin(n\,\theta_i)|\approx 1$.\\
Now we have implemented the goal of generating net inflaton $\phi$-particles from a symmetric initial conditions in natural inflations. Our next goal is transferring the $\phi$-particles into baryons, and associating $\alpha$ to the observable baryon-to-photon ratio $\eta$ in the next section.
\section{$\phi$-particles decay into baryons}
According to inflationary theory, at the late time of inflation, the inflaton field oscillates near the minimum of its effective potential and gradually
decays into standard model particles\cite{27}. This stage of the early universe is called ``reheating''. Almost all elementary particles populating
the universe are created during reheating, and these particles interact with each other and finally come to a state of thermal
equilibrium at a temperature $T_r$, which is called reheating temperature.
Now we assume that each $\phi$ particle carries a baryon number $B$, and it will decay into baryons through a process that conserves the
baryon number during the stage of reheating. We assume that all the
subsequent interactions also conserve the baryon number, so we have
\beq{}
(N_b-N_{\bar{b}})_f=B(N_\phi-N_{\bar{\phi}})_i,
\end{equation}
where the lower index $f$ means a final time on which reheating finishes, and $i$ stands for an initial time on which all the energy stored in the inflaton field is translated into $\phi$ particles, but $\phi$'s decay does not begin. By these symbols, we can write down $\eta$ in the following form:
\beq{}
\eta=\frac{(N_b-N_{\bar{b}})_f}{(N_{\gamma})_f}=B\frac{(N_\phi-N_{\bar{\phi}})_i}{(N_{\gamma})_f}=\alpha B\frac{(N_\phi+N_{\bar{\phi}})_i}{(N_{\gamma})_f}.
\label{etaFirst}
\end{equation}
Obviously, to calculate $\eta$, we need to work out the initial total number of $\phi$ particles and the photon number at late times.
At initial times, all the energy congesting the universe is provided by $\phi$ particles, so
\beq{}
m_\phi(N_\phi+N_{\bar{\phi}})_i=\frac{\Lambda^2}{f}(N_\phi+N_{\bar{\phi}})_i=V_{com}(a^3\varepsilon)_i .
\end{equation}
Using Friedmann equation, we can relate the energy density to the Hubble parameter as
\beq{}
(\varepsilon)_i=\frac{3m_{Pl}^2}{8\pi}(H^2)_i .
\end{equation}
While on the number of photons at late times, we can relate it with the temperature:
\beq{}
(N_\gamma)_f=V_{com}(a^3 n_\gamma)_f=V_{com}\frac{2\zeta(3)}{\pi^2}(a^3T^3)_f ,
\end{equation}
where $\zeta(3)\approx1.202$ is the so called Ap\'{e}ry's constant. Using this two results, we can rewrite the ratio $\eta$ in equation \eqref{etaFirst} as follows:
\beq{}
\eta=\frac{3\pi B\alpha}{16\zeta(3)}\frac{m_{Pl}^2 f}{\Lambda^2}\frac{(a^3H^2)_i}{(a^3T^3)_f}.
\label{etaSecond}
\end{equation}
Since we are not going to compute the result by detailed decaying processes,
we need to assume the decay of $\phi$ particles and the subsequent process of thermalization occurs very fast. So we can set both $(a^3H^2)_i$ and $(a^3T^3)_f$ to be values around the end of reheating and get an approximation for the final result. We insert an $\mathcal{O}(1)$ factor $\beta$ to account for the deviation caused by this assumption:
\beq{}
\eta=\frac{3\beta\pi B\alpha}{16\zeta(3)}\frac{m_{Pl}^2 f}{\Lambda^2}\frac{H^2_r}{T^3_r}.
\label{etaThird}
\end{equation}
The stage of reheating ends when Hubble parameter becomes smaller than the decay rate of $\phi$,
$H\lesssim \Gamma_\phi$\cite{27}. And the reheating temperature can be estimated by\cite{26}: $T_r\approx0.2\sqrt{\Gamma_\phi m_{Pl}}$.
Substituting these two relations to the approximate expression of $\eta$, we obtain:
\beq{}
\eta=\frac{3\pi\beta B\alpha}{0.2^3\times16\zeta(3)}\frac{m_{Pl}^{1/2} \Gamma_\phi^{1/2} f}{\Lambda^2}.
\label{etaFourth}
\end{equation}
Inserting the $\alpha$ expression \eqref{alphaFourth} obtained in the previous section into this equation, we will get our result for $\eta$:
\beq{}
\eta=-c_n\frac{3\pi\beta B\lambda}{0.2^3\times16\zeta(3)}\frac{m_{Pl}^{1/2} \Gamma_\phi^{1/2} f^{n+1}}{\Lambda^6}\sin(n\theta_i).
\label{etaFifth}
\end{equation}
Before further discussion of physical features of this expression for $\eta$, we should first determine the range of the symmetry breaking parameter $\lambda$. Since we assumed that the symmetry breaking term in the potential of $\phi$ is subdominant during the
inflation to ensure the feature of the natural inflation, we have to impose constraint
\beq{}
\lambda(\phi^n_i+\phi^{\ast n}_i)=\lambda \Phi^n_i \cos(n\,\theta_i)\ll \Lambda^4(1-\cos(\frac{\Phi_i}{f})),
\end{equation}
While to assure this constraint holds for all the possible value of $\theta_i$, the value of $\lambda$ has to be limited from the upper bound
\beq{}
\lambda\ll\lambda_{max}=\frac{\Lambda^4(1-\cos({\Phi_i/f}))}{\Phi_i^n}.
\end{equation}
With this constraint, we can test whether our result is physical acceptable.
Now we use the boundary value of $\lambda$ to work out the required $\Gamma_\phi$ for generating the observed $\eta\approx 6\times 10^{-10}$.
The expression of $\Gamma_{\phi,\mathrm{req}}$ can be derived from eq.(41):
\bea{}
\Gamma_{\phi, \mathrm{req}}\approx c_n^{-2}(\frac{\lambda_{max}}{\lambda})^2\frac{\eta^2(0.2^3\times16\zeta(3)^2)}{(3\pi)^2}\times
\nonumber\\
\frac{\Lambda^4f^{-2n-2}\Phi^{2n}_i}{m_{Pl}(1-\cos(\frac{\Phi_i}{f}))^2}(\beta B|\sin(n\theta_i)|)^{-2}.
\end{eqnarray}
In the previous sections, we have set $f=m_{Pl}$, $\Lambda=10^{16} \mathrm{GeV}$ and $\Phi_i=2.5 f$, so the expression can be simplified to:
\beq{}
\Gamma_{\phi, \mathrm{req}}\approx1.6\times 10^{-7} \mathrm{eV}\times 2.5^{2n}c_n^{-2}(\frac{\lambda_{max}}{\lambda})^2(\beta B|\sin(n\theta_i)|)^{-2}.
\label{GammaphiFirst}
\end{equation}
If we set an appropriate value for $\lambda$, for example, $\lambda=\frac{1}{10}\lambda_{max}$, and assume that $\beta B |\sin(n\theta_i)|\approx1$,
we will get the required decay rate for different $n$s. Our results, and those from \cite{17} with a magnitude correction for comparisons, are shown as follows:
\begin{eqnarray*}
\raisebox{3mm}{n~~~~} &\stackrel{\displaystyle\Gamma_{\phi,\mathrm{req}}~\mathrm{in}}{\displaystyle\mathrm{natural~inflation}} & \stackrel{\Gamma_{\phi,\mathrm{req}}~\mathrm{in}}{\mathrm{chaotic~inflation}}\\
n=3 & 6.1\times 10^{-5}\mathrm{eV} & ~~~~~~~~4\times 10^{-3}\mathrm{eV}\\
n=4 & 1.8\times 10^{-3}\mathrm{eV} & ~~~~~~~~2\times 10^{-1}\mathrm{eV}\\
n=5 & 9.0\times 10^{-2}\mathrm{eV} & ~~~~~~~~10^{1}\mathrm{eV}\\
n=6 & 3.0 ~\mathrm{eV} & ~~~~~~~~6\times10^2~\mathrm{eV}\\
n=7 & 9.5\times10 ~\mathrm{eV} & ~~~~~~~~2\times 10^4 \mathrm{eV}\\
n=8 & 2.6\times 10^3\mathrm{eV} & ~~~~~~~~9\times 10^5 \mathrm{eV}\\
n=9 & 6.5\times 10^4\mathrm{eV} & ~~~~~~~~3\times 10^7 \mathrm{eV}\\
n=10 & 1.5\times 10^6\mathrm{eV} & ~~~~~~~~10^9 \mathrm{eV}
\end{eqnarray*}
Obviously, larger power $n$ of symmetry breaking interaction requires larger decay width to give desired photon-baryon-ratio. While from derivations \eqref{etaFifth}-\eqref{GammaphiFirst}, we know that $\Gamma_\phi\propto\Lambda^4$, that is, higher energy scale of inflation requires larger width of inflaton decays, otherwise the theoretical photon-baryon-ratio will deviate remarkably from expectations.
With these results of $\Gamma_{\phi,\mathrm{req}}$, we can work out the corresponding reheating temperature by the
relation $T_r\approx 0.2\sqrt{\Gamma_\phi m_{Pl}}$. An important
condition is that the reheating temperature must be higher than the typical temperature of big bang nucleosynthesis $\sim$MeV.
For the lowest value of $\Gamma_{\phi,\mathrm{req}}$ in natural inflation, when $n=3$, the reheating temperature $T_r\approx 173$GeV, and for $n=10$, the corresponding $T_r\approx 2.7\times 10^7$GeV. All the $T_r$s in our model are much higher than MeV, this is obviously consistent
with the big bang nucleosynthesis, thus physically acceptable.
For all values of $n$ listed above, $\Gamma_{\phi,\mathrm{req}}$ in natural inflations is smaller than that in chaotic inflations, and the growth of $\Gamma_{\phi,\mathrm{req}}$ with the increasing of $n$ is slower than in chaotic inflation. These differences may be used to distinguish this two models in futures.
\section{Summary and discussion}
In this paper, we apply a variation of Affleck-Dine mechanism into natural inflation scenarios and generate the observed baryon-to-photon ratio. In this mechanism, the process of baryon asymmetry generation is unified with the stage of inflation and reheating. The baryon asymmetry is firstly implemented using the inflaton field with a weakly broken global $U(1)$ symmetry. It is in the second stage that the net inflaton $\phi$-particles decay into standard model particles.
By numerical calculations, we work out parameter $\alpha$ describing the asymmetric evolution of $\phi-\bar{\phi}$ particles during the natural inflation era and derive out formulas relating it with the baryon-to-photon ratio $\eta$. We calculate the decaying rate of $\phi$-particles required to generate the observed $\eta\approx 6\times10^{10}$. It is observed that
the reheating temperatures in this inflation model is much higher than the desired temperature of big bang nucleosynthesis. From this aspect, this model is physically acceptable. We also compare our results with those in chaotic inflation models. The differences between the two may be useful
for future distinguishing of them through observations.
As discussions, we note that parameter resonance phenomena\cite{26}, superheavy fermions production\cite{28}, detailed particle physics model implementation, dark matter particles formation and properties \textit{et al}. in this natural inflation + reheating mechanism are all interesting future directions.
\section*{Acknowledgments}
We thank very much to Wang Qian-jun, Meng Sun, Jian-feng Wu and Prof. Yong-chang Huang for meaningful discussions. This work is supported by Beijing Municipal Natural Science Foundation, Grant. No. Z2006015201001.
|
\section{Introduction}
Heat transport has attracted increasing interest recently in both
classical nonlinear lattices and quantum
systems\cite{NLi12,Lepri,Dhar,Liu13,Zotos}. In this field, a
particularly interesting problem is related to the issue of
Fourier's law. Considering for example the normal transport in
one-dimensional systems, Fourier's law states that
$j(x,t)=-\kappa\partial_xT(x,t)$, where $j(x,t)$ is the local heat
current, $T(x,t)$ is the local equilibrium temperature, and $\kappa$
is the heat conductivity. If we let $\varepsilon(x,t)$ denote the
local energy density, then the continuity equation reads
$\partial_t\varepsilon(x,t)+\partial_xj(x,t)=0$. Combining these
equations with $\frac{\partial\varepsilon}{\partial T}=c$, where $c$
is the specific heat per unit volume, we arrive at the energy
diffusion equation,
$\partial_t\varepsilon(x,t)=\frac{\kappa}{c}\partial^2_x\varepsilon(x,t)$.
In classical systems, it was shown that normal diffusion can be
characterized by the mean squared displacement of the Helfand moment
\cite{Helfand,Viscardy}, which is related to the autocorrelation
function of heat current and thus to the Green-Kubo formula. Beyond
the normal diffusion, some recent works have investigated the
relation between heat diffusion and conduction
\cite{Denisov,Li03,Zhao,Zaburdaev,Dhar13,YLi}. In particular, a
rigorous relationship between energy (heat) spread and heat
conduction has been established from statistical
principles\cite{Liu14}. Therein, an excess energy distribution is
introduced, and then the energy diffusion is characterized by the
mean square deviation (MSD) of energy, which is connected to the
autocorrelation function of heat current. Accordingly, how thermal
conductivity depends on the system size may be extracted from energy
diffusion in lattice systems \cite{Zhao,Cipriani,NLi}.
Simultaneously, heat transport in low-dimensional quantum systems
has also been investigated intensively
\cite{Zotos96,Zotos99,Narozhny,Meisner,Benz,Karrasch,Zotos04,Jung,Karrasch13,Znidaric}.
A commonly used method is the Green-Kubo formula within linear
response theory, where nonzero Drude weights usually indicate
ballistic transport. An interesting example is the ballistic energy
transport in the spin-$\frac{1}{2}$ XXZ chain due to the
conservation of the current operator
\cite{Zotos97,Klumper,Karrasch14}. Besides, quantum quench dynamics
or spreading of different densities (e.g., energy densities) has
also been studied\cite{Michel,Langer,Steinigeweg09}. To determine
whether the spread process is ballistical, diffusive, or of other
type, one can observe the time evolution of the spatial variance
$\sigma^2$ (or MSD) of certain nonequilibrium density. For ballistic
transport, the variance behaviors as $\sigma^2\sim t^2$ whereas for
diffusive transport $\sigma^2\sim t$. In Ref.
\onlinecite{Steinigeweg09}, a connection between the variance and
the current-autocorrelation function was proposed; however, it is
applicable at high temperatures. The general connection between the
spreading processes and transport properties such as heat
conductivity is not well understood yet.
In this paper, starting from an energy density distribution, we give
a general connection between the MSD of energy diffusion and the
autocorrelation function of energy current for quantum systems,
within the linear response theory. This offers a different way to
extract thermal conductivity from the energy spreading process. As
an example, we apply it to a spinless fermion model, and the
numerical results confirm this connection.
\section{Connection between MSD and current-autocorrelation function}
In the following, we restrict to the one-dimensional case. The
generalization to higher-dimensional systems is straightforward. The
system is typically described by a continuous Hamiltonian:
\begin{eqnarray}
H_0=\int h(x)dx.
\end{eqnarray}
At the infinite past an additional perturbation, $H'=-\int
\eta(x)h(x)dx$, is also applied to the system. Here $\eta(x)$ is
nonzero only in a local region. Thus the total Hamiltonian reads
$H=H_0+H'$. Before $t=0$, we suppose the system is described by a
canonical ensemble at temperature $T$. Then the partition function
is $Z={\rm\,Tr}( e^{-\beta H})$, where $\beta=1/k_BT$. At time $t=0$, the
perturbation is turned off suddenly. After that the quenched initial
nonequilibrium state begins to relax towards the equilibrium state,
and so does the local energy distribution. The local excess energy
at $t>0$ can be described by
\begin{eqnarray}
\delta\langle h(x,t)\rangle_{neq} \equiv \langle h(x,t)\rangle_{neq}
-\langle h(x)\rangle, \label{Eq_deltaH}
\end{eqnarray}
where $h(x,t)=e^{iH_0t/\hbar}h(x)e^{-iH_0t/\hbar}$. $\langle \cdot
\rangle_{neq}$ denotes the expectation value in the nonequilibrium
state, i.e., $\langle \cdot \rangle_{neq}={\rm\,Tr}(e^{-\beta H}\cdot)/Z$,
and $\langle \cdot \rangle$ denotes the equilibrium average,
$\langle \cdot \rangle={\rm\,Tr}(\rho_0\cdot)$ with $\rho_0=e^{-\beta
H_0}/{\rm\,Tr}(e^{-\beta H_0})$.
To evaluate $\langle h(x,t)\rangle_{neq}$,
we consider an operator
$U(\tau)=e^{H_0\tau/\hbar}e^{-H\tau/\hbar}$. The equation of motion
of $U(\tau)$ is
\begin{eqnarray}
-\hbar\frac{\partial U(\tau)}{\partial\tau}=H'(\tau)U(\tau),
\end{eqnarray}
where $H'(\tau)=e^{H_0\tau/\hbar}H'e^{-H_0\tau/\hbar}$. To the first
order of $H'$, the solution can be written as
$U(\tau)\approx1-\frac{1}{\hbar}\int_0^{\tau}d\tau'H'(\tau')$. Thus
we may have
\begin{eqnarray}
e^{-\beta H}&=&e^{-\beta H_0}U(\hbar\beta) \nonumber \\
&\approx&e^{-\beta H_0}-\frac{1}{\hbar}e^{-\beta
H_0}\int_0^{\hbar\beta}d\tau H'(\tau). \label{Eq_betaH}
\end{eqnarray}
Then we can obtain the partition function to the first order of
$H'$,
\begin{eqnarray}
Z/Z_0\approx1-\frac{1}{\hbar}\langle\int_0^{\hbar\beta}d\tau
H'(\tau)\rangle.\label{Eq_Z}
\end{eqnarray}
Substituting Eqs.~(\ref{Eq_betaH}) and (\ref{Eq_Z}) into
Eq.~(\ref{Eq_deltaH}), we can obtain after some algebra,
\begin{eqnarray}
\delta\langle h(x,t)\rangle_{neq}
&=&-\frac{1}{\hbar}\langle\int_0^{\hbar\beta}d\tau
H'(\tau)h(x,t)\rangle \nonumber \\
& &+\frac{1}{\hbar}\langle\int_0^{\hbar\beta}d\tau H'(\tau)\rangle
\langle h(x,t)\rangle. \label{delta_h}
\end{eqnarray}
The second term in Eq.~(\ref{delta_h}) is time-independent actually.
The probability distribution function is then defined as
\begin{eqnarray}
\rho_E(x,t)=\delta\langle h(x,t)\rangle_{neq}/\mathcal{N},
\end{eqnarray}
where $\mathcal{N}=\int dx \delta\langle h(x,t)\rangle_{neq}=Tc\int
dx'\eta(x')$ is a normalization constant; see appendix
\ref{appendixA}.
The mean square deviation for energy spread is then
\begin{eqnarray}
\langle \Delta x^2(t)\rangle_E &\equiv&\int(x-\langle
x\rangle_E)^2\rho_E(x,t)dx \nonumber \\
&=&\langle x^2(t)\rangle_E-\langle x\rangle^2_E.
\end{eqnarray}
Using the reasoning similar to Ref. \onlinecite{Liu14}, it can be
shown that $\langle x\rangle_E$ is a constant. For later
convenience, we introduce two correlation functions:
\begin{eqnarray}
C_{jj}(x't',xt)&=&\langle\int_0^{\hbar\beta}d\tau
j(x',t'-i\tau)j(x,t)\rangle
\end{eqnarray}
and
\begin{eqnarray}
C_{hh}(x't',xt)&=&\langle\int_0^{\hbar\beta}d\tau
h(x',t'-i\tau)h(x,t)\rangle,
\end{eqnarray}
where the current operator $j(x,t)$ is defined via the continuity
equation, $\partial_t h(x,t)+\partial_x j(x,t)=0$. For homogeneous
systems, these correlation functions are invariant under both
temporal translation and spacial translation, i.e.,
$C_{jj}(x't',xt)=C_{jj}(x-x',t-t')$; a similar relation holds for
$C_{hh}$. Further, we can have
$\partial^2_tC_{hh}(x't',xt)=\partial^2_xC_{jj}(x't',xt)$\cite{Liu14}.
Making use of the above equations, we can obtain
\begin{eqnarray}
\mathcal{N}\frac{d^2\langle x^2(t)\rangle_E}{dt^2}
&=&\frac{1}{\hbar}\int\! dx\int\! dx' x^2
\frac{d^2C_{jj}(x'0,xt)}{dx^2}\eta(x') \nonumber \\
&=&\frac{1}{\hbar}\int\! dx x^2 \frac{d^2C_{jj}(x,t)}{dx^2}\!\int\!
dx'\eta(x').
\end{eqnarray}
That means
\begin{eqnarray}
\frac{d^2\langle x^2(t)\rangle_E}{dt^2} =\frac{1}{\hbar Tc}\int\! dx
x^2 \frac{d^2C_{jj}(x,t)}{dx^2}. \label{connect}
\end{eqnarray}
Eq.~(\ref{connect}) is a rigorous result. Integrating by parts twice
and neglecting the boundary terms, we obtain the final result:
\begin{eqnarray}
\frac{d^2\langle x^2(t)\rangle_E}{dt^2}=\frac{2C_{JJ}(t)}{cT},
\label{central_eq}
\end{eqnarray}
where $ C_{JJ}(t)=\lim_{L\rightarrow \infty}\frac{1}{L}
\int_0^{\beta}d\lambda\langle J(-i\lambda\hbar)J(t)\rangle$ is the
current-current correlation function that appears in the Green-Kubo
formula for heat conductivity; see appendix \ref{appendixB}. Here
$J=\int dx j(x)$ and $L$ is the length of the system. It should be
pointed out that taking the limit $L\rightarrow \infty$ is
necessary, because in systems with a finite size the autocorrelation
function $C_{jj}(x,t)$ at low temperatures may not decay to zero as
$x$ approaches to boundaries (see the example below). In that case,
the boundary terms such as $C_{jj}(x,t)x|^{L/2}_{-L/2}$ need to be
taken into account explicitly.
It is straightforward to extend Eq.~(\ref{central_eq}) to other
conserved quantities of the form $\hat{Q}=\int dx \hat{q}(x)$,
\begin{eqnarray}
\frac{d^2\langle x^2(t)\rangle_Q}{dt^2}=\frac{2C_{J_qJ_q}(t)}{\beta
\sigma^2_Q}, \label{central_eq2}
\end{eqnarray}
where $\langle x^2(t)\rangle_Q$ is defined through the distribution
$\delta\langle \hat{q}(x,t)\rangle_{neq}$, and $\sigma^2_Q=(\langle
\hat{Q}^2\rangle-\langle \hat{Q}\rangle^2)/L$ is the fluctuation of
quantity $Q$. The total current $\hat{J}_q$ is given by
$\hat{J}_q=\int dx \hat{j}_q(x)$, and $\hat{j}_q(x)$ is defined via
the continuity equation $\partial_t \hat{q}(x,t)+\partial_x
\hat{j}_q(x,t)=0$. According to the time evolution of $\langle
x^2(t)\rangle_Q$, transport processes may be classified as diffusive
($\langle x^2(t)\rangle_Q\sim t^{\beta}$, $\beta=1$),
super-diffusive ($1<\beta\leq2$), and sub-diffusive ($0<\beta<1$).
\section{An example: spinless fermion model}
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{fig1}
\caption{(Color online) (a) The first derivative of the MSD with
respect to time $d\langle x^2(t)\rangle_E/dt$ as a function of
time. (b) Comparison of $\frac{d^2\langle x^2(t)\rangle_E}{dt^2}$
and $2C_{JJ}(t)/cT$. The parameters are $t_0=1$ and $T=0.1$, and we
choose $k_B=\hbar=1$. The system size is $L=100$. } \label{fig1}
\end{figure}
As an application of Eq.~(\ref{central_eq}), we consider a
noninteracting fermion model, which may also be viewed as a
spin-$\frac{1}{2}$ XY chain. For complicated systems, one may resort
to methods such as finite-temperature, real-time density matrix
renormalization group \cite{Karrasch}. The Hamiltonian we consider
reads
\begin{eqnarray}
H_0=-t_0\sum_{i=-L/2+1}^{L/2} (c^{\dagger}_ic_{i+1}+\mathrm{H.c.})
\equiv \sum_{i}h_i,
\end{eqnarray}
where $h_i$ is the local energy operator. The total size of the
system is $L$, and we adopt periodic boundary conditions. Via the
continuity equation, the current operator can be shown to be
$j_i=\frac{it_0^2}{\hbar}(c^{\dagger}_{i-1}c_{i+1}-c^{\dagger}_{i+1}c_{i-1})$.
The total energy current operator $J=\sum_i j_i$ commutes with
$H_0$, and thus is conserved. Note that usually energy current is
different from heat current \cite{Mahan}. However in the following
we set the chemical potential to zero, so these two currents are the
same in our case.
We assume a local perturbation, $H'=-\sum_i\eta_ih_i$, where
$\eta_i=0.2$ for $i=0$ and $\eta_i=0$ otherwise. To compute the MSD
of energy diffusion, we first evaluate Eq.~(\ref{delta_h}). In the
basis of single-particle eigenstates of $H_0$
($H_0|\alpha\rangle=\epsilon_{\alpha}|\alpha\rangle$), the operators
can be expressed as
\begin{eqnarray}
h_i(t)=\sum_{\alpha\beta}\langle \alpha|h_i|\beta\rangle
c^{\dagger}_{\alpha}c_{\beta}e^{i(\epsilon_{\alpha}-\epsilon_{\beta})t/\hbar}
\label{exp1}
\end{eqnarray}
and
\begin{eqnarray}
H'(\tau)=\sum_{\alpha\beta}\langle \alpha|H'|\beta\rangle
c^{\dagger}_{\alpha}c_{\beta}e^{(\epsilon_{\alpha}-\epsilon_{\beta})\tau/\hbar},\label{exp2}
\end{eqnarray}
where $c^{\dagger}_{\alpha}$ ($c_{\alpha}$) creates (destroys) a
particle occupying the state $|\alpha\rangle$. Substituting
Eqs.~(\ref{exp1}) and (\ref{exp2}) into Eq.~(\ref{delta_h}), we can
get after some algebra
\begin{eqnarray}
\delta\langle h_i(t)\rangle_{neq} =\sum_{\alpha\beta}\langle
\beta|H'|\alpha\rangle\langle \alpha|h_i|\beta\rangle
e^{i(\epsilon_{\alpha}-\epsilon_{\beta})t/\hbar}
\frac{f_{\alpha}-f_{\beta}}{\epsilon_{\alpha}-\epsilon_{\beta}}, \nonumber \\
\label{delta_hi}
\end{eqnarray}
where $f_{\alpha}=1/(1+e^{\epsilon_{\alpha}/k_BT})$ is the
Fermi-Dirac distribution with the chemical potential being zero and
we have used the identity
${\rm\,Tr}[\rho_0c^{\dagger}_{\beta}c_{\alpha}c^{\dagger}_{\gamma}c_{\delta}]
=\delta_{\alpha \gamma}\delta_{\beta
\delta}f_{\beta}(1-f_{\alpha})+\delta_{\alpha \beta}\delta_{\gamma
\delta}f_{\beta}f_{\gamma}$. The specific heat can be easily
evaluated from $C_V=\frac{\partial E}{\partial T}$, where
$E=\sum_{\alpha}\epsilon_{\alpha}f_{\alpha}$. In a similar way we
can evaluate $C_{JJ}(t)$, and the final result is
\begin{eqnarray}
C_{JJ}(t)&=&-\frac{1}{L}\sum_{\alpha \beta}\langle
\beta|\hat{J}|\alpha\rangle\langle \alpha|\hat{J}|\beta\rangle
\nonumber \\
& &
\times\frac{f_{\alpha}-f_{\beta}}{\epsilon_{\alpha}-\epsilon_{\beta}}
e^{i(\epsilon_{\alpha}-\epsilon_{\beta})t/\hbar}.
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{fig2}
\caption{(Color online) (a) Difference between the MSD of energy
diffusion and $2C_{JJ}(t)/cT$ as a function of the system size. (b)
and (c) $|C_{jj}(0,i)|$ in logarithm scale at time $t=4$. In (b)
the total size is $L=100$ whereas in (c) $L=500$. The temperature
is $T=0.01$. $k_B=\hbar=1$. } \label{fig2}
\end{figure}
In numerical simulations, we take $t_0=1$ as units of energy, and we
set $k_B=\hbar=1$. In Fig.~\ref{fig1}(a), we plot the first
derivative of $\langle x^2(t)\rangle_E$ with respect to time at a
high temperature $T=0.1$. The corresponding real temperature is of
order $10^3$ K, and a finite size $L=100$ is used here. We see that
$d\langle x^2(t)\rangle_E/dt$ linearly increases with time. This is
also reflected in Fig.~\ref{fig1}(b), where $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}$ is clearly a constant. Thus the transport
process is ballistic. To check the validity of
Eq.~(\ref{central_eq}), we plot $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}$ and $2C_{JJ}(t)/cT$ as functions of time in
Fig.~\ref{fig1}(b), and a good agreement can be observed.
At low temperatures, the finite size effect becomes prominent; i.e.,
there could be a big difference between $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}$ and $2C_{JJ}(t)/cT$ when the system size is
not large enough. To characterize this difference, we plot the
relative error
\begin{eqnarray}
\delta=|\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}-\frac{2C_{JJ}(t)}{cT}|/|\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}|
\end{eqnarray}
in Fig.~\ref{fig2}(a). As the size increases, the error $\delta$
decays to zero rapidly. The reason for the big difference at a small
size is that the current autocorrelation function $C_{jj}(x'=0,x=i)$
does not decay to zero at the boundaries. For $L=100$, $C_{jj}(0,i)$
takes an appreciably small value at the boundaries, while
$C_{jj}(0,i)$ becomes very small at the boundaries for $L=500$; see
Figs.~\ref{fig2}(b) and (c). Thus when integrating
Eq.~(\ref{connect}) to get Eq.~(\ref{central_eq}), we can not
neglect the boundary terms for small systems. When the boundary
terms are taken into account, we have found excellent agreement
between $\frac{d^2\langle x^2(t)\rangle_E}{dt^2}$ and
$2C_{JJ}(t)/cT$ plus boundary terms regardless of the system size.
\section{Conclusions}
In summary, within the linear response theory we have established a
connection between the MSD of energy diffusion and the
autocorrelation function of energy current for quantum systems,
i.e., $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}=\frac{2C_{JJ}(t)}{cT}$. It is straightforward
to extend it to other conserved quantities. As an example, we have
applied it to a spinless fermion model (or the spin-$1/2$ XY model).
We found that at high temperatures $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}$ is consistent with $2C_{JJ}(t)/cT$ even for
a comparatively small size $L=100$. However, at low temperatures,
there may be large difference between $\frac{d^2\langle
x^2(t)\rangle_E}{dt^2}$ and $2C_{JJ}(t)/cT$ when the system
size is small due to the ignorance of boundary terms. Indeed when the
boundary terms are included, we could still find excellent agreement
between $\frac{d^2\langle x^2(t)\rangle_E}{dt^2}$ and
$2C_{JJ}(t)/cT$ plus boundary terms regardless of the system size.
This connection thus offers an alternative way to extract
conductivity from the energy spreading process in quantum systems.
\begin{acknowledgments}
The work was supported by NSFC (Grant No. 11204180).
\end{acknowledgments}
|
\section{Introduction}
In the number field context, a prime $M$ is called a Mersenne prime if it is of the form $M = 2^p - 1$ for some prime $p$. The Mersenne primes are among the integers of the form $(1 + a)^m - 1$, where $a, m$ are positive integers. It is a classical result that if $(1 + a)^m - 1$ is a prime for some positive integers $a, m$ with $m \ge 2$, then it is necessary that $a = 1$ and $m = p$ for some prime $p$.
There are many strong analogies between number fields and function fields. We refer the reader to the excellent references \cite{Goss}, \cite{Rosen}, \cite{Thakur-book} for these analogies. The analogous pictures between number fields and function fields are clearly reflected when one considers the analogies between the two couples $(\bZ, \bQ)$ and $(\bF[T], \bF(T))$, where $\bF$ is a finite field. The aim of this article is to search for new analogous phenomena between number fields and function fields. Specifically we will study the notion of Mersenne primes in the Carlitz module context, and relate them to the arithmetic of cyclotomic function fields. We also study the acquaintances of Mersenne primes including Wieferich and non-Wieferich primes in the Carlitz module setting that were introduced by Thakur \cite{Thaur-Wieferich-primes} \cite{Thaur-Fermat-Wilson-congruences}.
Let us now introduce a Carlitz analogue of Mersenne primes. We begin by introducing some basic notation used here.
Let $q = p^s$, where $p$ is a prime and $s$ is a positive integer. Let $\bF_q$ be the finite field of $q$ elements. Let $A = \bF_q[T]$, and let $k = \bF_q(T)$. Let $\tau$ be the mapping defined by $\tau(x) = x^q$, and let $k\langle \tau \rangle$ denote the twisted polynomial ring. Let $C : A \rightarrow k\langle \tau \rangle$ ($a \mapsto C_a$) be the Carlitz module given by $C_T = T + \tau$. Let $R$ be a commutative $k$-algebra. The definition of the Carlitz module $C$ is equivalent to saying that $C_T(a) = Ta + a^q$ for every $a \in R$.
It is known that $C_m(x)$ is analogous to $(1 + x)^m - 1 \in \bZ[x]$. This analogy suggests the following definition: a prime in $A$ is called a \textit{Mersenne prime} if it is of the form $\alpha C_P(1)$, where $P$ is a monic prime in $A$ and $\alpha$ is a unit in $A$.
To draw an analogy between the above notion of Mersenne primes and that of Mersenne primes in the number field context, we prove in Section \ref{Section-Mersenne-primes} a Carlitz module analogue of the classical result in elementary number theory that was mentioned in the first paragraph of this introduction.
Let us now describe the content of the paper. In Section \ref{Section-Mersenne-primes}, we introduce the notions of Mersenne numbers and Mersenne primes in the Carlitz module context. As remarked in \cite{Murata-Pomerance}, it is not known whether there are infinitely many primes $p$ for which the Mersenne numbers $2^p - 1$ are composite. In contrast to the number field setting, we prove in the Carlitz module context that for every $q > 2$, there are infinitely many monic primes $\wp$ in $\bF_q[T]$ such that the Mersenne numbers $C_{\wp}(1)$ are composite.
In Section \ref{Section-Wieferich-primes}, we recall the notions of Wieferich and non-Wieferich primes in the Carlitz module context that were introduced by Dinesh Thakur \cite{Thaur-Wieferich-primes} \cite{Thaur-Fermat-Wilson-congruences}. Theorem \ref{Theorem-A-Mersenne-prime-is-a-non-Wieferich-prime-in-Carlitz-module-setting} shows that every Mersenne prime is a non-Wieferich prime, which is analogous to a similar statement in the number field context.
It is a classical result in elementary number theory that for a given odd prime $p$, every prime $q$ dividing the Mersenne number $M_p := 2^p - 1$ satisfies $q \equiv 1 \pmod{p}$. The classical proof of this result is based on the notion of the order of an element modulo a prime. In Section \ref{Section-The-Carlitz-annihilators-of-primes}, we prove a Carlitz module analogue of this result which states that for a given monic prime $P$ in $A$, every monic prime $Q$ dividing the Mersenne number $M_P := \alpha C_P(1)$ with $\alpha \in \bF_q^{\times}$ satisfies $Q \equiv 1 \pmod{P}$. In order to prove this result, we introduce in Section \ref{Section-The-Carlitz-annihilators-of-primes} a notion of the Carlitz annihilator of a prime that is analogous to that of the order of an element modulo a prime. In the last section, using the arithmetic of cyclotomic function fields, we prove a criterion for determining whether a Mersenne number is prime.
It is worth mentioning that Dinesh Thakur \cite{Thaur-Fermat-Wilson-congruences} found interesting relations linking Wieferich primes in the function field context to zeta values.
\subsection{Notation.}
\label{Subsection-Notation}
In addition to the notation introduced before, let us fix some basic notation and definitions used throughout the paper.
Every nonzero element $m \in A$ can be written in the form $m = \alpha_n T^n + \cdots + \alpha_1 T + \alpha_0$, where the $\alpha_i$ are elements in $\bF_q$ and $\alpha_n \ne 0$. When $m$ is of the form as above, we say that the degree of $m$ is $n$. In notation, we write $\deg(m) = n$. We use the standard convention that $\deg(0) = -\infty$.
For each $m \in A$, define $|m| := q^{\deg(m)}$. Note that $|m|$ is the number of elements of the finite ring $A/mA$. For basic properties of $|\cdot|$, we refer the reader to \cite{Rosen}.
Fix an algebraic closure $\bar{k}$ of $k$, and set
\begin{align*}
\Lambda := \{\lambda \in \bar{k}\; | \; \text{$C_m(\lambda) = 0$ for some nonzero $m \in A$} \}.
\end{align*}
For every nonzero element $m \in A$, define $\Lambda_m := \{\lambda \in \bar{k}\; | \; C_m(\lambda) = 0 \}$. We recall the following definition.
\begin{definition}
Let $m \in A$ be a polynomial of positive degree. The field $K_m = k(\Lambda_m)$ is called a \textit{cyclotomic function field}.
\end{definition}
For each polynomial $m$ in $A$ of positive degree, we define a \textit{primitive $m$-th root of $C_m$} to be a root of $C_m$ that generates the $A$-module $\Lambda_m$. Throughout the paper, for each $m \in A$ of positive degree, we fix a primitive $m$-th root of $C_m$, and denote it by $\lambda_m$. Let $\Phi_m$ be the $m$-th cyclotomic polynomial, that is, the monic irreducible polynomial over $k$ such that $\Phi_m(\lambda_m) = 0$.
For each polynomial $m \in A$ of positive degree, set
\begin{align*}
\cS_m := \{a \in A \; | \; \text{$\gcd(a, m) = 1$ and $0 \le \deg(a) < \deg(m)$}\}.
\end{align*}
Fix an element $m \in A$ of positive degree. For each element $a \in \cS_m$, let $\sigma_m^{(a)}$ be the $k$-automorphism of $K_m$ defined by $\sigma_m^{(a)}(\lambda_m) = C_a(\lambda_m)$. Let $\G_m$ denote the Galois group of $K_m$ over $k$. It is well-known \cite{Hayes} that $\G_m := \mathrm{Gal}(K_m/k) = \{\sigma_m^{(a)} \; | \; a \in \cS_m \}$.
\section{A Carlitz module analogue of Mersenne primes}
\label{Section-Mersenne-primes}
Let $m \in A$ be a monic polynomial of degree $d \ge 1$. By \cite[Proposition 12.11]{Rosen}, we can write $C_m(x) \in A[x]$ in the form
\begin{align}
\label{Equation-The-equation-of-Cm-x}
C_m(x) = mx + [m, 1]x^q + [m, 2]x^{q^2} + \cdots + [m, d - 1]x^{q^{d- 1}} + x^{q^d} \in A[x],
\end{align}
and hence
\begin{align*}
\dfrac{C_m(x)}{x} = m + [m, 1]x^{q - 1} + [m, 2]x^{q^2 - 1} + \cdots + [m, d - 1]x^{q^{d- 1} - 1} + x^{q^d - 1} \in A[x].
\end{align*}
It is well-known that $C_m(x) = \prod_{\substack{a | m \\ \text{$a$ monic}}}\Phi_a(x)$, where $\Phi_a(x) \in A[x]$ is the $a$-th cyclotomic polynomial.
If $m = P$ for some monic irreducible polynomial $P$, then it is well-known \cite{Hayes} that $\Phi_P(x) = C_P(x)/x$ is an Eisenstein polynomial at $P$, i.e., the polynomial $[m, i]$ is congruent to zero modulo $P$ for each $1 \le i \le d - 1$, and $m$ is divisible by $P$ but not divisible by $P^2$.
From the discussion above, we see that $C_m(x)$ is analogous to the polynomial $(1 + x)^m - 1 \in \bZ[x]$ in the classical cyclotomic theory. We now prove a lemma that naturally motivates the notions of Mersenne numbers and Mersenne primes in the Carlitz module setting.
\begin{lemma}
\label{Lemma-The-1st-lemma-that-motivates-the-definition-of-Mersenne-primes}
Let $m \in A$ be a monic polynomial of degree $d \ge 1$, and let $\alpha \in A$ be a polynomial of degree $h \ge 0$. Assume that $q \ge 3$ and $C_m(\alpha)$ is a monic prime in $A$. Then $m$ is a monic prime in $A$ and $\alpha$ belongs to $\bF_q^{\times}$.
\end{lemma}
\begin{proof}
By $(\ref{Equation-The-equation-of-Cm-x})$, we know that $C_m(\alpha) = \alpha Q$, where
\begin{align}
\label{Equation-The-equation-of-Q-in-motivation-lemma-for-Mersenne-primes}
Q = m + [m, 1]\alpha^{q - 1} + [m, 2]\alpha^{q^2 - 1} + \cdots + [m, d - 1]\alpha^{q^{d - 1} - 1} + \alpha^{q^d - 1} \in A.
\end{align}
Since $C_m(\alpha) = \alpha Q$ and $C_m(\alpha)$ is a prime, we deduce that $\alpha$ and $Q$ are nonzero.
If $\deg(\alpha) = h \ge 1$, then there exists at least one prime of positive degree in the prime factorization of $\alpha$. Since $C_m(\alpha)$ is a prime in $A$, it follows that $\alpha$ is prime, and thus $Q$ is a unit in $A$, that is, $Q$ belongs to $\bF_q^{\times}$. Therefore the degree of $Q$ is zero.
For each $1 \le i \le d - 1$, we see that the degree in $T$ of $\alpha^{q^i - 1}$ equals $h(q^i - 1)$, and thus the degree in $T$ of $[m, i]\alpha^{q^{i} - 1}$ is $q^i(d - i) + h(q^{i} - 1)$. For each $1 \le i \le d- 1$, we know from Bernoulli's inequality \cite[Theorem 42]{Hardy-Littlewood-Polya} that $h(q^{d - i} - 1) \ge q^{d - i} - 1 \ge (q - 1)(d - i) > d - i$, and thus
\begin{align*}
q^i(d - i) < hq^i(q^{d - i} - 1) = h(q^d - q^i) = h(q^d - 1) - h(q^i - 1).
\end{align*}
Hence we deduce that $h(q^d - 1) > q^i(d - i) + h(q^i - 1)$. Therefore the degree in $T$ of $\alpha^{q^d - 1}$ is greater than the degree in $T$ of $[m, i]\alpha^{q^{i} - 1}$ for each $1 \le i \le d - 1$.
Using Bernoulli's inequality and noting that $q \ge 3$, we know that $h(q^d - 1) \ge q^d - 1 \ge d(q - 1) > d$, and thus the degree in $T$ of $\alpha^{q^d - 1}$ is greater than the degree in $T$ of $m$. Therefore it follows from $(\ref{Equation-The-equation-of-Q-in-motivation-lemma-for-Mersenne-primes})$ that the degree in $T$ of $Q$ is $h(q^d - 1)$. This implies that $Q$ is not a unit in $A$, which is a contradiction. This contradiction implies that the degree of $\alpha$ is zero, and thus $\alpha$ belongs to $\bF_q^{\times}$.
We now prove that $m$ is a prime in $A$. Assume the contrary, that is, $m$ is not a prime in $A$. Hence $m$ can be written in the form $m = P_1P_2\cdots P_s$, where $s \ge 2$ and the $P_i$ are (not necessarily distinct) monic primes in $A$. For each $1 \le i \le s$, let $d_i \ge 1$ be the degree in $T$ of $P_i$. By \cite[Proposition 12.3.13]{Villa-Salvador}, we know that
\begin{align}
\label{Equation-The-factorization-of-Cm-alpha-in-cyclotomic-polynomials-Mersenne-primes-motivation}
C_m(\alpha) = \prod_{\substack{a | m \\ \text{$a$ monic}}}\Phi_a(\alpha),
\end{align}
where $\Phi_a(x) \in A[x]$ is the $a$-th cyclotomic polynomial for each monic element $a$ dividing $m$.
We know that
\begin{align}
\label{Equation-The-equation-of-Phi-P1-Mersenne-primes-motivation}
\Phi_{P_1}(\alpha) = \dfrac{C_{P_1}(\alpha)}{\alpha} = P_1 + [P_1, 1]\alpha^{q - 1} + [P_1, 2]\alpha^{q^2 - 1} + \cdots + [P_1, d_1 - 1]\alpha^{q^{d_1 - 1} - 1} + \alpha^{q^{d_1} - 1},
\end{align}
where $[P_1, i] \in A$ is a polynomial of degree $q^i(d_1 - i)$ for each $1 \le i \le d_1 - 1$.
For $1 \le i \le d_1 - 2$, we know from Bernoulli's inequality that
\begin{align*}
q^{d_1 - 1 - i} \ge 1 + (d_1 - 1 - i)(q - 1) > 1 + (d_1 - 1 - i) = d_1 - i,
\end{align*}
and it thus follows that $q^{d_1 - 1} > q^{i}(d_1 - i)$. Hence the degree in $T$ of $[P_1, d_1 - 1]$ is greater than the degree in $T$ of $[P_1, d_1 - i]$ for each $1 \le i \le d_1 - 2$. Similarly we can prove that $q^{d_1 - 1} > d_1$, and thus the degree in $T$ of $[P_1, d_1 - 1]$ is greater than the degree in $T$ of $P_1$. Since $\alpha$ is a unit in $A$, it follows from $(\ref{Equation-The-equation-of-Phi-P1-Mersenne-primes-motivation})$ that the degree in $T$ of $\Phi_{P_1}(\alpha)$ is $q^{d_1 - 1} \ge 1$. Similarly one can show that the degree in $T$ of $\Phi_{P_2}(\alpha)$ is $q^{d_2 - 1} \ge 1$. By $(\ref{Equation-The-factorization-of-Cm-alpha-in-cyclotomic-polynomials-Mersenne-primes-motivation})$, we can write
\begin{align*}
C_m(\alpha) = \Phi_{P_1}(\alpha)\Phi_{P_2}(\alpha)\prod_{\substack{a | m \\ \text{$a$ monic, $a \ne P_1, P_2$}}}\Phi_a(\alpha).
\end{align*}
Since $\Phi_{P_1}(\alpha), \Phi_{P_2}(\alpha)$ are non-units in $A$, we deduce from the last identity that $C_m(\alpha)$ is not prime in $A$, which is a contradiction. This contradiction establishes that $m$ is a prime in $A$.
\end{proof}
\begin{definition}
\label{Definition-Mersenne-numbers}
Let $q > 2$.
\begin{itemize}
\item [(i)] A Mersenne number in $A$ is a polynomial of the form $\alpha C_P(1)$, where $P$ is a monic prime in $A$ and $\alpha$ is an element in $\bF_q^{\times}$.
\item [(ii)] A Mersenne prime in $A$ is a prime of the form $\alpha C_P(1)$, where $P$ is a monic prime in $A$ and $\alpha$ is an element in $\bF_q^{\times}$.
\end{itemize}
\end{definition}
\begin{example}
\label{Example-Classification-of-Mersenne-primes-in-F-3-T}
Throughout this example, fix $q = 3$. Let $\wp_1 := T^2 + 1$, $\wp_2 := T^2 + 2T + 2$, $\wp_3 := T^2 + T + 2$. We know that $\{\wp_1, \wp_2, \wp_3\}$ consists of all monic irreducible polynomials of degree $2$ in $A$. We see that $M_{\wp_1} = C_{\wp_1}(1) = T^3 + T^2 + T + 2$, $M_{\wp_2} = C_{\wp_2}(1) = T^3 + T^2 + 2$, and $M_{\wp_3} = C_{\wp_3}(1) = T^3 + T^2 + 2T + 1$. Since the $M_{\wp_i}$ are primes in $A$, the set $\S_2 \subset A$ defined by
\begin{align*}
\S_2 := \{\alpha M_{\wp_i} \; | \; \text{$\alpha \in \bF_3^{\times}$ and $1 \le i \le 3$}\}
\end{align*}
consists of all Mersenne primes of the form $\alpha C_P(1)$, where $P$ is a monic prime in $\bF_3[T]$ of degree $2$ and $\alpha \in \bF_3^{\times}$.
\end{example}
\begin{remark}
\label{Remark-A-difficult-problem-to-determining-whether-C-P-1-is-a-prime-for-each-prime-P}
In the number field setting, it is not known whether there are infinitely many Mersenne primes. Furthermore we do not know whether there exist infinitely many prime numbers $p$ for which the Mersenne numbers $2^p - 1$ are composite (see, for example, \cite{Murata-Pomerance}). In the Carlitz module context with $q > 2$, the latter has an affirmative answer, i.e., for each $q > 2$, there exist infinitely many monic primes $\wp$ in $\bF_q[T]$ for which the Mersenne numbers $C_{\wp}(1)$ are composite. The rest of this section is devoted to proving this result.
\end{remark}
We recall a Carlitz module analogue of Fermat's little theorem which is a direct consequence of \cite[Proposition 2.4]{Hayes}.
\begin{lemma}
\label{Lemma-The-Carlitz-module-analogue-of-Fermat-little-theorem}
Let $P$ be a monic prime in $A$, and let $\alpha$ be a polynomial in $A$. Then $C_{P - 1}(\alpha) \equiv 0 \pmod{P}$.
\end{lemma}
\begin{corollary}
\label{Corollary-Fermat-little-theorem-for-alpha=1}
Let $P$ be a monic prime in $A$. Then $C_{P - 1}(1) \equiv 0 \pmod{P}$.
\end{corollary}
We recall a theorem of Hall's \cite{Hall} that plays a key role in the proof of Theorem \ref{Theorem-For-q->-2-there-are-infinitely-many-primes-P-such-that-C-P-1-is-composite}. For a proof of this result, see, for example, \cite[Theorem 4]{Pollack}.
\begin{theorem}
\label{Theorem-Theorem-of-Hall-about-twin-primes-in-function-fields}
$(\text{Hall})$
Assume that $q > 2$. Then there exists infinitely many polynomials $\wp$ in $\bF_q[T]$ such that $\wp, \wp + 1$ are monic primes.
\end{theorem}
\begin{theorem}
\label{Theorem-For-q->-2-there-are-infinitely-many-primes-P-such-that-C-P-1-is-composite}
Assume that $q > 2$. Then there exists infinitely many monic primes $\wp$ in $\bF_q[T]$ such that the Mersenne numbers $C_{\wp}(1)$ are composite.
\end{theorem}
\begin{proof}
It follows from Theorem \ref{Theorem-Theorem-of-Hall-about-twin-primes-in-function-fields} that there exist infinitely many primes $\wp, P \in A$ such that $P = \wp + 1$. Take such primes $\wp, P$ of degree $n \ge 2$ in $A$. We will prove that the Mersenne number $C_{\wp}(1)$ is composite.
By Corollary \ref{Corollary-Fermat-little-theorem-for-alpha=1}, we deduce that $C_{\wp}(1) = C_{P - 1}(1) \equiv 0 \pmod{P}$, and hence $P$ divides $C_{\wp}(1)$. Assume the contrary, that is, $C_{\wp}(1)$ is a prime. Since $P$ divides $C_{\wp}(1)$, it follows that $C_{\wp}(1) = \alpha P$ for some unit $\alpha$ in $A$. Hence
\begin{align}
\label{Equation-The-degree-of-C-wp-at-1-is-n}
n = \deg(P) = \deg(C_{\wp}(1)).
\end{align}
By \cite[Proposition 12.11]{Rosen} and since the degree of $\wp$ is $n$, we can write $C_{\wp}(x) \in A[x]$ in the form
\begin{align*}
C_{\wp}(x) = \wp x + [\wp, 1]x^q + [\wp, 2]x^{q^2} + \cdots + [\wp, n - 1]x^{q^{n - 1}} + x^{q^{n}},
\end{align*}
where $[\wp, i]$ is a polynomial of degree $q^i(n - i)$ for each $1 \le i \le n - 1$. Hence
\begin{align}
\label{Equation-C-wp-at-1}
C_{\wp}(1) = \wp + [\wp, 1] + [\wp, 2] + \cdots + [\wp, n - 1] + 1.
\end{align}
Using Bernoulli's inequality, we see that $q^{n - i - 1} \ge 1 + (q - 1)(n - i - 1)$ for every $0 \le i \le n - 2$. Since $q > 2$ and $n - i - 1 \ge 1$ for each $0 \le i \le n - 2$, we deduce that $1 + (q - 1)(n - i - 1) > n - i$, and thus $q^{n - i - 1} > n - i$. Therefore
\begin{align*}
\deg([\wp, n - 1]) = q^{n - 1} > q^i(n - i) = \deg([\wp, i])
\end{align*}
for every $0 \le i \le n - 2$. It thus follows from $(\ref{Equation-C-wp-at-1})$ that the degree of $C_{\wp}(1)$ is $q^{n - 1}$, which contradicts $(\ref{Equation-The-degree-of-C-wp-at-1-is-n})$. Thus $C_{\wp}(1)$ is composite, which proves our contention.
\end{proof}
\section{Wieferich primes and non-Wieferich primes}
\label{Section-Wieferich-primes}
In this section, we recall the notion of Wieferich primes in $A$ that was introduced by Thakur \cite{Thaur-Wieferich-primes} \cite{Thaur-Fermat-Wilson-congruences}. The aim of this section is to prove that a Mersenne prime is a non-Wieferich prime, which is analogous to a similar statement in the number field context. We begin by recalling the notions of Wieferich primes and non-Wieferich primes in the function field context.
\begin{definition}
\label{Definition-Wieferich-primes}
(see \cite{Thaur-Wieferich-primes} and \cite{Thaur-Fermat-Wilson-congruences})
Let $q > 2$, and let $\wp$ be a prime in $A$. Let $\alpha$ be the leading coefficient of $\wp$, and let $P$ be the unique monic prime in $A$ such that $\wp = \alpha P$. The prime $\wp$ is called \textit{a Wieferich prime} if $C_{P - 1}(1) \equiv 0 \pmod{P^2}$.
\end{definition}
\begin{definition}
\label{Definition-Non-Wieferich-primes}
(see \cite{Thaur-Wieferich-primes} and \cite{Thaur-Fermat-Wilson-congruences})
We maintain the same notation and assumptions as in Definition \ref{Definition-Wieferich-primes}. The prime $\wp$ is called \textit{a non-Wieferich prime} if $C_{P - 1}(1) \not\equiv 0 \pmod{P^2}$.
\end{definition}
In the number field case, it is a classical result that a Mersenne prime $p \in \bZ$ is a non-Wieferich prime. We now prove a Carlitz module analogue of this result.
\begin{theorem}
\label{Theorem-A-Mersenne-prime-is-a-non-Wieferich-prime-in-Carlitz-module-setting}
Assume that $q > 2$. Let $M_P = \alpha C_P(1)$ be a Mersenne prime, where $\alpha$ is a unit in $A$ and $P$ is a monic prime in $A$. Then $M_P$ is a non-Wieferich prime.
\end{theorem}
\begin{proof}
Let $\beta \in \bF_q^{\times}$ be the leading coefficient of $M_P$, and let $\wp$ be the unique monic prime in $A$ such that $M_P = \beta \wp$. Define
\begin{align}
\label{Equation-The-definition-of-gamma-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes}
\gamma = \alpha^{-1}\beta \in \bF_q^{\times}.
\end{align}
We see that
\begin{align}
\label{Equation-C-P-1-equals-alpha-inverse-times-beta-times-wp}
C_P(1) = \alpha^{-1}M_P = \gamma \wp.
\end{align}
By Corollary \ref{Corollary-Fermat-little-theorem-for-alpha=1}, we know that $\gamma \wp - 1 = C_P(1) - 1 = C_{P - 1}(1) \equiv 0 \pmod{P}$, and thus there exists a nonzero element $Q$ in $A$ such that
\begin{align}
\label{Equation-alpha-inverse-times-beta-wp-minus-1-equals-PQ}
\gamma \wp - 1 = PQ.
\end{align}
Assume the contrary, that is, $M_P$ is a Wieferich prime. Therefore we deduce that
\begin{align*}
C_{PQ}(1) = C_{\gamma \wp - 1}(1) = C_{\gamma\wp}(1) - 1 = \gamma C_{\wp}(1) - 1 \equiv \gamma - 1 \pmod{\wp^2}.
\end{align*}
Hence it follows from $(\ref{Equation-C-P-1-equals-alpha-inverse-times-beta-times-wp})$ that $C_Q(\gamma \wp) = C_Q(C_P(1)) = C_{PQ}(1) \equiv \gamma - 1 \pmod{\wp^2}$, and therefore
\begin{align}
\label{Equation-The-1st-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes}
C_Q(\gamma \wp) - (\gamma - 1) \equiv 0 \pmod{\wp^2}.
\end{align}
Let $r \ge 0$ be the degree of $Q$. If $r = 0$, we see that $Q$ is a unit in $\bF_q^{\times}$. Hence $C_Q(\gamma \wp) = Q\gamma \wp$, and it thus follows from $(\ref{Equation-The-1st-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes})$ that $Q\gamma \wp - (\gamma - 1) \equiv 0 \pmod{\wp^2}$. Thus $\wp^2$ divides $Q\gamma \wp - (\gamma - 1)$, which implies that
\begin{align*}
\deg(\wp) = \deg(Q\gamma \wp - (\gamma - 1)) \ge \deg(\wp^2) = 2\deg(\wp).
\end{align*}
Hence $\deg(\wp) = 0$, which is a contradiction since $\wp$ is a monic prime of positive degree. Thus $\deg(Q) = r \ge 1$.
We can write $C_Q(x) \in A[x]$ in the form $C_Q(x) = Qx + [Q, 1]x^q + \cdots + [Q, r - 1]x^{q^{r-1}} + [Q, r]x^{q^r}$, where $[Q, i] \in A$ is a polynomial of degree $q^i(r - i)$ for each $1 \le i \le r - 1$ and $[Q, r] \in \bF_q^{\times}$ is the leading coefficient of $Q$. From the equation of $C_Q(x)$, we deduce that
\begin{align}
\label{Equation-The-2nd-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes}
\dfrac{C_Q(\gamma \wp) - (\gamma - 1)}{\wp} = \dfrac{\gamma C_Q(\wp) - (\gamma - 1)}{\wp} = \gamma(Q + [Q, 1]\wp^{q - 1} + \cdots + [Q, r]\wp^{q^r - 1}) - \dfrac{(\gamma - 1)}{\wp}.
\end{align}
By $(\ref{Equation-The-1st-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes})$, we deduce that
\begin{align}
\label{Equation-The-3rd-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes}
\dfrac{C_Q(\gamma \wp) - (\gamma - 1)}{\wp} \equiv 0 \pmod{\wp},
\end{align}
and thus there is an element $R$ in $A$ such that $\dfrac{C_Q(\gamma \wp) - (\gamma - 1)}{\wp} = \wp R \in A$. We deduce from $(\ref{Equation-The-2nd-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes})$ that $\dfrac{(\gamma - 1)}{\wp} \in A$, and thus $\wp$ divides $\gamma - 1$. Since $\gamma - 1$ belongs to $\bF_q$ and $\wp$ is a monic prime of positive degree, we deduce that $\gamma - 1 = 0$, and hence $\gamma = 1$.
For $1 \le i \le r$, we see that $q^i - 1 \ge 1$, and it thus follows that $\wp^{q^i - 1} \equiv 0 \pmod{\wp}$. By $(\ref{Equation-The-2nd-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes})$, $(\ref{Equation-The-3rd-equation-in-lemma-that-Mersenne-primes-are-non-Wieferich-primes})$ and since $\gamma = 1$, we deduce from the last congruences that $0 \equiv \dfrac{C_Q(\wp)}{\wp} \equiv Q \pmod{\wp}$. Therefore there exists a nonzero element $P_1 \in A$ such that $P_1\wp = Q$.
By $(\ref{Equation-alpha-inverse-times-beta-wp-minus-1-equals-PQ})$ and since $\gamma = 1$, we deduce that $\wp - 1 = PQ = P(P_1\wp)$, and thus $\wp(1 - P_1P) = 1$. Therefore $\wp$ is a unit in $A$, which is a contradiction. Thus $M_P$ is a non-Wieferich prime.
\end{proof}
Thakur \cite{Thaur-Fermat-Wilson-congruences} recently found interesting relations linking Wieferich primes in the function field context to zeta values. Also in \cite{Thaur-Fermat-Wilson-congruences}, Thakur made the following conjecture that relates the degree of a Wieferich prime to the characteristic $p$.
\begin{conjecture}
$(\text{Thakur's conjecture})$
For every $p > 2$, the degree of a Wieferich prime in $A = \bF_q[T]$ is divisible by the characteristic $p$.
\end{conjecture}
\section{The Carlitz annihilators of primes}
\label{Section-The-Carlitz-annihilators-of-primes}
In this section, we introduce a notion of the Carlitz annihilator of a prime. By way of illustration, we prove a Carlitz module analogue of a classical result concerning congruences for prime divisors of the Mersenne numbers in elementary number theory.
For a monic prime $P \in A$, we know from Corollary \ref{Corollary-Fermat-little-theorem-for-alpha=1} that $C_{P - 1}(1) \equiv 0 \pmod{P}$. Hence using Proposition 1.6.5 and Lemma 1.6.8 in \cite{Goss}, the following result follows immediately.
\begin{proposition}
\label{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime}
Let $P$ be a monic prime in $A$ of positive degree. Then there exists a unique monic polynomial $\wp_{P} \in A$ of positive degree satisfying the following.
\begin{itemize}
\item [(i)] $C_{\wp_{P}}(1) \equiv 0 \pmod{P}$;
\item [(ii)] for any nonzero element $a \in A$, $\wp_{P}$ divides $a$ if and only if $C_a(1) \equiv 0 \pmod{P}$; and
\end{itemize}
\end{proposition}
Proposition \ref{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime} motivates the following definition that is a Carlitz module analogue of the order of an element in a finite group.
\begin{definition}
\label{Definition-The-Carlitz-module-order-of-a-prime-in-F-q}
Let $P$ be a monic prime in $A$ of positive degree, and let $\wp_P$ be the unique monic polynomial satisfying $(i), (ii)$ in Proposition \ref{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime}. The polynomial $\wp_P$ is called \textit{the Carlitz annihilator of $P$}.
\end{definition}
By Corollary \ref{Corollary-Fermat-little-theorem-for-alpha=1} and part $(ii)$ of Proposition \ref{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime}, the following result follows immediately.
\begin{lemma}
\label{Lemma-The-Carlitz-order-of-P-divides-P-minus-1}
Let $P$ be a monic prime in $A$, and let $\wp_P$ be the Carlitz annihilator of $P$. Then $\wp_P$ divides $P - 1$.
\end{lemma}
Return to the number field context, and let $M_p := 2^p - 1$ be a Mersenne number for some odd prime $p$. A classical result in elementary number theory says that any prime $q$ dividing $M_p$ satisfies $q \equiv 1 \pmod{p}$ (for a proof of this result, see, for example, \cite[Theorem 3]{Jaroma-Reddy}). We now prove a Carlitz module analogue of this result using the notion of the Carlitz annihilator of a prime in a similar manner that the notion of the order of an integer modulo a prime appears in the proof of the analogous result in the number field context mentioned above.
\begin{theorem}
\label{Theorem-Every-prime-dividing-a-Carlitz-module-Mersenne-number-is-congruent-to-1-mod-P}
Let $P$ be a monic prime in $A$, and let $M_P := \alpha C_P(1)$ be a Mersenne number, where $\alpha$ is an element in $\bF_q^{\times}$. Then $Q \equiv 1 \pmod{P}$ for any monic prime $Q$ dividing $M_P$.
\end{theorem}
\begin{proof}
Let $Q$ be any monic prime in $A$ such that $Q$ divides $M_P$. Thus $M_P = \alpha C_P(1) \equiv 0 \pmod{Q}$. Since $\alpha$ is a unit in $A$ and $Q$ is a monic prime, it follows that
\begin{align}
\label{Equation-The-1st-equation-in-the-result-about-being-congruent-to-1-mod-P-of-primes-dividing-Mersenne-numbers}
C_P(1) \equiv 0 \pmod{Q}.
\end{align}
Let $\wp_{Q}$ be the Carlitz annihilator of $Q$. By part $(ii)$ of Proposition \ref{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime} and $(\ref{Equation-The-1st-equation-in-the-result-about-being-congruent-to-1-mod-P-of-primes-dividing-Mersenne-numbers})$, we deduce that $\wp_Q$ divides $P$. Since $\wp_Q$ is a monic polynomial of positive degree and $P$ is a monic prime in $A$, it follows that $\wp_Q = P$. By Lemma \ref{Lemma-The-Carlitz-order-of-P-divides-P-minus-1}, we know that $\wp_Q$ divides $Q - 1$, and thus $P$ divides $Q - 1$. Therefore $Q \equiv 1 \pmod{P}$, which proves our contention.
\end{proof}
As a by-product of Theorem \ref{Theorem-Every-prime-dividing-a-Carlitz-module-Mersenne-number-is-congruent-to-1-mod-P}, one obtains the following result that precisely describes the Carlitz annihilator of any monic prime occurring in the prime factorization of a Mersenne number.
\begin{corollary}
\label{Corollary-The-Carlitz-orders-of-primes-in-the-prime-factorization-of-a-Mersenne-number}
Let $P$ be a monic prime in $A$, and let $M_P := \alpha C_P(1)$ be a Mersenne number, where $\alpha$ is an element in $\bF_q^{\times}$. Let $Q$ be a monic prime in $A$ such that $Q$ divides $M_P$, and let $\wp_Q$ be the Carlitz annilator of $Q$. Then $\wp_Q = P$.
\end{corollary}
\begin{proof}
Following the proof of Theorem \ref{Theorem-Every-prime-dividing-a-Carlitz-module-Mersenne-number-is-congruent-to-1-mod-P}, we derive Corollary \ref{Corollary-The-Carlitz-orders-of-primes-in-the-prime-factorization-of-a-Mersenne-number} immediately.
\end{proof}
The following result fully characterizes all monic primes $P$ for which $\wp_P$ is a prime.
\begin{corollary}
\label{Corollary-A-criterion-for-determining-when-wp-P-is-a-prime}
Let $P$ be a monic prime in $A$, and let $\wp_P$ be the Carlitz annihilator of $P$. Then $\wp_P$ is a prime if and only if $P$ divides $C_Q(1)$ for some monic prime $Q$.
\end{corollary}
\begin{proof}
Assume that $P$ divides $C_Q(1)$ for some monic prime $Q$. It follows from Corollary \ref{Corollary-The-Carlitz-orders-of-primes-in-the-prime-factorization-of-a-Mersenne-number} that $\wp_P = Q$, and thus $\wp_P$ is a prime.
Conversely assume that $\wp_P$ is a prime, say $Q$. By Proposition \ref{Proposition-The-result-motivating-the-definition-of-Carlitz-order-of-a-prime}, we see that $C_Q(1) = C_{\wp_P}(1) \equiv 0 \pmod{P}$, and thus $P$ divides $C_Q(1)$.
\end{proof}
\section{A criterion for determining whether a Mersenne number is prime}
\label{Section-A-criterion-for-determing-whether-a-Mersenne-number-is-prime}
In this section, we prove a criterion for determining whether a Mersenne number is prime. The criterion relies on primality of certain elements in cyclotomic function fields. For the notation used in this section, we refer the reader to Subsection \ref{Subsection-Notation}.
\begin{theorem}
\label{Theorem-A-criterion-for-determining-whether-a-Mersenne-number-is-prime}
Let $P$ be a monic prime in $A$. Let $M_P := \alpha C_P(1)$ be a Mersenne number, where $\alpha$ is an element in $\bF_q^{\times}$. Let $K_P$ be the $P$-th cyclotomic function field. Then $M_P$ is a prime in $A$ if and only if $1 - \lambda_P$ is a prime element in the ring of integers $\cO_P$ of $K_P$. Furthermore when $M_P$ is a prime, $M_P$ splits completely in $K_P$.
\end{theorem}
\begin{proof}
Throughout the proof, let $\Phi_P$ be the $P$-th cyclotomic polynomial. It is known \cite[Proposition 2.4]{Hayes} that $M_P = \alpha C_P(1) = \alpha \Phi_P(1)$.
Assume that $M_P$ is a prime in $A$. It follows that $\Phi_P(1)$ is a prime in $A$. We know that
\begin{align}
\label{Equation-Representation-of-Phi-P-1-as-the-norm-of-1-minus-lambda-P}
\Phi_P(1) = \prod_{\substack{0 \le \deg(a) < \deg(P)}}(1 - \sigma_P^{(a)}(\lambda_P)) = \Norm_{K_P/k}(1 - \lambda_P).
\end{align}
Let
\begin{align}
\label{Equation-The-prime-ideal-factorization-of-1-minus-lambda-P}
(1 - \lambda_P)\cO_P = \fB_1^{\epsilon_1}\fB_2^{\epsilon_2}\cdots\fB_h^{\epsilon_h}
\end{align}
be the prime ideal factorization of the principal ideal $(1 - \lambda_P)\cO_P$ in $\cO_P$, where the $\fB_i$ are distinct prime ideals in $\cO_P$ and the $\epsilon_i$ are positive integers. For each $1 \le i \le h$, let $\fp_i$ be the prime ideal of $k$ lying below $\fB_i$, and let $f_i := f(\fB_i/\fp_i)$ denote the relative degree of $\fB_i$ over $\fp_i$. Let $\N_{K_P/k}$ denote the norm map on ideals of $K_P$ over $k$.
We know from $(\ref{Equation-Representation-of-Phi-P-1-as-the-norm-of-1-minus-lambda-P})$, $(\ref{Equation-The-prime-ideal-factorization-of-1-minus-lambda-P})$ and \cite[Theorem 3.1.3]{Goldschmidt} that
\begin{align*}
\Phi_P(1)A = \Norm_{K_P/k}(1 - \lambda_P)A = \N_{K_P/k}((1 - \lambda_P)\cO_P) = \N_{K_P/k}(\prod_{i = 1}^h\fB_i^{\epsilon_i}) = \prod_{i = 1}^h\fp_i^{\epsilon_i f_i}.
\end{align*}
Since $\Phi_P(1)$ is a prime in $A$, we derive that $\Phi_P(1)A$ is a prime ideal of $A$. Thus it follows from the above equation that $h = \epsilon_1 = f_1 = 1$. Therefore we deduce from $(\ref{Equation-The-prime-ideal-factorization-of-1-minus-lambda-P})$ that $(1 - \lambda_P)\cO_P = \fB_1$, which is a prime ideal. Hence $1 - \lambda_P$ is a prime element in $\cO_P$.
Conversely suppose that $1 - \lambda_P$ is a prime element in $\cO_P$. Hence there is a prime ideal $\fB$ in $\cO_P$ such that $(1 - \lambda_P)\cO_P = \fB$. Let $\fp$ be the prime ideal of $k$ lying below $\fB$, and let $f := f(\fB/\fp)$ be the relative degree of $\fB$ over $\fp$. Since $A$ is a principal ideal domain, there exists a monic prime $\wp$ in $A$ such that $\fp = \wp A$. Using \cite[Theorem 3.1.3]{Goldschmidt} and repeating the same arguments as above, we deduce that
\begin{align*}
\Phi_P(1)A = \Norm_{K_P/k}(1 - \lambda_P)A = \N_{K_P/k}((1 - \lambda_P)\cO_P) = \N_{K_P/k}(\fB) = \fp^{f} = (\wp A)^f = \wp^f A.
\end{align*}
Thus $\wp^f$ divides $\Phi_P(1)$, and hence $\wp$ divides $\Phi_P(1)$. Hence $\wp$ divides $M_P$, and it thus follows from Theorem \ref{Theorem-Every-prime-dividing-a-Carlitz-module-Mersenne-number-is-congruent-to-1-mod-P} that $\wp \equiv 1 \pmod{P}$. By \cite[Theorem 12.10]{Rosen}, we deduce that $\wp A$ splits completely in $K_P$, and therefore $f = 1$. Hence $\Phi_P(1)A = \wp A$, which implies that $\Phi_P(1)$ is a prime in $A$. Therefore $M_P$ is a prime.
Finally when $M_P$ is a prime, we see that since $\alpha$ is a unit in $A$, there exists a monic prime $\wp$ in $A$ such that $M_PA = \alpha \Phi_P(1)A = \wp A$. Using the same arguments as above, we deduce that $\wp \equiv 1 \pmod{P}$, and therefore $M_P A = \wp A$ splits completely in $K_P$.
\end{proof}
\begin{remark}
Theorem \ref{Theorem-A-criterion-for-determining-whether-a-Mersenne-number-is-prime} is a Carlitz module analogue of the corollary to \cite[Theorem 3]{Helou} in the classical cyclotomic theory. Although we have adapted the ideas in the proof of the corollary in \cite{Helou} to the proof of Theorem \ref{Theorem-A-criterion-for-determining-whether-a-Mersenne-number-is-prime} in the Carlitz module context, there is a substantial difference between the two proofs. To be more specific, note that the congruence $\wp \equiv 1 \pmod{P}$ is a key step in both the proof of Theorem \ref{Theorem-A-criterion-for-determining-whether-a-Mersenne-number-is-prime} and that of the corollary to \cite[Theorem 3]{Helou}, where $\wp$ is any prime dividing the Mersenne number $M_P$. In the number field case, this congruence is a well-known result and can be derived using the notion of the orders of elements modulo primes (see \cite[Theorem 3]{Jaroma-Reddy} for a proof of this congruence). The congruence $\wp \equiv 1 \pmod{P}$ in the number field context can also be proved using some facts about primes dividing the norms of certain cyclotomic elements as shown in \cite{Helou}. In the Carlitz module context, we need to introduce the notion of the Carlitz annihilator of a prime as a replacement for that of the order of an element modulo a prime to derive the congruence $\wp \equiv 1 \pmod{P}$.
\end{remark}
\section*{Acknowledgements}
I am very grateful to Dinesh Thakur for many of his insights in function field arithmetic, making many useful comments, and pointing out some useful references. I thank the referee for useful comments. I thank my parents, Nguyen Ngoc Quang and Phan Thi Thien Huong, for their constant support. I was supported by a postdoctoral fellowship in the Department of Mathematics at University of British Columbia.
|
\section{Introduction}
\label{sec:intro}
High-energy collisions at CERN's Large Hadron Collider (LHC) provide new opportunities to search for physics beyond the Standard Model (SM).
This paper describes such a search in events containing a lepton (electron or muon) and
missing transverse momentum using 8~\tev\ \pp\ collision data collected with the ATLAS detector during
2012, corresponding to a total integrated luminosity of 20.3~\ifb.
The first new-physics scenario that is considered in this paper is the Sequential Standard Model (SSM),
the extended gauge model of ref.~\cite{ssm}. This model proposes the existence of additional heavy gauge bosons,
of which the charged ones are commonly denoted \wp. The \wp\ has the same couplings to fermions as the SM \w~boson
and a width that increases linearly with the \wp\ mass. The coupling of the \wp\ to $WZ$ is set to zero.
Similar searches~\cite{cdf:Wprime2010,cms:wprime2011-2,cms:wprime2013-1,atlas:wprime_2010_pub,atlas:wprime_2011-2_pub,atlas:wprime_2012_pub}
have been performed using $\sqrt{s} = 1.96$ TeV $p\bar{p}$ collision data by the CDF Collaboration,
$\sqrt{s} = 7$ TeV $pp$ collision data by the ATLAS Collaboration as well as $\sqrt{s} = 7$ TeV and $\sqrt{s} = 8$ TeV data by the CMS Collaboration.
The second new-physics scenario that is considered originates from ref.~\cite{wzstar_motivate} and proposes
the existence of charged partners, denoted \wstar, of the chiral boson excitations described in ref.~\cite{wzstar}.
The anomalous (mag\-netic-moment type) coupling of the \wstar\ leads to kinematic distributions significantly different
from those of the \wp\ as demonstrated in the previous ATLAS search~\cite{atlas:wprime_2012_pub}
that was performed using 7~\tev\ \pp\ collision data collected in 2011 corresponding to an integrated luminosity of 4.7~\ifb.
In the analysis presented in this paper the search region is expanded to higher masses and the sensitivity is considerably improved in the region covered by the previous search.
The third new-physics scenario considered is of direct production of weakly interacting candidate dark matter (DM) particles.
These particles can be pair-produced at the LHC, $pp\rightarrow\chi\bar{\chi}$, via a new intermediate state.
Since DM particles do not interact with the detector material, these events can be detected if there is associated initial-state radiation of a SM particle~\cite{Birkedal:2004xn,Goodman:2010yf,Bai:2010hh,Goodman:2010ku}.
The Tevatron and LHC collaborations have reported limits on the cross-section of $p\bar{p}/pp\rightarrow\chi\bar{\chi}+X$
where $X$ is a hadronic jet~\cite{cdfWIMPjet,atlasWIMPjet,cmsWIMPjet}, a photon~\cite{atlasWIMPphoton,cmsWIMPphoton},
a hadronically decaying \w\ or \z\ boson~\cite{atlasWIMPhadronicWZ} or a leptonically decaying \z\ boson~\cite{atlasWIMPleptonicZ}.
Previous LHC results have also been reinterpreted to set limits on the scenario where $X$ is a leptonically decaying \w\ boson \cite{monolep}.
This analysis is the first direct ATLAS search for this case. Limits are reported for the \mbox{DM--nucleon} scattering cross-section as well as the
mass scale, \mstar, of a new SM--DM interaction expressed in an effective field theory (EFT) as a four-point contact interaction
~\cite{Beltran:2010ww,Cao:2009uw,Rajaraman:2011wf,Fox:2011pm,Cheung:2012gi,Cotta:2012nj}.
As discussed in the literature, e.g. refs.~\cite{Buchmueller:2013dya,Busoni:2013lha}, the EFT formalism is not always an appropriate approximation
but this issue is not addressed any further in this paper. Four effective operators are used as a representative set based on the definitions in
ref.~\cite{Goodman:2010ku}: D1 scalar, D5 vector (both constructive and destructive interference cases are considered, the former denoted by D5c and the
latter by D5d) and D9 tensor.
The analysis presented here identifies event candidates in the electron and muon channels,
sets separate limits and then combines these assuming a common branching fraction for the two final states.
The kinematic variable used to identify the signal is the transverse mass
\begin{equation}
\mt = \sqrt{ 2 \pt \met (1 - \cos \varphi_{\ell\nu})},
\end{equation}
where \pt\ is the lepton transverse momentum, \met\ is the magnitude of the missing transverse momentum vector
and $\varphi_{\ell\nu}$ is the angle between the \pt\ and \met\ vectors.\footnote{
ATLAS uses a right-handed coordinate system with its origin at the nominal
interaction point in the centre of the detector and the $z$-axis along the
beam pipe. Cylindrical coordinates $(r,\varphi)$ are used in the transverse
plane, $\varphi$ being the azimuthal angle around the beam pipe. The pseudorapidity $\eta$ is
defined in terms of the polar angle $\theta$ by $\eta=-\ln\tan(\theta/2)$.
}
The main background to the $W'$, $W^{*}$ and DM signals comes from the tail of the \mt\ distribution from SM \w~boson production with
decays to the same final state.
Other relevant backgrounds are \z\ boson production with decays into two leptons where one lepton is not reconstructed,
\w\ or \z\ production with decays to $\tau$ leptons where a $\tau$ subsequently decays to either an electron or a muon, and
diboson production. These are collectively referred to as the electroweak (EW) background.
There is also a contribution to the background from \ttbar\ and single-top production, collectively referred to as the top background,
which is most important for the lowest \wps\ masses considered here, where it constitutes about 10\% of the background after event selection in the
electron channel and 15\% in the muon channel.
Other relevant strong-interaction background sources occur when a light or heavy hadron decays semileptonically or
when a jet is misidentified as an electron or muon. These are referred to as the multi-jet background\ in this paper.
\FloatBarrier
\section{The ATLAS detector}
\label{sec:det}
The ATLAS detector~\cite{atlas:detector} is a multi-purpose particle
physics detector with a forward-backward symmetric cylindrical
geometry and nearly 4$\pi$ coverage in solid angle.
The ATLAS detector has three major components: the inner tracking detector (ID),
the calorimeter and the muon spectrometer (MS). Tracks and vertices of charged particles are reconstructed with
silicon pixel and silicon microstrip detectors covering $|\eta|<2.5$ and straw-tube transition radiation detectors covering $|\eta| < 2.0$,
all immersed in a homogeneous 2~T magnetic field provided by a superconducting solenoid.
The ID is surrounded by a hermetic calorimeter that covers $|\eta| <$ 4.9 and provides three-dimensional reconstruction of particle showers.
The electromagnetic calorimeter is a liquid argon (LAr) sampling calorimeter, which uses lead absorbers for \mbox{$|\eta| <$ 3.2}
and copper absorbers in the very forward region. The hadronic sampling calorimeter
uses plastic scintillator tiles as the active material and iron absorbers in the region $|\eta| <$ 1.7. In the region
1.5 $ < |\eta| < $ 4.9, liquid argon is used as the active material, with copper and/or tungsten absorbers.
The MS surrounds the calorimeters and consists of three large superconducting toroid systems (each with eight coils)
together with multiple layers of trigger chambers up to $|\eta|<$ 2.4 and tracking chambers, providing precision track measurements, up to $|\eta|<$ 2.7.
\FloatBarrier
\section{Trigger and reconstruction}
\label{sec:trigAndReco}
The data used in the electron channel were recorded with a trigger requiring the presence of an energy cluster in the EM compartment of the calorimeter
(EM cluster) with $\et > 120\gev$.
For the muon channel, matching tracks in the MS and ID with combined $\pt > 36\gev$ are
used to select events. In order to compensate for the small loss in the selection efficiency at high \pt\ due to this matching,
events are also recorded if a muon with $\pt>40\gev$ and $|\eta|<1.05$ is found in the MS.
The average trigger efficiency (measured with respect to reconstructed objects) is above 99\% in the electron channel and
80\%--90\% in the muon channel for the region of interest in this analysis.
Each EM cluster with \mbox{$\et > 125\gev$} and $|\eta| < 1.37$ or $1.52 < |\eta| < 2.47$
is considered as an electron candidate if it is matched to an ID track.
The region $1.37 \leq |\eta| \leq 1.52$ exhibits degraded energy resolution due to the transition from
the central region to the forward regions of the calorimeters and is therefore excluded.
The track and the cluster must satisfy a set of identification criteria
that are optimised for the conditions of many proton--proton collisions in the same or nearby beam bunch crossings
(in-time or out-of-time pile-up, respectively)~\cite{atlas:egamreco}. These criteria require the shower profiles to be consistent with those expected
for electrons and impose a minimum requirement on the amount of transition radiation that is present. In addition,
to suppress background from photon conversions, a hit in the first layer of the pixel detector is required if an active pixel sensor is traversed.
The electron's energy is obtained from the calorimeter measurements while its direction is obtained from the associated track.
In the high-\et\ range relevant for this analysis, the electromagnetic calorimeter energy resolution is measured in data to be
1.2\% in the central region and 1.8\% in the forward region~\cite{atlas:egamma_perf}.
These requirements result in about a 90\% identification efficiency for electrons with \mbox{$\et > 125 \gev$}.
Muons are required to have a $\pt > 45\gev$, where the momentum of the muon is obtained by combining the ID and MS measurements.
To ensure an accurate measurement of the momentum, muons are required to have hits in three MS layers
and are restricted to the ranges $|\eta| < 1.0$ and $1.3 < |\eta| < 2.0$.
Some of the chambers in the region $1.0 < |\eta| < 1.3$ were not yet installed, hence the momentum resolution
of MS tracks is degraded in this region. Including the muon candidates with an \eta-range 2.0 $<|\eta|<$ 2.5 would lead to an increase in the signal selection efficiency
of up to 12\% for lower \wp\ masses and of up to 3\% for a \wp\ mass of 3 TeV.
However, the background levels in the signal region would increase by more than 15\%.
Therefore, the previously stated $\eta$ restrictions are retained.
For the final selection of good muon candidates, the individual ID and MS momentum measurements are required to be
in agreement within 5 standard deviations.
The average momentum resolution is about 15\%--20\% at $\pt = 1\tev$.
About 80\% of the muons in the \eta-range considered are reconstructed, with most of the loss coming from regions without
three MS layers.
The \met\ in each event is evaluated by summing over energy-calibrated physics objects
(jets, photons and leptons) and adding corrections for calorimeter deposits not associated with these objects~\cite{atlas:met2011}.
This analysis makes use of all of the $\sqrt{s} = 8\tev$ data collected in 2012 for which the relevant detector systems were operating properly
and all data quality requirements were satisfied. The integrated luminosity of the data used in this study is 20.3~\ifb\ for both the electron
and muon decay channels.
The uncertainty on this measurement is 2.8\%, which is derived following the methodology detailed in ref.~\cite{atlas:lumi7Tev}.
\FloatBarrier
\section{Monte Carlo simulation}
\label{sec:mcSim}
With the exception of the multi-jet background, which is estimated from data, expected signals and backgrounds are
evaluated using simulated Monte Carlo samples and normalised using the calculated cross-sections and the integrated luminosity of the data.
The \wp\ signal events are generated at leading order (LO) with \pythia\ v8.165~\cite{pythia,pythia8} using the MSTW2008 LO~\cite{mstw}
parton distribution functions (PDFs). \pythia\ is also used for the fragmentation and hadronisation of \wsl\ events that are
generated at LO with \calchep\ v3.3.6~\cite{calchep} using the CTEQ6L1 PDFs~\cite{cteq6l}.
DM signal samples are generated at LO with \madgraphFive\ v1.4.5~\cite{madgraph} using the MSTW2008 LO PDFs, interfaced to \pythia\ v8.165.
The \wzbg\ boson and \ttbar\ backgrounds are generated at next-to-leading order (NLO) with \powhegbox\ r1556 \cite{powheg}
using the CT10 NLO~\cite{CT10} PDFs. For the \wzbg\ backgrounds, fragmentation and hadronisation is performed with \pythia\ v8.165,
while for \ttbar\ \pythia\ v6.426 is used. The single-top background is generated at NLO with \mcatnlo\ v4.06~\cite{mcatnlo} using the CT10 NLO PDFs
for the $Wt$- and $s$-channels, and with \acerMC\ v3.8~\cite{acermc} using the CTEQ6L1 PDFs for the $t$-channel.
Fragmentation and hadronisation for the \mcatnlo\ samples are performed with \herwig\ v6.520~\cite{herwig}, using
\jimmy\ v4.31~\cite{jimmy} for the underlying event, whereas \pythia\ v6.426 is used for the \acerMC\ samples.
The $WW$, $WZ$ and $ZZ$ diboson backgrounds are generated at LO with \sherpa\ v1.4.1~\cite{sherpa} using the CT10 NLO PDFs.
The \pythia\ signal model for \wp\ has \vminusa\ SM couplings to fermions but does not include
interference between the \w\ and \wp. For both \wp\ and \wstar, decay channels beside $e\nu$ and $\mu\nu$,
notably $\tau\nu$, $ud$, $sc$ and $tb$, are included in the calculation of the widths but are not explicitly included as signal or background.
At high mass ($\mwp > 1~\tev$), the total width is about 3.5 \% of the pole mass, and
the branching fraction to each of the lepton decay channels is 8.2\%.
For all samples, final-state photon radiation from leptons is handled by
\photos~\cite{photos}. The ATLAS full detector simulation~\cite{atlas:sim} based on \geant4~\cite{geant} is used to
propagate the particles and account for the response of the detector.
For the underlying event, the ATLAS tune AUET2B~\cite{mc11tune} is used for \pythia\ 6 and AU2~\cite{mc11tune1} is used for \pythia\ 8, while
AUET2~\cite{mc11tune2} is used for the \herwig\ with \jimmy.
The effect of pile-up is incorporated into the simulation by overlaying additional minimum-bias
events generated with \pythia\ onto the generated hard-scatter events. Simulated events are weighted to match the
distribution of the number of interactions per bunch crossing observed in data, but are otherwise
reconstructed in the same manner as data.
The \wlnu\ and \zll\ cross-sections are calculated at next-to-next-to-leading order (NNLO) in QCD with ZWPROD~\cite{zwprod}
using MSTW2008 NNLO PDFs. Consistent results are obtained using VRAP v0.9~\cite{vrap} and FEWZ v3.1b2~\cite{fewz,fewz2}.
Higher-order electroweak corrections are calculated with MCSANC~\cite{mcsanc}.
Mass-dependent $K$-factors obtained from the ratios of the calculated higher-order cross-sections
to the cross-sections of the generated samples are used to scale $W^{+}$, $W^{-}$ and $Z$ backgrounds separately.
The \wpl\ cross-sections are calculated in the same way, except that the electroweak corrections beyond
final-state radiation are not included because the calculation for the SM \w\ cannot be applied directly.
Cross sections for \wsl\ are kept at LO due to the non-renormalisability of the model at higher orders in QCD.
The \ttbar\ cross-section is also calculated at NNLO including resummation of next-to-next-to-leading logarithmic (NNLL) soft gluon terms
obtained with \topPP\ v2.0~\cite{Cacciari:2011hy,Baernreuther:2012ws,Czakon:2012zr,Czakon:2012pz,Czakon:2013goa,Czakon:2011xx}
for a top quark mass of $172.5 \gev$.
The $W'$, $W^{*}$, and DM particle signal cross-sections are listed in tables~\ref{tab:xsec_sig} and~\ref{tab:xsec_sigdm}.
The most important background cross-sections are listed in table~\ref{tab:xsec_bg}.
\begin{table}[!htbp]
\caption{Predicted values of the cross-section times branching fraction (\xbr) for \wpl\ and \wsl.
The \xbr\ for \wpl\ are at NNLO while those for \wsl\ are at LO.
The values are given per channel, with $\ell=e$ or $\mu$.
}
\label{tab:xsec_sig}
\begin{center}
\begin{tabular}{c|l|l}
\hline
\hline
Mass & \multicolumn{1}{c|}{\wpl} & \multicolumn{1}{c}{\wsl} \\ \relax
[\gev] & \multicolumn{1}{c|}{\xbr\ [pb]} & \multicolumn{1}{c}{\xbr\ [pb]} \\
\hline
\phantom{0}300 & \phantom{}149.0 & \\
\phantom{0}400 & \phantom{0}50.2 & \phantom{0}37.6 \\
\phantom{0}500 & \phantom{0}21.4 & \phantom{0}16.2 \\
\phantom{0}600 & \phantom{0}10.4 & \phantom{00}7.95 \\
\phantom{0}750 & \phantom{00}4.16 & \phantom{00}3.17 \\
\phantom{}1000 & \phantom{00}1.16 & \phantom{00}0.882 \\
\phantom{}1250 & \phantom{00}0.389 & \phantom{00}0.294 \\
\phantom{}1500 & \phantom{00}0.146 & \phantom{00}0.108 \\
\phantom{}1750 & \phantom{00}0.0581 & \phantom{00}0.0423 \\
\phantom{}2000 & \phantom{00}0.0244 & \phantom{00}0.0171 \\
\phantom{}2250 & \phantom{00}0.0108 \phantom{0} & \phantom{00}0.00700 \\
\phantom{}2500 & \phantom{00}0.00509 \phantom{0} & \phantom{00}0.00290 \\
\phantom{}2750 & \phantom{00}0.00258 & \phantom{00}0.00120 \\
\phantom{}3000 & \phantom{00}0.00144 & \phantom{00}4.9$\times 10^{-4}$ \\
\phantom{}3250 & \phantom{00}8.9$\times 10^{-4}$ & \phantom{00}2.0$\times 10^{-4}$ \\
\phantom{}3500 & \phantom{00}5.9$\times 10^{-4}$ & \phantom{00}8.0$\times 10^{-5}$ \\
\phantom{}3750 & \phantom{00}4.2$\times 10^{-4}$ & \phantom{00}3.2$\times 10^{-5}$ \\
\phantom{}4000 & \phantom{00}3.1$\times 10^{-4}$ & \phantom{00}1.3$\times 10^{-5}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!htbp]
\caption{
Predicted values of \xbr\ for DM signal with different mass values, $m_{\chi}$.
The values of \mstar\ used in the calculation for a given operator are also shown.
The cross-sections are at LO, and the values are given for the sum of three lepton flavours $\ell=e,\mu,\tau$.
}
\label{tab:xsec_sigdm}
\begin{center}
\begin{tabular}{c|l|l|l|l}
\hline
\hline
& \multicolumn{4}{c}{DM production} \\
$m_{\chi}$ & \multicolumn{4}{c}{$\xbr$ [pb]} \\ \relax
[\gev] & \multicolumn{1}{c}{D1} & \multicolumn{1}{c}{D5d} & \multicolumn{1}{c}{D5c} & \multicolumn{1}{c}{D9}
\\
& \multicolumn{1}{c}{\mstar$\ = 10 \GeV$} & \multicolumn{1}{c}{\mstar$\ = 100 \GeV$} & \multicolumn{1}{c}{\mstar$\ = 1 \TeV$} & \multicolumn{1}{c}{\mstar$\ = 1 \TeV$}
\\
\hline
\phantom{000}1 & 439 & 72.2 & 0.0608 & 0.0966 \\
\phantom{0}100 & 332 & 70.8 & 0.0575 & 0.0870 \\
\phantom{0}200 & 201 & 58.8 & 0.0488 & 0.0695 \\
\phantom{0}400 & \phantom{0}64.6 & 32.9 & 0.0279 & 0.0365 \\
\phantom{}1000 & \phantom{00}1.60 & \phantom{0}2.37 & 0.00192 & 0.00227 \\
\phantom{}1300 & \phantom{00}0.213 & \phantom{0}0.454 & 0.000351 & 0.000412 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\caption{
Predicted values of \xbr\ for the leading backgrounds.
The value for $\ttbar\rightarrow\ell X$ includes all final states with at least one
lepton ($e$, $\mu$ or $\tau$).
The others are exclusive and are used for both $\ell=e$ and $\ell=\mu$.
All cross-sections are at NNLO.
}
\label{tab:xsec_bg}
\begin{center}
\begin{tabular}{l|l}
\hline
\hline
Process & $\xbr$ [pb] \\
\hline
\wlnu\ & \phantom{}12190\phantom{.0} \\
\zgll\ ($m_{\zg}>60 \gev$) & \phantom{0}1120\phantom{.0} \\
\ttbarl & \phantom{00}137.3 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
Uncertainties on the \wp\ cross-section and the \wzbg\ background cross-sections are estimated from variations of the renormalisation and factorisation
scales, PDF$+\alpha_{\mathrm{s}}$ variations and PDF choice.
The scale uncertainties are estimated by varying both the renormalisation and factorisation
scales simultaneously up or down by a factor of two. The resulting maximum variation from the two fluctuations is taken as the symmetric scale uncertainty.
The PDF$+\alpha_{\mathrm{s}}$ uncertainty is evaluated using 90\% confidence level (CL) eigenvector and 90\% CL $\alpha_{\mathrm{s}}$ variations of the
nominal MSTW2008 NNLO PDF set and combined with the scale uncertainty in quadrature.
The PDF choice uncertainty is evaluated by comparing the central values of the MSTW2008 NNLO, CT10 NNLO, NNPDF 2.3 NNLO~\cite{nnpdf},
ABM11 5N NNLO~\cite{abm11} and HERAPDF 1.5 NNLO~\cite{Radescu:2011cn} PDF sets.
The envelope of the PDF central value comparisons and the combination of the scale and PDF$+\alpha_{\mathrm{s}}$
uncertainties is taken as the total uncertainty on the differential cross-section as a function of the invariant mass of the lepton--neutrino system ($m_{\ell\nu}$).
The PDF and $\alpha_{\mathrm{s}}$ uncertainties on the $t\bar{t}$ cross-section are calculated using the PDF4LHC prescription~\cite{pdf4lhc}
with the MSTW2008 68\% CL NNLO, CT10 NNLO and NNPDF2.3 5f FFN PDF error sets added in quadrature to the scale uncertainty.
The systematic uncertainty arising from the variation of the top mass by $\pm$1~GeV is also added in quadrature.
An additional uncertainty on the differential cross-section due to the beam energy uncertainty is calculated as function of $m_{\ell\nu}$
for the charged-current Drell--Yan process with VRAP at NNLO using CT10 NNLO PDFs by taking a 0.66\% uncertainty on the
energy of each 4 TeV proton beam as determined in ref. \cite{Wenninger:1546734}.
The size of this uncertainty is observed to be about 2\% (6\%) at $m_{\ell\nu} =$ 2 (3) TeV.
The calculated uncertainties are propagated to both the $W$ and \wps\ processes in order to derive uncertainties on the background levels as well as the signal selection efficiencies in each signal region.
Uncertainties are not reported on the cross-sections for the \wstar\ due to the breakdown of higher-order corrections for non-renormalisable models.
However, uncertainties on the signal selection efficiency for the \wstar\ are evaluated using the same relative differential cross-section uncertainty as for the \wp.
Uncertainties on DM production are evaluated using 68\% confidence level eigenvector variations of the nominal MSTW2008 LO PDF set
as in~\cite{atlasWIMPhadronicWZ}.
\FloatBarrier
\section{Event selection}
\label{sec:eventSel}
The primary vertex for each event is required to have at least three tracks with $\pt > 0.4$~\gev\ and
to have a longitudinal distance less than 200~mm from the centre of the collision region.
There are on average 20.7 interactions per event in the data used for this analysis.
The primary vertex is defined to be the one with the highest summed track $\pt^2$.
Spurious tails in the \met\ distribution, arising from calorimeter noise and other detector problems are suppressed by checking the quality
of each reconstructed jet and discarding events containing reconstructed jets of poor quality, following the description given in ref.~\cite{atlas:2012:jet_cleaning}.
In addition, the ID track associated with the electron or muon is required to be compatible
with originating from the primary vertex by requiring that the transverse distance of closest approach, $d_0$, satisfies $|d_0|<1~(0.2)$~mm
and longitudinal distance, $z_0$, satisfies $|z_0|<5~(1)$~mm for the electron (muon).
Events are required to have exactly one electron candidate with \mbox{$\et > 125\gev$} or one muon candidate with $\pt > 45\gev$ satisfying
these requirements and the identification criteria described in section~\ref{sec:trigAndReco}.
In the electron channel, events having additional electrons with $\et > 20\gev$,
passing all electron identification criteria, are discarded. Similarly, in the muon channel,
events having additional muon candidates with a \pt\ threshold of 20 GeV are discarded.
To suppress the multi-jet background, the lepton is required to be isolated. In the electron channel,
the isolation energy is measured with the calorimeter in a cone \mbox{$\Delta R = \sqrt{(\Delta \eta)^2 + (\Delta \varphi)^2} = 0.2$}
around the electron track, and the requirement is $\Sigma E^{\mathrm{calo}}_{\mathrm{T}}< 0.007\times\et + 5 \gev$,
where the sum includes all calorimeter energy clusters in the cone excluding those that are attributed to the electron.
The scaling of the isolation requirement with the electron \et\ reduces the efficiency loss due to radiation from the electron at high \et.
In the muon channel, the isolation energy is measured using ID tracks with $\pttrack > 1$~\gev\ in a cone $\Delta R = 0.3$ around the muon track.
The isolation requirement is \mbox{$\sumpttrack < 0.05 \times \pt$}, where the muon track is excluded from the sum.
As in the electron channel, the scaling of the isolation requirement with the muon \pt\ reduces the efficiency loss due to radiation from the muon at high \pt.
An \met\ requirement is imposed to select signal events and to further suppress the contributions from the multi-jet and SM $W$ backgrounds.
In both channels, the requirement placed on the charged lepton
\pt\ is also applied to the \met: $\met > 125 \gev$ for the electron channel and $\met > 45 \gev$ for the muon channel.
The multi-jet background\ around the Jacobian peak of the \mt\ distribution is evaluated using the \emph{matrix method} as described in ref.~\cite{atlas:ttbar2010} in both the electron and muon channels. The high-mass tail of the distribution is then fitted by a power-law function in order to determine the level of the multi-jet background\
in the region used to search for new physics.
In the electron channel, the multi-jet background\ constitutes about 2\%--4\% of the total background at high \mt. Consistent results are obtained using the
\emph{inverted isolation} technique described in ref.~\cite{atlas:wprime_2010_pub}.
In the muon channel, the multi-jet background\ constitutes about 1\%--3\% of the total background at high \mt.
The uncertainty of the multi-jet background\ is determined by varying the selection requirements used to define the control region and
by varying the \mt\ threshold of the fitting range used in the extrapolation to high \mt.
The same reconstruction criteria and event selection are applied to both the data and simulated samples.
Figure~\ref{fig:final_mt} shows the \pt, \met, and \mt\ spectra for each channel after event selection
for the data, the expected background and three examples of \wp\ signals at different masses.
Prior to investigating if there is evidence for a signal, the agreement between the data and the predicted background
is established for events with \mt\ $<\ $252~\gev, the lowest \mt\ threshold used to search for new physics.
The optimisation of the \mt\ thresholds for event selection is described below.
The agreement between the data and expected background is good.
Table~\ref{tab:nbg} shows an example of how different sources contribute to the background
for $\mt>1500 \gev$, the region used to search for a \wp\ with a mass of 2000~\gev.
The \wlnu\ background is the dominant contribution for both the electron and muon channels.
The \zll\ background in the electron channel is smaller than in the muon channel due to calorimeters
having larger $\eta$ coverage than the MS, and the electron energy resolution being better than the muon
momentum resolution at high \pt.
\begin{table}[!htbp]
\caption{
Expected numbers of events from the various background
sources in each decay channel
for \mbox{$\mt>1500 \gev$}, the region used to
search for a \wp\ with a mass of 2000~\gev.
The \wlnu\ and \zll\ rows include the expected contributions from the $\tau$-lepton.
The uncertainties are statistical.
}
\label{tab:nbg}
\begin{center}
\begin{tabular}{l | l|l}
\hline
\hline
& \multicolumn{1}{c|}{$e\nu$} & \multicolumn{1}{c}{$\mu\nu$} \\
\hline
\tspace
\wlnu & 2.65\phantom{000} $\pm$ 0.10 & 2.28\phantom{000} $\pm$ 0.21 \\
\zll & 0.00163\phantom{} $\pm$ 0.00022 & 0.232\phantom{00} $\pm$ 0.005 \\
Diboson & 0.27\phantom{000} $\pm$ 0.23 & 0.46\phantom{000} $\pm$ 0.23 \\
Top & 0.0056\phantom{0} $\pm$ 0.0009 & 0.0017 \phantom{0}$\pm$ 0.0001 \\
Multi-jet & 0.066\phantom{00} $\pm$ 0.020 & 0.046 \phantom{00}$\pm$ 0.039 \\
\hline
\tspace
Total & 2.99\phantom{000} $\pm$ 0.25 & 3.01\phantom{000} $\pm$ 0.31 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.49\textwidth]{fig_01a}
\includegraphics[width=0.49\textwidth]{fig_01b}
\includegraphics[width=0.49\textwidth]{fig_01c}
\includegraphics[width=0.49\textwidth]{fig_01d}
\includegraphics[width=0.49\textwidth]{fig_01e}
\includegraphics[width=0.49\textwidth]{fig_01f}
\caption{Spectra of lepton \pt\ (top), \met\ (centre) and \mt\ (bottom)
for the electron (left) and muon (right) channels after the event selection.
The spectra of \pt\ and \met\ are shown with the requirement \mt\ $>$ 252 \gev.
The points represent data and the filled, stacked histograms show the predicted backgrounds.
Open histograms are \wpl\ signals added to the background with their masses in \gev\ indicated in parentheses in the legend.
The signal and background samples are normalised using the integrated luminosity of the data
and the NNLO cross-sections listed in tables~\ref{tab:xsec_sig} and~\ref{tab:xsec_bg},
except for the multi-jet background\ which is estimated from data. The error bars on the data points
are statistical. The ratio of the data to the total background prediction is shown below each of the distributions.
The bands represent the systematic uncertainties on the background including the ones arising from the statistical uncertainty of
the simulated samples.
\label{fig:final_mt}
}
\end{figure*}
\FloatBarrier
\section{Statistical analysis and systematic uncertainties}
\label{sec:statAnaAndSys}
A Bayesian analysis is performed to set limits on the studied processes.
For each candidate mass and decay channel, events are counted above an \mt\ threshold.
The optimisation of \mtmin\ is done separately for \wpl\ and \wsl.
For each candidate mass, the \mtmin\ values that minimise the expected cross-section limits are obtained in the electron and muon channels separately,
but for simplicity the lower value is used in both channels since this has a negligible impact on the final results.
A similar optimisation is performed when setting the limits on DM production, and in this case a single \mtmin\ is chosen for each operator.
The expected number of events in each channel is
\begin{equation}
\nexp = \effsig \lint \xbr + \nbg,
\end{equation}
where \lint\ is the integrated luminosity of the data sample, \effsig\ is the signal selection efficiency
defined as the fraction of signal events that satisfy the event selection criteria as well as $\mt > \mtmin$, \nbg\ is the expected number of background events,
and \xbr\ is the cross-section times branching fraction.
Using Poisson statistics, the likelihood to observe \nobs\ events is
\begin{equation}
{\cal L}(\nobs|\xbr) = \frac{(\lint \effsig \xbr + \nbg)^{\nobs} \mathrm{e}^{-(\lint \effsig \xbr + \nbg)}}{\nobs!}.
\label{eqn:lhood}
\end{equation}
Uncertainties are included by introducing nuisance parameters \npi, each with
a probability density function \pdfi, and integrating the product of
the Poisson likelihood with the probability density function.
The integrated likelihood is
\small
\begin{equation}
{\cal L}_B(\nobs|\xbr) = \int {\cal L}(\nobs|\xbr) \prod \pdfi d\npi,
\label{eqn:lhoodn}
\end{equation}
\normalsize
where a log-normal distribution is used for the \pdfi.
The nuisance parameters are taken to be: \lint, \effsig\ and \nbg, with the appropriate correlation accounted for between the first and the third parameters.
The measurements in the two decay channels are combined assuming the same branching fraction for each.
Equation (\ref{eqn:lhoodn}) remains valid with the Poisson likelihood replaced by the product
of the Poisson likelihoods for the two channels.
The integrated luminosities for the electron and muon channels are fully correlated.
For \wpsl\ the signal selection efficiencies and background levels are partly correlated with each other and
between the two channels due to the full correlation of the cross-section uncertainties.
If these correlations were not included, the observed \xbr\ limits would improve by 25\%--30\% for the lowest mass points,
a few percent for the intermediate mass points and by about 10\% for the highest mass points.
Bayes' theorem gives the posterior probability that the signal has signal strength \xbr:
\begin{equation}
\ppost(\xbr|\nobs) = N {\cal L}_{B}(\nobs|\xbr) \, \pprior(\xbr)
\label{eqn:ppost}
\end{equation}
where $\pprior(\xbr)$ is the assumed prior probability, here chosen to be flat in \xbr,
for $\xbr > 0$.
The constant factor $N$ normalises the total probability to one.
The posterior probability is evaluated for each mass and decay channel as well as for their combination,
and then used to set a limit on \xbr.
The inputs for the evaluation of ${\cal L}_B$ (and hence \ppost) are \lint, \effsig, \nbg, \nobs\ and the uncertainties on
the first three. The uncertainties on \effsig\ and \nbg\ account for experimental and theoretical systematic effects as well as the statistics of the simulated samples.
The experimental systematic uncertainties include those on the efficiencies of the electron or muon trigger, reconstruction and event/object selection.
Uncertainties in the lepton energy/momentum and \met, characterised by scale and resolution uncertainties, are also included.
Performance metrics are obtained {\it in-situ} using well-known processes such as $Z\rightarrow \ell\ell$~\cite{atlas:egamreco, Aad:2014rra, atlas:jetmet}.
Since most of these performance metrics are measured at relatively low \pt\ their values are extrapolated to
the high-\pt\ regime relevant to this analysis using MC simulation.
The uncertainties in these extrapolations are included but are too small to significantly affect the results.
Table~\ref{tab:syst_summary} summarises the uncertainties on the event selection efficiencies
and the expected number of background events for the \mbox{\wpl} signal with $\mwp=2000\gev$ using $\mt>1500 \gev$, and $W^{*}$
signal with $\mws=2000\gev$ using $\mt>1337 \gev$.
\begin{table}[!htbp]
\caption{
Relative uncertainties on the selection efficiency \effsig\ and expected number of background events \nbg\
for a \wp\ (upper part of the table) and $W^{*}$ (lower part of the table) with a mass of 2000~\gev.
The efficiency uncertainties include contributions from the trigger, reconstruction and event selection.
The last row gives the total relative uncertainties.
\label{tab:syst_summary}
}
\begin{center}
\begin{tabular}{l|rr||rr}
\hline
\hline
& \multicolumn{2}{c|}{\effsig} & \multicolumn{2}{c}{\nbg} \\
Source & \multicolumn{1}{c}{$e\nu$} & \multicolumn{1}{c|}{$\mu\nu$}
& \multicolumn{1}{c}{$e\nu$} & \multicolumn{1}{c}{$\mu\nu$} \\
\hline
\wpl\ & \multicolumn{2}{c}{} \\
\hline
Reconstruction and trigger efficiency & 2.5\% & 4.1\% & 2.7\% & 4.1\% \\
Lepton energy/momentum resolution & 0.2\% & 1.4\% & 1.9\% & 18\%\\
Lepton energy/momentum scale & 1.2\% & 1.8\% & 3.5\% & 1.5\%\\
\met\ scale and resolution & 0.1\% & 0.1\% & 1.2\% & 0.5\% \\
Beam energy & 0.5\% & 0.5\% & 2.8\% & 2.1\%\\
Multi-jet background & \begin{tabular}{c}{-}\end{tabular} & \begin{tabular}{c}{-}\end{tabular} & 2.2\% & 3.4\% \\
Monte Carlo statistics & 0.9\% & 1.3\% & 8.5\% & 10\% \\
Cross-section (shape/level) & 2.9\% & 2.8\% & 18\% & 15\% \\
\hline
Total & 4.2\% & 5.6\% & 21\% & 27\% \\
\hline
\hline
\wsl\ & \multicolumn{2}{c}{} \\
\hline
Reconstruction and trigger efficiency & 2.7\% & 4.1\% & 2.6\% & 4.0\% \\
Lepton energy/momentum resolution & 0.4\% & 0.9\% & 3.0\% & 17\%\\
Lepton energy/momentum scale & 2.4\% & 2.4\% & 3.1\% & 1.5\%\\
\met\ scale and resolution & 0.1\% & 0.4\% & 3.1\% & 0.6\% \\
Beam energy & 0.1\% & 0.1\% & 2.5\% & 1.9\%\\
Multi-jet background & \begin{tabular}{c}{-}\end{tabular} & \begin{tabular}{c}{-}\end{tabular} & 1.8\% & 2.6\% \\
Monte Carlo statistics & 1.2\% & 1.8\% & 6.7\% & 8.6\% \\
Cross-section (shape/level) & 0.2\% & 0.2\% & 17\% & 15\% \\
\hline
Total & 3.9\% & 5.1\% & 19\% & 25\% \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\FloatBarrier
\section{Results}
\label{sec:results}
The inputs for the evaluation of ${\cal L}_B$ are listed in tables~\ref{tab:limit_input},~\ref{tab:limit_inputws} and~\ref{tab:limitdm_input}.
The uncertainties on \effsig\ and \nbg\ account for all relevant experimental and theoretical effects except for the uncertainty on the integrated luminosity.
The latter is included separately and is correlated between signal and background.
The tables also list the predicted numbers of signal events, \nsig, with their uncertainties accounting for the uncertainties in
both \effsig\ and the cross-section calculation.
The maximum value for the signal selection efficiency is at $\mwp = 2000\gev$.
For lower masses, the efficiency falls because the relative \mt\ threshold, $\mtmin/\mwp$,
increases in order to reduce the background level. The contribution from $W'\rightarrow \tau\nu$ with a leptonically decaying $\tau$ is neglected.
It would increase the signal yield by 2\%--3\% for the highest masses.
The background level is estimated for each mass by summing over all of the background sources.
\begin{table*}[!htbp]
\caption{Inputs for the \wpl\ \xbr\ limit calculations.
The first three columns are the \wp\ mass, \mt\ threshold and decay channel.
The next two are the signal selection efficiency, \effsig, and the prediction
for the number of signal events, \nsig, obtained with this efficiency.
The last two columns are the expected number of background
events, \nbg, and the number of events observed in data, \nobs.
The uncertainties on \nsig\ and \nbg\ include contributions from the uncertainties on the
cross-sections but not from that on the integrated luminosity.
\label{tab:limit_input}
}
\begin{center}
\begin{tabular}{ccc| r@{$ ~\pm~$}l r@{$~\pm~$}l r@{$~\pm~$}l r}
\hline
\hline
\mwp & \mtmin & Channel & \multicolumn{2}{c}{\effsig} & \multicolumn{2}{c}{\nsig} & \multicolumn{2}{c}{\nbg} & \nobs \\ \relax
[\gev] & [\gev] & \\
\hline
\hline
\multirow{2}{*}{300} & \multirow{2}{*}{252}
& $e\nu$ & 0.228 & 0.009 & 688000 & 28000 & 12900 & 820 & 12717\\
& & $\mu\nu$ & 0.184 & 0.007 & 555000 & 21000 & 11300 & 770 & 10927\\ \hline
\multirow{2}{*}{400} & \multirow{2}{*}{336}
& $e\nu$ & 0.319 & 0.012 & 325000 & 12000 & 5280 & 360 & 5176\\
& & $\mu\nu$ & 0.193 & 0.007 & 196000 & 7500 & 3490 & 250 & 3317\\ \hline
\multirow{2}{*}{500} & \multirow{2}{*}{423}
& $e\nu$ & 0.325 & 0.013 & 141000 & 5700 & 2070 & 150 & 2017\\
& & $\mu\nu$ & 0.186 & 0.007 & 80900 & 3200 & 1370 & 100 & 1219\\ \hline
\multirow{2}{*}{600} & \multirow{2}{*}{474}
& $e\nu$ & 0.397 & 0.014 & 83800 & 2900 & 1260 & 96 & 1214\\
& & $\mu\nu$ & 0.229 & 0.009 & 48200 & 1900 & 827 & 64 & 719\\ \hline
\multirow{2}{*}{750} & \multirow{2}{*}{597}
& $e\nu$ & 0.393 & 0.013 & 33200 & 1100 & 456 & 45 & 414\\
& & $\mu\nu$ & 0.226 & 0.009 & 19100 & 750 & 305 & 30 & 255\\ \hline
\multirow{2}{*}{1000} & \multirow{2}{*}{796}
& $e\nu$ & 0.386 & 0.012 & 9080 & 290 & 116 & 15 & 101\\
& & $\mu\nu$ & 0.219 & 0.009 & 5160 & 220 & 84 & 10 & 58\\ \hline
\multirow{2}{*}{1250} & \multirow{2}{*}{1002}
& $e\nu$ & 0.378 & 0.012 & 2980 & 98 & 35.3 & 5.8 & 34\\
& & $\mu\nu$ & 0.210 & 0.009 & 1650 & 73 & 28.3 & 4.6 & 19\\ \hline
\multirow{2}{*}{1500} & \multirow{2}{*}{1191}
& $e\nu$ & 0.376 & 0.014 & 1110 & 40 & 13.2 & 2.5 & 14\\
& & $\mu\nu$ & 0.206 & 0.010 & 610 & 30 & 10.9 & 2.3 & 6\\ \hline
\multirow{2}{*}{1750} & \multirow{2}{*}{1416}
& $e\nu$ & 0.336 & 0.013 & 396 & 16 & 4.56 & 0.92 & 5\\
& & $\mu\nu$ & 0.182 & 0.010 & 214 & 12 & 4.3 & 1.1 & 0\\ \hline
\multirow{2}{*}{2000} & \multirow{2}{*}{1500}
& $e\nu$ & 0.370 & 0.015 & 183.0 & 7.7 & 2.99 & 0.61 & 3\\
& & $\mu\nu$ & 0.198 & 0.011 & 98.0 & 5.5 & 3.01 & 0.80 & 0\\ \hline
\multirow{2}{*}{2250} & \multirow{2}{*}{1683}
& $e\nu$ & 0.327 & 0.015 & 71.5 & 3.3 & 1.38 & 0.33 & 0\\
& & $\mu\nu$ & 0.173 & 0.011 & 37.9 & 2.3 & 1.44 & 0.33 & 0\\ \hline
\multirow{2}{*}{2500} & \multirow{2}{*}{1888}
& $e\nu$ & 0.262 & 0.018 & 27.1 & 1.8 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.140 & 0.012 & 14.4 & 1.2 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{2750} & \multirow{2}{*}{1888}
& $e\nu$ & 0.235 & 0.024 & 12.3 & 1.3 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.127 & 0.014 & 6.64 & 0.74 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3000} & \multirow{2}{*}{1888}
& $e\nu$ & 0.183 & 0.029 & 5.33 & 0.86 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.100 & 0.016 & 2.93 & 0.48 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3250} & \multirow{2}{*}{1888}
& $e\nu$ & 0.124 & 0.033 & 2.22 & 0.59 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.069 & 0.018 & 1.24 & 0.32 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3500} & \multirow{2}{*}{1888}
& $e\nu$ & 0.077 & 0.031 & 0.92 & 0.36 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.044 & 0.017 & 0.52 & 0.20 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3750} & \multirow{2}{*}{1888}
& $e\nu$ & 0.047 & 0.024 & 0.40 & 0.21 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.028 & 0.013 & 0.24 & 0.11 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{4000} & \multirow{2}{*}{1888}
& $e\nu$ & 0.031 & 0.018 & 0.20 & 0.11 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.019 & 0.010 & 0.121 & 0.061 & 0.61 & 0.15 & 0\\ \hline
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[!htbp]
\caption{Inputs for the \wsl\ \xbr\ limit calculations.
The columns are the same as in table \ref{tab:limit_input}.
\label{tab:limit_inputws}
}
\begin{center}
\begin{tabular}{ccc| r@{$~\pm~$}l r@{$~\pm~$}l r@{$~\pm~$}l r}
\hline
\hline
\mws & \mtmin & Channel & \multicolumn{2}{c}{\effsig} & \multicolumn{2}{c}{\nsig} & \multicolumn{2}{c}{\nbg} & \nobs \\ \relax
[\gev] & [\gev] & \\
\hline \hline
\multirow{2}{*}{400} & \multirow{2}{*}{317}
& $e\nu$ & 0.196 & 0.010 & 149000 & 7400 & 6630 & 440 & 6448\\
& & $\mu\nu$ & 0.111 & 0.005 & 84900 & 3700 & 4420 & 310 & 4230\\ \hline
\multirow{2}{*}{500} & \multirow{2}{*}{377}
& $e\nu$ & 0.246 & 0.011 & 80900 & 3500 & 3320 & 220 & 3275\\
& & $\mu\nu$ & 0.140 & 0.006 & 45900 & 1900 & 2210 & 160 & 2008\\ \hline
\multirow{2}{*}{600} & \multirow{2}{*}{448}
& $e\nu$ & 0.257 & 0.011 & 41400 & 1800 & 1630 & 120 & 1582\\
& & $\mu\nu$ & 0.144 & 0.006 & 23200 & 960 & 1080 & 79 & 938\\ \hline
\multirow{2}{*}{750} & \multirow{2}{*}{564}
& $e\nu$ & 0.248 & 0.011 & 15900 & 680 & 593 & 54 & 524\\
& & $\mu\nu$ & 0.143 & 0.006 & 9200 & 400 & 388 & 35 & 321\\ \hline
\multirow{2}{*}{1000} & \multirow{2}{*}{710}
& $e\nu$ & 0.302 & 0.013 & 5390 & 230 & 203 & 24 & 177\\
& & $\mu\nu$ & 0.174 & 0.007 & 3100 & 130 & 143 & 17 & 109\\ \hline
\multirow{2}{*}{1250} & \multirow{2}{*}{843}
& $e\nu$ & 0.337 & 0.013 & 2010 & 79 & 86 & 12 & 79\\
& & $\mu\nu$ & 0.191 & 0.008 & 1140 & 50 & 65.5 & 8.5 & 40\\ \hline
\multirow{2}{*}{1500} & \multirow{2}{*}{1062}
& $e\nu$ & 0.296 & 0.011 & 648 & 25 & 25.8 & 4.4 & 26\\
& & $\mu\nu$ & 0.164 & 0.007 & 360 & 16 & 20.9 & 3.8 & 12\\ \hline
\multirow{2}{*}{1750} & \multirow{2}{*}{1191}
& $e\nu$ & 0.324 & 0.013 & 278 & 11 & 13.2 & 2.5 & 14\\
& & $\mu\nu$ & 0.182 & 0.009 & 156.0 & 7.6 & 10.9 & 2.3 & 6\\ \hline
\multirow{2}{*}{2000} & \multirow{2}{*}{1337}
& $e\nu$ & 0.341 & 0.013 & 118.0 & 4.6 & 6.8 & 1.3 & 9\\
& & $\mu\nu$ & 0.186 & 0.010 & 64.6 & 3.3 & 5.8 & 1.4 & 3\\ \hline
\multirow{2}{*}{2250} & \multirow{2}{*}{1416}
& $e\nu$ & 0.391 & 0.014 & 55.5 & 2.0 & 4.56 & 0.92 & 5\\
& & $\mu\nu$ & 0.204 & 0.010 & 28.9 & 1.5 & 4.3 & 1.1 & 0\\ \hline
\multirow{2}{*}{2500} & \multirow{2}{*}{1683}
& $e\nu$ & 0.337 & 0.013 & 19.80 & 0.76 & 1.38 & 0.33 & 0\\
& & $\mu\nu$ & 0.179 & 0.010 & 10.50 & 0.57 & 1.44 & 0.33 & 0\\ \hline
\multirow{2}{*}{2750} & \multirow{2}{*}{1888}
& $e\nu$ & 0.322 & 0.013 & 7.84 & 0.31 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.161 & 0.011 & 3.92 & 0.27 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3000} & \multirow{2}{*}{1888}
& $e\nu$ & 0.382 & 0.015 & 3.80 & 0.15 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.185 & 0.011 & 1.84 & 0.11 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3250} & \multirow{2}{*}{1888}
& $e\nu$ & 0.437 & 0.018 & 1.770 & 0.073 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.218 & 0.014 & 0.880 & 0.056 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3500} & \multirow{2}{*}{1888}
& $e\nu$ & 0.474 & 0.025 & 0.766 & 0.040 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.229 & 0.016 & 0.371 & 0.027 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{3750} & \multirow{2}{*}{1888}
& $e\nu$ & 0.498 & 0.055 & 0.320 & 0.035 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.244 & 0.029 & 0.157 & 0.019 & 0.61 & 0.15 & 0\\ \hline
\multirow{2}{*}{4000} & \multirow{2}{*}{1888}
& $e\nu$ & 0.487 & 0.150 & 0.124 & 0.038 & 0.432 & 0.091 & 0\\
& & $\mu\nu$ & 0.242 & 0.073 & 0.062 & 0.019 & 0.61 & 0.15 & 0\\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[!htbp]
\caption{Inputs to the limit calculations on the pair production of DM particles for the operators D1, D5d, D5c and D9.
Expected number of signal events for each operator is calculated for a different value of the mass scale,
notably $M_{*}=$ 10 GeV for D1, $M_{*}=$ 100 GeV for D5d, and \mbox{$M_{*}=$ 1 TeV} for operators D9 and D5c.
The columns are the same as in table \ref{tab:limit_input}.
\label{tab:limitdm_input}
}
\begin{center}
\footnotesize
\begin{tabular}{ccc| ll l l |lcc}
\hline
\hline
$m_{\chi}$ & \mtmin & Channel & \multicolumn{2}{c}{\effsig} & \multicolumn{2}{c}{\nsig} & \multicolumn{2}{c}{\nbg} & \nobs \\
~[\gev] & [\gev] & \\
\hline
\hline
& & \multicolumn{8}{c}{D1 Operator}\\
\hline
\multirow{2}{*}{1} & \multirow{12}{*}{796}
& $e\nu$ & {$0.0294\pm0.0044$}& & {$87000\pm13000$}& & \multirow{12}{*}{$e\nu$} &\multirow{12}{*}{$116\pm15$} & \multirow{12}{*}{101}\\
& & $\mu\nu$& {$0.0177\pm0.0023$}& & {$52500\pm7000$}& & \multirow{12}{*}{$\mu\nu$}&\multirow{12}{*}{\phantom{0}$84\pm10$}& \multirow{12}{*}{58}\\
\multirow{2}{*}{100} &
& $e\nu$ & {$0.0396\pm0.0052$}& & {$89000\pm12000$}& \\
& & $\mu\nu$ & {$0.0252\pm0.0033$}& & {$56600\pm7500$}& \\
\multirow{2}{*}{200} &
& $e\nu$ & {$0.0484\pm0.0057$}& & {$65800\pm7700$}& \\
& & $\mu\nu$& {$0.0293\pm0.0034$}& & {$39900\pm4600$}& \\
\multirow{2}{*}{400} &
& $e\nu$ & {$0.0709\pm0.0071$}& & {$30900\pm3100$}& \\
& & $\mu\nu$& {$0.0398\pm0.0041$}& & {$17300\pm1800$}& \\
\multirow{2}{*}{1000} &
& $e\nu$ & {$0.0989\pm0.0100$}& & {\phantom{0}$1070\pm110$}& \\
& & $\mu\nu$& {$0.0621\pm0.0068$}& & {\phantom{00}$673\pm73$}& \\
\multirow{2}{*}{1300} &
& $e\nu$ & {$0.0964\pm0.0095$}& & {\phantom{00}$138\pm14$}& \\
& & $\mu\nu$& {$0.0522\pm0.0048$}& & {\phantom{ 0}$75.1\pm6.9$}& \\
\hline
& & \multicolumn{8}{c}{D5d Operator}\\
\hline
\multirow{2}{*}{1} & \multirow{12}{*}{597}
& $e\nu$ & {$0.0148\pm0.0016$}& & {\phantom{0}$7230\pm800$}& & \multirow{12}{*}{$e\nu$} & \multirow{12}{*}{$456\pm45$} & \multirow{12}{*}{414} \\
& & $\mu\nu$& {$0.0080\pm0.0011$}& & {\phantom{0}$3890\pm530$}& & \multirow{12}{*}{$\mu\nu$}& \multirow{12}{*}{$305\pm30$} & \multirow{12}{*}{255}\\
\multirow{2}{*}{100} &
& $e\nu$ & {$0.0158\pm0.0018$}& & {\phantom{0}$7580\pm850$}& \\
& & $\mu\nu$& {$0.0096\pm0.0012$}& & {\phantom{0}$4600\pm580$}& \\
\multirow{2}{*}{200} &
& $e\nu$ & {$0.0147\pm0.0015$}& & {\phantom{0}$5850\pm610$}& \\
& & $\mu\nu$& {$0.0086\pm0.0011$}& & {\phantom{0}$3420\pm430$}& \\
\multirow{2}{*}{400} &
& $e\nu$ & {$0.0190\pm0.0020$}& & {\phantom{0}$4220\pm440$}& \\
& & $\mu\nu$& {$0.0113\pm0.0013$}& & {\phantom{0}$2500\pm300$}& \\
\multirow{2}{*}{1000} &
& $e\nu$ & {$0.0281\pm0.0025$}& & {\phantom{00}$450\pm41$}& \\
& & $\mu\nu$& {$0.0177\pm0.0019$}& & {\phantom{00}$283\pm30$}& \\
\multirow{2}{*}{1300} &
& $e\nu$ & {$0.0291\pm0.0028$}& & {\phantom{ 0}$89.3\pm8.5$}& \\
& & $\mu\nu$& {$0.0167\pm0.0018$}& & {\phantom{ 0}$51.1\pm5.4$}& \\
\hline
& & \multicolumn{8}{c}{D5c Operator}\\
\hline
\multirow{2}{*}{1} & \multirow{12}{*}{843}
& $e\nu$ & {$0.0737\pm0.0047$} & & {$\phantom{0}30.3\pm1.9$}& &\multirow{12}{*}{$e\nu$} & \multirow{12}{*}{$86\pm12$} & \multirow{12}{*}{79}\\
& & $\mu\nu$& {$0.0435\pm0.0034$}& & {$\phantom{0}17.9\pm1.4$}& & \multirow{12}{*}{$\mu\nu$}& \multirow{12}{*}{\phantom{00}$65.5\pm8.5$} & \multirow{12}{*}{40}\\
\multirow{2}{*}{100} &
& $e\nu$ & {$0.0798\pm0.0050$}& & {$\phantom{0}31.0\pm1.9$}& \\
& & $\mu\nu$& {$0.0437\pm0.0034$}& & {$\phantom{0}17.0\pm1.3$}& \\
\multirow{2}{*}{200} &
& $e\nu$ & {$0.0762\pm0.0049$}& & {$\phantom{0}25.1\pm1.6$}& \\
& & $\mu\nu$& {$0.0461\pm0.0034$}& & {$\phantom{0}15.2\pm1.1$}& \\
\multirow{2}{*}{400} &
& $e\nu$ & {$0.0857\pm0.0055$}& & {$\phantom{0}16.2\pm1.0$}& \\
& & $\mu\nu$ & {$0.0532\pm0.0040$}& & {$\phantom{0}10.0\pm0.8$}& \\
\multirow{2}{*}{1000} &
& $e\nu$ & {$0.0987\pm0.0091$}& & {$\phantom{0}1.28\pm0.12$}& \\
& & $\mu\nu$& {$0.0636\pm0.0057$}& & {$\phantom{}0.824\pm0.074$}& \\
\multirow{2}{*}{1300} &
& $e\nu$ & {$0.1010\phantom{}\pm0.0095$}& & {$\phantom{}0.240\pm0.023$}& \\
& & $\mu\nu$& {$0.0589\pm0.0057$}& & {$\phantom{}0.140\pm0.014$}& \\
\hline
& & \multicolumn{8}{c}{D9 Operator}\\
\hline
\multirow{2}{*}{1} & \multirow{12}{*}{843}
& $e\nu$ & {$0.0851\pm0.0053$}& & {$\phantom{0}55.5\pm3.5$}& & \multirow{12}{*}{$e\nu$}&\multirow{12}{*}{$86\pm12$} & \multirow{12}{*}{79}\\
& & $\mu\nu$& {$0.0517\pm0.0035$}& & {$\phantom{0}33.8\pm2.3$}& &\multirow{12}{*}{$\mu\nu$}& \multirow{12}{*}{\phantom{00}$65.5\pm8.5$} & \multirow{12}{*}{40}\\
\multirow{2}{*}{100} &
& $e\nu$ & {$0.0950\pm0.0056$}& & {$\phantom{0}55.8\pm3.3$}& \\
& & $\mu\nu$& {$0.0529\pm0.0038$}& & {$\phantom{0}31.1\pm2.3$}& \\
\multirow{2}{*}{200} &
& $e\nu$ & {$0.1040\phantom{}\pm0.0062$}& & {\phantom{0}$48.9\pm2.9$}& \\
& & $\mu\nu$& {$0.0553\pm0.0039$}& & {$\phantom{0}26.0\pm1.8$}& \\
\multirow{2}{*}{400} &
& $e\nu$ & {$0.1030\phantom{}\pm0.0067$}& & {$\phantom{0}25.5\pm1.6$}& \\
& & $\mu\nu$& {$0.0578\pm0.0042$}& & {$\phantom{0}14.3\pm1.0$}& \\
\multirow{2}{*}{1000} &
& $e\nu$ & {$0.1070\pm0.0092$}& & {$\phantom{0}1.63\pm0.14$}& \\
& & $\mu\nu$& 0{$.0615\pm0.0055$}& & {$\phantom{}0.944\pm0.084$}& \\
\multirow{2}{*}{1300} &
& $e\nu$ & {$0.1020\pm0.0100$}& & {$\phantom{}0.285\pm0.029$}& \\
& & $\mu\nu$ & {$0.0573\pm0.0056$}& & {$\phantom{}0.160\pm0.016$}& \\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
\FloatBarrier
The number of observed events is generally in good agreement with the expected number of background events for all
mass bins. None of the observations for any mass point in either channel or their combination show a significant
excess above background, so there is no evidence for the observation of either \wpl\ or \wsl. A deficit in the number of observed events
with respect to the expected number of background events is observed in the muon channel. This deficit has at most a 2.2$\sigma$ local significance.
Tables~\ref{tab:limits_xbrfid_wp_and_ws1} and~\ref{tab:limits_xbrfid_wp_and_ws2} and figure~\ref{fig:final_limits} present the
95\% confidence level (CL) observed limits on \xbr\ for both \wpl\ and \wsl\
in the electron channel, the muon channel and their combination.
The tables also give the limits obtained without systematic uncertainties.
Limits with various subsets of the systematic uncertainties are shown for \wpl\ as a representative case.
The uncertainties on the signal selection efficiency have very little effect on the final limits, and the
background-level and luminosity uncertainties are important only for the lowest masses.
Figure~\ref{fig:final_limits} also shows the expected limits and the theoretical \xbr\ for a \wp\
and for a \wstar. Limits are evaluated by fixing the \wstar\ coupling strengths to give the same partial decay
widths as the \wp. The off-shell production of \wp\ degrades the acceptance at high mass,
worsening the limits. As discussed in chapter~\ref{sec:intro}, \wstar\ has different couplings with respect to \wp,
enhancing the production at the pole. Since the off-shell production is reduced with respect to \wp,
the \wstar\ limits do not show the same behaviour at high mass.
\begin{table}[!htbp]
\caption{
Observed upper limits on \xbr\ for \wp\ and \wstar\ with masses up to 2000 GeV. The first column is
the \wps\ mass and the following columns refer to the 95\% CL limits for the \wp\ with
headers indicating the nuisance parameters for which uncertainties
are included: S for the event selection efficiency (\effsig), B
for the background level (\nbg), and L for the integrated
luminosity (\lint). The column labelled SBL includes all
uncertainties neglecting correlations. Results are also presented when including the correlation of the
signal and background cross-section uncertainties, as well as the correlation of the background cross-section uncertainties for the combined limits
($\mathrm{SB_c}$, $\mathrm{SB_cL}$). The last two columns show
the limits for the $W^{*}$ without nuisance parameters and when including all nuisance parameters
with correlations.
\label{tab:limits_xbrfid_wp_and_ws1}
}
\begin{center}
\footnotesize
\begin{tabular}{cc|rrrrrr|rr}
\hline
\hline
\mwps ~[\gev] & Channel & \multicolumn{8}{c}{95\% CL limit on \xbr\ [fb]} \\
& & \multicolumn{6}{c}{\wp\ } & \multicolumn{2}{c}{$W^{*}$} \\
& &none & S & SB & SBL & $\mathrm{SB_c}$ & $\mathrm{SB_cL}$ & none & $\mathrm{SB_cL}$\\
\hline
\multirow{3}{*}{300}
& $e\nu$ & 29.0 & 29.1 & 304 & 342 & 305 & 343 \\
& $\mu\nu$ & 22.4 & 22.4 & 327 & 363 & 327 & 363 \\
& both & 14.2 & 14.2 & 219 & 269 & 290 & 331 \\
\hline
\multirow{3}{*}{400}
& $e\nu$ & 14.1 & 14.1 & 94.8 & 105 & 95.0 & 105 & 20.7 & 204 \\
& $\mu\nu$ & 12.6 & 12.6 & 91.3 & 102 & 91.4 & 102 & 25.1 & 233 \\
& both & 7.55 & 7.56 & 63.4 & 77.0 & 83.2 & 94.7 & 12.6 & 197 \\
\hline
\multirow{3}{*}{500}
& $e\nu$ & 9.14 & 9.18 & 38.7 & 42.2 & 38.8 & 42.4 & 17.3 & 87.5 \\
& $\mu\nu$ & 6.42 & 6.44 & 30.6 & 34.0 & 30.7 & 34.1 & 10.5 & 77.9 \\
& both & 4.26 & 4.26 & 22.3 & 27.0 & 29.8 & 33.9 & 7.54 & 77.7 \\
\hline
\multirow{3}{*}{600}
& $e\nu$ & 5.67 & 5.68 & 19.5 & 21.2 & 19.7 & 21.4 & 10.4 & 43.9 \\
& $\mu\nu$ & 4.38 & 4.40 & 15.5 & 17.0 & 15.6 & 17.1 & 7.11 & 32.8 \\
& both & 2.78 & 2.78 & 11.1 & 13.2 & 15.5 & 17.4 & 4.75 & 33.9 \\
\hline
\multirow{3}{*}{750}
& $e\nu$ & 2.95 & 2.95 & 8.25 & 8.71 & 8.35 & 8.81 & 4.23 & 14.9 \\
& $\mu\nu$ & 3.33 & 3.34 & 7.89 & 8.35 & 7.97 & 8.43 & 5.23 & 14.7 \\
& both & 1.73 & 1.73 & 5.06 & 5.63 & 7.01 & 7.52 & 2.51 & 12.8 \\
\hline
\multirow{3}{*}{1000}
& $e\nu$ & 1.84 & 1.85 & 3.25 & 3.34 & 3.29 & 3.38 & 2.69 & 6.01 \\
& $\mu\nu$ & 1.86 & 1.87 & 2.87 & 2.95 & 2.92 & 3.00 & 3.02 & 5.88 \\
& both & 1.03 & 1.04 & 1.86 & 1.96 & 2.48 & 2.58 & 1.57 & 4.94 \\
\hline
\multirow{3}{*}{1250}
& $e\nu$ & 1.63 & 1.64 & 2.06 & 2.09 & 2.09 & 2.12 & 2.29 & 3.65 \\
& $\mu\nu$ & 1.62 & 1.62 & 2.01 & 2.04 & 2.04 & 2.07 & 1.78 & 2.60 \\
& both & 0.990 & 0.991 & 1.30 & 1.34 & 1.54 & 1.57 & 1.16 & 2.53 \\
\hline
\multirow{3}{*}{1500}
& $e\nu$ & 1.27 & 1.28 & 1.40 & 1.41 & 1.42 & 1.43 & 1.99 & 2.39 \\
& $\mu\nu$ & 1.21 & 1.22 & 1.35 & 1.36 & 1.37 & 1.38 & 1.71 & 2.06 \\
& both & 0.775 & 0.777 & 0.879 & 0.890 & 0.967 & 0.979 & 1.14 & 1.63 \\
\hline
\multirow{3}{*}{1750}
& $e\nu$ & 0.964 & 0.967 & 0.993 & 0.997 & 1.01 & 1.01 & 1.48 & 1.64 \\
& $\mu\nu$ & 0.813 & 0.818 & 0.818 & 0.821 & 0.827 & 0.831 & 1.37 & 1.54 \\
& both & 0.521 & 0.522 & 0.533 & 0.537 & 0.563 & 0.567 & 0.889 & 1.10 \\
\hline
\multirow{3}{*}{2000}
& $e\nu$ & 0.721 & 0.724 & 0.735 & 0.738 & 0.743 & 0.746 & 1.34 & 1.40 \\
& $\mu\nu$ & 0.747 & 0.751 & 0.751 & 0.754 & 0.760 & 0.762 & 1.18 & 1.26 \\
& both & 0.415 & 0.416 & 0.422 & 0.424 & 0.439 & 0.441 & 0.831 & 0.922 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!htbp]
\caption{
Observed upper limits on \xbr\ for \wp\ and \wstar\ with masses above 2000 GeV.
The columns are the same as in table \ref{tab:limits_xbrfid_wp_and_ws1}.
\label{tab:limits_xbrfid_wp_and_ws2}
}
\begin{center}
\footnotesize
\begin{tabular}{cc|rrrrrr|rr}
\hline
\hline
\mwps ~[\gev] & Channel & \multicolumn{8}{c}{95\% CL limit on \xbr\ [fb]} \\
& & \multicolumn{6}{c}{\wp\ } & \multicolumn{2}{c}{$W^{*}$} \\
& &none & S & SB & SBL & $\mathrm{SB_c}$ & $\mathrm{SB_cL}$ & none & $\mathrm{SB_cL}$\\
\hline
\multirow{3}{*}{2250}
& $e\nu$ & 0.453 & 0.455 & 0.455 & 0.456 & 0.458 & 0.459 & 0.830 & 0.859 \\
& $\mu\nu$ & 0.853 & 0.859 & 0.859 & 0.862 & 0.866 & 0.869 & 0.726 & 0.734 \\
& both & 0.296 & 0.297 & 0.297 & 0.298 & 0.301 & 0.303 & 0.457 & 0.488 \\
\hline
\multirow{3}{*}{2500}
& $e\nu$ & 0.564 & 0.569 & 0.569 & 0.570 & 0.572 & 0.573 & 0.438 & 0.441 \\
& $\mu\nu$ & 1.06 & 1.07 & 1.07 & 1.08 & 1.08 & 1.08 & 0.828 & 0.837 \\
& both & 0.368 & 0.370 & 0.370 & 0.371 & 0.376 & 0.377 & 0.287 & 0.289 \\
\hline
\multirow{3}{*}{2750}
& $e\nu$ & 0.629 & 0.643 & 0.643 & 0.644 & 0.648 & 0.649 & 0.459 & 0.462 \\
& $\mu\nu$ & 1.16 & 1.19 & 1.19 & 1.20 & 1.21 & 1.21 & 0.917 & 0.928 \\
& both & 0.409 & 0.413 & 0.413 & 0.414 & 0.425 & 0.426 & 0.306 & 0.308 \\
\hline
\multirow{3}{*}{3000}
& $e\nu$ & 0.809 & 0.852 & 0.852 & 0.853 & 0.863 & 0.865 & 0.387 & 0.389 \\
& $\mu\nu$ & 1.47 & 1.55 & 1.55 & 1.56 & 1.58 & 1.58 & 0.798 & 0.807 \\
& both & 0.523 & 0.534 & 0.534 & 0.536 & 0.566 & 0.567 & 0.261 & 0.263 \\
\hline
\multirow{3}{*}{3250}
& $e\nu$ & 1.20 & 1.37 & 1.37 & 1.37 & 1.40 & 1.40 & 0.338 & 0.340 \\
& $\mu\nu$ & 2.14 & 2.45 & 2.45 & 2.45 & 2.52 & 2.52 & 0.678 & 0.687 \\
& both & 0.768 & 0.815 & 0.815 & 0.816 & 0.919 & 0.920 & 0.226 & 0.228 \\
\hline
\multirow{3}{*}{3500}
& $e\nu$ & 1.92 & 2.56 & 2.56 & 2.56 & 2.64 & 2.64 & 0.312 & 0.315 \\
& $\mu\nu$ & 3.37 & 4.38 & 4.38 & 4.39 & 4.56 & 4.57 & 0.645 & 0.655 \\
& both & 1.22 & 1.38 & 1.38 & 1.38 & 1.72 & 1.73 & 0.210 & 0.213 \\
\hline
\multirow{3}{*}{3750}
& $e\nu$ & 3.12 & 4.90 & 4.90 & 4.90 & 5.07 & 5.08 & 0.297 & 0.307 \\
& $\mu\nu$ & 5.32 & 7.85 & 7.85 & 7.86 & 8.22 & 8.24 & 0.605 & 0.630 \\
& both & 1.97 & 2.37 & 2.37 & 2.38 & 3.26 & 3.27 & 0.199 & 0.208 \\
\hline
\multirow{3}{*}{4000}
& $e\nu$ & 4.76 & 8.07 & 8.07 & 8.09 & 8.38 & 8.40 & 0.304 & 0.372 \\
& $\mu\nu$ & 7.75 & 12.0 & 12.0 & 12.0 & 12.6 & 12.6 & 0.613 & 0.749 \\
& both & 2.95 & 3.66 & 3.66 & 3.66 & 5.24 & 5.24 & 0.203 & 0.255 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.49\textwidth]{fig_02a}
\includegraphics[width=0.49\textwidth]{fig_02b}
\includegraphics[width=0.49\textwidth]{fig_02c}
\includegraphics[width=0.49\textwidth]{fig_02d}
\includegraphics[width=0.49\textwidth]{fig_02e}
\includegraphics[width=0.49\textwidth]{fig_02f}
\caption{Observed and expected limits on \xbr\ for \wp\ (left) and $W^{*}$ (right) at 95\% CL in the electron
channel (top), muon channel (centre) and the combination (bottom) assuming the same branching fraction for both channels.
The predicted values for \xbr\ and their uncertainties (except for $W^{*}$) are also shown.
The calculation of uncertainties on the \wp\ cross-sections is explained in section \ref{sec:mcSim}.
\label{fig:final_limits}
}
\end{figure*}
\begin{table}[!htbp]
\caption{Lower limits on the \wp\ and $W^{*}$ masses. The
first column is the decay channel ($e\nu$, $\mu\nu$ or both
combined) and the following give the expected (Exp.) and observed
(Obs.) mass limits.
\label{tab:limits_mass_wp_and_ws}
}
\vspace{3 mm}
\centering
\begin{tabular}{c|rr|rr}
\hline
\hline
& \multicolumn{2}{c}{\mwp\ [\tev]} & \multicolumn{2}{c}{\mws\ [\tev]} \\
Decay & Exp. & Obs. & Exp. & Obs. \\
\hline
$e\nu$ & 3.13 & 3.13 & 3.08 & 3.08 \\
$\mu\nu$ & 2.97 & 2.97 & 2.83 & 2.83 \\
Both & 3.17 & 3.24 & 3.12 & 3.21 \\
\hline
\hline
\end{tabular}
\end{table}
In figure~\ref{fig:final_limits} the intersection between the central theoretical prediction and
the observed limits provides the 95\% CL lower limits on the mass.
The expected and observed \wp\ and \wstar\ mass limits for the electron and muon decay channels
as well as their combination are listed in table~\ref{tab:limits_mass_wp_and_ws}.
The difference between the expected and observed combined mass limits originate from the
slight data deficit in each decay channel that are individually not significant.
The band around the theoretical prediction in figure~\ref{fig:final_limits} indicates the total theory uncertainty as described earlier in the text.
The mass limit for the \wp\ decreases by 50 \gev\ if the intersection between the lower theoretical prediction and the observed limit is used.
The uncertainties on \effsig, \nbg\ and \lint\ affect the derived mass limits by a similar amount.
Limits are also evaluated following the $\mathrm{CL_{s}}$ prescription~\cite{aread} using the profile likelihood ratio as the test statistic including
all uncertainties. The cross-section limits are found to agree within 10\% across the entire mass range, with only marginal impact on the mass limit.
The mass limits presented here are a significant improvement over those reported in previous ATLAS and CMS searches
\cite{atlas:wprime_2010_pub,atlas:wprime_2011-2_pub,atlas:wprime_2012_pub,cms:wprime2013-1}.
The results of the search for pair production of DM particles in association with a leptonically decaying $W$ boson
are shown in figures~\ref{fig:mstar_vs_mchi} and~\ref{fig:chi_nucleon_xsec}.
The former shows the observed limits on $M_{*}$, the mass scale of the unknown mediating interaction for the DM particle pair production,
whereas the latter shows the observed limits on the DM--nucleon scattering cross-section. Both are shown as a function of the DM particle mass, $m_{\chi}$,
and presented at 90\% CL. Results of the previous ATLAS searches for hadronically decaying $W/Z$~\cite{atlasWIMPhadronicWZ}, leptonically decaying
$Z$~\cite{atlasWIMPleptonicZ}, and $j+\chi\chi$~\cite{atlasWIMPjet} are also shown. The observed limits on $M_{*}$ as a function of $m_{\chi}$
are by a factor $\sim$1.5 stronger in the search for DM production in association with hadronically decaying $W$ with respect the ones
presented in this paper.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.8\textwidth]{fig_03}
\caption{Observed limits on $M_{*}$ as a function of the DM particle mass ($m_{\chi}$) at 90\% CL
for the combination of the electron and muon channel,
for various operators as described in the text.
For each operator, the values below the corresponding line are excluded.
No signal samples are generated for masses below 1 \GeV\ but the limits are expected to be stable down to arbitrarily small values.
Results of the previous ATLAS searches for hadronically decaying $W/Z$~\cite{atlasWIMPhadronicWZ} and
leptonically decaying $Z$~\cite{atlasWIMPleptonicZ} are also shown.
\label{fig:mstar_vs_mchi}
}
\end{figure*}
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.9\textwidth]{fig_04}
\caption{Observed limits on the DM--nucleon scattering cross-section as a function of $m_{\chi}$
at 90\% CL for spin-independent (left) and spin-dependent (right) operators in the EFT.
Results are compared with the previous ATLAS searches for hadronically decaying $W/Z$~\cite{atlasWIMPhadronicWZ},
leptonically decaying $Z$~\cite{atlasWIMPleptonicZ}, and $j+\chi\chi$~\cite{atlasWIMPjet}, and with direct detection searches by
CoGeNT~\cite{2011PhRvL.106m1301A}, XENON100~\cite{2012PhRvL.109r1301A},
CDMS~\cite{2014PhRvL.112d1302A,2014arXiv1402.7137A}, LUX~\cite{2013arXiv1310.8214L},
COUPP~\cite{2012PhRvD..86e2001B}, SIMPLE~\cite{2012PhRvL.108t1302F},
PICASSO~\cite{2012PhLB..711..153A} and IceCube~\cite{2012arXiv1212.4097I}.
The comparison between direct detection and ATLAS results is only possible within the limits of the
validity of the EFT~\cite{wimp14TeV}.
\label{fig:chi_nucleon_xsec}}
\end{figure*}
\FloatBarrier
\section{Conclusions}
\label{sec:conclusions}
A search is presented for new high-mass states decaying to a lepton (electron or muon)
plus missing transverse momentum using 20.3~\ifb\ of proton--proton collision data at
$\sqrt{s}=8\tev$ recorded with the ATLAS experiment at the Large Hadron Collider.
No significant excess beyond SM expectations is observed. Limits on \xbr\ are presented.
A \wp\ with SSM couplings is excluded for masses below 3.24~\tev\ at 95\% CL.
The exclusion for \wstar\ with equivalent couplings is 3.21~\tev.
For the pair production of weakly interacting DM particles in events with a leptonically decaying $W$, limits are set on
the mass scale, \mstar, of the unknown mediating interaction as well as on the DM--nucleon scattering cross-section.
\FloatBarrier
\section*{Acknowledgements}
\label{sec:acknowled}
We thank CERN for the very successful operation of the LHC, as well as the
support staff from our institutions without whom ATLAS could not be
operated efficiently.
We acknowledge the support of ANPCyT, Argentina; YerPhI, Armenia; ARC,
Australia; BMWF and FWF, Austria; ANAS, Azerbaijan; SSTC, Belarus; CNPq and FAPESP,
Brazil; NSERC, NRC and CFI, Canada; CERN; CONICYT, Chile; CAS, MOST and NSFC,
China; COLCIENCIAS, Colombia; MSMT CR, MPO CR and VSC CR, Czech Republic;
DNRF, DNSRC and Lundbeck Foundation, Denmark; EPLANET, ERC and NSRF, European Union;
IN2P3-CNRS, CEA-DSM/IRFU, France; GNSF, Georgia; BMBF, DFG, HGF, MPG and AvH
Foundation, Germany; GSRT and NSRF, Greece; ISF, MINERVA, GIF, I-CORE and Benoziyo Center,
Israel; INFN, Italy; MEXT and JSPS, Japan; CNRST, Morocco; FOM and NWO,
Netherlands; BRF and RCN, Norway; MNiSW and NCN, Poland; GRICES and FCT, Portugal; MNE/IFA, Romania;
MES of Russia and ROSATOM, Russian Federation; JINR; MSTD,
Serbia; MSSR, Slovakia; ARRS and MIZ\v{S}, Slovenia; DST/NRF, South Africa;
MINECO, Spain; SRC and Wallenberg Foundation, Sweden; SER, SNSF and Cantons of
Bern and Geneva, Switzerland; NSC, Taiwan; TAEK, Turkey; STFC, the Royal
Society and Leverhulme Trust, United Kingdom; DOE and NSF, United States of
America.
The crucial computing support from all WLCG partners is acknowledged
gratefully, in particular from CERN and the ATLAS Tier-1 facilities at
TRIUMF (Canada), NDGF (Denmark, Norway, Sweden), CC-IN2P3 (France),
KIT/GridKA (Germany), INFN-CNAF (Italy), NL-T1 (Netherlands), PIC (Spain),
ASGC (Taiwan), RAL (UK) and BNL (USA) and in the Tier-2 facilities
worldwide.
\bibliographystyle{JHEP}
|
\section{Introduction
\label{sec:intro}}
The spin--orbit angle between the stellar spin and the planetary orbital
axes, $\psi$, is supposed to be a unique observational probe of the
origin and evolution of planetary systems. The existence of
Jupiter-like planets with orbital periods less than a week strongly
indicates that the inward migration of those planets is a basic
ingredient of successful theories of planet formation and evolution. A
fairly popular scenario of the migration is based on the planet--disk
interaction,
in which planets are supposed to be on circular orbits whose orbital axes are parallel to the
stellar spin axis. On the other hand, scenarios such as planet--planet
scattering
or the Kozai mechanism
predict a broad range of eccentric and
oblique orbits. Thus the precise determination of $\psi$ and its
statistical distribution put a tight constraint on the viable migration
models \citep{Queloz2000, Winn2005}.
While the measurement of
$\psi$ is not easy, its projection onto the plane of the sky, $\lambda$,
has already been measured for more than 70 transiting planetary systems
via the Rossiter--McLaughlin (RM) effect \citep{2011exop.book...55W}, and is now
established as one of the most basic parameters that characterize
transiting planetary systems.
The RM effect was originally proposed to determine the projected
spin--orbit angle of eclipsing binary star systems \citep{R1924,M1924}.
\citet{Queloz2000} successfully applied the technique for the first
discovered transiting exoplanetary system, HD\,209458, and obtained
$\lambda = \pm \timeform{3D .9} {+18^\circ \atop -21^\circ}$.
In the quest for
improving the precision and accuracy, \citet{Ohta2005} presented an
analytic formula to describe the RM effect and studied in detail the
error budget and possible degeneracy among different parameters. This
allowed \citet{Winn2005} to revisit HD\,209458 with updated photometric
and spectroscopic data, and to obtain
$\lambda=-\timeform{4^D.4} \pm \timeform{1D.4}$,
improving the precision of the previous measurement by an order of
magnitude.
In doing so, \citet{Winn2005} pointed out that the analytic
approximation adopted by \citet{Ohta2005} leads to typically 10 percent
error in the predicted velocity anomaly amplitude, while the estimated
$\lambda$ is fairly reliable. This motivated \citet{Hirano2010} and
\citet{Hirano2011} to take into account stellar rotation,
macroturbulence, and thermal/pressure/instrumental broadenings in
modeling the stellar absorption line profiles. Those authors derived an
analytic formula for the velocity anomaly of the RM effect by maximizing
the cross-correlation function between the in-transit spectrum and the
stellar template spectrum. As a result, their analytic formulae reproduce
mock simulations within $\sim 0.5$ percent, enabling
the accurate and efficient multi-dimensional fit of parameters
characterizing the star and planet(s) of an individual system.
More importantly, \citet{Winn2005} clearly demonstrated the potential of
the RM effect to put strong quantitative constraints on the existing
and/or future planetary formation scenarios. Indeed, when HD\,209458
was the only known transiting planetary system, \citet{Ohta2005}
discussed that {\it ``Although unlikely, we may even speculate that a
future RM observation may discover an extrasolar planetary system in
which the stellar spin and the planetary orbital axes are anti-parallel
or orthogonal. Then it would have a great impact on the planetary
formation scenario, \ldots ''}. In reality, however, they were too
conservative. Among the 70 transiting planetary systems observed with
the RM effect, more than 30 systems exhibit significant misalignment
with $|\lambda|>\timeform{22D.5}$ [\,see, e.g., figure 7 of
\citet{Xue2014}\,]. This unexpected diversity of the spin--orbit angle
is not yet properly understood by the existing theories, and remains an
interesting challenge \citep[see,
][]{Fabrycky2007,Nagasawa2008,Winn2010b,Nagasawa2011,Hirano2012,Hirano2012b,
Lai2012,Albrecht2013,Masuda2013,Xue2014}.
It should be noted, however, that $\lambda$ differs from the true
spin--orbit angle $\psi$ due to the projection on the sky. In addition
to $\lambda$, $\psi$ also depends on the orbital inclination $i_{\rm
orb}$ and the obliquity of the stellar spin-axis $i_\star$:
\begin{equation}
\cos\psi = \cos i_\star \cos i_{\rm orb}
+ \sin i_\star \sin i_{\rm orb} \cos\lambda ,
\label{psi_3d}
\end{equation}
as illustrated in Figure \ref{angles}.
\begin{figure}
\centering
\includegraphics[width=8.5cm,clip]{Figures/angles_rev7b.eps}
\caption{Schematic illustration of geometric configuration of
star--planet systems. We choose a coordinate system centered on
the star, where the $XY$-plane is in the plane of the sky and
$+Z$-axis points towards the observer. The $+Y$-axis is chosen
along the sky-projected stellar spin and the $X$-axis is
perpendicular to both $Y$- and $Z$-axes, forming a right-handed
triad. Red and green arrows indicate on a unit
sphere, the angular momentum vectors of the stellar spin and
the planetary orbital motion, respectively. The stellar and
orbital inclinations, $i_\star$ and $i_{\rm orb}$, are measured
from the $+Z$-axis and in the range of $[\,0^\circ, 180^\circ]$.
The planetary orbital axis projected onto the sky plane is
specified by the projected spin--orbit angle, $\lambda$, which
is measured from the $+Y$-axis and in the range of $[0^\circ,
360^\circ]$. Note that $\lambda$ is measured in the
direction specified by the arrow. The angle AOC
between the stellar spin and the planetary orbit axis vectors,
$\psi$, is derived from the law of cosines for the spherical
triangles ABC, as given by Equation (\ref{psi_3d}).}
\label{angles}
\end{figure}
The main purpose of this paper is to establish a methodology to
determine $\psi$, instead of $\lambda$, through the
joint analysis of asteroseismology, transit lightcurve, and the RM
effect, and to present specific results for a couple of interesting
transiting planetary systems, HAT-P-7\footnote{ We would like to
emphasize the efforts made by Lund M. N. and his collaborators for their
work on HAT-P-7. This system turned out to be studied simultaneously and
independently by our respective teams.} (KIC\,10666592) and Kepler-25
(KIC\,4349452).
In the case of transiting planetary systems, $i_{\rm orb}$ can be
estimated from the transit lightcurve, and in any case is close to
$90^\circ$. Given the projected angle $\lambda$ measured from the RM
effect, the major uncertainty for $\psi$ comes from the unknown
stellar inclination $i_\star$.
There are two major and complementary approaches to estimate
$i_\star$, and hence $\psi$. One is to determine the line-of-sight
rotational velocity of the star, $v \sin i_\star$, either from the width
of absorption lines or from the RM effect. The observed $v\sin i_\star$
is then compared with an independent estimate of the equatorial velocity
of the star, $v$, to yield $i_\star$. \citet{2010ApJ...719..602S} used
an empirical relation for Sun-like stars to evaluate $v$ from their
masses and ages. Alternatively, one can determine the stellar spin
period photometrically from periodic variations in the lightcurve due to
the stellar activity, and then estimate $v$ assuming the stellar radius
\citep[e.g.,][]{Hirano2012}.
The other is asteroseismology \citep{Unno1989, Aerts2010} for which the
key principles are described in detail later, but briefly summarised
below. Thanks to space-borne instruments such as MOST
\citep{2003PASP..115.1023W}, CoRoT \citep{Baglin2006a,Baglin2006b} and
\emph{Kepler} \citep{Borucki2010}, asteroseismology now opens a
good opportunity to unveil the internal structure of many stars
with high precision. This is made possible through the
detection of oscillation modes propagating throughout the stars with
unprecedented precision, because of extraordinary low noise level
and uninterrupted extremely long-term data monitoring with short
sampling cadences, both of which are never available from the
ground-based observations \cite[e.g.][]{Appourchaux2008, Metcalfe2012,
Gizon2013}. More details about the recent development in
asteroseismology may be found in recent conference proceedings
such as \citet{2012ASPC..462.....S}, \citet{2013ASPC..479.....S}, and
\citet{2014IAUS..301.....G}. When coupled with non-seismic observables,
the asteroseismic observational information promises accurate
inference of fundamental properties of host stars
\citep[e.g.][]{Bazot2005, Carter2012}.
The stellar rotation affects the frequency spectrum of stellar
oscillation modes. It induces a multiplet fine structure, whose
frequency separation is dependent on the internal rotation profile of
the star as well as the stellar structure, for each mode. More
importantly, the apparent profiles of the rotationally induced frequency
multiplets are very sensitive to the inclination angle of the
stellar rotation axis with respect to the line-of-sight, $i_\star$.
In turn, one can infer $i_\star$ quite well from
asteroseismology.
One might wonder how commonly stellar oscillations that enable
asteroseismology can be detected among the host stars of exoplanet
systems. Remember that transiting planet hunting
preferentially select stars with small radii (i.e. low-mass stars
in the main sequence) in order to increase the relative transit
depth in the photometric lightcurve. Moreover, such low-mass
stars are suitable for the radial velocity follow-up not only because they are
more affected by orbital motion of planets but also because they have
sharp and narrow absorption lines due to their slow spin rotation
velocity. Such low-mass, cool stars have a thick convective envelope
(as in the case of the Sun) that sustains pulsations. Turbulent motion
with speeds close to that of sound near the stellar surface
stochastically generates acoustic waves, which propagate inside the star
until they are damped. The oscillations with frequencies close to
those of eigenmodes of the star are eventually sustained as many
acoustic modes. Therefore cool (\emph{i.e.} $\lessapprox 7000$ K)
host stars for exoplanets should commonly exhibit solar-like
oscillations, and thus consitute good targets for asteroseismology.
In this paper, we focus on two specific exoplanetary systems,
HAT-P-7 and Kepler-25; HAT-P-7 is the first example of a system
hosting a retrograde or a polar-orbit planet, while Kepler-25 is a
multi-transiting system with three planets, making them two interesting
examples. We show that joint analyses of asteroseismology,
transit lightcurve, and the RM effect provide stringent orbital parameter
estimates.
This paper is organised as follows. Section \ref{sec:previous}
summarizes the previous RM measurements and radial velocity (RV) data of
the two systems. Section \ref{sec:asteroseismology} presents a
detailed description of the basic principle of asteroseismology,
followed by our main results of asteroseismology for the two stars in
Sections \ref{sec:asteroseismology:HATP7} and
\ref{sec:asteroseismology:K25}. Sections \ref{sec:hat-p-7} and
\ref{sec:kepler-25} analyze the {\it Kepler} transit lightcurves and the
RV anomaly of the RM effect, using the asteroseismology results as a
prior, and show how the joint analysis improves the estimate{s} of
the system parameters. Section \ref{sec:summary} is devoted to further
discussion, and Section \ref{sec:9} summarises the present paper.
\section{Previous Spin--Orbit Measurements
\label{sec:previous}}
\subsection{HAT-P-7
\label{subsec:previous-hatp7}}
The HAT-P-7 system comprises a bright (\emph{V}=10.5) F6 star and a hot
Jupiter transiting the host star with a 2.2-d period \citep[hereafter
P08]{Pal2008}. In addition to the significant spin--orbit misalignment
first revealed by the Subaru spectroscopy \citep{Narita2009, Winn2009},
the fact that the system is in the {\it Kepler} field makes it very
attractive as an asteroseismology target.
Interestingly, there have been three independent measurements of the RM
effect for the HAT-P-7 system, which all indicate the significant {spin--orbit}
misalignment, but do not agree quantitatively. \citet{Winn2009}
(hereafter W09) performed the joint analysis of the spectroscopic and
photometric transit of HAT-P-7b to obtain $\lambda = \timeform{182D.5}
\pm \timeform{9D.4}$.
For RVs, they analyzed 17 spectra observed with the High Resolution
Spectrograph (HIRES) on the Keck I telescope as well as 69 spectra
observed with the High Dispersion Spectrograph (HDS) on the Subaru
telescope. Eight of the HIRES spectra were from P08 and taken in 2007,
while the other nine were obtained in 2009. Among 69 HDS spectra, 40
were obtained on 2009 July 1 that spanned a transit.
On the other hand, \citet{Narita2009} (hereafter N09) determined
$\lambda = \timeform{227D.4} {+\timeform{10D.5} \atop -\timeform{16D.3}}$
(equivalently $\lambda=- \timeform{132D.6} {+\timeform{10D.5}
\atop -\timeform{16D.3}}$) based on the eight HIRES RVs from P08
and 40 HDS spectra spanning the transit on 2008 May 30. Although they
fixed the transit parameters in the analysis of the RM effect, the
systematics from the uncertainties of these parameters do not seem to
explain the mild discrepancy with the W09 result, according to their
discussion (see cases 1 to 4 in section 4 of N09).
Later on, \citet{2012ApJ...757...18A} (hereafter A12) reported another
measurement of the RM effect, resulting in $\lambda = \timeform{155D}
\pm 14^\circ$. They analyzed 49 HIRES spectra spanning a transit on the
night 2010 July 23/24 with the priors on transit parameters and
ephemeris from the {\it Kepler} lightcurves.
In this paper, we use the same RV data published in each of the three
papers. Since the origin of the possible discrepancy in $\lambda$ is
not clear, we analyze each data set separately instead of
combining the three.
\subsection{Kepler-25
\label{subsec:previous-kepler25}}
The Kepler-25 system is one of the few multi-transiting planetary
systems with constrained $\lambda$. It consists of a relatively
bright ($K_{\rm p}=10.7$) host star, two short-period Neptune-sized
planets confirmed with Transit Timing Variations (TTVs)
\citep{2012MNRAS.421.2342S}, and one outer non-transiting planet
detected in long-term RV trend \citep{2014ApJS..210...20M}.
\citet{Albrecht2013} (hereafter A13) measured $\lambda = 7^\circ \pm
8^\circ$ for the larger transiting planet Kepler-25c based on the HIRES
spectra observed for two nights (2011 July 18/19 and 2012 May
31/June 1). Since the signal-to-noise ratio of the RV anomaly was small
due to the relatively small radius of Kepler-25c, they also analyzed the
time-dependent distortion of the spectral lines directly [known as the
``Doppler shadow'' method; see \citet{2010MNRAS.403..151C}] and obtained
a consistent result, $\lambda= -\timeform{0D.5} \pm \timeform{5D.7}$.
In this paper, we analyze the RVs around the above two transits from A13
alone because our focus is the determination of $\psi$.
\section{Asteroseismology \label{sec:asteroseismology}}
\subsection{Setting Up the Problem }
Due to its sensitivity to the stellar internal structure,
asteroseismology can achieve high-precision determinations of
stellar fundamental parameters \citep{Lebreton2009} (e.g.,
uncertainties of a few percent level for their mass and
radius). The stellar modelling using seismic observables mostly relies
on the stellar pulsation frequencies, usually extracted from the
analysis of the power spectrum of the stellar lightcurve.
For a spherically symmetric star, each eigenmode is characterised
by three quantum numbers; the angular degree $l$, the azimuthal
order $m$ ($-l \le m \le +l$), and the radial order $n$. The degree $l$
corresponds to the number of nodal surface lines, while the
azimuthal order $m$ specifies the surface pattern of the eigenfunction,
with $|m|$ being the number of longitude lines among the $l$ nodal
surface lines. The radial order $n$ corresponds to the number of nodal
surfaces along the radius. For a non-rotating star, both the radial
eigenfunction and the frequency of each mode are independent of $m$ and
show the $(2l+1)$-fold degeneracy. The eigenfrequency $\nu$
depends on $l$ and $n$ alone. Frequencies of high order, acoustic
(or p-) modes of the same low degree $(n \gg l \sim 1)$ are almost
equally spaced and separated on average by a frequency spacing
$\Delta\nu$:
\begin{equation}
\nu(n,l) = \Delta\nu \left(n+{{l}\over{2}}+\alpha \right)
+ \varepsilon_{n,l},
\label{eq:2}
\end{equation}
where $\alpha$ is a constant of order unity, and $\varepsilon_{n,l}$ is
a small correction. The spacing is related to the sound
velocity inside the star by
\begin{equation}
\Delta\nu = \left(2\int_0^{R_\star} {{1}\over{c(r)}}\,dr\right)^{-1}
\label{eq:3}
\end{equation}
and is sensitive to the mean stellar density $\rho_\star$.
Therefore, knowing the solar density $\rho_\odot=(1.4060 \pm
0.0005)\times 10^3$\,kg\,m$^{-3}$ and its frequency spacing
$\Delta\nu_\odot = 135.20 \pm 0.25$ $\mu$Hz\footnote{From frequencies of
\cite{Garcia2011b}.}, one can estimate the mean stellar density
from the scaling:
\begin{equation}
\rho_{\star,\mathrm{s}} = \rho_\odot (\Delta\nu/\Delta\nu_\odot)^2.
\label{eq:3b}
\end{equation}
The stellar rotation lifts the degeneracy among non-radial modes
($l\neq 0$), revealing a fine structure of modes identified by their
azimuthal order $m$. In the case of solar-like oscillations, acoustic
modes are excited stochastically by turbulent convection. This
mechanism is expected to generate almost the same amplitudes in
the rotationally split modes with the same $l$ and $n$. If this is
the case, in disk-integrated photometry as achieved by \emph{Kepler},
the height of the azimuthal modes in the power spectrum is
sensitive to the stellar inclination angle $i_\star$ due to a
geometrical projection effect [see \cite{Gizon2003} for more details]
and has been widely used to evaluate $i_\star$
\cite[e.g.][]{Benomar2009b, Appourchaux2012}. In turn, it enables us
to measure $\psi$ of exoplanets \citep{Chaplin2013,
2014ApJ...782...14V}, and indeed revealed a significant
spin--orbit misalignment for a red-giant host star system, Kepler-56
\citep{huber2013}. This dependence of visibility in the power spectrum
is expressed in terms of
\begin{equation}
\label{eq:legendre}
\mathcal{E}(l,m, i_\star)=\frac{(l-|m|)!}{(l+|m|)!} \left[P^{|m|}_l (\cos i_\star)\right]^2 ,
\label{eq:4}
\end{equation}
where $P^{|m|}_l$ is the associated Legendre function and the integral
of $\mathcal{E}(l,m, i_\star)$ over $\cos i_\star$ is normalised by
$(2l+1)^{-1}$.
Solar-like oscillators such as Kepler-25 and HAT-P-7 are typically slow
rotators for which the centrifugal force can be neglected. In addition,
there is no evidence of a strong magnetic field and we can safely neglect
it \citep{Reese2006, ballot2010}. If the internal
rotation of the star is independent of the latitude and the longitude,
the split frequencies are simply written as
\begin{equation}
\nu(n,l,m) = \nu(n,l) + m \,\delta\nu_{\rm s}(n,l),
\label{eq:5}
\end{equation}
where $\delta\nu_{\rm s}(n,l)$ is the rotational splitting [e.g. \cite{Appourchaux2008, Benomar2009, Chaplin2013}].
\begin{figure*}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/Ideal_Amplitude_Map_l1.eps, angle=90, width=8cm, height=6cm}}
\subfigure{\epsfig{figure=Figures/Ideal_Amplitude_Map_l2.eps, angle=90, width=8cm, height=6cm}}
\end{center}
\caption{The figures illustrate the dependence between stellar inclination and rotation. The relative power of azimuthal components $m$ for $l=1$ and $l=2$ in the ideal case
($\delta\nu_s \gg \Gamma$) is indicated by the color scale and is calculated using Equation \ref{eq:power}.
Horizontal dash lines indicate frequencies of the multiplets $m$. Each inclination is characterised by a unique $l=1$ and $l=2$ mode structure. This property is used to infer the stellar inclination. Zero degree corresponds to a star seen from a pole, 90 degree to a star seen from the equator.}
\label{fig:is-nus-ideal}
\end{figure*}
\begin{table*
\caption{Non-seismic observables of HAT-P-7 and Kepler-25. All but $v\sin i_\star$ are used for stellar modelling.}
\begin{center}
\begin{tabular}{cccrcc|l}
\toprule
Star & $L/L_\odot$ & $\log g$ (cgs) & [Fe/H] & $T_{\rm eff}$ (K) & $v\sin i_\star$ (km\,s$^{-1}$) & Source \\
\midrule
HAT-P-7 & $4.9 \pm 1.1$ & $4.070 \pm 0.06$ & $0.26 \pm 0.08$ & $6350 \pm 80$ & $3.8 \pm 0.5$ & \cite{Pal2008} \\
Kepler-25 & N/A & $4.278 \pm 0.03$ & $-0.04 \pm 0.10$ & $6270 \pm 79$ & $9.5 \pm 0.5$ & \cite{2014ApJS..210...20M}\\
\bottomrule
\end{tabular}
\end{center}
\label{tab:nonseism_obs}
\end{table*}
\renewcommand{\arraystretch}{0.7}
\begin{table}
\caption{Detected pulsation frequencies of HAT-P-7 and Kepler-25. The radial order $n$ is determined by the best models matching observables.}
\begin{center}
\begin{tabular}{c|crc|crc}
\toprule
& \multicolumn{3}{c}{HAT-P-7} & \multicolumn{3}{c}{Kepler-25} \\
$l$ & $n$ & $\nu_{n, l, m=0}$ & $\sigma$ & $n$ & $\nu_{n, l, m=0}$ & $\sigma$ \\
& & ($\mu$Hz) & ($\mu$Hz) & & ($\mu$Hz) & ($\mu$Hz) \\
\midrule
0 & 11 & 715.50 & 0.30 & 16 & 1691.79 & 0.52 \\
0 & 12 & 771.57 & 0.51 & 17 & 1788.58 & 0.23 \\
0 & 13 & 828.34 & 0.30 & 18 & 1884.39 & 0.36 \\
0 & 14 & 885.91 & 0.26 & 19 & 1981.33 & 0.18 \\
0 & 15 & 944.88 & 0.25 & 20 & 2080.08 & 0.32 \\
0 & 16 & 1004.77 & 0.22 & 21 & 2178.68 & 0.43 \\
0 & 17 & 1064.83 & 0.20 & 22 & 2277.00 & 0.32 \\
0 & 18 & 1123.17 & 0.23 & 23 & 2375.48 & 0.67 \\
0 & 19 & 1181.90 & 0.23 & 24 & 2472.91 & 0.59 \\
0 & 20 & 1240.53 & 0.27 & 25 & 2570.03 & 1.43 \\
0 & 21 & 1300.53 & 0.35 & & \\
0 & 22 & 1360.78 & 0.43 & & \\
0 & 23 & 1421.55 & 0.94 & & \\
0 & 24 & 1482.03 & 0.75 & & \\
0 & 25 & 1542.96 & 1.21 & & \\
\midrule
1 & 11 & 740.79 & 0.22 & 16 & 1736.27 & 0.79 \\
1 & 12 & 796.71 & 0.35 & 17 & 1832.49 & 0.20 \\
1 & 13 & 854.00 & 0.23 & 18 & 1929.17 & 0.28 \\
1 & 14 & 911.89 & 0.20 & 19 & 2026.97 & 0.28 \\
1 & 15 & 971.85 & 0.16 & 20 & 2125.46 & 0.32 \\
1 & 16 & 1031.54 & 0.15 & 21 & 2224.32 & 0.51 \\
1 & 17 & 1091.15 & 0.15 & 22 & 2323.04 & 0.32 \\
1 & 18 & 1149.92 & 0.17 & 23 & 2421.68 & 0.53 \\
1 & 19 & 1208.36 & 0.17 & 24 & 2521.29 & 0.63 \\
1 & 20 & 1267.82 & 0.23 & 25 & 2621.12 & 1.12 \\
1 & 21 & 1327.41 & 0.27 & & \\
1 & 22 & 1388.49 & 0.36 & & \\
1 & 23 & 1448.96 & 0.46 & & \\
1 & 24 & 1509.40 & 0.54 & & \\
1 & 25 & 1569.30 & 0.92 & & \\
\midrule
2 & 10 & 710.81 & 0.63 & 15 & 1683.26 & 3.88 \\
2 & 11 & 767.31 & 0.62 & 16 & 1779.57 & 2.17 \\
2 & 12 & 824.46 & 0.53 & 17 & 1875.12 & 1.39 \\
2 & 13 & 882.27 & 0.54 & 18 & 1972.55 & 0.67 \\
2 & 14 & 940.46 & 0.34 & 19 & 2071.55 & 0.77 \\
2 & 15 & 1000.17 & 0.49 & 20 & 2170.64 & 0.88 \\
2 & 16 & 1059.82 & 0.35 & 21 & 2269.90 & 1.14 \\
2 & 17 & 1118.74 & 0.31 & 22 & 2368.62 & 1.16 \\
2 & 18 & 1177.89 & 0.39 & 23 & 2467.03 & 1.38 \\
2 & 19 & 1236.33 & 0.42 & 24 & 2565.48 & 2.79 \\
2 & 20 & 1296.40 & 0.50 & & \\
2 & 21 & 1356.39 & 0.50 & & \\
2 & 22 & 1417.09 & 0.97 & & \\
2 & 23 & 1478.41 & 0.97 & & \\
2 & 24 & 1539.79 & 1.50 & & \\
\bottomrule
\end{tabular}
\end{center}
\label{tab:seism:freq}
\end{table}
\renewcommand{\arraystretch}{1.0}
It should be noted that because of the modes stochastic nature, each solar-like mode has a Lorentzian profile
in the power spectrum \citep{harvey1985}. Thus, the stellar oscillations can be expressed as a sum of Lorentzian over $n$, $l$ and $m$,
\begin{equation}
\label{eq:power}
P(\nu)= \sum_{n,l}\sum_{m=-l}^l
\frac{\mathcal{E}(l,m, i_\star)H(n,l)}
{1+4(\nu-\nu(n,l,m))^2/\Gamma^2(n,l,m)} .
\end{equation}
Each mode is therefore not only characterised by its frequency $\nu (l,n,m)$, but also by a height $H(n,l,m)=\mathcal{E}(l,m, i_\star)
H(n,l)$ and a full width at half maximum $\Gamma(n,l,m)$ (hereafter
called width). Here, $H(n,l)$ is the intrinsic height for the mode of
$n$ and $l$. The heights and the widths of the modes retain information
on, for example, the modes excitation mechanism and on non-adiabatic
processes.
In order to show the sensitivity of the asteroseismology analysis
to $i_\star$, we plot $P(\nu)$ in Figure \ref{fig:is-nus-ideal} as a
function of $i_\star$ and $\nu-\nu(n,l)$. The left and right panels
correspond to $l=1$ and $l=2$ modes, respectively. The plot is
color-coded according to the amplitude of $P(\nu)$ and for a given $n$ ({\it i.e,} a sum of
the $(2l+1)$ Lorentzian profiles). Figure \ref{fig:is-nus-ideal} is
presented simply for illustrative purpose, and is computed from
equations (\ref{eq:legendre}) and (\ref{eq:power}), assuming
$\delta\nu_{\rm s} = 1\,\mu$Hz and $\Gamma=0.25\mu$Hz. The condition
$\delta\nu_{\rm s} \gg \Gamma$ breaks the degeneracy among the
rotationally split $m$ components. As demonstrated in this figure, in
the case of $i_\star \simeq 0^\circ$, that is, when we see the star from
the pole, only the $m=0$ component is visible as a singlet for both of
$l=1$ and $l=2$ modes. On the other hand, in the case of $i_\star \simeq
90^\circ$, the rotational splitting appears as a doublet in the case of
$l=1$ and as a triplet in the case of $l=2$. Thus for a given value of
$i_\star$, the power is the result of a unique configuration of height for the
$m$ components, which enables us to infer the value of $i_\star$ from the
$l=1$ and $l=2$ mode profiles. Note, however, that because Equation
(\ref{eq:legendre}) depends on $|m|$
, solutions in the four quadrants of the trigonometric circle are degenerate and one cannot distinguish
between $i_\star$ and $(180^{\circ}-i_\star)$.
\subsection{Data processing and modeling
\label{sec:asteroseismology:method}}
The {\it Kepler} Space Telescope collected time series lightcurves of
about 160,000 stars over the 115 square degrees field-of-view from
its 372.5-d, heliocentric Earth-trailing orbit over its four-year
lifetime for 2009 -- 2013. Its major purpose was to find
extra-solar planets by detecting a small amount drop of the visual
brightness of their parent stars, caused by the transits of the planets
in front of the stars. So the photometric asteroseismology and planet
studies are synergistic. Four times per orbit the satellite was
scheduled to perform a roll to keep its solar panels facing the Sun, so
the data were divided into `Quarters' (1/4 of its 372.5-d heliocentric
orbit), denoted as Q$n$.
For HAT-P-7, we use Q0 to Q16 (1437 days in total) of
\emph{Kepler} data taken every 1-min (`Short Cadence' data; SC), while
for Kepler-25, we used Q5 to Q16 (1114 days) SC data. After
removing the transits from the lightcurve with a median
high-pass filter of an adequate frequency width, we compute the
power spectrum of each star following the method described in
\cite{Garcia2011}. The high-pass filter is efficient to remove the
signal of the transit in the power spectrum without altering the stellar
pulsation characteristics, since the orbital periods of the detected
planets around HAT-P-7 and Kepler-25 are of the order of days, while
stellar pulsation periods are in the minute range. To extract the
mode parameters, we perform a Lorentzian profile fit to each mode
that exhibits significant power. We use a Markov Chain Monte Carlo
(MCMC) method and a similar method to \cite{Benomar2009} but with a
smoothness condition on the frequencies. \citep[see][]{Benomar2013a, Benomar2014}.
The prior on the rotational splitting $\delta\nu_{\rm s}$ is uniform
between 0 and 8\,$\mu$Hz. The prior on $i_\star$ is chosen to be
uniform in $\cos i_\star$ for $0<\cos i_\star<1$, and is equivalent to
the random uniform distribution of $i_\star$. Because of the symmetries
in Equation (\ref{eq:legendre}), we only consider solutions of $0^\circ
\leq i_\star \leq 90^\circ$ in what follows.
Figures \ref{fig:spectrum:Hatp7} and \ref{fig:spectrum:K25} show the
resulting power spectra and their best-fit models for HAT-P-7 and
Kepler-25, respectively. The identified pulsation modes and the derived
pulsation frequencies of the central component of multiplets are listed
in Table\,\ref{tab:seism:freq}.
\begin{figure*
\begin{center}
\includegraphics[width=0.55\linewidth,angle=90]{Figures/HATP7-Spec_wlegends.eps}
\end{center}
\caption{\textbf{HAT-P-7.} Power spectrum over three radial orders for
modes with highest signal-to-noise ratio. The spectrum is shown after a
boxcar smooth over $0.08$ $\mu$Hz (grey) and $0.24$ $\mu$Hz (black). The
best fit is the solid red line. The inset shows all the extracted
modes.} \label{fig:spectrum:Hatp7}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.55\linewidth,angle=90]{Figures/K25-Spec_wlegends.eps}
\end{center}
\caption{\textbf{Kepler-25}. Power spectrum over three radial orders for
modes with highest signal-to-noise ratio. The spectrum is shown after a
boxcar smooth over $0.21$ $\mu$Hz (grey) and $0.83$ $\mu$Hz (black). The
best fit is the solid red line. The inset shows all the extracted
modes.} \label{fig:spectrum:K25}
\end{figure*}
Stellar models that simultaneously match non-seismic observables
(cf. Table \ref{tab:nonseism_obs}) and seismic observables (frequencies
in Table \ref{tab:seism:freq}) are found using the `astero' module of
the Modules for Experiments in Stellar Astrophysics (MESA) evolutionary
code \citep{paxton2011, paxton2013}. Stellar models are calculated
assuming a fixed mixing length parameter $\alpha_{\rm MLT}=2.0$
and an initial hydrogen abundance $X=0.7$. The opacities are
calculated using the MESA standard equation of state from the
opacity table in \cite{Asplund2009}. These are applicable for
stars with effective temperature $10^{3}\,\rm K< T_{\rm eff} <
10^{4.1}\,\rm K$. Nuclear reactions are set to include standard hydrogen and
helium burning; the pp-chain and the CNO cycle in addition to the triple
alpha reaction.
It is known that semiconvective zones are present in stars of $\sim
1.5\,M_\odot$. Since a small convective core in such stars expands due
to the growing importance of the CNO cycle, the opacity is larger at the
outer side of the convective boundary than at its inner side. We adopt
the M. Schwarzschild treatment to define the boundary between the
convective and radiative zones in such a case. With expected mass larger
than 1.2\,$M_\odot$, HAT-P-7 and Kepler-25 may have a convective
core. Then, the nature of the transition (e.g. sharp or smooth)
between convective and radiative regions may have a significant impact
on the seismic frequencies \citep[e.g.][]{Monteiro1994}. Thus, to
describe a possible extension of the convective zone inside the
radiative zone, we have included an overshoot. Diffusion was not
implemented. Mass $M_\star$, metallicity [Fe/H], helium abundance $Y$,
the coefficient for overshooting $\alpha_{\rm ov}$, and age are
treated as free parameters.
Eigenfrequencies are calculated assuming adiabaticity and using {\tt
ADIPLS} \citep{JCD2008b}. We apply surface effect corrections to the
frequencies, following the method of \cite{Kjeldsen2008}. The search
for the best model involves a simplex minimisation approach
\citep{Simplex} using the $\chi^2$ criteria. Uncertainties are then
estimated by evaluating the $\chi^2$ for solutions surrounding the best
model and by weighing the model parameters with Likelihood $\propto
\exp(-\chi^2/2)$.
\subsection{Mode Degree Identification
\label{sec:asteroseismology:modeI}}
Prior to modelling a star, it is important to identify the degree $l$ from the power spectrum.
In solar-like cool stars (K, G type) the
identification is often obvious and relies on the \'echelle diagram
\citep{Grec1983}. An \'echelle diagram is built by dividing the power spectrum into frequency bins of interval $\Delta$, that are stacked in order to form an image in which the power is color-coded. In this image, the Y-axis represents the central frequency of each bin, while the X-axis corresponds to the frequency modulo $\Delta$. Note that the central frequency of the bins is a discrete quantity and one could use instead an integer for the Y-axis.
Figures \ref{fig:EDs-HATP7} and \ref{fig:EDs-K25} are the corresponding
\'echelle diagram for HAT-P-7 and Kepler-25 stacked with
$\Delta=59.9\,\mu$Hz and $\Delta=97.8\,\mu$Hz, respectively. If we
choose $\Delta=\Delta\nu$, the excess power due to the modes of the same
degree $l$ should show up along an almost straight vertical
line\footnote{For HAT-P-7, $\Delta$ is chosen slightly different than
$\Delta\nu$ for a better rendering of the \'echelle diagram.}. This is
because p-modes of the same degree are almost regularly spaced in
frequency, as implied by Equation (\ref{eq:2}).
Equation (\ref{eq:2}) shows that $\nu({n,l})=\nu({n-1, l+2})$ as long as
$\varepsilon_{n,l}$ is small. Thus the eigenmodes of {($n$, $l=0$)} and
{($n-1$, $l=2$)} have approximately the same frequencies. The same is true
for ($n$, $l=1$) and ($n-1$, $l=3$). On the other hand, the pulsation
amplitude of the surface, and {consequently} the integrated
luminosity variation, {are} smaller for larger $l$ modes. Thus, the
detected photometric amplitudes of the pulsation are usually dominated
by $l=0$ and $l=1$, and $l \ge 3$ are often buried in the noise.
This is why the careful visual inspection of the relative height and
frequency of the power spectra {enable us to} identify the corresponding modes.
This approach works for Kepler-25, but not for HAT-P-7 in reality. The
power spectrum of HAT-P-7 exhibits significant mixture of $l=0$ and
$l=2$ modes, and it is hard to disentangle them by visual inspection.
In such a case that the modes of the same $l$ are almost regularly
spaced in frequency, there exist two possibilities: either (S1) the fit
misidentifies the modes, or (S2) the fit correctly identifies the mode.
As for the former, all modes of degrees $l=0$ and $l=2$ would be
misidentified as $l=1$ modes (and vice-versa). This problem of mode
identification is recurrent in F stars and was first encountered in a
star observed by CoRoT, HD\,49933 \citep{Appourchaux2008}.
The most likely solution among the two competitive solutions (S1) and
(S2) described above may be judged by the Bayes factor between S1
and S2 [see \cite{Benomar2009, Benomar2009b, Appourchaux2012} for more
details]. Using our MCMC samples, we evaluated the Bayes factor at
$10^6:1$ in favour of modes with frequencies listed in Table
\ref{tab:seism:freq}. According to \cite{Jeffreys1961} , the Bayes
factor $>100$ is ``Decisive'', and thus one can safely assume that the
mode identification is correct. We also note that use of the empirical
approach detailed in \cite{White2012} reproduces the same degree
identification.
Furthermore, there is not clear evidence for $l=3$ in the \'echelle
diagram. To verify this quantitatively, we attempted to detect modes of
degree $l = 3$ by comparing the Bayes factor between a model $M_{l \leq
3}$, that includes those modes, with a model $M_{l \leq 2}$ that does
not. We obtained a factor $\simeq 2:1$ and $\simeq 2.5:1$ for HAT-P-7
and Kepler-25 respectively, in favour of $M_{l \leq 2}$, which is the
simplest model. Thus modes of degree $l =3$ are not conclusively
detected.
\begin{table}[!h]
\caption{Stellar model characteristics for HAT-P-7 and Kepler-25 derived
with MESA. $\rho_{\star,\rm m}$ is the density derived from
modelling. $\rho_{\star,\rm s}$ is the density derived by rescaling the
Sun density using the average frequency separation $\Delta \nu$. }
\begin{center}
\begin{tabular}{c|c|c}
\toprule
parameter & HAT-P-7 & Kepler-25 \\
\midrule
$M_\star$ $(M_\odot)$ & $1.59 \pm 0.03$ & $1.26 \pm 0.03$ \\
$R_\star$ $(R_\odot)$ & $2.02 \pm 0.01$ & $1.34 \pm 0.01$ \\
$[{\rm Fe/H}]$ & $0.32 \pm 0.04$ & $0.11 \pm 0.03$ \\
$T_{\rm eff}$ (K) & $6310 \pm 15$ & $6354 \pm 27$ \\
Age (Myrs) & $1770 \pm 100$ & $2750 \pm 300$ \\
$\alpha_{\rm ov}$ & $0.000 {{+0.002}\atop{-0.000}}$ & $0.007 \pm 0.003$ \\
$L/L_\odot$ & $5.84 \pm 0.05$ & $2.64 \pm 0.07$ \\
$\log g$ (cgs) & $4.029 \pm 0.002$ & $4.285 \pm 0.003$ \\
$\rho_{\star,\rm m}$ (10$^3$\,kg\,m$^{-3}$) & $0.2708 \pm 0.0035$ & $0.7367 \pm 0.0137$ \\
$\rho_{\star,\rm s}$ (10$^3$\,kg\,m$^{-3}$) & $0.2696 \pm 0.0011$ & $0.7356 \pm 0.0030$ \\
\midrule
reduced $\chi^2$ & $1.73$ & $1.03$ \\
\bottomrule
\end{tabular}
\end{center}
\label{tab:output_model}
\end{table}
\section{Asteroseismology of HAT-P-7
\label{sec:asteroseismology:HATP7}}
The power spectrum of HAT-P-7 (Figure\,\ref{fig:spectrum:Hatp7}) shows a
broad range of modes, spanning over 15 different radial orders
with a high signal-to-noise ratio (Table \ref{tab:seism:freq}),
enabling us to infer modes properties with an unprecedented precision
for an F-star. Our asteroseismic analysis detected a total of 45 modes
of degree $l=0$, $1$ and $2$, for which the frequencies are listed in
Table \ref{tab:seism:freq}.
\subsection{Fundamental Properties}
The \'echelle diagram of HAT-P-7 (Figure\,\ref{fig:EDs-HATP7}) shows
clear departures from a straight line, which is mostly the signature of
the transition between the outer convective zone and the radiative
zone. This is because discontinuities within the structure translate
into steep gradients in the acoustic structure of a star, which induce
frequency modulations of periods related to the acoustic depth of the
discontinuities \citep[e.g.][]{Vorontsov1988, Monteiro1994,
Roxburgh2003}. In modelling HAT-P-7, it is therefore important to
find models that match not only the average frequency separation
$\Delta\nu$ (which is sensitive only to the mean density) but also all
individual frequencies accurately.
\begin{figure}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/ED_HatP7.eps, angle=90, width=9.5cm, height=7.5cm}}
\end{center}
\caption{\textbf{Left panel.} Difference between observed frequencies $\nu_{\rm obs}$ of HAT-P-7
and best model frequencies $\nu_{\rm m}$. $l=0, 1, 2$ are shown as orange, red and
black diamonds respectively. \textbf{Right panel.} \'Echelle diagram
showing the observed power spectrum (background), the observed frequencies (diamonds) and the frequencies from the best model (white circles).}
\label{fig:EDs-HATP7}
\end{figure}
\begin{figure}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/ED_K25.eps, angle=90, width=9.5cm, height=7.5cm}}
\end{center}
\caption{\textbf{Left panel.} Difference between observed frequencies $\nu_{\rm obs}$ of Kepler-25
and best model frequencies $\nu_{\rm m}$. $l=0, 1, 2$ are shown as orange, red and
black diamonds respectively. \textbf{Right panel.} \'Echelle diagram
showing the observed power spectrum (background), the observed frequencies (diamonds) and the frequencies from the best model (white circles).}
\label{fig:EDs-K25}
\end{figure}
Following the method described in Section
\ref{sec:asteroseismology:method}, we found that the best model implies
$M_\star=1.59 \pm 0.03\,M_\odot$ (cf. Table \ref{tab:output_model} for
main characteristics of the model), which is slightly greater than what
was reported in earlier seismic studies; \cite{JCD2010} used Q0 and Q1
{\it Kepler} data ($\approx 60$ days long) and reported $M_\star = 1.520
\pm 0.036\,M_\odot$. They fitted individual frequencies corrected from
the surface effects \citep{Kjeldsen2008} and they used the {\tt ASTEC}
evolutionary code with a method and physics similar to what we
adopted in the present paper\footnote{Opacity tables and some nuclear
reaction rates are different.}. \cite{Oshagh2013} carried out an
analysis of HAT-P-7 using {\it Kepler} Q0 to Q2 (144 days long). Their
approach slightly differs from ours as they did a non-adiabatic
frequency calculation. They reported $M_\star=1.415 \pm
0.020\,M_\odot$. Furthermore \cite{VanEylen2012} used {\it Kepler} data
from Q0 to Q11 and reported $M_\star=1.361 \pm 0.021\,M_\odot$. While
our model values are consistent with those quoted in \cite{JCD2010}
within $2\sigma$, the other estimates are significantly different. Thus
we discuss the issue below.
First of all, \cite{JCD2010} and our study result in consistent
mean stellar densities\footnote{Using an MCMC analysis, they found
$M_\star=1.520 \pm 0.036\,M_\odot$ and $R_\star =1.991 \pm
0.018\,R_\odot$, corresponding to $\rho_{\star,\rm m} = (0.2707 \pm
0.0010)\times 10^3$\,kg\,m$^{-3}$, while our model implies
$\rho_{\star,\rm m}= (0.2708 \pm 0.0035)\times 10^3$\,kg\,m$^{-3}$. } at
$1\sigma$. In contrast, \cite{Oshagh2013} obtain $\rho_{\star,\rm
m}=(0.2778 \pm 0.0059)\times 10^3$\,kg\,m$^{-3}$ and \cite{VanEylen2012}
$\rho_{\star,\rm m}= (0.2781 \pm 0.0017)\times 10^3$\,kg\,m$^{-3}$,
which are consistent within $1\sigma$. While the differences
between \cite{Oshagh2013} and the present study may be due to the
non-adiabatic treatment of model frequencies and to the data quality as
well, this cannot explain the low mass found by \cite{VanEylen2012}.
Nevertheless, although the model in figure 2 of \cite{VanEylen2012}
has a small value of $\chi^2$, it does not seem to reproduce accurately
their individual frequencies. Moreover their method of measuring the
frequencies differs from ours (frequencies are measured by taking the
frequency at maximum height of a smooth spectrum) and they reported
larger uncertainties than what we obtain here.
In order to see if the difference in methodology could explain the
apparent discrepancies, we looked for the best model (minimum $\chi^2$)
assuming $M_\star=1.36\,M_{\odot}$, to be coherent with
\cite{VanEylen2012}. The best model has a $\chi^2=24.6$, approximately
14 times higher than the best model shown in Table
\ref{tab:output_model} and does not reproduce accurately the individual
oscillation frequencies. The mean stellar density $\rho_{\star,\rm
m}=(0.2562 \pm 0.0002)\times 10^3$\,kg\,m$^{-3}$ is also significantly
different. Thus we conclude that mass of $\approx 1.36\,M_{\odot}$
is less favored than $\approx 1.59\,M_{\odot}$, from our seismic
observables.
The best-fit model of the present study implies that the HAT-P-7
has a convective core that extends up to 6.9\% of the stellar radius,
while the outer convective zone represents approximately 13.1\% of the
stellar radius. The central hydrogen abundance $X_{\rm c}=0.214$, which
corresponds to 32\% of its initial core hydrogen, indicates that the
star is at a late stage in its main sequence. Finally we note that the
best model of HAT-P-7 has no need of surface effect correction.
\subsection{Rotation and Inclination}
Figure \ref{fig:is-pdfs-HATP7} shows the joint probability density
function (PDF) of $\delta\nu_{\rm s}$ and $i_\star$,
$p(i_\star,\delta\nu_{\rm s})$, for HAT-P-7 as well
as their marginalised posterior PDF, $p(\delta\nu_{\rm s})$ and
$p(i_\star)$. As clearly illustrated, $i_\star$ of HAT-P-7 is not
tightly constrained. The most probable value is
$i_\star=\timeform{27D.3}{+\timeform{34D.9}\atop{-\timeform{17D.5}}}$
with a 68\% confidence interval. This suggests that the star is more
likely seen by its pole than by its equator, albeit with large
uncertainty. To understand why $i_\star$ is not well determined, we
show in Figure \ref{fig:is-nus-hatp7} the power $P(\nu)$ corresponding
to the modes of degree $l=1$ and $l=2$ as we did in Figure
\ref{fig:is-nus-ideal}, but we set the rotational splitting equal to the
observed median splitting ($\delta\nu_{\rm s} = 0.70\,\mu$Hz). The
width $\Gamma$ of each Lorentzian is fixed to the average width ($\Gamma
= 3$ $\mu$Hz) of the modes of the highest signal-to-noise ratio. In this
case, $\delta\nu_{\rm s} \ll \Gamma$ and the $m$ components cannot be
resolved. Thus, the mode profiles are almost insensitive to the stellar
inclination, contrary to the ideal case of well resolved modes as
illustrated in Figure \ref{fig:is-nus-ideal}.
\begin{figure*}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/HATP7-nus-abs_inc_COSIs.eps, angle=90, width=16cm, height=12cm}}
\end{center}
\caption{\textbf{Upper right.} Joint posterior probability distribution
of the stellar inclination and the rotation for HAT-P-7. Blue represents
region of lowest probability. Red areas are of highest
probability. Superimposed and using a dark grey dotted line, we show the
spectroscopic $v\sin i_\star$ from P08 with its $1\sigma$ uncertainty
intervals (grey dotted lines). \textbf{Upper left.} Marginalized
probability density function for the rotational splitting. \textbf{Lower
right.} Marginalized probability density function for the stellar
inclination. \textbf{Lower left.} Seismic $v\sin i_\star$, inferred
using the probability density for the rotational splitting, the
inclination and the radius of HAT-P-7.} \label{fig:is-pdfs-HATP7}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/HAT-P7_COSIs_Amplitude_Map_l1.eps, angle=90, width=8cm, height=6cm}}
\subfigure{\epsfig{figure=Figures/HAT-P7_COSIs_Amplitude_Map_l2.eps, angle=90, width=8cm, height=6cm}}
\end{center}
\caption{\textbf{HAT-P-7.} Relative power $P(\nu)$ of
azimuthal components for $l=1$ (three Lorentzian) and $l=2$ (five Lorentzian) at the median value of the
rotational splitting. Width of the Lorentzian is the average mode width
($\Gamma=3 \,\mu$Hz). Horizontal dash lines indicate frequencies of the
multiplets $m$. Vertical dot-dash lines indicate the credible interval
(orange) and the median (black) of the measured inclination (cf. Figure
\ref{fig:is-pdfs-HATP7}). These figures show that because
$\delta\nu_{\rm s} \ll \Gamma$, the profile of modes of degree $l=1$ and
$l=2$ are almost insensitive to stellar inclination.}
\label{fig:is-nus-hatp7}
\end{figure*}
Although $\delta\nu_{\rm s}$ is related to the average internal rotation
frequency\footnote{Each mode is sensitive to the rotation at a given
depth. Assuming a modest differential rotation, for low-degree p-modes,
$\delta\nu_{\rm s}$ is nearly equal to the surface rotation
frequency.}, it provides a good proxy to the surface rotation
frequency. Based on this idea, with the radius $R_\star$ derived by
stellar modelling, we calculated the seismic $v\sin i_\star = 2 \pi
R_\star\, \delta\nu_{\rm s} \sin i_\star$ (cf. Figure
\ref{fig:is-pdfs-HATP7}). We obtained $v\sin i_\star =
2.95^{+1.71}_{-1.98}$\,km\,s$^{-1}$, which is in agreement with $v\sin
i_\star = 3.8 \pm 0.5$\,km\,s$^{-1}$ obtained by P08.
The degeneracy in solutions due to the correlation between rotation and
inclination limits the precision. In our effort to improve our
constraint on the inclination angle, $i_\star$, we looked for signs of
surface rotation by computing the autocorrelation of the
timeseries. Solar-like stars may have long-lived surface stellar spots
at low latitude that can modulate the light flux periodically, thus
revealing the surface rotation period. Unfortunately, HAT-P-7 shows no
sign of activity. While this may indicate that the star is not
active, this is consistent with our interpretation of the small
inclination angle.
\section{Asteroseismology of Kepler-25
\label{sec:asteroseismology:K25}}
Kepler-25 is an F star that shows oscillations for which we detected 30
modes of degree $l=0$, $1$ and $l=2$ spanning over 10 radial orders but
with amplitudes smaller than HAT-P-7 (Figure \ref{fig:spectrum:K25}).
\subsection{Fundamental Properties}
The precision on the extracted seismic frequencies is lower by
approximately a factor two, compared with the case of HAT-P-7. As seen
in the \'echelle diagram (Figure\,\ref{fig:EDs-K25}) the range of
observed frequencies does not allow us to entirely retrieve the
oscillation pattern of the modes, which certainly reduces the accuracy
of the modelling.
A seismic analysis of Kepler-25 has already been carried out by
\cite{Huber2013b} using the empirical scaling relations among
mass, radius, effective temperature, the frequency spacing
$\Delta\nu$ and frequency at maximum power of the modes, $\nu_{\rm max}$
[see for example \cite{Huber2011} for more details]. They derived
$M_\star=1.19 \pm 0.06\,M_{\odot}$ and $R_\star=1.309 \pm
0.023\,R_{\odot}$.
For this star, the model with the minimum $\chi^2$ is found with
surface effect and with an exponent of $b=4.9$. It describes a star
with $M_\star=1.26 \pm 0.03\,M_{\odot}$ and $R_\star =1.34 \pm
0.01\,R_\odot$. This is consistent with the first estimates
by \cite{Huber2011}. The central hydrogen abundance of $X_{\rm c} =
0.329$ corresponds to 46.9\% of the initial hydrogen abundance,
suggesting a star in the middle of its main sequence stage. The star
has a small convective core, extending up to $7\%$ of the stellar radius
and an outer convective zone representing $20\%$ of the stellar radius.
\subsection{Rotation and Inclination}
Figure \ref{fig:is-pdfs-K25} plots $p(i_\star,\delta\nu_{\rm s})$,
for Kepler-25 as well as their marginalised posterior PDF,
$p(\delta\nu_{\rm s})$ and $p(i_\star)$. We obtain
$i_\star=\timeform{66D.7}{+\timeform{12D.1}\atop{-\timeform{7D.4}}}$
within a $68\%$ confidence interval. The precision on $i_\star$
is much higher than for HAT-P-7, despite a lower signal-to-noise ratio.
This is because the rotational splitting is at least twice greater
($\delta\nu_s \simeq 1.72\,\mathrm{\mu Hz}$). The multiplets of each degree are
disentangled ($\delta\nu_s \approx \Gamma \simeq 2.5 \mu$Hz), and the
mode profile $\mathcal{E}(l,m, i_\star)$ becomes very sensitive to the
stellar inclination (cf. Figure \ref{fig:is-nus-K25}).
The radius derived from the best-fit model allows us to directly compare
the spectroscopically determined radial velocity, $v\sin i_\star$,
quoted by \citet{2014ApJS..210...20M} against our value. As shown in
Figure\,\ref{fig:is-pdfs-K25}, the spectroscopic $v\sin i_\star$ is
consistent with the maximum location of the joint
PDF. Moreover, the rotational kernels of the $l=1$ and $l=2$ modes
show that the measured rotational splitting is as much sensitive to the
rotation in the convective envelope as into the radiative zone. The
modes are however not sensitive to the rotation in the inner convective
region. This indicates that the radiative layer and the outer convective
region are rotating uniformly, with the same velocity as the surface.
Finally, note that $H(n,l,m)$ autocorrelation of the timeseries does not show
evidence for stellar activity.
\begin{figure*}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/K25_corel_nus_inc_with_spectro.eps, angle=90, width=16cm, height=12cm}}
\end{center}
\caption{\textbf{Upper right.} Joint posterior probability distribution
of the stellar inclination and the rotation for Kepler-25. Blue
represents region of lowest probability. Red areas are of highest
probability. Superimposed and using a dark grey dotted line, we show the
spectroscopic $v\sin i_\star$ quoted by \citet{2014ApJS..210...20M} with
its $1\sigma$ uncertainty intervals (grey dotted lines). \textbf{Upper
left.} Marginalised probability density function for the rotational
splitting. \textbf{Lower right.} Marginalised probability density
function for the stellar inclination. \textbf{Lower left.} Seismic
$v\sin i_\star$, inferred using the probability density for the
rotational splitting, the inclination and the radius of Kepler-25.}
\label{fig:is-pdfs-K25}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\subfigure{\epsfig{figure=Figures/K25_Amplitude_Map_l1.eps, angle=90, width=8cm, height=6cm}}
\subfigure{\epsfig{figure=Figures/K25_Amplitude_Map_l2.eps, angle=90, width=8cm, height=6cm}}
\end{center}
\caption{\textbf{Kepler-25.} Relative power {$P(\nu)$} of
azimuthal components for $l=1$ {(three Lorentzian)} and $l=2$ {(five Lorentzian)} at the median value of the
rotational splitting. Width of the Lorentzian is the average mode width
($\Gamma=2.5 \,\mu$Hz). Horizontal dash lines indicate frequencies of
the multiplets $m$. Vertical dotted lines indicate the credible interval
(orange) and the median (black) of the measured inclination (cf. Figure
\ref{fig:is-pdfs-K25}). The rotation is fast enough to distinguish the
Lorentzian profiles of each azimuthal order. This allows an accurate
determination of the stellar inclination.} \label{fig:is-nus-K25}
\end{figure*}
\section{Joint Analysis of the HAT-P-7 System
\label{sec:hat-p-7}}
{In this section and the next, we combine $i_\star$ from asteroseismology and $\lambda$ from the RM effect
to constrain the three-dimensional spin--orbit angle $\psi$.
Since the seismic $v\sin i_\star$ and $\rho_\star$ are also complementary to those from the RM effect and transit photometry,
we reanalyze the RM effect and the whole available {\it Kepler} lightcurves simultaneously,
incorporating the constraints on $i_\star$, $v\sin i_\star$, and $\rho_\star$ described in the previous sections as the prior knowledge.
The method and results are presented in this section for HAT-P-7 and in the next section for Kepler-25.}
For the HAT-P-7 system, the combination of the asteroseismology and {\it Kepler} lightcurves provides
a unique opportunity to tightly constrain the orbital eccentricity of HAT-P-7b,
especially because the occultation (secondary eclipse) is clearly detected for this giant and close-in planet.
Therefore, we first describe how the transit and {occultation} lightcurves constrain the planetary orbit
in Section \ref{ssec:hatp7-lc}, before reporting the joint analysis for $\psi$ in Section \ref{ssec:hatp7-joint}.
\subsection{Analysis of Transit and Occultation Lightcurves
\label{ssec:hatp7-lc}}
\subsubsection{Data Processing and Revised Ephemeris}
In the following analysis, we use the {\it Kepler} short-cadence
Pre-search Data Conditioned Simple Aperture Photometry (PDCSAP) fluxes
through Q0 to Q17 retrieved from the NASA exoplanet
archive.\footnote{\texttt{http://exoplanetarchive.ipac.caltech.edu}}
First, lightcurves are detrended and normalized by fitting a third-order
polynomial to the out-of-transit fluxes around $\pm 0.5$ days of every
transit center. Here, the central time and the duration of each transit
are determined from the central time of the first observed transit
calculated from the linear ephemeris, $t_0$, the orbital period, $P$,
and the duration taken from the archive. We iterate the polynomial fit
until all the {$>5\sigma$} outliers are excluded. In this process, we
remove the transits whose baselines cannot be determined reliably due to
the data gap around the ingress or egress.
Second, we fit each detrended and normalized transit with the lightcurve
model by \citet{2009ApJ...690....1O} to determine its central time. We
fix the planet-to-star radius ratio, $R_{\rm p}/R_\star$, the ratio of
the semi-major axis to the stellar radius, $a/R_\star$, the cosine of
the orbital inclination, $\cos i_{\rm orb}$ {at those values} from
the archive, adopt the coefficients for the quadratic limb-darkening
law, $u_1$ and $u_2$, from \citet{2012ApJ...751..112J}, and assume zero
orbital eccentricity ($e$). Since only the out-of-transit outliers were
removed in the first step, we also iteratively remove in-transit
{$>5\sigma$} outliers. The resulting transit times are used to
phase fold all the transits and to improve the transit parameters and
orbital period $P$.
Using these revised transit parameters, we again fit each transit
lightcurve for its central time and total duration. Here we assume
$e=0$, fix the values of $u_1$, $u_2$, $a/R_\star$, $R_{\rm p}/R_\star$,
and $P$, and float only central transit time and $\cos i_{\rm orb}$.
From these transit times, we calculate the revised ephemeris
$t_0(\mathrm{BJD}) - 2454833 = 121.3585049(49)$ and $P =
2.204735427(13)$ days by linear regression. Since we find no systematic
TTVs, hereafter we assume that the orbit of HAT-P-7b is described by the
strictly periodic Keplerian orbit with $t_0$ and $P$ obtained above.
\subsubsection{Orbital Eccentricity and Mean Stellar Density from the Phase-folded Transit and Occultation\label{sssec:hatp7-lc-e}}
The top and middle panels of Figure \ref{best_tra_occ} respectively show
the transit and occultation lightcurves stacked using the revised
ephemeris. The lightcurves are binned {into $1$-minute bins} and the
uncertainty of {the flux at the $i$-th bin},
$\sigma_{i,\mathrm{MAD}}$, is calculated as $1.4826 \times
\mathrm{median\ absolute\ deviation}$ divided by the square root of the
number of data points in the bin \citep{1969drea.book.....B}. Solid
lines are the best-fit lightcurves obtained from the simultaneous fit to
both lightcurves. We use the transit model by
\citet{2002ApJ...580L.171M}, and binned model fluxes are calculated by
averaging fluxes sampled at 0.1-minute interval. { In this figure,
the transit and occultation are shifted in time by $t_{{\rm c,\,tra}}$
and $P/2+ t_{{\rm c,\,tra}}$, respectively, where $t_{{\rm c,\,tra}}$ is
the central time of the phase-folded transit lightcurve. This parameter
is introduced to take into account the uncertainty in $t_0$, and the
best-fit value of $t_{{\rm c,\,tra}}$ is indeed within that uncertainty
(see Table \ref{tab:hatp7-joint}).} In the transit residuals (top
panel), we reproduce the anomaly first reported by
\citet{2013ApJ...764L..22M}, who attributed it to the planet-induced
gravity darkening.
Since the asymmetry of the planetary orbit alters the relative duration
of the transit and occultation, as well as their time interval, one can
tightly constrain the orbital eccentricity from the combination of
transits and occultations; see equations (33) and (34) in
\citet{2011exop.book...55W} for instance. {The bottom panel of
Figure \ref{best_tra_occ} illustrates this subtle effect by comparing
the best-fit transit and occultation lightcurves. Here the depth of the
occultation is scaled by $\delta$, the occultation depth divided by $(R_{\rm
p}/R_\star)^2$, for ease of comparison. In this panel, the
egress of the occultation occurs slightly later than that of the
transit, while the difference is smaller for their ingresses. In other
words, our best-fit model indicates that the occultation duration is
longer than the transit one and that the center of occultation deviates
from $P/2$.} These are most likely due to the asymmetry of the orbit
introduced by the slight but non-zero eccentricity, as well as the time
delay of $4.5\times10^{-4}$ days due to the finite speed of light (twice
the orbital semi-major axis divided by the speed of light; calculated
for $M_\star = 1.59\,M_\odot$). In fact, with the non-zero eccentricity
and the above light-travel time included, the simultaneous fit to the
phase-folded transit and occultation lightcurves give tight constraints
on the planet's eccentricity, $e\cos\omega = 0.00026\pm0.00015$ and
$e\sin\omega = 0.0041\pm0.0022$, where $\omega$ is the argument of
periastron measured from the plane of the sky.
Since $e\sin\omega$ and $a/R_\star$ are degenerate in determining the
transit durations, the tight constraint on $e\sin\omega$ also allows the
accurate determination of $a/R_\star$, and hence the mean stellar
density $\rho_\star$ independently from asteroseismology
\citep{2003ApJ...585.1038S}. We obtain $a/R_\star=4.131\pm0.009$ from
the above fit, and then deduce $\rho_\star = (0.275\pm0.002)\times
10^3\,\mathrm{kg\,m}^{-3}$ from
\begin{equation}
\rho_\star = \frac{3\pi}{GP^2} \left(\frac{a}{R_\star}\right)^3 \left(1+\frac{M_{\rm p}}{M_\star}\right)^{-1} ,
\end{equation}
where $G$ denotes the gravitational constant, and $M_{\rm p}/M_\star
\sim 10^{-3}$ can be neglected. This value is larger than
$\rho_{\star,\rm s}$ based on the seismic scaling relation by
$2.4\sigma$, but consistent with $\rho_{\star,\rm m}$ from the stellar
model at the $1\sigma$ level (see Table \ref{tab:output_model}). For
this reason, we adopt the constraints from the stellar model as the
prior information in the following joint fit. The choice of the prior,
however, does not affect the spin--orbit angle determination, but only
slightly changes the values of $a/R_\star$, $\rho_\star$, $\cos i_{\rm
orb}$, and $e\sin\omega$. The slight discrepancy between $\rho_\star$
from the seismic scaling relation ($\rho_{\star,\rm s}$) and that from
transit and occultation implies that the current precision of the {\it
Kepler} photometry has reached the level that could permit an
independent test of the seismic scaling relation for the mean stellar
density.
\begin{figure}
\centering
\includegraphics[width=8.2cm,clip]{Figures/best_lc_rev.eps}
\caption{Phase-folded transit (top) and occultation (middle)
lightcurves. Points are the binned fluxes ($1\,\mathrm{min}$)
and solid lines show the best-fit model lightcurves. Vertical
dashed and dotted lines correspond to the four contact points;
see figure 2 of \citet{2011exop.book...55W} for their
definitions. In the bottom panel, we compare the durations and
central times of best-fit transit and occultation lightcurves.
Occulation is shifted by $P/2$ in time in the middle and the
bottom panels, and its depth is scaled by $\delta$ in the bottom
panel for ease of comparison.} \label{best_tra_occ}
\end{figure}
\subsection{Joint Analysis
\label{ssec:hatp7-joint}}
\subsubsection{Method
\label{sssec:hatp7-joint-method}}
In this subsection, we report the joint MCMC analysis of phase-folded
transit and occultation lightcurves (cf. Section \ref{ssec:hatp7-lc})
and RVs (cf. Section \ref{subsec:previous-hatp7}) making use of the
prior constraints on the mean stellar density $\rho_\star$, projected
stellar rotational velocity $v\sin i_\star$, and stellar inclination
$i_\star$ obtained from asteroseismology in Sections
\ref{sec:asteroseismology}--\ref{sec:asteroseismology:K25}. As
discussed in Section \ref{ssec:hatp7-lc}, the precise constraint on
$\rho_\star$ (equivalent to that on $a/R_\star$) helps to lift the
degeneracy between $a/R_\star$ and $e\sin\omega$, thus resulting in
{improved constraints} on these {two} parameters. In
addition, $v\sin i_\star$ is the key parameter for the RM effect along
with $\lambda$, and so the constraints on $v\sin i_\star$ help us to
better determine $\lambda$ from the observed RM signal. Finally,
$i_\star$ is crucial in determining the three-dimensional spin--orbit
angle $\psi$ via Equation (\ref{psi_3d}), which is the {major goal}
of this paper.
In order to properly handle the possible correlation among $\lambda$,
$v\sin i_\star$, and $i_\star$, we adopt the joint probability
distribution for $v\sin i_\star$ and $i_\star$ as the prior in our MCMC
analysis and directly calculate the posterior distribution for $\psi$ by
floating $i_\star$ as well.
It should be noted here that our observables {do not determine} the
sign of $\cos i_\star$ or $\cos i_{\rm orb}$, due to the symmetry with
respect to the plane of the sky. In order to take into account this
inherent degeneracy, we randomly change the sign of the first term in
Equation (\ref{psi_3d}) in computing $\psi$. Since the probability
distribution of $\rho_\star$ is almost independent {of} those of
$v\sin i_\star$ and $i_\star$, we include the constraint on this
parameter as an independent Gaussian with the central value and width of
$\rho_{\star,\rm m}$ {listed} in Table \ref{tab:output_model}.
We adopt the same model (including non-zero eccentricity and
light-travel time) for transit and occultation as in Section
\ref{ssec:hatp7-lc}. {The observed} RVs are modeled as
\begin{equation}
v_{\star, \mathrm{model}}(t) = v_{\star,\mathrm{orb}}(t) + v_{\star, \mathrm{RM}}(t) + \gamma_i + \dot \gamma (t - t_0).
\end{equation}
Here,
\begin{equation}
v_{\star,\mathrm{orb}} = K_\star \left[ \cos(\omega + f) + e \cos \omega \right]
\end{equation}
is the stellar orbital RVs for the Keplerian orbit, where $K_\star$ is
the RV semi-amplitude and $f$ is the true anomaly of the planet. The
$\gamma_i$ ($i=1, 2$) are the constant offsets for RVs from Keck/HIRES
($i=1$) and Subaru/HDS ($i=2$), and $\dot{\gamma}$ accounts for the
linear trend in the observed RVs in the W09 data set \citep{Winn2009,
2012PASJ...64L...7N, 2014ApJ...785..126K}. Finally, anomalous RVs due
to the RM effect, $v_{\star, \mathrm{RM}}$, are modeled following
\citet{Hirano2011}. The parameters {characterizing} the RM model
{include} $v\sin i_\star$ (projected rotational velocity of the
star), $\beta$ (Gaussian dispersion of spectral lines), $\gamma$
(Lorentzian dispersion of spectral lines), $\zeta$ (macroturbulence
dispersion of spectral lines), $u_{1\mathrm{RM}}+u_{2\mathrm{RM}}$, and
$u_{1\mathrm{RM}}-u_{2\mathrm{RM}}$ (coefficients for the quadratic
limb-darkening law in the RM effect). We do not take into account the
effect of convective blueshift \citep{2011ApJ...733...30S}, as its
typical amplitude ($\sim 1\,\mathrm{m\,s^{-1}}$) is smaller than the
(jitter-included) precision of {the} RVs analyzed here.
We impose the non-seismic priors as well {on} some of the model
parameters. For the ephemeris, we use the Gaussian priors $t_0
(\mathrm{BJD}) -2454833 = 121.3585049 \pm 0.0000049$ and $P =
2.204735427 \pm 0.000000013\,\mathrm{days}$ obtained from the transit
lightcurves. The priors on the RM parameters {($\beta$, $\gamma$, $\zeta$,
$u_{1\rm RM} + u_{2\rm RM}$, and $u_{1\rm RM} - u_{2\rm RM}$)} are almost the same as in
A12. Namely, we fix $\beta=3\,\mathrm{km\,s^{-1}}$ and
$\gamma=1\,\mathrm{km\,s^{-1}}$, and assume Gaussian prior $\zeta = 5.18
\pm 1.5\,\mathrm{km\,s^{-1}}$. We fix the value of
$u_{1\mathrm{RM}}-u_{2\mathrm{RM}}$ at $-0.023$ from the tables of
\citet{2000A&A...363.1081C} for the Johnson V band and the ATLAS model.
The value is obtained using the {\tt jktld}
tool\footnote{\texttt{http://www.astro.keele.ac.uk/jkt/codes/jktld.html}}
for the parameters $T_{\rm eff}=6350\,\mathrm{K}$, $\log
g\,(\mathrm{cgs})=4.07$, and $[{\rm Fe/H}]=0.3$. The value of
$u_{1\mathrm{RM}}+u_{2\mathrm{RM}}$ is floated around the tabulated
value of $0.70$ assuming the Gaussian prior of width $0.10$. In
addition, we impose an additional Gaussian prior on $v\sin i_\star$
based on the spectroscopic value in Table \ref{tab:nonseism_obs},
because the seismic constraint on this parameter is independent {of}
the spectroscopic $v\sin i_\star$. We assume uniform priors for the
other $13$ fitting parameters {listed} in Table
\ref{tab:hatp7-joint} (top and middle blocks).
In the joint fit, we assume the same values of stellar jitter as used in
the original papers{;} $9.3\,\mathrm{m\,s^{-1}}$ for the W09 set,
$3.8\,\mathrm{m\,s^{-1}}$ for the Keck/HIRES RVs of the N09 set, and
$6.0\,\mathrm{m\,s^{-1}}$ for the A12 set. In order to prevent the
transit and occultation lightcurves from placing unreasonably tight
constraints compared to RVs, we also {increase the errors quoted
for} photometric data (evaluated in Section \ref{sssec:hatp7-lc-e}) as
$\sigma_i = \sqrt{ \sigma_{i,\mathrm{MAD}}^2 + \sigma_{\rm r}^2}$.
Here, $\sigma_{\rm r} = 5.8 \times 10^{-6}$ is a parameter analogous to
the RV jitter and chosen so that the reduced $\chi^2$ of the lightcurve
fit becomes unity. This prescription is also motivated by the following
two facts. First, $\sigma_{i,\mathrm{MAD}}$ tends to underestimate the
true uncertainty because it neglects the effect of correlated noise.
Indeed, when the number of data points is sufficiently large,
uncertainties are dominated by the correlated or ``red'' noise component
\citep{2006MNRAS.373..231P}. Second, the systematic residuals of the
best-fit transit model (top panel of Figure \ref{best_tra_occ}) suggest
other effects that are not taken into account in our model [e.g.,
possible planet-induced gravity darkening discussed by
\citet{2013ApJ...764L..22M}]. Placing too much weights on such features
could bias the transit parameters.\\
\subsubsection{Results}
Constraints on the system parameters from the joint analysis are
summarized in Table \ref{tab:hatp7-joint}. The ``parameters mainly
derived from lightcurves/RVs'' are the model (fitted) parameters, while
the ``derived quantities'' are the parameters derived from the model
parameters (along with $M_\star$ and $R_\star$ in Table
\ref{tab:output_model} for $M_{\rm p}$, $R_{\rm p}$, and $\rho_{\rm
p}$). While our result {is in a reasonable agreement} with previous
studies \citep[c.f.,][]{2013ApJ...764L..22M, 2013ApJ...772...51E,
2013ApJ...774L..19V}, it provides two major improvements.
First, we determine the orbital eccentricity of HAT-P-7b essentially
from the photometry (i.e., transit, occultation, and asteroseismology)
alone. A similar method has recently been {employed} by
\citet{2014ApJ...782...14V} to constrain the planet's orbital
eccentricity using the seismic stellar density \citep[see
also][]{2012ApJ...756..122D, 2014MNRAS.440.2164K}, but here we show that
this method is {also useful} for such a low-eccentricity orbit.
Furthermore, our result is even more precise and reliable because
{it takes into account} the independent constraint on $\rho_\star$ and $e$
from the occultation lightcurve.
Second, we obtain the probability distribution for the three-dimensional
spin--orbit angle $\psi$ in a consistent manner. In the case of
HAT-P-7, the constraint on $\psi$ is not very strong because the modest
splitting of the azimuthal modes only allows a weak constraint on
$i_\star$ (see Figure \ref{fig:is-nus-hatp7}). Nevertheless, we find
that the peak values of $\psi$ shift towards $90^\circ$ compared to
those obtained from the ``random'' $i_\star$ (uniform in $\cos i_\star$)
in all three data sets, as shown in Figure \ref{psi_hatp7}. Moreover,
the methodology presented here can be applied to other systems, for some
of which asteroseismology may be able to tightly constrain $i_\star$
unlike HAT-P-7. We will show that this is indeed the case for the
Kepler-25 system in the next section.
\renewcommand{\arraystretch}{0.7}
\begin{table*}
\caption{Parameters of the HAT-P-7 System from the Joint Analysis.}
\centering
\label{tab:hatp7-joint}
\small
\begin{tabular}{cccc}
\toprule
Parameter & Value (W09) & Value (N09) & Value (A12)\\
\midrule
\multicolumn{4}{c}{\it Parameters mainly derived from lightcurves (transit, occultation, asteroseismology)}\\
\midrule
$t_0(\mathrm{BJD})-2454833$ & \multicolumn{3}{c}{$121.3585049 \pm 0.0000049$}\\
$P$ (days) & \multicolumn{3}{c}{$2.204735427 \pm 0.000000013$}\\
$e\cos\omega$ & $0.00024\pm0.00020$ & $0.00024\pm0.00020$ & $0.00025\pm0.00020$\\
$e\sin\omega$ & $0.0053_{-0.0021}^{+0.0022}$ & $0.0057_{-0.0026}^{+0.0025}$ & $0.0049_{-0.0030}^{+0.0026}$\\
$u_1$ & $0.3540\pm0.0034$ & $0.3544_{-0.0034}^{+0.0033}$ & $0.3545_{-0.0035}^{+0.0034}$\\
$u_2$ & $0.1670_{-0.0054}^{+0.0055}$ & $0.1663_{-0.0053}^{+0.0055}$ & $0.1661_{-0.0055}^{+0.0056}$\\
$\rho_\star$ ($10^3\,\mathrm{kg\,m^{-3}}$) & $0.2736\pm0.0016$ & $0.2731_{-0.0018}^{+0.0021}$ & $0.2737_{-0.0018}^{+0.0024}$\\
$\cos i_{\rm orb}$ & $0.12149_{-0.00057}^{+0.00056}$ & $0.12166_{-0.00068}^{+0.00063}$ & $0.12145_{-0.00081}^{+0.00061}$\\
$R_{\rm p}/R_\star$ & $0.077589_{-0.000021}^{+0.000020}$ & $0.077593\pm0.000020$ & $0.077591_{-0.000021}^{+0.000020}$\\
$\delta$ & \multicolumn{3}{c}{$0.01171\pm0.00010$}\\
{$t_{{\rm c,\,tra}}$} (days) & \multicolumn{3}{c}{$-0.0000044_{-0.0000042}^{+0.0000041}$}\\
$i_\star$ ($^\circ$) & $31_{-16}^{+33}$ & $33_{-20}^{+34}$ & $33_{-20}^{+34}$\\
\midrule
\multicolumn{4}{c}{\it Parameters mainly derived from RVs}\\
\midrule
$K_\star$ ($\mathrm{m\,s^{-1}}$) & $211.7\pm2.3$ & $213.2\pm1.8$ & $214.0\pm4.6$\\
$\gamma_1$ ($\mathrm{m\,s^{-1}}$) & $-15.5\pm3.0$ & $-37.5\pm1.5$ & $10.4_{-1.6}^{+1.5}$ \\
$\gamma_2$ ($\mathrm{m\,s^{-1}}$) & $-9.7\pm1.7$ & $-16.9\pm1.4$ & --\\
$\dot{\gamma}$ ($\mathrm{m\,s^{-1}\,yr^{-1}}$) & $21.5\pm2.5$ & -- & --\\
$\lambda$ ($^\circ$) & $186_{-11}^{+10}$ & $220.3_{-9.3}^{+8.2}$ & $157_{-13}^{+14}$\\
$v\sin i_\star$ ($\mathrm{km\,s^{-1}}$)& $4.15_{-0.39}^{+0.38}$ & $3.17\pm0.33$ & $3.17_{-0.34}^{+0.33}$\\
$\beta$ ($\mathrm{km\,s^{-1}}$) & \multicolumn{3}{c}{$3.0$ (fixed)}\\
$\gamma$ ($\mathrm{km\,s^{-1}}$) & \multicolumn{3}{c}{$1.0$ (fixed)}\\
$\zeta$ ($\mathrm{km\,s^{-1}}$) & $5.3\pm1.5$ & $5.5\pm1.5$ & $5.5\pm1.5$\\
$u_{1\mathrm{RM}}+u_{2\mathrm{RM}}$ & \multicolumn{3}{c}{$0.70\pm0.10$}\\
$u_{1\mathrm{RM}}-u_{2\mathrm{RM}}$ & \multicolumn{3}{c}{$-0.23$ (fixed)}\\
\midrule
\multicolumn{4}{c}{\it Derived quantities}\\
\midrule
$\psi$ ($^\circ$) & $122_{-18}^{+30}$ & $115_{-16}^{+19}$ & $120_{-18}^{+26}$\\
$a/R_\star$ & $4.1269_{-0.0078}^{+0.0082}$ & $4.1245_{-0.0092}^{+0.0103}$ & $4.1277_{-0.0090}^{+0.0121}$\\
transit impact parameter ($R_\star$) & $0.4987\pm0.0013$ & $0.4989\pm0.0013$ & $0.4988_{-0.0014}^{+0.0013}$\\
$T_{14,\mathrm{tra}}$ (days) & $0.164301\pm0.000022$ & $0.164303\pm0.000023$ & $0.164300\pm0.000023$\\
$T_{23,\mathrm{tra}}$ (days) & $0.133042_{-0.000048}^{+0.000049}$ & $0.133034_{-0.000048}^{+0.000047}$& $0.133037_{-0.000048}^{+0.000052}$\\
$T_{\mathrm{tra}}$ (days) & $0.148672_{-0.000024}^{+0.000025}$ & $0.148668\pm0.000024$ & $0.148669_{-0.000024}^{+0.000025}$\\
occultation impact parameter ($R_\star$) & $0.5040_{-0.0023}^{+0.0022}$ & $0.5047_{-0.0028}^{+0.0025}$ & $0.5039_{-0.0033}^{+0.0024}$\\
$T_{14,\mathrm{occ}}$ (days) & $0.16555_{-0.00050}^{+0.00051}$ & $0.16566_{-0.00061}^{+0.00058}$ & $0.16547_{-0.00070}^{+0.00060}$\\
$T_{23,\mathrm{occ}}$ (days) & $0.13385_{-0.00033}^{+0.00034}$ & $0.13392_{-0.00040}^{+0.00039}$ & $0.13379_{-0.00046}^{+0.00041}$\\
$T_{\mathrm{occ}}$ (days) & $0.14970_{-0.00041}^{+0.00042}$ & $0.14979_{-0.00051}^{+0.00048}$ & $0.14963_{-0.00058}^{+0.00050}$\\
occultation depth (ppm) & \multicolumn{3}{c}{$70.5\pm0.6$}\\
$M_{\rm p} (M_{\rm J})$ & $1.86\pm0.03$ & $1.87\pm0.03$ & $1.88\pm0.05$\\
$R_{\rm p} (R_{\rm J})$ & \multicolumn{3}{c}{$1.526\pm0.008$}\\
$\rho_{\rm p}$ ($10^3\,\mathrm{kg\,m^{-3}}$) & $0.65\pm0.01$ & $0.66\pm0.01$ & $0.66\pm0.02$\\
\bottomrule
\\
\multicolumn{4}{l}{\parbox{0.9\textwidth}{
Note --- The quoted best-fit values are the medians of their MCMC posteriors,
and uncertainties exclude 15.87\% of values at upper and lower extremes.
The $T_{ij}$ ($i, j = 1, 2, 3, 4$) is the duration between the two contact points $i$ and $j$
[see figure 2 of \citet{2011exop.book...55W} for their definitions], and $T=(T_{14} + T_{23})/2$.
The subscript ``tra" refers to transits and ``occ" to occultations.
}
}
\end{tabular}
\end{table*}
\renewcommand{\arraystretch}{1.0}
\begin{figure}
\centering \includegraphics[bb= 50 60 200 300,
width=8.5cm]{Figures/psi_hatp7.eps}
\caption{Probability distributions for the three-dimensional spin--orbit
angle $\psi$ of HAT-P-7b for the W09 (top), N09 (middle), and
A12 (bottom) data sets. Solid red lines show the posteriors
from the joint analysis, while the black ones are the
probability distributions obtained from uniform $\cos i_\star$
and the posteriors of $\lambda$ and $i_{\rm orb}$ from the joint
analysis (Table \ref{tab:hatp7-joint}). The median, $1\sigma$
lower limit, and $1\sigma$ upper limit for each distribution are
shown in vertical dotted lines. A small bump around $\psi
\approx 95^\circ$ in each panel originates from the fact that
each posterior shown here is the superposition of the two
inherently degenerate configurations with the opposite signs of
$\cos i_\star \cos i_{\rm orb}$; see the discussion in the
second paragraph of Section \ref{sssec:hatp7-joint-method}.}
\label{psi_hatp7}
\end{figure}
\section{Joint Analysis of the Kepler-25 System
\label{sec:kepler-25}}
\subsection{Method
\label{ssec:method-kepler-25}}
We repeat almost the same analysis for Kepler-25c as in Section
\ref{sec:hat-p-7}. There are, however, several differences in the
lightcurve and RV analyses as described below, mainly due to the
multiplicity of the Kepler-25 system and relatively small
signal-to-noise ratio of the Kepler-25c's transit:
\begin{enumerate}
\item We phase-fold the transits using the actually observed transit
times rather than those calculated from the linear ephemeris. This is
because the transit times of Kepler-25c ($P = 12.7$ days) exhibit
significant TTVs due to the proximity to the $2:1$ mean-motion resonance
with Kepler-25b ($P = 6.2$ days). This is why we do not allow
{$t_{{\rm c, \,tra}}$}, the central time of the phase-folded
transit, to be a free parameter. We adopt $\sigma_{\rm r} = 1.6 \times
10^{-5}$ based on the $\chi^2$ of the lightcurve fit.
\item The {occultation} of Kepler-25c was not detected and not taken
into account in the following analysis.
\item As the quality of the transit lightcurve of Kepler-25c is not so
good as that of HAT-P-7b, we could not determine the
limb-darkening coefficients very well.
For this reason, we impose the prior $u_1-u_2=-0.0015\pm0.50$ based on the tables of \citet{2000A&A...363.1081C}, and choose
{$u_1+u_2$ and $u_1-u_2$, instead of $u_1$ and $u_2$, as free
parameters.}
We {made sure} that the choice of the confidence interval for
$u_1-u_2$ does not affect the constraint on $\psi$.
\item In order to take into account the other planets in the RV fit, we
allow the orbital semi-amplitude $K_\star$ and RV offset $\gamma$
for each of the nights in 2011 and 2012 to be free parameters, as
in A13. RV jitters are fixed at $3.3\,\mathrm{m\,s^{-1}}$.
\item We do not fit the orbital eccentricity but fix $e=0$, because we
do not analyze the occultation nor RVs throughout the orbit
\citep{2014ApJS..210...20M}.
\item We assume the independent Gaussian priors
$u_{1\mathrm{RM}}+u_{2\mathrm{RM}} = 0.69\pm0.10$ and $\zeta =
4.85\pm1.5\,\mathrm{km\,s^{-1}}$ from A13, and fix
$u_{1\mathrm{RM}}-u_{2\mathrm{RM}} = -0.0297$ from the tables of
\citet{2000A&A...363.1081C}.
\end{enumerate}
\subsection{Results
\label{ssec:results-kepler-25}}
In the case of the Kepler-25 system, the uncertainty in $\psi$ is
significantly reduced by virtue of the seismic information. This
situation is clearly illustrated in Figure \ref{psi_kep25}, which
compares the posterior probability distribution for $\psi$ from the
joint fit (solid red line) to that based on $\lambda$ and $i_{\rm orb}$
from the joint fit and the uniform $\cos i_\star$ (solid black line).
The corresponding system parameters are summarized in Table
\ref{tab:kepler25-joint}. They are basically consistent with
those {obtained} by A13, except for the increased precision in the
transit parameters.
\begin{figure}
\centering
\includegraphics[width=8.5cm,clip]{Figures/psi_kep25.eps}
\caption{Probability distributions for the three-dimensional
spin--orbit angle $\psi$ of Kepler-25c. The solid red line
shows the posterior from the joint analysis, while the black one
is the probability distribution obtained from $\lambda$ and
$i_{\rm orb}$ in Table \ref{tab:kepler25-joint} and uniform
$\cos i_\star$. The median, $1\sigma$ lower limit, and
$1\sigma$ upper limit for each distribution are shown in
vertical dotted lines.} \label{psi_kep25}
\end{figure}
Interestingly, our result suggests a spin--orbit misalignment for
Kepler-25c with more than $2\sigma$ significance. In order to check the
robustness of this result, we also calculate the probability
distribution of $\psi$ for the seismic $i_\star$ and an independent
Gaussian {$\lambda=-\timeform{0D.5}\pm\timeform{5D.7}$} from the
Doppler tomography. We obtain
$\psi=\timeform{23D.7}_{-\timeform{11D.3}}^{+\timeform{8D.0}}$ in this
case, which still {points to} the spin--orbit misalignment
marginally. If confirmed, this will be the first example of the
spin--orbit misalignment in the multi-transiting system around a
main-sequence star\footnote{The first spin--orbit misalignment in the
multi-transiting system was confirmed by \citet{huber2013} around a red
giant star Kepler-56; they also used asteroseismology.}. The
implication of this result will be discussed in Section
\ref{subsec:summary_K25}.
\begin{table}
\caption{Parameters of the Kepler-25 System from the Joint Analysis (see also the note in Table \ref{tab:hatp7-joint}).}
\centering
\label{tab:kepler25-joint}
\begin{tabular}{cc}
\toprule
Parameter & Value (A13)\\
\midrule
\multicolumn{2}{c}{\it Parameters mainly derived from lightcurves}\\
\multicolumn{2}{c}{\it (transit, asteroseismology)}\\
\midrule
$t_0(\mathrm{BJD})-2454833$ & $127.646558_{-0.000094}^{+0.000096}$\\
$P$ (days) & $12.7203724_{-0.0000013}^{+0.0000014}$\\
$u_1 + u_2$ & $0.550\pm0.018$\\
$u_1 - u_2$ & $-0.27\pm0.44$\\
$\rho_\star$ ($10^3\,\mathrm{kg\,m^{-3}}$) & $0.733_{-0.012}^{+0.013}$\\
$\cos i_{\rm orb}$ & $0.04788_{-0.00038}^{+0.00036}$\\
$R_{\rm p}/R_\star$ & $0.03590_{-0.00046}^{+0.00054}$\\
$i_\star$ ($^\circ$) & $65.4_{-6.4}^{+10.6}$ \\
\midrule
\multicolumn{2}{c}{\it Parameters mainly derived from RVs}\\
\midrule
$K_{\star,2011}$ ($\mathrm{m\,s^{-1}}$) & $-13\pm22$\\
$K_{\star,2012}$ ($\mathrm{m\,s^{-1}}$) & $-37\pm30$\\
$\gamma_{2011}$ ($\mathrm{m\,s^{-1}}$) & $-3.5\pm1.3$\\
$\gamma_{2012}$ ($\mathrm{m\,s^{-1}}$) & $2.0\pm1.4$ \\
$\lambda$ ($^\circ$) & $9.4 \pm 7.1$ \\
$v\sin i_\star$ ($\mathrm{km\,s^{-1}}$)& $9.34_{-0.39}^{+0.37}$ \\
$\beta$ ($\mathrm{km\,s^{-1}}$) & \multicolumn{1}{c}{$3.0$ (fixed)}\\
$\gamma$ ($\mathrm{km\,s^{-1}}$) & \multicolumn{1}{c}{$1.0$ (fixed)}\\
$\zeta$ ($\mathrm{km\,s^{-1}}$) & $4.9\pm1.5$\\
$u_{1\mathrm{RM}}+u_{2\mathrm{RM}}$ & $0.69\pm0.10$\\
$u_{1\mathrm{RM}}-u_{2\mathrm{RM}}$ & \multicolumn{1}{c}{$-0.0297$ (fixed)}\\
\midrule
\multicolumn{2}{c}{\it Derived quantities}\\
\midrule
$\psi$ ($^\circ$) & $26.9_{-9.2}^{+7.0}$\\
$a/R_\star$ & $18.44\pm0.11$\\
transit impact parameter ($R_\star$) & $0.8826\pm0.0018$\\
$T_{14,\mathrm{tra}}$ (days) & $0.11925\pm0.00025$\\
$T_{23,\mathrm{tra}}$ (days) & $0.08528_{-0.00069}^{+0.00065}$\\
$T_{\mathrm{tra}}$ (days) & $0.10226_{-0.00037}^{+0.00036}$\\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion
\label{sec:summary}}
\subsection{HAT-P-7
\label{subsec:summary_hatp7}}
From asteroseismology alone, we obtain
$i_\star=\timeform{27D}_{-\timeform{18D}}^{+\timeform{35D}}$ for HAT-P-7
(Figure \ref{fig:is-pdfs-HATP7}). This constraint, combined with the
{\it Kepler} lightcurves and the three independent RM measurements,
yields $\psi=\timeform{122D}_{\timeform{-18D}}^{\timeform{+30D}}$ and
$i_\star=\timeform{31D}^{+\timeform{33D}}_{-\timeform{16D}}$,
$\psi=\timeform{115D}_{\timeform{-16D}}^{\timeform{+19D}}$ and
$i_\star=\timeform{33D}^{+\timeform{34D}}_{-\timeform{20D}}$, and
$\psi=\timeform{120D}_{\timeform{-18D}}^{\timeform{+26D}}$ and
$i_\star=\timeform{33D}_{-\timeform{20D}}^{+\timeform{34D}}$ for the RVs
from W09, N09, and A12, respectively (Figure \ref{psi_hatp7} and Table
\ref{tab:hatp7-joint}). Although the resulting constraints are not very
strong due to the modest splittings of azimuthal modes (see Figure
\ref{fig:is-nus-hatp7}), our results suggest that the orbit of HAT-P-7b
is closer to the polar configuration rather than retrograde as $\lambda$
may imply.
It is worth noting that the suggested discrepancies in $\lambda$ and $v\sin i_\star$ in three data sets (cf. Section \ref{subsec:previous-hatp7})
still persist in our analysis.
For a fair comparison with the A12 result, we repeat the same analyses for the W09 and N09 data only including RVs taken over the same night,
but the values of $\lambda$ and $v\sin i_\star$ do not change significantly.
Since we have used the same model of the RM effect and the same priors from the {\it Kepler} photometry for the three sets of data,
our results confirm that the discrepancy comes from the RV data themselves.
As A12 discussed, such a discrepancy may originate from some physics that is not included in the current model of the RM effect,
but its origin is beyond the scope of this paper.
As a by-product of the spin--orbit analysis, we have found that HAT-P-7b
has a {small} but non-zero orbital eccentricity, $e=0.005\pm0.001$
(weighted mean of the three data sets), {which is consistent} with
$e=0.0055_{-0.0033}^{+0.007}$ obtained by \citet{2014ApJ...785..126K}.
Our constraint on $e$ comes from the duration and mid-time of the
occultation of HAT-P-7b relative to those of the transit, along with the
constraint on the mean stellar density $\rho_\star$ from
asteroseismology. This approach is justified by the fact that
$\rho_\star$ from the transit and occultation alone shows a reasonable
agreement with the model stellar density $\rho_{\star,\rm m}$ derived
independently from asteroseismology. The origin of this non-zero $e$
may {deserve further theoretical consideration} because the tides
are expected to damp $e$ rapidly for such a close-in planet like
HAT-P-7b.
\subsection{Kepler-25\label{subsec:summary_K25}}
For Kepler-25, we obtain
$i_\star=\timeform{65D.4}_{-\timeform{6D.4}}^{+\timeform{10D.6}}$ from
the joint analysis. This is slightly better than
$i_\star=\timeform{66D.7}^{+\timeform{12D.1}}_{-\timeform{7D.4}}$ from
asteroseismology alone (Figure \ref{fig:is-pdfs-K25}), mainly due to the
prior on $v\sin i_\star$ from spectroscopy. The constraint on $i_\star$
is better than HAT-P-7 despite the lower signal-to-noise ratio, because
of the greater rotational splitting (see Figure \ref{fig:is-nus-K25}).
This allows us to tightly constrain the spin--orbit angle of Kepler-25c
as $\psi = \timeform{26D.9}_{-\timeform{9D.2}}^{+\timeform{7D.0}}$
(Figure \ref{psi_kep25}). {Our finding is important in two
aspects; 1) this is the first quantitative measurement of $\psi$,
instead of $\lambda$, for multi-planetary systems, except for the
Solar system. 2) Kepler-25 is the first system that exhibits the
significant spin--orbit misalignment among the {multi-transiting systems}
with a main-sequence host star, while it is the second example if we
consider the systems with a red-giant host star, Kepler-56.}
{The spin--orbit misalignment in systems with multiple transiting planets is particularly
interesting for the following reason.}
Considering the transit probabilities of multiple planets, planets' orbital planes are likely to
be coplanar in multi-transiting systems, and hence presumably trace
their natal protoplanetary disks. The spin--orbit misalignment in such
systems, therefore, could be a clue to the processes that tilt a stellar
spin relative to its protoplanetary disk
\citep[e.g.,][]{2010MNRAS.401.1505B, 2011MNRAS.412.2790L,
2012Natur.491..418B}.
{In this context, the} orbital inclinations of the other two planets (Kepler-25b and
Kepler-25d) relative to that of Kepler-25c would help the interpretation
of the observed misalignment. They may be constrained from the analysis
of TTVs and Transit Duration Variations (TDVs), along with orbital RVs
to constrain the orbit of the outer non-transiting planet d. In this
paper, we did not model these phenomena because our main concern is the
determination of the spin--orbit angle. It should be noted, however,
that the independent information on $\rho_\star$ from asteroseismology
benefits the TTV analysis as well because TTVs are sensitive to the mean
stellar density and orbital eccentricity of the planets
\citep[e.g.,][]{2012Natur.487..449S, 2014ApJ...783...53M}.\\
{
It is also interesting to note that both HAT-P-7 and Kepler-25 are relatively hot stars with $T_{\rm eff}
\gtrsim 6300\,\mathrm{K}$ and in line with the observed trend that the
spin--orbit misalignments are preferentially found around stars with
$T_{\rm eff} > 6250\,\mathrm{K}$ \citep{Winn2010b}.
Although \cite{2012ApJ...758L...6R} suggested that temporal variations of
the stellar rotation due to internal gravity waves could explain this empirical trend,
we found no evidence to support this scenario for the two systems.
Regarding} {HAT-P-7, we compared
the rotational splitting from Figure \ref{fig:is-pdfs-HATP7} with that
from Q0 to Q2 \citep[results from the study of][]{Oshagh2013}, but
found no evidence of significant variations. Although results using only
Q0 to Q2 have large uncertainties, this may indicate that the rotation
remains constant over time. Moreover, we tightly constrained the
rotation of Kepler-25 and showed that outer layers certainly rotate at
constant velocity. This is incompatible with the scenario suggested by \cite{2012ApJ...758L...6R}.}
\section{Summary
\label{sec:9}}
The major purpose of the present paper is two-fold. The first is to
develop and describe a detailed methodology of determining the
three-dimensional spin--orbit angle $\psi$ for transiting planetary
systems. The other is to demonstrate the power of the methodology by
applying to the two {specific systems}, HAT-P-7 and
Kepler-25.
The application of asteroseismology to exoplanetary systems is now
becoming popular. It is particularly useful in determining the stellar
inclination $i_\star$ with respect to the line-of-sight. Combined with
the orbital inclination $i_{\rm orb}$ determined for transiting systems,
and with the projected spin-orbit angle $\lambda$ via the spectroscopic
observation of the Rossiter-McLaughlin effect, the joint analysis
{presented in this paper} indeed enables the determination of
$\psi$, rather than $\lambda$. While the observed distribution of
$\lambda$ for more than 70 transiting systems [e.g., figure 7 of
\citet{Xue2014}] already put tight constraints on planetary migration
scenarios, that of $\psi$ is even more useful because it is free from
the projection effect. As we discussed, HAT-P-7 seems to host a
polar-orbit planet instead of a retrograde one as {naively}
suspected from the observed $\lambda \approx \timeform{180D}$. The
determination of $\psi$ is also important for multi-transiting planetary
systems where all the planets are supposed to share the same orbital
plane; large $\psi$ in such a system indicates that the stellar
obliquity experiences significant tilt with respect to the
protoplanetary disk that would be the orbital plane of the planets. This
turned out to be the case for Kepler-25 as we discussed in the previous
section.
{While it may be premature to consider the statistics at this point, it is tempting to note that two out of the {six multi-transiting systems with measured spin--orbit angles are shown to be significantly misaligned.
The misaligned cases are Kepler-25c ($\psi = \timeform{26D.9}_{-\timeform{9D.2}}^{+\timeform{7D.0}}$) and Kepler-56 \citep[$i_\star = \timeform{47D} \pm \timeform{6D}$,][]{huber2013}, while the aligned cases are Kepler-30 \citep[$\lambda \lesssim \timeform{10D}$,][]{2012Natur.487..449S}, Kepler-50 and Kepler-65 \citep[$i_\star = \timeform{82D}^{+\timeform{8D}}_{-\timeform{7D}}$ and $i_\star = \timeform{81D}^{+\timeform{9D}}_{-\timeform{16D}}$,][]{Chaplin2013} and Kepler-89d (KOI-94d) with $\lambda = -\timeform{6D}^{+\timeform{13D}}_{-\timeform{11D}}$ \citep{Hirano2012b} or $-\timeform{11D} \pm \timeform{11D}$ (A13). }
Even if the spin--orbit misalignment is rare, the physical mechanism for its origin is an interesting theoretical question. If it indeed turns out to be fairly common, it will pose a serious challenge to all viable theories of the formation and evolution of multi-planetary systems.}
In addition to the determination of $\psi$, the joint analysis improves
the accuracy and precision of numerous system parameters for a specific
target. In turn, any discrepancy among the separate analyses strongly
points to a certain physical process which needs to be taken into
account in the detailed modeling. This would open a new window for the
exploration of the origin and evolution of planetary systems.
\bigskip
We are grateful to Simon Albrecht and Josh Winn for providing us with the radial velocity data of Kepler-25.
We thank NASA and the {\it Kepler} team for their revolutionary data.
O.B. is supported by Japan Society for Promotion of Science (JSPS)
Fellowship for Research (No. 25-13316). K.M. is supported {by JSPS Research Fellowships for Young Scientists (No. 26-7182) and}
by the Leading Graduate Course for Frontiers of Mathematical Sciences and Physics.
Y.S. gratefully acknowledges the support from the Grant-in Aid for
Scientific Research by JSPS (No. 24340035).
\bibliographystyle{apj}
|
\section{Early history: Euler, Abel, Jacobi, Riemann}
Our story, like many others in mathematics, can be traced back at least
to Euler who studied elliptic integrals of the form
$$
\int \frac{dx}{\sqrt{x^3+ax^2+bx+c}}.
$$
The study of integrals of algebraic functions was further developed
by Abel and Jacobi. From our point of view the next major step
was taken by Riemann. Instead of dealing with a multi-valued function
like $\sqrt{x^3+ax^2+bx+c} $, Riemann looks at the
complex algebraic curve
$$
C:=\bigl\{(x,y): y^2=x^3+ax^2+bx+c\bigr\}\subset \cc^2.
$$
Then the above integral becomes
$$
\int_{\Gamma} \frac{dx}{y}
$$
for some path $\Gamma$ on the
algebraic curve $C$.
More generally, a polynomial $g(x,y)$ implicitly defines
$y:=y(x)$ as a multi-valued function of $x$ and
for any meromorphic function $h(u,v)$, the multi-valued integral
$$
\int h\bigl(x, y(x)\bigr)\ dx
$$
becomes a single valued integral
$$
\int_{\Gamma} h\bigl(x, y\bigr)\ dx
$$
for some path $\Gamma$ on the
algebraic curve
$C(g):=\bigl(g(x,y)=0\bigr)\subset \cc^2$.
Substitutions that transform one integral associated
to a polynomial $g_1$ into another integral associated
to a $g_2$
can be now seen as algebraic maps between the
curves $C(g_1) $ and $C(g_2) $.
Riemann also went further. As a simple example, consider
the curve $C$ defined by $(y^2=x^3+x^2)$ and notice that
$(t^3-t)^2\equiv(t^2-1)^3+(t^2-1)^2$.
Thus the substitution
$x=t^2-1,\ y=t^3-t$ (with inverse $t=y/x$) allows us to transform any integral
$$
\int h\bigl(x, \sqrt{x^3+x^2}\bigr)\ dx
\qtq{into}
\int h\bigl(t^2-1, t^3-t\bigr)\cdot 2tdt.
$$
To put it somewhat differently, the map
$$
t\mapsto \bigl(x=t^2-1, y=t^3-t\bigr)
\qtq{and its inverse}
(x,y)\mapsto t=y/x
$$
establish an isomorphism
$$
\left\{
\begin{array}{c}
\mbox{meromorphic functions}\\
\mbox{on the curve $(y^2=x^3+x^2)$}
\end{array}
\right\}
\leftrightarrow
\left\{
\begin{array}{c}
\mbox{meromorphic functions}\\
\mbox{on the complex plane $\cc$}
\end{array}
\right\}.
$$
It is best to work with meromorphic functions on $\cc$ that are
also meromorphic at infinity; these live naturally on the
Riemann sphere $\cc\pp^1$. We can now state Riemann's
fundamental theorem as follows.
\begin{thm}[Riemann, 1851] \label{riemann.thm}
For every algebraic curve
$C\subset \cc^2$ there is a unique, compact Riemann surface $S$
and a meromorphic map $\phi:S\map C$ with meromorphic inverse
$\phi^{-1}:C\map S$ such that
$$
f_C\mapsto f_S:=f_C\circ \phi\qtq{and} f_S\mapsto f_C:=f_S\circ \phi^{-1}
$$
establish an isomorphism between the meromorphic function theory on
$C$ and the meromorphic function theory on
$S$.
\end{thm}
\section{Main questions, informally}
We can now give an initial formulation of the two main problems
that we consider; the precise versions are stated in
Sections \ref{blocks.sec} and \ref{moduli.sec}.
The first is a direct higher-dimensional analog of the
results of Riemann. (See Section \ref{sec.defn} for basic definitions.)
\begin{main-ques} \label{prob1}
Given an algebraic variety $X$, is there another algebraic
variety $X^{\rm m}$ such that
\begin{enumerate}
\item the meromorphic function theories of $X$ and of $X^{\rm m}$
are isomorphic and
\item the geometry of $X^{\rm m}$ is the ``simplest'' possible?
\end{enumerate}
\end{main-ques}
Riemann's theorem says that, in dimension 1, ``simplest''
should mean smooth and compact, but in higher dimensions
smoothness is not the right notion. One of the hardest aspects
of the theory was to understand what the correct concept
of ``simplest'' should be.
So far we have dealt with individual algebraic varieties.
A salient feature of algebraic geometry is that by
continuously varying the coefficients of the defining polynomials
we get continuously varying families of algebraic varieties.
We can thus study how to transform a family $\{X_t:t\in T\}$
of varieties into its ``simplest'' form.
A tempting idea is to take the ``simplest'' forms
$\{X_t^{\rm m}:t\in T\}$ obtained previously.
Unfortunately, this fails already in dimension 1.
Starting with a family of curves $\{C_t:t\in T\}$,
the corresponding Riemann surfaces $\{S_t:t\in T\}$
form a continuously varying family over a dense
open subset $T^0\subset T$ but not everywhere.
For curves the correct answer was found by Deligne and Mumford in 1969.
We use the guidance provided by this 1-dimensional case
and the answer to the first Main Question to answer the
second.
\begin{main-ques} \label{prob2}{\ }
What are the ``simplest'' families of algebraic varieties?
How can one transform an arbitrary family into one of the ``simplest''
families?
\end{main-ques}
\section{What are algebraic varieties?}\label{sec.defn}
Here we quickly recall the basic concepts and definitions that we use.
For general introductory texts, see \cite{shaf, MR0453732, Harris95}.
An {\it affine algebraic set} in $\cc^N$ is
the common zero-set of some polynomials
$$
\begin{array}{rcl}
X^{\rm aff}&=&X^{\rm aff}(f_1,\dots, f_r)\\
&=&\bigl\{(x_1,\dots, x_N):
f_1(x_1,\dots, x_N)=\cdots=f_r(x_1,\dots, x_N)=0\bigr\}\subset \cc^N.
\end{array}
$$
It is especially easy to visualize {\it hypersurfaces} $X(f)\subset \cc^N$
defined by 1 equation.
Usually we count complex dimensions,
thus $\dim \cc^N=N$ and $\dim X$ is one half of the
usual topological dimension of $X$. In low dimensions we talk about
{\it curves, surfaces, 3-folds.}
Thus, somewhat confusingly, an algebraic curve is a
(possibly singular) Riemann surface.
An affine algebraic set $X$ is called {\it irreducible}
if it can not be written as a union of two algebraic sets in a nontrivial way.
Such sets are
called {\it affine algebraic varieties.} Every algebraic set $X$ is a
finite union of algebraic varieties $X=\cup_iX_i$ such that
$X_i\nsubseteq X_j$ for $i\neq j$. Such a decomposition is unique,
up to permuting the indices.
Thus from now on we are interested mainly
in algebraic varieties.
For example, the irreducible components of a hypersurface $X(f)$
correspond to the irreducible factors of $f$, thus $X(f)$
is irreducible iff $f$ is a power of an
irreducible polynomial.
An affine algebraic set $X^{\rm aff} $ is compact iff it is 0-dimensional,
thus it is almost always better to work with the closure
of $X^{\rm aff}$ in the complex projective space
$$
X:=X^{\rm proj}\subset \cc\pp^N.
$$
Thus we get {\it projective algebraic sets} and
{\it projective varieties.} Finally, a {\it quasi-projective variety}
is an open subset $U$ of a projective variety $X$ whose complement
$X\setminus U$ is a projective algebraic set. Note that $U$ is a
``very large'' subset of $X$, in particular $U$ is dense in $X$.
This is a key feature of algebraic geometry: all open subsets are
``very large.''
On a complex projective space $\cc\pp^N$ the
homogeneous coordinates $(x_0{:}\cdots{:}x_N)$
are defined only up to multiplication by a scalar.
Thus one can not evaluate a polynomial
$p(x_0,\dots, x_N)\in \cc[x_0,\dots, x_N]$,
at a point of $\cc\pp^N$. However, if $p$ is homogeneous of degree $d$
then
$$
p(\lambda x_0,\dots, \lambda x_N)=\lambda^dp(x_0,\dots, x_N).
$$
Thus the zero set of $p$ is well-defined and a
quotient of two homogeneous polynomials of the same degree
$$
f(x_0,\dots, x_N)=\frac{p_1(x_0,\dots, x_N)}{p_2(x_0,\dots, x_N)}
$$
is also well-defined (except where $p_2$ vanishes).
These are the {\it rational functions}
on $\cc\pp^N$. By restriction, we get rational functions
on any projective variety $X\subset \cc\pp^N$.
At first sight these seem downright antiquated definitions; a modern
theory ought to be local. That is, one should
consider varieties that are locally defined by analytic functions
and work with meromorphic functions on them. However, we know that
every meromorphic function on $\cc\pp^1$ is rational and
the same holds in all dimensions.
\begin{thm}[Chow, 1949; Serre, 1956]
Let $M\subset \cc\pp^N$ be a closed subset that
can be {\em locally} given as the common zero set of analytic functions. Then
\begin{enumerate}
\item $M$ is algebraic, that is, it can be {\em globally} given as
the common zero set of homogeneous polynomials and
\item every meromorphic function $f$ on $M$ is algebraic,
that is, $f$ can be {\em globally} given as
the quotient of two homogeneous polynomials of the same degree.
\end{enumerate}
\end{thm}
Now we come to a key feature of algebraic geometry.
There are two competing notions of ``map'' and
two competing notions of ``isomorphism.''
\begin{defn}[Map and morphism] Let $X\subset \cc\pp^N$ be an algebraic variety
and $f_0,\dots, f_M$ nonzero rational functions on $X$.
They define a {\it map} (or rational map)
$$
{\mathbf f}:X\map \cc\pp^M
\qtq{given by} p\mapsto \bigl(f_0(p){:}\cdots {:} f_M(p)\bigr)\in \cc\pp^M.
$$
To start with, ${\mathbf f }$ is only defined
at a point $p$ if
none of the $f_i$ has a pole at $p$ and
not all of the $f_i$ vanish at $p$. However,
since the projective coordinates are defined only up to a scalar multiple,
$(gf_0,\dots, gf_M)$ define the same map for any rational function $g$,
thus it can happen that ${\mathbf f }$ is everywhere defined.
In this case it is called a {\it morphism.}
A map is denoted by $\map$ and a morphism by $\to$.
For example, projecting $\cc\pp^2$ from the origin $(0{:}0{:}1)$ to the
line at infinity is given by
$$
\pi:(x{:}y{:}z)\mapsto \bigl(\tfrac{x}{z} : \tfrac{y}{z}\bigr)=
\bigl(\tfrac{x}{y} : 1\bigr)=\bigl(1 : \tfrac{y}{x}\bigr).
$$
Thus $\pi$ is defined everywhere except at $(0{:}0{:}1)$.
\end{defn}
\begin{defn}[Isomorphism]
Two quasi-projective varieties $X\subset \cc\pp^N$ and $Y\subset \cc\pp^M$
are {\it isomorphic} if there are morphisms
$$
f:X\to Y \qtq{and} g:Y\to X
$$
that are inverses of each other. Isomorphism is denoted by $X\cong Y$.
\end{defn}
We will think of isomorphic varieties as being essentially the same.
Using maps instead of morphisms in the above definition yields
the notion of birational equivalence. This notion is
unique to algebraic geometry; it has no known analog in
topology or differential geometry.
\begin{defn}[Birational equivalence]\label{bir.eq.defn}
Two quasi-projective varieties $X\subset \cc\pp^N$ and $Y\subset \cc\pp^M$
are {\it birational}
(in old terminology, birationally isomorphic)
if there are rational maps
$$
f:X\map Y \qtq{and} g:Y\map X
$$
such that the following equivalent conditions hold.
\begin{enumerate}
\item $\phi_Y\mapsto \phi_X:=\phi_Y\circ f$ and
$\phi_X\mapsto \phi_Y:=\phi_X\circ g$
establish an isomorphism between the meromorphic (=rational) function theory on
$X$ and the meromorphic (=rational) function theory on $Y$.
\item There are algebraic subsets $Z\subsetneq X$
and $W\subsetneq Y$ such that
$(X\setminus Z)\cong (Y\setminus W)$.
\end{enumerate}
As an example, consider the affine surface $S:=(x^2+y^2=z^3)\subset \cc^3$.
It is birational to $\cc^2_{uv}$ as shown by the rational maps
$$
f:(x,y,z)\map \bigl(\tfrac{x}{z}, \tfrac{y}{z}\bigr)
\qtq{and}
g:(u,v)\to \bigl(u(u^2+v^2), v(u^2+v^2), u^2+v^2\bigr).
$$
Here $f$ is not defined if $z=0$ while $g$ is everywhere defined
but it maps the pair of lines $(u=\pm iv)$ to the
origin $(0,0,0)$. Thus
$$
S\setminus(z=0)\cong \cc^2\setminus(u^2+v^2=0)
\qtq{but} S\not\cong \cc^2.
$$
\end{defn}
\begin{brot} Let $X, Y$ be algebraic varieties that are
birational to each other.
Many questions of algebraic geometry about $X$
can be answered by
\begin{itemize}
\item first studying the same question on $Y$ and then
\item studying a similar question involving the
lower dimensional algebraic sets $Z$ and $W$ as in (\ref{bir.eq.defn}.2).
\end{itemize}
The aim of the Minimal Model Program is to exploit this
in two steps.
\begin{itemize}
\item Given a question and a variety $X$,
find a variety $Y$ that is
birational to $X$ such that the
geometry of $Y$ is ``best adapted'' to studying
the particular question. This is a variant of the first Main Question.
\item Set up the appropriate dimension induction to deal with
the exceptional sets $Z\subset X$ and $W\subset Y$.
\end{itemize}
\end{brot}
{\it Important aside.}
More generally, if we decompose an algebraic variety into
disjoint locally closed pieces, then the collection of the
pieces carries a lot of information about the variety.
I would like to stress that this is a rather
noteworthy fact about algebraic geometry. For instance, if we decompose a
simplicial complex into its simplices, then usually the only information
we retain is the dimension and the Euler characteristic.
By contrast, all the homology groups of a smooth, projective algebraic
variety can be recovered from
the pieces. This is a key consequence of Hodge Theory, as
formulated by Deligne, and is a starting point of
Grothedieck's theory of motives.
\section{Classical results}
After the study of algebraic curves, two main avenues of
investigations were pursued. One direction focused on the local study
of varieties with a main aim of resolving them completely.
The other direction aimed to understand the global structure
of algebraic surfaces. These are both still very active research areas.
We recall a few of the main results that are relevant for the
general theory. For detailed treatments and for
references see \cite{bpv, k-res}.
\subsection*{Resolution of singularities} {\ }
Riemann's theorem says that
every singular algebraic curve $C$ is birational to a smooth, compact curve
(or Riemann surface). The first steps toward answering the Main Questions
in higher dimensions focused on this problem: {\it Is every algebraic variety
birational to a smooth, projective variety?}
\begin{defn} A variety $X\subset \cc^N$ is smooth and has dimension $d$
at a point $p\in X$ iff the following equivalent conditions hold.
\begin{enumerate}
\item $X\subset \cc^N\cong\rr^{2N}$ is a $C^{\infty}$-submanifold
of (real) dimension $2d$ near $p$.
\item One can choose coordinates $z_1,\dots, z_N$ and
equations $f_1,\dots, f_{N-d}$ of $X$ such that
$(f_1=\cdots = f_{N-d}=0)$ coincides with $X$ near $p$ and
the Jacobian matrix
$\bigl(\partial f_i/\partial z_j: 1\leq i,j\leq N-d\bigr)$ is invertible at $p$.
\item There are holomorphic functions $\phi_i=\phi_i(w_1,\dots, w_d)$
defined near the origin and constants $c_i$ such that
$$
(w_1,\dots, w_d)\mapsto
\bigl(\phi_1({\bf w}), \dots, \phi_{N-d}({\bf w}), w_1+c_1, \dots, w_d+c_d\bigr)
$$
maps a small {\it ball}
${\mathbf 0}\in \bb^d(\epsilon)\subset \cc^d$ onto a neighborhood of $p\in X$.
\end{enumerate}
In the latter case we view $(w_1,\dots, w_d) $ as
{\it local analytic coordinates} on $X$ near $p$.
(It is an ever present technical problem that there is
no good notion of local algebraic coordinates. Open algebraic neighborhoods
are too large to admit a single-valued coordinate system.)
\end{defn}
On an algebraic variety $X$ the set of singular points turns out
to be an algebraic subset, denoted by $\sing X\subset X$.
For every variety $X$, a generalization of Riemann's method (\ref{riemann.thm})
produces a new variety $X^{\rm n}\to X$,
called the {\it normalization} of $X$, such that
$\sing \bigl(X^{\rm n}\bigr)$ has codimension $\geq 2$ in $X^{\rm n}$.
Thus, in higher dimensions, one usually works with
{\it normal varieties} whose singular set has codimension $\geq 2$.
To make the singular set even smaller, or to get rid of it
completely, turned out to be very difficult. The final result
was established by Hironaka in 1964.
\begin{thm}[Resolution of singularities]\label{res.sing.thm}
For every algebraic variety $X$, there are (very many)
smooth, projective varieties $X^{\rm sm}$
birational to $X$.
If $X$ is projective, one can arrange to have
a morphism $f:X^{\rm sm}\to X$ that is an isomorphism
over $X\setminus \sing X$.
\end{thm}
\subsection*{Algebraic surfaces} {\ }
By resolution of singularities, any projective surface $S$
is birational to a smooth projective surface $S^{\rm sm}$, but,
in contrast with the theory of curves, there are many such
smooth projective surfaces $S^{\rm sm}$.
We can thus reformulate the first Main Question:
Is there a ``simplest'' one
among all smooth projective surfaces birational to $S$?
To answer this question, first we study how to make a
smooth projective surface more ``complicated.''
\begin{defn}[Blowing-up] Let $S$ be a smooth algebraic surface
and $p\in S$ a point. {\it Blowing-up} is an operation
that creates a new smooth surface $B_pS$ by removing $p$ and replacing it
with a $\cc\pp^1$ corresponding to all the tangent directions
of $S$ at $p$. Collapsing the new $\cc\pp^1$ to a point gives
a morphism $\pi:B_pS\to S$.
In local coordinates, it can be described as follows.
Start with the unit ball $\bb^2_{xy}\subset \cc^2_{xy}$
and $\cc\pp^1_{st}$ where the subscripts name the coordinates. Set
$$
B_{\mathbf 0}\bb^2_{xy}:=(xt-ys=0)\subset \bb^2_{xy}\times \cc\pp^1_{st}.
$$
Let $\pi:B_{\mathbf 0}\bb^2_{xy}\to \bb^2_{xy}$
denote the coordinate projection.
If $(x,y)\neq (0,0)$ then $\pi^{-1}(x,y)$
is the single point $(x,y)\times(x{:}y)$. However, if
$x=y=0$ then $(s{:}t)$ can be arbitrary, thus
$\pi^{-1}(0,0)\cong \cc\pp^1_{st}$.
Note that $s/t=x/y$ is the natural coordinate on $\cc\pp^1_{st}$,
thus blowing up is akin to switching to polar coordinates
since the polar angle $\theta$ equals $\tan^{-1}(x/y)$.
\end{defn}
One can blow up any number of points of $S$ and then repeat
by blowing up some of the new points of $B_pS$. Thus
blowing up is a cheap way to get infinitely many new smooth surfaces out of one.
\begin{defn} A smooth projective surface is called {\it minimal}
if it can not be obtained from another smooth, projective surface
by blowing up.
\end{defn}
This notion allows us to get a very good analog of Riemann's theorem
\ref{riemann.thm}.
\begin{thm}[Enriques, 1914 ; Kodaira, 1966]\label{min.mod.surf.thm}
For every projective, algebraic surface $S$,
exactly one of the following holds.
\begin{enumerate}
\item (Minimal model) There is a unique, minimal surface $S^{\rm m}$
birational to $S$.
\item $S$ is birational to $C\times \cc\pp^1$
for a unique, smooth, projective curve $C$.
\end{enumerate}
\end{thm}
\begin{say}[Du~Val singularities]\label{duval.say}
It was also gradually understood that instead of working with the
minimal model $S^{\rm m}$, it is sometimes better to use a
slightly singular {\it canonical model} $S^{\rm can}$.
The resulting singularities were first classified by
Du~Val in 1934; the list is quite short, ranging from the simplest
$(x^2+y^2+z^2=0)$ to the most complicated $(x^2+y^3+z^5=0)$.
They are also called {\it rational double points.}
Their importance was not generally
recognized until the 1960's when they were rediscovered from many
different points of view;
see \cite{durfee} for a
survey.
\end{say}
\section{The first Chern class and the Ricci curvature}
The first Chern class,
which is closely related to the Ricci curvature,
carries much of the important information about the
structure of a variety. We follow the differential geometry sign conventions;
algebraic geometers usually work with the {\it canonical class,}
which is (a slight refinement of) the negative of the first Chern class.
\begin{say}[Complex volume forms] A measure on $\rr^n$ can be
identified with an $n$-form
$$
s(x_1,\dots, x_n)\cdot dx_1\wedge\dots\wedge dx_n.
$$
Thus a measure on a real manifold $M$ is an $n$-form
that in local coordinates can be written as above.
Similarly, on a smooth variety $X$ of dimension $n$
a {\it complex volume form}
is an $n$-form $\omega$
that in local holomorphic coordinates can be written as
$$
h(z_1,\dots, z_n)\cdot dz_1\wedge\dots\wedge dz_n.
$$
Thus a complex volume form $\omega$ gives a real volume form
$\left(\frac{\sqrt{-1}}{2}\right)^n\omega\wedge\bar\omega$
where the constant comes from the formula
$$
dz\wedge d\bar z=(dx+\sqrt{-1}dy)\wedge(dx-\sqrt{-1}dy)=
-2\sqrt{-1}\ dx\wedge dy.
$$
(There is usually an additional $\pm$, depending on one's
orientation conventions.)
\end{say}
From the point of view of differential geometry,
one would like to use $C^{\infty}$ complex volume forms,
that is, the $h(z_1,\dots, z_n) $ should be nowhere zero $C^{\infty}$-functions.
Algebraic geometry, however, prefers meromorphic volume forms
where the $h(z_1,\dots, z_n) $ are meromorphic functions.
(See (\ref{jac.sing}.1) for some explicit examples.)
Thus the ideal situation is when a complex volume form
is given by nowhere zero holomorphic functions
$h(z_1,\dots, z_n) $. This is possible only for
Calabi--Yau
varieties; they form a very special but important subclass
(\ref{blocks.ssc.2}).
Thus in general we try to understand how to connect
$C^{\infty}$ and meromorphic volume forms.
On the differential geometry side the key notion is the
curvature which defines the Chern form.
\begin{defn} [Chern form and Chern class]
Let $\omega$ be a $C^{\infty}$ complex volume form.
The {\it first Chern form} or {\it Ricci curvature form} of $(X,\omega)$
is the 2-form
$$
\tilde c_1(X,\omega):= \frac{\sqrt{-1}}{\pi}
\partial\bar\partial \log |h(z_1,\dots, z_n)|=
\frac{\sqrt{-1}}{\pi}\sum_{ij}\frac{\partial^2 \log |h({\mathbf z})|}
{\partial z_i\partial \bar z_j}dz_i\wedge d\bar z_j.
$$
As a 2--form, this depends on the choice of the volume form $\omega$, but
it gives a
well defined De~Rham cohomology class $c^{\rr}_1(X)\in H_{DR}^2(X, \rr)$
which actually lifts to an integral cohomology class
$$
c_1(X)\in H^2(X, \zz),
$$
called the {\it first Chern class} of $X$.
\end{defn}
\begin{defn}[Algebraic degree]
Let $X$ be a smooth, projective variety and $C\subset X$
an algebraic curve. It is not hard to see that there is
always a meromorphic volume form $\omega_m$ that is defined and nonzero
at all but finitely many points of $C$. We define the
{\it degree} of $\omega_m$ on $C$ as
$$
\deg_C\omega_m:=
\#(\mbox{zeros of $\omega_m$ on $C$})-\#(\mbox{poles of $\omega_m$ on $C$}),
$$
where both zeros and poles are counted with multiplicities.
\end{defn}
The Chern form and the algebraic degree are connected by the
Gauss--Bonnet theorem.
\begin{thm} \label{gauss-b.thm}
Let $X$ be a smooth, projective variety.
Let $\omega_r$ be a $C^{\infty}$ complex volume form and
$\omega_m$ a meromorphic volume form. Then, for every
algebraic curve $C\subset X$
$$
\int_C c_1(X)=\int_C \tilde c_1(X, \omega_r)=-\deg_C\omega_m.
\eqno{(\ref{gauss-b.thm}.1)}
$$
\end{thm}
(The minus sign comes from the happenstance that differential geometers
prefer to work with the tangent bundle while the volume forms
use the (determinant of the) cotangent bundle.)
\subsection*{Positivity/negativity and complex differential geometry}{\ }
In differential geometry it is especially nice to
work with metrics whose curvature is everywhere
positive (or everywhere zero or everywhere negative) but
these rarely exist. A usual weakening is to work with
K\"ahler metrics that satisfy the
{\it Einstein condition}: the Ricci curvature
should be a constant multiple of the metric;
see \cite[Chap.19]{MR2325093} for definitions and an introduction.
If this {\it Einstein constant} is positive, then
in (\ref{gauss-b.thm}.1) we integrate an everywhere
positive form. Thus $\int_C c_1(X) $ is positive for every curve $C$.
We hope that in this case there are meromorphic volume forms
with poles (but no zeros).
Similarly, if the Einstein constant is negative, then
in (\ref{gauss-b.thm}.1) we integrate an everywhere
negative form. Thus $\int_C c_1(X) $ is negative for every curve $C$.
We hope that in this case there are holomorphic volume forms
(usually with zeros).
Algebraic geometry can be used to understand the numbers
$\deg_C\omega_m$, hence the values of the integrals $\int_C c_1(X) $.
It is a very difficult task to use the
positivity/negativity of the integrals $\int_C c_1(X) $
to obtain a K\"ahler metric with positive/negative Einstein constant.
For smooth varieties Aubin and Yau proved existence in 1977
when $\int_C c_1(X) $ is always negative or when
$\int_C c_1(X) $ is always zero.
The singular case is treated in \cite{MR2505296, ber-g}.
The positive curvature case is more
subtle; a complete answer is not yet known.
While our approach to the structure of varieties is guided by
these curvature considerations, in algebraic geometry we can understand
only the algebraic degree of the first Chern class. Thus we look at the
functional
$$
C\mapsto \int_Cc_1(X)
$$
and focus on those varieties where this is everywhere
negative (or everywhere zero or everywhere positive).
The Main Conjecture then asserts that every variety can be built up
from these special varieties in a rather clear process.
\section{The Main Conjecture}\label{blocks.sec}
On a typical variety $X$, the Chern class $c_1(X)$ is positive on
some curves and negative on others, in a rather unpredictable way.
Using the first Chern class and the theory of algebraic surfaces as our guide,
we focus on three basic ``especially simple'' types of
smooth, projective varieties.
These are the ``building blocks'' of all algebraic varieties.
\begin{say}[{\bf Negatively curved}]\label{blocks.ssc.1}
These are the varieties where
$\int_Cc_1(X)$ is negative for every curve $C\subset X$.
This is the largest class of the three.
\end{say}
\begin{say}[{\bf Flat or Calabi--Yau}]\label{blocks.ssc.2}
Here $\int_Cc_1(X)$ is zero for every curve $C\subset X$.
They play an especially important role in string theory and mirror symmetry;
see \cite{MR1711184, MR2003030} for introductions.
\end{say}
\begin{say}[{\bf Positively curved or Fano}]\label{blocks.ssc.3}
Here $\int_Cc_1(X)$ is positive for every curve.
There are few of these varieties in each dimension, but
they occur especially frequently in applications.
\end{say}
A simple set of examples to keep in mind is the following.
A smooth hypersurface $X_d\subset \cc\pp^n$ of degree $d$ is
negatively curved if $d>n+1$, flat if $d=n+1$ and
positively curved if $d<n+1$.
A variety in any of these 3 classes is considered ``simplest,''
but we do not yet have enough ``simplest'' varieties for
answering the first Main Question.
For example, taking products of these we get examples where $c_1(X)$
has different signs on different curves. Two of these
possible ``mixed types'' are relevant for us.
Consider a product $X:=N\times F$ of a negatively curved and of a flat variety.
It is clear that $\int_Cc_1(X)\leq 0$ for every curve $C\subset X$
and $\int_Cc_1(X)= 0$ only if $C$ lies in a fiber of the first projection
$N\times F\to N$. This observation leads to the 4th class.
\begin{say}[{\bf Semi-negatively curved or Kodaira--Iitaka type}]
\label{blocks.ssc.4}
Here $\int_Cc_1(X)\leq 0$ for every curve $C\subset X$
and there is a unique morphism $I_X:X\to I(X)$ such that
$\int_Cc_1(X)=0$ iff $C$ is contained in a fiber of $I_X$.
This includes the classes \ref{blocks.ssc.1}--\ref{blocks.ssc.2}:
$I_X$ is an isomorphism for negatively curved varieties and
a constant map in the flat case.
In the intermediate cases, when
$0<\dim I(X)<\dim X$, almost all fibers of $I_X$ are
Calabi--Yau varieties. Thus one can view these as
families of (lower dimensional) Calabi--Yau varieties parametrized by
the (lower dimensional) variety $I(X)$. If we understand families
of (lower dimensional) varieties well enough, we understand $X$.
(This is one of the reasons we are interested in
the second Main Question.)
Furthermore, in these cases $I(X)$ is negatively curved
in a ``suitable sense,'' though we do not yet have a
final agreed-upon definition of what this means.
\end{say}
Next consider a product $X:=N\times P$ of a negatively curved and of a
positively curved variety. If a curve $C$
lies in a fiber of the first projection then
$\int_Cc_1(X)> 0$, but there are many other such curves.
Nonetheless, the first projection
is uniquely determined by $X$ and this
leads to the definition of the 5th class.
\begin{say}[{\bf Positive fiber type}]\label{blocks.ssc.5}
I really would like to say that in these cases
there is a unique morphism $m_X:X\to M(X)$ such that
$M(X)$ is semi-negatively curved and
$c_1(X)$ is positive on all the fibers.
(To avoid trivial cases, we also assume that $\dim M(X)<\dim X$.)
This, unfortunately, still does not give enough ``simplest'' varieties for
the first Main Question.
It took quite some time to arrive at the correct definition,
to be discussed in Section \ref{ratconn.sec}.
\end{say}
We can now state a precise version of
the first Main Question.
\begin{main-conj} \label{main-conj.1}
Every algebraic variety $X$ is birational to a variety
$X^{\rm m}$ that is either of type
{\rm (\ref{blocks.ssc.4})} or of type {\rm (\ref{blocks.ssc.5})}.
\end{main-conj}
\noindent{\bf Complement.} $X^{\rm m}$
-- especially in case
(\ref{blocks.ssc.4}) -- is called a
{\it minimal model} of $X$.
In the semi-negatively curved case $I\bigl(X^{\rm m}\bigr)$
is unique but $X^{\rm m} $ itself is not. However, it is quite well understood
how the different $X^{\rm m} $ are related to each other.
(This is the story of {\it flops,} see \cite{MR1159257, MR2827807}.)
By contrast, in case (\ref{blocks.ssc.5})
it is very hard to determine
when two such varieties $X^{\rm m}_1 $ and $X^{\rm m}_2 $ are birational.
\smallskip
\noindent{\bf Caveat.} While the Main Conjecture is expected to be true,
in general
one has to allow {\it terminal} singularities --
to be defined in (\ref{terminal.defn}) --
on $X^{\rm m}$.
This was a rather
difficult point historically
since over a century of experience suggested
that singularities should be avoided.
For surfaces terminal = smooth, thus the issue of singularities
did not come up in Theorem \ref{min.mod.surf.thm}.
By now the correct classes of
singularities have been established and, for many questions we consider,
they do not seem to cause any problems. We describe these singularities
in Section \ref{sing.sec}.
\begin{say}[Traditional names] \label{trad.names.say}
A variety $X$ is said to be
of {\it general type} if $\dim I\bigl(X^{\rm m}\bigr)=\dim X$.
In this case $X\map I\bigl(X^{\rm m}\bigr)$ is birational and
$I(X):=I\bigl(X^{\rm m}\bigr)$ is called the
{\it canonical model} of $X$; it
has canonical singularities (\ref{can-lc.defn}).
We see in Section \ref{moduli.sec}
that the second Main Question
has a good answer for families of
canonical models.
The {\it Kodaira dimension} of a variety $X$ is the dimension
of $I\bigl(X^{\rm m}\bigr)$.
The Kodaira dimension is defined to be
$-\infty$ for the class (\ref{blocks.ssc.5}).
The Main Conjecture is usually broken down into two parts
that are, in principle, independent of each other.
The first part separates the classes \ref{blocks.ssc.4}\ and \ref{blocks.ssc.5}
from each other and the second part
provides the structural description in case \ref{blocks.ssc.4}.
These forms first appear in Reid's paper \cite[Sec.4]{r-mmc3}.
\medskip
\noindent \ref{trad.names.say}.1
{\it Minimal Model Conjecture.}
Every algebraic variety $X$ is birational to a
variety
$X^{\rm m}$ such that
either $c_1\bigl(X^{\rm m}\bigr)$ is semi-negative
or there is a morphism to a lower dimensional variety
$\pi:X^{\rm m}\to S$ such that
$\int_Cc_1\bigl(X^{\rm m}\bigr)>0$ if $C$ is contained in a fiber of $\pi$.
(In the second case the map $\pi$ need not be
unique and it does not give the best structural description.)
\medskip
\noindent \ref{trad.names.say}.2
{\it Abundance Conjecture.}
If $c_1(Y)$ is semi-negative then
there is a unique morphism $I_Y:Y\to I(Y)$ such that
$\int_Cc_1(Y)=0$ iff $C$ is contained in a fiber of $I_Y$.
\end{say}
\section{Rationally connected varieties}\label{ratconn.sec}
Before we consider minimal models, we describe
the structure we expect for varieties in the
5th class (\ref{blocks.ssc.5}).
An introduction aimed at non-specialists is given in
\cite{k-simplest}. More detailed accounts are in
\cite{ar-ko, rc-book}.
Clebsch and Max Noether noticed around 1860--1870
that, when the numerical invariants suggest that
a surface could be birational to $\cc\pp^2$, then it is.
The final result along these lines was established by
Castelnuovo in 1896.
Analogous questions in higher dimension turned out to be
much harder. Fano classified smooth positively curved 3--folds
around 1930. (He missed some cases though, so did subsequent ``complete''
lists produced in the 1970's and then in the 1980's. The (hopefully)
final list was not established until 2003.)
This is, however, one area where the singularities do matter;
we still do not know all positively curved 3--folds
with terminal singularities.
It appears that instead of global descriptions
we should focus on {\it rational curves} in a variety;
these are the images of morphisms $\phi:\cc\pp^1\to X$.
For a projective variety $X$, the following dichotomy is quite easy
to establish.
i) either the rational curves cover a subset of $X$
which is {\it meager}
(that is, a
countable union of nowhere dense closed subsets)
ii) or the rational curves cover all of $X$.
These two cases correspond to the alternatives in
the Main Conjecture.
That is, if
$X$ is birational to a semi-negatively curved variety
then rational curves cover a meager subset and, conjecturally,
the converse also holds.
The correct approach to the best structural description
of the 5th class \ref{blocks.ssc.5}
was not discovered until 1992
(Koll\'ar--Miyaoka--Mori \cite{KMM92a}).
The key observation is that
we should even change the class \ref{blocks.ssc.3}.
Instead of a curvature description,
we should focus on rational curves contained in a variety.
\begin{defn}
A projective variety $X$ is called
{\it rationally connected} if, for any number of points $x_1,\dots, x_r\in X$,
there is a
morphism $\phi:\cc\pp^1\to X$ whose image passes through $x_1,\dots, x_r$.
\end{defn}
I claim that rationally connected varieties constitute the
``correct'' birational version of being
positively curved. This is not a precise mathematical assertion
since not every rationally connected variety is birational to
a positively curved variety, not even when singularities are allowed.
Rather, the assertion is that any answer to the first Main Question
needs to work with rational connectedness
instead of positivity of curvature.
\begin{say}[Supporting evidence]\label{supp.evid.rc.say}{\ }
It is easy to see that $\cc\pp^n$ is rationally connected.
More generally, every positively curved variety is
rationally connected (Nadel \cite{Nadel91}, Campana \cite{Campana92b},
Koll\'ar--Miyaoka--Mori \cite{KMM92c}, Zhang \cite{MR2208131}).
Being rationally connected is invariant under smooth deformations
and birational maps \cite{KMM92c}.
Rationally connected varieties share key arithmetic properties of
rational varieties over $p$-adic fields (Koll\'ar \cite{k-loc}),
finite fields (Koll\'ar--Szab\'o \cite{MR2019976}, Esnault
\cite{MR1943746})
and function fields of curves (Graber--Harris--Starr \cite{ghs},
de~Jong--Starr \cite{MR1981034}).
The loop space of a rationally connected variety is
also rationally connected (Lempert--Szab\'o \cite{MR2372727}).
\end{say}
The notion of rational connectedness allows us to give the
correct description of the class \ref{blocks.ssc.5}.
A weaker variant is proved in \cite{KMM92a};
the form below combines this with \cite{ghs}.
\begin{thm} Let $X$ be a variety that is covered by rational curves.
Then there is a unique
(up to birational equivalence) map
$m_X:X\map M(X)$ such that
\begin{enumerate}
\item almost all fibers of $m_X$ are rationally connected and
\item rational curves cover only a meager subset of $M(X)$.
\end {enumerate}
\end{thm}
There are
two main open geometric problems about rationally connected varieties.
The first concerns a topological characterization.
In its naive form the question asks: What can we tell about a variety
from its underlying topological space? It seems that the answer is: not much.
However, the underlying topological space of a smooth variety
carries a natural symplectic structure and this seems to incorporate
much more information.
\begin{conj} {\rm \cite[Conj.4.2.7]{k-lowdeg}} Being rationally connected is a
property of the
underlying symplectic structure.
\end{conj}
For partial results see \cite{k-lowdeg, MR2521651}.
The other problem asks if we could strengthen the
definition of rationally connected varieties.
Note that $\cc\pp^n$ contains not just many rational curves
but also many higher dimensional rational
subvarieties (hyperplanes, hyperquadrics, ...).
Maybe this is also a general property of rationally connected varieties?
As far as I know, 3-dimensional rationally connected varieties
always contain rational surfaces. I believe, however,
that this is not the case in higher dimension.
\begin{conj} {\rm \cite[Prob.56]{k-simplest}}
Many rationally connected varieties
do not contain any rational surface.
\end{conj}
\section{Minimal Models}
This is a short history of {\it Mori's program}, also called
the {\it Minimal Model Program} and frequently abbreviated as {\it MMP}.
For general introductions see \cite{CKM88, km-book}
or the technically more detailed \cite{ka-ma-ma, MR2359340, hac-kov}.
\begin{say}[{\bf Iitaka's program, 1970--85}]
This approach predates
the Main Conjecture. At the beginning it was not
even suspected that the Main Conjecture could be true, in fact,
lacking the right class of singularities,
it was assumed that the Main Conjecture
would fail for most varieties. Thus the aim of Iitaka's program was
to sort varieties into 5 broad types that (as we now know) exactly
correspond to the ones in (\ref{blocks.ssc.1}--\ref{blocks.ssc.5}).
The main contributors were,
in rough historical order,
Iitaka, Ueno, Fujita, Kawamata, Viehweg and Koll\'ar;
see \cite{ueno, Mori87} for surveys.
\end{say}
\begin{say}[{\bf Canonical and terminal singularities, Reid 1980--83}]
Reid was studying higher dimensional
analogs of Du~Val singularities of surfaces (\ref{duval.say});
obtaining rather complete
descriptions in dimension 3.
It was quite important that when
Mori's program lead to singularities, the relevant classes were
already there and were known to be well behaved.
An especially readable account is
\cite{r-ypg}.
\end{say}
\begin{say}[{\bf The birth of Mori's program, 1981--88}]
Mori's groundbreaking paper \cite{Mori82} introduces 3 new ideas.
If $c_1(X)$ is not semi-negative then, by definition,
$c_1(X)$ is positive on some curve $C\subset X$.
Mori first proves that there is such a rational curve;
that is, there is a morphism $\phi:\cc\pp^1\to X$ such that
$c_1(X)$ is positive on its image.
It is quite remarkable that
the proof goes through algebraic geometry over finite fields.
To this day there is no proof known that avoids this; in particular
this step is not yet known for complex manifolds that are not algebraic.
Second, he identifies the ``most positive'' such maps
$\phi:\cc\pp^1\to X$; this is called {\it extremal ray} theory.
Third, in dimension 3 he gives a complete description of
all extremal rays and the resulting map $X\to X_1$
that removes the ``most positive'' part of $X$.
The program now seems clear (at least in dimension 3).
Repeat the procedure for $X_1$
and prove that after finitely many steps we end up
with $X\to X_1\to\cdots\to X_r$ such that $c_1(X_r)$ is semi-negative.
This is called {\it Mori's program} or
{\it Minimal Model program.}
There are two, rather formidable, problems.
In many cases the new variety $X_1$ is smooth but
sometimes it is singular. Luckily, these singularities have been
studied by Reid, at least in dimension 3.
Still, it is necessary to
establish the above 3 steps for singular varieties.
This was accomplished rather rapidly by
Kawamata, Reid, Shokurov and Koll\'ar.
The program was first written down in \cite[Sec.4]{r-mmc3}.
The more serious problem is that in some cases
taking the contraction $X_i\to X_{i+1}$ is clearly not the right step.
Instead we have to take a step back and construct a new
variety $X_i^+ $ that sits in a {\it flip diagram}
$$
\xymatrix{%
X_i \ar@{-->}[rr]^{\phi_i} \ar[rd]_(.45){p_i}
&& X_i^+ \ar[ld]^(.35){p_i^+} \\
& X_{i+1} &
}
$$
Geometrically, we start with $X_i$, find an especially
badly behaving $\cc\pp^1\cong C_i\subset X_i$ and remove it.
Then we compactify the resulting $X_i\setminus C_i$
by attaching another curve $C_i^+\cong \cc\pp^1$ but differently.
The key difference is a sign change:
$$
\int_{C_i} c_1\bigl(X_i\bigr)>0\qtq{but} \int_{C_i^+} c_1\bigl(X_i^+\bigr)<0.
$$
This operation is called a {\it flip.}
For more about flips, see \cite{MR1159257, MR2827807}.
Flips are reminiscent of {\it Dehn surgery} in 3--manifold topology
where we remove a circle and put it back differently.
In dimension 3 the existence of flips is proved in a very difficult paper
by Mori \cite{mori-MR924704}, which completes the program in this case.
A detailed description of 3--dimensional flips is given in
\cite{MR1149195}. The list is rather lengthy; this makes it unlikely that
a similarly complete answer will ever be worked out in higher dimensions.
\end{say}
\begin{say}[{\bf Log variants: Kawamata, Shokurov, 1984--1992}]
The Iitaka program established that for many results one can
work with cohomology classes in $H^2(X,\rr)$ that are close enough to
the first Chern class.
This turned out to be a very powerful tool. By choosing
the perturbations appropriately, we can focus our attention
on one or another part of a variety. These are somewhat technical questions
but by now we understand how to work with them and
most applications of the Minimal Model Program use a perturbed case.
\end{say}
\begin{say}[{\bf Abundance: Kawamata, Miyaoka, 1987--1992}]
Even for surfaces, the Abundance Conjecture \ref{trad.names.say}.2
is a rather subtle result.
It is even harder for 3--folds. The proofs use many special
properties of surfaces; this is why the higher dimensional cases
are still not well understood.
A rather complete account of the 3--dimensional
methods is given in \cite{k-etal}.
\end{say}
\begin{say}[{\bf Inductive approach in low dimensions: Shokurov, 1992--2003}]
In retrospect, the
key development of the decade was an inductive approach to flips.
A detailed treatment of the 3--dimensional case is given in
\cite{k-etal}. For the rest of the nineties progress was slow, culminating
in a treatment of 4--dimensional flips.
There were many technical difficulties to overcome and the
importance of these methods was not fully appreciated at
first since the dimension reduction leads to a much more complicated
problem that seems to fail in higher dimensions.
\end{say}
\begin{say}[{\bf The Corti seminar, 2003--2005}]
Over the course of several years a group led by Corti
developed the previous ideas further and integrated them with the
rest of the program \cite{MR2359340}.
This provided the bridge to the general case.
\end{say}
\begin{say}[{\bf The general type case: Hacon and M\textsuperscript{c}Kernan,
2005--2010}]
The real breakthrough was achieved in \cite{MR2359343}
where the existence of flips in dimension $n$ was reduced to
an instance of the MMP in dimension $n-1$. This left a
series of global questions to resolve. The paper
\cite{bchm} settled everything for varieties of general type.
A good introduction is in \cite{MR2743824}.
At about the same time
Siu started to develop an analytic approach which aims to get
$I(X^{\rm m})$, without going through the
individual steps; see
\cite{MR2827808} for an overview.
An algebraic variant of this is in \cite{cas-laz}.
\end{say}
\begin{say}[{\bf Abundance: Hacon and Xu, 2012--}]
Although the Abundance conjecture is known in very few cases,
there has been significant progress when $\dim I(X)$ is
expected to be close to $\dim X$. The log version of the special case when
$\dim I(X)=\dim X$ is especially important for applications
in moduli theory. These have been settled in \cite{MR3032329, MR2955764}.
\end{say}
\begin{say}[{\bf Positive characteristic, Hacon and Xu, Birkar, Patakfalvi,
2012--}]
Mori's original works are very geometric
and these ideas quickly lead to a simple proof of the
2-dimensional case of the Main Conjecture in positive characteristic.
However, subsequent developments
rely very heavily on Kodaira-type vanishing theorems that
are known to fail in positive characteristic, although no actual
failure is known in the cases used by the program.
The 3--dimensional case was recently settled in
\cite{2013arXiv1302.0298H, 2013arXiv1311.3098B}.
Substantial parts of the Iitaka program are proved
in positive characteristic in \cite{2013arXiv1308.5371P}.
\end{say}
\begin{say}[{\bf Open problems}] From our point of view, the main open problem
is to complete the missing parts of the Main Conjecture.
It is known that the MMP always runs,
that is, the sequence of contractions and flips
$X=X_1\map X_2\map \cdots$ exists.
The problem is that
it is not clear how to prove that the process eventually stops.
In the 3--dimensional case, Mori's approach provides a
rather complete description of the steps of the MMP.
This gives many ways to show that each step
improves various invariants and that eventually the process stops.
By contrast, the method of Hacon--M\textsuperscript{c}Kernan produces the steps
of the MMP in a rather indirect way. We have very little
information about the steps beyond their existence.
\end{say}
\section{Singularities of the Minimal Model Program} \label{sing.sec}
So far we have been sweeping the singularities of the
minimal models under a rug, but it is time for a look at them.
Understanding the correct class of singularities
is crucial in the development of the structure theory of algebraic varieties.
This is a somewhat technical subject with many
difficult questions and methods but by now we understand
these singularities well enough that in many questions
they do not cause any problems. A rather complete treatment is given in
\cite{kk-singbook}. Here I focus on the main ideas behind the definitions.
Given a variety $Y$, one frequently looks at a resolution
of singularities $f:X\to Y$ as in Theorem \ref{res.sing.thm}
and translates problems on $Y$ to
questions on $X$. Then the hard part is to interpret the answer obtained on
$X$ in terms of $Y$. Here the key seems to be the inverse function
theorem.
\begin{say}[{\bf The inverse function theorem}]
The classical inverse function theorem says that if
${\mathbf f}:=(f_1,\dots, f_n):\rr^n_{\mathbf x}\to \rr^n_{\mathbf y}$
is a differentiable map then ${\mathbf f} $
has a local inverse at a point $p\in \rr^n_{\mathbf x}$
iff the Jacobian determinant
$$
\operatorname{Jac}({\mathbf f}):=
\det\left(\frac{\partial f_i}{\partial x_j}\right)
$$
does not vanish at $p$. We can also think about it
in terms of the ``standard'' volume forms
$\omega_{\mathbf x}:=dx_1\wedge\cdots\wedge dx_n$ and
$\omega_{\mathbf y}:=dy_1\wedge\cdots\wedge dy_n$.
Then
$$
{\mathbf f}^*\omega_{\mathbf y}=
\operatorname{Jac}({\mathbf f})\cdot \omega_{\mathbf x},
$$
thus the vanishing/non-vanishing of the Jacobian tells us
how the pull-back of the ``standard'' volume form of the target
compares to the ``standard'' volume form of the source.
Note that the Jacobian itself depends on the choice of the coordinates,
but its vanishing or non-vanishing depends only on ${\mathbf f} $.
In the complex analytic setting
one can use the ``standard'' complex volume forms
$\omega_{\mathbf z}:=dz_1\wedge\cdots\wedge dz_n$ and
$\omega_{\mathbf w}:=dw_1\wedge\cdots\wedge dw_n$
on the unit balls $\bb^n_{\mathbf z}\subset \cc^n_{\mathbf z}$
and $\bb^n_{\mathbf w}\subset \cc^n_{\mathbf w}$.
Given a holomorphic map
${\mathbf f}:=(f_1,\dots, f_n):\bb^n_{\mathbf z}\to \bb^n_{\mathbf w}$
we get that
$$
{\mathbf f}^*\omega_{\mathbf w}
=\det\left(\frac{\partial f_i}{\partial z_j}\right)\cdot \omega_{\mathbf z}
=:\operatorname{Jac}({\mathbf f})\cdot \omega_{\mathbf z},
$$
and ${\mathbf f} $
has a local inverse
iff $\operatorname{Jac}({\mathbf f}) $
does not vanish at $p$.
\end{say}
\begin{say}[{\bf The Jacobian in the singular case}]\label{jac.sing}
Let $X$ be a normal algebraic variety
and $p\in X$ a singular point. It is quite easy to see
that if $\omega_1, \omega_2$ are two holomorphic volume forms on
$X\setminus\sing X$ in a neighborhood of a singular point $p\in \sing X$
then
there is a unique holomorphic function $\phi$ such that
$\omega_1=\phi\cdot\omega_2$ and $\phi(p)\neq 0$.
Thus all holomorphic volume forms on $X\setminus\sing X$
have the same asymptotic
behavior near $\sing X$.
The local existence of such forms is a slightly technical question,
so let us just focus on an example.
If
$Y=\bigl(f(w_1,\dots, w_{n+1})=0\bigr)\subset \cc^{n+1}$
is a hypersurface then the ``standard'' volume form is given by
$$
\omega_Y=(-1)^i
\frac{dw_1\wedge \cdots\wedge dw_{i-1}\wedge dw_{i+1}\wedge\cdots\wedge dw_{n+1}}
{\partial f/\partial w_i}.
\eqno{(\ref{jac.sing}.1)}
$$
(It is easy to check that this is independent of $i$. Note also
that $\omega_Y $ is not defined when all of the
$\partial f/\partial w_i $ vanish; which happens exactly on $\sing Y$.)
Thus if $f:\bb^n_{\mathbf z}\to Y$ is
holomorphic then we can define the {\it Jacobian} of $f$
by the formula
$$
\operatorname{Jac}(f):=\frac{f^*\omega_Y}{ \omega_{\mathbf z}}.
$$
Note that due to the denominators in (\ref{jac.sing}.1),
in general $\operatorname{Jac}(f) $ can have poles.
For example, consider the singularity
$Y_{n,d}:=\bigl(w_1^d+\cdots+w_n^d=w_{n+1}^d\bigr)\subset \cc^{n+1}$
and a holomorphic map $f:\bb^n_{\mathbf z}\to Y$ given by
$$
f:(z_1,\dots, z_n)\to
\bigl(z_1, z_1z_2, \dots, z_1z_n, z_1\sqrt[d]{1+ z_2^d+\cdots+z_n^d}\bigr).
\eqno{(\ref{jac.sing}.2)}
$$
Then $\omega_{Y_{d,n}}=-d^{-1}w_1^{1-d} dw_2\wedge \cdots\wedge dw_{n+1}$ and
we easily compute that the Jacobian of $f$ has a zero/pole of order $n-d$
along the hyperplane $(z_1=0)$.
As in the classical case, the Jacobian of $f$ depends on the choice of
the ``standard'' volume forms but the vanishing/non-vanishing
or the order of vanishing of the Jacobian depends only on $f$.
We can now define terminal singularities; these form the smallest
possible class needed for the Main Conjecture.
\end{say}
\begin{defn}\label{terminal.defn}
A normal variety $Y$ has
{\it terminal} singularities iff the
inverse function theorem holds for $Y$. That is, if
$f:\bb^n_{\mathbf z}\to Y$ does not have a local inverse at $p\in \bb^n_{\mathbf z}$
then $\operatorname{Jac}(f)$ vanishes at $p$.
(There is a small problem when the exceptional set of $f$ is
too small, we can ignore it for now.)
\end{defn}
For canonical models and for moduli questions, two more types of
singularities are needed.
\begin{defn}\label{can-lc.defn}
A normal variety $Y$ has
{\it canonical} singularities iff
$\operatorname{Jac}(f) $ is holomorphic for every $f:\bb^n_{\mathbf z}\to Y$
and {\it log-canonical} singularities iff
$\operatorname{Jac}(f) $ has at most simple poles
for every $f$.
\end{defn}
The above computations suggest (and it is indeed true) that
$Y_{n,d}$ (as in \ref{jac.sing}.2) is terminal iff $d<n$, canonical iff
$d\leq n$ and log canonical iff
$d\leq n+1$.
\begin{say}[Local volume of $Y$ near $\sing Y$] A good way to think about
these singularities is as follows. Pick a point $p\in \sing Y$
and let $\omega_Y$ be a ``standard'' local complex volume form.
Then $(\sqrt{-1}/2)^n \omega_Y\wedge\bar\omega_Y$ is a real volume form
and we can ask about the
{\it local volume of $X$}, that is,
$\int_U (\sqrt{-1}/2)^n \omega_Y\wedge\bar\omega_Y$
for a suitably small neighborhood $p\in U\subset X$.
If $Y$ has a canonical singularity near $p$ then the
local volume is finite. In the log-canonical case
the local volume is infinite but barely. If $g$ is any
holomorphic function vanishing on $\sing Y$ then
$\int_U|g|^{\epsilon}(\sqrt{-1}/2)^n \omega_Y\wedge\bar\omega_Y$
is finite for every $\epsilon>0$.
\end{say}
\begin{say}[{\bf Intermediate differential forms}]
On an $n$-dimensional variety we have so far considered
holomorphic $n$-forms only but for several questions
one also needs to understand the pull-back $f^*\eta$ of
lower degree differential forms
as well. This proved to be surprisingly difficult but
almost all local questions were settled by
Greb--Kebekus--Kov\'acs--Peternell \cite{GKKP10}.
\end{say}
\section{Moduli of varieties of general type}\label{moduli.sec}
Let ${\mathbf X}$ be a class of projective varieties,
for instance curves or surfaces of a certain type.
The theory of moduli aims to find ``optimal'' ways
to write down all varieties in the class ${\mathbf X}$.
This is a large theory with many aspects. The 3 volumes of
\cite{mod-hand} contain surveys of most of the active areas.
Here my aim is to focus on just one of them: the
moduli of varieties of general type.
Introductions are given in
\cite{MR2483953, hac-kov, k-modsurv} while a detailed treatment should be
in \cite{k-modbook}.
We start with the historically first example.
\begin{exmp}[Elliptic curves] They can all be given by
an affine equation
$$
E(a,b,c):=(y^2=x^3+ax^2+bx+c)\subset \cc^2;
$$
the corresponding projective curve has a unique
point $[p]$ at infinity.
Here $c_1(E)=0$, so it is best to think of this as
elliptic curves with a marked point $[p]$.
The curve $E(a,b,c) $ is smooth iff the
discriminant of the cubic
$$
\Delta(a,b,c):=18abc-4a^3c+a^2b^2-4b^3-27c^2\qtq{is not zero.}
$$
Two such curves
are isomorphic iff there is an affine-linear transformation
$(x,y)\mapsto (\alpha^2 x+\beta, \alpha^3 y)$
that transforms one equation into the other.
All these transformations form a (2-dimensional) group $G$.
Thus we get the following.
{\it Version 1.}
The isomorphism classes of all elliptic curves are in one-to-one
correspondence with the orbits of $G$ on
$\cc^3\setminus(\Delta(a,b,c)=0)$.
Next we need to identify the $G$-orbits.
The key is the $j$-invariant
$j\bigl(E(a,b,c)\bigr):= 2^{8}(a^2-3b)^3/
\Delta(a,b,c)$.
(The factor $2^{8}$ is not important for us, it is
there for number-theoretic reasons.)
It is not very hard to work out the following.
{\it Version 2.}
Two elliptic curves are isomorphic iff they have the same
$j$-invariant.
We can restate this as follows:
{\it Version 3.}
The moduli space of elliptic curves is the complex line
${\mathcal M}_1\cong \cc$
and the value $j(E)$ of the $j$-invariant gives the point
in ${\mathcal M}_1$ that corresponds to $E$.
The only sensible compactification of $\cc$ is $\cc\pp^1$,
so what corresponds to the point at infinity? This should be
a curve where the discriminant of the cubic $x^3+ax^2+bx+c$ vanishes.
That is, when $x^3+ax^2+bx+c $ has a multiple root.
There are 2 types of such cubics. If there is a triple root
we get $y^2=x^3$, a cuspidal curve. If there is a double root
we get $y^2=x^3+x^2$, a nodal curve. In this case the correct choice is to
go with the nodal curve.
\end{exmp}
\begin{say}[The main steps of a moduli theory] \label{main.steps.moduli}
We hope to do something similar with more general
algebraic varieties. We proceed in several steps.
{\it Step 1.} Identify a class of projective varieties ${\mathbf X}$
that should have a ``good'' moduli theory.
We aim to prove that such a theory exists for
negatively curved varieties as in (\ref{blocks.ssc.1}).
We allow canonical singularities, thus this includes
canonical models of varieties of general type.
(It seems that in most other
cases there is no ``good'' compactified moduli theory, unless
some additional structure is added on, for instance an ample divisor
as in \cite{ale-abvar}.)
{\it Step 2.} Add some extra data (also called rigidification) first.
A typical extra datum
is an embedding
$j:X\DOTSB\lhook\joinrel\to \pp^N$ for some $N$.
Use the additional data to get a moduli space
with a universal family
$$
{\mathbf U}_{\mathbf X,j}\subset \pp^N\times {\mathbf M}_{\mathbf X,j}
\qtq{with projection}
\pi_{\mathbf X,j}:{\mathbf U}_{\mathbf X,j}\to {\mathbf M}_{\mathbf X,j}
$$
such that
every pair $(X,j)$ occurs exactly once among the fibers of $\pi_{\mathbf X,j}$.
(It is not easy to show that one can choose a fixed $N$ that works for all
varieties in a given class. For smooth varieties
this was proved by Matsusaka in 1972; the general case was settled
recently by Hacon and Xu.)
{\it Step 3.} Next we get rid of the extra data. Usually
we have to take a quotient
by a Lie group like $\mathrm{GL}(N+1, \cc)$.
This can be hard but, if everything works out,
at the end we have
$$
{\mathbf U}_{\mathbf X}:={\mathbf U}_{\mathbf X,j}/\mathrm{GL}(N+1, \cc),\quad
{\mathbf M}_{\mathbf X}:={\mathbf M}_{\mathbf X,j}/\mathrm{GL}(N+1, \cc)
$$
and a morphism
$\pi_{\mathbf X}:{\mathbf U}_{\mathbf X}
\to {\mathbf M}_{\mathbf X}$. (See Step 6 for the possible dependence on $N$.)
{\it Step 4.} In almost all cases, the resulting spaces are not compact
and compactifying them in a ``good'' way is difficult.
The key step is to
identify the limits of families of varieties in ${\mathbf X}$
that should give a ``good'' compact moduli theory.
There is no a priori reason to believe that such a choice exists
or that it is unique. Finding the right choice in higher dimension
was the last conceptual step in the program.
For canonical models of varieties of general type we have the
``right'' answer, see (\ref{mod.KSB}) and (\ref{mod.gen}).
{\it Step 5.} We have to go back and redo Steps 1--3 for this more
general class of objects to get a compactified moduli theory
$$
\bar\pi_{\mathbf X}:\overline{\mathbf U}_{\mathbf X}
\to \overline{\mathbf M}_{\mathbf X}.
$$
{\it Step 6.} An extra issue that arises is that
the compactifications could also depend on the
dimension of the $\pp^N$ chosen in Step 2. This does not seem to happen
for ${\mathbf M}_{\mathbf X}$ itself (at least for $N$ large enough)
but it does happen for $\overline{\mathbf M}_{\mathbf X} $
for some of the proposed
variants.
{\it Step 7.} Finally, if everything works out,
we would like to study the properties of ${\mathbf M}_{\mathbf X}$,
$\overline{\mathbf M}_{\mathbf X}$ and to use these to prove further theorems.
\end{say}
Next we review the historical development of the
higher dimensional theory.
\begin{say}[{\bf Geometric Invariant Theory: Mumford, 1965}]
Riemann probably knew that all smooth, compact Riemann surfaces of
a given genus $g$ form a nice family, but
the moduli spaces ${\mathcal M}_g$ were first rigorously constructed by
Teichm\"uller in 1940 as an analytic space and by
Mumford in 1965 as an algebraic variety.
Mumford's book \cite{git} presents a program to construct moduli spaces
under rather general conditions and uses it to obtain ${\mathcal M}_g$.
Using these methods, moduli spaces were
constructed for surfaces (Gieseker, 1977) and
for higher dimensions (Viehweg, 1990).
The correct compactification of these moduli spaces
was much less clear. In principle,
GIT provides an answer, but the resulting compactification
might depend on the embedding dimension
chosen in (\ref{main.steps.moduli}.Step 2).
Recently Wang--Xu \cite{xu-wang}
prove that, for surfaces and in higher dimensions,
the GIT compactification does depend on the embedding dimension.
(The current examples, however, do not exclude the possibility that
some variant of the GIT approach does provide an answer
that is independent of the embedding dimension.)
\end{say}
\begin{say}[{\bf Compact moduli of curves: Deligne and Mumford, 1969}]
\label{comp.dm.say}
The optimal compactification of ${\mathcal M}_g$ is constructed
in \cite{del-mum}. In the boundary
$\overline{\mathcal M}_g\setminus {\mathcal M}_g$
we should allow reducible curves $C=\cup_i C_i$ that satisfy two
restrictions.
({\it Local property}) $C$ has only {\it nodes} as singularities.
In suitable local analytic coordinates these are given
by an equation $(xy=0)\subset \cc^2$.
As in (\ref{jac.sing}) the ``standard'' volume form on a node is given by
$\frac{dx}{x}$ (on the line $(y=0)$) and by
$-\frac{dy}{y}$ (on the line $(x=0)$). These forms
have a simple pole at the singularity, corresponding to the restriction on
log canonical singularities in (\ref{can-lc.defn}).
({\it Global property}) Instead of each $c_1(C_i)$ being negative,
we assume that each $c_1(C_i)-D_i$ is negative where $D_i$ is the sum of the
nodes that lie on $C_i$. (Thus we allow $C_i\cong \cc\pp^1$, as long as at least
3 nodes also lie on $C_i$.)
\end{say}
\begin{say}[{\bf Compact moduli of surfaces: Koll\'ar and Shepherd--Barron, 1988}]
\label{mod.KSB}
It was clear from the Mumford--Gieseker approach that one should work
with the {\it canonical models} of surfaces
of general type (as in \ref{duval.say}) in order to get a good moduli theory,
but the correct class of singular limits was not known.
An approach using minimal models was proposed in
\cite{ksb}: given a family of canonical models
over a punctured disc $S^*\to \Delta^*$,
first construct any compactification whose
central fiber is a reduced simple normal crossing divisor
and then take the (relative) canonical model. It is not hard to see that
this gives a unique limit. This says that at the boundary
of the moduli space
we should allow {\it stable surfaces:}
reducible surfaces $S=\cup_i S_i$ that satisfy two
restrictions.
({\it Local property}) $S$ has so--called {\it semi-log canonical}
singularities. What are these? First of all, aside from
finitely many points $S$ is either smooth or has two local branches meeting
transversally, like $(xy=0)\subset \cc^3$. These are the natural
generalizations of nodes. Then we can have log canonical singularities
(\ref{can-lc.defn}). Finally, it can happen that several
$S_i$ come together at a point and each of them has a
log-canonical singularity there. An explicit list is given in
\cite{ksb}.
({\it Global property}) Instead of each $c_1(S_i)$ being negative,
we assume that each $c_1(S_i)-D_i$ is negative where $D_i$ is the sum of the
double curves that lie on $S_i$.
Another interesting issue that arises is that not every
deformation of such singular surfaces is allowed.
It turns out that even basic numerical invariants of a surface
can jump if we allow arbitrary deformations. To avoid this,
\cite{ksb} identifies a restricted deformation theory
(called $\qq$G-condition)
that produces the correct boundary.
This answers our second Main Question:
First, the ``simplest'' families of surfaces of general type are
families $f:S_M\to M$ whose fibers are
stable surfaces (and satisfy the $\qq$G-condition).
Second, every family of surfaces of general type
is birational to such a ``simplest'' family, at least after a generically
finite-to-one change of the base $M$.
The projectivity of the resulting moduli spaces was
proved in \cite{k-proj}.
\end{say}
\begin{say}[{\bf Moduli of pairs: Alexeev, Kontsevich, 1994}]
Frequently we are interested in understanding all subvarieties $X$ of a given
variety $Y$. All is well if $X$ is smooth, but it is less clear
how to handle singular subvarieties. Various methods have been proposed,
going back to Cayley in 1860.
Alexeev proposed in \cite{ale-pairs} that instead of working with
very singular subvarieties, one should look at
morphisms $X\to Y$ that mimic (\ref{mod.KSB}); see also
\cite{MR2275608}. Independently,
Kontsevich developed this approach for curves \cite{MR1291244}.
The latter since became a standard tool in quantum cohomology theory.
\end{say}
\begin{say}[{\bf Quotient theorems: Keel, Koll\'ar, Mori, 1997}]
Step 3 of (\ref{main.steps.moduli}) leads to the general problem of
taking the quotient of a variety by a group. In our cases we have the extra
information that every point has a finite stabilizer.
In the sixties
Artin and Seshadri proved several quotient theorems,
especially when all stabilizers are trivial.
The general
results needed for the moduli theory were established in
\cite{k-quot, ke-mo}. This is a quite subtle subject since the
resulting quotients are so called {\it algebraic spaces,} a
concept somewhat more general than varieties (or even schemes).
Using the ideas of \cite{k-proj} one can then show that,
in the cases of interest to us, the
quotients are in fact projective (Fujino, Kov\'acs, M\textsuperscript{c}Kernan).
\end{say}
\begin{say}[{\bf Moduli in higher dimensions}]\label{mod.gen}
The general theory follows the outlines of (\ref{mod.KSB})
with some key differences.
First, when \cite{ksb} was written, minimal models were known
to exist only in dimension 3. The higher dimensional theory
needs several results that were established only recently \cite{MR3032329}.
Second, it turned out to be quite difficult to understand how
the irreducible components of a reducible variety $X=\cup_i X_i$
glue together. For curves, as in (\ref{comp.dm.say})
the well-defined residue of the 1-form $\frac{dx}{x}$
is a key ingredient. The current approach in higher dimension
relies on a new Poincar\'e--type residue theory for log canonical singularities;
see \cite[Chap.4]{kk-singbook}.
The full theory should be written up in \cite{k-modbook}.
\end{say}
\begin{say}[{\bf Explicit examples: Alexeev, 2002--}]
While the above methods provide a complete answer in principle,
it has been very difficult to work out a full description in
concrete cases. The first such examples were Abelian varieties
\cite{ale-abvar} and plane curves (Hacking, \cite{hacking}). Recent
surface examples are in \cite{MR2956036}.
\end{say}
\begin{say}[{\bf Hyperbolicity: Kov\'acs, Viehweg, Zuo, 2000--2010}]
So far, very little is known about
the moduli spaces of surfaces and higher dimensional varieties
in general. The local structure of these spaces can be arbitrarily
complicated \cite{MR2227692}.
Hyperbolicity properties of the moduli of smooth curves
were conjectured by Shafarevich in 1962 and later
extended to higher dimensions in
\cite{MR1713524, MR1978884, MR2393082, MR2726098}.
\end{say}
\begin{say}[{\bf Degeneration of Fano varieties: Xu, 2007--}]
We know much less about the moduli of Fano (=positively curved) varieties.
Most of the geometric works deal with extending families
$g^*: X^*\to \Delta^*$ over a punctured disc across the puncture.
Two questions turned out to be
especially interesting:
understanding the combinatorial structure of
the central fiber $X_0$ for arbitrary limits
and
finding limits where $X_0$ is especially simple.
A series of papers \cite{k-ax, hog-xu, dkx} shows that
the combinatorial structure of $X_0$ is contractible;
this answers an old conjecture of J.~Ax.
Recently, \cite{xu-li} shows that there are limits where $X_0$ itself is
a (singular) Fano variety, as conjectured by Tian.
\end{say}
\begin{ack} I thank J.~Fickenscher, A.~Fulger, J.M.~Johnson, S.~Kov\'acs,
T.~Murayama, Zs.~Patakfalvi and
N.~Sheridan
for helpful suggestions.
Partial financial support was provided by the NSF under grant number
DMS-0968337.
\end{ack}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def\dbar{\leavevmode\hbox to
0pt{\hskip.2ex \accent"16\hss}d} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$''${$''$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
An $r$-uniform hypergraph $H$ ($r$-graph for short) is a collection of $r$-element subsets of a vertex set $V(H)$.
Given $r$-graphs $G$ and $H$, the ramsey number $r(G, H)$ is the minimum $n$ such that every red/blue-edge coloring of the complete $r$-graph $K_n^{(r)}:={[n] \choose r}$ contains a red copy of $G$ or a blue copy of $H$ (often we will write $K_n$ for $K_n^{(r)}$).
Determining these numbers for graphs ($r=2$) is known to be notoriously difficult, indeed the order of magnitude (for fixed $t$) of $r(K_t, K_s)$ is wide open when $t \ge 4$. The case $t=3$ is one of the celebrated results in graph Ramsey theory:
\begin{equation} \label{r3s} r(K_3,K_s)=\Theta(s^2/\log s).\end{equation}
The upper bound was proved by Ajtai-Koml\'os-Szemer\'edi~\cite{AKS} as one of the first applications of the semi-random method in combinatorics (simpler proofs now exist due to Shearer~\cite{S83, S91}). The lower bound, due to Kim~\cite{K}, was also achieved by using the semi-random or nibble method. More recently, the first author~\cite{B} showed that a lower bound for $r(K_3, K_s)$ could also be obtained by the triangle-free process, which is a random greedy algorithm. This settled a question of Spencer on the independence number of the triangle-free process. Still more recently, Bohman-Keevash~\cite{BK2} and Fiz Pontiveros-Griffiths-Morris~\cite{FGM} have analyzed the triangle-free process more carefully and improved the constants obtained so that the gap between the upper and lower bounds for $r(K_3, K_s)$ is now asymptotically a multiplicative factor of 4.
Given the difficulty of these basic questions in graph Ramsey theory, one would expect that the corresponding questions for hypergraphs are hopeless. This is not always the case. Hypergraphs behave quite differently for asymmetric Ramsey problems, for example, there exist $K_4^{(3)}$-free 3-graphs on $n$ points with independence number of order $\log n$, so $r(K_4^{(3)}, K_s^{(3)})$ is exponential in $s$ unlike the graph case.
Consequently, to obtain $r$-graph results parallel to (\ref{r3s}), one must consider problems $r(G, K_s)$ where $G$ is much sparser than a complete graph. A recent result in this vein due to Kostochka-Mubayi-Verstra\"ete~\cite{KMV} is that there are positive constants $c_1, c_2$ with
$$\frac{c_1 s^{3/2}}{(\log s)^{3/4}} < r(C_3^{(3)}, K_s^{(3)})< c_2 s^{3/2}$$
where $C_3^{(3)}$ is the loose triangle, comprising 3 edges that have pairwise intersections of size one and have no point in common. The authors in \cite{KMV} conjectured that $r(C_3^{(3)}, K_s^{(3)})=o(s^{3/2})$ and the order of magnitude remains open. Another result of this type for hypergraphs due to Phelps and R\"odl~\cite{PR} is that $r(P_2^{(3)}, K_s^{(3)})=\Theta(s^2/\log s)$, where $P_t^{(3)}$ is the tight path with $t$ edges. Recently, the second author and Cooper~\cite{CM} prove that for fixed $t \ge 4$, the behavior of this Ramsey number changes and we have
$r(P_t^{(3)}, K_s^{(3)})=\Theta(s^2)$; the growth rate for $t=3$ remains open. These are the only nontrivial hypergraph results of polynomial Ramsey numbers, and in this paper we add to this list with an extension of (\ref{r3s}).
\begin{definition} An $r$-uniform triangle $T^{(r)}$ is a set of
$r+1$ edges $b_1, \ldots, b_r, a$ with $b_i \cap b_j=R$ for all $i<j$ where $|R|=r-1$ and $a = \cup_i (b_i-R)$. In other words, $r$ of the edges share a common $(r-1)$-set of vertices, and the last edge contains the remaining point in all these previous edges.
\end{definition}
When $r=2$, then $T^{(2)}=K_3$, so in this sense $T^{(r)}$ is a generalization of a graph triangle. We may view a $T^{(r)}$-free $r$-graph
as one in which all neighborhoods are independent sets, where the neighborhood of an $R \in {V(H)\choose r-1}$ is
$\{x: R \cup \{x\} \in H\}$. Frieze and the first two authors~\cite{BFM} proved that for fixed $r \ge 2$, there are positive constants $c_1$ and $c_2$ with
$$c_1\frac{s^r}{(\log s)^{r/(r-1)}}<r(T^{(r)}, K_s^{(r)})< c_2 s^r.$$
They conjectured that the upper bound could be improved to $o(s^r)$ and believed that the log factor in the lower bound could also be improved. Kostochka-Mubayi-Verstra\"ete~\cite{KMV} partially achieved this by improving the upper bound to
$$r(T^{(r)}, K_s^{(r)}) = O(s^r/\log r)$$
and believed that the log factor was optimal.
In this paper we verify this assertion by analyzing the $T^{(r)}$-free (hyper)graph process.
This process begins with an empty hypergraph $G(0)$ on $n$ vertices. Given $G(i-1)$, the hypergraph $G(i)$ is then formed by adding an edge $e_i$ selected uniformly at random from the $r$-sets of vertices which neither form edges of $G(i-1)$ nor create a copy of $T^{(r)}$ in the hypergraph $G(i-1) + e_i$. The process terminates with a maximal $T^{(r)}$-free graph $G(M)$ with a random number $M$ of edges. Our main result is the following:
\begin{theorem} \label{main}
For $r \ge 3$ fixed
the $T^{(r)}$-free process on $n$ points produces an $r$-graph with independence
number $ O\left( ( n \log n)^{1/r} \right) $ with high probability.
\end{theorem}
\noindent
This result together with the aformentioned result of
Kostochka-Mubayi-Verstra\"ete give the following generalization of (\ref{r3s})
to hypergraphs.
\begin{corollary}
For fixed $r \ge 3$ there are positive constants $c_1$ and $c_2$ with
$$c_1\frac{s^r}{\log s}<r(T^{(r)}, K_s^{(r)})< c_2 \frac{s^r}{\log s}.$$\end{corollary}
Graph processes that iteratively add edges chosen uniformly at random subject to the
condition that some graph property is maintained have been used to generate interesting
combinatorial objects in a number of contexts. In addition to the lower bound on
the Ramsey number $ r(K_3, K_s) $ given by the triangle-free graph process
(discussed above), the $H$-free graph process gives the best known lower bound
on the Ramsey number $ r(K_t,K_s) $ for $ t \ge 4$ fixed and the best known lower bound
on the Tur\'an numbers for some bipartite graphs \cite{BK}. The process that forms
a subset of $ {\mathbb Z}_n$ by iteratively choosing elements to be members of the set
uniformly at random subject to the condition that the set does not contains a $k$-term
arithmetic progression produces a set that has interesting properties with respect to
the Gowers norm \cite{BB}.
The $T^{(r)}$-free (hyper)graph process can be viewed as an instance of the random greedy hypergraph
independent set process. Let $ H$ be a hypergraph. An independent
set in $ H $ is a set of vertices that contains no edge of $H$. The random
greedy independent set process forms such a set by starting with an empty set of vertices and
iteratively choosing vertices uniformly
at random subject to the condition that the set of chosen vertices continues to be
an independent set. We study the random greedy
independent set process for the hypergraph ${\mathcal H}_{T^{(r)}}$ which has vertex set $ \binom{[n]}{r} $
and edge set consisting of all copies of $ T^{(r)}$ on vertex set $[n]$. Note that, since an independent
set in ${\mathcal H}_{T^{(r)}}$ gives a $T^{(r)}$-free $r$-graph on point set $[n]$, the random
greedy independent set process on ${\mathcal H}_{T^{(r)}}$ is equivalent to the $T^{(r)}$-free process.
Our analysis of the
$T^{(r)}$-free process is based on recent work on the random greedy hypergraph independent set process
due to Bennett and Bohman \cite{BB}.
The remainder of the paper is organized as follows. In the following Section we establish some
notation and recall the necessary facts from \cite{BB}. The proof of Theorem~\ref{main} is given
in the Section that follows, modulo the proofs of some technical lemmas. These lemmas are proved
in the final Section by application of the differential equations method for proving dynamic
concentration.
\section{Preliminaries }
Let $\ca{H}$ be a hypergraph on vertex set $V=V(\ca{H})$. For each set of vertices $A \subseteq V$, let $N_{\ca{H}}(A)$ denote the {\bf neighborhood} of $A$ in $\ca{H}$, the family of all sets $Y \subseteq V\setminus A$ for which $A \cup Y \in \ca{H}$. We then define the {\bf degree} of $A$ in $\ca{H}$ to be $d_{\ca{H}}(A)=|N_{\ca{H}}(A)|$. For a nonnegative integer $a$, we define $\Delta_a(\ca{H})$ to be the maximum of $d_{\ca{H}}(A)$ over all $A \in \bin{V}{a}$. Next, for a pair of (not necessarily disjoint) sets $A,B \subseteq V$, we define the {\bf codegree} of $A$ and $B$ to be the number of sets $X \subseteq V \setminus (A \cup B)$ for which $A \cup X, B \cup X$ both lie in $\ca{H}$.
Recall that we define $G(i)$ to be the $r$-graph produced through $i$ steps of the $T^{(r)}$-free process.
We let $\ca{F}_i$ denote the natural filtration determined by the process (see \cite{B}, for example). We also simplify our notation somewhat and write $N_i(A)$ in place of $N_{G(i)}(A)$, $d_i(A)$ in place of $d_{G(i)}(A)$, etc., when appropriate.
The $r$-graph $G(i)$ partitions $\bin{[n]}{r}$ into three sets $E(i),O(i),C(i)$. The set $E(i)$ is simply the set of $i$ edges chosen in the first $i$ steps of the process. The set $O(i)$ consists of the {\bf open} $r$-sets: all $e \in \bin{n}{r} \setminus E(i)$ for which $G(i)+e$ is $T^{(r)}$-free. The $r$-sets in $C(i) := \bin{[n]}{r}\setminus (E(i) \cup O(i))$ are {\bf closed}. Finally, for each open $r$-set $e \in O(i)$, we define the set $C_e(i)$ to consist of all open $r$-sets $f \in O(i)$ such that the graph $G(i)+e+f$ contains a copy of $T^{(r)}$ using both $e$ and $f$ as edges. (That is, $C_e(i)$ consists of the open $r$-sets whose selection as the next edge $e_{i+1}$ would result in $e \in C(i+1)$.)
We now introduce some notation in preparation for our application of the results in \cite{BB}.
Set
\begin{align*}
N &:= \bin{n}{r} & D &:= (r+1)\cdot \bin{n-r}{r-1} & s &:= \frac{N}{D^{1/r}}.
\end{align*}
Note that $N$ is the size of the vertex set of the hypergraph ${\mathcal H}_{T^{(r)}}$ and $D$ is the vertex degree
of ${\mathcal H}_{T^{(r)}}$ (in other words, every $r$-set in $[n]$ is in $D$ copies of $ T^{(r)}$). The
parameter $s$ is the `scaling' for the length of the process. This choice is motivated by the heuristic that
$E(i)$ should be pseudorandom; that is, $E(i)$ should resemble in some ways a collection of $r$-sets chosen uniformly at random (without any further condition). If this is indeed the case
then the probability that a given $r$-set is open would be roughly
\[ \left( 1 - \left(\frac{i}{N}\right)^r \right)^D \approx \exp \left\{ - \left(\frac{i}{N}\right)^r D \right\} \]
and a substantial number of $r$-sets are closed when roughly $s$ edges have been added. In order to
discuss the evolution in more detail, we pass to a limit by introducing a
continuous time variable $t$ where $t=t(i) = i/s$.
The evolution of key parameters of the process closely
follow trajectories given by the functions
\[ q(t):= \exp\left\{ -t^r \right\} \; \; \; \text{ and } \; \; \; c(t):= -q'(t) = rt^{r-1}q(t). \]
We introduce small constants $ \zeta, \gamma $ such that $ \zeta \ll \gamma \ll 1/r$.
(The notation $ \alpha \ll \beta $ here means that $ \alpha $ is chosen to be sufficiently small relative to $ \beta $.)
The point where we stop tracking the process is given by
\[\im :=\zeta \cdot ND^{-1/r}(\log^{1/r} N) \;\;\;\mbox{ and } \;\;\; \tm :=\im/s = \zeta \log^{1/r} N.\]
For $i^* \ge 0$, let $\ca{T}_{i^*}$ denote the event that the following estimates hold for all steps $0 \le i \le i^*$:
\begin{equation}
\label{eq:open}
|O(i)| = \lp q(t) \pm N^{-\gamma} \rp N
\end{equation}
and for every open $r$-set $e \in O(i)$
\begin{equation}
\label{eq:degree}
|C_e(i)| = \lp c(t) \pm N^{-\gamma}\rp D^{1/r}.
\end{equation}
It follows from the results of Bohman and Bennett that $ \ca{T}_{\im}$ holds
with high probability.
\begin{quote}
\begin{proof}
This follows from the estimates for the random greedy hypergraph independent set process given in \cite{BB} applied to the $(r+1)$-uniform hypergraph $ \ca{ H}_{T^{(r)}}$. Verification of the conditions of Theorem 1.1 in \cite{BB} for this hypergraph is routine. (Note that $ \Delta_\ell (\ca{H}_{T^{(r)}}) = \Theta( n^{r - \ell})$ and $ \Gamma_r( \ca{H}_{T^{(r)}}) = 0$.) The estimates (\ref{eq:open}) and (\ref{eq:degree}) above then follow from those on $|V(i)|$ and $d_2(v,i)$
given by (5) and (6) in \cite{BB}.
\end{proof}
\end{quote}
\noindent
Note that the fact that $ \ca{T}_{\im} $ holds with high probability does not
prove that the independence number of $ G(M)$ is $ O \lp ( n \log n)^{1/r} \rp$ with high probability.
This is proved below.
We will also make use of the following fact regarding
$r$-graphs that appear as subgraphs of the $T^{(r)}$-free process.
\begin{lemma}[\cite{BB} Lemma 4.2]
\label{lem:subgraph}
Fix a constant $L$ and suppose $e_1,\ldots,e_L \in \bin{[n]}{r}$ form a $T^{(r)}$-free hypergraph. Then for all steps $j \le \im$,
\[\pr{\{e_1,\ldots,e_L\} \subseteq E(j)} = (j/N)^L\cdot (1 + o(1)).\]
\end{lemma}
We conclude this Section by noting that the desired bound on the independence number
of $G(M)$ can be viewed as a pseudorandom property of the $r$-graph $ G(i) $. Indeed, if $ G(i)$
resembles a collection of $r$-sets chosen uniformly at random then the expected number of
independent sets of size $k$ would be
\[ \binom{n}{k} \left( 1 - \frac{ \binom{k}{r}}{ \binom{n}{r}} \right)^i
= \exp \left\{ \Theta \left( k \log n \right) - \Theta \left( i \frac{k^r}{n^r} \right) \right\}. \]
If the process lasts through $ i = \Theta( N D^{-1/r} ( \log^{1/r} N) ) = \Theta( n^{r -1 + 1/r} \log^{1/r} n) $
steps then we would anticipate an independence number of $ O\left( (n \log n)^{1/r} \right) $. In the remainder of the paper
we make this heuristic calculation rigorous.
\section{Independence number: Proof of Theorem~\ref{main}}
We expand the list of constants given in the previous section by introducing
large constants $\kappa$ and $W$, and small constant $\epsilon $ such that
\begin{equation}
\label{eq:constants}
\frac{1}{\kappa} \ll \zeta \ll \frac{1}{W} \ll \ep \ll \gamma.
\end{equation}
In the course of the argument we introduce dynamic concentration phenomena
that will stated in terms of the error function
\[f(t) := \exp \left\{ W(t^{r} + t)\right\}.\]
Define the constant $\lambda:= \frac{\kappa - \gamma}{2}$, and then let
\begin{align*}
k &:= \kappa (n \log n)^{1/r} & \mbox{ and } & & \ell &:= \lambda (n \log n)^{1/r},
\end{align*}
noting that as $\gamma$ is small, $k \approx 2\l$.
Our aim is to show that the independence number of $G(\im)$ is at most $k$ with high probability. To do so, we will show that provided $\kappa$ is suitably large, w.h.p. for every step $0 \le i \le \im$, every $k$-element set of vertices has at least $ \Omega \lp q(t) \bin{k}{r} \rp$ open $r$-sets. As equation (\ref{eq:open}) establishes $ (1+o(1)) q(t) N$ open $r$-sets in total w.h.p., the probability that
$\ca{T}_{\im}$ holds and a given $k$-set remains independent over all $\im$ steps is then at most
\[ \prod_{i=1}^{\im} \lp 1 - \Omega \lp \frac{q(t)k^r}{q(t)N} \rp \rp = \lp 1 - \Omega \lp \frac{\kappa^r \log n}{n^{r-1}} \rp \rp^{\im} = \exp\left\{-\zeta\kappa^r\cdot \Omega(n^{1/r} \log^{1+1/r} n) \right\},\]
where our $ O( \cdot), \Omega( \cdot), \Theta( \cdot ) $ notation does not
suppress any constant that appears in (\ref{eq:constants}).
Since $$n^k =\exp \left\{ \kappa \cdot O (n^{1/r} \log^{1+1/r} n) \right\},$$ this suffices by the union bound, provided $\kappa$ is suitably large
with respect to $r$ and $ \zeta $.
There is a significant obstacle to proving that every set of $k$ vertices contains the `right' number
of open $r$-sets. Note that all $r$-sets within the neighborhood of an $(r-1)$-set are closed.
(To be precise, if $ A \in \binom{[n]}{r-1} $ then $ \binom{N_i(A)}{r} \subseteq C(i) $). So a
set of $k$ vertices that has a large intersection with the neighborhood of an
$(r-1)$-set does not have the `right' number of open $r$-sets.
To overcome this obstacle, we extend the
argument in \cite{B} for bounding the independence number of the
triangle-free process. Our argument has two steps:
\begin{enumerate}
\item We apply the differential equations method for establishing dynamic concentration to show that
unless a certain `bad' condition occurs, a pair of disjoint $\l$-sets will have the `right' number of
open $r$-sets that are contained in the union of the pair of $\l$-sets and intersect both $\l$-sets, that is about $q(t)\cdot [\bin{2\l}{r}-2\bin{\l}{r}]$ open $r$-sets.
Note that $\bin{2\l}{r}-2\bin{\l}{r} > \frac{1}{3}\bin{k}{r}$, say, as $\gamma$ is small.
\item We then argue that w.h.p., every $k$-set contains a (disjoint) pair of $\l$-sets which is `good', i.e., for which the bad condition does not occur.
\end{enumerate}
We formalize this with the notion of $r$-sets which are open `with respect to' a pair of disjoint $\l$-sets.
\begin{definition}
Fix a disjoint pair $A,B \in \bin{[n]}{\l}$. The stopping time $\tau_{A,B}$ is the minimum of $\im$ and the first step $i$ for which there exists a $(r-1)$-set $X$ such that
$$ N_i(X)\cap A \ne \emptyset, \ \ \ \ N_i(X)\cap B \ne \emptyset, \ \ \ \text{ and } \ \ \ |N_i(X)\cap (A \cup B)| \ge k/n^{2\ep}.$$
\end{definition}
\begin{definition}
For each step $i \ge 0$, we say that an $r$-set $e \subseteq A \cup B$ is {\bf open with respect to the pair $A,B$ in $G(i)$} if $e \cap A \ne \emptyset$, $e \cap B \ne \emptyset$, and either
$\bullet$ $e \in O(i)$ {\em or}
$\bullet$ $e \in O(i-1) \cap C(i)$ and $i=\tau_{A,B}$.
\noindent Let $Q_{A,B}(i)$ count the number of $r$-sets which are open with respect to the pair $A,B$ in $G(i)$. \end{definition}
\begin{lemma}\label{lem:relopentraj}
With high probability, for every disjoint pair $A,B \in \bin{[n]}{\l}$ and all steps $0 \le i \le \tau_{A,B}$,
\begin{equation}\label{eq:relopenpairs}
Q_{A,B}(i) = \lp q(t) \pm \frac{f(t)}{n^{\ep}}\rp \cdot \left[\bin{2\l}{r}-2\bin{\l}{r}\right].
\end{equation}
\end{lemma}
\begin{lemma}\label{lem:largeopensets}
With high probability, for every step $0 \le i < \im$ and every set $K \in \bin{[n]}{k}$, there exists a pair of disjoint $\l$-sets $A,B$ contained in $K$ for which $\tau_{A,B} > i$.
\end{lemma}
\noindent
Lemmas \ref{lem:relopentraj} and \ref{lem:largeopensets}, respectively, complete steps
1 and 2 of the proof outlined above. The `bad' condition for a pair $A,B$ of disjoint
$\ell$-sets is the event
that we have reached the stopping time $ \tau_{A,B}$; that is, the bad condition is that there
is some $ (r-1)$-set whose neighborhood intersects both $A$ and $B$ and has large intersection
with $A \cup B$. Note that if $ i < \tau_{A,B} $ then $ Q_{A,B}$ is equal to the number of
open $r$-sets that are contained in $ A \cup B$ and intersect both $A$ and $B$. Thus,
Lemma~\ref{lem:relopentraj} says that if we do not have the `bad' condition then we have the `right' number of
such sets. Lemma~\ref{lem:largeopensets} then says that every $k$-set contains a pair disjoint pair $ A,B$ of
$ \ell$-sets for which the `bad' condition does not hold. Taken together,
Lemmas~\ref{lem:relopentraj}~and~\ref{lem:largeopensets} yield that w.h.p., for every
step $0 \le i < \im$, every $k$-set contains at least
$q(t)(1 + o(1))[\bin{2\l}{r}-2\bin{\l}{r}] = \Omega \lp q(t) \binom{k}{r} \rp$ open $r$-sets, as required.
We now prove Lemma~\ref{lem:largeopensets} modulo the proof of Lemma~\ref{lem:maxdeg} which bounds the
maximum degree of an $(r-1)$-set. Lemmas~\ref{lem:relopentraj}~and~\ref{lem:maxdeg} are proved in the
next Section.
\begin{proof}[Proof of Lemma~\ref{lem:largeopensets}]
We require a bound on the maximum degree of $(r-1)$-sets of vertices. For each step $i \ge 0$ let $\ca{D}_{i}$ denote the event that $\Delta_{r-1}(G(i)) \le \ep (n \log n)^{1/(r-1)}$.
\begin{lemma}\label{lem:maxdeg}
$\ca{T}_{\im} \land \ca{D}_{\im}$ holds with high probability.
\end{lemma}
\noindent
The proof of Lemma~\ref{lem:maxdeg} is given in the next Section.
Fix a step $0 \le i < \im$, and a set $K \in \bin{[n]}{k}$. Note that, by Lemma~\ref{lem:maxdeg},
we may assume that $\ca{D}_i$ holds. We also note that the maximum co-degree of a pair of sets
$ A,B \in \binom{[n]}{r-1} $ is at most $ 5r$ with high probability. This follows from Lemma~\ref{lem:subgraph}
and the union bound:
\begin{multline}
\label{eq:codegree}
\Pr \lp \exists A,B \in \binom{[n]}{r-1} \text{ with co-degree } 5r \rp \le \binom{n}{r-1} \binom{n}{r-1} n^{5r} \left(\frac{i}{N}\right) ^{10r} \\ = n^{8-3r +o(1)} = o(1).
\end{multline}
Given these two facts (i.e. these degree and co-degree bounds for $ (r-1)$-sets),
the remainder of the proof is deterministic.
To begin, define the set
\[\ca{X}:=\left\{X \in \bin{[n]}{r-1} : |N_i(X)\cap K| \ge k/n^{2\ep}\right\}.\]
\begin{claim}
$|\ca{X}| < 2 n^{2\ep}$.
\end{claim}
\begin{quote}
\begin{proof}
Suppose $\exists \ca{Y} \subseteq \ca{X}$ with $|\ca{Y}| = 2n^{2\ep}$. Let $N = \bigcup_{Y \in \ca{Y}} (N_i(Y) \cap K)$. By inclusion-exclusion,
\[k \ge |N| \ge |\ca{Y}|\cdot (k /n^{2\ep}) - |\ca{Y}|^2 5r \ge 2k-20rn^{4\ep},\]
a contradiction as $\ep$ is small and $k = n^{1/r + o(1)}$.
\end{proof}
\end{quote}
Next, we `discard' from $K$ the vertices which are common neighbors of $(r-1)$-sets in $\ca{X}$: let
$$K_{bad}:=\{v \in K : \exists X,Y \in \ca{X} \mbox{ with } X \ne Y \mbox { and } v \in N_i(X)\cap N_i(Y)\}$$ and $K_{good}:=K\setminus K_{bad}$. Then
\[|K_{bad}| \le |\ca{X}|^2 5r\le 20rn^{4\ep} < \frac{\gamma}{2}\cdot (n \log n)^{1/r},\]
say, for large $n$.
We find disjoint $\l$-subsets $A,B$ of $K_{good}$ as follows, noting $|K_{good}| \ge 2\ell + (\gamma/2)(n \log n)^{1/r}$. For each subset $\ca{Y}\subseteq \ca{X}$, let
\[N({\ca{Y}}) = \bigcup_{Y \in \ca{Y}} N_i(Y) \cap K_{good}.\]
Now, choose a maximal subset $\ca{X}^* \subseteq \ca{X}$ subject to $|N(\ca{X}^*)| \le \l$. If $\ca{X}^*=\ca{X}$, then let $A,B$ be $\l$-sets satisfying $N(\ca{X}^*)\subseteq A \subseteq K_{good}$ and $B \subseteq K_{good}\setminus A$.
Otherwise, pick any set $X^*\in \ca{X}\setminus \ca{X}^*$, so
\[\ell < |N(\ca{X}^* \cup \{X^*\})| < \ell + \ep (n \log n)^{1/r};\]
let $A \subseteq N(\ca{X}^* \cup \{X^*\})$ and $B \subseteq K_{good}\setminus N(\ca{X}^* \cup \{X^*\})$ be $\ell$-sets.
Observe now that if $\ca{X}^*=\ca{X}$, then $N_i(X)\cap B=\emptyset$ for all $X \in \ca{X}$. Otherwise, if $X \in \ca{X}^*\cup \{X^*\}$, $N_i(X)\cap B=\emptyset$, but if $X \in \ca{X}\setminus (\ca{X}^*\cup \{X^*\})$ then $N_i(X)\cap A=\emptyset$ as we are working within $K_{good}$. In either case, for every $(r-1)$-set $X$ for which $|N_i(X) \cap (A \cup B)| \ge k/n^{2\ep}$ holds, either $N_i(X)\cap A = \emptyset$ or $N_i(X)\cap B = \emptyset$, and $\tau_{A,B} > i$ follows.
\end{proof}
\section{Dynamic Concentration}
In this section we prove Lemmas~\ref{lem:relopentraj}~and~\ref{lem:maxdeg}. Both of
these statements assert dynamic concentration of key parameters of the $T^{(r)}$-free process. We apply
the differential equations method for proving dynamic concentration, which we now briefly sketch.
Suppose we have a combinatorial stochastic process based on a ground set of size $n$ that generates
a natural filtration $ \ca{F}_0, \ca{F}_1, \dots $. Suppose further that we
have a sequence of random variables $ A_0, A_1, \dots$ and that we would like to prove a dynamic concentration
statment of the form
\begin{equation}
\label{eq:dynamic}
A_i \le T_i + E_i \ \ \text{ for all } \ \ 0 \le i \le m(n) \ \ \text{ with high probability},
\end{equation}
where $ T_0, T_1, \dots $ is the expected trajectory of the sequence of random variables $A_i$
and $ E_0, E_1, \dots $ is a sequence of error functions. (One is often interested in proving a lower
bound on $A_i$ in conjunction with (\ref{eq:dynamic}). The argument for proving this is essentially
the same as the upper bound argument
that we discuss here.) We often make this statement in the context
of a limit that we define in terms of a continuous time $t$ given by $ t = i/s$ where $ s$ is the
{\bf time scaling} of the process. The limit of the expected trajectory is determined
by setting $ T_i = f(t) S(n) $ where $ S = S(n)$ is
the {\bf order scaling} of the random variable $A_i$. Given these assumptions we should have
\[ \ev{A_{i+1} - A_i \mid \ca{F}_i } = T_{i+1} - T_i = \left[ f(t+1/s) - f(t) \right] S \approx f'(t) \cdot \frac{S}{s}.\]
Thus the trajectory is determined by the expected one-step change in $ A_i$.
We prove (\ref{eq:dynamic}) by applying facts regarding the probability of large deviations in
martingales with bounded differences. In particular, we consider the sequence
\[ D_i = A_i - T_i - E_i. \]
Note that if we set $ T_0 = A_0$ (which is often the natural initial condition) then
$ D_0 = - E_0 $. If we can establish that the sequence $ D_i$ is a supermartingale
and $ E_0$ is sufficiently large then it should be unlikely that $ D_i$ is ever positive,
and (\ref{eq:dynamic}) follows. In order to complete such a proof we
show that the sequence $D_i$ is a supermartingale, a fact that is sometimes called the {\bf trend
hypothesis} (see Wormald \cite{W}). The trend hypothesis will often impose a condition that the sequence of
error functions $ E_i$ is growing sufficiently quickly (i.e. the derivative of the limit of error
function is sufficiently large). We then show that the one-step changes in $ D_i$ are bounded in
some way (this is sometimes called the {\bf boundedness hypothesis}). This puts us in the
position to apply a martingale inequality. In order to get good bounds from the martingale inequality one
generally needs to make $ E_0$ large.
In this section we appeal to the following pair of martingale inequalities (see \cite{B}).
For positive reals $b,B$, the sequence $A_0,A_1,\ldots$ is said to be {\bf $(b,B)$-bounded} if $A_i - b \le A_{i+1} \le A_i + B$ for all $i \ge 0$.
\begin{lemma}\label{lem:submartinbd}
Suppose $b \le B/10$ and $0<a<bm$. If $A_0,A_1,\ldots$ is a $(b,B)$-bounded submartingale, then $\pr{A_m \le A_0 - a} \le \exp \left\{ -a^2/3b m B \right\}$.
\end{lemma}
\begin{lemma}\label{lem:supmartinbd}
Suppose $b \le B/10$ and $0<a<bm$. If $A_0,A_1,\ldots$ is a $(b,B)$-bounded supermartingale, then $\pr{A_m \ge A_0 + a} \le \exp \left\{ -a^2/3b m B \right\}$.
\end{lemma}
\noindent
Our applications of these Lemmas make use of stopping times. Formally speaking, a stopping time
is simply a postive integer-valued random variable $ \tau$ for
which $ \{ \tau \le n\} \in {\mathcal F}_n $. In other words,
$\tau$ is a stopping time if the event $ \tau \le n$ is determined by the first $n$ steps of the
process. We consider the stopped process $ ( D_{ i \wedge \tau }) $, where
$ x \wedge y := \min \{x,y\} $, in the place of the sequence $ D_0, D_1, \dots$. Our stopping time $ \tau$
is the first step in the process when any condition on some short list of conditions
fails to hold, where the condition
$ D_i \le 0 $ is one of the conditions in the list. Note that, since the variable $ ( D_{ i \wedge \tau }) $ does
not change once we reach the stopping time $ \tau$, we can assume that all conditions in the list
hold when we are proving the trend and boundedness hypotheses. Also note that if the stopping time
$\tau'$ is simply the minimum of $ \im$ and the first step for which $ D_i > 0$ then
$ \{ D_{\im \wedge \tau'} > 0 \} $ contains the event $ \{ \exists i \le \im: D_i > 0 \}$.
\subsection{Proof of Lemma \ref{lem:maxdeg}.}
For each set $A \in \bin{[n]}{r-1}$ and step $i \ge 0$, let $O_A(i):=\{e \in O(i) : A \subseteq e\}$, and
$Q_A(i) = |O_A(i)|$. We define sequences of random variables
\begin{align*}
Y^+_A(i) &:= q(t)\cdot n - Q_A(i) + f(t)\cdot n^{1-\ep},\\
Y^-_A(i) &:= q(t)\cdot n - Q_A(i) - f(t)\cdot n^{1-\ep},\\
Z_A(i) &:= d_i(A) - t\cdot D^{-1/r} n - f(t)q(t)^{-1}\cdot n^{1/r-\ep},
\end{align*}
Finally, we define the stopping time $\tau$ to be the minimum of $\bin{n}{r}$, the first step $i$ where $\ca{T}_i$ fails, or where any of $Y^+_A(i) < 0$, $Y^-_A(i) > 0$, or $Z_A(i) > 0$ holds for some $A \in \bin{[n]}{r-1}$.
To prove Lemma~\ref{lem:maxdeg}, we show that for each $A \in \bin{[n]}{r-1}$,
\begin{align}
\label{eq:YAplusbd} \pr{Y^+_A(\im \land \tau) < 0} &= o(n^{-(r-1)}),\\
\label{eq:YAminusbd} \pr{Y^-_A(\im \land \tau) > 0} &=o(n^{-(r-1)}), \mbox{ and }\\
\label{eq:ZAbd} \pr{Z_A(\im \land \tau) > 0} &=o(n^{-(r-1)}).
\end{align}
Consider the event
$ \tau \le \im $. This event is the union of the event that $ \ca{ T}_{\im} $ fails and
the event that
there exists $ A \in \bin{[n]}{r-1}$ such that $ Y^+_A(\im \land \tau) < 0 $ or
$ Y^-_A(\im \land \tau) > 0 $ or $ Z_A(\im \land \tau) > 0 $.
Since $\ca{T}_{\im}$ holds with high probability, it follows from (\ref{eq:YAplusbd})--(\ref{eq:ZAbd}) and
the union bound that w.h.p. $\tau > \im$. In particular, $Z_A(i) \le 0$ for
all $(r-1)$-sets $A$ and steps $0 \le i \le \im$.
It then follows -- since $\zeta \ll \min\{1/W,\ep \}$ implies that we may bound $f(\tm) < n^{\ep/2}$, say -- that we have
\[\Delta_{r-1}(G(\im)) \le \tm D^{-1/r}n + f(\tm) n^{1/r-\ep/2} = \zeta \cdot O( (n \log n)^{1/r}) \le \ep (n \log n)^{1/r},\]
for $n$ sufficiently large. (We remark in passing that
the bounds on $Y_A^{\pm}(i)$ given when $i<\tau$ are necessary for our proof of the bounds on $Z_A(i)$.)
For the remainder of this argument, fix a set $A \in \bin{[n]}{r-1}$. We first prove \eqref{eq:YAplusbd} and \eqref{eq:YAminusbd}.
\begin{claim} \label{clm:YApmbds}
For $n$ sufficiently large, the variables $Y^+_A(0),\ldots,Y^+_A(\im \land \tau)$ form an $(O(n/s),O(n^{1 - \frac{1}{2r}}))$-bounded submartingale, and the variables $Y^-_A(0),\ldots,Y^-_A(\im \land \tau)$ form an $(O(n/s),O(n^{1 - \frac{1}{2r}}))$-bounded supermartingale.
\end{claim}
\begin{proof}
We begin by fixing a step $0 \le i \le \im$, and we assume that $i < \tau$. Throughout we write $t=t(i)$, and note $t(i+1)=t+s^{-1}$ and that $s^{-1}=D^{1/r}/N=\Theta(n^{1-1/r-r})$.
To aid the calculations to follow, we begin by estimating the quantity $$\Xi := f(t+s^{-1}) - f(t).$$ Since $f(t)=\exp(Wt^r + Wt)$, $f'(t)$ and $f''(t)$ are products of $f(t)$ with polynomials in $t$. As $\zeta \ll \max\{1/W,\ep\}$, $\tm$ is polylogarithmic in $n$, and $n$ is large, we have the crude bounds $ f(t) \le n^{\epsilon/2} $ and
$f''(t) \le n^{o(1)} f'(t) $. Thus, by Taylor's Theorem,
\begin{equation}\label{eq:Xiestimate}
\left|\Xi - \frac{f'(t)}{s}\right| = O\lp \frac{ \max_{ t^* \le \tm } f''(t^*) }{s^2}\rp= o\lp \frac{f'(t)}{s}\rp.
\end{equation}
Observe now that we may write
\[Y_A^{\pm}(i+1)-Y_A^{\pm} (i) = (q(t+s^{-1})-q(t))\cdot n - \lp Q_A(i+1)-Q_A(i)\rp \pm \Xi\cdot n^{1-\ep}.\]
(Note that this stands for the pair of equations in which each $\pm$ is replaced with $+$ or with $-$, respectively.)
We begin by establishing the boundedness claims: it is routine to verify that $c(t)$ and $c'(t)$
are bounded over the reals, implying
\begin{equation}\label{eq:qtdifferentialest}
|q(t+s^{-1})-q(t) - c(t)\cdot s^{-1}| = O(s^{-2}),
\end{equation}
and so
\[0 \ge \lp q(t + s^{-1})-q(t)\rp \cdot n \ge -O\lp \frac{n}{s} \rp.\]
As we have the bound $|f'(t)| = n^{\ep/2 + o(1) } $ and (\ref{eq:Xiestimate}), we have
\( |\Xi|\cdot n^{1-\ep}= o( n/s)\),
and the lower bound in the boundedness claims follows. To establish the upper bounds,
it remains to bound $Q_A(i)-Q_A(i+1)$. Consider the `next' edge $e_{i+1} \in O(i)$ and observe that $$Q_A(i)-Q_A(i+1) = |\lp\{e_{i+1}\} \cup C_{e_{i+1}}(i)\rp \cap O_A(i)|.$$ We bound $|C_{e_{i+1}}(i)\cap O_A(i)|$ by considering five cases depending on $|e_{i+1}\cap A|$:\\
\noindent {\bf Case 1: $|e_{i+1} \cap A| = 0$.} Let $f \in O_A(i)\cap C_{e_{i+1}}(i)$: then $f = A \cup \{v\}$ for some vertex $v$, and since $G(i)+e_{i+1}+f$ contains a copy of $T^{(r)}$, $v \in e_{i+1}$ must hold. (Recall that every pair of edges in $T^{(r)}$ either shares exactly one or $r-1$ vertices.) In this case, $|C_{e_{i+1}}(i)\cap O_A(i)| \le |e_{i+1}|=r$.\\
\noindent {\bf Case 2: $|e_{i+1} \cap A| = r-1$.} In this case, we may write $e_{i+1} = A \cup \{u_1\}$. Now, let $f =A \cup \{v\} \in O_A(i)\cap C_{e_{i+1}}(i)$: since $f \cap e_{i+1}=A$ and $f \in C_{e_{i+1}}(i)$, there must exist vertices $u_2,\ldots,u_{r-1} \in N_i(A)$ so that $\{u_1,\ldots,u_{r-1},v\} \in E(i)$. As then $v \in N_i(\{u_1,\ldots,u_{r-1}\})$, we may bound the number of such choices of $v$ (and hence of $f$) in this case above by $\Delta_{r-1}(G(i))^{r-1}\le \zeta^{r-1} (n \log n)^{(r-1)/r}$. (Note the bound on the maximum degree follows as $Z_A(i) \le 0$ since $i < \tau$.)\\
\noindent {\bf Case 3: $|e_{i+1} \cap A| = 1$.} Write $A = \{x_1, \dots, x_{r-1}\}$, where we take $e_{i+1}\cap A = \{x_1\}$. Let $f = A \cup \{v\} \in C_{e_{i+1}}(i) \cap O_A(i)$, and suppose $v \notin e_{i+1}$ (as there are at most $r-1$ such $v$), so $f \cap e_{i+1}=\{x_1\}$. Consider a copy of $T^{(r)}$ in $G(i)+e_{i+1} + f$ using both $e_{i+1}$ and $f$ as edges: without loss of generality, we may assume that one of $e_{i+1},f$ maps to the edge $b_1$ of $T^{(r)}$, the other to the edge $a$.
If $e_{i+1}$ maps to $b_1$, then the $(r-1)$-set $e_{i+1}\setminus \{x_1\}$ maps to the common intersection $B$ of $b_1,\ldots,b_r$. Consequently $v \in N_i(e_{i+1}\setminus \{x_1\})$ must hold, and so there are at most $\Delta_{r-1}(G(i))$ such $r$-sets $f \in C_{e_{i+1}}(i) \cap O_A(i)$.
Otherwise, if $e_{i+1}$ maps to the edge $a$ and $f$ maps to $b_1$, then $\{x_2,\dots, x_{r-1},v\}$ maps to the common intersection $B$. Thus, for each $u \in e_{i+1}\setminus \{x_1\}$ we have $\{u,x_2, \dots, x_{r-1},v\} \in E(i)$, implying $v \in N_i(\{u,x_2, \dots, x_{r-1}\})$ and (as $e_{i+1}$ is fixed), there are again at most $\Delta_{r-1}(G(i))$ such choices of $f$. Thus, in this case we have $|C_{e_{i+1}}(i) \cap O_A(i)| \le 2 + 2\Delta_{r-1}(G(i)) = n^{1/r + o(1)}$.\\
\noindent {\bf Case 4: $1 < |e_{i+1} \cap A| = r-2$.} Let $f = A \cup \{v\} \in O_A(i) \cap C_{e_{i+1}}(i)$. Since $|f \cap e_{i+1}| \ge |A \cap e_{i+1}| > 1$, $|f \cap e_{i+1}| = r-1$ must hold, implying $v \in e_{i+1}$ and so $|O_A(i) \cap C_{e_{i+1}}(i)| \le r$ as in Case 1. \\
\noindent {\bf Case 5: $2 \le |e_{i+1} \cap A| \le r-3$.} In this case, $|C_{e_{i+1}}(i)\cap O_A(i)| = 0$, as every $f \in
O_A(i)$ satisfies $1 \le |f \cap e_{i+1}| \le r-2$.\\[5mm]
\noindent From the cases above it follows that $Q_A(i)-Q_A(i+1) = n^{(r-1)/r + o(1)}$, and combining the above bounds, it follows that the sequences $Y^{\pm}_A(0),\ldots,Y^{\pm}_A(\im \land \tau)$ are $(O(n/s),O(n^{ 1 - \frac{1}{2r}}))$-bounded.
We turn now to the sub- and supermartingale claims: all expectation calculations to follow are implicitly conditioned on the history of the process up to step $i$, and we recall that we assume $i < \tau$. For each open $r$-set $f \in O_A(i)$, we have $f \notin O_A(i+1)$ if and only if $e_{i+1} \in C_f(i) \cup \{f\}$. Thus,
\begin{align*}
\ev{Y^{\pm}_A((i+1)) - Y^{\pm}_A(i)} &= (q(t+s^{-1})-q(t))\cdot n + \sum_{f \in O_A(i)} \frac{|C_f(i)|+1}{|O(i)|} \pm \Xi\cdot n^{1-\ep}.
\end{align*}
To establish the submartingale claim, consider the following chain of inequalities:
\begin{align*}
\sum_{f \in O_A(i)} \frac{|C_f(i)|+1}{|O(i)|} &\ge (q(t)-f(t) n^{-\ep})\cdot n \cdot \frac{(c(t) - N^{-\gamma})\cdot D^{1/r} }{(q(t) + N^{-\gamma})\cdot N}\\
&= \lp 1 - \frac{N^{-\gamma} +f(t)n^{-\ep}}{q(t)+N^{-\gamma}}\rp(c(t)-N^{-\gamma})\cdot \frac{n}{s}.\\
&\ge \lp 1 - 2q(t)^{-1}f(t)n^{-\ep}\rp(c(t)-N^{-\gamma})\cdot \frac{n}{s}\\
&\ge \lp c(t) - 2c(t)q(t)^{-1}f(t)n^{-\ep}-N^{-\gamma}\rp\cdot \frac{n}{s}\\
&\ge \lp c(t) - (2c(t)q(t)^{-1}+1)\cdot f(t)n^{-\ep}\rp\cdot \frac{n}{s}.
\end{align*}
The first inequality follows from the bounds given by (\ref{eq:open}) and (\ref{eq:degree}) on the event $\ca{T}_i$ and as $Y_A^-(i) \le 0$, since $i<\tau$. In the second and fourth inequalities we bounded $N^{-\gamma} < f(t)n^{-\ep}$, valid as $f(t) \ge 1$ and $\ep \ll \gamma$. Thus, applying this bound and \eqref{eq:qtdifferentialest} gives
\begin{align*}
\ev{Y^+_A(i+1) - Y^+_A(i)} &\ge \Xi\cdot n^{1-\ep} - (2c(t)q(t)^{-1}+1)f(t)\frac{n^{1-\ep}}{s}-O\lp \frac{1}{s^2}\rp\\
&\ge \Xi\cdot n^{1-\ep} - (2c(t)q(t)^{-1}+2)f(t)\frac{n^{1-\ep}}{s}\\
&= \lp (1 + o(1))f'(t)- (2c(t)q(t)^{-1}+2)f(t)\rp \cdot\frac{n^{1-\ep}}{s}
\end{align*}
by \eqref{eq:Xiestimate}. Since $f'(t)=(Wrt^{r-1}+W)f(t)$ and $2c(t)q(t)^{-1}= 2rt^{r-1}$, this final bound is nonnegative for large $n$ as $W$ is large, and so $Y_A^+(0),\ldots,Y_A^+(\im\land \tau)$ forms a submartingale.
We similarly bound $\ev{Q_A(i) - Q_A(i+1)}$ above to establish the supermartingale claim: as $1 < N^{-\gamma}D^{1/r}$ for large $n$, and as $\ca{T}_i$ holds and $Y_A^+(i) \ge 0$,
\begin{align*}
\sum_{f \in O_A(i)} \frac{|C_f(i)|+1}{|O(i)|} &\le (q(t)+f(t) n^{-\ep})\cdot n \cdot \frac{(c(t) + 2N^{-\gamma})\cdot D^{1/r} }{(q(t) - N^{-\gamma})\cdot N}\\
&= \lp 1 + \frac{N^{-\gamma} +f(t)n^{-\ep}}{q(t)-N^{-\gamma}}\rp(c(t)+2N^{-\gamma})\cdot \frac{n}{s}\\
&\le \lp 1 + 4q(t)^{-1}f(t)n^{-\ep}\rp(c(t)+2N^{-\gamma})\cdot \frac{n}{s}\\
&\le \lp c(t) + (4c(t)q(t)^{-1}+4)f(t)n^{-\ep}\rp\cdot \frac{n}{s}.
\end{align*}
In addition to the bound $N^{-\gamma} \le f(t)n^{-\ep}$ used above, in the second inequality, we bounded $q(t)-N^{-\gamma} \ge q(t)/2$, and in the final we bounded $2N^{-\gamma}(1+4q(t)^{-1}f(t)n^{-\ep}) \le 4f(t)n^{-\ep}$ as $q(t)^{-1}f(t)n^{-\ep} \le 1$ which holds as $2W\zeta^r<\epsilon$ and $n$ is large.
Thus,
\begin{align*}
\ev{Y^-_A(i+1) - Y^-_A(i)} &\le -\Xi\cdot n^{1-\ep} + (4c(t)q(t)^{-1}+4)f(t)\frac{n^{1-\ep}}{s}+O\lp \frac{1}{s^2}\rp\\
&\le -\Xi\cdot n^{1-\ep} + (4c(t)q(t)^{-1}+5)f(t)\frac{n^{1-\ep}}{s}\\
&= \lp -(1 + o(1))f'(t)+ (4c(t)q(t)^{-1}+5)f(t)\rp \cdot\frac{n^{1-\ep}}{s},
\end{align*}
and again, as $W$ is large, this is strictly negative for $n$ sufficiently large. Thus, the sequence $Y_A^-(0),\ldots,Y_A^-(\im\land \tau)$ forms a supermartingale, completing the proof.
\end{proof}
Since $Q_A(0)=n-r+1$, $Y_A^+(0) = r-1 + n^{1-\ep}$ and $Y_A^-(0) = r-1-n^{1-\ep}$. Applying Lemmas \ref{lem:submartinbd} and \ref{lem:supmartinbd}, respectively, we have
\begin{align*}
\pr{Y^+_A(\im \land \tau) < 0} &\le \exp\left\{ - \Omega \lp \frac{ n^{2-2\ep} }{ \frac{n}{s} \cdot \zeta s\log^{1/r} N \cdot n^{1 - \frac{1}{2r}})} \rp \right\}\\
&= \exp\left\{ - n^{\frac{1}{2r}-2\ep + o(1)} \right\}\\
&< \exp\left\{ - n^{\frac{1}{4r}} \right\}
\end{align*}
(valid for large $n$ as $\ep$ is small), and an identical calculation yields
\begin{align*}
\pr{Y^-_A(\im \land \tau) > 0} &\le \exp\left\{ - n^{ \frac{1}{4r}} \right\}.
\end{align*}
We have established \eqref{eq:YAplusbd} and \eqref{eq:YAminusbd}.
It remains to prove \eqref{eq:ZAbd}.
\begin{claim} \label{clm:degbdmartingale}
The variables $Z_A(0),\ldots,Z_A(\im \land \tau)$ form a $(2n/N,2)$-bounded supermartingale.
\end{claim}
\begin{proof}
We begin by fixing a step $0 \le i \le \im$, and we assume that $i < \tau$. Throughout we write $t=t(i)$. Let $f_1(t)=f(t)q(t)^{-1} = \exp((W+1)t^{r} + Wt)$, and let $\Xi_1 := f_1(t+s^{-1})-f_1(t)$.
By the same reasoning given in Claim \ref{clm:YApmbds}, we may bound $|f_1(t)| < n^{\ep/2}$, say, for large $n$,
and $ f_1''(t) \le n^{o(1)} f_1'(t) $, and so
\begin{equation}\label{eq:Xi1estimate}
\left|\Xi_1 - \frac{f_1'(t)}{s}\right| = O\lp \frac{\max_{t^* < \tm} f_1'' (t^*)}{s^2}\rp=o\lp \frac{f_1'(t)}{s}\rp.
\end{equation}
Next, we observe that
\[Z_A(i+1)-Z_A(i) = d_{i+1}(A)-d_i(A) - \frac{n}{N} - \Xi_1 \cdot n^{1/r-\ep}.\]
The boundedness claim then follows for $n$ sufficiently large as $0 \le d_A(i+1)-d_A(i) \le 1$ and as
\[|\Xi_1|\cdot n^{1/r-\ep} \le n^{\ep/2 + o(1)}\cdot n^{1/r-\ep}\cdot s^{-1} < n/N \]
as $s^{-1} =D^{1/r}/N = \Theta(n^{1-1/r}/N)$.
Turning to the supermartingale condition, observe that $d_{i+1}(A)=d_i(A)+1$ if and only if $e_{i+1}$ lies in the set of open $r$-sets counted by $Q_A(i)$. Conditioned on the history of the process up to step $i$, it follows that
\begin{align}
\nonumber \ev{Z_A(i+1)-Z_A(i)} &= \frac{Q_A(i)}{|O(i)|} - \frac{n}{N} - \Xi_1\cdot n^{1/r-\ep} \\
\nonumber &\le \frac{(q(t) + f(t) n^{-\ep})\cdot n}{(q(t)-N^{-\gamma})\cdot N} - \frac{n}{N} - \Xi_1\cdot n^{1/r-\ep}\\
\nonumber &= \frac{N^{-\gamma} + f(t)n^{-\ep}}{(q(t)-N^{-\gamma})} \cdot\frac{n}{N} - \Xi_1\cdot n^{1/r-\ep}\\
\nonumber &\le (N^{-\gamma} + f(t) n^{-\ep})\cdot 2q(t)^{-1}\cdot \frac{n}{N} - \Xi_1\cdot n^{1/r-\ep}\\
\nonumber &= (2q(t)^{-1}N^{-\gamma} + 2f_1(t) n^{-\ep})\cdot \frac{n}{N} - \Xi_1\cdot n^{1/r-\ep}\\
\label{eq:ZAlowerbd} &\le 4f_1(t) \cdot n^{-\ep}\cdot \frac{n}{N} - \Xi_1\cdot n^{1/r-\ep}
\end{align}
Note that the first inequality holds as $\ca{T}_i$ and $Y_A^+(i) \ge 0$ since $i < \tau$, the second as $q(t)-N^{-\gamma} \ge q(t)/2$ since $\zeta \ll \gamma$, and the final as $N^{-\gamma} \le f(t)\cdot n^{-\ep}$, since $f(t) \ge 1$ and $\ep \ll \gamma$. Noting that for large $n$, $D \ge n^{r-1}/r^r$ and so $s^{-1} \ge n^{1-1/r}/(rN)$, by \eqref{eq:Xi1estimate} we have
\begin{align*}
\Xi_1 \cdot n^{1/r-\ep} &= (1+o(1))\cdot \frac{f_1'(t)}{s}\cdot n^{1/r-\ep} \\
&\ge (1+o(1))\cdot \frac{Wf_1(t)\cdot n^{1-1/r}}{rN}n^{1/r-\ep}\\
&> \frac{W}{2r}\cdot f_1(t) \cdot n^{-\ep} \cdot \frac{n}{N}.
\end{align*}
Thus, since we assume $W$ is large, the supermartingale condition follows now from \eqref{eq:ZAlowerbd}.
\end{proof}
Finally, to show \eqref{eq:ZAbd}, we apply Lemma \ref{lem:supmartinbd} to yield
\begin{align*}
\pr{Z_A(\im \land \tau) > 0} &\le \exp \left\{ - \Omega \lp \frac{n^{2/r-2\ep}}{ \frac{n}{N} \cdot \zeta s\log^{1/r} N } \rp \right\} \\
&= \exp\left\{ - \frac{n^{2/r-2\ep}}{ n^{1 - (r-1)/r + o(1)} } \right\} \\
&= \exp\left\{ - n^{1/r - 2\ep- o(1)} \right\}
\end{align*}
which suffices as $\ep$ is small. This completes the proof of Lemma \ref{lem:maxdeg}.
\subsection{Proof of Lemma \ref{lem:relopentraj}} We begin by letting
\[S = S(n) = \bin{2\ell}{r} - 2\bin{\ell}{r},\]
and we note that $S = \Theta(k^r)$.
We fix a pair $A,B$ of disjoint $\ell$-element subsets of $[n]$, and define the following sequences of random variables: for each step $i \ge 0$, let
\begin{align*}
X^+(i) &= q(t)\cdot S - Q_{A,B}(i) + f(t)\cdot Sn^{-\ep}, \mbox{ and } \\
X^-(i) &= q(t)\cdot S - Q_{A,B}(i) - f(t)\cdot Sn^{-\ep}.
\end{align*}
We next define the stopping time $\tau^*$ to be the minimum of $\tau_{A,B}$ and the first step $i$ for which $X^+(i) \le 0$, $X^-(i) \ge 0$, or the event $\ca{T}_i$ fails to hold.
\begin{claim}\label{clm:lem1mart}
The sequence $X^+(0),\ldots,X^+(\im \land \tau^*)$ forms a $(O(k^r/s),O(k^{r-1}/n^{4\ep}))$-bounded submartingale, and the sequence $X^-(0),\ldots,X^-(\im \land \tau^*)$ forms a $(O(k^r
/s),O(k^{r-1}/n^{4\ep}))$-bounded supermartingale.
\end{claim}
\begin{proof}
We fix a step $0 \le i \le \im$, and we suppose that $i < \tau^*$. Throughout we write $t=t(i)$, and note $t(i+1)=t+s^{-1}$ and that $s^{-1}=D^{1/r}/N=\Theta(n^{1-1/r-r})$.
To aid the calculations to follow, we begin by estimating the quantity $\Xi := f(t+s^{-1}) - f(t)$. Recall equation
(\ref{eq:Xiestimate}):
\begin{equation*}
\left|\Xi - \frac{f'(t)}{s}\right| = O\lp \frac{ \max_{ t^* \le \tm } f''(t^*) }{s^2}\rp= o\lp \frac{f'(t)}{s}\rp.
\end{equation*}
Observe that we may write
\[X^{\pm}(i+1)-X^{\pm} (i) = (q(t+s^{-1})-q(t))\cdot S - \lp Q_{A,B}(i+1)-Q_{A,B}(i)\rp \pm \Xi\cdot Sn^{-\ep}.\]
(As above, this stands for the pair of equations in which each $\pm$ is replaced with $+$ or with $-$, respectively.)
We begin by establishing the boundedness claims:
by (\ref{eq:qtdifferentialest}) and
as $S = \Theta(k^r)$, we have
\[0 \ge \lp q(t + s^{-1})-q(t)\rp \cdot S \ge -O\lp \frac{k^r}{s} \rp.\]
Next, bounding $|f'(t)| \le n^{\ep/2 + o(1)}$,
\[|\Xi|\cdot Sn^{-\ep} \le n^{-\ep/2+o(1)}\cdot \frac{k^r}{s} \]
In order to establish the boundedness part of the claim,
it remains to bound the quantity $Q_{A,B}(i+1)-Q_{A,B}(i)$.
Let $O_{A,B}(i)$ denote the set of $r$-sets that are
open with respect to the pair $A,B$ in $G(i)$, and let $O_{\tau}$ denote the set of
all open $r$-sets whose selection as $e_{i+1}$ would result in $\tau_{A,B}=i+1$.
Now, if $e_{i+1} \in O_{\tau}$, then $Q_{A,B}(i+1)-Q_{A,B}(i) = 0$ by definition, and, otherwise, we have
\[Q_{A,B}(i+1)-Q_{A,B}(i) = -|O_{A,B}(i) \cap ( C_{e_{i+1}}(i) \cup \{e_{i+1}\})|.\]
It suffices, then, to bound the quantity $|C_e(i) \cap O_{A,B}(i)|$ for all $e \in O(i)\setminus O_{\tau}$: fix such an open $r$-set $e$. Now, for any $f \in C_e(i) \cap O_{A,B}(i)$, there is a copy $T_{r,f}$ of $T^{(r)}$ in the graph $G(i)+e+f$ using both $e$ and $f$ as edges. Up to isomorphism, there are only three possibilities for the pair $(e,f)$ in that copy: $(e,f)$ maps to $(b_1,b_2)$, or to $(b_1,a)$, or to $(a,b_1)$. We treat these three cases separately.\\
\noindent {\bf Case 1: $(e,f)$ maps to $(b_1,b_2)$.} In this case, the $r-1$ vertices that map to the set $R$ lie entirely in $e$, and $f$ is the union of those $r-1$ vertices along with another vertex lying in $A \cup B$. Thus, we may bound the total number of such $f$ above by $rk$.\\
\noindent {\bf Case 2: $(e,f)$ maps to $(b_1,a)$.} Let $R' = e-f$, the set of $r-1$ vertices shared by all edges $b_j$ in this copy of $T^{(r)}$. Then $f-e \subseteq N_i(R')$: since $f \cap A \ne \emptyset$ and $f \cap B \ne \emptyset$ (as $f \in O_{A,B}(i)$), and since $e \notin O_{\tau}$, it follows that $|N_i(R') \cap( A \cup B)| \le k/n^{2\ep}$. Thus, for a fixed such choice of $R'$ there are fewer than $(k/n^{2\ep})^{r-1}$ such open $r$-sets $f$, yielding a total bound of at most $r(k/n^{2\ep})^{r-1}$.\\
\noindent {\bf Case 3: $(e,f)$ maps to $(a,b_1)$.} There exists an $(r-1)$-set $R' \subseteq A \cup B$ and a vertex $v \in e$ so that $f = R' \cup \{v\}$ and so that $e\setminus \{v\} \subseteq N_i(R')$. To bound the number of such $f$, it suffices to bound the number of $(r-1)$-sets $R' \subseteq A \cup B$ for which $N_i(R')$ contains $(r-1)$ vertices from $e$.
To that end, fix a vertex $v \in e$ and let $\ca{H}_v$ denote the $(r-1)$-uniform hypergraph on $(A \cup B) \setminus e $ whose edges are the $(r-1)$-subsets $X$ for which $N_i(X) \supseteq e\setminus \{v\}$. We claim that
\[\Delta_{r-2}(\ca{H}_v) < 4r.\]
Suppose to the contrary that this does not hold: then there exist an $(r-2)$-set $Y \subseteq (A \cup B)\setminus e$ and vertices $x_1,x_2,\ldots,x_{4r}\in (A\cup B)\setminus (Y \cup e )$ so that for each for each vertex $u \in e\setminus \{v\}$, $\{u\} \cup Y \cup \{x_j\} \in E(i)$ for $1 \le j \le 4r$. It follows from Lemma~\ref{lem:subgraph} that such a
configuration does not appear in $ G(i)$. Indeed, as this configuration spans $ 6r-3 $ vertices and has $ 4r(r-1) $ edges, the
probability that such a configuration appears is at most
\[ n^{6r-3} \left( \frac{i}{N} \right)^{4r(r-1)} = n^{6r-3 - 4 (r-1)^2 + o(1)} = o(1). \]
It follows that $|\ca{H}_v| < 4r \bin{k}{r-2}$, and thus the total number of such open $r$-sets $f$ as above is less than $4r^2k^{r-2}$.\\
As $\ep$ is small and as $k = n^{1/r + o(1)}$, it follows that for large $n$ we have
\[ |C_e(i) \cap O_{A,B}(i)| \le rk + r\cdot (k/n^{2\ep})^{r-1} + 4r^2k^{r-2} = O(k^{r-1}/n^{2\ep(r-1)}),\]
and as $r \ge 3$ we conclude that
\[0 \ge Q_{A,B}(i+1)-Q_{A,B}(i) = -O(k^{r-1}/n^{4\ep}).\]
Thus, it follows that the sequences $X^{\pm}(0), \ldots, X^{\pm}(\im \land \tau^*)$ are $(O(k^r/s), O(k^{r-1}/n^{4\ep}))$-bounded as claimed.\\
We now turn to the sub- and supermartingale claims, and we remark that all expectation and probability calculations to follow are implicitly conditioned on the history of the process up to step $i$. We begin by bounding the expected value of $Q_{A,B}(i+1)-Q_{A,B}(i)$. Recall that we assume $i < \tau_{A,B}$ and that $O_{\tau} \subseteq O(i)$ consists of the open $r$-sets whose selection as $e_{i+1}$ would yield $\tau_{A,B} = i+1$. We claim that
\begin{equation}\label{eq:otauest}
|O_{\tau}| \le 4n^{2\ep}\cdot k\,
\end{equation}
To see this, let
\[\ca{R}:=\left\{X \in \bin{[n]}{r-1} : |N_i(X)\cap (A \cup B)| \ge k/(2n^{2\ep})\right\}.\]
Then $|\ca{R}| < 4n^{2\ep}$, which can be argued as follows. Suppose by way of contradiction that $\exists \ca{S} \subseteq \ca{R}$ with $|\ca{S}| = 4n^{2\ep}$. Let $N = \bigcup_{Y \in \ca{S}} (N_i(Y) \cap (A \cup B))$. By inclusion-exclusion and the fact that Lemma~\ref{lem:subgraph} implies that the co-degree of any pair of $(r-1)$-sets is at most $5r$ (see (\ref{eq:codegree})), we have
\[k \ge |N| \ge |\ca{S}|\cdot k /(2n^{2\ep}) - |\ca{S}|^2 5r \ge 2k-80rn^{4\ep},\]
a contradiction as $\ep$ is small and $k = n^{1/r + o(1)}$. To deduce \eqref{eq:otauest} it suffices to observe that each open $r$-set $e \in O_{\tau}$ can be written $e = \{v\} \cup X$ for some vertex $v \in A \cup B$ and $(r-1)$-set $X$ satisfying $|N_i(X) \cap (A \cup B)|\ge k/n^{2\ep}-1$ (and thus $X \in \ca{R}$).
Conditioning on the event $e_{i+1} \notin O_{\tau}$ then yields
\begin{align*}
\ev{Q_{A,B}(i+1)-Q_{A,B}(i)}
&= -\sum_{e \in O_{A,B}(i)} \frac{|C_e(i)\setminus O_{\tau}|}{|O(i)|}
\end{align*} by linearity of expectation. Consequently,
\[\ev{X^{\pm}(i+1)-X^{\pm}(i)} = (q(t+s^{-1})-q(t))\cdot S + \sum_{e \in O_{A,B}(i)} \frac{|C_e(i)\setminus O_{\tau}|}{|O(i)|} \pm \Xi\cdot Sn^{-\ep}.\]
To establish the submartingale claim, we note first that as $r \ge 3$ and $\ep \ll \gamma \ll 1/r$, from \eqref{eq:otauest} we have $|O_{\tau}| = n^{1/r + 2\ep + o(1)} < N^{-\gamma}\cdot D^{1/r}$.
Now, as $i < \tau^*$, $\ca{T}_i$ and $X^-(i) \le 0$ hold, we have
\begin{align*}
\sum_{e \in O_{A,B}(i)} \frac{|C_e(i)\setminus O_{\tau}|}{|O(i)|} &\ge \lp q(t)-\frac{f(t)}{n^{\ep}}\rp \cdot S \cdot \frac{(c(t)-2N^{-\gamma})D^{1/r}}{(q(t)+N^{-\gamma})N}\\
&= \lp 1 - \frac{N^{-\gamma} +f(t)n^{-\ep}}{q(t)+N^{-\gamma}}\rp(c(t)-2N^{-\gamma})\cdot \frac{S}{s}\\
&\ge \lp 1 - 2q(t)^{-1}f(t)n^{-\ep}\rp(c(t)-2N^{-\gamma})\cdot \frac{S}{s}\\
&\ge \lp c(t) - 2c(t)q(t)^{-1}f(t)n^{-\ep}-2N^{-\gamma}\rp\cdot \frac{S}{s}\\
&\ge \lp c(t) - (2c(t)q(t)^{-1}+1)f(t)n^{-\ep}\rp\cdot \frac{S}{s}.
\end{align*}
Note that these bounds follow for large $n$ since $f(t) \ge 1$ and $\ep \ll \gamma$ imply $N^{-\gamma} \le f(t)n^{-\ep}/2$. Applying this and \eqref{eq:qtdifferentialest} gives
\begin{align*}
\ev{X^+(i+1) - X^+(i)} &\ge \Xi\cdot Sn^{-\ep} - (2c(t)q(t)^{-1}+1)f(t)\frac{Sn^{-\ep}}{s}-O\lp \frac{1}{s^2}\rp\\
&\ge \Xi\cdot Sn^{-\ep} - (2c(t)q(t)^{-1}+2)f(t)\frac{Sn^{-\ep}}{s}\\
&= \lp (1 + o(1))f'(t)- (2c(t)q(t)^{-1}+2)f(t)\rp \cdot\frac{Sn^{-\ep}}{s}
\end{align*}
by \eqref{eq:Xiestimate}. Since $f'(t)=(Wrt^{r-1}+W)f(t)$ and $2c(t)q(t)^{-1}= 2rt^{r-1}$, this final bound is nonnegative for large $n$ as $W$ is large, and so $X^+(0),\ldots,X^+(\im\land \tau)$ forms a submartingale.
Turning to the supermartingale claim, we take a similar approach and begin by noting as $\ca{T}_i$ holds and $X^+(i) \ge 0$,
\begin{align*}
\sum_{e \in O_{A,B}(i)} \frac{|C_e(i)\setminus O_{\tau}|}{|O(i)|} &\le \lp q(t)+\frac{f(t)}{n^{\ep}}\rp \cdot S \cdot \frac{(c(t)+N^{-\gamma})D^{1/r}}{(q(t)-N^{-\gamma})N}\\
&= \lp 1 + \frac{N^{-\gamma} +f(t)n^{-\ep}}{q(t)-N^{-\gamma}}\rp(c(t)+N^{-\gamma})\cdot \frac{S}{s}\\
&\le \lp 1 + 2q(t)^{-1}f(t)n^{-\ep} \rp(c(t)+N^{-\gamma})\cdot \frac{S}{s}\\
&\le \lp c(t) + (2c(t)q(t)^{-1}+1)f(t)n^{-\ep} \rp\cdot \frac{S}{s}.
\end{align*}
The supermartingale condition then follows in essentially the same way as the submartingale condition above.
\end{proof}
Now, as $X^+(0) = Sn^{-\ep}$, $X^-(0) = -Sn^{-\ep}$, $S = \Theta(k^r)$ and $\im = s\cdot n^{o(1)}$, it follows from Claim \ref{clm:lem1mart} and Lemmas \ref{lem:submartinbd} and \ref{lem:supmartinbd} that
\[\pr{X^+(\im \land \tau^*) \le 0} \le \exp \left\{ - \Omega \lp \frac{ S^2 n^{-2 \ep}}{ \frac{ k^r}{s} \cdot \frac{ k^{r-1}}{n^{4 \ep}} \cdot s n^{o(1)}} \rp \right\} = \exp\left\{ - k \cdot n^{2\ep - o(1)} \right\}. \]
Simillarly, we have
\[\pr{X^-(\im \land \tau^*) \ge 0} \le \exp \left\{ - k \cdot n^{2\ep - o(1)} \right\} .\]
Since there are fewer than $n^{2k} = \exp \{ 2k \log n \}$ choices of the pair of sets $A$ and $B$, Lemma \ref{lem:relopentraj} follows from the union bound.
\bibliographystyle{plain}
|
\section{Introduction}
One-dimensional (1D) quantum wires can be defined in a two-dimensional electron gas (2DEG) using split-gate nanostructure devices~\cite{Thornton1986}.
The conductance of a quantum wire displays plateaus as a function of gate voltage, quantized in units of the conductance quantum $G_0 = 2e^2/h$~\cite{Wharam1988, vanWees1988}, which can be understood within a non-interacting framework. However, an anomalous feature appears near $0.7G_0$--the `0.7 structure' or `0.7 anomaly'~\cite{Thomas1996, Micolich2011}--which occurs as a direct result of electron-electron interactions. This intriguing conductance anomaly continues to inspire efforts to explain its occurrence~\cite{Bauer2013, Iqbal2013}. Many theories have been proposed, including spontaneous spin polarization~\cite{Thomas1996, Wang1996}, quasi-bound state formation and associated Kondo effect~\cite{Cronenwett2002, Meir2002, Rejec2006, Iqbal2013}, and enhanced electron interactions as electrons slow on passing through the 1D channel, due to the potential barrier~\cite{Sloggett2008, Bauer2013, Lunde2009}.
The conductance value of the 0.7 structure ($G_{0.7}$) can vary between $0.5$ and $0.9G_0$~\cite{Thomas1998, Thomas2000, Nuttinck2000, Wirtz2002, Reilly2001, Chiatti2006, Lee2006, Pyshkin2000, Hashimoto2001}, and depends on temperature ($T$), magnetic field ($B$), carrier density in the 1D channel, and device geometry.
The $T$ and $B$ dependence of the 0.7 structure are well established~\cite{Thomas1996}. However, differing results have been reported regarding the dependence on density (summarized in Ref.~\cite{Burke2012}) and upon the length of the 1D channel~\cite{Iqbal2013, Reilly2001}.
These data were obtained from a variety of 1D devices and confining potentials. This highlights the important role that the confining potential plays in determining $G_{0.7}$~\cite{Reilly2005}.
Due to limited resources available, many previous experimental studies reproduce results on a handful of devices at most.
This prevents statistically significant statements being formulated, and possibly leads to trends being overlooked.
Here, a statistical study of the 0.7 structure is presented, using an array of 256 individual split gates with the same dimensions.
Experiments were performed during a single cooldown, where each split gate was measured separately by means of a multiplexing scheme described in Ref.~\cite{Al-Taie2013}.
Despite the identical device design, fabrication and measurement conditions, large differences exist in the appearance and value of the 0.7 anomaly.
Systematic methods are employed to estimate $G_{0.7}$ and the curvature of the potential barrier in the transport direction.
The value of $G_{0.7}$ appears to be highly sensitive to the specifics of the 1D potential, which differs between even nominally identical split gates.
\begin{figure*}
\includegraphics[width=16cm,height=10cm,keepaspectratio]{Fig1}
\caption{\label{Fig1}
(Color online) (a) Typical trace of conductance $G$ as a function of the voltage applied to the split gate $V_{sg}$, from $V_{sg}$ = 0 to $V_p$. The vertical (horizontal) dashed line indicates $V_d$ ($G_d$).
The inset shows a schematic diagram of a split gate, where S and D correspond to the source and drain contacts, respectively.
(b) Scatter plot of $G_d$ against $V_p$ for which $r=-0.81$, indicating a strong negative correlation.
(c) Histogram of $V_d$ from 240 split gates, for a bin size of 1 mV.
(d) Color scale of $V_d$ as a function of spatial location, where each box represents a split gate in the array. Clear boxes indicate devices which did not define a 1D channel.
The boxes are equally spaced in the $x$ and $y$ directions for convenience, although in reality the split gates have horizontal and vertical pitch lengths of 100 and 130 $\mu$m, respectively.
}
\end{figure*}
The outline of the article is as follows.
First, the methods of fabricating and measuring the sample are described in Sec. II.
We then introduce the properties of the 1D conductance data in Sec. III, highlighting variations which may be related to local fluctuations in the density, thus characterizing the homogeneity of the wafer.
Next, we investigate the 0.7 anomaly and its dependence on properties of 1D conductance in Sec. IV. The method of estimating $G_{0.7}$ is described, to systematically analyze the large data set.
Since the 0.7 anomaly appears exceedingly sensitive to the 1D potential profile, variations in the potential are quantified by developing a model in Sec. V to estimate the curvature of the barrier in the transport direction, through fitting the data with the transmission probability from a saddle-point model.
Finally, in Sec. VI we acknowledge the role of disorder in giving rise to other anomalous features in conductance at unexpected values, close to $0.5G_0$.
\section{Methods}
The sample was fabricated on a modulation-doped GaAs/AlGaAs high electron mobility transistor (HEMT), in which the 2DEG is formed $90$ nm below the wafer surface. The 2D carrier density ($n_{2D}$) and mobility ($\mu$) of the 2DEG were measured to be $1.7\times10^{11}$ cm$^{-2}$ and $0.94\times10^6$ cm$^2$V$^{-1}$s$^{-1}$, respectively.
The split gates were arranged in a rectangular array, with pitch lengths of 100 and 130 $\mu$m in the two perpendicular directions. Each split gate was $400$ nm long and $400$ nm wide, defined using electron-beam (\emph{e}-beam) lithography [a schematic diagram of a split gate is shown in the inset of Fig.~\ref{Fig1}(a)]. Two-terminal measurements were performed at $T=1.4$ K, using an ac excitation voltage of 100 $\mu$V at $77$ Hz. Fifteen split gates failed to define a 1D channel, due to damage to one or both arms of the split gate, which is likely to have occurred during fabrication.
\section{Properties of 1D conductance}
Figure~\ref{Fig1}(a) shows a typical conductance trace as a function of the voltage applied to the split gate ($V_{sg}$). Conductance $G$ is plotted from $V_{sg}=0$ to pinch off voltage $V_p$.
There is an initial drop in $G$ before a quasi-1D channel forms at $V_{sg} = V_d$ (the 1D definition voltage). This is marked by a sudden change in gradient of $G$ as a function of $V_{sg}$.
The definition conductance ($G_d$), $V_p$ and $V_d$ are indicated by arrows in Fig.~\ref{Fig1}(a).
Figure~\ref{Fig1}(b) shows a scatter plot of $G_d$ against $V_p$ for 240 devices (241 were measured, however, the conductance of one dropped to zero without defining a 1D channel).
The degree of correlation can be quantified using the Pearson product-moment correlation coefficient ($r$), where $r=1$ ($r=-1$) corresponds to a perfect positive (negative) correlation, and $r=0$ corresponds to no correlation.
There is a strong negative correlation between $G_d$ and $V_p$ in Fig.~\ref{Fig1}(b), for which $r=-0.81$.
Since $G$ in these devices is determined by the number of 1D subbands, a higher $G_d$ suggests that there are more 1D subbands in the channel. The subband spacing may therefore be smaller, requiring the channel to be wider on definition. A stronger electric field (more negative voltage) will be required to fully deplete a wider channel, as reflected in Fig.~\ref{Fig1}(b).
No correlations were apparent between $V_d$ and $V_p$ ($r=-0.07$), or $G_d$ and $V_d$ ($r=0.06$).
Figure~\ref{Fig1}(c) shows a histogram of $V_d$, for a bin size of 1 mV. The mean $\bar{V}_d=-0.234$ V and standard deviation $\sigma_{V_d}=10.4$ mV, corresponding to $4.4 \%$ of the mean. Variations in density $n_{2D}$ across the array of split gates may be estimated using $V_d$ from each device (since $V_d$ is the voltage at which the 2D region beneath the gates is depleted).
By considering the capacitance between the split gate and the 2DEG, $V_d$ is related to $n_{2D}$ by
\begin{equation}
V_d = \frac{edn_{2D}}{\epsilon}\ ,
\end{equation}
where $e$ is the electronic charge, $d$ is the depth of the 2DEG, and $\epsilon$ is the dielectric constant of the material.
This equation is valid for gate width $w$ $\gg d$; here $w = 400$ nm and $d = 90$ nm.
For $\bar{V}_d=-0.234$ V, Eg. (1) gives $n_{2D}=1.74\times10^{11}$ cm$^{-2}$ (using $\epsilon \approx 12 \epsilon_0$ for $33 \%$ AlGaAs), and $\sigma_{n_{2D}} = 8\times10^{9}$ cm$^{-2}$. For comparison, conventional Hall bar measurements on two nearby sections of wafer yielded $1.65\times10^{11}$ cm$^{-2}$ and $1.69\times10^{11}$ cm$^{-2}$.
In this approximation, the capacitance due to the finite density of states in the 2DEG is ignored, since it is small with respect to the geometric capacitance. If it is assumed that changes in density are the only reason for differences in $V_d$, the distribution of $V_d$ is directly proportional to fluctuations in $n_{2D}$ (i.e. standard deviation $\sigma_{V_d} \approx 4\%$ corresponds to the same variation in $n_{2D}$).
Figure~\ref{Fig1}(d) shows a color scale (color online) of $V_d$ as a function of the position of each device in the array. On the chip, split gates are separated by 100 and 130 $\mu$m in the horizontal and vertical directions, respectively, whereas boxes are equally spaced in Fig.~\ref{Fig1}(d) for convenience. Clear boxes indicate split gates for which $V_d$ could not be determined.
Plotting $V_d$ as a function of spatial location illustrates how density fluctuations in a HEMT structure can be investigated on a micron scale.
This technique approximately characterizes the homogeneity of a wafer, since variations on this length scale not shown by conventional Hall bar measurements.
\section{0.7 structure}
\begin{figure}
\includegraphics[width=12cm,height=14cm,keepaspectratio]{Fig2}
\caption{\label{Fig2} (Color online) (a)-(c) Conductance against $V_{sg}$ for three nominally identical split gates. Well-developed conductance anomalies occur below $G_0$ in panels (a) and (b), indicated by the arrows. A much weaker shoulder-like feature occurs near $0.7G_0$ in panel (c), marked by the arrow.
(d) Conductance as a function of $V_{sg}$ for a fourth example split gate (solid line).
First, second and third derivatives $dG/dV_{sg}$, $d^2G/dV_{sg}^2$, and $d^3G/dV_{sg}^3$ are shown by the dashed, dotted, and dash-dotted curves, respectively. The $dG/dV_{sg}$ trace is offset vertically for clarity, and all derivative data are scaled vertically in order to be shown on the plot.
The dashed horizontal line shows $G_{0.7}$, corresponding to the local minimum in $dG/dV_{sg}$.
The upper and lower error bounds of $G_{0.7}$ are shown by the horizontal dotted lines, given by the nearest maxima in $d^2G/dV_{sg}^2$ and $d^3G/dV_{sg}^3$, respectively.
(e), (f) Conductance against $V_{sg}$ for two example split gates showing evidence of disorder. In panel (e), the conductance is no longer quantized in units of $G_0$, while in panel (f) the second plateau is unusually weak and both the second and third plateaus occur below the expected values (the arrow indicates a 0.7 structure).
}
\end{figure}
The conductance data from the array of split gates display a large variation in the appearance of the 0.7 structure.
Figures~\ref{Fig2}(a)-(c) show $G$ as a function of $V_{sg}$ for three example devices, corrected for series resistance $R_s$ (to ensure consistency, $R_s=1/G$ at $V_{sg}=0$).
In Figs.~\ref{Fig2}(a) and 2(b), well-defined structures occur near $0.7G_0$, marked by the arrows.
The feature in Fig.~\ref{Fig2}(b) is particularly pronounced.
A much weaker, shoulder-like structure is shown in Fig.~\ref{Fig2}(c), indicated by the arrow.
\begin{figure*}
\includegraphics[width=18cm,height=10cm,keepaspectratio]{Fig3}
\caption{\label{Fig3} (Color online) (a) Histogram of $G_{0.7}$, for a bin size of $0.005G_0$. Data are obtained from 36 devices, where $G_{0.7}$ is defined as the local minimum in $dG/dV_{sg}$.
(b)-(f) Scatter plots of $G_{0.7}$ against $\Delta R_{0.5}$, $W_1$, $V_p$, $G_d$ and $V_d$, respectively, where $\Delta R_{0.5} = \Delta V_{sg}$ from $G=0$ to $0.5G_0$, and $W_1=\Delta V_{sg}$ between $G=0.5$ and $1.5G_0$. No strong correlations are apparent; $r=0.07$, $0.07$, $-0.11$, $0.10$ and $0.33$ for (b), (c), (d), (e) and (f), respectively.
Error bounds on the estimate of $G_{0.7}$ are given by the width of the 0.7 anomaly~\cite{noteG07bounds}.
}
\end{figure*}
\subsection{Method of estimating $G_{0.7}$}
A systematic method of estimating $G_{0.7}$ is required to further analyze the data.
There is no stated definition of the conductance value that should be assigned to the 0.7 structure~\cite{note3}; however, it was the location of the flattest part of the feature--most often observed near $0.7G_0$--which led to it being given its particular name~\cite{Thomas2014}.
Therefore we present data where $G_{0.7}$ is defined as the local minimum in $dG/dV_{sg}$.
Figure~\ref{Fig2}(d) shows $G$ as a function of $V_{sg}$ (solid line) for another example device.
The first, second and third derivatives $dG/dV_{sg}$, $d^2G/dV_{sg}^2$, and $d^3G/dV_{sg}^3$ are shown by the dashed, dotted, and dash-dotted curves, respectively. The $dG/dV_{sg}$ trace is offset vertically for clarity.
The $V_{sg}$ at which $d^2G/dV_{sg}^2=0$ (corresponding to the local minimum in $dG/dV_{sg}$), is shown by the dashed vertical line. This gives our estimate of $G_{0.7}$ (indicated by the arrow). The `width' of the 0.7 plateau is estimated as $\Delta V_{sg}$ between the closest maximum in $d^3G/dV_{sg}^3$ to the left and closest maximum in $d^2G/dV_{sg}^2$ to the right of $d^2G/dV_{sg}^2=0$ (indicated by the vertical dotted lines).
This defines the bounds of our estimate of $G_{0.7}$, shown by the horizontal dotted lines~\cite{noteG07bounds}.
The value of $G_{0.7}$ can be obtained for a limited number of devices using this method, due to variations in shape of the 0.7 structure. If the anomaly is not sufficiently pronounced--for example in Fig.~\ref{Fig2}(c)--$G_{0.7}$ cannot be estimated (since there is no clear minimum in $dG/dV_{sg}$).
However, the resulting benefit is that the data set is reduced to devices for which the conductance characteristics are very similar. Since the strength of the 0.7 structure is related to the relative energy scales within the 1D system, these should therefore be similar for the data remaining.
In addition, we were careful to discard data which showed evidence of disorder at low $G$ (below $3G_0$), since this may affect $G_{0.7}$.
Various disorder effects were observed. In some instances the quantization of $G$ was significantly affected and no 0.7 structure existed [shown in Fig.~\ref{Fig2}(e)].
In other cases, a 0.7 structure was observed in addition to disorder effects (these effects included missing or weakened plateaus, deviations in plateau values from multiples of $G_0$, unusually weak quantization, and resonant features ranging from Coulomb blockade~\cite{Liang1997} to phase-coherent resonances~\cite{Kirczenow1989}).
Figure~\ref{Fig2}(f) shows $G$ as a function of $V_{sg}$ for a device in which the second and third plateaus occur below the expected values, and the second plateau is weak. The arrow indicates a strong 0.7 structure.
Such data were discarded because we cannot rule out disorder affecting $G_{0.7}$.
After discarding data which showed evidence of disorder, data from 98 split gates remained ($41\%$ of the 241 measured).
An estimate of $G_{0.7}$ is obtained for 36 of these 98 devices~\cite{07note}.
This highlights a key benefit of our multiplexing technique: By measuring many devices, we can discard $85\%$ of the data and still retain a data set sufficient for statistical analysis (36 is currently the largest number of devices for which $G_{0.7}$ has been estimated from measurements in a single cooldown).
We expect a lower rejection ratio for a sample fabricated on a wafer with higher mobility.
While we have have attempted to remove the effect of disorder, thermal broadening of energy levels may mask other disorder effects ($T=1.4$ K).
Measurements were performed at this $T$ in order to observe a well-defined 0.7 anomaly, since we anticipate any statistical (anti)correlation between $G_{0.7}$ and other parameters to be most clear at $T$ for which the 0.7 anomaly is strongest.
\subsection{Dependence of $G_{0.7}$ on properties of 1D conductance}
Figure~\ref{Fig3}(a) shows a histogram of the $36$ counts of $G_{0.7}$ (for a bin size of $0.005G_0$), in which $G_{0.7}$ ranges from $\approx 0.63$ to $0.84G_0$.
The mean $\bar{G}_{0.7} =0.75G_0$, and standard deviation $\sigma_{0.7} = 0.05G_0$.
The spread of $G_{0.7}$ is within that which has been reported previously~\cite{Burke2012}.
However, Fig.~\ref{Fig3}(a) represents data from devices with a geometrically identical design, which have undergone exactly the same fabrication process, and were measured during a single cooldown at a constant $T$.
Therefore the difference of more than $0.2G_0$ in $G_{0.7}$ is quite remarkable.
Even though each split gate is patterned with a geometrically-identical design, the shape of the 1D potential profile may vary from device to device. This occurs for a number of reasons including fluctuations in $n_{2D}$ (standard deviation of $n_{2D}$ is estimated to be $\approx 4\%$ of the mean, Sec. III), and/or the existence of impurities close to the 1D channel.
The differences in $G_{0.7}$ suggest it is highly dependent on the shape of the potential profile, and minor variations thereof.
In Figs.~\ref{Fig3}(b) to (f), $G_{0.7}$ is plotted against various properties of the 1D conductance trace.
Specifically, Figs.~\ref{Fig3}(b), (c), (d), (e), and (f) show $G_{0.7}$ against $\Delta R_{0.5}$, $W_1$, $V_p$, $G_d$, and $V_d$, respectively, where $\Delta R_{0.5}=\Delta V_{sg}$ from $G=0$ to $0.5G_0$ (corresponding to the steepness of the initial rise in $G$ towards the 0.7 anomaly), and $W_1 = \Delta V_{sg}$ between $G = 0.5$ and $1.5G_0$ (estimating the width of the first conductance plateau).
These properties of the 1D conductance trace are determined by or reflect physical conditions of the system. For example, $V_p$ indicates the strength of the electrostatic field at pinch off; this field is weaker for values of $V_p$ closer to zero, such that the confinement potential is generally shallower. Lower electron densities also often result in $V_p$ closer to zero.
As discussed in Sec. III, $G_d$ and $V_d$ depend on the initial number of 1D subbands in the 1D channel, and fluctuations in $n_{2D}$, respectively.
Additionally, the length of the conductance plateaus depends on the 1D subband spacing, and steepness of the transitions between plateaus depends on the length of the potential barrier in the transport direction (discussed in Sec. V).
A relationship between $G_{0.7}$ and any of these properties may illuminate physical conditions which govern $G_{0.7}$.
However, no correlations are apparent in Figs.~\ref{Fig3}(b)-(f); [$r=0.07$, $0.07$, $-0.11$, $0.10$ and $0.33$, for 3(b), 3(c), 3(d), 3(e), and 3(f), respectively].
Correlations are perhaps hidden because although the properties of conductance may be primarily related to a particular parameter, they are also subject to other influences.
These data illustrate that $G_{0.7}$ is governed by a combination of conditions and is highly sensitive to the specific potential profile within each device.
\begin{figure}
\includegraphics[width=16cm,height=11cm,keepaspectratio]{Fig4}
\caption{\label{Fig4}
(Color online) (a) The solid line shows the measured $G$ as a function of $V_{sg}$ from an example device [same as Fig. 2(d)]. Dot-dashed lines show a fit to the data ($G_n$) for individual subbands $n=1$, 2, and 3, using a modified saddle-point model~\cite{Buttiker1990}. The points at which $G_n=0.5$ are aligned with the corresponding $G = (n-0.5)G_0$ values on the experimental data, and the dashed line shows $\sum_n G_n$.
(b) Lever arm $\alpha$ as a function of 1D subband index $n$. The inset shows a grayscale diagram of the transconductance $dG/dV_{sg}$ as a function of $V_{sg}$ and $V_{sd}$. The dark (white) regions correspond to high (low) transconductance. The conductance values of low transconductance regions are labeled in units of $2e^2/h$, and $\Delta E_{1,2}$ is given by the point at which the two dashed lines cross.
}
\end{figure}
\begin{figure*}
\includegraphics[width=18cm,height=12cm,keepaspectratio]{Fig5}
\caption{\label{Fig5}
(Color online) (a) Scatter plot of $\hbar\omega_{x,1}$ against $\Delta R_{0.5}$, where $\Delta R_{0.5} = \Delta V_{sg}$ from $G=0$ to $0.5G_0$. Error bounds are obtained by finding $\hbar\omega_{x,1}$ for the upper and lower estimates of $\alpha_1$. The strong correlation illustrates the accuracy of the fit to the experimental data.
(b), (c) Scatter plots of $\Delta R_{0.5}$ against $G_d$, and $\hbar\omega_{x,1}$ against $G_d$, respectively. Panel (b) shows a reasonably distinct diagonal cutoff such that there are no data points in the upper-left section of the plot. This is weakly reflected in panel (c).
(d)-(f) Scatter plots of $\hbar\omega_{x,1}$ against $V_d$, $V_p$ and $W_1$, respectively, where $W_1=\Delta V_{sg}$ between $G=0.5$ and $1.5G_0$.
}
\end{figure*}
\section{Quantifying the 1D confining potential}
To quantify the conditions of confinement within each device one can measure the subband spacing using dc bias spectroscopy~\cite{Patel1991}. This type of individual characterization is time consuming, and an automated routine of extracting information from the conductance trace is preferred, because of large data set.
We therefore fit the data with a transmission probability based on the saddle-point model~\cite{Buttiker1990} in order to estimate the harmonic oscillator energy $E_x =\hbar\omega_x$, which describes the curvature of the potential barrier in the transport direction.
\subsection{Model}
Figure~\ref{Fig4}(a) shows $G$ as a function of $V_{sg}$ (solid line) for an example device.
The dashed line shows a fit to the data for the transmission probability $T_n=(1+e^{-\pi\epsilon_n})^{-1}$, where $n$ is the 1D subband index, $\epsilon_n = 2[E-E_n-V_0]/\hbar\omega_{x,n}$, $V_0$ is the potential at the center of the 1D channel, and $E_n$ is the energy of the bottom of the $n^{th}$ 1D subband (relative to $V_0$).
We deviate from a strict saddle-point model which assumes 1D subbands are equally spaced, since this is not the case for real devices.
Additionally, we use a subband-dependent $\hbar\omega_{x,n}$ to achieve a better fit to the data.
The conductance is calculated independently for $n = 1$, $2$, and $3$ using
\begin{equation}
G_n = G_0 \int\;dE\; \left(-\frac{\partial f}{\partial E}\right)\;T_n\ ,
\end{equation}
where $f$ is the Fermi-Dirac distribution $f=(1+e^{(E-\mu)/k_BT})^{-1}$.
The conductance increases by $2e^2/h$ when chemical potential $\mu$ rises above the bottom of the 1D subband. We choose a reference frame in which each subband edge is initially at $\mu = 0$, i.e., $V_0$ and $E_n = 0$. The integration is performed between limits $\pm50k_BT$, for $T = 1.4$ K.
The $G_n$ for each subband is then individually scaled by $\alpha_n^{-1}$, where $\alpha$ is a lever arm relating $E$ and $V_{sg}$ obtained from dc bias spectroscopy measurements. We follow the method of estimating $\alpha$ described in the supplementary material of Ref.~\cite{Srinivasan2013}; $\Delta E = \alpha e\Delta V_{sg}$, and $\alpha = \partial V_{sd}/\partial V_{sg}$.
Figure~\ref{Fig4}(b) shows the average values of $\alpha_n$ for each subband from dc bias measurements on four split gates. The error bars are the maximum and minimum estimates of $\alpha_n$.
A grayscale diagram of transconductance $dG/dV_{sg}$ as a function of $V_{sd}$ and $V_{sg}$ from one of the devices is shown as an inset to Fig.~\ref{Fig4}(b). The dark (white) regions correspond to high (low) transconductance, and conductance values of low transconductance regions are labeled in units of $2e^2/h$.
The data are corrected for series resistance (also following the method described in the supplementary material of Ref.~\cite{Srinivasan2013}).
To automate the fitting routine, $G_n$ for each split gate is scaled by the average $\alpha_n$. For simplicity, we also use a constant $\alpha_n$ for each subband, although in reality it varies with $V_{sg}$.
We believe the use of an average $\alpha_n$ to be the most significant source of error in estimating $\hbar\omega_{x,n}$.
The 1D subband spacings were also obtained from the dc bias data. The average $\Delta E_{n,n+1} = 2.8, 1.9$ and $1.7$ meV, for $n = 1, 2$ and $3$, respectively.
Since this measurement was performed at $T = 1.4$ K, no feature appears near $0.85 G_0$ at small dc bias \cite{Patel1991, Kristensen2000}. Therefore, $\Delta E_{1,2}$ was estimated as the crossing point of the tranconductance peaks separating the $0.25$ and $G_0$ regions, and the $G_0$ and $1.5 G_0$ regions.
This is illustrated by the dashed lines on the grayscale [inset, Fig.~\ref{Fig4}(b)].
After scaling, the points at which $G_n=0.5 G_0$ are aligned with the corresponding points on the experimental data, i.e., 0.5, 1.5, and $2.5 G_0$ for subbands $n = 1$, $2$ and $3$, respectively.
A fitting routine is used to find the minimum difference squared between the experimental and calculated conductances. For $n=2$ ($3$) the fit was performed between $G=1.01$ ($G=2.01$) and $2 G_0$ ($3 G_0$), with fitting parameter $\hbar\omega_{x,2}$ ($\hbar\omega_{x,3}$). For $n=1$, the fit was performed on the lower half of the riser to the first plateau (from $0.01$ to $0.5 G_0$), to avoid the 0.7 structure.
Figure~\ref{Fig4}(a) shows $G$ after the fitting has been performed, where dot-dashed lines show $G_n$ for individual subbands. The dashed line shows the sum of these data, which overlays the measured $G$ (solid line) well.
The only fitting parameters used in the model are $\hbar\omega_{x,n}$ for $n=1$, $2$, and $3$. Since this is a non interacting model, there is no 0.7 structure in the fit to the conductance data.
\begin{figure}
\includegraphics[width=8cm,height=10cm,keepaspectratio]{Fig6}
\caption{\label{Fig6} (Color online) (a) Scatter plot of $G_{0.7}$ against $\hbar\omega_{x,1}$ ($r=0.12$).
Error bounds in $\hbar\omega_{x,1}$ are obtained from the upper and lower estimates of lever arm $\alpha_1$. The bounds on $G_{0.7}$ are related to the width of the conductance anomaly.
}
\end{figure}
\subsection{Dependence of $\hbar\omega_{x,1}$ on properties of 1D conductance}
Using the method described above, $\hbar\omega_{x,n}$ is estimated for the 98 split gates which did not show evidence of disorder below $3G_0$.
The mean $\hbar\omega_{x,1}=1.95$, $\hbar\omega_{x,2}=1.92$ and $\hbar\omega_{x,3}=1.56$ meV, for which the standard deviations $\sigma_{\hbar\omega_{x,1}}=0.41$, $\sigma_{\hbar\omega_{x,2}}=0.26$ and
$\sigma_{\hbar\omega_{x,3}}=0.22$ meV (corresponding to $21\%$, $14\%$ and $14\%$ of the mean, respectively).
This highlights the differences that exist in 1D potential from device to device, despite the lithographically identical design.
The steepness of the initial rise in $G$ towards the 0.7 anomaly is given by $\Delta R_{0.5}=\Delta V_{sg}$ from $G=0$ to $0.5G_0$. Figure~\ref{Fig5}(a) shows a scatter plot of $\hbar\omega_{x,1}$ against $\Delta R_{0.5}$. This is a helpful check of the quality of the fit, since the fitting was performed over this range of $G$. There is a strong degree of correlation (Pearson product-moment correlation $r = 0.9$).
Thus, the steepness of the transition to $G_0$ for the fitted $G$ accurately reflects that of measured data, indicating that the fit is reasonable. Error bounds are given by finding $\hbar\omega_{x,1}$ for the upper and lower values of $\alpha_1$ shown in Fig.~\ref{Fig4}(b).
Figure~\ref{Fig5}(b) shows $\Delta R_{0.5}$ against the 1D definition conductance $G_d$.
While there is no apparent correlation, there seems to be relatively distinct diagonal cutoff above which there are no data points (in the top-left triangular section of the plot). Thus, a sharper initial rise in conductance tends to correspond to a lower $G_d$.
In Fig.~\ref{Fig5}(c), $\hbar\omega_{x,1}$ is plotted against $G_d$. The diagonal cutoff is reflected here, although less distinctly.
Figures~\ref{Fig5}(d),~\ref{Fig5}(e), and~\ref{Fig5}(f) show $\hbar\omega_{x,1}$ against $V_d$, $V_p$, and $W_1$, respectively.
No correlations are apparent between $\hbar\omega_{x,1}$ and these other properties of the 1D conductance trace (no correlations were also evident between $\hbar\omega_{x,2}$ or $\hbar\omega_{x,3}$ and these properties).
It is possible that trends may be masked by errors in the lever arm $\alpha$. However, the spread in $\hbar\omega_{x,1}$ for most values of $G_d$, $V_d$, $V_p$ and $W_1$ is larger than the estimated error.
\subsection{Dependence of $G_{0.7}$ on $\hbar\omega_x$}
Figure~\ref{Fig6} shows a scatter plot of $G_{0.7}$ against $\hbar\omega_{x,1}$.
For these data $r=0.12$, indicating an exceedingly weak correlation.
Unfortunately, any correlation which may exist is likely to be masked by errors in the estimate of $\hbar\omega_{x,1}$.
We believe the error in $G_{0.7}$ to be less significant since $G_{0.7}$ is given by a well-defined point (the local minimum in $dG/dV_{sg}$), and bounds in $G_{0.7}$ are instead related to the width of the conductance anomaly~\cite{noteG07bounds}.
The error in $\hbar\omega_{x,1}$ could be reduced by using the correct $\alpha$ for each device. This requires dc bias measurements to be performed for every device.
A specific correlation between $G_{0.7}$ and $\omega_{x}$ is predicted by certain models for the origin of the 0.7 structure. For example, the 1D Kondo effect occurs when electrons are localized within a 1D channel~\cite{Sfigakis2008}. In the 1D Kondo scenario, for large $\omega_x$ the Kondo temperature ($T_K$) is also high~\cite{Meir2002, Rejec2006, Hirose2003}. Thus, an increase in $\omega_x$ ($T_K$) should cause $G_{0.7}$ to increase at a given temperature, since $G \propto [1-(T/T_K)^2]$. However, other models expect the opposite trend. It has been proposed that the 0.7 structure is related to enhanced interactions as the electrons slow down on passing through the 1D barrier~\cite{Sloggett2008, Bauer2013, Lunde2009}. Of these models, only one~\cite{Sloggett2008} studies the high-temperature dependence of $G_{0.7}$ with $\omega_x$.
If the 1D barrier is a saddle-point potential~\cite{Buttiker1990}, the value of $G_{0.7}$ is predicted to decrease as $\omega_x$ increases.
We do not observe a strong enough trend in our data to support one theory above another.
\begin{figure}[t]
\includegraphics[width=8cm,height=10cm,keepaspectratio]{Fig7}
\caption{\label{Fig7} (Color online) Conductance as a function of $V_{sg}$ for an example split gate, where an anomalous feature appears near $0.5G_0$, indicated by the arrow.
}
\end{figure}
\section{Anomalous conductance features near $G=0.5G_0$}
Conductance anomalies were also observed at $G$ values lower than the range shown in Fig~\ref{Fig3}(a).
Figure~\ref{Fig7} shows $G$ as a function of $V_{sg}$ for a device in which an anomalous feature occurs at $0.5G_0$ (marked by the arrow).
Of the 241 split gates measured, $\approx 2\%$ showed conductance anomalies at this value; it is therefore unusual to find features at $0.5G_0$ in 1D devices on GaAs/AlGaAs heterostructures for $B=0$ T (the magnetic field at which these measurements were performed).
These data were not included in the analysis because of evidence of disorder; for example, in Fig.~\ref{Fig7} the plateaus do not appear at correct values~\cite{SeriesRes}. The co-existence of disorder effects may be responsible for the lowering of the conductance of the 0.7 structure to $0.5G_0$.
A statistical measurement makes it possible to distinguish the ``normal'' characteristics from device-specific effects, which may be related to disorder. Devices which display unusual conductance anomaly can be investigated for rare physical phenomena, while very clean devices can be used to identify standard behavior.
The reproducibility of conductance characteristics has been investigated by thermally cycling the sample.
It was found that many of the split gates which showed evidence of disorder did so on both cooldowns~\cite{Al-Taie2013B}.
The aim of the current article is to compare properties of 1D conductance from a large number of devices on a single cooldown.
The reproducibility of these properties on thermal cycling warrants a further, separate study.
The sample was also illuminated with a light-emitting diode (LED), which increased $n_{2D}$ and $\mu$ to $2.9\times10^{11}$ cm$^{-2}$ and $2.2\times10^6$ cm$^2$V$^{-1}$s$^{-1}$, respectively. This resulted in longer, better-defined plateaus in conductance, since the 1D confining potential becomes stronger and the 1D subband spacing increases. However, many devices showed occasional structure in conductance which had the appearance of resonant transmission through the quantum wire, consistent with an enhancement of resonant effects due to sharper confinement~\cite{Kirczenow1989}.
We did not investigate $G_{0.7}$ or $\hbar\omega_x$ in this case since the estimates are very likely to be affected by the resonances.
\section{Conclusion}
Across an array of nominally identical split gates (measured at a single $T$), significant fluctuations were seen to exist in the 1D potential. These have been quantified by estimating the curvature of barrier in the transport direction ($\hbar\omega_{x}$) for each device.
Large variations were observed in both the appearance of the 0.7 structure
and the value at which it occurred. The 0.7 structure appears to be extremely sensitive to the specific 1D potential in each device. Measuring many devices has enabled a statistical study to be performed. No correlations were apparent between $G_{0.7}$ and $\hbar\omega_{x}$, or other properties of the 1D conductance trace.
A specific set of physical conditions combine to give a particular conductance trace for a split gate.
With the current analysis, the effect of individual factors influencing the conductance properties cannot be separated.
Thus, parameters which may govern $G_{0.7}$ and $\hbar\omega_x$ have not been identified.
This may become possible by performing dc bias spectroscopy for each device in order to accurately measure 1D subband spacing and lever arm $\alpha$, thereby giving a better estimate of $\hbar\omega_x$.
The confining potential will be affected by fluctuations in the background potential due to the ionized dopants (leading to local density variations) and the existence of impurities (giving rise to disorder effects). We removed data which showed evidence of disorder from the analysis, although a fuller study requires $B$- and $T$-dependent measurements (since disorder effects will be masked at the temperature at which the measurements were performed). We have shown that disorder does give rise to conductance anomaly at unexpected values, e.g., close to $0.5G_0$.
Disorder effects can be reduced by fabricating samples on a wafer with higher electron mobility, or using an undoped heterostructure and electrostatically inducing the 2DEG~\cite{See2012, Harrell1999, Sarkozy2009}.
This work was supported by the Engineering and Physical Sciences Research Council Grant No. EP/I014268/1. The authors thank T.-M. Chen, C. J. B. Ford, I. Farrer, E. T. Owen and K. J. Thomas for useful discussions, and R. D. Hall for \emph{e}-beam exposure.
$*$ Corresponding author. E-mail address: <EMAIL>
|
\section{Introduction}
This paper is a continuation of our earlier work \cite{M-M} on
$q$-frequent hypercyclicity, which coincides with frequent
hypercyclicity for $q=1$. We study here this concept for linear maps
defined on Banach algebras of operators on Banach and Hilbert
spaces. Hypercyclicity in spaces of operators was initiated by K.C. Chan \cite{C} and subsequently studied by J. Bonet, F. Martinez-Gimenez and A. Peris \cite{BFP}, K.C. Chan and R. Taylor \cite{CT}, F. Martinez-Gimenez and A. Peris \cite{FP}
and H. Petersson \cite{H}. Indeed, left multiplication operators $\mathfrak{L}_R(S)=RS$
were considered in \cite{BFP},\cite{C},\cite{CT} and \cite{FP} and their
general form $C_{R,T}(S)=RST$ was studied in \cite{H}. A collective work in \cite{BFP},\cite {C} and \cite{CT} states that a bounded operator
$R$ on a separable Banach space $X$ satisfies the Hypercyclicity
Criterion if and only if the left multiplication operator $\mathfrak{L}_R$ is hypercyclic on
$\mathcal{L}(X)$ in the topology of pointwise convergence.
This result holds for the topology of uniform convergence on compact subsets if $X^*$ is separable and $X$ has the
approximation property, see \cite{BFP}. In \cite{H} H. Petersson proved that if $T$ satisfies the Hypercyclicity Criterion in a separable Hilbert space, then $\mathfrak{L}_T$ as well as the conjugate operator $C_T$ is hypercyclic on the Schatten von Neumann classes $S_p(H)$, $1\leq p<\infty$ and $\mathcal{K}(H)$.
In Section 3 we provide a sufficient criterion for $C_{R,T}$ to be $q$-frequently
hypercyclic on the algebra of compact operators on Banach spaces and
give applications to the unilateral and bilateral shift operators, and
in Section 4 we continue the study in the space $S_p(H)$. Finally in Section 5, using
an Eigenvalue Criterion, we characterize frequent hypercyclicity of certain maps of the form
$C_{R,T}$ defined on spaces of operators on the classical Hardy space and $\ell^p$.
\section{Preliminaries}
A continuous operator $T$ on a topological vector space (TVS) $X$ is said to be
$\mathbf{hypercyclic}$ if the set $\{T^nx: n\geq1\}$ is dense in $X$
for some $x\in X$. For $q\in \mathbb{N}$ (\textit{the set of natural
numbers}), $T$ is said to be $q$-$\mathbf{frequently}$
$\mathbf{hypercyclic}$ (see \cite{M-M}) if there exists a vector
$x\in X$ such that the set $\{n\in \mathbb{N}: T^nx\in U\}$ has
positive $q$-lower density for each non-empty open set $U\subset X$,
where the $q$-lower density of $A\subset \mathbb{N}$ is defined as
\begin{center}
$\displaystyle \underline{\text{$q$-dens}}(A)=\liminf_{N\rightarrow\infty}$ $\frac
{\text{card}\{n\in A:~ n\leq N^q\}}{N}$.
\end{center}
For $q=1$, the above notion is known as frequent hypercyclicity of
an operator, studied in \cite{BG}, \cite{BGE} and \cite{BE}. If $T$
is frequently hypercyclic, then it is $q$-frequently hypercyclic for
all $q\in \mathbb{N}$, however, the converse is not true, cf.
\cite{M-M}.
Let $X$ and $Y$ be separable Banach spaces. The space of all
bounded (resp. of all compact) operators from $X$ to $Y$ is denoted
by $\mathcal{L}(X,Y)$ (resp. $\mathcal{K}(X,Y)$). We shall use the
symbols $\mathcal{L}(X)$ and $\mathcal{K}(X)$ for $\mathcal{L}(X,X)$
and $\mathcal{K}(X,X)$ respectively. The real subspace of
$\mathcal{L}(H)$, of all self-adjoint operators on a separable
infinite dimensional Hilbert space $H$ is denoted by
$\mathcal{S}(H)$ and is equipped with the topology of uniform
convergence on compact subsets (COT). Also, for $p\in [1,\infty)$
the Schatten von Neumann class $S_p(H)$ is defined as the space
of all operators $T\in \mathcal{L}(H)$ for which the approximation
numbers $(a_n(T))\in \ell^p$, where
\begin{center}
$a_n(T)=\inf \{\|T-F\|:~ rank(F)<n\}, n\geq 1$,
\end{center}
See \cite{DJT} and \cite{BST} for more details on the Schatten
von Neumann classes.
For $R\in \mathcal{L}(X)$, the left and right multiplication
operators are respectively defined as $\mathfrak{L}_R(S)=RS$ and
$\mathfrak{R}_R(S)=SR$. Also, if $R$ is a bounded operator on a
Hilbert space, then the conjugate operator $C_R$ is defined as
$C_R(S)=RSR^*$.
Recall that a Banach space $X$ is said to have the \textbf{approximation
property (AP)} if the identity operator on $X$ can be approximated by
finite rank operators uniformly on compact subsets of $X$; that is,
for any $\epsilon >0$ and $K\subset X$ compact, there exists an
operator $F$ of finite rank such that $\|F(x)-x\|<\epsilon$,
$\forall x\in K$. If $X$ has the AP, then finite rank operators are
norm-dense in $\mathcal{K}(Y,X)$ for all Banach spaces $Y$, cf. \cite{DJT} and \cite{LT}.
A series $\sum_{n\geq 1} x_{n,j}$ in an $F$-space is said to be \textbf{unconditionally convergent uniformly} in $j\geq 0$ if for every $\delta >0$, there exists $N\in \mathbb{N}$ such that $\|\sum_{n\in F} x_{n,j}\|<\delta$ for all finite sets $F\subset [N,\infty)$ and $j$. We will make use of the following inequality in our subsequent work: let $(\lambda_n)$ be a scalar sequence and $\sum_{n\geq 1} x_n$ a series in a Banach space. Then if $F\subset \mathbb{N}$ is finite, we have
\begin{equation}\label{qq}
\displaystyle \left \|\sum_{n\in F} \lambda_n x_n \right \| \leq 4~\sup_{n\in F}|\lambda_n| \sup_{G\subseteq F}\left \|\sum_{n\in G} x_n \right \|,
\end{equation}
cf. \cite{Oho} (See also \cite{KG}).
\section{$q$-Frequent Hypercyclicity in $\mathcal{K}(X)$,
$\mathcal{L}(X)$ and $\mathcal{S}(H)$}
In this section, we first obtain a sufficient criterion for the
$q$-frequent hypercyclicity of $C_{R,T}$ on the Banach algebras
$\mathcal{L}(X,Y)$ and $\mathcal{K}(X,Y)$, where $R\in
\mathcal{L}(Y)$ and $T\in \mathcal{L}(X)$. The next result is already known for $q=1$, cf. \cite{PE}, Remark 9.10. However, following the proof of the frequent hypercyclicity criterion given in Theorem 6.18 of \cite{BM}, we outline the proof for a given $q\in \mathbb{N}$.
\begin{thm} \textbf{$(q$-$FHC$ $Criterion)$} \label{1.3} Let $X$ be a separable $F$-space and $D$ be a dense set in $X$. If for each $x\in D$,
there exists a sequence $(x_n)_{n\geq 0}$ in $X$ such that $x_0=x$ and
\begin{itemize}
\item[(a)] $\sum_{n=0}^{r} T^{r^q-(r-n)^q}(x)$ converges unconditionally, uniformly in $r\geq 0$,
\item[(b)] $\sum_{n\geq0} x_{(n+r)^q-r^q}$ converges unconditionally, uniformly in $r\geq 0$; and
\item[(c)] $T^{n^q}x_{n^q}=x$, $T^{n^q}x_{m^q}=x_{m^q-n^q}$ for $m>n\geq 0$,
\end{itemize}
then $T$ is $q$-frequently hypercyclic on $X$.
\end{thm}
\begin{proof}
Without loss of generality, one may assume that $D$ is countable. Write $D=\{x_k:k\in \mathbb{N}\}$ and fix $(\epsilon_k)$ such that $k\epsilon_k+\sum_{j\geq k+1}\epsilon_j \rightarrow 0$. For each $x_k$, there exists a sequence $(x_{n,k})_{n\geq 0}$ with the conditions in the hypotheses being satisfied. By (a) and (b), it is possible to find an increasing sequence $(N_k)$ of natural numbers such that $\|\sum_{n\in F} T^{r^q-(r-n)^q}(x_i)\| <\epsilon_k$ for $F\subset [N_k,\infty)\cap \{1,...,r\}$, and $\|\sum_{n\in G} x_{(n+r)^q-r^q,i}\|<\epsilon_k$ for $G\subset [N_k,\infty)$, uniformly in $r\geq 0$, where $1\leq i\leq k$. By Lemma 6.19 of \cite{BM}, corresponding to $(N_k)$, we find a sequence $(J_k)$ of subsets of $\mathbb{N}$ such that $\underline{dens}(J_k)>0$, $\min (J_k)\geq k$ and $|m-n|\geq N_k+N_j$ for all $m\in J_k$, $n\in J_j$ and $m\neq n$. With these properties, the vector $x=\sum_{\ell \geq 1}\sum_{n\in J_\ell} x_{n^q,\ell}$ is a frequently hypercyclic vector for the sequence $(T^{n^q})$, and thus it is a $q$-frequently hypercyclic vector for $T$.
\end{proof}
Let $y\otimes x^*$ be the one-rank operator $x\rightarrow x^*(x)y$, where $y\in Y$
and $x^*\in X^*$ and $(\mathcal{I}(X,Y),\|.\|_{\mathcal{I}(X,Y)})$ be a
Banach space of operators from $X$ to $Y$ such that the set of
all finite-rank operators is $\|.\|_{\mathcal{I}(X,Y)}$-dense in
$\mathcal{I}(X,Y)$ and $\|y\otimes x^*\|_{\mathcal{I}(X,Y)}=\|y\|\|x^*\|$ for all $y\in Y$ and $x^*\in
X^*$. We have the following result concerning the
separability of $\mathcal{L}(X,Y)$ with respect to the topologies
SOT and COT, and of $\mathcal{K}(X,Y)$ in the operator norm topology.
\begin{prop}\label{P}
Let $X$ and $Y$ be separable Banach spaces. Then the following are true.\\
$(1)$ If $D$ is a countable dense subset of $Y$ and $\Phi$ is a
countable weak$^*$-dense subset of $X^*$, then the set
\begin{equation*}\label{gdphi}
\displaystyle \mathcal{G}_{D,\Phi}=\left \{\sum_{n=1}^{N}y_n\otimes
x_n^*: y_n\in D, x_n^*\in \Phi, N\in \mathbb{N} \right\}
\end{equation*}
is a countable SOT-dense subset of
$\mathcal{L}(X,Y)$.\\
$(2)$ If $X^*$ is separable and $\Phi$ is norm-dense, then the above
set $\mathcal{G}_{D,\Phi}$ is $\|.\|_{\mathcal{I}(X,Y)}$-dense in
$\mathcal{I}(X,Y)$.\\
$(3)$ Suppose that $X^*$ is separable and $Y$ has the AP. If $\Phi$
is norm-dense, then $\mathcal{G}_{D,\Phi}$ is norm-dense in
$\mathcal{K}(X,Y)$ and COT-dense in $\mathcal{L}(X,Y)$.
\end{prop}
\begin{proof}
The proof of $(1)$ is similar to the case of $X=Y$, proved in \cite{PE}, p. 277. Further, by the properties of $\mathcal{I}(X,Y)$ mentioned above, part $(2)$ follows since we can approximate every operator of finite rank by elements of $\mathcal{G}_{D,\Phi}$ in the norm $\|.\|_{\mathcal{I}(X,Y)}$. To get part $(3)$, let us assume that $Y$ has the AP. Then the space $\mathcal{F}(X,Y)$ of all finite-rank
operators is norm-dense in $\mathcal{K}(X,Y)$ for every Banach space $X$. Moreover, $\|v\otimes u^*\|_{op}=\|v\|\|u^*\|$ for all $v\in Y$ and $u^*\in X^*$.
\end{proof}
Using the above proposition, we prove
\begin{thm}\label{FHC}
Let $R\in \mathcal{L}(Y)$ and $T \in \mathcal{L}(X)$ for separable
Banach spaces $X$ and $Y$ and $q\in \mathbb{N}$. Let $\mathcal{D}$
be a norm-dense set in $Y$ and $\Phi$ be a countable weak*-dense set
in $X^*$. Suppose that for each $(y,x^*)\in \mathcal{D}\times \Phi$, there exist sequences $(y_n)_{n\geq 0}$ in $Y$ and
$(x^*_n)_{n\geq 0}$ in $X^*$ such that
\begin{itemize}
\item[\emph{(a)}] the series $\displaystyle \sum_{n=0}^{r} R^{r^q-(r-n)^q}(y)\otimes (T^*)^{r^q-(r-n)^q}(x^*)$ and
$\displaystyle \sum_{n=1}^{\infty} y_{(n+r)^q-r^q}\otimes x^*_{(n+r)^q-r^q}$
are unconditionally convergent in $(\mathcal{L}(X,Y),\|.\|_{op})$, uniformly in $r\geq 0$; and
\item[\emph{(b)}] $R^{n^q}y_{n^q}=y$, $(T^*)^{n^q}x^*_{n^q}=x^*$,
$R^{n^q}y_{m^q}=y_{m^q-n^q}$, and $(T^*)^{n^q}x^*_{m^q}=x^*_{m^q-n^q}$ for all $m>n\geq 0$.
\end{itemize}
Then the following assertions hold.\\ $\emph{(i)}$ If
$T^*(\Phi)\subseteq \Phi$, $\{x_n^*\}\subseteq \Phi$, then
$C_{R,T}$ is $q$-frequently hypercyclic on $(\mathcal{L}(X,Y),\emph{SOT})$.\\
$\emph{(ii)}$ If $Y$ has the AP and the set $\Phi$ is
norm-dense in $X^*$, then $C_{R,T}$ is $q$-frequently hypercyclic on
$\left(\mathcal{K}(X,Y),\|.\|_{op}\right)$ and
$\left(\mathcal{L}(X,Y),\emph{COT}\right)$.
\end{thm}
\begin{proof}
(i) As $Y$ is a separable Banach space, we may assume that $D$ is countable.
Let
\begin{equation*}
\displaystyle
\mathcal{L}_\Phi=\overline{\text{span}}^{\|.\|_{op}}\left \{y\otimes
x^*: y\in Y, x^*\in \Phi\right \},
\end{equation*}
where the closure is taken in the operator norm $\|.\|_{op}$. Then
$\mathcal{L}_\Phi$ is a separable Banach space since the set
\begin{center}
$\displaystyle \mathcal{G}_{D,\Phi}=\left \{ \sum_{n=1}^
{N}y_n\otimes x_n^*: y_n\in D, x_n^*\in \Phi, N\in
\mathbb{N}\right\}$
\end{center}
is countable and norm-dense in $\mathcal{L}_\Phi$.
If $G=\sum_{j\leq N} y_j\otimes x_k^*$, then $C_{R,T}(G)=\sum_{j\leq N} R(y_j)\otimes
T^*(x^*_j)$. As the map $C_{R,T}$ is continuous and
$T^*(\Phi)\subseteq \Phi$, it follows that $C_{R,T}$ takes
$\mathcal{L}_\Phi$ to itself.
To establish the $q$-frequent hypercyclicity of the operator
$C_{R,T}$ in SOT, we first show that $C_{R,T}$ satisfies the
conditions of Theorem \ref{1.3} in the space
$\mathcal{L}_\Phi$. Let $F=\sum_{j=1} ^{k} y_j\otimes x_j^*\in \mathcal{G}_{\mathcal{D},\Phi}$. For each $(y_j,x^*_j)$, let $(y_{j,n})$ and $(x^*_{j,n})$ be some sequences, respectively in $Y$ and $\Phi$, as in the hypothesis. Then
\begin{center}
$\displaystyle F_{n}=\sum_{j\leq k}y_{j,n}\otimes x^*_{j,n} \in \mathcal{L}_\Phi$, $n\geq 0$.
\end{center}
Therefore, by the assumption (a) of our theorem, both the series\\
$\sum_{n\leq r}(C_{R,T})^{r^q-(r-n)^q}(F)=\sum_{j\leq k} \sum_{n\leq r}R^{r^q-(r-n)^q}(y_j)\otimes (T^*)^{r^q-(r-n)^q}(x^*_j)$ and $\sum_{n\geq 0}F_{(n+r)^q-r^q}=\sum_{j\leq k}\sum_{n\geq 0}y_{j,(n+r)^q-r^q}\otimes x^*_{j,(n+r)^q-r^q} $ converge
unconditionally in $\mathcal{L}_\Phi$ with respect to the operator
norm, uniformly in $r\geq 0$.
Further, we have
\begin{align*}
\displaystyle (C_{R,T})^{n^q}F_{m^q}&=\sum_{j=1} ^{k}
R^{n^q}y_{j,m^q}\otimes
(T^*)^{n^q}x^*_{j,m^q}\\&=\begin{cases}
\sum_{j=1} ^{k} y_{j,m^q-n^q}\otimes x^*_{j,m^q-n^q},~~~~ m>n\\
~~F,~~~~~~~~m=n.
\end{cases}
\end{align*}
by hypotheses. Thus $(C_{R,T})^{n^q}F_{m^q}=F_{m^q-n^q}$ if $m>n$ and
$(C_{R,T})^{n^q}F_{n^q}=F, n\geq0$. So $C_{R,T}$ is $q$-frequently hypercyclic on $\mathcal{L}_\Phi$ with respect to the
operator norm topology. As $\mathcal{G}_{D,\Phi}$ is SOT-dense in $\mathcal{L}(X,Y)$ by
Proposition \ref{P}, the operator $C_{R,T}$ is $q$-frequently
hypercyclic on $(\mathcal{L}(X,Y),SOT)$. This establishes part (i).
Let us now prove part (ii). If $Y$ has the AP and the set $\Phi$ is
norm-dense in $X^*$, then by Proposition \ref{P}(3),
$\mathcal{G}_{D,\Phi}$ is dense in $\mathcal{K}(X,Y)$ with respect
to the operator norm and so $\mathcal{L}_\Phi=\mathcal{K}(X,Y)$. Consequently, $C_{R,T}$ is $q$-frequently hypercyclic
on $(\mathcal{K}(X,Y),\|.\|_{op})$ and $(\mathcal{L}(X,Y),\text{COT})$.
\end{proof}
For applications of Theorem \ref{FHC}, we require the following lemmas.
\begin{lem}\label{lomma1}
Let $X$ and $Y$ be Banach spaces. If $\displaystyle
\sum_{n=1}^{\infty} u_{n,j}$ is unconditionally convergent in $Y$, uniformly in $j\geq 0$, and $\{u^*_{n,k}\}\subset X^*$ is such that
$\{u^*_{n,k}: n\geq N_0, k\geq 1\}$ is norm-bounded for some $N_0\in \mathbb{N}$,
then $\sum_{n} u_{n,j}\otimes u^*_{n,k}$ is unconditionally convergent in
$(\mathcal{L}(X,Y),\|.\|_{op})$, uniformly in $j,k\in \mathbb{N}$.
\end{lem}
\begin{proof}
By definition, for a given $\epsilon>0$, one can choose a natural number $N>N_0$ such that
$\|\sum_{n\in F} u_{n,j} \| < \epsilon$ for every finite set set $F\subset [N,\infty)\cap \mathbb{N}$ and all $j\geq 1$.
Let $M\in \mathbb{R}$ be such that $\|u^*_{n,k}\|\leq M$ $\forall$ $n\geq N_0,k\geq 1$. Using the inequality \eqref{qq}, it follows that
\begin{center}
$\displaystyle \left \|\sum_{n\in F} u^*_{n,k}(x) u_{n,j} \right \| \leq 4M~\sup_{G\subseteq F} \left \|\sum_{n\in G} u_{n,j} \right \|<4M\epsilon$,
\end{center}
for $j,k\geq 1$ and $\|x\|\leq 1$. Thus
\begin{center}
$\displaystyle \left \|\sum_{n\in F} u_{n,j}\otimes u^*_{n,k} \right \|_{op}<4M\epsilon$.
\end{center}
\end{proof}
\begin{lem}\label{lomma2}
Let $X$ and $Y$ be Banach spaces. If $\displaystyle
\sum_{n=1}^{j} u_{n,j}$ is unconditionally convergent in $Y$, uniformly in $j\in \mathbb{N}$, and
$\{u^*_{n,j}: n\geq N_0, k\geq 1\}$ is norm-bounded in $X^*$ for some $N_0\in \mathbb{N}$,
then the series $\sum_{n}^{j} u_{n,j}\otimes u^*_{n,j}$ is unconditionally convergent in
$(\mathcal{L}(X,Y),\|.\|_{op})$, uniformly in $j\in \mathbb{N}$.
\end{lem}
Recalling the $q$-FHC Criterion from Theorem
\ref{1.3}, we prove the $q$-frequent hypercyclicity of the
left multiplication operator $\mathfrak{L}_R(S)=RS$. This strengthens a result of A. Bonilla and K.-G. Grosse-Erdmann \cite{BE} about the SOT-frequent
hypercyclicity of $\mathfrak{L}_R$.
\begin{cor} \label{left}
Let $X$ be a separable Banach space and $R\in \mathcal{L}(X)$
satisfy the $q$-FHC Criterion. Then the following hold.\\
$\emph{(i)}$ $\mathfrak{L}_R$ is $q$-frequently hypercyclic on $(\mathcal{L}(X),\text{SOT})$.\\
$\emph{(ii)}$ If $X^*$ is separable and $X$ has the AP, then the
$\mathfrak{L}_R$ is $q$-frequently hypercyclic on
$(\mathcal{K}(X),\|.\|_{op})$ and $(\mathcal{L}(X),\text{COT})$.
\end{cor}
\begin{proof}
In Theorem \ref{FHC}, let $X=Y$ and $T$ be the identity operator on
$X$. Since $X$ is separable, the dual $X^*$ is weak$^*$-separable. So, we can choose $\Phi$ to be any
countable weak$^*$-dense subset of $X^*$. Since $R$ satisfies
the $q$-FHC Criterion, we find a set $D$ satisfying the conditions (a) and (b) of
Theorem \ref{1.3}. That is, for $x\in D$, there exists $(x_n)$ in $X$ with $x_0=x$ such that
\begin{equation*}
\sum_{n\leq r} R^{r^q-(r-n)^q}(x)~and~\sum_{n\geq 1}x_{(n+r)^q-r^q}~are~unconditionally
~~convergent~uniformly~in~r\geq 0,
\end{equation*}
and
\begin{equation*}
R^{n^q}x_{n^q}=x,~ R^{n^q}x_{m^q}=x_{m^q-n^q},~~ m>n,~n\geq 0.
\end{equation*}
Now we verify the condition $(a)$ of Theorem
\ref{FHC}. For each $x^*\in \Phi$, let $x_n^*=x$, $n\geq 0$. By Lemmas \ref{lomma1} and \ref{lomma2} we get (i).
To see (ii), take $\Phi$ as any norm-dense subset of $X^*$ in
Theorem \ref{FHC} and apply Theorem \ref{FHC}(ii).
\end{proof}
Similarly, one can prove the following results. We observe that if $\sum_{n} x_{n,j}$ is unconditionally convergent, uniformly in $j$, then there exists $N\in \mathbb{N}$ such that the set
$\{x_{n,j}:n\geq N,j\geq 1\}$ is bounded.
\begin{cor} \label{right} Suppose $X$ is a separable Banach space and $T\in
\mathcal{L}(X)$. Then the following are true.\\
\emph{(1)} Let $\Phi$ be a countable weak$^*$-dense subset of $X^*$. Suppose that for each $x^*\in \Phi$,
there exists $(x_n^*)$ in $\Phi$ with properties that $x^*_0-x^*$, the series $\sum_{n\leq r}(T^*)^{r^q-(r-n)^q}(x^*)$ and $\sum_{n\geq 1} x^*_{(n+r)^q-r^q}$ are unconditionally convergent in $(X^*,\|.\|)$, uniformly in $r$; and
$(T^*)^{n^q}x^*_{n^q}=x^*$, $(T^*)^{n^q}x^*_{m^q}=x^*_{m^q-n^q}$, $m>n$. If $T^*(\Phi)\subset \Phi$, then $\mathfrak{R}_T$ is $q$-frequently
hypercyclic on $(\mathcal{L}(X),\text{SOT})$. \\
$\emph{(2)}$ Assume that $X^*$ is separable and $X$ has the AP. If
$T^*$ satisfies the $q$-FHC Criterion, then
$\mathfrak{R}_T$ is $q$-frequently hypercyclic on
$(\mathcal{K}(X),\|.\|_{op})$ and $(\mathcal{L}(X),\text{COT})$.
\end{cor}
\begin{prop}\label{corollaryiso}
Let $R,T\in \mathcal{L}(X)$, where $R$ satisfies the $q$-FHC Criterion, and let $\Phi$ be a countable weak$^*$-dense
set in $X^*$. If for each $f\in \Phi$, there exists a bounded $(f_n)$ in $X^*$ such that $f_0=0$ and the set $\{(T^*)^{n}(f):n\geq0\}$ is bounded; and $(T^*)^{n^q}f_{n^q}=f$, $(T^*)^{n^q}f_{m^q}=f_{m^q-n^q}$ for $m>n\geq 0$, then the following hold.\\ $(1)$ If $T^*(\Phi)\subseteq \Phi$, then $C_{R,T}$ is
$q$-frequently hypercyclic on $(\mathcal{L}(X),\text{SOT})$.\\
$(2)$ If $\Phi$ is norm-dense, and $X$ has the AP, then $C_{R,T}$ is
$q$-frequently hypercyclic on $(\mathcal{K}(X),\|.\|_{op})$ and
$(\mathcal{L}(X),\text{COT})$.
\end{prop}
Let us now establish the $q$-frequent hypercyclicity of $C_{R,U}$ for a unitary $U$ and of the conjugate
operator $C_R(S)=RSR^*$.
\begin{cor}\label{conjugate}
Let $R$ satisfy the $q$-FHC Criterion in a
separable Hilbert space $H$. Then the operators $C_R$ and $C_{R,U}$
are $q$-frequently hypercyclic on $(\mathcal{K}(H),\|.\|_{op})$ and
$(\mathcal{L}(H),\text{COT})$, where $U\in \mathcal{L}(H)$ is a
unitary operator.
\end{cor}
\begin{proof}
Since $H$ is a separable Hilbert space, it has the AP. As in the proof of the above results, one can find
a dense set $D$ of $H$ satisfying the conditions in Theorem
\ref{1.3}. Now, in Proposition \ref{corollaryiso} take
$T=R^*$ and $\Phi=D$. The result follows.
\end{proof}
We now proceed to some concrete applications of Theorem \ref{FHC}.
We provide sufficient conditions on the weights $(w_n)$ and
$(\mu_n)$ for the map $C_{B_w,F_\mu}$ to be $q$-frequently
hypercyclic on different Banach algebras of operators on $\ell^p$,
$1\leq p<\infty$, where the backward shift $B_w$ and the forward
shift $F_\mu$ are respectively given by $B_w(e_0)=0$,
$B_w(e_n)=w_ne_{n-1}$, $n\geq 1$ and $F_\mu (e_n)=\mu_{n+1}
e_{n+1}$, $n\geq 0$. Here $\{e_n\}_{n\geq 0}$ is the standard basis
in $\ell^p$.
\begin{prop}\label{unifhc}
If $\displaystyle \lim_{n\rightarrow
\infty}|w_1w_2..w_{(n+r)^q-r^q+i}\mu_1\mu_2..\mu_{(n+r)^q-r^q+j}|=\infty$, uniformly in $r\geq 0$, for all
$i,j\in \mathbb{N}_0$, then $C_{B_w,F_\mu}$ is $q$-frequently
hypercyclic on $(\mathcal{L}(\ell^1),\emph{SOT})$,
$(\mathcal{K}(\ell^p),\|.\|_{op})$ and
$(\mathcal{L}(\ell^p),\emph{COT})$, where $1<p<\infty$.
\end{prop}
\begin{proof}
In Theorem \ref{FHC}, let $X=Y=\ell^p$. To prove the result for
$(\mathcal{L}(\ell^1),\text{SOT})$, let us write $\Phi_0$ for the
linear span of $\{e^*_n:n\geq0\}$ over rationals in $\ell^\infty$
and $\mathcal{D}$ for $\text{span}\{e_n:n\geq0\}$ in $\ell^1$.
Consider the maps $S_w$ and $J_\mu$ given by $S_w(e_n)=\frac {1} {w_{n+1}}e_{n+1}$ and
$J_\mu(e^*_n)=\frac {1} {\mu_{n+1}} e^*_{n+1}$, $n\geq0$. Note that $B_wS_w$ and $F_\mu^*J_\mu$ are identity operators, and
\begin{equation}\label{plus}
\displaystyle S_w^m(e_n)=\frac {1}
{w_{n+1}w_{n+2}...w_{n+m}}e_{n+m}~~\text{and}~~\displaystyle
J_\mu^m(e^*_n)=\frac {1} {\mu_{n+1}\mu_{n+2}...\mu_{n+m}} e^*_{n+m}.
\end{equation}
Put
\begin{center}
$\displaystyle \Phi= \bigcup_{j\geq 0}\Phi_j$, where $\displaystyle
\Phi_{j+1}=\bigcup_{n,k\geq 0}(F_\mu^*)^nJ_\mu^k(\Phi_j)$, $j\geq
0$.
\end{center}
Then the set $\Phi$ is weak$^*$-dense in $\ell^\infty$ as
$\Phi_0\subseteq \Phi$. Clearly $\Phi$ is countable. Further
$F_\mu^*(\Phi)\subseteq \Phi$ and $J_\mu(\Phi)\subseteq \Phi$.
We only consider the series
\begin{center}
$\displaystyle \sum_{n=1}^{r} B^{r^q-(r-n)^q}_w(e_i)\otimes
{(F^*_\mu)}^{r^q-(r-n)^q}(e_j^*)$ and $\displaystyle\sum_{n=1}^{\infty}
S^{(n+r)^q-r^q}_w(e_i)\otimes J^{(n+r)^q-r^q}_\mu(e_j^*)$
\end{center}
for $i,j\geq 0$. As $B_w^n(e_i)=0$ for sufficiently large $n$ and $r^q-(r-n)^q\geq n$, the
first series converges unconditionally in the operator norm, uniformly in $r\geq 0$. For the latter series, it suffices to prove that $\displaystyle \sum_{n\geq 1} a_{n,r} e_{(n+r)^q-r^q+i}\otimes e^*_{(n+r)^q-r^q+j}$ converges unconditionally, uniformly in the operator norm
if $\lim_{n\rightarrow \infty} |a_{n,r}|=0$, uniformly in $r$. But this is immediate as, for $x=(x_n)\in \ell^p$, $1\leq p<\infty$, we have
\begin{center}
$\displaystyle \|\sum_{n\in F} a_{n,r} e_{(n+r)^q-r^q+i}\otimes
e_{(n+r)^q-r^q+j}^*\|=\left \|\sum_{n\in F} a_{n,r}
x_{(n+r)^q-r^q+j} e_{(n+r)^q-r^q+i} \right\|= \big (\sum_{n\in F} |a_{n,r}
x_{(n+r)^q-r^q+j}|^p \big)^{1/p}\leq \max_{n\in F}|a_{n,r}| \|x\|$.
\end{center}
Therefore $C_{B_w,F_\mu}$ is $q$-frequently hypercyclic on
$(\mathcal{L}(\ell^1),SOT)$ by Theorem \ref{FHC}(i).
Since $\ell^p$ has the AP and the set $\Phi$ constructed above is
norm-dense in $\ell^{p^\prime}$, where $1<p<\infty$,
$1/p+1/p^\prime=1$ and $\ell^{p^\prime}$ is the dual of $\ell^p$,
the operator $C_{B_w,F_\mu}$ is $q$-frequently hypercyclic on
$(\mathcal{K}(\ell^p),\|.\|_{op})$ and $(\mathcal{L}(\ell^p),COT)$
by Theorem \ref{FHC}(ii).
\end{proof}
\begin{xrem}
It is evident from the proof of the above proposition that the
result holds for any Banach sequence space $E$ with the AP such that
span $\{e_n:n\geq0\}$, span $\{e^*_n\}$ over rationals are
norm-dense in $E$, $E^*$ respectively and $\displaystyle
\sum_n(w_1w_2..w_{i+(n+r)^q-r^q}\mu_1\mu_2..\mu_{j+(n+r)^q-r^q})^{-1}e_{i+(n+r)^q-r^q}\otimes
e^*_{j+(n+r)^q-r^q}$ converges unconditionally in the operator norm, uniformly in $r$ for all
$i,j\geq 0$.
\end{xrem}
Our next aim is to obtain the bilateral version of Proposition
\ref{unifhc}. For $a=(a_n)_{n\in\mathbb{Z}}$ of a bounded sequence
of nonzero scalars, we define the bilateral backward shift $T_a$ on
the sequence space $\ell^p(\mathbb{Z})$, $1\leq p<\infty$, as
$T_a(e_n)=a_ne_{n-1}$ and the forward shift $S_a$ as
$S_a(e_n)=a_ne_{n+1}$, $n\in \mathbb{Z}$, where $\{e_n\}_{n\in
\mathbb{Z}}$ is the standard basis in $\ell^p(\mathbb{Z})$. Then we
have
\begin{prop}\label{bifhc}
Suppose that, for all $i,j\in \mathbb{Z}$, $\displaystyle \lim_{n\rightarrow
\infty}|a_1a_2..a_{(n+r)^q-r^q+i}b_1b_2..b_{(n+r)^q-r^q+j}|=\infty~~ and~~ \lim_{n\rightarrow \infty}
|a_ia_{i-1}..a_{i-r^q+(r-n)^q+1}b_jb_{j-1}..b_{j-r^q+(r-n)^q+1}|=0$, uniformly in ~$r\in \mathbb{N}_0$.
Then $C_{T_a,S_b}$ is $q$-frequently hypercyclic
on $(\mathcal{L}(\ell^1(\mathbb{Z})),SOT)$, $(\mathcal{K}(\ell^p(\mathbb{Z}),\|.\|_{op}))$ and
$(\mathcal{L}(\ell^p(\mathbb{Z})),COT)$, $1<p<\infty$.
\end{prop}
\begin{proof}
We apply Theorem \ref{FHC}. Choose
$X=Y=\ell^p(\mathbb{Z})$, $\mathcal{D}$=
span$\{e_n:n\in\mathbb{Z}\}$ and $\Phi_0=$
span$\{e^*_n:n\in\mathbb{Z}\}$ over rationals. Define the maps $S$
and $J$ as $S(e_n)=\frac {1} {a_{n+1}} e_{n+1}$ and $J(e_n^*)=\frac {1}
{b_{n+1}} e^*_{n+1}$ for $n\in \mathbb{Z}$. Let
\begin{center}
$\displaystyle \Phi= \bigcup_{j\geq 0}\Phi_j$, where $\displaystyle
\Phi_{j+1}=\bigcup_{n,k\geq 0}{S_b^*}^nJ^k(\Phi_j)$, $j\geq 0$.
\end{center}
The set $\mathcal{D}$ is norm-dense in $\ell^p(\mathbb{Z})$ for $p\in
[1,\infty)$, $\Phi$ is weak$^*$-dense in $\ell^\infty(\mathbb{Z})$ and norm-dense in
$\ell^{p^\prime}(\mathbb{Z})$, where $1/p+1/{p^\prime}=1$ and $1<
p<\infty$. Moreover $T_aS$ is the identity operator on $\mathcal{D}$
and $S_b^*J$ is the identity on $\Phi$. As in Proposition
\ref{unifhc}, $J(\Phi)\subseteq \Phi$ and $S_b^*(\Phi)\subseteq
\Phi$. Further for $n\geq 1$ and $i,j\in \mathbb{Z}$, we have
\begin{center}
~$T_a^n(e_i)=a_ia_{i-1}..a_{i-n+1}e_{i-n}$,~
$(S_b^*)^n(e_j^*)=b_jb_{j-1}..b_{j-n+1}e^*_{j-n}$,
\end{center}
\begin{center}
$S^n(e_i)=\frac {1} {a_{i+1}a_{i+2}..a_{i+n}} e_{i+n}$~ and~
$J^n(e_j^*)=\frac {1} {b_{j+1}b_{j+2}..b_{j+n}}e_{j+n}^*$.
\end{center}
Let $S_n=S^n$ and $J_n=J^n$. Proceeding as in the proof of
Proposition \ref{unifhc}, one can show that the series $\sum_{n\leq
r} T_a^{r^q-(r-n)^q}(e_i)\otimes (S_b^*)^{r^q-(r-n)^q}(e_j^*)$ and $\sum_{n\geq 1}
S_{(n+r)^q-r^q}(e_i)\otimes J_{(n+r)^q-r^q}(e_j^*)$ are unconditionally convergent
in the operator norm, uniformly in $r\geq 0$, by the hypothesis. This proves part (1).
Also since $\ell^p(\mathbb{Z})$ has the AP, and the above set $\Phi$
is norm-dense in $\ell^{p^\prime}(\mathbb{Z})$ ($1<p<\infty$ and
$\frac {1} {p}+\frac {1} {p^\prime}=1$), Theorem \ref{FHC}(ii)
yields (2).
\end{proof}
For the next result, let $\mathbb{C}^N$ be considered as a vector
space over $\mathbb{C}$, and for
$\lambda=(\lambda_1,...,\lambda_N)\in \mathbb{C}^N$, let $D_\mu$ be
the diagonal operator $D_\lambda(f_j)=\lambda_j f_j$ on
$\mathbb{C}^N$ with respect to the standard basis
$\{f_1,f_2,..,f_N\}$ of $\mathbb{C}^N$, where $N\in\mathbb{N}$. Then
we have
\begin{prop}Suppose that $|\lambda_j|\geq 1$ for each $1\leq j\leq N$.\\
$(1)$ If $\displaystyle \sum_{n\geq 1} \frac{1}
{|\mu_1\mu_2...\mu_{(n+r)^q-r^q+i}|}<\infty$ uniformly in $r$, for each $i\geq 0$, then
$C_{D_\lambda,F_\mu}$ is $q$-frequently
hypercyclic on $(\mathcal{L}(\ell^1,\mathbb{C}^N),\text{SOT})$.\\
$(2)$ If $1<p<\infty$ and $\displaystyle \sum_{n\geq 1} \frac {1}
{|\mu_1\mu_2...\mu_{(n+r)^q-r^q+i}|^p}<\infty$ uniformly in $r$, for all $i\geq 0$, then
$C_{D_\lambda,F_\mu}$ is $q$-frequently hypercyclic on
$(\mathcal{K}(\ell^p,\mathbb{C}^N),\|.\|_{op})$ and
$(\mathcal{L}(\ell^p,\mathbb{C}^N),COT)$.
\end{prop}
\begin{proof}
In Theorem \ref{FHC}, take $X=\ell^p$ and $Y=\mathbb{C}^N$. Consider
the set $\Phi$ and the maps $J_{n}$ as in the proof of Proposition
\ref{unifhc}. Choose $D=\mathbb{C}^N$ and $S_{n}=S_\lambda^{n}$,
where $S_\lambda (f_j)= \frac {1} {\lambda_j} f_j$. For a fixed
$i\geq 0$, the series $\sum_{n\leq r} D_\lambda^{r^q-(r-n)^q} (f_j)\otimes (F_\mu^*)^{r^q-(r-n)^q}(e_i)$ is clearly
unconditionally convergent in the operator norm, uniformly in $r$. Since
$S_\lambda ^n(f_j)= {\lambda_j^{-n}} f_j$ and $|\lambda _j|\geq 1$,
the set $\{S_{n}(f_j):n=0,1,2,...\}$ becomes bounded.
Consequently, the series $\sum_{n\geq 1} S_{(n+r)^q-r^q}(f_j) \otimes J_{(n+r)^q-r^q}(e_i)$
converges unconditionally, uniformly in $r$ by Lemma \ref{lomma1}.
\end{proof}
So far, we have considered applications of Theorem \ref{FHC} to
maps on Banach algebras of operators. Now we turn to the $q$-frequent
hypercyclicity of the conjugate operator $C_R(S)=RSR^*$ defined on the real
subspace $\mathcal{S}(H)$ of $\mathcal{L}(H)$, consisting of all
self-adjoint operators on a separable Hilbert space $H$. H. Petersson \cite{H} showed that if $R$ satisfies the Hypercyclicity Criterion, then $C_R$ is hypercyclic on the norm-closure of span $\{h\otimes h:h\in H\}$ over $\mathbb{R}$, and hence COT-hypercyclic on $\mathcal{S}(H)$. A standard application of
the $q$-FHC criterion, using Lemmas \ref{lomma1} and \ref{lomma2} yield the following:
\begin{prop} \label{FHCselfadjoint}
If $R\in \mathcal{L}(H)$ satisfies the $q$-FHC Criterion, then $C_R$ is $q$-frequently hypercyclic on
$(\mathcal{S}(H),\emph{COT})$.
\end{prop}
\section{$q$-Frequent Hypercyclicity in $S_p(H)$}
In this section we provide a sufficient condition for $C_{R,T}$ to be $q$-frequently hypercyclic on $S_p(H)$ for a separable Hilbert space $H$. Let us begin
with the $q$-frequent hypercyclicity of the left multiplication operator $\mathfrak{L}_R$ on $S_p(H)$, $1\leq p<\infty$, which is an easy application of the $q$-FHC Criterion.
\begin{prop} \label{left-schatten}
Suppose that $R$ is an operator satisfying the $q$-Frequent
Hypercyclicity Criterion in a Banach space $X$ with separable dual.\\
$(1)$ If $(\mathcal{I}(X),\|.\|_\mathcal{I})$ is a separable Banach
ideal in $\mathcal{L}(X)$ such that the finite rank operators on $X$
are $\|.\|_\mathcal{I}$-dense in $\mathcal{I}(X)$, and $\|x\otimes
x^*\|_\mathcal{I}=\|x\|\|x^*\|$ for all $x\in X$ and $x^*\in X^*$,
then the left multiplication operator $\mathfrak{L}_R$ is
$q$-frequently
hypercyclic on $\mathcal{I}(X)$.\\
$(2)$ If $X$ is a separable Hilbert space, then $\mathfrak{L}_R$ is
frequently hypercyclic on $S_p(X)$, $1\leq p<\infty$.
\end{prop}
\begin{proof}
The proof of (1) is omitted. To see that (2) is true, recall that
$\|u\otimes v\|_p=\|u\| \|v\|$ for all $u,v \in X$ and the finite rank operators on $X$ form a
dense subspace of $S_p(X)$.
\end{proof}
For the main theorem of this section, we state the following result on summability of a series in $S_p(H)$, cf. \cite{AC}, p. 152.
\begin{lem}\label{Lemma1}
Let $\displaystyle \{T_n\}_{n=1}^{\infty}\subset \mathcal{L}(H)$ be
such that $T_n^*T_m=T_nT_m^*=0$ whenever $m\neq n$. Then for $1\leq
p <\infty$
\begin{center}
$\|\sum_{n} T_n\|_p^p= \sum_{n} \|T_n\|_p^p$.
\end{center}
\end{lem}
We prove:
\begin{thm}\label{FHC-Schatten}
Let $1\leq p<\infty$, $R,T\in \mathcal{L}(H)$ and $D_1,D_2 \subset
H$. Let $D_1$ and $D_2$ both span dense subspaces of $H$. If for each $(x,y)\in
D_1\times D_2$, there exist sequences $(x_n,y_n)\in H\times H$ with $(x_0,y_0)=(x,y)$ and
\begin{itemize}
\item[\emph{(a)}] $\displaystyle \sum_{n=1}^{r} \|R^{r^q-(r-n)^q}(x)\|^p\|(T^*)^{r^q-(r-n)^q}(y)\|^p<\infty$ and $\displaystyle
\sum_{n=1}^{\infty} \|x_{(n+r)^q-r^q}\|^p\|y_{(n+r)^q-r^q}\|^p<\infty$ uniformly in $r\geq 0$,
\item[\emph{(b)}] $\big<R^{n}(x),R^{m}(x)\big>=\big<S_{n}(x),S_{m}(x)\big>=\big<(T^*)^{n}(y),(T^*)^{m}(y)\big>\\
=\big<J_{n}(y),J_{m}(y)\big>=0,$ for $m\neq n$; and
\item[\emph{(c)}]$R^{n^q}x_{n^q}=x$, $(T^*)^{n^q}y_{n^q}=y$, $R^{n^q}x_{m^q}=x_{m^q-n^q}$,
$(T^*)^{n^q}y_{m^q}=y_{m^q-n^q}$, $\forall m>n\geq 0$,
\end{itemize}
then $C_{R,T}$ is $q$-frequently hypercyclic on $(S_p(H),\|.\|_p)$.
\end{thm}
\begin{proof}
Let $\Delta=\text{span}\{x\otimes y: x\in D_1, y\in D_2\}$. Note
that $\Delta$ can also be written as the span of the set $\{x\otimes
y: x\in \text{span}D_1, y\in \text{span} D_2\}$. Since span$D_1$ and
span$D_2$ are dense in $H$, it can be proved that $\Delta$ is dense
in $S_p(H)$, $1\leq p <\infty$.
Let $F=\sum_{k=1}^{N} a_k x_k\otimes y_k \in \Delta$. Corresponding to $x_k$ and $y_k$, we obtain sequences $(x_{k,n})$ and $(y_{k,n})$ as in the hypothesis.
Set $F_n=\sum_{k=1}^{N} a_k x_{k,n}\otimes y_{k,n}$. Consider the series $\sum_{n\leq r} C_{R,T}^{r^q-(r-n)^q}(F)=\sum_{k=1}^{N} a_k
\sum_{n=1}^{\infty} R^{r^q-(r-n)^q}(x_k)\otimes (T^*)^{r^q-(r-n)^q}(y_k)$ and
$\sum_n F_{(n+r)^q-r^q}=\sum_{k=1}^{N}a_k \sum_{n=1}^{\infty}
x_{k,(n+r)^q-r^q}\otimes y_{k,(n+r)^q-r^q}$. It suffices to prove that $\sum_{n\leq r}R^{r^q-(r-n)^q}(x)\otimes
(T^*)^{r^q-(r-n)^q}(y)$ and $\sum_n x_{k,(n+r)^q-r^q}\otimes y_{k,(n+r)^q-r^q}$ are unconditionally convergent in
$S_p(H)$, uniformly in $r$, for all $x\in D_1$ and $y\in D_2$.
Write $T_{n,r}=R^{r^q-(r-n)^q}(x)\otimes (T^*)^{r^q-(r-n)^q}(y)$, $n\geq 1$. Then
$T_{n,r}^*=(T^*)^{r^q-(r-n)^q}(y)\otimes R^{r^q-(r-n)^q}(x)$. If $<.>$ is the inner
product in $H$, then $T_{n,r}^*(z)=\big<z,R^{r^q-(r-n)^q}(x)\big>(T^*)^{r^q-(r-n)^q}(y)$
and $T_{m,r}(z)=\big<z,(T^*)^{r^q-(r-m)^q}(y)\big>R^{r^q-(r-m)^q}(x)$ and so
\begin{center}
$T_{n,r}^*T_{m,r}(z)=\big<z,(T^*)^{r^q-(r-m)^q}(y)\big>\big<R^{r^q-(r-m)^q}x,R^{r^q-(r-n)^q}x\big>(T^*)^{r^q-(r-n)^q}(y).$
\end{center}
Similarly
\begin{center}
$T_{n,r}T_{m,r}^*(z)=\big<z,R^{r^q-(r-m)^q}x\big>\big<(T^*)^{r^q-(r-m)^q}y,(T^*)^{r^q-(r-n)^q}y\big>R^{r^q-(r-n)^q}x$.
\end{center}
From part $(b)$ in the hypotheses, we get $T^*_{n,r}T_{m,r}=T_{n,r}T_{m,r}^*=0$,
$m\neq n$. Since $\|u\otimes v\|_p=\|u\|\|v\|$ for all $u,v\in H$,
Lemma \ref{Lemma1} and the hypothesis $(a)$ yield that
$\sum_{n\leq r}R^{r^q-(r-n)^q}(x)\otimes (T^*)^{r^q-(r-n)^q}(y)$ is unconditionally convergent in
$S_p(H)$, uniformly in $r$. Similarly, one can obtain that $\sum_n
x_{k,(n+r)^q-r^q}\otimes y_{k,(n+r)^q-r^q}$ is unconditionally convergent in $S_p(H)$, uniformly in $r$.
Thus the condition $(a)$ of Theorem \ref{FHC} is
satisfied by $C_{R,T}$ in $S_p(H)$.
\end{proof}
Next, from Theorem \ref{FHC-Schatten}, we obtain conditions on the
weight sequences $(w_n)$ and $(\mu_n)$ that are sufficient for the
$C_{B_w,F_\mu}$ on $S_p(\ell^2)$ and $C_{T_a,S_b}$ on
$S_p(\ell^2(\mathbb{Z}))$ to be $q$-frequently hypercyclic, $1\leq
p<\infty$.
\begin{prop}\label{unifhcS}
$(1)$ If $\displaystyle
\sum_{n=1}^{\infty}|w_1w_2..w_{(n+r)^q-r^q+i}\mu_1\mu_2..\mu_{(n+r)^q-r^q+j}|^{-p}<\infty$ uniformly in $r\geq 0$
for all $i,j\geq0$, then $C_{B_w,F_\mu}$ is $q$-frequently hypercyclic
on $S_p(\ell^2)$.\\
$(2)$ If $\displaystyle \sum_{n=1}^{\infty}|a_1a_2..a_{(n+r)^q-r^q+i}b_1b_2..b_{(n+r)^q-r^q+j}|^{-p}<\infty$ and\\ $\displaystyle
\sum_{n=0}^{r}{|a_ia_{i-1}..a_{i-r^q+(r-n)^q+1}b_jb_{j-1}..b_{j-r^q+(r-n)^q+1}
|^p}<\infty$ uniformly in $r\geq 0$ for all $i,j\in\mathbb{Z}$, then $C_{T_a,S_b}$ is
$q$-frequently hypercyclic on $S_p(\ell^2(\mathbb{Z}))$.
\end{prop}
\begin{proof}
To prove part (1), choose $D_1=D_2=\{e_n: n\geq 0\}$, the standard
orthonormal basis in $\ell^2$. Let $S_n$ and $J_n$ be the maps as
considered in the proof of Proposition \ref{unifhc}, i.e., $S_{n}(e_i)=S_w^{n}(e_i)=\frac {1}
{w_{i+1}..w_{i+n}} e_{i+n}$ and $J_{n}(e_j)=J_\mu^{n}(e_j)=\frac {1}
{\mu_{j+1}..\mu_{j+n}}e_{j+n}$. Note that $\big<S_n(e_j),S_m(e_j)\big>=\big<J_n(e_j),J_m(e_j)\big>=0$ for
$n\neq m$. As $B_w^n(e_i)=0$ for sufficiently large $n$, we have that\\ $\sum_{n\leq r} \|B_w^{r^q-(r-n)^q}(e_i)
\|^p\|(F_\mu^*)^{r^q-(r-n)^q}(e_j)\|^p<\infty$ uniformly in $r$. Moreover, from the
hypothesis, we have $\sum_n \|S_{(n+r)^q-r^q}(e_i)\|^p
\|J_{(n+r)^q-r^q}(e_j)\|^p<\infty$ uniformly in $r$. Now Theorem \ref{FHC-Schatten} yields
the result.
To prove part (2), consider the maps $S_n$ and $J_n$ in the proof of
Proposition \ref{bifhc} and choose $D_1=D_2=\{e_n:n\in
\mathbb{Z}\}$, the standard orthonormal basis in
$\ell^2(\mathbb{Z})$. Then, for $i,j\in \mathbb{Z}$, we have
\begin{equation*}
S_n(e_i)=S^n(e_i)=\frac {1} {a_{i+1}a_{i+2}..a_{i+n}}
e_{i+n}~~\text{and}~~ J_n(e_j)=J^n(e_j)=\frac {1}
{b_{j+1}b_{j+2}..b_{j+n}}e_{j+n},
\end{equation*}
and
\begin{equation*}
T_a^n(e_i)=a_ia_{i-1}...a_{i-n}
e_{i-n-1}~~\text{and}~~(S_b^*)^n(e_j)=b_jb_{j-1}...b_{j-n}
e_{j-n-1}.
\end{equation*}
It follows by the hypotheses that the series $\displaystyle \sum_{n\geq 1} \|S_{(n+r)^q-r^q}(e_i)\|^p
\|J_{(n+r)^q-r^q}(e_j)\|^p$
as well as\\ $\displaystyle\sum_{n\leq r} \|T_a^{r^q-(r-n)^q}(e_i)\|^p
\|(S_b^*)^{r^q-(r-n)^q}(e_j)\|^p $ converges uniformly $r\geq 0$.
\end{proof}
\section{Frequent Hypercyclicity in $S_p(H^2(\mathbb{D}))$ and $\mathcal{N}(\ell^p)$}
This section includes results on frequent hypercyclicity of specific
operators of the form $C_{R,T}$ defined on the $p$th Schatten von-Neumann class of operators on the Hardy space $H^2(\mathbb{D})$,
$1\leq p<\infty$, as well as on the space $\mathcal{N}(\ell^p)$ of
all nuclear operators on $\ell^p$, $1<p<\infty$. Let us recall the \textbf{Eigenvalue Criterion}, due to S. Grivaux.
\begin{prop} \cite{Grivaux} \label{Grivaux}
Let $X$ be a separable, complex Banach space and $T\in \mathcal{L}(X)$. If for every countable subset $D$ of the unit circle $S^1$, the set $\bigcup_{\alpha \in S^1\setminus D} Ker(T-\alpha I)$ spans a dense subspace of $X$, then $T$ is frequently hypercyclic.
\end{prop}
Corresponding to a sequence $\beta=(\beta_n)$,
$\beta_n>0$, $n\geq 0$, let $\big(H^\beta(\mathbb{D}),<.>\big)$ be a
Hilbert space of complex functions, analytic on the open unit disc
$\mathbb{D}$ such that the evaluation mappings $f\rightarrow f(z)$
are continuous at each $z\in \mathbb{D}$, i.e. there exists $k_z\in
H^\beta(\mathbb{D})$ such that $f(z)=\big<f,k_z\big>$ for each $f\in
H^\beta(\mathbb{D})$. Such a function $k_z$ is called \textbf{a reproducing
kernel} at $z\in \mathbb{D}$. Also, assume that
$\{e_n\}_{n=0}^{\infty}$ forms an orthonormal basis for
$H^\beta(\mathbb{D})$, where $e_n(z)=\beta_n z^n$. Note that when
$\beta_n=1$ for all $n\geq 0$, we have the Hardy space
$H^2(\mathbb{D})$.
Let $M_\varphi$ be the multiplication operator $f(z)\rightarrow
\varphi(z) f(z)$ on $H^\beta(\mathbb{D})$, corresponding to an
analytic function $\varphi$ on $\mathbb{D}$ and $M^*_\varphi$ be the
Hilbert space adjoint of $M_\varphi$. Our aim is to establish the
frequent hypercyclicity of $C_{M^*_\varphi,M_\psi}$ on
$S_p(H^\beta(\mathbb{D}))$, where $M_\varphi$ and $M_\psi$ are
bounded multiplication operators on $H^\beta(\mathbb{D})$
corresponding to the analytic functions $\varphi$ and $\psi$ on
$\mathbb{D}$. Let us first prove
\begin{lem}\label{lem}
Let $\varphi$ and $\psi$ be non-zero analytic functions on
$\mathbb{D}$ such that at least one of them is non-constant and
$|\varphi(z)\psi(w)|=1$ for some $z,w \in \mathbb{D}$. Then
\begin{center}
\emph{span} $\{k_z\otimes k_w: \overline{\varphi(z)}\psi(w)\in
S^1\setminus D\}$
\end{center}
is dense in the space $S_1(H^\beta(\mathbb{D}))$ for every countable
set $D\subset S^1$.
\end{lem}
\begin{proof}
We write $\mathcal{H}=H^\beta(\mathbb{D})$ and
$\overline{\varphi(\mathbb{D})}=\{\overline{\varphi(z)}:z\in
\mathbb{D}\}$. By the open mapping theorem for analytic functions,
the set
$\overline{\varphi(\mathbb{D})}\psi(\mathbb{D})=\{\overline{\varphi(z)}\psi(w):
z,w\in \mathbb{D}\}=\bigcup_{z\in
\mathbb{D}}(\overline{\varphi(z)}\psi(\mathbb{D}))$ is non-empty and
open. Hence there exists an open arc $\Gamma$ in $S^1$ such that
$\Gamma \subset \overline {\varphi(\mathbb{D})}\psi(\mathbb{D})$; let us assume that
this arc $\Gamma$ is the maximal one.
Consider the set
\begin{center}
$U\times V=\{(z,w)\in \mathbb{D}\times \mathbb{D}:
\overline{\varphi(z)}\psi(w)\in \Gamma\setminus D\}$.
\end{center}
We claim that $U$ is uncountable, and for each $z\in U$, there
exists an uncountable set $V_1\subseteq V$ such that
\begin{equation} \label{kuthira}
\overline{\varphi(z)}\psi(w) \in \Gamma\setminus D,
~~\text{for}~\text{all}~~ w\in V_1.
\end{equation}
To prove this, assume that both $\varphi$ and $\psi$ are
non-constant. In this case, $\overline{\varphi(\mathbb{D})}$ is a
non-empty open set. If $|\overline{\phi(z)} \psi(w)|=1$ for some
$(z,w)\in \mathbb{D}\times \mathbb{D}$, then
$\psi(w)\overline{\varphi(\mathbb{D})}$ is non-empty and open and
so, we can find an arc $\Gamma_1 \subset \Gamma$ such that
$\Gamma_1\subset \psi(w)\overline{\varphi(\mathbb{D})}$ as above.
Since $D$ is countable and $\Gamma_1 \setminus D \subset
\psi(w)\overline {\varphi(\mathbb{D})}$, the set $U$ has to be
uncountable. Now fix $z\in U$. Then the set
$\overline{\varphi(z)}\psi(\mathbb{D})$, being non-empty and open,
contains $\Gamma_2 \setminus D$ for some sub-arc $\Gamma_2$ of
$\Gamma$. This proves the claim when $\varphi$ and $\psi$ are non-constant.
Now assume that $\psi$ is constant, say $\psi=c$ and $\varphi$ is
non-constant. We can proceed as above to prove that the set $U$ is
uncountable since $c\overline{\varphi(\mathbb{D})}$ is a non-empty
open set containing $\Gamma \setminus D$. Also, since $\psi$ is
constant, we can take $V_1=V=\mathbb{D}$. Finally when $\varphi$ is
constant and $\psi$ is non-constant, we proceed similarly to get the
result. Hence our claim is established.
We now show that $\Lambda=$ span$\{k_z\otimes k_w: z\in U, w\in V\}$
is dense in the space $S_1(H)$. Recall that the trace of $A\in S_1(\mathcal{H})$ is given by
$tr(A)=\sum_{n\geq0}\big<Ae_n,e_n\big>$, where $e_n(z)=\beta_nz^n$, $n\geq 0$. Also we have $S_1(H)^*=\mathcal{L}(H)$
with respect to the duality-pairing $(A,T)=tr(AT)$, $T\in
S_1(H)$ and $A\in \mathcal{L}(H)$.\\ Let $A\in \mathcal{L}(H)$ be such that $tr(AT)=0$ for all
$T\in\Lambda$. For $T=k_z\otimes k_w$, we have
$Te_n=\big<e_n,k_w\big>k_z=e_n(w)k_z=\beta_nw^nk_z$ and
\begin{align*}
tr(AT)&=\sum_{n\geq 0}\big<ATe_n,e_n\big>\\
&=\sum_{n\geq 0}\beta_n w^n \big<k_z,A^*e_n\big>\\
&=\sum_{n\geq 0}\beta_n\overline{(A^*e_n)(z)}w^n.
\end{align*}
\noindent Since $AT\in S_1(H)$, the above power series is well-defined for all
$z,w\in \mathbb{D}$. Hence it is an analytic function in the
variable $w$ for a fixed $z\in \mathbb{D}$. For $z\in U$, there exists an uncountable set $V_1$ such that
\eqref{kuthira} holds. Since $V_1$ is uncountable, it has a limit point in
$\mathbb{D}$. As $\beta_n>0$ for all $n\geq 0$ and $tr(AT)=0$, it
follows that the coefficients of
the above power series are all zero, i.e., $A^*(e_n)(z)=0$ for all $n\geq
0$. Similarly, since $z \in U$ is arbitrary and $U$ is uncountable,
we have $A^*(e_n)=0$, $\forall$ $n\geq 0$. As $\{e_n:n\geq 0\}$ spans a dense subspace of $\mathcal{H}$,
we conclude that $A=0$. Therefore the set $\Lambda$ is dense in $S_1(\mathcal{H})$. The
proof is now complete.
\end{proof}
The above lemma yields
\begin{thm}\label{ghardy}
Let $\varphi$ and $\psi$ be non-zero analytic functions on
$\mathbb{D}$ such that the corresponding multiplication operators
are bounded on $H^\beta(\mathbb{D})$. If one of the maps $\varphi$
and $\psi$ is non-constant and $|\varphi(z)\psi(w)|=1$ for some
$z,w\in \mathbb{D}$, then $C_{M^*_\varphi,M_\psi}$ is frequently
hypercyclic on $(S_p(H^\beta(\mathbb{D})),\|.\|_p)$,
$(\mathcal{K}(H^\beta(\mathbb{D})),\|.\|_{op})$, and
$((\mathcal{L}(H^\beta(\mathbb{D})),\text{COT}))$.
\end{thm}
\begin{proof}
For $z,w\in \mathbb{D}$, consider $k_z\otimes k_w\in S_1(H)$. Since $M^*_\varphi(k_z)=\overline{\varphi(z)} k_z$, we have
\begin{center}
$C_{M^*_\varphi,M_\psi}(k_z\otimes
k_w)=\overline{\varphi(z)}\psi(w) (k_z\otimes k_w)$.
\end{center}
Thus $k_z\otimes k_w$ is an eigen vector for
$C_{M^*_\varphi,M_\psi}$ corresponding to the eigen value
$\overline{\varphi(z)}\psi(w)$. Now by Lemma \ref{lem}, span
$\{k_z\otimes k_w:\overline{\varphi(z)}\psi(w)\in S^1\setminus D\}$
is dense in $S_1(H^\beta(\mathbb{D}))$ for any countable set
$D\subset S^1$. Hence $C_{M^*_\varphi,M_\psi}$ is frequently
hypercyclic on $(S_1(H^\beta(\mathbb{D})),\|.\|_1)$ by Proposition \ref{Grivaux}. Since the embeddings
$S_1(H^\beta(\mathbb{D}))\hookrightarrow S_p(H^\beta(\mathbb{D}))
\hookrightarrow \mathcal{K}(H^\beta(\mathbb{D})) \hookrightarrow
(\mathcal{L}(H^\beta(\mathbb{D})),COT)$ are continuous and have
dense range, it follows that $C_{M^*_\varphi,M_\psi}$ is frequently
hypercyclic on each of the these spaces.
\end{proof}
As noted in the beginning of this section, the Hardy space
$H^2(\mathbb{D})$ is a special case of $H^\beta(\mathbb{D})$. In
this case, we have the following characterization for
$C_{M^*_\varphi,M_\psi}$ on $S_p(H^2(\mathbb{D}))$ and
$\mathcal{K}(H^2(\mathbb{D}))$.
\begin{thm}\label{hardy}
Let $\varphi$ and $\psi$ be non-zero, bounded and analytic on
$\mathbb{D}$, with one of them being non-constant. Then
$C_{M^*_\varphi,M_\psi}$ is frequently hypercyclic on
$(S_p(H^2(\mathbb{D})),\|.\|_p)$,
$(\mathcal{K}(H^2(\mathbb{D})),\|.\|_{op})$ and
$(\mathcal{L}(H^2(\mathbb{D})),COT)$ if $|\varphi(z)\psi(w)|=1$ for
some $z,w\in \mathbb{D}$. Conversely, if $C_{M^*_\varphi,M_\psi}$ is
frequently hypercyclic on
$(\mathcal{K}(H^2(\mathbb{D})),\|.\|_{op})$ or
$(S_p(H^2(\mathbb{D})),\|.\|_{p})$, then $|\varphi(z)\psi(w)|=1$ for
some $z,w\in \mathbb{D}$.
\end{thm}
\begin{proof}
We know that $M_\varphi \in
\mathcal{L}(H^2(\mathbb{D}))$ if and only if $\varphi$ is bounded on
$\mathbb{D}$. Thus by the preceding theorem, if $|\varphi(z)\psi(w)|=1$ for some $z,w \in \mathbb{D}$, then
$C_{M^*_\varphi,M_\psi}$ is frequently hypercyclic.
Let $G=\varphi(\mathbb{D})\psi(\mathbb{D})$.
Assume that the converse is not true, i.e., $G\cap S^1= \phi$. Then
$G\subseteq \mathbb{D}$ or $G\subseteq
\mathbb{C}\backslash\overline{\mathbb{D}}$ since $G$ is non-empty
and open, where $\overline{\mathbb{D}}$ is the closed unit disc in
$\mathbb{C}$. In case $G\subseteq \mathbb{D}$, then
$\|\varphi\|_{\infty}\|\psi\|_{\infty}=\sup_{z\in
\mathbb{D}}|\varphi(z)|\sup_{w\in \mathbb{D}}|\psi(w)|=\sup_{z,w \in
\mathbb{D}} |\varphi (z)\psi (w)|\leq 1$, and consequently,
\begin{center}
$\|C_{M_\varphi^*,M_\psi}\| \leq \|M^*_\varphi\|\|M_\psi\| \leq
\|\varphi\|_{\infty}\|\psi\|_{\infty} \leq 1$.
\end{center}
In the latter case when $G\subseteq
\mathbb{C}\setminus\overline{\mathbb{D}}$, we have $\displaystyle
\inf_{z,w\in \mathbb{D}}|\varphi(z)\psi(w)|\geq 1$, and so,
$\displaystyle \inf_{z\in \mathbb{D}} |\varphi(z)|>0$ and
$\displaystyle \inf_{w\in \mathbb{D}}|\psi(w)|>0$. Thus the
functions $\varphi^{-1}$ and $\psi^{-1}$ are bounded and analytic on
$\mathbb{D}$, and the corresponding operators $M_{\varphi^{-1}}$,
$M_{\psi^{-1}}$ are bounded on $H^2(\mathbb{D})$. Moreover
$(M_\varphi)^{-1}=M_{\varphi^{-1}}$ and
$(M^*_\varphi)^{-1}=M^*_{\varphi^{-1}}$. Then we observe that
\begin{center}
$\|{(C_{M_\varphi^*,M_\psi})}^{-1}\|=\|C_{M_{\varphi^{-1}}^*,M_{\psi^{-1}}}\|\leq
\|M^*_{\varphi^{-1}}\|\|M_{\psi^{-1}}\|\leq
\|\varphi^{-1}\|_{\infty}\|\psi^{-1}\|_{\infty}\leq 1$.
\end{center}
\noindent Thus, in both the cases, the operator $C_{M^*_\varphi, M_\psi}$ is not
hypercyclic on $\mathcal{K}(H^2(\mathbb{D}))$. This contradiction proves that $G\cap S^1\neq \phi$.
\end{proof}
As a special case, we state below the frequent hypercyclicity of the conjugate operator
$C_{M^*_\varphi}$ for the Hardy space $H=H^2(\mathbb{D})$.
\begin{prop}
If $\varphi$ is non-constant and $|\varphi(z)|=1$ for some $z\in
\mathbb{D}$, then the conjugate map $C_{M^*_\varphi}$ is frequently
hypercyclic on $(S_p(H),\|.\|_p)$, $(\mathcal{K}(H),\|.\|_{op})$ and
$(\mathcal{L}(H),COT)$. Conversely, if $C_{M^*_\varphi}$ is
frequently hypercyclic on $(\mathcal{K}(H),\|.\|_{op})$, then $\varphi$ is
non-constant and $\varphi(\mathbb{D})\cap S^1\neq \phi$.
\end{prop}
\begin{proof}
Immediate from the preceding theorem.
\end{proof}
The above characterization is not true for multiplication operators
on all Hilbert spaces of analytic functions, e.g. consider
\begin{exa}
Let $\displaystyle \mathcal{H}=\big \{ f(z)=\sum_{n\geq 0} a_n z^n:
\|f\|^2=\sum_{n\geq 0}(n+1)^2|a_n|^2<\infty \big \}$. Then $\mathcal{H}$ is
a reproducing kernel Hilbert space and the multiplication operator $M_\varphi$ corresponding to
$\varphi(z)=z$ acts as the shift
\begin{center}
$\displaystyle M_\varphi(e_n)(z)=ze_n(z)=\frac{1} {n+1} z^{n+1}=
\frac {n+2} {n+1} e_{n+1}(z)$
\end{center}
with respect to the orthonormal basis $e_n(z)=\frac {1} {n+1} z^n$, $n\geq 0$. Now $w_1w_2...w_n=n+1$ implies
that $\sum_n \frac {1} {(w_1w_2..w_n)^2} <\infty$ and consequently,
the adjoint $M^*_\varphi$ satisfies the FHC Criterion on $\mathcal{H}$. Hence by Corollary
\ref{conjugate}, $C_{M^*_\varphi}$ is frequently hypercyclic on the
spaces $(S_p(H),\|.\|_p)$, $(\mathcal{K}(\mathcal{H}),\|.\|_{op})$
and $(\mathcal{L}(\mathcal{H}),COT)$. However there are no $z,w\in
\mathbb{D}$ such that $|\varphi(z)\varphi(w)|=1$.
\end{exa}
In the spirit of Theorem \ref{ghardy}, let us prove a similar result
about $C_{\phi(B),\psi(F)}$ defined on spaces of operators on
$\ell^p$, $1<p<\infty$, where $\varphi(B)$ and $\varphi(F)$ are
functions of the unweighted backward and forward shifts
respectively. If $\varphi(z)=\sum_{n\geq 0} a_n z^n$ is an analytic
function on some neighborhood of the closed disc
$\overline{\mathbb{D}}$, then $\varphi(B)=\sum_n a_n B^n$ and
$\varphi(F)=\sum_n a_n F^n$ are bounded operators on $\ell^p$,
$1\leq p<\infty$. Moreover, the Banach space adjoint of $\varphi(B)$
is $\phi(F)$ and that of $\varphi(F)$ is $\varphi(B)$. Note that if $f_\lambda =(1,\lambda,\lambda^2,...)$ and $|\lambda|<1$, then $\varphi(B) f_\lambda =\varphi(\lambda )f_\lambda$. In \cite{DE}, R. Delaubenfels and H. Emamirad proved that if $\varphi(\mathbb{D})\cap S^1\neq \phi$, then $\varphi(B)$ is hypercyclic on $\ell^p$. We now have the following result, which can be proved using Proposition \ref{Grivaux}.
\begin{prop}
If $\varphi$ is non-constant and $\varphi(\mathbb{D})\cap S^1 \neq
\phi$, then $\varphi(B)$ is frequently hypercyclic on $\ell^p$,
$1\leq p<\infty$.
\end{prop}
Let $\mathcal{N}(\ell^p)$ denote the space of all nuclear operators on $\ell^p$. Then the trace $tr(T)=\sum_n
x_n^*(x_n)$ of $T=\sum_n x_n\otimes x_n^*\in \mathcal{N}(\ell^p)$,
$1<p<\infty$. Then the dual of $\mathcal{N}(\ell^p)$ is identified
with $\mathcal{L}(\ell^p)$ via the trace-duality $(S,T)=tr(TS)$, where $T\in \mathcal{N}(\ell^p)$ and $S\in
\mathcal{L}(\ell^p)$, cf. \cite{BST}, Theorem 16.50. The one-rank operator $f_\lambda \otimes f_\mu$ on $\ell^p$ is given
by $x\rightarrow f_\mu(x) f_\lambda$.
\begin{lem}\label{lemma-nuclear}
Let $\varphi$ and $\psi$ be non-zero functions analytic on some
neighborhoods of the closed disc $\overline{\mathbb{D}}$ with one of
them being non-constant and $|\varphi(\lambda)\psi(\mu)|=1$ for some
$\lambda,\mu \in \mathbb{D}$. Then
\begin{equation*}
\emph{span} \{f_\lambda \otimes f_\mu: \varphi(\lambda)\psi(\mu)\in
S^1\setminus D\}
\end{equation*}
is dense in $\mathcal{N}(\ell^p)$ for every countable set $D\subset
S^1$ and $1<p<\infty$.
\end{lem}
\begin{proof}
Invoking the proof of Lemma \ref{lem}, we can find an arc $\Gamma
\subseteq S^1\cap \varphi(\mathbb{D})\psi(\mathbb{D})$. Write
$U\times V=\{(\lambda,\mu)\in \mathbb{D}\times \mathbb{D}:
\varphi(\lambda)\psi(\mu)\in \Gamma\setminus D\}$. Then, for each
$\mu \in V$, there exists an uncountable set $U_1\subset \mathbb{D}$ such that
for $\lambda\in U_1$ we have $\varphi(\lambda)\psi(\mu) \in \Gamma
\setminus D$.
Let us now prove that $\Delta=$ span $\{f_\lambda \otimes f_\mu:
\lambda \in U, \mu \in V \}$ is dense in $\mathcal{N}(\ell^p)$. For this, let $S \in \mathcal{L}(\ell^p)$ such that $tr(TS)=0$ for all $T\in \Delta$. In particular, if $T=f_\lambda \otimes f_\mu$ for $\lambda \in U_1$, then $tr(f_\lambda \otimes S^*f_\mu)= (S^*f_\mu)(f_\lambda)=0$. Since $U_1$ has limit points in $\mathbb{D}$, we have that span$\{f_\lambda: \lambda \in U_1\}$ is dense in $\ell^p$. Thus $S^*(f_\mu)=0$ for all $\mu
\in V$. As span$\{f_\lambda: \mu \in V \}$ is dense in $\ell^{p^*}$,
the dual of $\ell^p$, $S=0$.
\end{proof}
Using the above lemma, we prove the frequent hypercyclicity of
$C_{\varphi(B),\psi(F)}$ as follows.
\begin{thm}\label{prop-lp}
Suppose $\phi$ and $\psi$ are non-zero analytic maps on some
neighborhoods of the closed disc $\overline{\mathbb{D}}$ such that
$|\phi(z)\psi(w)|=1$ for some $z,w\in \mathbb{D}$, with one of
$\varphi$ and $\psi$ being non-constant. Then $C_{\phi(B),\psi(F)}$
is frequently hypercyclic on $(\mathcal{N}(\ell^p),\|.\|_{nu})$,
$(\mathcal{K}(\ell^p),\|.\|_{op})$ and $(\mathcal{L}(\ell^p),COT)$
for $1<p<\infty$.
\end{thm}
\begin{proof}
Let $\lambda, \mu \in \mathbb{D}$. Since
$\phi(B)(f_\lambda)=\phi(\lambda)f_\lambda$, we get
\begin{center}
$C_{\phi(B),\psi(F)}(f_\lambda \otimes
f_\mu)=\varphi(B)(f_\lambda)\otimes \psi(B)(f_\mu)=\phi(\lambda)
\psi(\mu) (f_\lambda \otimes f_\mu)$,
\end{center}
where $f_\lambda=(1,\lambda,\lambda^2,...)$. From Lemma
\ref{lemma-nuclear}, it follows that span $\{f_\lambda \otimes f_\mu:
\varphi(\lambda)\psi(\mu)\in S^1\setminus D\}$ is dense in $\mathcal{N}(\ell^p)$ and the proof is complete by Proposition
\ref{Grivaux}.
\end{proof}
\normalsize
\baselineskip=17pt
\subsection*{Acknowledgements}
The second author acknowledges the Council of Scientific and
Industrial Research INDIA for a research fellowship.
|
\section{Tables with $k$-factors for the gluon-gluon channel}
\label{sec:append}
In this Appendix we document the numerical results for the comparisons
between the NNLO threshold approximation and the NNLO exact
calculation in the gluons-only channel. In the following tables we
show for each $p_T$ bin of each experiment in columns 2 and 3 the
experimental cross section together with its experimental uncertainty
computed as described in Sect.~\ref{sec:bench}. Additionally we give
NNLO/NLO gluons-only $k$-factors with both NNLO and NLO results
computed in the exact calculation (column 4) and in the threshold
approximation (column 5). The percentage wise relative difference
between the two is given in column 6. For completeness we give also
the NNLO threshold $k$-factor using the NLO exact calculation in the
denominator (column 7) and using the approximate NLO threshold
calculation in the denominator (column 8). Their percentage wise
relative difference is given in column 9.
\newpage
\subsection{CMS jets}
\vspace{-0.2cm}
\begin{table}[H]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$|\eta|<0.5$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 2.6410e+03 & 7.6605 & 1.1604 & 1.2129 & 4.52 & 1.1685 & 1.2129 & 3.80\tabularnewline
\hline
2 & 1.2183e+03 & 7.1542 & 1.1494 & 1.1932 & 3.81 & 1.1483 & 1.1932 & 3.91\tabularnewline
\hline
3 & 5.8988e+02 & 6.8006 & 1.1507 & 1.1771 & 2.29 & 1.1327 & 1.1771 & 3.92\tabularnewline
\hline
4 & 2.9063e+02 & 6.4464 & 1.1393 & 1.1649 & 2.25 & 1.1214 & 1.1649 & 3.88\tabularnewline
\hline
5 & 1.5159e+02 & 6.1977 & 1.1372 & 1.1545 & 1.52 & 1.1123 & 1.1545 & 3.79\tabularnewline
\hline
6 & 8.1069e+01 & 6.1166 & 1.1315 & 1.1453 & 1.22 & 1.1032 & 1.1453 & 3.82\tabularnewline
\hline
7 & 4.3880e+01 & 6.0166 & 1.1326 & 1.1406 & 0.71 & 1.0994 & 1.1406 & 3.75\tabularnewline
\hline
8 & 2.3974e+01 & 6.069 & 1.1486 & 1.1333 & -1.33 & 1.0931 & 1.1333 & 3.68\tabularnewline
\hline
9 & 1.3433e+01 & 5.8344 & 1.129 & 1.1295 & 0.04 & 1.0907 & 1.1295 & 3.56\tabularnewline
\hline
10 & 7.5741e+00 & 5.9214 & 1.1364 & 1.1253 & -0.98 & 1.0881 & 1.1253 & 3.42\tabularnewline
\hline
11 & 4.3578e+00 & 5.9103 & 1.1417 & 1.125 & -1.46 & 1.0872 & 1.125 & 3.48\tabularnewline
\hline
12 & 2.5573e+00 & 5.977 & 1.1356 & 1.122 & -1.20 & 1.0848 & 1.122 & 3.43\tabularnewline
\hline
13 & 1.4881e+00 & 6.0889 & 1.1383 & 1.1208 & -1.54 & 1.0866 & 1.1208 & 3.15\tabularnewline
\hline
14 & 8.5545e-01 & 6.2194 & 1.1269 & 1.1192 & -0.68 & 1.0856 & 1.1192 & 3.10\tabularnewline
\hline
15 & 5.0729e-01 & 6.3535 & 1.1374 & 1.119 & -1.62 & 1.0849 & 1.119 & 3.14\tabularnewline
\hline
16 & 3.0281e-01 & 6.5094 & 1.1345 & 1.1191 & -1.36 & 1.0885 & 1.1191 & 2.81\tabularnewline
\hline
17 & 1.7834e-01 & 6.685 & 1.1432 & 1.1193 & -2.09 & 1.0875 & 1.1193 & 2.92\tabularnewline
\hline
18 & 1.0597e-01 & 6.88 & 1.1398 & 1.1228 & -1.49 & 1.0909 & 1.1228 & 2.92\tabularnewline
\hline
19 & 6.2975e-02 & 7.1088 & 1.1458 & 1.1215 & -2.12 & 1.0915 & 1.1215 & 2.75\tabularnewline
\hline
20 & 3.7135e-02 & 7.4039 & 1.1518 & 1.1244 & -2.38 & 1.0939 & 1.1244 & 2.79\tabularnewline
\hline
21 & 2.1920e-02 & 7.7549 & 1.1466 & 1.1266 & -1.74 & 1.0981 & 1.1266 & 2.60\tabularnewline
\hline
22 & 1.2961e-02 & 8.0812 & 1.1445 & 1.129 & -1.35 & 1.1022 & 1.129 & 2.43\tabularnewline
\hline
23 & 7.4565e-03 & 8.5314 & 1.1682 & 1.1347 & -2.87 & 1.11 & 1.1347 & 2.23\tabularnewline
\hline
24 & 4.1735e-03 & 9.069 & 1.1489 & 1.1393 & -0.84 & 1.1093 & 1.1393 & 2.70\tabularnewline
\hline
25 & 2.3067e-03 & 9.7573 & 1.1701 & 1.1455 & -2.10 & 1.1219 & 1.1455 & 2.10\tabularnewline
\hline
26 & 1.4581e-03 & 10.5009 & 1.1676 & 1.1506 & -1.46 & 1.1312 & 1.1506 & 1.71\tabularnewline
\hline
27 & 7.9732e-04 & 11.6712 & 1.1657 & 1.1583 & -0.63 & 1.1371 & 1.1583 & 1.86\tabularnewline
\hline
28 & 3.3575e-04 & 13.8014 & 1.176 & 1.1651 & -0.93 & 1.1434 & 1.1651 & 1.90\tabularnewline
\hline
29 & 1.7796e-04 & 16.6186 & 1.1895 & 1.1746 & -1.25 & 1.1563 & 1.1746 & 1.58\tabularnewline
\hline
30 & 9.4376e-05 & 20.234 & 1.1972 & 1.1853 & -0.99 & 1.1703 & 1.1853 & 1.28\tabularnewline
\hline
31 & 4.3007e-05 & 24.8406 & 1.2009 & 1.1982 & -0.22 & 1.1879 & 1.1982 & 0.87\tabularnewline
\hline
32 & 1.6149e-05 & 31.2363 & 1.2171 & 1.2165 & -0.05 & 1.2071 & 1.2165 & 0.78\tabularnewline
\hline
33 & 2.0397e-06 & 74.2223 & 1.2492 & 1.2532 & 0.32 & 1.2508 & 1.2532 & 0.19\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 11.38\% & 15.6\% & 15.1\% & -0.5\% & 12\% & 15.1\% & 2.7\%\tabularnewline
\hline
\end{tabular}{\caption{\label{tab:kcms1} Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CMS 2011 $\sqrt{s}=7$ TeV dataset~\cite{Chatrchyan:2012bja} in the rapidity slice $|\eta|<0.5$.}
}
\end{table}
\begin{table}[h]
\begin{centering}
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.5<|\eta|<1.0$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta\%$\tabularnewline
\hline
\hline
1 & 2.5139e+03 & 7.5804 & 1.1542 & 1.2592 & 9.10 & 1.2159 & 1.2592 & 3.56\tabularnewline
\hline
2 & 1.1242e+03 & 7.2061 & 1.1388 & 1.2379 & 8.70 & 1.193 & 1.2379 & 3.76\tabularnewline
\hline
3 & 5.3237e+02 & 6.7691 & 1.149 & 1.2203 & 6.21 & 1.176 & 1.2203 & 3.77\tabularnewline
\hline
4 & 2.6596e+02 & 6.399 & 1.138 & 1.2079 & 6.14 & 1.1626 & 1.2079 & 3.90\tabularnewline
\hline
5 & 1.3774e+02 & 6.2308 & 1.1373 & 1.1955 & 5.12 & 1.1501 & 1.1955 & 3.95\tabularnewline
\hline
6 & 7.2171e+01 & 6.0829 & 1.131 & 1.1853 & 4.80 & 1.1406 & 1.1853 & 3.92\tabularnewline
\hline
7 & 3.8690e+01 & 6.054 & 1.1417 & 1.178 & 3.18 & 1.1321 & 1.178 & 4.05\tabularnewline
\hline
8 & 2.1313e+01 & 6.0825 & 1.1254 & 1.1722 & 4.16 & 1.1274 & 1.1722 & 3.97\tabularnewline
\hline
9 & 1.1977e+01 & 5.8952 & 1.1289 & 1.165 & 3.20 & 1.1207 & 1.165 & 3.95\tabularnewline
\hline
10 & 6.7334e+00 & 5.9827 & 1.1307 & 1.1593 & 2.53 & 1.1157 & 1.1593 & 3.91\tabularnewline
\hline
11 & 3.8290e+00 & 6.0043 & 1.1362 & 1.1575 & 1.87 & 1.113 & 1.1575 & 4.00\tabularnewline
\hline
12 & 2.2126e+00 & 6.1054 & 1.1361 & 1.1545 & 1.62 & 1.1104 & 1.1545 & 3.97\tabularnewline
\hline
13 & 1.2718e+00 & 6.2587 & 1.1404 & 1.1509 & 0.92 & 1.109 & 1.1509 & 3.78\tabularnewline
\hline
14 & 7.3168e-01 & 6.4106 & 1.1304 & 1.1502 & 1.75 & 1.1075 & 1.1502 & 3.86\tabularnewline
\hline
15 & 4.1588e-01 & 6.5276 & 1.1398 & 1.1476 & 0.68 & 1.104 & 1.1476 & 3.95\tabularnewline
\hline
16 & 2.4081e-01 & 6.738 & 1.126 & 1.1483 & 1.98 & 1.1037 & 1.1483 & 4.04\tabularnewline
\hline
17 & 1.4048e-01 & 6.9955 & 1.1478 & 1.1502 & 0.21 & 1.1089 & 1.1502 & 3.72\tabularnewline
\hline
18 & 8.1789e-02 & 7.2416 & 1.1614 & 1.1508 & -0.91 & 1.106 & 1.1508 & 4.05\tabularnewline
\hline
19 & 4.7098e-02 & 7.5424 & 1.1451 & 1.1519 & 0.59 & 1.1087 & 1.1519 & 3.90\tabularnewline
\hline
20 & 2.7347e-02 & 7.8865 & 1.1501 & 1.1525 & 0.21 & 1.11 & 1.1525 & 3.83\tabularnewline
\hline
21 & 1.5332e-02 & 8.2945 & 1.1652 & 1.1549 & -0.88 & 1.1169 & 1.1549 & 3.40\tabularnewline
\hline
22 & 8.2629e-03 & 8.7298 & 1.1592 & 1.1597 & 0.04 & 1.1205 & 1.1597 & 3.50\tabularnewline
\hline
23 & 4.3435e-03 & 9.3245 & 1.1578 & 1.1633 & 0.48 & 1.1248 & 1.1633 & 3.42\tabularnewline
\hline
24 & 2.3794e-03 & 10.0224 & 1.1859 & 1.1668 & -1.61 & 1.1308 & 1.1668 & 3.18\tabularnewline
\hline
25 & 1.4168e-03 & 10.8797 & 1.1695 & 1.1746 & 0.44 & 1.139 & 1.1746 & 3.13\tabularnewline
\hline
26 & 7.6208e-04 & 11.9879 & 1.1779 & 1.1803 & 0.20 & 1.1519 & 1.1803 & 2.47\tabularnewline
\hline
27 & 3.2068e-04 & 14.3091 & 1.1843 & 1.1898 & 0.46 & 1.164 & 1.1898 & 2.22\tabularnewline
\hline
28 & 1.2875e-04 & 18.0032 & 1.1881 & 1.1986 & 0.88 & 1.1732 & 1.1986 & 2.17\tabularnewline
\hline
29 & 7.1113e-05 & 23.8937 & 1.2097 & 1.21 & 0.02 & 1.1869 & 1.21 & 1.95\tabularnewline
\hline
30 & 1.5280e-05 & 24.9748 & 1.224 & 1.2273 & 0.27 & 1.2092 & 1.2273 & 1.50\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 9.08\% & 15.4\% & 17.7\% & 2.1\% & 13.8\% & 17.7\% & 3.5\%\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{\label{tab:kcms2}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CMS 2011 $\sqrt{s}=7$ TeV dataset~\cite{Chatrchyan:2012bja} in the rapidity slice $0.5<|\eta|<1.0$.}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.0<|\eta|<1.5$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta\%$\tabularnewline
\hline
\hline
1 & 2.1264e+03 & 8.5345 & 1.1464 & 1.3606 & 18.68 & 1.3193 & 1.3606 & 3.13\tabularnewline
\hline
2 & 9.4841e+02 & 7.9455 & 1.154 & 1.3343 & 15.62 & 1.2906 & 1.3343 & 3.39\tabularnewline
\hline
3 & 4.4306e+02 & 7.4404 & 1.1424 & 1.3102 & 14.69 & 1.2648 & 1.3102 & 3.59\tabularnewline
\hline
4 & 2.1618e+02 & 7.0173 & 1.1386 & 1.2929 & 13.55 & 1.2443 & 1.2929 & 3.91\tabularnewline
\hline
5 & 1.1088e+02 & 6.8155 & 1.1384 & 1.2802 & 12.46 & 1.23 & 1.2802 & 4.08\tabularnewline
\hline
6 & 5.8338e+01 & 6.6866 & 1.1465 & 1.267 & 10.51 & 1.2159 & 1.267 & 4.20\tabularnewline
\hline
7 & 3.0596e+01 & 6.6694 & 1.1275 & 1.2556 & 11.36 & 1.2032 & 1.2556 & 4.36\tabularnewline
\hline
8 & 1.5963e+01 & 6.7146 & 1.1548 & 1.2484 & 8.11 & 1.1951 & 1.2484 & 4.46\tabularnewline
\hline
9 & 8.7895e+00 & 6.5443 & 1.1409 & 1.2403 & 8.71 & 1.1867 & 1.2403 & 4.52\tabularnewline
\hline
10 & 4.8448e+00 & 6.61 & 1.1311 & 1.2355 & 9.23 & 1.1807 & 1.2355 & 4.64\tabularnewline
\hline
11 & 2.6745e+00 & 6.6891 & 1.128 & 1.2317 & 9.19 & 1.1747 & 1.2317 & 4.85\tabularnewline
\hline
12 & 1.4883e+00 & 6.8523 & 1.1442 & 1.2273 & 7.26 & 1.1677 & 1.2273 & 5.10\tabularnewline
\hline
13 & 8.1723e-01 & 7.0222 & 1.1385 & 1.2264 & 7.72 & 1.1708 & 1.2264 & 4.75\tabularnewline
\hline
14 & 4.5003e-01 & 7.2568 & 1.1408 & 1.2243 & 7.32 & 1.1691 & 1.2243 & 4.72\tabularnewline
\hline
15 & 2.5032e-01 & 7.4706 & 1.152 & 1.2231 & 6.17 & 1.1688 & 1.2231 & 4.65\tabularnewline
\hline
16 & 1.3720e-01 & 7.7755 & 1.1439 & 1.2243 & 7.03 & 1.1629 & 1.2243 & 5.28\tabularnewline
\hline
17 & 7.3806e-02 & 8.0491 & 1.1394 & 1.2248 & 7.50 & 1.1658 & 1.2248 & 5.06\tabularnewline
\hline
18 & 3.9477e-02 & 8.4148 & 1.1554 & 1.2269 & 6.19 & 1.17 & 1.2269 & 4.86\tabularnewline
\hline
19 & 2.0829e-02 & 8.8651 & 1.1597 & 1.2311 & 6.16 & 1.1772 & 1.2311 & 4.58\tabularnewline
\hline
20 & 1.0519e-02 & 9.3896 & 1.1663 & 1.2338 & 5.79 & 1.1742 & 1.2338 & 5.08\tabularnewline
\hline
21 & 5.2512e-03 & 9.9822 & 1.1661 & 1.2401 & 6.35 & 1.1848 & 1.2401 & 4.67\tabularnewline
\hline
22 & 2.6204e-03 & 10.7199 & 1.1669 & 1.2429 & 6.51 & 1.1969 & 1.2429 & 3.84\tabularnewline
\hline
23 & 1.2672e-03 & 11.7867 & 1.2002 & 1.2528 & 4.38 & 1.2041 & 1.2528 & 4.04\tabularnewline
\hline
24 & 5.3787e-04 & 13.4865 & 1.1923 & 1.2589 & 5.59 & 1.2036 & 1.2589 & 4.59\tabularnewline
\hline
25 & 2.6212e-04 & 15.7051 & 1.2056 & 1.2696 & 5.31 & 1.2149 & 1.2696 & 4.50\tabularnewline
\hline
26 & 1.4076e-04 & 20.3736 & 1.2039 & 1.2788 & 6.22 & 1.24 & 1.2788 & 3.13\tabularnewline
\hline
27 & 8.7777e-06 & 28.6887 & 1.2271 & 1.2983 & 5.80 & 1.2735 & 1.2983 & 1.95\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 9.6\% & 15.7\% & 25.7\% & 8.6\% & 20.6\% & 25.7\% & 4.3\%\tabularnewline
\hline
\end{tabular}\caption{\label{tab:kcms3}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CMS 2011 $\sqrt{s}=7$ TeV dataset~\cite{Chatrchyan:2012bja} in the rapidity slice $1.0<|\eta|<1.5$.}
\end{table}
\begin{table}[h]
\begin{centering}
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.5<|\eta|<2.0$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta\%$\tabularnewline
\hline
\hline
1 & 1.6383e+03 & 9.8283 & 1.1528 & 1.5131 & 31.25 & 1.4825 & 1.5131 & 2.06\tabularnewline
\hline
2 & 7.2786e+02 & 9.4308 & 1.1491 & 1.4836 & 29.11 & 1.444 & 1.4836 & 2.74\tabularnewline
\hline
3 & 3.3039e+02 & 9.2082 & 1.1511 & 1.4595 & 26.79 & 1.414 & 1.4595 & 3.22\tabularnewline
\hline
4 & 1.5735e+02 & 8.903 & 1.1492 & 1.4359 & 24.95 & 1.3865 & 1.4359 & 3.56\tabularnewline
\hline
5 & 7.8425e+01 & 8.987 & 1.1577 & 1.4236 & 22.97 & 1.3671 & 1.4236 & 4.13\tabularnewline
\hline
6 & 3.8482e+01 & 9.0353 & 1.1385 & 1.4091 & 23.77 & 1.3517 & 1.4091 & 4.25\tabularnewline
\hline
7 & 2.0336e+01 & 9.2203 & 1.1418 & 1.4005 & 22.66 & 1.337 & 1.4005 & 4.75\tabularnewline
\hline
8 & 1.0188e+01 & 9.5409 & 1.155 & 1.3918 & 20.50 & 1.3271 & 1.3918 & 4.88\tabularnewline
\hline
9 & 5.3698e+00 & 9.5504 & 1.1448 & 1.3903 & 21.44 & 1.3193 & 1.3903 & 5.38\tabularnewline
\hline
10 & 2.7229e+00 & 9.8758 & 1.1593 & 1.3885 & 19.77 & 1.3187 & 1.3885 & 5.29\tabularnewline
\hline
11 & 1.3901e+00 & 10.1572 & 1.1481 & 1.3876 & 20.86 & 1.3145 & 1.3876 & 5.56\tabularnewline
\hline
12 & 7.2674e-01 & 10.6264 & 1.1346 & 1.3882 & 22.35 & 1.3096 & 1.3882 & 6.00\tabularnewline
\hline
13 & 3.5895e-01 & 11.0987 & 1.1555 & 1.3901 & 20.30 & 1.3134 & 1.3901 & 5.84\tabularnewline
\hline
14 & 1.8085e-01 & 11.7454 & 1.1618 & 1.3921 & 19.82 & 1.3157 & 1.3921 & 5.81\tabularnewline
\hline
15 & 8.4817e-02 & 12.1609 & 1.1716 & 1.3991 & 19.42 & 1.3165 & 1.3991 & 6.27\tabularnewline
\hline
16 & 3.8799e-02 & 12.7936 & 1.1609 & 1.4051 & 21.04 & 1.3176 & 1.4051 & 6.64\tabularnewline
\hline
17 & 1.7787e-02 & 13.5213 & 1.1789 & 1.4144 & 19.98 & 1.3336 & 1.4144 & 6.06\tabularnewline
\hline
18 & 8.1318e-03 & 14.3923 & 1.1836 & 1.427 & 20.56 & 1.3349 & 1.427 & 6.90\tabularnewline
\hline
19 & 3.5729e-03 & 15.4557 & 1.1845 & 1.4395 & 21.53 & 1.36 & 1.4395 & 5.85\tabularnewline
\hline
20 & 1.4427e-03 & 16.8527 & 1.1866 & 1.4546 & 22.59 & 1.391 & 1.4546 & 4.57\tabularnewline
\hline
21 & 5.0721e-04 & 19.2402 & 1.2051 & 1.4808 & 22.88 & 1.4011 & 1.4808 & 5.69\tabularnewline
\hline
22 & 1.8119e-04 & 23.5419 & 1.2075 & 1.5142 & 25.40 & 1.4325 & 1.5142 & 5.70\tabularnewline
\hline
23 & 7.2472e-05 & 30.0012 & 1.2098 & 1.5437 & 27.60 & 1.5079 & 1.5437 & 2.37\tabularnewline
\hline
24 & 1.0486e-05 & 37.2034 & 1.2099 & 1.5484 & 27.98 & 1.5125 & 1.5484 & 2.37\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 13.85\% & 16.7\% & 43.7\% & 23.1\% & 37.1\% & 43.7\% & 4.8\%\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{\label{tab:kcms4}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CMS 2011 $\sqrt{s}=7$ TeV dataset~\cite{Chatrchyan:2012bja} in the rapidity slice $1.5<|\eta|<2.0$.}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.0<|\eta|<2.5$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta\%$\tabularnewline
\hline
\hline
1 & 1.1132e+03 & 14.1934 & 1.1456 & 1.7432 & 52.16 & 1.7273 & 1.7432 & 0.92\tabularnewline
\hline
2 & 4.7126e+02 & 14.1554 & 1.1492 & 1.7157 & 49.30 & 1.6835 & 1.7157 & 1.91\tabularnewline
\hline
3 & 2.1037e+02 & 14.375 & 1.1575 & 1.6903 & 46.03 & 1.6466 & 1.6903 & 2.65\tabularnewline
\hline
4 & 9.5500e+01 & 14.4644 & 1.1525 & 1.6788 & 45.67 & 1.6254 & 1.6788 & 3.29\tabularnewline
\hline
5 & 4.2179e+01 & 14.8869 & 1.159 & 1.6734 & 44.38 & 1.608 & 1.6734 & 4.07\tabularnewline
\hline
6 & 1.9811e+01 & 15.3918 & 1.165 & 1.6704 & 43.38 & 1.5994 & 1.6704 & 4.44\tabularnewline
\hline
7 & 9.1229e+00 & 16.0282 & 1.1427 & 1.6725 & 46.36 & 1.5931 & 1.6725 & 4.98\tabularnewline
\hline
8 & 3.9568e+00 & 16.9358 & 1.154 & 1.6787 & 45.47 & 1.5933 & 1.6787 & 5.36\tabularnewline
\hline
9 & 1.7580e+00 & 17.2414 & 1.1752 & 1.6929 & 44.05 & 1.5988 & 1.6929 & 5.89\tabularnewline
\hline
10 & 7.6681e-01 & 18.0971 & 1.1567 & 1.7049 & 47.39 & 1.6091 & 1.7049 & 5.95\tabularnewline
\hline
11 & 3.1497e-01 & 18.9133 & 1.1775 & 1.7329 & 47.17 & 1.6324 & 1.7329 & 6.16\tabularnewline
\hline
12 & 1.2120e-01 & 19.9428 & 1.1814 & 1.753 & 48.38 & 1.6388 & 1.753 & 6.97\tabularnewline
\hline
13 & 4.6228e-02 & 21.1426 & 1.203 & 1.7832 & 48.23 & 1.6904 & 1.7832 & 5.49\tabularnewline
\hline
14 & 1.6192e-02 & 22.6955 & 1.1958 & 1.8329 & 53.28 & 1.7136 & 1.8329 & 6.96\tabularnewline
\hline
15 & 5.3245e-03 & 23.3028 & 1.1988 & 1.9006 & 58.54 & 1.7871 & 1.9006 & 6.35\tabularnewline
\hline
16 & 1.6421e-03 & 25.1682 & 1.2098 & 2 & 65.32 & 1.9336 & 2 & 3.43\tabularnewline
\hline
17 & 3.2618e-04 & 28.8786 & 1.1516 & 2.109 & 83.14 & 2.0596 & 2.109 & 2.40\tabularnewline
\hline
18 & 9.2622e-05 & 36.9091 & 1.1519 & 2.0432 & 77.38 & 1.9141 & 2.0432 & 6.74\tabularnewline
\hline
19 & 8.1124e-06 & 52.9953 & 1.1825 & 1.7888 & 51.27 & 1.6879 & 1.7888 & 5.98\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 21.35\% & 16.9\% & 78.2\% & 52.5\% & 70.2\% & 78.2\% & 4.7\%\tabularnewline
\hline
\end{tabular}\caption{\label{tab:kcms5}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CMS 2011 $\sqrt{s}=7$ TeV dataset~\cite{Chatrchyan:2012bja} in the rapidity slice $2.0<|\eta|<2.5$.}
\end{table}
\clearpage
\subsection{ATLAS jets at $\sqrt{s}=7$ TeV}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$|\eta|<0.3$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 4.7475e+06 & 23.3268 & 1.1405 & 2.1444 & 88.02 & 2.176 & 2.1444 & -1.45\tabularnewline
\hline
2 & 7.2424e+05 & 18.5066 & 1.0848 & 1.8308 & 68.77 & 1.8118 & 1.8308 & 1.05\tabularnewline
\hline
3 & 1.4949e+05 & 13.9274 & 1.0779 & 1.6208 & 50.37 & 1.5817 & 1.6208 & 2.47\tabularnewline
\hline
4 & 3.8485e+04 & 12.0152 & 1.0383 & 1.4933 & 43.82 & 1.4502 & 1.4933 & 2.97\tabularnewline
\hline
5 & 8.6061e+03 & 12.261 & 1.0486 & 1.3936 & 32.90 & 1.3497 & 1.3936 & 3.25\tabularnewline
\hline
6 & 1.4949e+03 & 10.6946 & 1.0675 & 1.3075 & 22.48 & 1.2645 & 1.3075 & 3.40\tabularnewline
\hline
7 & 2.5400e+02 & 11.2422 & 1.0427 & 1.2411 & 19.03 & 1.1976 & 1.2411 & 3.63\tabularnewline
\hline
8 & 6.3400e+01 & 14.0325 & 1.0346 & 1.2048 & 16.45 & 1.1655 & 1.2048 & 3.37\tabularnewline
\hline
9 & 2.0700e+01 & 13.1113 & 1.0487 & 1.1798 & 12.50 & 1.1422 & 1.1798 & 3.29\tabularnewline
\hline
10 & 5.9600e+00 & 12.046 & 1.0438 & 1.1609 & 11.22 & 1.1278 & 1.1609 & 2.93\tabularnewline
\hline
11 & 1.3300e+00 & 12.7419 & 1.0634 & 1.1447 & 7.65 & 1.1077 & 1.1447 & 3.34\tabularnewline
\hline
12 & 3.4700e-01 & 14.693 & 1.0546 & 1.1308 & 7.23 & 1.1026 & 1.1308 & 2.56\tabularnewline
\hline
13 & 6.4400e-02 & 16.3403 & 1.0499 & 1.1258 & 7.23 & 1.0916 & 1.1258 & 3.13\tabularnewline
\hline
14 & 1.0100e-02 & 24.3375 & 1.056 & 1.1211 & 6.16 & 1.0934 & 1.1211 & 2.53\tabularnewline
\hline
15 & 1.1400e-03 & 43.0001 & 1.0774 & 1.1227 & 4.20 & 1.105 & 1.1227 & 1.60\tabularnewline
\hline
16 & 4.0000e-04 & 64.3242 & 1.0591 & 1.1327 & 6.95 & 1.1114 & 1.1327 & 1.92\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 19.8\% & 6.2\% & 33.5\% & 25.3\% & 30.5\% & 33.5\% & 2.5\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas71}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $|\eta|<0.3$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.3<|\eta|<0.8$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 4.7041e+06 & 23.275 & 1.0949 & 2.2051 & 101.40 & 2.2597 & 2.2051 & -2.42\tabularnewline
\hline
2 & 7.3737e+05 & 18.9232 & 1.0858 & 1.8855 & 73.65 & 1.882 & 1.8855 & 0.19\tabularnewline
\hline
3 & 1.5051e+05 & 14.1032 & 1.0638 & 1.6706 & 57.04 & 1.643 & 1.6706 & 1.68\tabularnewline
\hline
4 & 3.7475e+04 & 12.0853 & 1.0535 & 1.542 & 46.37 & 1.5039 & 1.542 & 2.53\tabularnewline
\hline
5 & 8.4343e+03 & 11.4745 & 1.0429 & 1.436 & 37.69 & 1.3944 & 1.436 & 2.98\tabularnewline
\hline
6 & 1.4545e+03 & 11.7405 & 1.0502 & 1.3483 & 28.39 & 1.3059 & 1.3483 & 3.25\tabularnewline
\hline
7 & 2.4343e+02 & 12.0632 & 1.0399 & 1.2758 & 22.68 & 1.2315 & 1.2758 & 3.60\tabularnewline
\hline
8 & 6.0600e+01 & 13.2759 & 1.0255 & 1.2361 & 20.54 & 1.1958 & 1.2361 & 3.37\tabularnewline
\hline
9 & 1.9600e+01 & 13.3945 & 1.0273 & 1.2129 & 18.07 & 1.168 & 1.2129 & 3.84\tabularnewline
\hline
10 & 5.6400e+00 & 14.0202 & 1.0355 & 1.1914 & 15.06 & 1.1495 & 1.1914 & 3.65\tabularnewline
\hline
11 & 1.1900e+00 & 14.0031 & 1.0427 & 1.1729 & 12.49 & 1.1413 & 1.1729 & 2.77\tabularnewline
\hline
12 & 3.1600e-01 & 15.1863 & 1.0344 & 1.159 & 12.05 & 1.1238 & 1.159 & 3.13\tabularnewline
\hline
13 & 6.6000e-02 & 17.3743 & 1.0413 & 1.1513 & 10.56 & 1.1176 & 1.1513 & 3.02\tabularnewline
\hline
14 & 7.8000e-03 & 23.7402 & 1.0486 & 1.1489 & 9.57 & 1.118 & 1.1489 & 2.76\tabularnewline
\hline
15 & 1.3800e-03 & 38.8769 & 1.0662 & 1.1505 & 7.91 & 1.1214 & 1.1505 & 2.59\tabularnewline
\hline
16 & 2.3500e-04 & 73.3069 & 1.0726 & 1.1595 & 8.10 & 1.1264 & 1.1595 & 2.94\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 20.4\% & 5.2\% & 37.2\% & 30.1\% & 34.3\% & 37.2\% & 2.5\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas72}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $0.3<|\eta|<0.8$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.8<|\eta|<1.2$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 4.0000e+06 & 27.1532 & 1.1068 & 2.3304 & 110.55 & 2.4438 & 2.3304 & -4.64\tabularnewline
\hline
2 & 6.7959e+05 & 20.6534 & 1.1082 & 2.0116 & 81.52 & 2.0448 & 2.0116 & -1.62\tabularnewline
\hline
3 & 1.4040e+05 & 15.5795 & 1.0721 & 1.7867 & 66.65 & 1.7808 & 1.7867 & 0.33\tabularnewline
\hline
4 & 3.4242e+04 & 12.8432 & 1.0481 & 1.6531 & 57.72 & 1.6305 & 1.6531 & 1.39\tabularnewline
\hline
5 & 7.6263e+03 & 12.4546 & 1.0456 & 1.5445 & 47.71 & 1.5136 & 1.5445 & 2.04\tabularnewline
\hline
6 & 1.2929e+03 & 12.3535 & 1.0497 & 1.4457 & 37.73 & 1.4051 & 1.4457 & 2.89\tabularnewline
\hline
7 & 2.0707e+02 & 12.6587 & 1.0332 & 1.367 & 32.31 & 1.3204 & 1.367 & 3.53\tabularnewline
\hline
8 & 5.0100e+01 & 13.8718 & 1.0586 & 1.323 & 24.98 & 1.2765 & 1.323 & 3.64\tabularnewline
\hline
9 & 1.5900e+01 & 14.1359 & 1.0352 & 1.2929 & 24.89 & 1.2424 & 1.2929 & 4.06\tabularnewline
\hline
10 & 4.3800e+00 & 14.2512 & 1.0362 & 1.2711 & 22.67 & 1.2224 & 1.2711 & 3.98\tabularnewline
\hline
11 & 9.1800e-01 & 15.507 & 1.0553 & 1.2485 & 18.31 & 1.2052 & 1.2485 & 3.59\tabularnewline
\hline
12 & 2.1400e-01 & 16.8993 & 1.055 & 1.2346 & 17.02 & 1.1773 & 1.2346 & 4.87\tabularnewline
\hline
13 & 4.0700e-02 & 20.2256 & 1.0547 & 1.2328 & 16.89 & 1.1807 & 1.2328 & 4.41\tabularnewline
\hline
14 & 2.7000e-03 & 32.8892 & 1.0711 & 1.232 & 15.02 & 1.1665 & 1.232 & 5.62\tabularnewline
\hline
15 & 4.6600e-04 & 69.3513 & 1.0877 & 1.2384 & 13.85 & 1.1774 & 1.2384 & 5.18\tabularnewline
\hline
16 & 1.0200e-04 & 113.3008 & 1.1179 & 1.2575 & 12.49 & 1.21 & 1.2575 & 3.93\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas73}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $0.8<|\eta|<1.2$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.2<|\eta|<2.1$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 3.8144e+06 & 28.3609 & 1.0859 & 2.6535 & 144.36 & 2.9287 & 2.6535 & -9.40\tabularnewline
\hline
2 & 5.7938e+05 & 23.1683 & 1.079 & 2.3145 & 114.50 & 2.4513 & 2.3145 & -5.58\tabularnewline
\hline
3 & 1.1122e+05 & 17.3887 & 1.0492 & 2.0782 & 98.07 & 2.1363 & 2.0782 & -2.72\tabularnewline
\hline
4 & 2.7245e+04 & 14.9045 & 1.0528 & 1.9272 & 83.05 & 1.9512 & 1.9272 & -1.23\tabularnewline
\hline
5 & 5.9592e+03 & 13.5442 & 1.0351 & 1.8033 & 74.22 & 1.7962 & 1.8033 & 0.40\tabularnewline
\hline
6 & 9.6768e+02 & 12.2931 & 1.036 & 1.688 & 62.93 & 1.6567 & 1.688 & 1.89\tabularnewline
\hline
7 & 1.5253e+02 & 13.6862 & 1.0465 & 1.585 & 51.46 & 1.5396 & 1.585 & 2.95\tabularnewline
\hline
8 & 3.4141e+01 & 14.5209 & 1.0276 & 1.533 & 49.18 & 1.4709 & 1.533 & 4.22\tabularnewline
\hline
9 & 9.6768e+00 & 15.9416 & 1.0315 & 1.4963 & 45.06 & 1.4328 & 1.4963 & 4.43\tabularnewline
\hline
10 & 2.3500e+00 & 16.5515 & 1.0449 & 1.4705 & 40.73 & 1.3991 & 1.4705 & 5.10\tabularnewline
\hline
11 & 3.9100e-01 & 18.4577 & 1.0499 & 1.4427 & 37.41 & 1.3634 & 1.4427 & 5.82\tabularnewline
\hline
12 & 7.2500e-02 & 21.7 & 1.0403 & 1.4169 & 36.20 & 1.3226 & 1.4169 & 7.13\tabularnewline
\hline
13 & 9.5100e-03 & 27.6168 & 1.0597 & 1.4036 & 32.45 & 1.3252 & 1.4036 & 5.92\tabularnewline
\hline
14 & 4.8600e-04 & 51.7804 & 1.1121 & 1.4055 & 26.38 & 1.3384 & 1.4055 & 5.01\tabularnewline
\hline
15 & 6.9300e-05 & 112.4636 & 1.1182 & 1.4388 & 28.67 & 1.347 & 1.4388 & 6.82\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 26.8\% & 5.8\% & 71.1\% & 61.6\% & 69.7\% & 71.1\% & 2.1\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas74}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $1.2<|\eta|<2.1$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.1<|\eta|<2.8$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 2.8750e+06 & 35.9715 & 1.0832 & 3.2222 & 197.47 & 3.8805 & 3.2222 & -16.96\tabularnewline
\hline
2 & 4.2396e+05 & 30.3757 & 1.0753 & 2.8748 & 167.35 & 3.2745 & 2.8748 & -12.21\tabularnewline
\hline
3 & 8.2474e+04 & 20.5697 & 1.0547 & 2.625 & 148.89 & 2.8555 & 2.625 & -8.07\tabularnewline
\hline
4 & 1.7629e+04 & 16.4221 & 1.0525 & 2.463 & 134.01 & 2.6046 & 2.463 & -5.44\tabularnewline
\hline
5 & 3.5102e+03 & 13.8806 & 1.0482 & 2.3323 & 122.51 & 2.4063 & 2.3323 & -3.08\tabularnewline
\hline
6 & 5.0816e+02 & 14.279 & 1.0436 & 2.2307 & 113.75 & 2.2373 & 2.2307 & -0.29\tabularnewline
\hline
7 & 5.9899e+01 & 16.5848 & 1.0417 & 2.1677 & 108.09 & 2.1112 & 2.1677 & 2.68\tabularnewline
\hline
8 & 1.0101e+01 & 19.9705 & 1.0456 & 2.1603 & 106.61 & 2.055 & 2.1603 & 5.12\tabularnewline
\hline
9 & 1.8687e+00 & 21.4024 & 1.046 & 2.1806 & 108.47 & 2.0411 & 2.1806 & 6.83\tabularnewline
\hline
10 & 2.7879e-01 & 27.9207 & 1.0446 & 2.2199 & 112.51 & 2.0491 & 2.2199 & 8.34\tabularnewline
\hline
11 & 1.6600e-02 & 37.8399 & 1.0667 & 2.3796 & 123.08 & 2.1761 & 2.3796 & 9.35\tabularnewline
\hline
12 & 3.1300e-04 & 96.5422 & 1.2447 & 2.7499 & 120.93 & 2.769 & 2.7499 & -0.69\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 29.3\% & 7.1\% & 146.7\% & 130.3\% & 153.8\% & 146.7\% & -1.2\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas75}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $2.1<|\eta|<2.8$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.8<|\eta|<3.6$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 2.0323e+06 & 44.9498 & 1.049 & 3.9296 & 274.60 & 5.1988 & 3.9296 & -24.41\tabularnewline
\hline
2 & 2.8404e+05 & 41.0901 & 1.0693 & 3.6019 & 236.85 & 4.4167 & 3.6019 & -18.45\tabularnewline
\hline
3 & 4.4737e+04 & 29.701 & 1.0441 & 3.3733 & 223.08 & 3.8837 & 3.3733 & -13.14\tabularnewline
\hline
4 & 8.9271e+03 & 22.2178 & 1.0492 & 3.2522 & 209.97 & 3.6012 & 3.2522 & -9.69\tabularnewline
\hline
5 & 1.3505e+03 & 22.0491 & 1.0557 & 3.2034 & 203.44 & 3.4039 & 3.2034 & -5.89\tabularnewline
\hline
6 & 1.1959e+02 & 22.3605 & 1.0341 & 3.2345 & 212.78 & 3.2729 & 3.2345 & -1.17\tabularnewline
\hline
7 & 5.6837e+00 & 33.596 & 0.9974 & 3.4406 & 244.96 & 3.2884 & 3.4406 & 4.63\tabularnewline
\hline
8 & 2.9490e-01 & 43.4456 & 1.1062 & 3.9922 & 260.89 & 3.7523 & 3.9922 & 6.39\tabularnewline
\hline
9 & 9.2626e-03 & 93.156 & 1.1287 & 5.659 & 401.37 & 4.8168 & 5.659 & 17.48\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 39.2\% & 5.9\% & 274.3\% & 251.9\% & 295.9\% & 274.3\% & -4.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas76}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $2.8<|\eta|<3.6$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$3.6<|\eta|<4.4$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 9.1000e+05 & 64.4667 & 1.035 & 4.9707 & 380.26 & 7.3006 & 4.9707 & -31.91\tabularnewline
\hline
2 & 9.4945e+04 & 67.1388 & 1.0528 & 4.802 & 356.12 & 6.401 & 4.802 & -24.98\tabularnewline
\hline
3 & 9.7033e+03 & 50.5717 & 1.03 & 4.8858 & 374.35 & 5.9346 & 4.8858 & -17.67\tabularnewline
\hline
4 & 1.0000e+03 & 40.8999 & 1.0079 & 5.1638 & 412.33 & 5.8942 & 5.1638 & -12.39\tabularnewline
\hline
5 & 5.8478e+01 & 38.009 & 1.0187 & 5.8867 & 477.86 & 6.2008 & 5.8867 & -5.07\tabularnewline
\hline
6 & 5.8913e-01 & 52.7968 & 0.9229 & 9.4063 & 919.21 & 8.9487 & 9.4063 & 5.11\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 52.3\% & 1.1\% & 485.2\% & 486.7\% & 578\% & 485.3\% & -14.5\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas77}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2010 $\sqrt{s}=7$ TeV dataset~\cite{Aad:2010ad} in the rapidity slice $3.6<|\eta|<4.4$.}
}
\end{table}
\clearpage
\subsection{ATLAS jets at $\sqrt{s}=2.76$ TeV}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$|\eta|<0.3$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.4864e+03 & 24.2727 & 1.0721 & 1.7853 & 66.52 & 1.7372 & 1.7853 & 2.77\tabularnewline
\hline
2 & 2.1374e+02 & 20.195 & 1.0639 & 1.5548 & 46.14 & 1.501 & 1.5548 & 3.58\tabularnewline
\hline
3 & 3.5667e+01 & 14.3507 & 1.0484 & 1.4088 & 34.38 & 1.3558 & 1.4088 & 3.91\tabularnewline
\hline
4 & 7.7000e+00 & 12.8376 & 1.0613 & 1.3275 & 25.08 & 1.2789 & 1.3275 & 3.80\tabularnewline
\hline
5 & 1.4206e+00 & 13.8038 & 1.0452 & 1.2638 & 20.91 & 1.2186 & 1.2638 & 3.71\tabularnewline
\hline
6 & 2.0276e-01 & 12.2458 & 1.0618 & 1.2108 & 14.03 & 1.1715 & 1.2108 & 3.35\tabularnewline
\hline
7 & 2.5778e-02 & 15.7923 & 1.0393 & 1.174 & 12.96 & 1.1431 & 1.174 & 2.70\tabularnewline
\hline
8 & 4.2737e-03 & 24.0877 & 1.0293 & 1.1584 & 12.54 & 1.1282 & 1.1584 & 2.68\tabularnewline
\hline
9 & 1.2081e-03 & 41.7325 & 1.06 & 1.1464 & 8.15 & 1.1147 & 1.1464 & 2.84\tabularnewline
\hline
10 & 4.3152e-04 & 48.5072 & 1.0478 & 1.1419 & 8.98 & 1.1144 & 1.1419 & 2.47\tabularnewline
\hline
11 & 7.1101e-05 & 116.4275 & 1.0627 & 1.1474 & 7.97 & 1.1292 & 1.1474 & 1.61\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 31.3\% & 5\% & 30.2\% & 23.4\% & 26.3\% & 30.2\% & 3.04\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas21}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $|\eta|<0.3$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.3<|\eta|<0.8$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.4818e+03 & 23.9425 & 1.0863 & 1.8582 & 71.06 & 1.8202 & 1.8582 & 2.09\tabularnewline
\hline
2 & 2.0736e+02 & 19.9388 & 1.0719 & 1.6216 & 51.28 & 1.5715 & 1.6216 & 3.19\tabularnewline
\hline
3 & 3.2989e+01 & 14.9857 & 1.0246 & 1.464 & 42.89 & 1.4122 & 1.464 & 3.67\tabularnewline
\hline
4 & 7.3589e+00 & 12.8073 & 1.0586 & 1.3779 & 30.16 & 1.3263 & 1.3779 & 3.89\tabularnewline
\hline
5 & 1.3990e+00 & 12.9025 & 1.0398 & 1.3104 & 26.02 & 1.2621 & 1.3104 & 3.83\tabularnewline
\hline
6 & 1.8980e-01 & 13.5593 & 1.0415 & 1.2539 & 20.39 & 1.2101 & 1.2539 & 3.62\tabularnewline
\hline
7 & 2.1495e-02 & 15.6309 & 1.0365 & 1.2075 & 16.50 & 1.1686 & 1.2075 & 3.33\tabularnewline
\hline
8 & 3.3768e-03 & 23.7035 & 1.0441 & 1.189 & 13.88 & 1.153 & 1.189 & 3.12\tabularnewline
\hline
9 & 9.8283e-04 & 35.6697 & 1.048 & 1.1799 & 12.59 & 1.1437 & 1.1799 & 3.17\tabularnewline
\hline
10 & 3.6303e-04 & 47.7796 & 1.0478 & 1.1774 & 12.37 & 1.1431 & 1.1774 & 3.00\tabularnewline
\hline
11 & 8.8899e-05 & 81.2158 & 1.052 & 1.1804 & 12.21 & 1.1443 & 1.1804 & 3.15\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 27.5\% & 5\% & 34.7\% & 28.1\% & 30.5\% & 34.7\% & 3.3\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas22}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $0.3<|\eta|<0.8$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.8<|\eta|<1.2$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.2854e+03 & 26.9324 & 1.0757 & 2.0263 & 88.37 & 2.0208 & 2.0263 & 0.27\tabularnewline
\hline
2 & 1.7380e+02 & 20.4538 & 1.0475 & 1.7743 & 69.38 & 1.7394 & 1.7743 & 2.01\tabularnewline
\hline
3 & 2.8617e+01 & 16.2391 & 1.0381 & 1.6024 & 54.36 & 1.5549 & 1.6024 & 3.05\tabularnewline
\hline
4 & 6.1844e+00 & 14.789 & 1.0356 & 1.5034 & 45.17 & 1.4511 & 1.5034 & 3.60\tabularnewline
\hline
5 & 1.1827e+00 & 13.0982 & 1.0486 & 1.4242 & 35.82 & 1.3704 & 1.4242 & 3.93\tabularnewline
\hline
6 & 1.5232e-01 & 14.5727 & 1.0454 & 1.3549 & 29.61 & 1.3006 & 1.3549 & 4.17\tabularnewline
\hline
7 & 1.7313e-02 & 18.1585 & 1.0547 & 1.3066 & 23.88 & 1.2502 & 1.3066 & 4.51\tabularnewline
\hline
8 & 2.3051e-03 & 27.3074 & 1.039 & 1.2871 & 23.88 & 1.2309 & 1.2871 & 4.57\tabularnewline
\hline
9 & 6.5778e-04 & 46.1972 & 1.0514 & 1.2798 & 21.72 & 1.2298 & 1.2798 & 4.07\tabularnewline
\hline
10 & 5.2980e-05 & 126.204 & 1.0551 & 1.275 & 20.84 & 1.2148 & 1.275 & 4.96\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 32.4\% & 4.91\% & 48.3\% & 41.3\% & 43.6\% & 48.3\% & 3.5\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas23}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $0.8<|\eta|<1.2$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.2<|\eta|<2.1$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.0631e+03 & 31.3513 & 1.0773 & 2.4321 & 125.76 & 2.5278 & 2.4321 & -3.79\tabularnewline
\hline
2 & 1.3804e+02 & 25.4028 & 1.0641 & 2.145 & 101.58 & 2.1559 & 2.145 & -0.51\tabularnewline
\hline
3 & 2.1351e+01 & 17.9496 & 1.0639 & 1.9372 & 82.08 & 1.904 & 1.9372 & 1.74\tabularnewline
\hline
4 & 4.3594e+00 & 14.1602 & 1.0473 & 1.8043 & 72.28 & 1.7536 & 1.8043 & 2.89\tabularnewline
\hline
5 & 7.3959e-01 & 14.6909 & 1.0501 & 1.705 & 62.37 & 1.6378 & 1.705 & 4.10\tabularnewline
\hline
6 & 8.3714e-02 & 15.7487 & 1.0419 & 1.6173 & 55.23 & 1.5365 & 1.6173 & 5.26\tabularnewline
\hline
7 & 9.2384e-03 & 20.4022 & 1.0441 & 1.5472 & 48.19 & 1.4631 & 1.5472 & 5.75\tabularnewline
\hline
8 & 9.1525e-04 & 32.1716 & 1.0217 & 1.5095 & 47.74 & 1.4002 & 1.5095 & 7.81\tabularnewline
\hline
9 & 9.0354e-05 & 81.2976 & 1.0604 & 1.4933 & 40.82 & 1.4035 & 1.4933 & 6.40\tabularnewline
\hline
10 & 2.4030e-05 & 125.8652 & 1.0772 & 1.4978 & 39.05 & 1.4001 & 1.4978 & 6.98\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 37.9\% & 5.5\% & 76.9\% & 67.5\% & 71.8\% & 76.8\% & 3.6\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas24}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $1.2<|\eta|<2.1$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.1<|\eta|<2.8$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 7.1034e+02 & 40.3484 & 1.0701 & 3.2055 & 199.55 & 3.558 & 3.2055 & -9.91\tabularnewline
\hline
2 & 8.4674e+01 & 32.6038 & 1.0175 & 2.8889 & 183.92 & 3.0465 & 2.8889 & -5.17\tabularnewline
\hline
3 & 1.0968e+01 & 22.655 & 1.0937 & 2.6712 & 144.24 & 2.702 & 2.6712 & -1.14\tabularnewline
\hline
4 & 1.8989e+00 & 18.4278 & 1.052 & 2.5724 & 144.52 & 2.5342 & 2.5724 & 1.51\tabularnewline
\hline
5 & 2.2594e-01 & 20.7194 & 1.0344 & 2.5309 & 144.67 & 2.4325 & 2.5309 & 4.05\tabularnewline
\hline
6 & 1.1698e-02 & 24.7426 & 1.0123 & 2.5444 & 151.35 & 2.3683 & 2.5444 & 7.44\tabularnewline
\hline
7 & 5.4844e-04 & 51.9687 & 1.0582 & 2.7866 & 163.33 & 2.5084 & 2.7866 & 11.09\tabularnewline
\hline
8 & 6.5250e-05 & 123.3975 & 1.0044 & 4.8469 & 382.57 & 3.8137 & 4.8469 & 27.09\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 41.8\% & 4.3\% & 200\% & 189.3\% & 187.1\% & 200\% & 4.4\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas25}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $2.1<|\eta|<2.8$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.8<|\eta|<3.6$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 3.6430e+02 & 50.2722 & 1.0306 & 4.2642 & 313.76 & 5.0833 & 4.2642 & -16.11\tabularnewline
\hline
2 & 3.1697e+01 & 48.5635 & 1.0433 & 4.0624 & 289.38 & 4.4793 & 4.0624 & -9.31\tabularnewline
\hline
3 & 2.4769e+00 & 41.0702 & 1.033 & 4.0504 & 292.10 & 4.1676 & 4.0504 & -2.81\tabularnewline
\hline
4 & 1.9495e-01 & 36.0109 & 1.0466 & 4.2394 & 305.06 & 4.144 & 4.2394 & 2.30\tabularnewline
\hline
5 & 4.9813e-03 & 37.8164 & 1.0301 & 4.9744 & 382.90 & 4.6503 & 4.9744 & 6.97\tabularnewline
\hline
6 & 1.6560e-04 & 92.2274 & 1.238 & 1.1865 & -4.16 & 1.0938 & 1.1865 & 8.48\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 50.9\% & 7\% & 279\% & 263.2\% & 293.6\% & 279\% & -1.8\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas26}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $2.8<|\eta|<3.6$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$3.6<|\eta|<4.4$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 5.9628e+01 & 102.8591 & 1.0091 & 6.5656 & 550.64 & 8.3872 & 6.5656 & -21.72\tabularnewline
\hline
2 & 1.7051e+00 & 153.8379 & 0.9774 & 7.567 & 674.20 & 8.59 & 7.567 & -11.91\tabularnewline
\hline
3 & 1.1701e-02 & 148.6082 & 0.7618 & 26.0327 & 3317.26 & 16.5523 & 26.0327 & 57.28\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 135.1\% & -9\% & 13.4\% & 1514\% & 1017.7\% & 1238.8\% & 7.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:katlas27}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the ATLAS 2011 $\sqrt{s}=2.76$ TeV dataset~\cite{Aad:2013lpa} in the rapidity slice $3.6<|\eta|<4.4$.}
}
\end{table}
\clearpage
\subsection{CDF jets}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$|\eta|<0.1$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.2332e+01 & 18.3447 & 1.1842 & 1.2089 & 2.09 & 1.1569 & 1.2089 & 4.49\tabularnewline
\hline
2 & 5.8392e+00 & 16.3363 & 1.1805 & 1.1919 & 0.97 & 1.1439 & 1.1919 & 4.20\tabularnewline
\hline
3 & 2.5648e+00 & 15.1286 & 1.1815 & 1.1771 & -0.37 & 1.1301 & 1.1771 & 4.16\tabularnewline
\hline
4 & 1.1293e+00 & 14.1836 & 1.1586 & 1.1679 & 0.80 & 1.1205 & 1.1679 & 4.23\tabularnewline
\hline
5 & 4.9055e-01 & 13.6802 & 1.1719 & 1.1587 & -1.13 & 1.1155 & 1.1587 & 3.87\tabularnewline
\hline
6 & 2.1819e-01 & 13.5437 & 1.1882 & 1.1527 & -2.99 & 1.1124 & 1.1527 & 3.62\tabularnewline
\hline
7 & 8.8655e-02 & 13.4586 & 1.1679 & 1.1489 & -1.63 & 1.1074 & 1.1489 & 3.75\tabularnewline
\hline
8 & 3.4626e-02 & 13.8317 & 1.1696 & 1.148 & -1.85 & 1.1115 & 1.148 & 3.28\tabularnewline
\hline
9 & 1.3375e-02 & 14.4361 & 1.1688 & 1.1453 & -2.01 & 1.1085 & 1.1453 & 3.32\tabularnewline
\hline
10 & 5.0470e-03 & 15.5616 & 1.1852 & 1.1504 & -2.94 & 1.1185 & 1.1504 & 2.85\tabularnewline
\hline
11 & 1.7391e-03 & 16.953 & 1.1856 & 1.1543 & -2.64 & 1.1279 & 1.1543 & 2.34\tabularnewline
\hline
12 & 5.7740e-04 & 18.58 & 1.1889 & 1.166 & -1.93 & 1.1357 & 1.166 & 2.67\tabularnewline
\hline
13 & 1.7438e-04 & 20.4884 & 1.2169 & 1.1773 & -3.25 & 1.1554 & 1.1773 & 1.90\tabularnewline
\hline
14 & 4.6164e-05 & 23.1129 & 1.2243 & 1.1985 & -2.11 & 1.1727 & 1.1985 & 2.20\tabularnewline
\hline
15 & 1.2831e-05 & 27.0039 & 1.2479 & 1.2243 & -1.89 & 1.2038 & 1.2243 & 1.70\tabularnewline
\hline
16 & 2.4676e-06 & 36.4214 & 1.2859 & 1.2655 & -1.59 & 1.2588 & 1.2655 & 0.53\tabularnewline
\hline
17 & 2.1933e-07 & 58.6463 & 1.3422 & 1.3305 & -0.87 & 1.3374 & 1.3305 & -0.52\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 20.28\% & 20.3\% & 18.6\% & -1.4\% & 15.4\% & 18.6\% & 2.8\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kcdf1}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CDF $\sqrt{s}=1.96$ TeV dataset~\cite{Abulencia:2007ez} in the rapidity slice $|\eta|<0.1$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.1<|\eta|<0.7$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.1773e+01 & 17.3309 & 1.1882 & 1.2291 & 3.44 & 1.1767 & 1.2291 & 4.45\tabularnewline
\hline
2 & 5.3141e+00 & 15.7537 & 1.1818 & 1.2111 & 2.48 & 1.161 & 1.2111 & 4.32\tabularnewline
\hline
3 & 2.3850e+00 & 14.375 & 1.18 & 1.1974 & 1.47 & 1.1478 & 1.1974 & 4.32\tabularnewline
\hline
4 & 1.0304e+00 & 13.4666 & 1.1811 & 1.1861 & 0.42 & 1.1371 & 1.1861 & 4.31\tabularnewline
\hline
5 & 4.5067e-01 & 12.9087 & 1.1752 & 1.1733 & -0.16 & 1.1276 & 1.1733 & 4.05\tabularnewline
\hline
6 & 1.9389e-01 & 12.6869 & 1.1704 & 1.1693 & -0.09 & 1.1247 & 1.1693 & 3.97\tabularnewline
\hline
7 & 8.0186e-02 & 12.7669 & 1.1643 & 1.1627 & -0.14 & 1.1203 & 1.1627 & 3.78\tabularnewline
\hline
8 & 3.1674e-02 & 13.1506 & 1.177 & 1.1608 & -1.38 & 1.1213 & 1.1608 & 3.52\tabularnewline
\hline
9 & 1.1795e-02 & 14.0111 & 1.1735 & 1.1591 & -1.23 & 1.1202 & 1.1591 & 3.47\tabularnewline
\hline
10 & 4.3733e-03 & 15.1797 & 1.1781 & 1.1627 & -1.31 & 1.1276 & 1.1627 & 3.11\tabularnewline
\hline
11 & 1.5238e-03 & 16.7691 & 1.1924 & 1.1665 & -2.17 & 1.1321 & 1.1665 & 3.04\tabularnewline
\hline
12 & 4.7637e-04 & 18.7902 & 1.1939 & 1.1767 & -1.44 & 1.1439 & 1.1767 & 2.87\tabularnewline
\hline
13 & 1.4064e-04 & 21.2221 & 1.2045 & 1.1898 & -1.22 & 1.164 & 1.1898 & 2.22\tabularnewline
\hline
14 & 3.6648e-05 & 24.4097 & 1.2254 & 1.2092 & -1.32 & 1.187 & 1.2092 & 1.87\tabularnewline
\hline
15 & 7.2081e-06 & 28.8382 & 1.2483 & 1.2386 & -0.78 & 1.2265 & 1.2386 & 0.99\tabularnewline
\hline
16 & 1.1827e-06 & 36.0876 & 1.2804 & 1.2838 & 0.27 & 1.2738 & 1.2838 & 0.79\tabularnewline
\hline
17 & 9.3577e-08 & 53.4882 & 1.3445 & 1.3473 & 0.21 & 1.3545 & 1.3473 & -0.53\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 20.1\% & 20.4\% & 20.1\% & -0.2\% & 16.7\% & 20.1\% & 2.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kcdf2}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CDF $\sqrt{s}=1.96$ TeV dataset~\cite{Abulencia:2007ez} in the rapidity slice $0.1<|\eta|<0.7$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.7<|\eta|<1.1$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.0490e+01 & 17.5576 & 1.1918 & 1.308 & 9.75 & 1.2514 & 1.308 & 4.52\tabularnewline
\hline
2 & 4.7944e+00 & 16.2211 & 1.1866 & 1.2863 & 8.40 & 1.2302 & 1.2863 & 4.56\tabularnewline
\hline
3 & 2.1471e+00 & 14.8956 & 1.1863 & 1.2656 & 6.68 & 1.2091 & 1.2656 & 4.67\tabularnewline
\hline
4 & 9.1016e-01 & 14.2023 & 1.1731 & 1.2457 & 6.19 & 1.1906 & 1.2457 & 4.63\tabularnewline
\hline
5 & 3.8312e-01 & 13.8172 & 1.1749 & 1.2356 & 5.17 & 1.1795 & 1.2356 & 4.76\tabularnewline
\hline
6 & 1.6165e-01 & 13.8845 & 1.185 & 1.2254 & 3.41 & 1.1698 & 1.2254 & 4.75\tabularnewline
\hline
7 & 6.4513e-02 & 14.1844 & 1.1743 & 1.2187 & 3.78 & 1.1642 & 1.2187 & 4.68\tabularnewline
\hline
8 & 2.3995e-02 & 14.8484 & 1.1752 & 1.2118 & 3.11 & 1.1564 & 1.2118 & 4.79\tabularnewline
\hline
9 & 8.5956e-03 & 15.8953 & 1.1928 & 1.2121 & 1.62 & 1.1561 & 1.2121 & 4.84\tabularnewline
\hline
10 & 2.9408e-03 & 17.3788 & 1.1894 & 1.2161 & 2.24 & 1.1641 & 1.2161 & 4.47\tabularnewline
\hline
11 & 9.2721e-04 & 19.3332 & 1.2042 & 1.2221 & 1.49 & 1.1706 & 1.2221 & 4.40\tabularnewline
\hline
12 & 2.4773e-04 & 22.1375 & 1.211 & 1.2373 & 2.17 & 1.1894 & 1.2373 & 4.03\tabularnewline
\hline
13 & 6.0900e-05 & 26.0269 & 1.2306 & 1.2529 & 1.81 & 1.2116 & 1.2529 & 3.41\tabularnewline
\hline
14 & 1.1095e-05 & 31.8232 & 1.2662 & 1.2838 & 1.39 & 1.2579 & 1.2838 & 2.06\tabularnewline
\hline
15 & 1.5785e-06 & 40.2833 & 1.2995 & 1.3305 & 2.39 & 1.3149 & 1.3305 & 1.19\tabularnewline
\hline
16 & 2.3408e-07 & 57.7304 & 1.3417 & 1.3985 & 4.23 & 1.3902 & 1.3985 & 0.60\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 21.9\% & 21.1\% & 25.9\% & 3.9\% & 21.3\% & 25.9\% & 3.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kcdf3}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CDF $\sqrt{s}=1.96$ TeV dataset~\cite{Abulencia:2007ez} in the rapidity slice $0.7<|\eta|<1.1$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.1<|\eta|<1.6$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 9.3784e+00 & 17.4829 & 1.1948 & 1.4397 & 20.50 & 1.3791 & 1.4397 & 4.39\tabularnewline
\hline
2 & 3.8388e+00 & 16.4669 & 1.1745 & 1.4086 & 19.93 & 1.3458 & 1.4086 & 4.67\tabularnewline
\hline
3 & 1.6185e+00 & 15.6789 & 1.1808 & 1.3816 & 17.01 & 1.3162 & 1.3816 & 4.97\tabularnewline
\hline
4 & 6.5171e-01 & 15.764 & 1.1735 & 1.3624 & 16.10 & 1.293 & 1.3624 & 5.37\tabularnewline
\hline
5 & 2.7261e-01 & 14.9935 & 1.18 & 1.3457 & 14.04 & 1.2759 & 1.3457 & 5.47\tabularnewline
\hline
6 & 1.0535e-01 & 15.4667 & 1.1844 & 1.3325 & 12.50 & 1.2643 & 1.3325 & 5.39\tabularnewline
\hline
7 & 3.8212e-02 & 16.014 & 1.1864 & 1.3278 & 11.92 & 1.2543 & 1.3278 & 5.86\tabularnewline
\hline
8 & 1.3067e-02 & 17.3512 & 1.188 & 1.3254 & 11.57 & 1.2507 & 1.3254 & 5.97\tabularnewline
\hline
9 & 3.9695e-03 & 19.2511 & 1.2008 & 1.3257 & 10.40 & 1.2544 & 1.3257 & 5.68\tabularnewline
\hline
10 & 1.0991e-03 & 22.068 & 1.2123 & 1.3346 & 10.09 & 1.263 & 1.3346 & 5.67\tabularnewline
\hline
11 & 2.6483e-04 & 25.8205 & 1.2328 & 1.3472 & 9.28 & 1.2774 & 1.3472 & 5.46\tabularnewline
\hline
12 & 5.1499e-05 & 30.4573 & 1.2529 & 1.372 & 9.51 & 1.317 & 1.372 & 4.18\tabularnewline
\hline
13 & 8.2305e-06 & 37.7877 & 1.2743 & 1.4069 & 10.41 & 1.3684 & 1.4069 & 2.81\tabularnewline
\hline
14 & 1.1920e-06 & 58.2323 & 1.3041 & 1.4916 & 14.38 & 1.4652 & 1.4916 & 1.80\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 23.1\% & 21.0\% & 37.2\% & 13.4\% & 30.9\% & 37.2\% & 4.8\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kcdf4}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CDF $\sqrt{s}=1.96$ TeV dataset~\cite{Abulencia:2007ez} in the rapidity slice $0.7<|\eta|<1.1$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.6<|\eta|<2.1$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 5.9423e+00 & 16.3059 & 1.1994 & 1.6725 & 39.44 & 1.6053 & 1.6725 & 4.19\tabularnewline
\hline
2 & 2.4098e+00 & 15.1148 & 1.1845 & 1.6347 & 38.01 & 1.5606 & 1.6347 & 4.75\tabularnewline
\hline
3 & 9.4782e-01 & 14.659 & 1.1936 & 1.6136 & 35.19 & 1.5306 & 1.6136 & 5.42\tabularnewline
\hline
4 & 3.4801e-01 & 14.9126 & 1.1938 & 1.5986 & 33.91 & 1.5103 & 1.5986 & 5.85\tabularnewline
\hline
5 & 1.2369e-01 & 15.9119 & 1.1909 & 1.589 & 33.43 & 1.4929 & 1.589 & 6.44\tabularnewline
\hline
6 & 3.9807e-02 & 17.6115 & 1.2015 & 1.587 & 32.08 & 1.4921 & 1.587 & 6.36\tabularnewline
\hline
7 & 1.1696e-02 & 20.0832 & 1.2017 & 1.5956 & 32.78 & 1.4887 & 1.5956 & 7.18\tabularnewline
\hline
8 & 2.8706e-03 & 23.1154 & 1.2069 & 1.6111 & 33.49 & 1.5118 & 1.6111 & 6.57\tabularnewline
\hline
9 & 5.7736e-04 & 27.0941 & 1.2306 & 1.6507 & 34.14 & 1.5464 & 1.6507 & 6.74\tabularnewline
\hline
10 & 8.7540e-05 & 33.0822 & 1.2463 & 1.7251 & 38.42 & 1.6413 & 1.7251 & 5.11\tabularnewline
\hline
11 & 9.2235e-06 & 42.8557 & 1.259 & 1.9175 & 52.30 & 1.832 & 1.9175 & 4.67\tabularnewline
\hline
12 & 5.9441e-07 & 60.5582 & 1.1491 & 1.4305 & 24.49 & 1.2008 & 1.4305 & 19.13\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 25.1\% & 20.5\% & 63.5\% & 35.6\% & 53.4\% & 63.5\% & 6.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kcdf5}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the CDF $\sqrt{s}=1.96$ TeV dataset~\cite{Abulencia:2007ez} in the rapidity slice $1.6<|\eta|<2.1$.}
}
\end{table}
\clearpage
\subsection{D0 jets}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$|\eta|<0.4$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 2.3282e+04 & 11.2934 & 1.1922 & 1.2223 & 2.52 & 1.1709 & 1.2223 & 4.39\tabularnewline
\hline
2 & 8.8665e+03 & 10.2301 & 1.1835 & 1.2008 & 1.46 & 1.1499 & 1.2008 & 4.43\tabularnewline
\hline
3 & 3.7927e+03 & 9.7392 & 1.1769 & 1.1858 & 0.76 & 1.1366 & 1.1858 & 4.33\tabularnewline
\hline
4 & 1.7678e+03 & 9.41 & 1.1867 & 1.1751 & -0.98 & 1.1283 & 1.1751 & 4.15\tabularnewline
\hline
5 & 8.9746e+02 & 9.2724 & 1.1765 & 1.1663 & -0.87 & 1.1209 & 1.1663 & 4.05\tabularnewline
\hline
6 & 4.8546e+02 & 9.0215 & 1.1694 & 1.1603 & -0.78 & 1.118 & 1.1603 & 3.78\tabularnewline
\hline
7 & 2.7591e+02 & 8.9452 & 1.1649 & 1.1581 & -0.58 & 1.1164 & 1.1581 & 3.74\tabularnewline
\hline
8 & 1.5967e+02 & 8.8119 & 1.1811 & 1.1544 & -2.26 & 1.1138 & 1.1544 & 3.65\tabularnewline
\hline
9 & 8.6959e+01 & 8.8544 & 1.1769 & 1.1529 & -2.04 & 1.1135 & 1.1529 & 3.54\tabularnewline
\hline
10 & 4.3619e+01 & 8.8333 & 1.1732 & 1.1497 & -2.00 & 1.1126 & 1.1497 & 3.33\tabularnewline
\hline
11 & 2.1406e+01 & 8.9002 & 1.162 & 1.1499 & -1.04 & 1.1151 & 1.1499 & 3.12\tabularnewline
\hline
12 & 9.8640e+00 & 9.0221 & 1.1743 & 1.1513 & -1.96 & 1.1138 & 1.1513 & 3.37\tabularnewline
\hline
13 & 4.7402e+00 & 9.192 & 1.1808 & 1.1524 & -2.41 & 1.1176 & 1.1524 & 3.11\tabularnewline
\hline
14 & 2.4291e+00 & 9.3741 & 1.1825 & 1.1555 & -2.28 & 1.1228 & 1.1555 & 2.91\tabularnewline
\hline
15 & 1.1835e+00 & 9.6935 & 1.1832 & 1.161 & -1.88 & 1.133 & 1.161 & 2.47\tabularnewline
\hline
16 & 5.1120e-01 & 10.1172 & 1.1955 & 1.1667 & -2.41 & 1.1399 & 1.1667 & 2.35\tabularnewline
\hline
17 & 2.1106e-01 & 10.7344 & 1.1904 & 1.1772 & -1.11 & 1.1536 & 1.1772 & 2.05\tabularnewline
\hline
18 & 8.2381e-02 & 11.5402 & 1.2108 & 1.1908 & -1.65 & 1.167 & 1.1908 & 2.04\tabularnewline
\hline
19 & 2.9294e-02 & 12.8752 & 1.2178 & 1.2101 & -0.63 & 1.1908 & 1.2101 & 1.62\tabularnewline
\hline
20 & 9.9894e-03 & 15.0811 & 1.2423 & 1.2305 & -0.95 & 1.2148 & 1.2305 & 1.29\tabularnewline
\hline
21 & 3.0638e-03 & 19.2471 & 1.2626 & 1.2598 & -0.22 & 1.2491 & 1.2598 & 0.86\tabularnewline
\hline
22 & 5.9961e-04 & 31.0056 & 1.2957 & 1.2935 & -0.17 & 1.2905 & 1.2935 & 0.23\tabularnewline
\hline
23 & 9.0820e-05 & 45.5599 & 1.3339 & 1.3499 & 1.20 & 1.3532 & 1.3499 & -0.24\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 12.9\% & 20.1\% & 19\% & -0.9\% & 15.8\% & 19\% & 2.8\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd01}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $|\eta|<0.4$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.4<|\eta|<0.8$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 2.1728e+04 & 11.47 & 1.1893 & 1.2639 & 6.27 & 1.2097 & 1.2639 & 4.48\tabularnewline
\hline
2 & 8.1627e+03 & 10.5165 & 1.1796 & 1.2379 & 4.94 & 1.1857 & 1.2379 & 4.40\tabularnewline
\hline
3 & 3.4953e+03 & 10.061 & 1.1852 & 1.2216 & 3.07 & 1.1693 & 1.2216 & 4.47\tabularnewline
\hline
4 & 1.6199e+03 & 9.826 & 1.1805 & 1.2085 & 2.37 & 1.1571 & 1.2085 & 4.44\tabularnewline
\hline
5 & 8.0500e+02 & 9.7325 & 1.1707 & 1.1968 & 2.23 & 1.1472 & 1.1968 & 4.32\tabularnewline
\hline
6 & 4.3791e+02 & 9.526 & 1.1688 & 1.1876 & 1.61 & 1.141 & 1.1876 & 4.08\tabularnewline
\hline
7 & 2.4307e+02 & 9.5132 & 1.1705 & 1.1831 & 1.08 & 1.1368 & 1.1831 & 4.07\tabularnewline
\hline
8 & 1.4182e+02 & 9.4416 & 1.1838 & 1.1812 & -0.22 & 1.1356 & 1.1812 & 4.02\tabularnewline
\hline
9 & 7.6800e+01 & 9.4736 & 1.185 & 1.1781 & -0.58 & 1.1311 & 1.1781 & 4.16\tabularnewline
\hline
10 & 3.8278e+01 & 9.5848 & 1.1688 & 1.1749 & 0.52 & 1.1274 & 1.1749 & 4.21\tabularnewline
\hline
11 & 1.8237e+01 & 9.6905 & 1.1803 & 1.1711 & -0.78 & 1.1299 & 1.1711 & 3.65\tabularnewline
\hline
12 & 8.1536e+00 & 9.9031 & 1.1959 & 1.1731 & -1.91 & 1.1335 & 1.1731 & 3.49\tabularnewline
\hline
13 & 3.7998e+00 & 10.1606 & 1.1779 & 1.1756 & -0.20 & 1.1333 & 1.1756 & 3.73\tabularnewline
\hline
14 & 1.8641e+00 & 10.4222 & 1.1909 & 1.1779 & -1.09 & 1.1395 & 1.1779 & 3.37\tabularnewline
\hline
15 & 8.7747e-01 & 10.8379 & 1.1958 & 1.1864 & -0.79 & 1.1428 & 1.1864 & 3.82\tabularnewline
\hline
16 & 3.5585e-01 & 11.4567 & 1.1953 & 1.1931 & -0.18 & 1.1557 & 1.1931 & 3.24\tabularnewline
\hline
17 & 1.3900e-01 & 12.3298 & 1.2031 & 1.2028 & -0.02 & 1.169 & 1.2028 & 2.89\tabularnewline
\hline
18 & 5.3148e-02 & 13.4911 & 1.2334 & 1.2191 & -1.16 & 1.1874 & 1.2191 & 2.67\tabularnewline
\hline
19 & 1.5559e-02 & 15.4698 & 1.2426 & 1.2365 & -0.49 & 1.2067 & 1.2365 & 2.47\tabularnewline
\hline
20 & 3.7822e-03 & 19.2421 & 1.2659 & 1.2614 & -0.36 & 1.2462 & 1.2614 & 1.22\tabularnewline
\hline
21 & 7.4934e-04 & 28.0381 & 1.2838 & 1.2993 & 1.21 & 1.2852 & 1.2993 & 1.10\tabularnewline
\hline
22 & 6.1715e-05 & 46.9948 & 1.3418 & 1.3509 & 0.68 & 1.3478 & 1.3509 & 0.23\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 13.5\% & 20.4\% & 21.3\% & 0.74\% & 17.4\% & 21.3\% & 3.4\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd02}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $0.4<|\eta|<0.8$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$0.8<|\eta|<1.2$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.7739e+04 & 12.0673 & 1.188 & 1.3436 & 13.10 & 1.2873 & 1.3436 & 4.37\tabularnewline
\hline
2 & 6.6481e+03 & 11.369 & 1.198 & 1.3133 & 9.62 & 1.256 & 1.3133 & 4.56\tabularnewline
\hline
3 & 2.8083e+03 & 11.1644 & 1.1829 & 1.2915 & 9.18 & 1.2328 & 1.2915 & 4.76\tabularnewline
\hline
4 & 1.2952e+03 & 11.1887 & 1.1656 & 1.2741 & 9.31 & 1.2158 & 1.2741 & 4.80\tabularnewline
\hline
5 & 6.6021e+02 & 10.2182 & 1.1784 & 1.26 & 6.92 & 1.2008 & 1.26 & 4.93\tabularnewline
\hline
6 & 3.4070e+02 & 10.2663 & 1.1861 & 1.2537 & 5.70 & 1.1963 & 1.2537 & 4.80\tabularnewline
\hline
7 & 1.7052e+02 & 10.111 & 1.1803 & 1.244 & 5.40 & 1.1864 & 1.244 & 4.86\tabularnewline
\hline
8 & 7.5323e+01 & 10.2419 & 1.1874 & 1.2387 & 4.32 & 1.1796 & 1.2387 & 5.01\tabularnewline
\hline
9 & 3.6419e+01 & 10.2693 & 1.1763 & 1.2342 & 4.92 & 1.1732 & 1.2342 & 5.20\tabularnewline
\hline
10 & 1.8092e+01 & 10.5054 & 1.1751 & 1.2296 & 4.64 & 1.1729 & 1.2296 & 4.83\tabularnewline
\hline
11 & 8.5257e+00 & 10.7795 & 1.1833 & 1.2312 & 4.05 & 1.1703 & 1.2312 & 5.20\tabularnewline
\hline
12 & 3.6738e+00 & 11.229 & 1.2057 & 1.2332 & 2.28 & 1.1763 & 1.2332 & 4.84\tabularnewline
\hline
13 & 1.6477e+00 & 11.7495 & 1.2011 & 1.2365 & 2.95 & 1.1837 & 1.2365 & 4.46\tabularnewline
\hline
14 & 7.6533e-01 & 12.3378 & 1.2072 & 1.2427 & 2.94 & 1.1883 & 1.2427 & 4.58\tabularnewline
\hline
15 & 3.5918e-01 & 13.1141 & 1.2198 & 1.2517 & 2.62 & 1.1981 & 1.2517 & 4.47\tabularnewline
\hline
16 & 1.4486e-01 & 14.1806 & 1.2162 & 1.2605 & 3.64 & 1.216 & 1.2605 & 3.66\tabularnewline
\hline
17 & 4.1615e-02 & 16.025 & 1.2417 & 1.2793 & 3.03 & 1.2257 & 1.2793 & 4.37\tabularnewline
\hline
18 & 1.0516e-02 & 18.9392 & 1.2682 & 1.3022 & 2.68 & 1.2729 & 1.3022 & 2.30\tabularnewline
\hline
19 & 2.4549e-03 & 24.0553 & 1.2867 & 1.3393 & 4.09 & 1.3242 & 1.3393 & 1.14\tabularnewline
\hline
20 & 1.6393e-04 & 34.0392 & 1.3245 & 1.3893 & 4.89 & 1.3804 & 1.3893 & 0.64\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 13.7\% & 20.9\% & 27.2\% & 5.3\% & 22.2\% & 27.2\% & 4.2\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd03}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $0.8<|\eta|<1.2$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.2<|\eta|<1.6$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.5249e+04 & 13.2561 & 1.1923 & 1.4726 & 23.51 & 1.4114 & 1.4726 & 4.34\tabularnewline
\hline
2 & 5.2668e+03 & 12.6595 & 1.1838 & 1.4355 & 21.26 & 1.3718 & 1.4355 & 4.64\tabularnewline
\hline
3 & 2.1975e+03 & 12.5006 & 1.1882 & 1.4087 & 18.56 & 1.3407 & 1.4087 & 5.07\tabularnewline
\hline
4 & 9.6320e+02 & 11.5469 & 1.1926 & 1.3883 & 16.41 & 1.3191 & 1.3883 & 5.25\tabularnewline
\hline
5 & 4.6424e+02 & 11.3927 & 1.1795 & 1.3728 & 16.39 & 1.299 & 1.3728 & 5.68\tabularnewline
\hline
6 & 2.3508e+02 & 11.4617 & 1.1722 & 1.3617 & 16.17 & 1.2904 & 1.3617 & 5.53\tabularnewline
\hline
7 & 1.1078e+02 & 11.6263 & 1.1891 & 1.3568 & 14.10 & 1.2848 & 1.3568 & 5.60\tabularnewline
\hline
8 & 4.5867e+01 & 11.9463 & 1.18 & 1.3485 & 14.28 & 1.274 & 1.3485 & 5.85\tabularnewline
\hline
9 & 1.9989e+01 & 12.0667 & 1.1913 & 1.3506 & 13.37 & 1.2714 & 1.3506 & 6.23\tabularnewline
\hline
10 & 9.2228e+00 & 12.4861 & 1.1974 & 1.3529 & 12.99 & 1.2775 & 1.3529 & 5.90\tabularnewline
\hline
11 & 3.7713e+00 & 13.0025 & 1.1846 & 1.3524 & 14.17 & 1.2758 & 1.3524 & 6.00\tabularnewline
\hline
12 & 1.2448e+00 & 14.6819 & 1.2096 & 1.3665 & 12.97 & 1.2928 & 1.3665 & 5.70\tabularnewline
\hline
13 & 3.5744e-01 & 15.9733 & 1.203 & 1.3826 & 14.93 & 1.3065 & 1.3826 & 5.82\tabularnewline
\hline
14 & 1.0992e-01 & 17.9723 & 1.2341 & 1.4011 & 13.53 & 1.342 & 1.4011 & 4.40\tabularnewline
\hline
15 & 2.9682e-02 & 20.5192 & 1.2357 & 1.4297 & 15.70 & 1.3731 & 1.4297 & 4.12\tabularnewline
\hline
16 & 7.2732e-03 & 24.2715 & 1.2644 & 1.4718 & 16.40 & 1.4296 & 1.4718 & 2.95\tabularnewline
\hline
17 & 3.9206e-04 & 32.6071 & 1.307 & 1.5641 & 19.67 & 1.5451 & 1.5641 & 1.23\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 15.3\% & 20.6\% & 40\% & 16.1\% & 33.6\% & 40.1\% & 4.9\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd04}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $1.2<|\eta|<1.6$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$1.6<|\eta|<2.0$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 1.1684e+04 & 13.2001 & 1.1967 & 1.6629 & 38.96 & 1.6007 & 1.6629 & 3.89\tabularnewline
\hline
2 & 4.0542e+03 & 12.7548 & 1.1875 & 1.6233 & 36.70 & 1.5502 & 1.6233 & 4.72\tabularnewline
\hline
3 & 1.5651e+03 & 12.432 & 1.2011 & 1.5954 & 32.83 & 1.5135 & 1.5954 & 5.41\tabularnewline
\hline
4 & 6.6951e+02 & 12.5395 & 1.1844 & 1.5807 & 33.46 & 1.4936 & 1.5807 & 5.83\tabularnewline
\hline
5 & 2.9506e+02 & 12.6582 & 1.1809 & 1.5703 & 32.97 & 1.4813 & 1.5703 & 6.01\tabularnewline
\hline
6 & 1.3622e+02 & 12.9305 & 1.2123 & 1.5655 & 29.13 & 1.4714 & 1.5655 & 6.40\tabularnewline
\hline
7 & 5.6908e+01 & 13.3802 & 1.1986 & 1.5665 & 30.69 & 1.4713 & 1.5665 & 6.47\tabularnewline
\hline
8 & 2.0321e+01 & 14.1627 & 1.2159 & 1.5769 & 29.69 & 1.4742 & 1.5769 & 6.97\tabularnewline
\hline
9 & 6.3418e+00 & 15.2376 & 1.2074 & 1.5934 & 31.97 & 1.4867 & 1.5934 & 7.18\tabularnewline
\hline
10 & 1.9329e+00 & 16.6521 & 1.2138 & 1.6201 & 33.47 & 1.5083 & 1.6201 & 7.41\tabularnewline
\hline
11 & 6.8901e-01 & 18.3248 & 1.239 & 1.6514 & 33.28 & 1.5426 & 1.6514 & 7.05\tabularnewline
\hline
12 & 2.2119e-01 & 20.585 & 1.2254 & 1.7028 & 38.96 & 1.59 & 1.7028 & 7.09\tabularnewline
\hline
13 & 4.5679e-02 & 24.6268 & 1.2369 & 1.7912 & 44.81 & 1.7293 & 1.7912 & 3.58\tabularnewline
\hline
14 & 6.7894e-03 & 31.7718 & 1.1985 & 2.075 & 73.13 & 2.0602 & 2.075 & 0.72\tabularnewline
\hline
15 & 4.0224e-04 & 44.9061 & 1.1371 & 1.4991 & 31.84 & 1.2847 & 1.4991 & 16.69\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 18.4\% & 20.2\% & 64.5\% & 36.8\% & 55.1\% & 64.5\% & 6.4\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd05}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $1.6<|\eta|<2.0$.}
}
\end{table}
\begin{table}[h]
\centering{}%
\begin{tabular}{|c|cc||cc|c||cc|c|}
\hline
\multicolumn{3}{|c||}{$2.0<|\eta|<2.4$} & \multicolumn{2}{c|}{gg $k$-factor} & & \multicolumn{2}{c|}{gg $k$-factor, NNLO$^{{\rm thr}}/$} & \tabularnewline
\hline
$p_{T}$ bin & $\sigma_{{\rm Exp}}$ & $\sigma_{{\rm Exp}}^{{\rm error}}$ (\%) & Exact & Thr & $\delta\%$ & NLO$^{{\rm exact}}$ & NLO$^{{\rm thr}}$ & $\delta$\%\tabularnewline
\hline
\hline
1 & 7.2148e+03 & 14.7391 & 1.2004 & 1.9522 & 62.63 & 1.8857 & 1.9522 & 3.53\tabularnewline
\hline
2 & 2.2811e+03 & 14.7294 & 1.196 & 1.9255 & 60.99 & 1.8391 & 1.9255 & 4.70\tabularnewline
\hline
3 & 7.4168e+02 & 14.8434 & 1.1859 & 1.9162 & 61.58 & 1.8166 & 1.9162 & 5.48\tabularnewline
\hline
4 & 2.7132e+02 & 15.442 & 1.2224 & 1.9309 & 57.96 & 1.8063 & 1.9309 & 6.90\tabularnewline
\hline
5 & 9.9553e+01 & 16.0892 & 1.2034 & 1.9457 & 61.68 & 1.8335 & 1.9457 & 6.12\tabularnewline
\hline
6 & 3.8789e+01 & 16.9895 & 1.1997 & 1.9761 & 64.72 & 1.853 & 1.9761 & 6.64\tabularnewline
\hline
7 & 1.4726e+01 & 18.0928 & 1.222 & 2.011 & 64.57 & 1.8989 & 2.011 & 5.90\tabularnewline
\hline
8 & 5.7305e+00 & 19.3456 & 1.208 & 2.0719 & 71.51 & 1.9374 & 2.0719 & 6.94\tabularnewline
\hline
9 & 1.8183e+00 & 21.2101 & 1.2115 & 2.1508 & 77.53 & 2.0074 & 2.1508 & 7.14\tabularnewline
\hline
10 & 3.8345e-01 & 24.4537 & 1.1939 & 2.352 & 97.00 & 2.2418 & 2.352 & 4.92\tabularnewline
\hline
11 & 8.1686e-02 & 28.3113 & 1.0476 & 3.0379 & 189.99 & 2.9002 & 3.0379 & 4.75\tabularnewline
\hline
12 & 1.0941e-02 & 35.1504 & 1.2291 & 1.5741 & 28.07 & 1.3978 & 1.5741 & 12.61\tabularnewline
\hline
13 & 3.6139e-04 & 52.8355 & 1.4345 & 0.2947 & -79.46 & 0.279 & 0.2947 & 5.63\tabularnewline
\hline
\hline
\multicolumn{2}{|c|}{Average} & 22.5\% & 21.2\% & 93.4\% & 62.9\% & 82.3\% & 93.4\% & 6.3\%\tabularnewline
\hline
\end{tabular}{\small \caption{\label{tab:kd06}Numerical results for the gluons-only exact and approximate NNLO $k$-factors as described in Sect.~\ref{sec:append}
for the D0 $\sqrt{s}=1.96$ TeV dataset~\cite{Abazov:2011vi} in the rapidity slice $2.0<|\eta|<2.4$.}
}
\end{table}
\section{Benchmark predictions for jet production}
\label{sec:bench}
\subsection{LHC and Tevatron jet data}
In order to provide realistic comparisons, based on real data which is
already included in the extractions of parton distribution functions
(PDFs)~\cite{Ball:2012cx}, we have selected recent data sets obtained
during the LHC Run-I and the Tevatron Run-II. Using data from both
colliders provides the possibility of investigating differences and
similarities between datasets for different collision energies and
kinematic coverage. A summary of the experimental data included in our
analysis is presented in Table~\ref{tab:experiments}. As we will show
in Sect.~\ref{sec:lhc} and Sect.~\ref{sec:tev}, the region of validity
of the threshold approximation is very dependent on the experimental
details.
From the LHC experiments we have included the CMS measurements of
the double differential jet cross sections at $\sqrt{s}=7$
TeV~\cite{Chatrchyan:2012bja}, where jets are reconstructed up to
$|\eta|<2.5$. We have also included the ATLAS measurements of
inclusive jet cross sections at $\sqrt{s}=7$ TeV~\cite{Aad:2010ad}
and $\sqrt{s}=2.76$ TeV~\cite{Aad:2013lpa}, where the rapidity
coverage reaches $|\eta|<4.4$. For both LHC experiments jets are
reconstructed with the anti-$k_{t}$ algorithm. The main differences
between CMS and ATLAS data is the choice of jet resolution parameter R,
which is R=0.7 for CMS and R=0.4 for ATLAS, and the $p_{T}$
coverage which for CMS covers the very high $p_{T}$ region, reaching
2~TeV, while ATLAS measures very low $p_{T}$ jets starting
from 20~GeV.
Concerning the Tevatron data, we have included the
most recent CDF Run-II $k_{t}$ jets~\cite{Abulencia:2007ez} and the D0
Run-II cone data~\cite{Abazov:2011vi}. In contrast to LHC data, the
center of mass energy of both sets is $\sqrt{s}=1.96$ TeV and their
coverage in rapidity and $p_{T}$ is smaller than ATLAS and CMS
experiments. It is important to highlight that CDF uses the $k_{t}$
algorithm to do the jet reconstruction, while D0 presents
data reconstructed with the MidPoint cone algorithm which
is infrared unsafe at NNLO.
\begin{table}
\begin{centering}
\begin{tabular}{cc|c|c|c|c|c|c}
\hline \multicolumn{2}{c|}{Data Set} & Points & $\sqrt{s}$ (TeV) & $R$ & $\eta$ coverage
& $p_{T}$ coverage & Ref.\tabularnewline
\hline \hline
\multirow{3}{*}{LHC} & CMS 2011 & 133 & 7 & 0.7 & $|\eta|<2.5$ & $[114,2116]$ GeV & ~\cite{Chatrchyan:2012bja} \tabularnewline
\cline{2-8} & ATLAS 2010 & 90 & 7 & 0.4 & $|\eta|<4.4$ & $[20,1500]$ GeV & ~\cite{Aad:2010ad} \tabularnewline
\cline{2-8} & ATLAS 2011 & 59 & 2.76 & 0.4 & $|\eta|<4.4$ & $[20,500]$ GeV & ~\cite{Aad:2013lpa} \tabularnewline
\hline \hline
\multirow{2}{*}{Tevatron} & CDF $k_{t}$& 76 & 1.96 & 0.7 & $|\eta|<2.1$ &$[54,700]$ GeV & ~\cite{Abulencia:2007ez} \tabularnewline
\cline{2-8} & D0 cone & 110 & 1.96 & 0.7 & $|\eta|<2.4$ & $[50,665]$ GeV & ~\cite{Abazov:2011vi} \tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{\label{tab:experiments}Jet data included in the current
analysis with the respective kinematic information.}
\end{table}
\subsection{Theoretical predictions}
\label{sec:benchTH}
Theoretical predictions presented in this work are computed
exclusively with the central value of the {\tt
NNPDF23\_nnlo\_as\_0118} set, presented by the NNPDF collaboration
in Ref.~\cite{Ball:2012cx}. This set is used for predictions at all
perturbative orders. However, we are interested in comparing
predictions at the same order and thus the choice of the input PDF is
only marginally relevant.
At LO and NLO full exact predictions are available and these have been computed with the {\tt
FastNLO}~\cite{Kluge:2006xs} interface for CMS, CDF and D0 and with
the {\tt APPLgrid}~\cite{Carli:2010rw} tables for the ATLAS
predictions. Tables used with both interfaces have been computed with
{\tt NLOjet++} program~\cite{Catani:1996vz,Nagy:2003tz}. Predictions at NNLO are computed using the exact fixed-order results in the gluon-gluon channel
and with the threshold approximation code.
To obtain the exact predictions at NNLO we use the parton level Monte Carlo NNLOJET code recently presented in
Ref.~\cite{Currie:2013dwa,Ridder:2013mf,Currie:2014upa} interfaced with {\tt libHFILL}\footnote{Available at: \url{http://libhfill.hepforge.org/}}, a histogram library
developed for this work and compatible with all MCs programs which
allows the automatic construction of jet $p_{T}$ distributions
from event weights, using the binning and kinematic regions presented
in Table~\ref{tab:experiments}. The Monte Carlo uncertainties
presented for the exact predictions are below the percent level.
In this code, the $gg\to gg+X$ at full colour and the $q\bar{q}\to gg+X$~\cite{Currie:2013vh,Currie:2014upa} contributions at leading colour
are available at NNLO and the current limitations are the missing
partonic contributions for $qg$ and $qq$ scattering.
To obtain the approximate NNLO predictions based on threshold
resummation we use the threshold approximation
code~\cite{deFlorian:2013qia} which implements predictions for all
channels. We have also used the narrow-jet approximation code (NJA)
presented in Ref.~\cite{Jager:2004jh,Mukherjee:2012uz}, which computes
in analytic form the single jet inclusive cross section at LO and NLO
in the narrow-jet limit where both the matrix elements and phase space
are expanded around the narrow-jet limit. The original version of the
NJA and threshold codes have been improved through comparisons with
exact calculations and updated in order to use the {\tt
LHAPDF}~\cite{Whalley:2005nh} interface to PDFs and by including the
bottom and the anti-quark PDFs contributions to the total
luminosity. After these modifications both codes show full agreement
at LO with the exact calculations and can be used for comparisons with
experimental data.
For all predictions the value of $\alpha_s$ is provided by the PDF set through the {\tt LHAPDF}~\cite{Whalley:2005nh} interface.
For the exact calculation we generate predictions using two different dynamical renormalisation and factorisation scales. One choice evaluates the
fixed-order single jet inclusive cross section using $\mu_R=\mu_F=\mu=p_{T1}$ where for each event the renormalisation and factorisation
scales are set equal to each other and equal to the $p_T$ of the leading jet in the event. The $p_T$ of the leading jet is obtained after clustering all parton final-state momenta
into jet momenta using the appropriate jet algorithm employed by each experimental setup. Whenever the jet algorithm determines a partonic clustering the 4D recombination scheme is applied, i.e.,
the 4-momenta of the jet's constituents is added, producing a list of final-state jets that are ordered in $p_T$ at the end of the clustering procedure.
As a second choice we computed the fixed-order single jet inclusive cross section using $\mu_R=\mu_F=\mu=p_{T}$ where in this case each jet in every event
is binned with the weight evaluated at the scale $p_T$ of the jet. While at LO the two final state partons generate two jets with equal transverse momentum
$p_{T1}=p_{T2}=p_{T}$ and the two scale choices coincide, radiative corrections can generate subleading jets and the effects of the different scale choice in the theory prediction
become apparent at NLO and NNLO.
The approximate threshold and the NJA predictions use
$\mu_R=\mu_F=\mu=p_{T}$, where the $p_T$ of the observed jet is
generated by the MC integration and the general structure of the
resummed cross section applies specifically to the anti-$k_T$
algorithm~\cite{deFlorian:2013qia}.
In the next section we present the theoretical predictions for all the experimental setups in Table~\ref{tab:experiments} at each order in perturbation
theory up to NNLO and perform a benchmark comparison of the various approximations.
When assessing the validity of the threshold prediction
at NNLO we will suggest a rejection criteria which is to exclude approximate predictions which are more than 10\%
off the exact prediction. We will apply this criteria in the gluons-only channel to
help determine exactly the regions in the experimental setup where the threshold approximation is applicable. In Fig.~\ref{fig:ggfraction} it is shown,
using the full LO prediction, the relative contribution of the $gg$-channel
in the various ($p_T$,$|\eta|$) regions of each experimental setup to the full result. We observe that in the high-$p_T$ region of each experiment the
gluon-gluon scattering contribution is highly suppressed. The high-$p_{T}$ region corresponds precisely to the threshold region where the phase space
available for additional radiation is limited and for this reason the dominance of the gluon-gluon channel cannot be used as an argument to assess
the validity of the threshold approximation. However, by looking at the relative error of NLO-threshold vs. NLO exact for different partonic channels we
observe a convergence of the threshold approximation to the exact result at high-$p_T$ in each channel. Since the same comparison can be performed
reliably at NNLO in the $gg$-channel only, the criterium mentioned above to exclude approximate predictions is not a recommendation
and its purpose is to fix a level of accuracy when using approximate predictions, while the full calculation is not avaliable.
In Sect.~\ref{sec:PDFfit} we will discuss the effects of being more restrictive or flexible with this choice.
\begin{figure}[t]
\begin{centering}
\includegraphics[scale=0.35]{scatterCMS_lo}\includegraphics[scale=0.35]{scatterATLAS7_lo}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.35]{scatterATLAS2_lo}\includegraphics[scale=0.35]{scatterCDF_lo}
\par\end{centering}
\caption{
\label{fig:ggfraction} Relative percentage-wise contribution of the $gg$-channel to the full all-channel hadronic jet production
as a function of $p_{T}$ and $|y|$ for CMS, ATLAS 7 TeV and 2.76 TeV and CDF bins.}
\end{figure}
\section{Conclusion and outlook}
\label{sec:conclusion}
The purpose of this paper is to compare in the gluons-only channel predictions at NNLO for the single jet
inclusive cross section based on the exact NNLO calculation published in~\cite{Currie:2013dwa,Ridder:2013mf}
with an approximate NNLO calculation based on threshold resummation published in~\cite{deFlorian:2013qia}.
This comparison is performed using the same experimental setups employed by the Tevatron and LHC experiments in their jet analysis.
With these results we deliver an updated description of the state of the art of the accuracy of the theoretical predictions
for the single jet inclusive cross section in QCD at hadron colliders, and in particular revise contradictory statements
in the literature~\cite{deFlorian:2013qia}.
We observe that when the predictions are compared using the same central scale choice for the renormalisation and factorisation
scales the disagreements are larger than previously quoted. Concerning the regions of validity of the NNLO approximation
we conclude, based on a criteria of excluding approximate prediction which are more than 10\% off the exact prediction,
that the threshold approximation code provides predictions that are reasonably close to the exact calculation at large $p_{T}$
and central rapidity regions. We observed smaller differences between the exact calculation
and the approximate NNLO threshold calculations at the Tevatron than
at the LHC. It is important to highlight that
threshold predictions produced integrated over rapidity, as shown
in~\cite{deFlorian:2013qia}, are dominated by the central rapidity
regions and provide stable results. However, when looking at specific
rapidity bins the threshold predictions are, in some cases, far from the exact computation. This
remark is important and invites caution when using the threshold
approximation for the determination of PDFs.
As an exercise and to test this observation we performed a PDF fit
including approximate NNLO corrections. We observed that as expected
the resulting fit quality is dependent on the criteria which is
employed to exclude/include approximate NNLO corrections. A more
conservative criteria has the effect of excluding a larger amount of
the experimental data points that go into the fit, and favours regions
where the approximate prediction gives smaller NNLO corrections in
agreement with the exact calculation.
Finally we conclude that with the current results, there is no trivial
way to determine the $p_{T}$ value for which the
threshold approximation predictions are reliable. The only possible
prescription is to check the relative difference to the exact
computation bin by bin, and admitting a tolerance which can be
correlated to the real data uncertainty. As we have shown, the regions of validity
of the threshold approximation are very dependent on the experimental setups
that we have analysed and are very likely to be different for the future high-energy Run-II of the LHC.
As a further improvement it would be interesting to repeat such study
when the NNLO exact prediction becomes available for all channels.
\acknowledgments
We thank Stefano Forte for intensive discussions and reflections about
the results presented in this document, Werner Vogelsang for providing
the threshold and NJA approximations codes and provide examples on how
to setup and run the respective codes. We also thank Juan Rojo and the
NNPDF Collaboration for discussions about PDF fits with jet data. JP
acknowledges Nigel Glover and Thomas Gehrmann for comments on the
manuscript.
JP and SC acknowledge support by an Italian PRIN2010 grant, and for SC
also by an European Investment Bank EIBURS grant, and by the European
Commission through the HiggsTools Initial Training Network
PITN-GA-2012-316704. JP thanks the Dipartimento di Fisica, Universita
di Milano-Bicocca for their kind hospitality.
\section{Introduction}
\label{sec:intro}
Single jet inclusive and dijet observables are the most fundamental QCD processes measured at hadron colliders.
They probe the basic parton-parton scattering in QCD and thus allow for a determination of the parton distribution functions
in the proton and for a direct probe of the strong coupling constant up to the highest energy scales that can be attained in collider experiments.
In particular, gluon scattering is a direct contribution to the production of high-$p_T$ jets.
For this reason, jet data is included in PDF fits with the goal of assessing the gluon distribution in the proton at medium to large values
of the momentum fraction $x$.
Improvements at the level of accuracy of the theoretical predictions for the single jet inclusive
cross section beyond next-to-leading order (NLO) in QCD perturbation theory have been achieved recently.
First, the exact next-to-next-to leading order (NNLO) prediction for the gluons-only channel has been published in~\cite{Currie:2013dwa,Ridder:2013mf}.
Second, an approximate NNLO prediction based on threshold resummation is presented in~\cite{deFlorian:2013qia}.
In this work we perform a systematic study comparing theoretical
predictions at leading (LO), next-to-leading (NLO) and next-to-next-to-leading
order (NNLO), to recent data from the LHC and Tevatron
experiments. The aim of this study is to understand and characterize the validity
of the NNLO threshold approximation~\cite{deFlorian:2013qia} by comparing it to the exact
computation in the gluon-gluon channel~\cite{Currie:2013dwa,Ridder:2013mf}. In
Ref.~\cite{deFlorian:2013qia} the threshold approximation is compared to the
exact calculation in the gluon-gluon channel showing, after integration
over rapidity, a good agreement at large $p_{T}$.
However, for small $p_{T}$ regions it tends to diverge from the exact
computation. Our objective is to determine the
experimental regions where this breakdown of the threshold
approximation occurs. A rejection criteria to exclude approximate predictions will be suggested based on the gluon-gluon
channel which is dominant in the small $p_{T}$ region. In this region the full NNLO
computation is dominated by the gluon-gluon channel and therefore the predictions from the
exact NNLO calculation in this channel are reliable to determine the kinematic regions for which the threshold
terms become accurate. Moreover, and contrary to the study made in Ref.~\cite{deFlorian:2013qia}, we will
compare both predictions using the same factorisation and renormalisation scales in both calculations and discuss
the effects of making different scale choices in the theory predictions.
In order to obtain the exact NNLO predictions, the calculation in~\cite{Currie:2013dwa,Ridder:2013mf} used
the antenna subtraction scheme~\cite{GehrmannDeRidder:2005cm} to perform the cancellation of IR singularities between real
and virtual corrections at NNLO~\cite{Glover:2010im,GehrmannDeRidder:2011aa,Ridder:2012dg,Currie:2013vh}. For hadron collider observables this includes
contributions due to radiative corrections from partons in the initial state~\cite{Daleo:2006xa,Daleo:2009yj,Boughezal:2010mc,Gehrmann:2011wi,GehrmannDeRidder:2012ja}.
The cancellation of IR singularities is achieved analytically in all intermediate steps of the calculation thereby producing a strong check on the correctness of the calculation.
In this calculation the exact two-loop~\cite{Glover:2001af,Glover:2001rd,Bern:2002tk}, one-loop~\cite{Bern:1993mq} and tree-level~\cite{Mangano:1990by} QCD matrix elements
at NNLO are included in a parton-level generator NNLOJET, which integrates them over the exact full phase space to compute any infrared
safe two-jet observable to NNLO accuracy. For the purposes of the present study we compute the single jet inclusive cross section $pp\to j +X$
where we require to observe at least one jet in the final state and integrate inclusively any additional radiation.
In Ref.~\cite{deFlorian:2013qia} approximate NNLO results for the same observable were derived using the formalism of threshold resummation for single jet production
in hadron-hadron collisions. This formalism was first developed in~\cite{Kidonakis:1998bk,Kidonakis:1998nf,Laenen:1998qw} and predictions
in a scheme where jets are assumed to be massless at the partonic threshold were produced in~\cite{Kidonakis:2000gi}.
In a study performed in~\cite{Kumar:2013hia} it was shown that the NLO terms in this scheme~\cite{Kidonakis:2000gi}
fail to match a full NLO calculation even in a regime where threshold logarithms should dominate. On the other hand in Ref.~\cite{deFlorian:2013qia} the structure of the threshold logarithms allows jets to have a non-vanishing
invariant mass at threshold. We will compare our predictions at NLO and NNLO with the predictions obtained in the latter scheme~\cite{deFlorian:2013qia}.
In this framework the threshold limit is defined by the vanishing of the invariant mass of the system that recoils against the observed jet $s_4=P_X^2\to0$.
In this limit the phase space available for additional soft radiation is restricted such that the higher $k$th order coefficient functions are dominated by large
logarithmic corrections,
\begin{equation}
\alpha_s^k w_{ab}^{(k)}\to\alpha_s^k \left(\frac{\log^m(z)}{z}\right)_{+},\qquad m\le2k-1,\qquad z=\frac{s_4}{s}.
\end{equation}
The next-to-leading logarithmic (NLL) threshold calculation~\cite{deFlorian:2013qia}
then performs a systematic resummation of these logarithmic enhanced contributions for all partonic channels
to all orders in the strong coupling $\alpha_s$, by determining the
three leading logarithmic contributions $\propto$ $(\log^3(z)/z)_+$,
$(\log^2(z)/z)_+$, $(\log(z)/z)_+$ and keeping full dependence of the
cross section on the jet rapidity~\cite{deFlorian:2013qia}. The soft
contribution $\delta(z)$ as well as non-enhanced regular terms in $z$
of NNLO accuracy are not computed in this approach. For this reason
it has been in shown in~\cite{deFlorian:2013qia} that different
approximate NNLO predictions can be derived from the threshold
formalism if the variables used in the computation differ away from
the threshold $z=0$ limit (but are otherwise identical at $z=0$). This
effect can lead to a significant change in the shape of the
approximate NNLO threshold prediction~\cite{deFlorian:2013qia} and
increases its uncertainty.
Together with the comparison between the predictions at NNLO obtained
in the threshold formalism and in the exact fixed-order calculation,
we also provide the NNLO/NLO $k$-factors relevant for the Tevatron and
the LHC experiments. The paper is organized as follows. In
Sect.~\ref{sec:bench} we present the jet data selected for the
comparison, and the setup of the computational tools used for the
generation of the theoretical predictions. In Sect.~\ref{sec:lhc} and
Sect.~\ref{sec:tev} we show the results for the LHC and the Tevatron
experiments respectively. In Sect.~\ref{sec:PDFfit} we present as an
exercise an aNNLO PDF fit, where ``a'' stands for approximate, i.e. we
use only approximate NNLO $k$-factors computed in Sect.~\ref{sec:lhc}
and Sect.~\ref{sec:tev} and not the full NNLO predictions, since the
exact full prediction is not available yet. In this paper the aNNLO
notation is used in order to reserve the terminology NNLO PDF fit for
when the full NNLO predictions for jet production are available. In
Sect.~\ref{sec:conclusion} we present our conclusions and directions
for future work. An appendix is enclosed which provides tables with
$k$-factors in the gluon-gluon channel at the LHC and the Tevatron.
\section{LHC jet data}
\label{sec:lhc}
For each experiment in Table~\ref{tab:experiments} we have generated all channel full LO, NLO and approximate NNLO predictions
and compared them directly against the experimental data. In
Figure~\ref{fig:cmsfull} we show an example of this analysis for the
first rapidity bin of the CMS jets 2011 dataset. On the left plot of
Fig.~\ref{fig:cmsfull}, the full channel theoretical predictions are
normalized to the CMS data, corrected by non-perturbative corrections,
where uncertainties are estimated from the diagonal of the covariance
matrix, which is extracted by considering systematic uncertainties
additively.
From this plot we observe that the data is well described by the NLO predictions. We note that the NNPDF2.3 set used in this comparison is obtained from a NNLO fit that
includes jet data from the Tevatron and the LHC for which the corresponding theory predictions are known presently only to NLO accuracy. For this reason,
higher order theory effects beyond NLO are not taken into account in the jet prediction used in the fit. As a result, we observe that the approximate NNLO prediction based on threshold resummation
predicts a cross section above the data indicating the need to consistently include NNLO jet predictions in NNLO PDF fits of jet data. Part of this excess could be due to the inherent
approximated nature of this prediction and for this reason we aim to disentangle in the next sections the regions which correspond to a breakdown of the threshold approximation.
On the right plot of Fig.~\ref{fig:cmsfull} we quantify the size of
the higher order corrections by computing ratios of higher order cross
sections over the leading order one. These $k$-factors show that NLO
corrections vary between 20\% and 45\% with respect to the LO
prediction with the approximate NNLO threshold corrections varying
between 40\% and 70\%. We also observe that the NLO/LO $k$-factors
using the threshold (in green) and NJA (in red) codes are in good
agreement with the exact computation (in blue).
\begin{figure}
\begin{centering}
\includegraphics[scale=0.8]{cms-bin1}
\par\end{centering}
\caption{\label{fig:cmsfull} Comparison between CMS
data and theoretical predictions computed with the exact and
approximate codes for the first bin in $\eta$. On the left
plot the theoretical predictions are normalized to the CMS measurement. On the
right plot, $k$-factors for higher order cross sections over the leading order (LO) are presented. }
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.4]{cms-bin1-ratio}
\par\end{centering}
\caption{\label{fig:cmsfullratio} Ratios between exact and approximate
predictions at the same order in perturbation theory. The level of agreement at NLO takes into account all partonic channels in the predictions.
In the same plot we present the
exact NLO/LO $k$-factor (all partonic channels included) and the NNLO/NLO $k$-factors produced
by the threshold approximation code (all partonic channels included).}
\end{figure}
%
A more detailed comparison between the all channels exact and approximate predictions
is presented in Figure~\ref{fig:cmsfullratio}. We conclude that
as expected the NLO threshold (in green) and the NJA (in red) predictions converge to the exact computation
for high values of $p_{T}$. The high-$p_T$ region corresponds precisely to the threshold region $s_4=0$ where
the phase space available for additional radiation is limited. In the low-$p_T$ region we observe instead
an instability in the approximate predictions.
In the same figure we present the exact NLO/LO $k$-factor (in blue) and with
the approximate NNLO prediction we have constructed NNLO/NLO $k$-factors using two different choices for the
denominator: the approximate NLO threshold (in light magenta) and the NLO exact
(in dark magenta). As we can see
the effects due to this choice are at the few percent level, negligible in
comparison to the size of the approximate NNLO threshold correction.
We have performed this exercise for all experiments, however, in the
next subsections we limit the analysis to the gluon-gluon channel because we are
interested in determining the regions where the NNLO threshold
$k$-factors are in agreement with the exact computation, which is
available for that channel. The full report with all plots and tables
of $k$-factors is available on-line at:
\begin{center}
{\bf \url{http://libhfill.hepforge.org/JetStudy2014}~}
\end{center}
and a short summary of the main results which identify the experimental regions where approximate predictions
should be discarded is provided in Sect.~\ref{sec:PDFfit}.
\subsection{CMS jets}
In Figure~\ref{fig:cmsratio} we show the ratios between predictions in the gluons-only channel
computed with the codes presented in Sec.~\ref{sec:bench} using the
rapidity and $p_{T}$ bins of the CMS jets 2011 dataset.
The LO, NLO, NNLO predictions labelled exact are obtained from the Monte Carlo
NNLOJET presented in~\cite{Currie:2013dwa,Ridder:2013mf,Currie:2014upa} and are compared
with the NJA code~\cite{Jager:2004jh,Mukherjee:2012uz}
at NLO and with the threshold code~\cite{deFlorian:2013qia} at NLO and NNLO.
For all predictions we set the renormalisation and factorisation scales equal to each other
and equal to the $p_T$ of each individual jet in every event ($\mu_R=\mu_F=p_T$)
For all plots we first check the agreement at NLO between all
codes and then the agreement of the NNLO/NLO $k$-factors obtained with the threshold
approximation code and the exact computation. As we did in the
previous section, we provide two definitions for the NNLO threshold
$k$-factor by dividing the approximate NNLO threshold predictions with
the approximate NLO threshold predictions (in light magenta long-dashed curves) and also
by dividing them with the exact
NLO prediction (in dark magenta long-dashed curves). The NNLO/NLO $k$-factor
using the exact computation at NLO and NNLO is plotted in long-dashed black curves. The distance
between the long-dashed black curve and the long-dashed magenta curves in Figure~\ref{fig:cmsratio} and subsequent Figures
indicates the level of disagreement between the $k$-factors produced by the exact NNLO computation
and the approximate NNLO threshold computations.
By looking at Figure~\ref{fig:cmsratio} we conclude that at NLO the NJA
code shows percent level differences at small $p_{T}$. Similarly the NLO
threshold code shows percent level differences at small $p_T$ which rise to
5\% at central $p_{T}$ for the last bins in rapidity.
Concerning the NNLO predictions we looked at the NNLO $k$-factors and
relative differences bin by bin. The relative difference between the exact computation and the threshold computations
is documented in Tables~\ref{tab:kcms1} to~\ref{tab:kcms5} where, for each rapidity slice
of the experiment, we show for each $p_T$ bin, the experimental cross section, the experimental
error, the gluons-only exact NNLO and threshold NNLO $k$-factors together with their relative
percentage wise difference and finally the percentage wise relative difference between the two possible NNLO gluons-only
threshold $k$-factors.
We first notice that for the entire kinematic range of the experiment the choice of the
denominator for the NNLO threshold $k$-factor produces relative differences which are
much smaller than the difference to the exact $k$-factor.
When comparing either of these with the exact NNLO computation we find for all rapidity bins an instability
at low-$p_T$ in the approximate NNLO results. In these regions the approximate NNLO threshold $k$-factor starts to rise
generating large perturbative corrections. While for the first two bins in rapidity $|\eta|<1.0$, the relative differences with
the exact calculation are below 10\% we observe strong deviations for $|\eta|>1.0$.
Using the rejection criteria suggested at the end of Section~\ref{sec:benchTH} we conclude that for CMS
the NNLO threshold prediction is not applicable for the rapidity slices $|\eta|>1.5$ as
relative differences with respect to the exact computation are larger than 20\% and can rise up to 60\%.
Furthermore, due to the instability of the approximate prediction at low-$p_T$,
for the rapidity slice $1.0<|\eta|<1.5$ the first seven $p_T$ bins should be excluded.
As mentioned in the introduction, the comparison between the exact
fixed-order calculation and the threshold approximation performed
in~\cite{deFlorian:2013qia} used different central scale choices for
each prediction. The predictions from the threshold resummation
formalism were obtained using $\mu_R=\mu_F=p_T$, where $p_T$ is the
individual jet $p_T$ while the fixed-order calculation used
$\mu_R=\mu_F=p_{T1}$, where $p_{T1}$ is the transverse momentum of the
leading jet in each event. In order to eliminate this inconsistency,
and also to study the impact of the central scale choice in the
fixed-order predictions, we show in Figure~\ref{fig:CMSPTvsPT1} NLO
and NNLO gluons only cross sections evaluated at the two different
scales for the first rapidity slice of the CMS experiment. We observe
that at high-$p_T$ subleading jets tend to be soft and in this region
$p_{T1}\sim p_{T}$ and the predictions using either scale choice
coincide. In the low-$p_T$ region we observe low-$p_T$ jets
accompanying a high-$p_T$ object. In this case $p_{T} < p_{T1} $ and
we observe an increase in the NLO prediction of about 5\% when using
the scale $\mu_R=\mu_F=p_T$. This effect is due to the fact that the
virtual contribution which has Born kinematics remains identical with
either scale choice while the real radiation contribution is enhanced
when its weight is computed at a lower scale. At NNLO we observe
instead a reduction of the size of the fixed-order prediction at
low-$p_T$. As a consequence we conclude that the NNLO/NLO $k$-factor
is typically smaller in the low-$p_T$ region with the scale choice
$\mu_R=\mu_F=p_T$ as compared to the choice $\mu_R=\mu_F=p_{T1}$. The
resulting reduced NNLO $k$-factor for the exact prediction then shows
that the disagreement between the exact calculation and the threshold
calculation is enhanced when both calculations are performed using the
same central scale choice. This observation is rapidity independent as
emissions in events with a high-$p_T$ central object can produce
low-$p_T$ jets entering the single jet inclusive $p_T$ distribution in
the forward regions.
\begin{figure}[!h]
\begin{centering}
\includegraphics[scale=0.4]{cms-bin1gg-ratio}\includegraphics[scale=0.4]{cms-bin2gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{cms-bin3gg-ratio}\includegraphics[scale=0.4]{cms-bin4gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{cms-bin5gg-ratio}
\par\end{centering}
\caption{\label{fig:cmsratio} Ratios between exact and approximate
predictions at the same order (LO and NLO) in perturbation theory in the gluons-only channel.
In the same plot we present the
exact NLO/LO and NNLO/NLO $k$-factors (gluons-only channel) and the NNLO/NLO $k$-factors produced
by the threshold approximation code (gluons-only channel) for the CMS 2011 jet binning.}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.4]{CMS-NLO-PTvsPT1}\includegraphics[scale=0.4]{CMS-NNLO-PTvsPT1}
\par\end{centering}
\caption{\label{fig:CMSPTvsPT1} NLO (left) and NNLO (right) exact $gg$-channel predictions for CMS evaluated with the renormalisation and factorisation scales
$\mu_{R}=\mu_{F}=p_T$ and $\mu_{R}=\mu_{F}=p_{T1}$. In the lower pads we present the relative differences due to the different central scale choice.}
\end{figure}
\subsection{ATLAS jets at $\sqrt{s}=7$ TeV and $\sqrt{s}=2.76$ TeV}
The main difference between ATLAS and CMS jets is the different
kinematic coverage as shown in Table~\ref{tab:experiments}. ATLAS provide jet data at two distinct center of
mass energies and kinematic ranges.
Figures~\ref{fig:atlas7ratio} and~\ref{fig:atlas2ratio} present the gluons-only theoretical predictions
for ATLAS jets at $\sqrt{s}=7$ TeV and $\sqrt{s}=2.76$ TeV
respectively. For all predictions we set the renormalisation and factorisation scales equal to each other
and equal to the $p_T$ of each individual jet in every event ($\mu_R=\mu_F=p_T$). At NLO the NJA (narrow-jet approximation) prediction shows
good agreement with the exact NLO result across the entire kinematic range. We note that the ATLAS
experiment employs R=0.4 for the anti-$k_t$ jet clustering procedure and therefore the agreement
between NJA approximation and the exact calculation is expected and observed to improve for smaller values of R.
The threshold prediction at NLO is also in good agreement with the exact calculation but shows an evident
instability at small $p_T$ where discrepancies vary between 10\% and 40\% increasing with rapidity.
The predictions at NNLO are in worse agreement when compared to the
results obtained at CMS. There is a constant gap between the threshold and the exact
predictions for ATLAS jets at $\sqrt{s}=7$ TeV and $\sqrt{s}=2.76$ TeV. The convergence
of the threshold approximation code to the exact
computation is not evident as it is for CMS because the maximum $p_{T}$ values are
smaller for ATLAS.
On Tables~\ref{tab:katlas71} to~\ref{tab:katlas77} in the Appendix we document
the NNLO $k$-factor results. The general behaviour is similar to CMS:
differences between the exact NNLO computation and the approximate NNLO threshold computation
are large at small $p_{T}$ (between 20\% and 80\%) and increase with
rapidity. For the rapidity range $0.8<|\eta|<2.1$ and for all $p_T$ the disagreement between the two predictions is between $10\%-100\%$.
For the rapidity regions $|\eta|>2.1$ the disagreement is larger than $100\%$ for all $p_T$. An application of the rejection criteria suggested before
excludes approximate predictions for the $p_{T}$ points for the first ten bins with
$|\eta|<0.3$ and the first thirteen points with $0.3<|\eta|<0.8$. In the regions of rapidity $|\eta|>0.8$ we should discard all points as differences
between the exact and approximate calculation are for all values of jet $p_T$ much larger than 10\%.
In particular, and contrary to statements made in Ref.~\cite{deFlorian:2013qia}, the large NNLO $k$-factors observed in the approximate threshold calculation (Figure~\ref{fig:atlas7ratio})
of the order 5 or so at $\eta\sim4$ are not present in the exact NNLO calculation. As can be seen in Figure~\ref{fig:atlas7ratio} while the exact NNLO $k$-factor decreases
with rapidity, the NNLO $k$-factor of the approximate NNLO calculation increases with rapidity.
The comparison between the NNLO $\sqrt{s}=7$ TeV ATLAS $k$-factors and the
NNLO $\sqrt{s}=2.76$ TeV ATLAS $k$-factors show a moderate dependence on the center of mass
energy, with the predictions at $\sqrt{s}=2.76$ TeV giving slightly smaller NNLO $k$-factors.
With the results documented in Tables~\ref{tab:katlas21} to~\ref{tab:katlas27} and by
extending the rejection criteria suggested to $\sqrt{s}=7$
TeV to ATLAS at $\sqrt{s}=2.76$ TeV we conclude that
the approximate calculation gives acceptable predictions only for the first rapidity slice $|\eta|<0.5$ after
removing the first eight $p_{T}$ bins.
\begin{figure}[H]
\begin{centering}
\includegraphics[scale=0.38]{atlas7-bin1gg-ratio}\includegraphics[scale=0.38]{atlas7-bin2gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas7-bin3gg-ratio}\includegraphics[scale=0.38]{atlas7-bin4gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas7-bin5gg-ratio}\includegraphics[scale=0.38]{atlas7-bin6gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas7-bin7gg-ratio}
\par\end{centering}
\caption{\label{fig:atlas7ratio}Ratio plots for $gg$-channel predictions together with NNLO $k$-factors for ATLAS jets at $\sqrt{s}=7$ TeV kinematics.}
\end{figure}
\begin{figure}[H]
\begin{centering}
\includegraphics[scale=0.38]{atlas2-bin1gg-ratio}\includegraphics[scale=0.38]{atlas2-bin2gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas2-bin3gg-ratio}\includegraphics[scale=0.38]{atlas2-bin4gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas2-bin5gg-ratio}\includegraphics[scale=0.38]{atlas2-bin6gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.38]{atlas2-bin7gg-ratio}
\par\end{centering}
\caption{\label{fig:atlas2ratio}Ratio plots for $gg$-channel predictions together with NNLO $k$-factors for ATLAS jets at $\sqrt{s}=2.76$ TeV kinematics.}
\end{figure}
We conclude this section by comparing the exact fixed-order predictions in the $gg$-channel evaluated at the scales $\mu_R=\mu_F=p_{T1}$ and $\mu_R=\mu_F=p_{T}$.
In Figure~\ref{fig:atlas7PTvsPT1} we show NLO and NNLO gluons only cross sections evaluated at the two different scales for the first rapidity slice of the ATLAS experiment. The
qualitative effect is similar to the one showed for CMS in the previous section. By choosing $\mu_R=\mu_F=p_T$ as a central scale we observe that at low-$p_T$ the NLO prediction
increases by about 10\% while the NNLO prediction is reduced by around 20\%, with respect to the results obtained using $\mu_R=\mu_F=p_{T1}$.
Predictions at high-$p_T$ are as expected identical with either scale choice. We note that the due to the $p_T$ coverage by ATLAS being more extreme than at CMS
(nearly two orders of magnitude in $p_T$ covered by ATLAS) we are at low-$p_T$ more often in the kinematical regions where
$p_{T} \ll p_{T1}$ and therefore the effects that result from changing the central scale choice for the predictions
from $\mu=p_{T1}$ to $\mu=p_T$ are enhanced.
In Figure~\ref{fig:kfactPTvsPT1} we show the exact NLO/LO and NNLO/NLO
$k$-factors for each scale choice. Interestingly we observe that for
the first bin in $p_T$ the perturbative series behaves for
$\mu_{R}=\mu_{F}=p_T$ as 14\% NLO correction with respect to LO and
14\% NNLO corrections with respect to NLO. This compares with 3\% NLO
corrections with respect to LO and 54\% NNLO corrections with respect
to NLO when using $\mu_{R}=\mu_{F}=p_{T1}$. Therefore, the
convergence of the perturbative series in the fixed order calculation
is improved using $\mu_{R}=\mu_{F}=p_T$ as the NNLO/NLO $k$-factor is
smaller than the NLO/LO $k$-factor for all $p_{T}$ and rapidity.
In Figure~\ref{fig:scale} we study the scale dependence of the exact calculation
in the $gg$-channel at each order in perturbation theory for jets with $|y|<4.4$ and
80 GeV$<p_{T}<$ 97 GeV. For this study we have employed the same kinematical
setup used in~\cite{Ridder:2013mf} and changed only the renomalisation and factorization central scales
to generate the predictions to be $\mu_R=\mu_F=p_T$ instead of the leading jet $p_{T1}$.
We observe again in this kinematical setup that changing the central scale choice to $\mu_R=\mu_F=p_T$
results in the NNLO/NLO $k$-factor becoming smaller as the blue curve (NNLO prediction)
is closer to the red curve (NLO prediction) with this scale choice.
We observe also a reduction of the scale dependence of the cross section at NNLO.
Similarly to the results presented for CMS in the previous section we
conclude that the disagreement between the exact calculation and the
threshold calculation is enhanced when both calculations are performed
using the same central scale choice. In particular, by cutting away
kinematical regions where the disagreement between both calculations
is larger than 10\% we keep only data points for which the NNLO
$k$-factors are typically smaller $\sim1.1-1.2$. As we will shown in
Sect.~\ref{sec:PDFfit} this has the consequence of improving the
$\chi^2$ of an aNNLO PDF fit including the approximate NNLO single jet
inclusive prediction.
\begin{figure}
\begin{centering}
\includegraphics[scale=0.38]{ATLAS-NLO-PTvsPT1}\includegraphics[scale=0.38]{ATLAS-NNLO-PTvsPT1}
\par\end{centering}
\caption{\label{fig:atlas7PTvsPT1} NLO (left) and NNLO (right) exact $gg$-channel predictions for ATLAS evaluated with the renormalisation and factorisation scales
$\mu_{R}=\mu_{F}=p_T$ and $\mu_{R}=\mu_{F}=p_{T1}$. In the lower pads we present the relative differences due to the different central scale choice.}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.38]{atlas7-bin1gg-PTvsPT1}
\par\end{centering}
\caption{\label{fig:kfactPTvsPT1} NLO/LO and NNLO/NLO exact $k$-factors for the $gg$-channel evaluated with the renormalisation and factorisation scales
$\mu_{R}=\mu_{F}=p_T$ and $\mu_{R}=\mu_{F}=p_{T1}$.}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.38]{scale-PT}\includegraphics[scale=0.38]{scale-PT1}
\par\end{centering}
\caption{Scale dependence of the inclusive jet cross section for $pp$ collisions at $\sqrt{s}=8~$TeV for
the anti-$k_T$ algorithm with $R=0.7$ at NNLO (blue), NLO (red) and LO (green) and with $|y|<4.4$ and
80 GeV$<p_{T}<$ 97 GeV for different central scale choices:
$\mu_R=\mu_F=p_{T}$ on the left and $\mu_R=\mu_F=p_{T1}$ on the right.}
\label{fig:scale}
\end{figure}
\section{Exclusion criteria summary}
\label{sec:PDFfit}
In the previous sections we performed a comparison at NLO and NNLO
between the exact predictions and the approximate predictions from the
threshold resummation formalism at the same order. We performed this
comparison for all experimental hadron-hadron collider setups at the
LHC and the Tevatron from which a wealthy dataset of inclusive jet
data has been delivered.
This comparison was performed in the $gg$-channel where the exact NNLO
results were delivered first~\cite{Currie:2013dwa,Ridder:2013mf} and
therefore can be used to identify the regions where the approximate
NNLO prediction emulates the exact results. As a result of this
exercise we include in this work the full set of NLO/LO and NNLO/NLO
$k$-factors based on both approaches for the $gg$-channel and also the
approximate NNLO all-channel prediction. Using these predictions we
can identify kinematical regions where discrepancies between the two
results in the $gg$-channel are larger than $\delta$=10\%. In such
regions we conclude that the results of the approximate prediction
should not be trusted and for this reason, leads to an exclusion of
part of the full dataset of single jet inclusive cross section
measurements at hadron-hadron colliders that can be analysed. We
studied the effect of being more restrictive or relaxed with this
criteria and summarise experiment by experiment the resulting
exclusion regions of experimental data as a function of $\delta$ in
Tables~\ref{tab:PTpoints} to~\ref{tab:PTpointsCDF}. As expected,
being more restrictive and demanding a smaller relative difference
between exact and threshold $k$-factors leads to an increased
exclusion region of data points. Using $\delta$=5\% results in
excluding more than half the data points from CMS and all data points
from ATLAS. This information is also reproduced in
Figure~\ref{fig:exclusion} where we show the relative difference
between the exact and approximate NNLO $k$-factor in the $gg$-channel
in the $(p_{T},|y|)$ plane of each experiment.
\begin{figure}[t]
\begin{centering}
\includegraphics[scale=0.35]{scatterCMS}\includegraphics[scale=0.35]{scatterATLAS7}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.35]{scatterATLAS2}\includegraphics[scale=0.35]{scatterCDF}
\par\end{centering}
\caption{
\label{fig:exclusion} Percentage-wise relative difference $|\delta|$
between the exact and approximate $gg$-channel NNLO $k$-factors as a
function of $p_{T}$ and $|y|$ for CMS, ATLAS 7 TeV and 2.76 TeV and
CDF bins. Disagreements larger than $|\delta|=100\%$ are represented
in red.}
\end{figure}
With this information at hand we aim in this section to perform
template aNNLO PDF fits of jet data including the approximate NNLO
corrections. We choose to perform four fits using a criteria of less
than 5\%, 7.5\% 10\% and 15\% relative disagreement between the exact
and the threshold $gg$-channel $k$-factors in order to exclude
approximate corrections, and assess the impact of these choices on the
quality of the fit. The kinematical regions which survive each cut are
introduced in the fit through the full channel threshold
approximation\footnote{Full channel approximate $k$-factors available
at: \url{http://libhfill.hepforge.org/JetStudy2014}}. The study of
the effect in the fit of performing these cut variations also provides a
way for an empiric determination of the best exclusion criteria. As
mentioned in Sect.~\ref{sec:tev} data from the D0 experiment will not
be included in this exercise.
To perform the fits we have used the latest {\tt NNPDF} fitting
technology~\cite{Ball:2012cx} to fit partial subsets of jet data
determined by the relative difference criteria $\delta$. Moreover we
have also eliminated a few points at large rapidity bins where the
full channel $k$-factor is orders of magnitude larger that the
$gg$-channel $k$-factor and considered only datasets which after
applying the cuts contain at least 2 data points.
In Tables~\ref{tab:PTpoints} to~\ref{tab:PTpointsCDF} we present as a
function of $\delta$ the experimental $\chi^2/\textrm{dof}$ obtained
when fitting the respective subset of data points. For CMS the fits
tend to include more data points, in particular in high-$p_{T}$
regions and fairly central $|y|<1.0$ jets. In these regions, the jet
data is probing kinematics not constrained by other data and we
observe that the $\chi^2/\textrm{dof}$ has a small $|\delta|$
dependence. For CDF a large fraction of data points is included in the
fit. In this case, however, we noticed that the $\chi^2/\textrm{dof}$
improves when $\delta$ is reduced. Finally, for the ATLAS data, we
observe that by reducing $\delta$ a large fraction of data points is
excluded, and this results in large $\chi^2/\textrm{dof}$
fluctuations. In conclusion, with these results we cannot find a
precise exclusion criteria. However, we suggest a possible compromise
of $|\delta|=10\%$ which allows the inclusion of some data from all
experiments, providing a reasonable and stable $\chi^2/\textrm{dof}$
in all cases. In this way within the tolerance error chosen,
perturbative QCD corrections can be included in PDF fits and result in
a reduction of the gluon PDF uncertainty at high-$x$.
Finally we would like to point out that by computing the
$qg$~\cite{nnlojet:2014hp} and $qq$~\cite{nnlojet:2014hp} scattering
processes to exact NNLO accuracy in the fixed-order calculation along
the lines followed for $gg$~\cite{Currie:2013dwa,Ridder:2013mf} and
$q\bar{q}$~\cite{Currie:2013vh,Currie:2014upa} scattering, we avoid
the necessity of introducing a rejection criteria to exclude/include
approximate higher order predictions. Instead, the exact prediction
from the fixed-order calculation allows the use of the full dataset of
jet cross section measurements in global NNLO PDF fits of jet
data. Moreover we can test the resulting fit quality in the
description of any fully differential 2-jet observable at NNLO. We
leave this study for future work.
\begin{table}[H]
\begin{centering}
\begin{tabular}{l|c|c|c|l}
\hline
CMS 2011 & $N_{{\rm dat}}$ & $\chi^{2}/$dof & \multicolumn{2}{c}{Exclusion regions ($y,p_{T}$)}\tabularnewline
\hline
\hline
\multirow{2}{*}{$|\delta|<15\%$} & \multirow{2}{*}{88} & \multirow{2}{*}{1.81} & $1.0<|y|<1.5$ & $p_{T}<153$ GeV\tabularnewline
& & & $|y|>1.5$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<10\%$} & \multirow{2}{*}{83} & \multirow{2}{*}{1.89} & $1.0<|y|<1.5$ & $p_{T}<272$ GeV\tabularnewline
& & & $|y|>1.5$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{3}{*}{$|\delta|<7.5\%$} & \multirow{3}{*}{77} & \multirow{3}{*}{1.89} & $0.5<|y|<1.0$ & $p_{T}<153$ GeV\tabularnewline
& & & $1.0<|y|<1.5$ & $p_{T}<395$ GeV\tabularnewline
& & & $|y|>1.5$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{3}{*}{$|\delta|<5\%$} & \multirow{3}{*}{59} & \multirow{3}{*}{1.83} & $0.5<|y|<1.0$ & $p_{T}<220$ GeV\tabularnewline
& & & $1.0<|y|<1.5$ & $p_{T}<737$ GeV, $p_{T}>790$ GeV \tabularnewline
& & & $|y|>1.5$ & all $p_{T}$ bins\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{
\label{tab:PTpoints} Summary of exclusion regions in $p_T$ and rapidity $|y|$ as a
function of the relative difference between exact and threshold
$k$-factors for the gluon-gluon channel for the CMS 133 data points.
In the table we quote the $\chi^2/\textrm{dof}$ for aNNLO PDF fits performed
with the full channel approximated $k$-factors. $N_{{\rm dat}}$ represents the number of experimental
data points included in the fit.}
\end{table}
\begin{table}[H]
\begin{centering}
\begin{tabular}{l|c|c|c|l}
\hline
ATLAS 2.76 TeV & $N_{{\rm dat}}$ & $\chi^{2}/$dof & \multicolumn{2}{c}{Exclusion regions ($y,p_{T}$)}\tabularnewline
\hline
\hline
\multirow{3}{*}{$|\delta|<15\%$} & \multirow{3}{*}{10} & \multirow{3}{*}{2.15} & $0.0<|y|<0.3$ & $p_{T}<110$ GeV\tabularnewline
& & & $0.3<|y|<0.8$ & $p_{T}<210$ GeV\tabularnewline
& & & $|y|>0.8$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<10\%$} & \multirow{2}{*}{3} & \multirow{2}{*}{0.35} & $0.0<|y|<0.3$ & $p_{T}<260$ GeV\tabularnewline
& & & $|y|>0.3$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{1}{*}{$|\delta|<7.5\%$} & \multirow{1}{*}{-} & \multirow{1}{*}{-} & all $|y|$ bins & all $p_{T}$ bins\tabularnewline
\hline
\multirow{1}{*}{$|\delta|<5\%$} & \multirow{1}{*}{-} & \multirow{1}{*}{-} & all $|y|$ bins & all $p_{T}$ bins\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{Summary of exclusion regions in $p_T$ and rapidity $|y|$ as a
function of the relative difference between exact and threshold
$k$-factors for the gluon-gluon channel for the ATLAS 2.76 TeV 59 data points.
In the table we quote the $\chi^2/\textrm{dof}$ for aNNLO PDF fits performed
with the full channel approximated $k$-factors. $N_{{\rm dat}}$ represents the number of experimental
data points included in the fit.}
\end{table}
\begin{table}[H]
\begin{centering}
\begin{tabular}{l|c|c|c|l}
\hline
ATLAS 7 TeV & $N_{{\rm dat}}$ & $\chi^{2}/$dof & \multicolumn{2}{c}{Exclusion regions ($y,p_{T}$)}\tabularnewline
\hline
\hline
\multirow{4}{*}{$|\delta|<15\%$} & \multirow{4}{*}{16} & \multirow{4}{*}{1.82} & $0.0<|y|<0.3$ & $p_{T}<260$ GeV\tabularnewline
& & & $0.3<|y|<0.8$ & $p_{T}<400$ GeV\tabularnewline
& & & $0.8<|y|<1.2$ & $p_{T}<1$ TeV\tabularnewline
& & & $|y|>1.2$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{3}{*}{$|\delta|<10\%$} & \multirow{3}{*}{9} & \multirow{3}{*}{1.58} & $0.0<|y|<0.3$ & $p_{T}<400$ GeV\tabularnewline
& & & $0.3<|y|<0.8$ & $p_{T}<800$ GeV\tabularnewline
& & & $|y|>0.8$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<7.5\%$} & \multirow{2}{*}{5} & \multirow{2}{*}{2.02} & $0.0<|y|<0.3$ & $p_{T}<500$ GeV\tabularnewline
& & & $|y|>0.8$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<5\%$} & \multirow{2}{*}{1} & \multirow{2}{*}{-} & $0.0<|y|<0.3$ & $p_{T}<1$ TeV, $p_{T}>1.2$ TeV\tabularnewline
& & & $|y|>0.3$ & all $p_{T}$ bins\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{Summary of exclusion regions in $p_T$ and rapidity $|y|$ as a
function of the relative difference between exact and threshold
$k$-factors for the gluon-gluon channel for the ATLAS 7 TeV 90 data points.
In the table we quote the $\chi^2/\textrm{dof}$ for aNNLO PDF fits performed
with the full channel approximated $k$-factors. $N_{{\rm dat}}$ represents the number of experimental
data points included in the fit.}
\end{table}
\begin{table}[H]
\begin{centering}
\begin{tabular}{l|c|c|c|l}
\hline
CDF & $N_{{\rm dat}}$ & $\chi^{2}/$dof & \multicolumn{2}{c}{Exclusion regions ($y,p_{T}$)}\tabularnewline
\hline
\hline
\multirow{2}{*}{$|\delta|<15\%$} & \multirow{2}{*}{60} & \multirow{2}{*}{2.32} & $1.1<|y|<1.6$ & $p_{T}<96$ GeV\tabularnewline
& & & $|y|>1.6$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<10\%$} & \multirow{2}{*}{52} & \multirow{2}{*}{1.86} & $1.1<|y|<1.6$ & $p_{T}<224$ GeV, $p_{T}>298$ GeV\tabularnewline
& & & $|y|>1.6$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<7.5\%$} & \multirow{2}{*}{48} & \multirow{2}{*}{1.37} & $0.7<|y|<1.1$ & $p_{T}<72$ GeV\tabularnewline
& & & $|y|>1.1$ & all $p_{T}$ bins\tabularnewline
\hline
\multirow{2}{*}{$|\delta|<5\%$} & \multirow{2}{*}{45} & \multirow{2}{*}{1.28} & $0.7<|y|<1.1$ & $p_{T}<110$ GeV\tabularnewline
& & & $|y|>1.1$ & all $p_{T}$ bins\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{\label{tab:PTpointsCDF} Summary of exclusion regions in $p_T$ and rapidity $|y|$ as a
function of the relative difference between exact and threshold
$k$-factors for the gluon-gluon channel for the CDF 76 data points.
In the table we quote the $\chi^2/\textrm{dof}$ for aNNLO PDF fits performed
with the full channel approximated $k$-factors. $N_{{\rm dat}}$ represents the number of experimental
data points included in the fit.}
\end{table}
\section{Tevatron jet data}
\label{sec:tev}
\subsection{CDF jets}
The gluons-only theory predictions for the CDF setup are presented in
Figure~\ref{fig:cdfratio}. We observe that the level of agreement at NLO is
better with respect to the LHC comparisons presented in the previous
section. As before, Tables~\ref{tab:kcdf1} to~\ref{tab:kcdf5} show the
$k$-factors for the $gg$-channel, where non-perturbative corrections have
been applied to the experimental data as performed for ATLAS and CMS.
At NNLO the situation has improved since differences between exact and
threshold approximation results are smaller than what we have observed
for the LHC experiments. With a rejection criteria of excluding points where the disagreement
is larger than 10\% we observe that the last two rapidity slices for $|\eta|>1.1$ should be excluded.
The main differences between the Tevatron and LHC setups are the different
center of mass energies, the projective particles and kinematic ranges
which are shorter at the Tevatron than at the LHC. For these reasons, the threshold approximation
code provides at the Tevatron predictions closer to the exact NNLO calculation.
\begin{figure}
\begin{centering}
\includegraphics[scale=0.4]{cdf-bin1gg-ratio}\includegraphics[scale=0.4]{cdf-bin2gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{cdf-bin3gg-ratio}\includegraphics[scale=0.4]{cdf-bin4gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{cdf-bin5gg-ratio}
\par\end{centering}
\caption{\label{fig:cdfratio} Ratios between exact and approximate
predictions at the same order (LO and NLO) in perturbation theory in the gluons-only channel.
In the same plot we present the
exact NLO/LO and NNLO/NLO $k$-factors (gluons-only channel) and the NNLO/NLO $k$-factors produced
by the threshold approximation code (gluons-only channel) for the CDF $\sqrt{s}=1.96$ TeV jet binning~\cite{Abulencia:2007ez}.}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.4]{d0con-bin1gg-ratio}\includegraphics[scale=0.4]{d0con-bin2gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{d0con-bin3gg-ratio}\includegraphics[scale=0.4]{d0con-bin4gg-ratio}
\par\end{centering}
\begin{centering}
\includegraphics[scale=0.4]{d0con-bin5gg-ratio}\includegraphics[scale=0.4]{d0con-bin6gg-ratio}
\par\end{centering}
\caption{\label{fig:d0ratio} Ratios between exact and approximate
predictions at the same order (LO and NLO) in perturbation theory in the gluons-only channel.
In the same plot we present the
exact NLO/LO and NNLO/NLO $k$-factors (gluons-only channel) and the NNLO/NLO $k$-factors produced
by the threshold approximation code (gluons-only channel) for the D0 $\sqrt{s}=1.96$ TeV jet binning~\cite{Abazov:2011vi}.}
\end{figure}
\subsection{D0 jets}
The results of the comparison for the D0 experimental setup are
presented in Figure~\ref{fig:d0ratio} with the respective
Tables~\ref{tab:kd01} to~\ref{tab:kd06}. As for CDF, the NLO
prediction is in good agreement for all rapidity and $p_{T}$
bins. The NNLO predictions behave similarly to the CDF results, with
the threshold approximation code providing acceptable predictions at
least for the first three rapidity slices. Predictions for the D0
experiment have been generated using the $k_{t}$ algorithm for the jet
reconstruction instead of the MidPoint cone used for the measurement
because this algorithm is IR unsafe at NNLO.
This represents a drawback towards analysing jet data from D0 at
NNLO. At the moment D0 data has been included in PDF fits where the IR
finiteness at NLO of the MidPoint cone jet algorithm
allows the perturbative computation to be performed at this order. A possible solution to include a perturbative prediction at NNLO for the D0 data could be identifying
a relationship between the MidPoint algorithm and an IR safe cone algorithm such as SISCone. As discussed in~\cite{Salam:2007xv} relative differences between the
two algorithms are expected to the start at ${\cal O}(\alpha_s^4)$ when we have 3 particles in a common neighbourhood plus one to balance the momentum. By producing an
inclusive jet $p_T$ spectrum using just tree-level $2\to4$ diagrams, differences at the level of 1-2\% are observed at the Tevatron in~\cite{Salam:2007xv} between the predictions of both jet algorithms
when the SISCone algorithm is employed with cone parameters R=0.7 and $f$=0.5. A detailed study applying this prescription for the NNLO inclusive jet prediction
is beyond the scope of the current work.
|
\section{Introduction}
Markov automata (MAs, for short) have been introduced in~\cite{EHZ10}
as a continuous-time version of Segala's probabilistic
automata~\cite{Seg95b}. Closed under operators such as parallel
composition and hiding, they provide a compositional formalism for
concurrent soft real time systems. A transition in an MA is either
labelled with a positive real number representing the rate of a
negative exponential distribution, or with an action. An action
transition leads to a discrete probability distribution over states.
MAs can thus model action transitions as in labelled transition
systems, probabilistic branching as found in (discrete time) Markov
chains and Markov decision processes, as well as delays that are
governed by exponential distributions as in continuous-time Markov
chains.
The semantics of MAs has been recently investigated in quite some
detail. Weak and strong (bi)simulation semantics have been presented
in~\cite{EHZ10,EHZ10b}, whereas it is shown
in~\cite{DBLP:journals/iandc/DengH13} that weak bisimulation provides
a sound and complete proof methodology for reduction barbed
congruence. A process algebra with data for the efficient modelling
of MAs, accompanied with some reduction techniques using static
analysis, has been presented in~\cite{MAPA}, and model checking of MAs
against Continuous Stochastic Logic (CSL) is discussed
in~\cite{HatefiH12}. Although the MA model raises several challenging
theoretical issues, both from a semantical and from an analytical
point of view, our main interest is in their practical applicability.
As MAs extend Hermanns' interactive Markov chains
(IMCs)~\cite{Hermanns02}, they inherit IMC application domains,
ranging from GALS hardware designs~\cite{CosteHLS09} and dynamic fault
trees~\cite{DBLP:journals/tdsc/BoudaliCS10} to the standardised
modelling language AADL~\cite{Bozzano,HaverkortKRRS10}. The
additional feature of probabilistic branching yields additional
expressivity and thereby enriches the spectrum of application contexts
further. This expressivity also makes them a natural semantic model
for other formalisms. Among others, MAs are expressive enough to
provide a natural operational model for generalised stochastic Petri
nets (GSPNs)~\cite{MCB84} and stochastic activity networks
(SANs)~\cite{MeyerMS85}, both popular modelling formalisms for
performance and dependability analysis. Let us briefly motivate this
by considering GSPNs. Whereas in SPNs all transitions are subject to
an exponentially distributed delay, GSPNs also incorporate immediate
transitions, transitions that happen instantaneously.
\begin{figure}[t!]
\centering
{}\hfill
\subfigure[]{
\scalebox{0.95}{\input{./confused-net.tikz}}
\label{fig:gspn-confused}
} \hfill
\subfigure[]{
\tikzstyle{every picture}=[thick, scale=0.78, transform shape]
\tikzstyle{every loop}=[->]
\tikzstyle{every scope}=[>=latex]
\centering
\begin{tikzpicture}[node distance=2.9cm,scale=0.9,every node/.style={transform shape}]
\tikzstyle{every label}=[label distance=0pt]
\node[initial,state, ellipse, minimum width=43pt,fill=gray!40!white] (p12) {$p_1,p_2$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,above right of=p12,xshift=-0.5cm, yshift=-.5cm] (p23) {$p_2,p_3$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,below right of=p12,xshift=-0.5cm, yshift=.5cm] (p15) {$p_1,p_5$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,right of=p23] (p4) {$p_4$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,right of=p4] (p6) {$p_6$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,right of=p15] (p35) {$p_3,p_5$};
\node[state, ellipse, minimum width=43pt,fill=gray!40!white,right of=p35] (p37) {$p_3,p_7$};
\path[->]
(p4) edge[thin,double] node[auto] {$\lambda_1$} (p6)
(p35) edge[thin,double] node[auto] {$\lambda_2$} (p37);
\path[->]
(p12) edge[thin,bend left=20] node[auto] {$\tau$} (p23)
(p12) edge[thin,bend right=20] node[auto,swap] {$\tau$} (p15);
\path[->]
(p15) edge[thin] node[auto] {$\tau$} (p35);
\path[->]
(p23) edge[thin] node[auto] {$\dfrac{w_3}{w_2+w_3}$} (p4)
(p23) edge[thin] node[auto,swap,yshift=0.3cm] {$\dfrac{w_2}{w_2+w_3}$} (p35);
\path[-]
(p23) edge[thin] node[inner sep=0mm,pos=0.2] (a1) {} (p4)
(p23) edge[thin] node[inner sep=0mm,pos=0.2] (b1) {} (p35);
\path[-,shorten <=-.4pt,shorten >=-.4pt] (a1) edge [thin,bend left] (b1) node[right,yshift=-.6cm] {$\tau$} ;
\end{tikzpicture}
\hfill{}
\label{fig:gspn-ma-semantics}
}
\caption{(a) Confused GSPN \cite[Fig.\ 21]{Mar95} with partial weights and (b) its MA semantics.}
\end{figure}
The traditional GSPN semantics yields a continuous-time Markov chain
(CTMC), i.e., an MA without action transitions. However, that
semantics is restricted to a subclass of GSPNs, namely those that are
\emph{confusion free}. Confusion~\cite{Mar95} is related to the
presence of nondeterminism. Confused GSPNs are traditionally
considered as semantically ambiguous and thus precluded from any kind
of analysis. This gap is particularly disturbing because several
published semantics for higher-level modelling formalisms---e.g., UML,
AADL, WSDL---map onto GSPN{}s without ensuring the mapping to be free
of confusion, therefore possibly inducing confused models.
It has recently been detailed in~\cite{Katoen12,EHKZ13} that MAs are a
natural semantic model for \emph{every} GSPN. To give some intuitive
insight into this achievement, consider the GSPN in
Fig.~\ref{fig:gspn-confused}. This net is confused: In Petri net
jargon, the transitions $t_1$ and $t_2$ are not in conflict, but
firing transition $t_1$ leads to a conflict between $t_2$ and $t_3$,
which does not occur if $t_2$ fires before $t_1$. Though decisive, the
firing order between $t_1$ and $t_2$ is
not determined. Transitions $t_2$
and $t_3$ are weighted so that in a marking $\{ p_2, p_3 \}$ in which
both transitions are enabled, $t_2$ fires with probability
$\frac{w_2}{w_2{+}w_3}$ and $t_3$ with its complement
probability.
The weight of transition $t_1$ is not relevant; we assume $t_1$ is not
equipped with a weight.
Classical GSPN semantics and analysis algorithms cannot cope with this
net due to the presence of confusion (i.e., nondeterminism).
Figure~\ref{fig:gspn-ma-semantics} depicts the MA semantics of this
net. Here, states correspond to sets of net places that contain a
token. In the initial state, there is a nondeterministic choice
between the transitions $t_1$ and $t_2$. Note that the presence of
weights is naturally represented by discrete probabilistic branching
as reflected in the outgoing transition from state $\{ p_2, p_3 \}$.
One can show that the MA semantics conservatively extends the
classical semantics, in the sense that the former and the latter are
weakly bisimilar~\cite{EHKZ13} on confusion-free GSPNs.
Thus, if transition $t_1$ in our example is assigned some weight $w_1$,
the GSPN has no confusion. This would be reflected in the MA semantics
by replacing the nondeterministic branching in state $\{ p_1, p_2 \}$ by
a single transition, yielding $\{ p_2, p_3 \}$ with probability $\frac{w_1}{w_1{+}w_2}$
and state $\{ p_1, p_5 \}$ with the complement probability.
This paper focuses on the quantitative analysis of MAs---and thus
implicitly of (possibly confused) GSPNs, of AADL specifications containing error models, and
so on.
We present analysis algorithms for three objectives: expected time, long-run average, and timed (interval) reachability.
As the model exhibits nondeterminism, we focus on maximal and minimal values for all three objectives.
We show that expected-time and long-run average objectives can be efficiently reduced to well-known problems on MDPs such as stochastic shortest path, maximal end-component decomposition, and long-run ratio objectives.
This generalises (and slightly improves) the results reported in~\cite{DBLP:conf/nfm/GuckHKN12} for IMCs to MAs.
Secondly, we present a discretisation algorithm for timed interval reachability objectives which extends~\cite{DBLP:conf/tacas/ZhangN10}.
Finally, we present the \toolname\ tool chain, an easily accessible publicly available tool chain\footnote{Stand-alone download as well as web-based interface available from \url{http://fmt.cs.utwente.nl/~timmer/mama}.} for the specification, mechanised simplification---such as confluence reduction~\cite{ConfluenceMA}, a form of on-the-fly partial-order reduction---and quantitative evaluation of MAs.
We describe the overall architectural design, as well as the tool components, and report on empirical results obtained with \toolname\ on a selection of case studies taken from different domains.
The experiments give insight into the effectiveness of the reduction techniques in \toolname\ and demonstrate that MAs provide the basis of a very expressive stochastic timed modelling approach without sacrificing the ability of time and memory efficient numerical evaluation.
\subsubsection*{Organisation of the paper}
We introduce Markov automata in Section~\ref{sec:ma}.
Section~\ref{section:expected} considers the evaluation of expected-time properties.
Section~\ref{section:longrun} discusses the analysis of long-run properties, and
Section~\ref{section:timed} focuses on timed reachability properties with time-interval bounds.
Implementation details of our tool, a compositional modelling formalism as well as experimental results are discussed in detail in Section~\ref{sec:tool}.
Section~\ref{sec:conc} concludes the paper.
We provide the proofs for our main results in the appendix.
\section{Preliminaries}
\label{sec:ma}
\subsection{Markov automata}
An MA is a transition system with two types of transitions: probabilistic (as in PAs) and Markovian transitions (as in CTMCs).
Let $\mathit{Act}$ be a countable universe of actions with internal action $\tau \in \mathit{Act}$, and $\distr(S)$ denote the set of discrete probability distribution functions over the countable set $S$.
Let $\alpha$, $\beta$ range over $\mathit{Act}$ and $\mu, \nu$ over $\distr(S)$.
Actions such as $\alpha$ can be used for interaction with other MAs~\cite{EHZ10}.
This does not apply to the internal action $\tau$, which is executed autonomously.
\begin{defi}[Markov automaton]\label{def:mas}
A \emph{Markov automaton (MA)} is a tuple $\mathcal{M} = ( S, A, \it{\, },$ $\mt{\, }, s_0)$ where
$S$ is a nonempty, finite set of states with \emph{initial state} $s_0 \in S$, $A \subseteq \mathit{Act}$ is a finite set of actions with $\tau \in A$, and
\begin{itemize}
\item $\it{\, } \subseteq S \times A \times \distr(S)$ is the \emph{probabilistic} transition relation, and
\item $\mt{\, }\ \subseteq S \times \mathbb{R}_{> 0} \times S$ is the \emph{Markovian} transition relation.
\end{itemize}
\end{defi}
We abbreviate $(s, \alpha, \mu) \in \it{\, }$ by $s \it{\alpha} \mu$ and $(s, \lambda, s') \in \ \mt{\, }$ by $\smash{s \mt{\lambda} s'}$.
An MA can evolve via its probabilistic and Markovian transitions.
If~$s \it{\alpha} \mu$, it can leave state $s$ by executing the action $\alpha$, after which the probability of going to some state $s' \in S$ is given by $\mu(s')$.
If $\smash{s \mt{\lambda} s'}$ is the only transition emanating from $s$, a state transition from $s$ to $s'$ can occur after an exponentially distributed delay with rate $\lambda$.
That is to say, the expected delay from $s$ to $s'$ is $\frac{1}{\lambda}$.
If $\smash{s \mt{\lambda} s'}$ and $s \it{\tau} \mu$ for some $\mu$, however, always the $\tau$-transition is taken and never the Markovian one.
This is the \emph{maximal progress} assumption~\cite{EHZ10}.
The rationale behind this assumption is that internal (i.e., $\tau$-labelled) transitions are not subject to interaction and thus can happen immediately, whereas the probability of a Markovian transition to immediately happen is zero.
Thus, $\smash{s \mt{\lambda} s'}$ almost never fires instantaneously.
Note that the maximal progress assumption does not apply in case $\smash{s \mt{\lambda} s'}$ and $s \it{\alpha} \mu$ with $\alpha \neq \tau$, as $\alpha$-transitions -- unlike $\tau$-transitions -- can be used for synchronisation and thus be subject to a delay.
In this case, the transition $\smash{s \mt{\lambda} s'}$ may happen with positive probability.
The semantics of several Markovian transitions in a state is as follows.
For a state with one or more Markovian transitions, let $\mathbf{R}(s,s') = \smash{\sum \{ \lambda \mid s \mt{\lambda} s' \}}$ be the total rate of moving from state~$s$ to state~$s'$, and let $E(s) = \sum_{s' \in S} \ \mathbf{R}(s,s')$ be the total outgoing rate of $s$.
If $s$ has more than one outgoing Markovian transition, a \emph{competition} between its Markovian transitions exists.
Then, the probability of moving from $s$ to state $s'$ within $d$ time units is
\[
\frac{\mathbf{R}(s,s')}{E(s)} \cdot \left( 1 - e^{-E(s){\cdot}d} \right).
\]
After a delay of at most $d$ time units (second factor) in state $s$, the MA moves to a direct successor state $s'$ with probability $\mathbf{P}(s, s') = \frac{\mathbf{R}(s,s')}{E(s)}$.
Note that also in this case, the maximal progress assumption applies: if $s \it{\tau} \mu$ and $s$ has several Markovian transitions, only the $\tau$-transition can occur and no delay occurs in $s$.
The behaviour of an MA in states with only Markovian transitions is thus the same as in CTMCs~\cite{DBLP:journals/tse/BaierHHK03}.
Fig.~\ref{fig:MA} depicts a sample MA.
Note that this MA only contains $\tau$-labelled probabilistic transitions; by maximal progress, any state has only Markovian transitions or only $\tau$-labelled transitions.
In case several $\tau$-transitions emanate from a state, a nondeterministic choice between these transitions exists.
\begin{figure}[t]
\tikzstyle{every picture}=[thick, scale=0.78, transform shape]
\tikzstyle{every loop}=[->]
\tikzstyle{every scope}=[>=latex]
\centering\begin{tikzpicture}[node distance=2.5cm]
\tikzstyle{every label}=[label distance=0pt]
\node[state, ellipse, minimum size=0pt,fill=gray!40!white] (s000)
{$0,0,0$};
\node[] (sInit) [left of=s000, node distance=1.5cm] {};
\draw[->] (sInit) edge [] node [auto, swap] {} (s000);
\node[state, minimum size=0pt, draw=white] (s100) [above right of=s000]
{};
\node[state, node distance=0.3cm, ellipse, minimum
size=0pt,fill=gray!40!white] (s100n) [below of=s100] {$1,0,0$};
\node[state, minimum size=0pt, draw=white] (s010) [below right of=s000]
{};
\node[state, node distance=0.3cm, ellipse, minimum
size=0pt,fill=gray!40!white] (s010n) [above of=s010] {$0,1,0$};
\node[state, ellipse, minimum size=0pt,fill=gray!40!white] (s001) [below
right of=s100] {$0,0,1$};
\node[state, minimum size=0pt, draw=white] (s101) [above right of=s001]
{};
\node[state, node distance=0.3cm, ellipse, minimum
size=0pt,fill=gray!40!white] (s101n) [below of=s101] {$1,0,1$};
\node[state, minimum size=0pt, draw=white] (s011) [below right of=s001]
{};
\node[state, node distance=0.3cm, ellipse, minimum
size=0pt,fill=gray!40!white] (s011n) [above of=s011] {$0,1,1$};
\node[state, ellipse, minimum size=0pt,fill=gray!40!white] (s111) [below
right of=s101] {$1,1,1$};
\node[state, ellipse, minimum size=0pt,fill=gray!40!white] (s110) [right
of=s111] {$1,1,0$};
\draw[->, double,thin] (s000) -- node [auto] {$\lambda_1$}
(s100n);
\draw[->, double,thin] (s000) -- node [auto, swap]{$\lambda_2$}
(s010n);
\draw[->,thin] (s100n) edge [] node [auto, swap] {$\frac{9}{10}$} (s001);
\draw[->,thin] (s100n) edge [bend right=10] node [auto, swap] {$\frac{1}{10}$}
(s101n);
\draw[thin] (s100n) + (1.25\arclength, -5pt) arc ( -10 : -39.70265826695821
: 1.25\arclength);
\path[thin] (s100n) +(31pt, -11.192068269362267pt) node {\phantom{.}} node
{$\tau$} ;
\draw[->,thin] (s010n) edge [] node [auto] {$\frac{9}{10}$} (s001);
\draw[->,thin] (s010n) edge [bend left=10] node [auto] {$\frac{1}{10}$}
(s011n);
\draw[thin] (s010n) + (1.25\arclength, 5pt) arc ( 10 : 39.70265826695821 :
1.25\arclength);
\path[thin] (s010n) +(31pt, 11.192068269362267pt) node {\phantom{.}} node
{$\tau$} ;
\draw[->, double,thin] (s001) -- node [auto, swap] {$\mu$} (s000);
\draw[->, double,thin] (s001) -- node [auto, swap] {$\lambda_1$}
(s101n);
\draw[->, double,thin] (s001) -- node [auto] {$\lambda_2$}
(s011n);
\draw[->, double,thin] (s101n) -- node [auto, swap] {$\lambda_2$}
(s111);
\draw[->, double,thin] (s101n) [bend right=10] to node [auto,
swap] {$\mu$} (s100n);
\draw[->, double,thin] (s011n) [bend left=10] to node [auto]
{$\mu$} (s010n);
\draw[->, double,thin] (s011n) -- node [auto] {$\lambda_1$}
(s111);
\draw[->, double,thin] (s111) -- node [auto] {$\mu$} (s110);
\draw[->,thin] (s110) edge [in=0, out=120] node [pos=0.65, auto, swap]
{$\frac{9}{10}$} (s101n);
\draw[->,thin] (s110) edge [in=30, out=120] node [pos=0.65, auto, swap]
{$\frac{1}{10}$} (s111);
\draw[thin] (s110) + (-20pt,27pt) arc ( 145 : 172 : 0.75\arclength);
\path[thin] (s110) +(-27pt, 25pt) node {\phantom{.}} node {$\tau$} ;
\draw[->,thin] (s110) edge [in=0, out=-120] node [pos=0.65, auto]
{$\frac{9}{10}$} (s011n);
\draw[->,thin] (s110) edge [in=-30, out=-120] node [pos=0.65, auto]
{$\frac{1}{10}$} (s111);
\draw[thin] (s110) + (-20pt,-27pt) arc ( -145 : -172 : 0.75\arclength);
\path[thin] (s110) +(-27pt, -25pt) node {\phantom{.}} node {$\tau$} ;
\end{tikzpicture}
\caption{A queueing system (taken from~\cite{MAPA}), consisting of a server and two stations.
Each state is represented as a tuple $(s_1,s_2,j)$, with $s_i$ the number of jobs in station $i$, and $j$ the number of jobs
in the server.
The two stations have incoming requests with rates $\lambda_1, \lambda_2$, which are stored until fetched by the server.
If both stations contain a job, the server chooses nondeterministically (in state (1,1,0)).
Jobs are processed with rate $\mu$, and when polling a station, with probability $\frac{1}{10}$ the job is erroneously kept in
the station after being fetched.
For simplicity we assume that each component can hold at most one~job.}
\label{fig:MA}
\end{figure}
\subsection{Actions}
Actions different from $\tau$ can be used to compose MAs from smaller MAs using parallel composition.
For instance, $\mathcal{M}_1 \mathbin{||}_H \mathcal{M}_2$ denotes the parallel composition of MA $\mathcal{M}_1$ and $\mathcal{M}_2$ in which actions in the set $H \subseteq \mathit{Act}$ with $\tau \not\in H$ need to be executed by both MAs simultaneously, and actions not in $H$ are performed autonomously by $\mathcal{M}_i$.
In this paper, we will not cover the details of such composition operation (see~\cite{EHZ10}); it suffices to understand that the distinction between $\tau$ and $\alpha \neq \tau$ is relevant when composing MAs from component MAs.
We assume in the sequel that the MAs to be analysed are \emph{single, monolithic} MAs.
These MAs are not subject to any interaction with other MAs.
Hence, we assume that all transitions are labelled by $\tau$-actions.
(This amounts to the assumption that prior to the analysis all actions needed to compose several MAs are explicitly turned into internal actions by hiding.)
Due to the maximal progress assumption, the outgoing transitions of each state are either all probabilistic or all Markovian.
We can therefore partition the states into a set of probabilistic states, denoted $\mbox{\sl PS} \subseteq S$, and a set of Markovian states, denoted $\mbox{\sl MS} \subseteq S$.
We denote the set of enabled actions in $s$ with $\mathit{Act}(s)$, where $\mathit{Act}(s) = \{ \alpha \in A \mid \exists \mu \in \distr(S) \; . \; s \it{\alpha}\mu\}$ if $s\in \mbox{\sl PS}$, and $\mathit{Act}(s)=\{\bot\}$ otherwise.
\subsection{Paths}
A path in an MA is an infinite sequence $ \pi \ = \ s_0 \it{\sigma_0, \mu_0, t_0} s_1 \it{\sigma_1, \mu_1, t_1} \ldots $ with $s_i \in S$,
$\sigma_i = \tau$ or $\sigma_i = \bot$, $\mu_i \in \distr(S)$ and $t_i \in \mathbb{R}_{\geq 0}$.
For $\sigma_i = \tau$, $s_i \it{\sigma_i, \mu_i, t_i} s_{i+1}$ denotes that after residing $t_i = 0$ time units in $s_i$, the MA moved via action $\sigma_i$ to $s_{i{+}1}$ with probability $\mu_i(s_{i{+}1})$.
In case $\sigma_i = \bot$, $s_i \it{\bot, \mu_i, t_i} s_{i+1}$ denotes that after residing $t_i$ time units in $s$, a Markovian transition led to $s_{i+1}$ with probability $\mu_i(s_{i+1}) = \mathbf{P}(s_i,s_{i+1})$.
For $t \in \mathbb{R}_{\geq 0}$, let $\pi@t$ denote the \emph{sequence} of states that $\pi$ occupies at time $t$.
Due to instantaneous probabilistic transitions, $\pi@t$ is a sequence of states, as an MA may occupy various states at the same time instant.
Let $\mathit{Paths}$ denote the set of infinite paths and $\mathit{Paths}^*$ be the set of finite prefixes thereof (called finite paths).
The time elapsed along the infinite path $\pi$ is given by $\sum_{i=0}^{\infty}t_i$.
Path $\pi$ is Zeno whenever this sum converges.
As the probability of a Zeno path in an MA that only contains Markovian transitions is zero~\cite[Prop.\ 1]{DBLP:journals/tse/BaierHHK03}, an MA is non-Zeno if and only if no SCC with only probabilistic states is reachable (with positive probability).
As such SCC contains no Markovian transitions, it can be traversed infinitely often without any passage of time.
In the rest of this paper, we assume MAs to be non-Zeno.
\subsection{Policies}
Nondeterminism occurs when there is more than one probabilistic transition emanating from a state.
To define a probability space over sets of infinite paths, we adopt the approach as for MDPs~\cite{Put05} and resolve the nondeterminism by a \emph{policy}.
A policy is a function that yields for each finite path ending in state~$s$ a probability distribution over the set of enabled transitions in $s$.
Formally, a policy is a function $D \colon \mathit{Paths}^* \to \distr((\mathit{Act} \mathrel{\cup} \{ \bot \}) \times \distr(S))$. Of course, policies should only choose from available transitions, so we require for each path $\pi$ ending in a state $s_n$ that $D(\pi)(\alpha, \mu) > 0$ implies $s_n \it{\alpha} \mu$
and $D(\pi)(\bot, \mu) > 0$ implies that $s_n$ is Markovian and $\mu = \mathbf{P}(s_n,\cdot)$.
Let $\GMS$ (generic measurable) denote the most general class of such policies that are still measurable; see~\cite{NSK09} for details on measurability.
In general, a policy randomly picks an enabled action and probability distribution in the final state of a given path.
This is also known as a \emph{history-dependent randomised} policy.
If a policy always selects an action and probability distribution according to a Dirac distribution, it is called a \emph{deterministic} policy.
Policies are also classified based on the level of information they use for the resolution of nondeterminism.
In the most general setting, a policy may use all information in a finite path, e.g., the states along the path, their ordering in the path, the amount of time spent in each state, and so forth.
A \emph{stationary} policy only bases its decision on the current state, and not on anything else.
That is, $D$ is stationary whenever $D(\pi_1) = D(\pi_2)$ for any finite paths $\pi_1$ and $\pi_2$ that have the same last state.
A stationary deterministic policy can be viewed as a function $D \colon \mbox{\sl PS} \rightarrow \mathit{Act} \times \distr(S)$ that maps each probabilistic state $s$ to an action $\alpha\in\mathit{Act}$ and probability distribution $\mu\in\distr(S)$ such that $s\it{\alpha}\mu$; such policies always take the same decision every time they are in the same state.
A \emph{time-abstract} policy resolves nondeterminism based on the alternating sequence of states and transitions visited so far, but not on the state residence times.
Let $\TAS$ denote the set of time-abstract policies.
For more details on different classes of policies (and their relationship) on models such as MAs, we refer to~\cite{NSK09}.
Like for MDPs~\cite{Put05}, a stationary or time-abstract policy on an MA induces a countable stochastic process that is equivalent to a (continuous-time) Markov chain.
Using a standard cylinder-set construction on infinite paths in such Markov chains~\cite{DBLP:journals/tse/BaierHHK03} we obtain a $\sigma$-algebra of subsets of $\mathit{Paths}$; given a policy~$D$ and an initial state~$s$, a measurable set of paths is equipped with probability measure $\Pr_{s,D}$.
To ease the development of the theory, and without loss of generality, we assume that each internal action induces a unique probability distribution. Note that this is no restriction: if there are multiple $\tau$-transitions emerging from a state $s\in \mbox{\sl PS}$, we may replace the $\tau$ by internal actions $\tau_1$ to $\tau_{n}$, where $n$ is the out-degree of $s$.
\subsection{Stochastic shortest path (SSP) problems}
As some objectives on MAs can be reduced to SSP problems, we briefly introduce them.
An MDP is a tuple $(S,A,\mathbf{P},s_0)$ where $S$ is a finite set of states, $A \subseteq \mathit{Act}$ is a set of actions, $\mathbf{P} \colon S \times A \times S \to [0,1]$ such that for each state $s$ and each $\alpha$, $\sum_{s' \in S} \mathbf{P}(s,\alpha,s') \in \set{0,1}$, and $s_0 \in S$ is the initial state.
It is assumed that in each state at least one action is enabled, i.e., $\mathbf{P}(s,\alpha,s') > 0$ for each $s$, for some~$\alpha$.
A \emph{non-negative SSP problem} is a tuple $(S,A,\mathbf{P},s_0, G,c,g)$ where the first four elements represent its underlying MDP accompanied by a set $G \subseteq S$ of goal states, cost function $c \colon (S \setminus G) \times A \to \mathbb{R}_{\geq 0}$ and terminal cost function $g\colon G \to \mathbb{R}_{\geq 0}$.
A path through an MDP is an alternating sequence $s_0 \it{\alpha_0} s_1 \it{\alpha_1} \ldots$ such that $\mathbf{P}(s_i, \alpha_i, s_{i{+}1}) > 0$, for all $i$.
The accumulated cost along a path~$\pi$ through the MDP before reaching $G$, denoted by $C_G(\pi)$, is $\sum_{j{=}0}^{k{-}1} c(s_j,\alpha_j) + g(s_k)$ where $k$ is the state index of reaching $G$.
If $\pi$ does not reach $G$, then $C_G(\pi)$ equals $\infty$.
As standard in MDPs~\cite{Put05}, nondeterminism between different actions in a state is resolved using policies; similar to the notion for MAs, a stationary deterministic policy is a function $D\colon\mbox{\sl PS}\to\mathit{Act}$.
Let $\mathit{cR}^{\min}(s, \diamondsuit G)$ denote the minimum expected cost reachability of $G$ in the SSP (under all policies) when starting from $s$.
It is a well-known result that stationary policies suffice to achieve $\mathit{cR}^{\min}(s, \diamondsuit G)$.
This expected cost can be obtained by solving an LP (linear programming) problem~\cite{BerTsi91}.
\section{Expected time objectives}
\label{section:expected}
Let $\mathcal{M}$ be an MA with state space $S$ and $G \subseteq S$ a set of goal states.
Define the (extended) random variable $V_G \colon \mathit{Paths} \rightarrow \mathbb{R}_{\geq 0}^{\infty}$ as the elapsed time before first visiting some state in $G$. That is, for an infinite path $\pi = s_0 \smash{\xrightarrow{\sigma_0,\mu_0,t_0}} s_1 \smash{\xrightarrow{\sigma_1,\mu_1,t_1}} \cdots$, let $V_G(\pi) = \min \left\{ t \in \mathbb{R}_{\geq 0} \mid G \cap \pi@t \not= \emptyset \right\}$ where $\min (\emptyset) = \infty$.
(With slight abuse of notation we use $\pi@t$ as the set of states occurring in the sequence $\pi@t$.)
The minimal expected time to reach $G$ from $s \in S$ is defined by
\begin{align*}
\mathit{eT}^{\min}(s, \diamondsuit G) \ = \
\inf_{D\in \GMS} \mathbb{E}_{s,D}(V_G) \ = \
\inf_{D\in \GMS} \int_{\mathit{Paths}} \hspace{-2ex} V_G(\pi) \cdot \Pr\nolimits_{s,D}(\mathrm{d}\pi)
\end{align*}
where $D$ is a generic measurable policy on $\mathcal{M}$.
(In the sequel, we assume that $\mathit{eT}^{\min}$ is a function indexed by $G$.)
Note that by definition of $V_G$, only the amount of time before entering the first $G$-state is relevant.
Hence, we may turn all $G$-states into absorbing without affecting the expected time reachability.
It is done via replacing all of their emanating transitions by a single Markovian self loop (a Markovian transition to the state itself) with an arbitrary rate.
In the remainder of this section we assume all goal states to be absorbing.
Let $\mu^s_\alpha$ be the distribution such that $s \it{\alpha} \mu^s_\alpha$.
As we assume that all action labels of the transitions emanating a state are unique (by numbering them), this distribution is unique.
\begin{restatable}{thm}{thmExpectedReachability}\label{thm_expected_reachability}
The function $\mathit{eT}^{\min}$ is a fixpoint of the Bellman operator
{\small \begin{align*}
\left[L(v)\right](s) = \begin{cases}
\displaystyle \frac{1}{E(s)} + \sum_{s' \in S} \mathbf{P}(s,s') \cdot v(s') & \text{ if } s \in \mbox{\sl MS} \setminus G \\
\displaystyle \min_{\alpha \in \textit{\footnotesize Act}(s)} \sum_{s' \in S} \mu^s_\alpha(s') \cdot v(s') & \text{ if } s \in \mbox{\sl PS} \setminus G \\
\displaystyle 0 & \text{ if } s \in G,
\end{cases}
\end{align*}}%
where $\mathit{Act}(s)=\{\tau_i \mid s\it{\tau_i}\mu\}$ and $\mu_\alpha^s \in \distr(S)$ is as formerly defined.
\end{restatable}
We will later see that $\mathit{eT}^{\min}$ is in fact the unique fixpoint of the Bellman operator.
Let us explain the above result.
For a goal state, the expected time obviously is zero.
For a Markovian state $s \not\in G$, the minimal expected time to reach some state in $G$ is the expected sojourn time in $s$ (which equals $\frac{1}{E(s)}$) plus the expected time to reach some state in $G$ via one of its successor states.
For a probabilistic state, an action is selected that minimises the expected time according to the distribution $\mu^s_\alpha$ corresponding to $\alpha$ in state $s$.
The characterisation of $\mathit{eT}^{\min}(s, \diamondsuit G)$ in Thm.\,\ref{thm_expected_reachability} allows us to
reduce the problem of computing the minimum expected time reachability in an MA \mbox{to a non-negative
SSP problem~\cite{BerTsi91,deAlf99}.}
This goes as follows.
\begin{defi}[SSP for minimum expected time reachability]\label{def_expected_time_reachability_ssp}
The SSP of MA $\mathcal{M} = \left( S, A, \it{\, }, \mt{\, }, s_0 \right)$ for the expected time reachability of $G \subseteq S$ is
\[
{\sf ssp}_{et}(\mathcal{M}) = \left( S, A \cup \left\{ \bot \right\}, \mathbf{P}, s_0, G,c, g \right)
\]
where $g(s) = 0$ for all $s \in G$ and
{\small \begin{align*}
\mathbf{P}(s, \sigma, s') & = \begin{cases}
\frac{\mathbf{R}(s,s')}{E(s)} & \text{if } s \in \mbox{\sl MS}, \sigma = \bot \\
\mu^s_\sigma(s') & \text{if } s \in \mbox{\sl PS}, s \it{\sigma} \mu^s_\sigma \\
0 & \text{otherwise, and}
\end{cases} &
c(s,\sigma) & = \begin{cases}
\frac{1}{E(s)} & \text{if } s \in \mbox{\sl MS} \setminus G, \sigma = \bot \\
0 & \text{otherwise.}
\end{cases}
\end{align*}}
\end{defi}
Terminal costs are zero.
Transition probabilities are defined in the standard way.
The cost of a Markovian state is its expected sojourn time, whereas that of a probabilistic one is zero.
\begin{restatable}{thm}{thmExpectedTimeReachabilityReduction}\label{thm_expected_time_reachability_reduction}
Given an MA $\mathcal{M}$, $\mathit{eT}^{\min}(s, \diamondsuit G)$ equals $\mathit{cR}^{\min}(s, \diamondsuit G)$ in
$\mbox{\sf ssp}_{et}(\mathcal{M})$.
\end{restatable}
Thus there is a stationary deterministic policy on $\mathcal{M}$ yielding $\mathit{eT}^{\min}(s, \diamondsuit G)$.
Moreover, the uniqueness of the minimum expected cost of an SSP~\cite{BerTsi91,deAlf99} now yields that
$\mathit{eT}^{\min}(s, \diamondsuit G)$ is the unique fixpoint of $L$ (see Thm.~\ref{thm_expected_reachability}). This follows from the fact that the Bellman operator defined in Thm~\ref{thm_expected_reachability} equals the Bellman operator for $\mathit{cR}^{\min}(s, \diamondsuit G)$.
The uniqueness result enables the usage of standard solution techniques such as value iteration and linear programming to compute $\mathit{eT}^{\min}(s, \diamondsuit G)$.
For maximum expected time objectives, a similar fixpoint theorem is obtained, and it can be proven that those objectives correspond to the maximal expected reward in the SSP problem defined above.
Thus far, we have assumed MAs to be non-Zeno, i.e., they do not contain a reachable cycle solely consisting of probabilistic transitions.
However, the above notions can all be extended to deal with such Zeno cycles, by, e.g., setting the minimal expected time of states in Zeno BSCCs that do not contain $G$-states to be infinite (as such states cannot reach $G$).
Similarly, the maximal expected time of states in Zeno end components (that do not contain $G$-states) can be defined as infinity, as in the worst case these states will~never~reach~$G$.
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\mathbf{1}}{\mathbf{1}}
\newcommand{\textit{LRA}}{\textit{LRA}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\paths}{\mathit{Paths}}
\section{Long-run objectives}
\label{section:longrun}
Let $\mathcal{M}$ be an MA with state space $S$ and $G \subseteq S$ a set of goal states.
Let $\mathbf{1}_G$ be the characteristic function of $G$ on finite sequences, i.e., $\mathbf{1}_G(\pi) = 1$ if and only if $s \in G$ for some $s$ in $\pi$.
Following the ideas of \cite{DBLP:conf/lics/Alfaro98,LHK01}, the fraction of time spent in $G$ on an infinite path $\pi$ in $\mathcal{M}$ up to time bound $t \in \mathbb{R}_{\geq 0}$ is given by the random variable
$A_{G,t}(\pi) \ = \ \frac{1}{t} \int_0^t \mathbf{1}_G(\pi@u)\,\mathrm{d}u$.
Taking the limit $t \rightarrow \infty$, we obtain the random variable
\begin{align*}
A_{G}(\pi) \ = \ \lim_{t\to\infty} A_{G,t}(\pi) \ = \ \lim_{t\to\infty} \frac{1}{t}\int_0^t \mathbf{1}_G(\pi@u)\,\mathrm{d}u.
\end{align*}
The expectation of $A_{G}$ for policy $D$ and initial state $s$ yields the corresponding long-run average time spent in $G$:
\begin{align*}
\textit{LRA}^D(s, G)
=
\mathbb{E}_{s,D}(A_{G})
=
\int_{\mathit{Paths}} \hspace{-2ex} A_{G}(\pi) \cdot {\Pr}_{s,D}(\mathrm{d}\pi).
\end{align*}
The minimum long-run average time spent in $G$ starting from state $s$ is then:
\begin{align*}
\textit{LRA}^{\min}(s,G) \ = \ \inf_{D\in \GMS}\ \textit{LRA}^D(s, G) \ = \ \inf_{D\in \GMS} \mathbb{E}_{s,D}(A_{G}).
\end{align*}
Note that $\mathbf{1}_G(\pi@u) = 1$ if and only if $\pi@u$ is a sequence containing at least one state in $G$. For the long-run average analysis, we assume w.l.o.g.\ that $G\subseteq \mbox{\sl MS}$, as the long-run average time spent in any probabilistic state is always 0.
This claim follows directly from the fact that probabilistic states are instantaneous, i.e.\ their sojourn time is $0$ by definition.
Note that in contrast to the expected time analysis, $G$-states cannot be made absorbing in the long-run average analysis.
First we need to introduce \emph{maximal end components}. A sub-MA of MA $\mathcal{M}$ is a pair $(S',K)$ where $S' \subseteq S$ and $K\colon S' \to 2^A$ is a function such that: (i) $K(s) \neq \emptyset$, (ii) $s \in S'$ and $\alpha \in K(s)$ and $s \it{\alpha} \mu$ with $\mu(s') > 0$ implies $s' \in S'$, and (iii) $s \in S'$ and $\smash{s \mt{\lambda} s'}$ implies $s' \in S'$.
A sub-MA $(S',K)$ is contained in a sub-MA $(S'',K')$ if $S' \subseteq S''$ and $K(s) \subseteq K'(s)$ for all $s \in S'$.
An \emph{end component} is a sub-MA whose underlying graph is strongly connected; it is \emph{maximal} w.r.t.\ $K$ if it is not contained in any other end component $(S'',K')$ of $\mathcal{M}$.
In the remainder of this section, we discuss in detail how to compute the minimum long-run average fraction of time spent in $G$ in an MA $\mathcal{M}$ with initial state $s_0$.
The general idea is the following three-step procedure:
\begin{enumerate}
\item Determine the maximal end components $\{ \mathcal{M}_1, \ldots, \mathcal{M}_k \}$ of MA $\mathcal{M}$.
\item Determine $\textit{LRA}^{\min}(G)$ in maximal end component $\mathcal{M}_j$ for all $j \in \{ 1, \ldots, k \}$.
\item Reduce the computation of $\textit{LRA}^{\min}(s_0,G)$ in MA $\mathcal{M}$ to an SSP problem.
\end{enumerate}
The first phase can be performed by a graph-based algorithm~\cite{deA97_thesis,DBLP:conf/soda/ChatterjeeH11}, whereas the last two phases boil down to solving (distinct) LP problems.
\subsection{Unichain MA}
We first show that for unichain MAs
computing $\textit{LRA}^{\min}(s, G)$ can be reduced to determining long-run ratio objectives in MDPs.
The notion of unichain is standard in MDPs~\cite{Put05} and is adopted to MAs in a straightforward manner.
An MA is unichain if for any stationary deterministic policy the induced stochastic process consists of a single ergodic class plus a possibly non-empty set of transient states\footnote{State $s$ is \emph{transient} if and only if the probability of the set of paths that start from $s$ but never return back to it is positive, otherwise it is \emph{recurrent}. An MA is \emph{ergodic} if for all stationary deterministic policies the induced stochastic process consists of a single recurrent class.}.
Let us first explain the long-run ratio objectives. Let $M=(S,A,\mathbf{P},s_0)$ be an MDP.
Assume w.l.o.g.\ that for each $s\in S$ there exists $\alpha\in A$ such that $\mathbf{P}(s,\alpha,s') > 0$ for some $s'\in S$.
Let $c_1, c_2\colon S \times A \to \mathbb{R}_{\geq 0}$ be cost functions.
The operational interpretation is that a cost $c_1(s,\alpha)$ is incurred when selecting action $\alpha$ in state $s$, and similar for $c_2$.
Our interest is the \emph{ratio} between $c_1$ and $c_2$ along a path.
The \emph{long-run ratio} $\mathcal{R}$ between the accumulated costs $c_1$ and $c_2$ along the infinite path $\pi = s_0 \it{\alpha_0} s_1 \it{\alpha_1} \ldots$ in the MDP $M$ is defined by:
$$
\mathcal{R}(\pi) \ = \ \displaystyle \lim_{n \to \infty} \dfrac{\sum_{i=0}^{n-1} c_1(s_i,\alpha_i)}{\sum_{j=0}^{n-1} c_2(s_j,\alpha_j)}.
$$
The minimum long-run ratio objective for state $s$ of MDP $M$ is defined by:
\begin{align*}
R^{\min}(s) \ = \
\inf_D \mathbb{E}_{s,D}(\mathcal{R}) \ = \
\inf_D \sum_{\pi \in \paths} \mathcal{R}(\pi) \cdot \text{Pr}_{s,D}(\pi).
\end{align*}
Here, $\paths$ is the set of paths in the MDP, $D$ is a stationary deterministic MDP-policy, and $\Pr$ is the probability measure on MDP-paths.
From~\cite[Th.\ 6.14]{deA97_thesis}, it follows that $R^{\min}(s)$ can be obtained by solving the following LP problem with real variables $k$ and non-negative $x_s$ for each $s \in S$: Maximise $k$ subject to:
$$
x_s \, \leq \, c_1(s,\alpha) - k \cdot c_2(s,\alpha) + \sum_{s' \in S} \mathbf{P}(s,\alpha,s') \cdot x_{s'} \quad
\mbox{ for each } s \in S, \alpha \in A.
$$
We now transform an MA into an MDP with two cost functions as follows.
\begin{defi}[From MA to 2-cost MDPs]\label{def:MAtomdp}
Let $\mathcal{M} = \left( S, A, \it{\, }, \mt{\, }, s_0 \right)$ be an MA and $G \subseteq S$ a set of goal states.
The MDP $\mbox{\sf mdp}(\mathcal{M}) = (S, A\cup\{\bot\}, \mathbf{P}, s_0)$, where $\mathbf{P}$ is defined as in Def.~\ref{def_expected_time_reachability_ssp},
is extended with cost functions $c_1$ and $c_2$ defined by:
{\small \begin{align*}
c_1(s,\sigma) & = \begin{cases}
\frac{1}{E(s)} & \text{if } s \in \mbox{\sl MS} \cap G \wedge \sigma = \bot \\
0 & \text{otherwise,}
\end{cases}
&
c_2(s,\sigma) & = \begin{cases}
\frac{1}{E(s)} & \text{if } s \in \mbox{\sl MS} \wedge \sigma = \bot \\
0 & \text{otherwise.}
\end{cases}
\end{align*}}
\end{defi}
Observe that cost function $c_2$ keeps track of the average residence
time in state~$s$ whereas $c_1$ only does so for states in $G$.
Furthermore, $\mathcal{R}$ is well-defined in this setting, since the
cost functions $c_1$ and $c_2$ are obtained from non-Zeno MA. In
other words, the probability of the set of paths with
ill-defined long-run ratio is zero.
\begin{restatable}{thm}{unichainTheorem}\label{thm_unichain_theorem}
For unichain MA $\mathcal{M}$, $LRA^{\min}(s,G)$ equals $R^{\min}(s)$ in $\mbox{\sf mdp}(\mathcal{M})$.
\end{restatable}
To summarise, computing the minimum long-run average fraction of time that is spent in some goal state in $G \subseteq S$ in a unichain MA $\mathcal{M}$ equals the minimum long-run ratio objective in an MDP with two cost functions.
The latter can be obtained by solving an LP problem.
Observe that for any two states $s$, $s'$ in a unichain MA, $\textit{LRA}^{\min}(s,G)$ and $\textit{LRA}^{\min}(s',G)$ coincide.
We therefore omit the state and simply write $\textit{LRA}^{\min}(G)$ when considering unichain MAs.
\subsection{Arbitrary MA}
Let $\mathcal{M}$ be an MA with initial state $s_0$ and maximal end components $\{ \mathcal{M}_1, \ldots,$ $\mathcal{M}_k \}$ for $k > 0$ where MA $\mathcal{M}_j$ has state space $S_j$.
\newtheorem{lemma}[thm]{Lemma}
\begin{restatable}{lemma}{lemMEC}\label{lem:lemMEC}
Let $\mathcal{M}$ be a maximal end component and $D$ a stationary deterministic policy inducing a multichain on $\mathcal{M}$. Then there exists a stationary deterministic policy $D'$ inducing a unichain on $\mathcal{M}$ such that the long-run ratio is at least as good as for $D$.
\end{restatable}
Therefore, we can say that each $\mathcal{M}_j$ induces a unichain MA for the optimal long-run ratio.
Using this decomposition of $\mathcal{M}$ into maximal end components, we obtain the following result:
\newif\ifnote\notetrue
\begin{restatable}{thm}{lemLRASSP}\ifnote\label{lem:LRA_SSP}\footnote{This theorem corrects a small flaw in the corresponding theorem for IMCs in~\cite{DBLP:conf/nfm/GuckHKN12}.}\fi
For MA $\mathcal{M} = (S, A, \it{\,}, \mt{\, }, s_0)$ with MECs $\{ \mathcal{M}_1, \ldots, \mathcal{M}_k \}$
with state spaces $S_1, \dots, S_k \subseteq S$, and set of goal states $G \subseteq S$:
\begin{align*}
\textit{LRA}^{\min}(s_0,G) & = \inf_{D\in GM}\sum_{j=1}^{k} \textit{LRA}^{\min}_j(G) \cdot {\Pr}_{s_0,D}(\diamondsuit \Box S_j),
\end{align*}
where ${\Pr}_{s_0,D}(\diamondsuit \Box S_j)$ is the probability to eventually reach and continuously stay in some states in $S_j$ from $s_0$ under policy $D$ and $\textit{LRA}^{\min}_j(G)$ is the LRA of $G \cap S_j$ in unichain~MA~$\mathcal{M}_j$.
\end{restatable}
\ifnote\notefalse
Computing the minimal LRA for arbitrary MAs is now reducible to a non-negative SSP problem.
This proceeds as follows.
In MA $\mathcal{M}$, we replace each maximal end component $\mathcal{M}_j$ by two fresh states $q_j$ and $u_j$.
Intuitively, $q_j$ represents $\mathcal{M}_j$ whereas $u_j$ can be seen as the gate to and from $\mathcal{M}_j$.
Thus, state $u_j$ has a Dirac transition to $q_j$ as well as all probabilistic transitions leaving $S_j$.
Let $U$ denote the set of $u_j$ states and $Q$ the set of $q_j$ states.
\iffalse
\begin{defi}[SSP for long-run average]\label{defSSP}
Let $\mathcal{M}$ be an MA with maximal end components $\mathcal{M}_1$ through $\mathcal{M}_k$ with $S_i$ the state space of $\mathcal{M}_i$.
The SSP of MA $\mathcal{M}$ for the LRA in $G \subseteq S$ is
$$
\mbox{\sf ssp}_{lra}(\mathcal{M}) =
\left( \left(S \setminus \smash{\bigcup_{i=1}^k} S_i \right) \cup U \cup Q, A \cup \{ \bot \}, \mathbf{P}', s_0, Q, c, g \right),
$$
where $g(q_i) = \textit{LRA}^{\min}_i(G)$ for $q_i \in Q$ and $c(s,\sigma) = 0$ for all $s$ and $\sigma\in A\cup\{\bot\}$.
$\mathbf{P}'$ is defined as follows.
$\mathbf{P}'(s,s')$ equals $\mathbf{P}(s,s')$ for all $s,s' \in S'$ where $S' = S \setminus \smash{\bigcup_{i=1}^k} S_i$.
For $u_j \in U$:
{\small
\begin{align*}
\mathbf{P}'(u_j, \tau, s') & = \mathbf{P}(S_j, \tau, s') \quad \text{\!if } s' \in S' \setminus S_j
&
\mbox{\!\!and\!\! } \quad \mathbf{P}'(u_i, \tau, u_j) & = \mathbf{P}(S_i, \tau, S_j) \quad \text{\!for } i \neq j.
\end{align*}}%
Finally, we have: $\mathbf{P}'(q_j,\bot,q_j) = 1 = \mathbf{P}'(u_j, \bot,q_j)$ and $\mathbf{P}'(s, \sigma, u_j) = \mathbf{P}(s, \sigma, S_j)$.
\end{defi}
Here, $\mathbf{P}(s,\alpha, S')$ is a shorthand for $\sum_{s' \in S'} \mathbf{P}(s,\alpha,s')$; similarly, we use $\mathbf{P}(S',\alpha,s') = \sum_{s \in S'} \mathbf{P}(s, \alpha, s')$.
The terminal costs of the new $q_i$-states are set to $\textit{LRA}^{\min}_i(G)$.
\fi
For simplicity of the definition we assume w.l.o.g. that each probabilistic state induces a $\tau$-transition with an index of the state. Further, the $\tau$-transitions of each state $s_k\in \mbox{\sl PS}$ are numbered from $1$ to $n_{s_k}\in\mathbb{N}$, where $n_{s_k}$ is the number of probability distributions induced by $\tau_{s_k}$. Thus, we denote an action in state $s_k$ with $\tau_{{s_k}_l}$ with $l\in \{1\ldots n_{s_k}\}$.
\begin{defi}[SSP for long-run average]\label{defSSP}
The SSP of MA $\mathcal{M}$ for the LRA in $G \subseteq S$ is $\mbox{\sf ssp}_{lra}(\mathcal{M}) = \left( (S \setminus \smash{\bigcup_{i=1}^k} S_i) \cup U \cup Q, A \cup \{ \bot \}, \mathbf{P}', s_0, Q, c, g \right)$, where $g(q_i) = \textit{LRA}^{\min}_i(G)$ for $q_i \in Q$ and $c(s,\sigma) = 0$ for all $s$ and $\sigma\in A \cup\{\bot\}$. $\mathbf{P}'$ is defined as follows. Let $S' = S \setminus \smash{\bigcup_{i=1}^k} S_i$. $\mathbf{P}'(s,\sigma,s')$ equals $\mathbf{P}(s,\sigma,s')$ for all $s,s' \in S'$ and $\sigma\in A\cup\{\bot\}$.
For the new states $u_j$:
{\small
\begin{align*}
\mathbf{P}'(u_j, \tau_{{s_k}_l}, s') & = \mathbf{P}(s_k, \tau_{{s_k}_l}, s') \quad \text{\!if } s' \in S' \wedge s_k \in S_j\wedge l\in\{1\dotso n_{s_k}\}
&
\mbox{\!\!and\!\! }\\
\quad \mathbf{P}'(u_i, \tau_{{s_k}_l}, u_j) & = \mathbf{P}(s_k, \tau_{{s_k}_l}, S_j) \quad \text{\!if } s_k \in S_i\wedge l\in\{1\dotso n_{s_k}\}\wedge \tau_{{s_k}_i}\not\in A_i
\end{align*}}%
Finally, we have: $\mathbf{P}'(q_j,\bot,q_j) = 1 = \mathbf{P}'(u_j, \bot,q_j)$ and $\mathbf{P}'(s, \sigma, u_j) = \mathbf{P}(s, \sigma, S_j)$.
\end{defi}
Here, $\mathbf{P}(s,\alpha, S')$ is a shorthand for $\sum_{s' \in S'} \mathbf{P}(s,\alpha,s')$ and $A_i$ denotes the action set of maximal end component $\mathcal{M}_i$.
The terminal costs of the new $q_i$-states are set to $\textit{LRA}^{\min}_i(G)$.
\begin{figure}[t!]
\centering
\subfigure[\scriptsize{A sample Markov automaton.}]{\label{fig:example_mec}
\begin{tikzpicture}[scale=0.9, every node/.style={transform shape}]
\node[state, initial] (s0) {$s_0$};
\node[state] (s3) [right of=s0,node distance=2cm] {$s_1$};
\node[state] (s5) [below of=s3,node distance=2cm] {$s_3$};
\node[state, accepting] (s6) [right of=s3,node distance=2cm] {$s_2$};
\node[state] (s8) [left of=s5,node distance=2cm] {$s_5$};
\node[state] (s9) [right of=s5,node distance=2cm] {$s_4$};
\node[rectangle] (tmp) [left of=s3, node distance=.38cm] {};
\node[rectangle] (tmp2) [left of=s5, node distance=.38cm] {};
\node[rectangle,draw,green,fit=(tmp)(tmp2)(s6)(s9)] (mec1) {};
\node[rectangle,draw,red,fit=(s8)] (mec2) {};
\path[->] (s0) edge[double,thin] node[sloped,above] {$2$} (s3);
\path[->] (s3) edge[thin] node[left] {$0.6$} (s5);
\path[->] (s3) edge[thin] node[sloped,above] {$0.4$} (s6);
\path[->]
(s3) edge[thin] node[inner sep=0mm,pos=0.2] (a1) {} (s5)
(s3) edge[thin] node[inner sep=0mm,pos=0.2] (b1) {} (s6);
\path[-,shorten <=-.4pt,shorten >=-.4pt] (a1) edge [thin,bend right] (b1) node[right,yshift=.2cm,xshift=.4cm] {$\alpha$} ;
\path[->] (s5) edge[thin] node[sloped,above] {$\alpha,1$} (s8);
\path[->] (s5) edge[thin] node[sloped, above] {$\beta,1$} (s9);
\path[->] (s8) edge[thin,double,loop left] node[left] {$1$} (s8);
\path[->] (s9) edge[thin,double] node[right] {$3$} (s6);
\path[->] (s6) edge[thin,double, bend right=60] node[sloped,above] {$1$} (s3);
\end{tikzpicture}
}
\hspace{.3cm}
\subfigure[\scriptsize{Induced SSP for MA in Figure~\ref{fig:example_mec}.}]{\label{fig:example_ssp}
\begin{tikzpicture}[scale=0.9, every node/.style={transform shape}]
\node[state, initial] (s0) {$s_0$};
\node[state,fill=green] (u2) [right of=s0,node distance=2cm] {$u_1$};
\node[state,fill=green] (q2) [right of=u2,node distance=2cm] {$q_1$};
\node[state,fill=red] (u1) [below of=u2,node distance=2cm] {$u_2$};
\node[state,fill=red] (q1) [right of=u1,node distance=2cm] {$q_2$};
\path[->] (s0) edge[thin] node[sloped,above] {$\bot,1$} (u2);
\path[->] (u2) edge[thin] node[sloped,above] {$\bot,1$} (q2);
\path[->] (u2) edge[thin] node[left] {$\alpha,1$} (u1);
\path[->] (u1) edge[thin] node[sloped,above] {$\bot,1$} (q1);
\path[->] (q2) edge[thin,loop right] node[right] {$\bot,1$} (q2);
\path[->] (q1) edge[thin,loop right] node[right] {$\bot,1$} (q1);
\end{tikzpicture}
}
\caption {Example for Definition~\ref{defSSP}.}
\end{figure}
\begin{exa}
Consider the MA $\mathcal{M}$ from Figure~\ref{fig:example_mec}, having MECs $\mathcal{M}_1$ with $S_1=\{s_1,s_2,s_3,s_4\}$ and $\mathcal{M}_2$ with $S_2=\{s_5\}$. For the simplification of the action notation, we use $\alpha$ and $\beta$ instead of $\tau$. Let $G = \{ s_2 \}$. By Definition~\ref{defSSP}, $\mbox{\sf ssp}_{lra}(\mathcal{M})$ is defined as follows. As $k{=}2$, $U=\{u_1,u_2\}$ and $Q=\{q_1,q_2\}$. Hence, $S_{\mbox{\sf ssp}} = \{s_0,u_1,u_2,q_1,q_2\}$. First consider $s,s'\in S'$. Since, $S'=\{s_0\}$ and there exists no transition from $s_0$ to $s_0$ we can omit the first rule. Now consider all outgoing transitions from MECs. For $\mathcal{M}_1$ there exists a transition from $s_3 \it{\alpha,1} s_5$ in the underlying MA, where $s_3\in S_1$ and $s_5 \in S_2$. It follows that $\mathbf{P}'(u_1,\alpha,u_2)=\mathbf{P}(s_3,\alpha,S_2)=1$.
Now consider all states in $U$ and $Q$ and add new transitions with $\mathbf{P}(u_i,\bot,q_i) = \mathbf{P}(q_i,\bot,q_i)=1$ for $i=1,2$. Finally, consider all states $s \in S_{\mbox{\sf ssp}} \cap S$ with a transition into a MEC. Hence, $\mathbf{P}'(s_0,\bot,u_1)=\mathbf{P}(s_0,\bot,s_1)=1$. The MDP of $\mbox{\sf ssp}_{lra}(\mathcal{M})$ is depicted in Figure~\ref{fig:example_ssp}.
\label{ex:append}
\end{exa}
\begin{restatable}{thm}{thmLRASSP}\label{thm:LRA_SSP}
For MA $\mathcal{M}$, $\textit{LRA}^{\min}(s_0,G)$ equals $cR^{\min}(s_0 ,\diamondsuit Q)$ in SSP $\mbox{\sf ssp}_{lra}(\mathcal{M})$.
\end{restatable}
To summarise, computing the minimum long-run average fraction of time that is spent in some goal states in $G \subseteq S$ in an arbitrary MA $\mathcal{M}$ starting in state $s_0$ equals the minimum expected cost of an SSP.
\section{Timed reachability objectives}
\label{section:timed}
This section presents an algorithm that approximates time-bounded reachability probabilities
in MAs. We start with a fixpoint characterisation, and then explain how
these probabilities can be approximated using a discretisation technique.
\subsection{Fixpoint characterisation}
Our goal is to come up with a fixpoint characterisation for the maximum (or minimum) probability to reach a set of goal states in a time interval. Let $\mathcal{I}$ and $\mathcal{Q}$ be the set of all nonempty nonnegative real intervals with real and rational bounds, respectively.
For interval $I \in \mathcal{I}$ and $t \in \mathbb{R}_{\ge 0}$, let $I \ominus t = \left \{ x - t \mid x \in I \wedge x \ge t \right \}$.
Given MA $\mathcal{M}$, $I \in \mathcal{I}$ and a set $G \subseteq S$ of goal states, the set of all paths that reach some goal states within interval $I$ is denoted by $\diamondsuit^{I} \, G$.
Let $p^{\mathcal{M}}_{\max}(s,\diamondsuit^I \, G)$ be the maximum probability of reaching $G$ within interval $I$ if starting in state $s$ at time $0$.
Here, the maximum is taken over all possible generic measurable policies.
The next lemma provides a characterisation of $p_{\max}^{\mathcal{M}}(s,\diamondsuit^I \, G)$ as a fixpoint.
\begin{lem}
\label{fpc:ma}
Let $\mathcal{M}$ be an MA, $G \subseteq S$ and $I \in \mathcal{I}$ with $\inf I=a$ and $\sup I=b$. Then, $p^{\mathcal{M}}_{\max}(s,\diamondsuit^I \, G)$ is the least fixpoint of the higher-order operator
$\Omega\colon (S \times \mathcal{I} \rightarrow [0,1])
\rightarrow (S \times \mathcal{I} \rightarrow [0,1])$,
which for $s \in \mbox{\sl MS}$ is given by:
\begin{align*}
\Omega(F)(s,I)&=
\begin{cases}
\displaystyle\int_{0}^{b}E(s)\mathrm{e}^{-E(s)t}\sum_{s' \in S}\mathbf{P}(s,s')F(s',I \ominus t)\:\mathrm{d} t & \!s \notin G \\
\displaystyle\mathrm{e}^{-E(s)a} + \int_{0}^{a}E(s)\mathrm{e}^{-E(s)t}\sum_{s' \in
S}\mathbf{P}(s,s')F(s',I \ominus t)\:\mathrm{d} t & \!s \in G
\end{cases}
\intertext{and for $s \in \mbox{\sl PS}$ is defined by:}
\Omega(F)(s,I)&=
\begin{cases}
1 & s \in G \wedge 0\in I \\
\max_{\alpha\in\mathit{Act}(s)} \sum_{s' \in S}
\mu^s_\alpha(s')F(s',I) & \mathrm{otherwise.}
\end{cases}
\end{align*}
\end{lem}
The proof of Lemma~\ref{fpc:ma} is a slight adaptation
of the proof of~\cite[Thm.~4]{Fu13}, where it has been also shown
that $p^{\mathcal{M}}_{\max}(s,\diamondsuit^I \, G)$ is Lipschitz
continuous and thus measurable. The characterisation is a simple
generalisation of that for IMCs~\cite{DBLP:conf/tacas/ZhangN10},
reflecting the fact that taking an action from a probabilistic state
leads to a distribution over the states (rather than a single
state). The above characterisation yields a Volterra integral
equation system which is in general not directly tractable
\cite{DBLP:journals/tse/BaierHHK03}. To tackle this problem, we
approximate the fixpoint characterisation using discretisation,
extending ideas developed in \cite{DBLP:conf/tacas/ZhangN10}.
\subsection{Discretisation}
We
split the time interval into equally-sized discretisation steps,
each of length $\delta$. The discretisation step is assumed to be
small enough such that with high probability it carries at most one
Markovian transition. This allows us to construct a discretised MA
(dMA), a variant of a semi-MDP, obtained by summarising the
behaviour of the MA at equidistant time points. Paths in a dMA can
be seen as time-abstract paths in the corresponding MA, implicitly
counting discretisation steps, and thus discrete time.
\begin{defi}
Given MA \defma and discretisation step
$\delta\in\mathbb{R}_{>0}$, $\mathcal{M}_{\delta}=( S, A, \it{\,
},$ $\mt{\, }_{\delta}, s_0)$ is the dMA induced from $\mathcal{M}$
with respect to $\delta$, with $\mt{}_{\delta} \, = \{ \, (s,
\mu^{s})\mid s \in \mbox{\sl MS} \, \}$, where
\begin{equation*}
\mu^s (s') =
\begin{cases}
(1-\mathrm{e}^{-E(s)\delta})\mathbf{P}(s,s') & \mbox{if } s' \neq s\\
(1-\mathrm{e}^{-E(s)\delta})\mathbf{P}(s,s') + \mathrm{e}^{-E(s)\delta} & \mbox{otherwise.}
\end{cases}
\end{equation*}
\end{defi}
Using the above fixpoint characterisation, it is now possible to
relate reachability probabilities in the MA $\mathcal{M}$ to
reachability probabilities in its dMA $\mathcal{M}_{\delta}$.
\begin{restatable}{thm}{thmTBR}\label{thm:tbr}
Given MA $\mathcal{M}=(S,A, \it{}, \mt{}, s_0)$, $G \subseteq S$, interval
$I=[0,b] \in \mathcal{Q}$ with $b \ge 0$ and $\lambda = \max_{s \in \scriptsize\mbox{\sl MS}}E(s)$.
Let $\delta\in\mathbb{R}_{>0}$ be such that $b=k_b\delta$ for some $k_b \in \mathbb{N}$.
Then, for all $s \in S$ it holds that
\begin{equation*}
p^{\mathcal{M}_{\delta}}_{\max}(s, \diamondsuit^{[0,k_b]} \, G)
\ \leq \
p^{\mathcal{M}}_{\max}(s, \diamondsuit^{[0,b]} \, G)
\ \leq \
p^{\mathcal{M}_{\delta}}_{\max}(s, \diamondsuit^{[0,k_b]} \, G) + 1 - e^{- \lambda b}\big(1+ \lambda \delta\big)^{k_b}.
\end{equation*}
\end{restatable}
This theorem can be extended to intervals with non-zero lower
bounds; for the sake of brevity, the details are omitted here.
The remaining problem is to compute
$p^{\dMAM}_{\max}(s,\diamondsuit^{[0,k_b]}\,G)$, which is the
maximum probability to reach some goal state in dMA \dMAM within the
step bound $k_b$ from initial state $s$. Let
$\diamondsuit^{[0,k_b]}\,G$ be the set of infinite (time-abstract)
paths of \dMAM that reach some state in $G$ within $k_b$ steps; the
objective is then formalised by
$p^{\dMAM}_{\max}(s,\diamondsuit^{[0,k_b]}\,G)=\sup_{D \in \TAS}
{\Pr}_{s,D}(\diamondsuit^{[0,k_b]}\,G)$ where we recall that $\TAS$ denotes the
set of time-abstract policies. Our algorithm
is now an adaptation (to dMA) of the well-known value iteration
\mbox{scheme for MDPs.}
The algorithm proceeds by backward unfolding of the dMA in an
iterative manner, starting from the goal states. Each iteration
intertwines the analysis of Markovian states and of probabilistic states.
The key idea is that a path from probabilistic states to $G$ is split
into two parts: reaching Markovian states from probabilistic states in zero
time and reaching goal states from Markovian states in interval $[0,j]$,
where $j$ is the step count of the iteration. The former computation
can be reduced to an unbounded reachability problem in the MDP induced
by probabilistic states with rewards on Markovian states. For the latter,
the algorithm operates on the previously computed reachability
probabilities from all Markovian states up to step count $j$. We can
generalise this recipe from step-bounded reachability to step
interval-bounded reachability; details are described
in~\cite{HatefiH12}.
\iffalse
\begin{algorithm}[t]
\caption{Computing maximum step bounded
reachability}\label{dtma:reach}
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\SetKwFunction{iPhase}{$i^*$-phase}\SetKwFunction{mPhase}{$m$-phase}
\Input{dMA $\mathcal{D}$, goal states $G \subseteq S$, step bound $k \in \mathbb{N}$}
\Output{The vector $\left( p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,k]} \, G) \right)_{s \in S}$}
\BlankLine
\Begin {
make all $s \in G$ of $\mathcal{D}$ absorbing \;
\lForEach {$s \in G$} {$p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,0]} \, G) := 1$} \;
\lForEach {$s \in S \setminus G$}{$p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,0]} \, G) := 0$}\;
\For {$j := 0$ \KwTo $k-1$} {
// $i^*$-phase for step $j$ \;
\lForEach {$s \in \mbox{\sl PS}$} {
$p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,j]} \, G) := \displaystyle\sup_{D \in \TAS} \sum_{s' \in \mbox{\scriptsize MS}} {\Pr}_{D,s}(\diamondsuit^{[0,0]} \, \{ s' \})p^{\mathcal{D}}_{\max}(s',\diamondsuit^{[0,j]} \, G)$ \;
}
// $m$-phase for step $j+1$ \;
\lForEach {$s \in \mbox{\sl MS}$} { $p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,j{+}1]} \, G) = \sum_{s' \in S} \DiBrPr(s, \bot, s')p^{\mathcal{D}}_{\max}(s',\diamondsuit^{[0,j]} \, G)$ \;}
} // $i^*$-phase for step $k$ \; \lForEach {$s \in \mbox{\sl PS}$} {
$p^{\mathcal{D}}_{\max}(s,\diamondsuit^{[0,k]} \, G) :=
\displaystyle\sup_{D \in \TAS} \sum_{s' \in \mbox{\scriptsize MS}}
{\Pr}_{D,s}(\diamondsuit^{[0,0]} \,
\{s'\})p^{\mathcal{D}}_{\max}(s',\diamondsuit^{[0,k]} \, G)$ \; } }
\end{algorithm}
\subsubsection*{Step-bounded reachability}
We consider the case $[0,k], \; k \in \mathbb{N}_0$ and apply a
modified value iteration algorithm to compute reachability
probabilities. The algorithm (cf.\ Algorithm~\ref{dtma:reach}) proceeds by backward unfolding of the
dMA in an iterative manner, starting from the goal states. Initially, all states in $G$ are made absorbing, i.e., all
transitions are removed, and replaced by a discretised Markovian self-loop
(a transition to a Dirac distribution over the source state).
The initial value of probability vector
($p^{\mathcal{D}}_{\max}(\cdot,[0,0])$) is set to $1$ for goal states
and to $0$ otherwise.
In the main loop of the algorithm, each iteration consists of an
\textit{$m$-phase} and an \textit{$i^*$-phase}. In an
\textit{$m$-phase}, reachability probabilities from Markovian states are
updated, while an $i^*$-phase updates reachability probabilities from
probabilistic states. The algorithm then proceeds by
intertwining $m$-phases and $i^*$-phases consecutively for $k$ steps.
After the $j$-th execution of the \textit{$i^*$-phase} (and thus
after $j{-}1$ executions of the $m$-phase), the elements of
$p^{\mathcal{D}}_{\max}(\cdot,[0,j-1])$ are up to date. Therefore, a
final \textit{$i^*$-phase} is needed.
In an \textit{$m$-phase}, we update the reachability probabilities
starting from Markovian states by taking the only discretised Markovian
transitions. This step is exactly the same as a standard value
iteration step for a DTMC. Thus in step $j$, the $m$-phase updates
the reachability probability for the goal states using the
reachability probabilities calculated in step $j{-}1$.
In an \textit{$i^*$-phase}, we maximise the reachability
probabilities starting from probabilistic states to Markovian states by
taking arbitrarily many probabilistic transitions. The number of
transitions does not matter, because they take zero time. In the
algorithm, the $j$-th iteration operates on the previously computes
reachability probabilities from all Markovian states up to step $j$. The
$i^*$-phase then computes the probability to reach Markovian states
within interval $[0,0]$ which implies that only probabilistic
transitions can be taken.
The key issue is that the path from probabilistic
states to $G$ is split into two parts: reaching Markov
states from probabilistic states in zero time and reaching goal states
from Markovian states in interval $[0,j]$. Due to the memoryless
property, the result comes from the product of the probability measure of
these two sets of paths which needs to be maximised over all possible
schedulers, i.e. $p^{\mathcal{D}}_{\max}(s,[0,j]) = \sup_{D \in \TAS}
\sum_{s' \in {\scriptsize\mbox{\sl MS}}} {\Pr}_{D,s}(\diamondsuit^{[0,0]} \, \{ s'
\})p^{\mathcal{D}}_{\max}(s',[0,j])$. Because the time is not
important in the computation of
${\Pr}_{D,s}(\diamondsuit^{[0,0]} \, \{ s' \})$, we can see
the problem as unbounded reachability with rewards on Markovian states
($p^{\mathcal{D}}_{\max}(s',[0,j])$). It can be shown that stationary
schedulers are sufficient for computing minimum and maximum
objectives~\cite[Chapter 10]{BaierK08} and the objectives can again be
computed by using a value iteration algorithm.
We can generalize this recipe to compute step-bounded reachability
to step interval-bounded reachability, see~\cite{HatefiH12}.
\fi
\section{Tool chain and case studies}
\label{sec:tool}
This section describes the implementation of the algorithms discussed, together with the modelling features resulting in our \toolname~tool chain.
Also, we present two case studies that provide empirical evidence of the strengths and weaknesses \mbox{of the \toolname~tool chain.}
\subsection{Modelling}\label{sec:modelling}
As argued in the introduction, MAs can be used as a semantical model for various modelling formalisms.
We use the process-algebraic specification language MAPA (Markov Automata Process Algebra)~\cite{MAPA,markPhd}.
This language contains the usual process algebra operators, can treat data as first-class citizens, and supports several reduction techniques for MA specifications.
In fact, it turns out to be beneficial to map a language (like GSPNs) to MAPA so as to profit from these reductions.
The MAPA language supports algebraic processes featuring data, nondeterministic choice, action prefix with probabilistic choice, rate prefix, conditional behaviour and process instantiation (allowing recursion). Using MAPA processes as basic building blocks, the language also supports the modular construction of large systems via top-level parallelism, encapsulation, hiding and renaming.
The operational semantics of a MAPA specification yields an MA; for a detailed exposition of the syntax and semantics we refer to~\cite{MAPA,markPhd}.
To enable state space reduction and generation, our tool chain uses a linearised normal form of MAPA referred to as MLPE (Markovian Linear Probabilistic process Equation).
In this format, there is precisely one process which consists of a nondeterministic choice between a set of symbolic transitions, making MLPEs easy to translate to MAs.
Every MAPA specification can be translated efficiently into an MLPE while preserving strong bisimulation~\cite{MAPA}.
\subsubsection*{Reduction techniques}
On MLPEs, several reduction techniques have been defined. Some of them simplify the MLPE to improve readability and speed up state space generation, while others really modify it in such a way that the underlying MA gets smaller. Being defined on the specification, these reductions eliminate the need to ever generate the original unreduced state space. We briefly discuss six such techniques.
\begin{itemize}
\item
\emph{Maximal progress reduction} removes Markovian transitions from states also having $\tau$-transitions (motivated by the maximal progress assumption).
\item
\emph{Constant elimination}~\cite{KPST11} replaces parameters that remain forever constant by their initial (and hence permanent) value.
\item
\emph{Expression simplification}~\cite{KPST11} evaluates functions for which all parameters are constants and applies basic laws from logic.
\item
\emph{Summation elimination}~\cite{KPST11} removes trivial nondeterministic choices often arising from synchronisations.
\item
\emph{Dead-variable reduction}~\cite{PT09} detects parts of the specification in which the value of some variable is irrelevant: it will be overwritten before being used for all possible futures. When reaching such a part, the variable is reset to its initial value.
\item
\emph{Confluence reduction}~\cite{ConfluenceMA} detects spurious nondeterminism resulting from parallel composition. It denotes a subset of the probabilistic transitions of a MAPA specification as confluent, meaning that they can safely be given priority if enabled together with other transitions.
\end{itemize}
\subsection{\toolname~tool chain}
\begin{figure}[b]
\centering\begin{tikzpicture}[scale=0.78, transform shape]
\node[state, rectangle, minimum width=50pt, minimum height=20pt] (s_0) {SCOOP};
\node[state, rectangle, minimum width=50pt, minimum height=20pt] (s_2) [right of=s_0, node distance=4.5cm] {IMCA};
\node[state, rectangle, draw=white, minimum width=50pt, minimum height=20pt] (s_3) [right of=s_2, node distance=2.75cm] {Results};
\node[state, rectangle, draw=white] (s_4) [above of=s_0, node distance=1.2cm] {MAPA spec + Property};
\draw[draw=white] (s_0) -- node [auto,swap] {Goal states} (s_2);
\draw[->] (s_0) -- node [auto] {MA} (s_2);
\draw[->, in=295, out=245, loop] (s_0) edge node [auto,swap] {Reduce} (s_0);
\draw[->] (s_2) -- node [auto] {} (s_3);
\draw[->] (s_4) -- (s_0);
\node[state, dashed, rectangle, minimum width=50pt, minimum height=20pt] (s_5) [left of=s_0, node distance=4.5cm] {GEMMA};
\draw[draw=white] (s_5) -- node [auto,swap] {Property} (s_0);
\draw[->, dashed] (s_5) -- node [auto] {MAPA spec} (s_0);
\node[state, rectangle, draw=white] (s_6) [above of=s_5, node distance=1.2cm] {GSPN + Property};
\draw[->, dashed] (s_6) -- (s_5);
\end{tikzpicture}
\caption{Analysing Markov automata using the \toolname~tool chain.}
\label{fig:approach}
\end{figure}
Our tool chain consists of several tool components: SCOOP~\cite{Timmer11,MAPA}, IMCA~\cite{DBLP:conf/nfm/GuckHKN12}, and GEMMA~\cite{gemma}, see Figure~\ref{fig:approach}.
The tool chain comprises about 8,000 LOC (without comments).
SCOOP (written in Haskell) supports the generation of MAs from MAPA specifications by a translation into the MLPE format.
It implements all the reduction techniques described above.
The capabilities of the IMCA tool component (written in {\tt C++}) have been lifted to expected time and long-run objectives for MAs, and extended with timed reachability objectives.
It also supports (untimed) reachability objectives which are not treated further here.
A prototypical translator from GSPNs to MAs, in fact MAPA specifications, has been realised (the GEMMA component, written in Haskell).
We connected the three components into a single tool chain, by making SCOOP export the (reduced) state space of an MLPE in the IMCA input language.
Additionally, SCOOP has been extended to translate properties, based on the actions and parameters of a MAPA specification, to a set of goal states in the underlying MA.
That way, in one easy process, systems and their properties can be modelled in MAPA, translated to an optimised MLPE by SCOOP, exported to the IMCA tool and then analysed.
\subsection{Case studies}
This section presents experiments with \toolname. All experiments were conducted on a 2.5 GHz Intel Core i5 processor with 4GB RAM, running Mac OS X~10.8.3.
\subsubsection*{Processor grid}
First, we consider a model of a $2 \times 2$ concurrent processor architecture.
Using GEMMA \cite{gemma}, we automatically derived the MA model from the GSPN model in~\cite[Fig.~11.7]{Mar95}.
Previous analysis of this model required weights for all immediate transitions, which necessitates having complete knowledge of the mutual behaviour of all these transitions.
We allow a weight assignment to just a (possibly empty) subset of the immediate transitions---reflecting the practical scenario of only knowing the mutual behaviour for a selection of the transitions.
For this case study we indeed kept weights for only a few of the transitions, obtaining probabilistic behaviour for them and nondeterministic behaviour for the others.
Table~\ref{tab:grid_tb} reports on the time-bounded and time-interval bounded probabilities for reaching a state such that the first processor has an empty task queue. We vary the degree of multitasking $K$, the error bound $\epsilon$ and the interval~$I$. For each setting, we report the number of states $|S|$ and goal states $|G|$, and the
generation time with SCOOP (both with and without the reductions from Section~\ref{sec:modelling}).
The runtime demands grow with both the upper and lower time
bound, as well as with the required accuracy. The model size also
affects the per-iteration cost and thus the overall complexity of
reachability computation.
Note that the reductions speed-up the analysis times by a factor between $1.8$ and $2.5$: even more than the reduction in state space size. This is due to the fact that these techniques significantly reduce the degree of nondeterminism.
Table~\ref{tab:grid} displays the results for expected time until an empty task queue, as well as the long-run average that a processor is active.
In contrast to~\cite{Mar95}, which fixes all nondeterminism and obtains, for instance, an LRA of $0.903$ for $K=2$, we are now able to retain nondeterminism and provide the more informative interval $[0.8810, 0.9953]$.
Again, SCOOP's reduction techniques significantly improve runtimes.
\begin{table}[t]
\centering
{\vskip15pt
\hspace{-0.5cm}\scalebox{.75}{\begin{tabular}{c||ccc|ccc|cc|rrr|rrr}
& \multicolumn{3}{c|}{unreduced} & \multicolumn{3}{c|}{reduced}\\
$K$ & $|S|$ & $|G|$ & time & $|S|$ & $|G|$ & time & $\epsilon$ & $I$ & \smash{\begin{turn}{35}$p^{\min}(s_0, \diamondsuit^{I} G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $p^{\max}(s_0, \diamondsuit^{I} G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex}\\
\hline
\hline
\multirow{4}{*}{2} & \multirow{4}{*}{2{,}508} & \multirow{4}{*}{1{,}398} & \multirow{4}{*}{0.6} &
\multirow{4}{*}{1{,}789} & \multirow{4}{*}{1{,}122} & \multirow{4}{*}{0.8} &
$10^{-2}$ & $[0,3]$ & 0.91\phantom{0} & 58.5 & 31.0 & 0.95\phantom{0} & 54.9 & 21.7\\
& & & & & & &
$10^{-2}$ & $[0,4]$ & 0.96\phantom{0} & 103.0 & 54.7 & 0.98\phantom{0} & 97.3 & 38.8\\
& & & & & & &
$10^{-2}$ & $[1,4]$ & 0.91\phantom{0} & 117.3 & 64.4 & 0.96\phantom{0} & 109.9 & 49.0\\
& & & & & & &
$10^{-3}$ & $[0,3]$ & 0.910 & 580.1 & 309.4 & 0.950 & 544.3 & 218.4\\
\hline
\multirow{4}{*}{3} & \multirow{4}{*}{10{,}852} & \multirow{4}{*}{4{,}504} & \multirow{4}{*}{3.1} &
\multirow{4}{*}{7{,}201} & \multirow{4}{*}{3{,}613} & \multirow{4}{*}{3.5} &
$10^{-2}$ & $[0,3]$ & 0.18\phantom{6} & 361.5 & 202.8 & 0.23\phantom{1} & 382.8 & 161.1\\
& & & & & & &
$10^{-2}$ & $[0,4]$ & 0.23\phantom{6} & 643.1 & 360.0 & 0.30\phantom{1} & 681.4 & 286.0\\
& & & & & & &
$10^{-2}$ & $[1,4]$ & 0.18\phantom{6} & 666.6 & 377.3 & 0.25\phantom{1} & 696.4 & 317.7\\
& & & & & & &
$10^{-3}$ & $[0,3]$ & 0.176 & 3{,}619.5 & 2{,}032.1 & 0.231 & 3{,}837.3 & 1{,}611.9\\
\hline
4 & 31{,}832 & 10{,}424 & 9.8 & 20{,}021 & 8{,}357 & 10.5 &
$10^{-2}$ & $[0,3]$ & 0.01\phantom{6} & 1{,}156.8 & 614.9 & 0.03\phantom{1} & 1{,}196.5 & 486.4\\
\hline
\end{tabular}}
}
\caption{Interval reachability probabilities for the grid. (Time in seconds.)}
\label{tab:grid_tb}
\end{table}
\begin{table}[t]
\centering
{
\vskip10mm\hspace{-0.3cm}\scalebox{.8}{\begin{tabular}{c||rrr|rrr|rrr|rrr}
$K$ & \smash{\begin{turn}{35}$eT^{\min}(s_0, \Diamond G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $eT^{\max}(s_0, \Diamond G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex} &
\smash{\begin{turn}{35}$\textit{LRA}^{\min}(s_0, G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $\textit{LRA}^{\max}(s_0, G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex}
\\
\hline
\hline
2 & 1.0000 & 0.3 & 0.1 & 1.2330 & 0.7 & 0.3 & 0.8110 & 1.3 & 0.7 & 0.9953 & 0.5 & 0.2\\
3 & 11.1168 & 18.3 & 7.7 & 15.2768 & 135.4 & 40.6 & 0.8173 & 36.1 & 16.1 & 0.9998 & 4.7 & 2.6\\
4 & 102.1921 & 527.1 & 209.9 & 287.8616 & 6{,}695.2 & 1{,}869.7 & 0.8181 & 505.1 & 222.3 & 1.0000 & 57.0 & 34.5\\
\hline
\end{tabular}}
}
\caption{Expected times and long-run averages for the grid. (Time in seconds.)}
\label{tab:grid}
\end{table}
\begin{table}[t]
\centering
{
\vskip15pt\hspace{-0.4cm}\scalebox{.75}{\begin{tabular}{cc||ccc|ccc|cc|rrr|rrr}
&& \multicolumn{3}{c|}{unreduced} & \multicolumn{3}{c|}{reduced}\\
$Q$ & $N$ & $|S|$ & $|G|$ & time & $|S|$ & $|G|$ & time & $\epsilon$ & $I$ & \smash{\begin{turn}{35}$p^{\min}(s_0, \diamondsuit^{I} G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $p^{\max}(s_0, \diamondsuit^{I} G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex}\\
\hline
\hline
\multirow{2}{*}{2} & \multirow{2}{*}{3} & \multirow{2}{*}{1{,}497} & \multirow{2}{*}{567} & \multirow{2}{*}{0.4} & \multirow{2}{*}{990} & \multirow{2}{*}{324} & \multirow{2}{*}{0.2} &
$10^{-3}$ & $[0,1]$ & 0.277 & 4.7 & 2.9 & 0.558 & 4.6 & 2.5\\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.486 & 22.1 & 14.9 & 0.917 & 22.7 & 12.5\\
\hline
\multirow{2}{*}{2} & \multirow{2}{*}{4} & \multirow{2}{*}{4{,}811} & \multirow{2}{*}{2{,}304} & \multirow{2}{*}{1.0} & \multirow{2}{*}{3{,}047} & \multirow{2}{*}{1{,}280} & \multirow{2}{*}{0.6} &
$10^{-3}$ & $[0,1]$ & 0.201 & 25.1 & 14.4 & 0.558 & 24.0 & 13.5 \\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.344 & 106.1 & 65.8 & 0.917 & 102.5 & 60.5\\
\hline
\hline
\multirow{2}{*}{3} & \multirow{2}{*}{3} & \multirow{2}{*}{14{,}322} & \multirow{2}{*}{5{,}103} & \multirow{2}{*}{3.0} & \multirow{2}{*}{9{,}522} & \multirow{2}{*}{2{,}916} & \multirow{2}{*}{1.7} &
$10^{-3}$ & $[0,1]$ & 0.090 & 66.2 & 40.4 & 0.291 & 60.0 & 38.5\\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.249 & 248.1 & 180.9 & 0.811 & 241.9 & 158.8\\
\hline
\multirow{2}{*}{3} & \multirow{2}{*}{4} & \multirow{2}{*}{79{,}307} & \multirow{2}{*}{36{,}864} & \multirow{2}{*}{51.6} & \multirow{2}{*}{50{,}407} & \multirow{2}{*}{20{,}480} & \multirow{2}{*}{19.1} &
$10^{-3}$ & $[0,1]$ & 0.054 & 541.6 & 303.6 & 0.291 & 578.2 & 311.0\\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.141 & 2{,}289.3 & 1{,}305.0 & 0.811 & 2{,}201.5 & 1{,}225.9\\
\hline
\hline
\multirow{2}{*}{4} & \multirow{2}{*}{2} & \multirow{2}{*}{6{,}667} & \multirow{2}{*}{1{,}280} & \multirow{2}{*}{1.1} & \multirow{2}{*}{4{,}745} & \multirow{2}{*}{768} & \multirow{2}{*}{0.8} &
$10^{-3}$ & $[0,1]$ & 0.049 & 19.6 & 14.0 & 0.118 & 19.7 & 12.8\\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.240 & 83.2 & 58.7 & 0.651 & 80.9 & 53.1\\
\hline
\multirow{2}{*}{4} & \multirow{2}{*}{3} & \multirow{2}{*}{131{,}529} & \multirow{2}{*}{45{,}927} & \multirow{2}{*}{85.2} & \multirow{2}{*}{87{,}606} & \multirow{2}{*}{26{,}244} & \multirow{2}{*}{30.8} &
$10^{-3}$ & $[0,1]$ & 0.025 & 835.3 & 479.0 & 0.118 & 800.7 & 466.1\\
& & & & & & & &
$10^{-3}$ & $[1,2]$ & 0.114 & 3{,}535.5 & 2{,}062.3 & 0.651 & 3{,}358.9 & 2{,}099.5\\
\hline
\end{tabular}
}
}
\caption{Interval reachability probabilities for the polling system. (Time in seconds.)}
\label{tab:poll_job_tb}
\end{table}
\begin{table}[t]
\centering
{\vskip20pt
\hspace{-0.3cm}\scalebox{.75}{\begin{tabular}{cc||rrr|rrr|rrr|rrr}
$Q$ & $N$ & \smash{\begin{turn}{35}$eT^{\min}(s_0, \Diamond G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $eT^{\max}(s_0, \Diamond G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex} &
\smash{\begin{turn}{35}$\textit{LRA}^{\min}(s_0, G)$ \end{turn}}\hspace*{-10ex}& \smash{\begin{turn}{35}time(unred)\end{turn}}\hspace*{-8ex} & \smash{\begin{turn}{35}time(red)\end{turn}}\hspace*{-6ex} &\smash{\begin{turn}{35} $\textit{LRA}^{\max}(s_0, G)$ \end{turn}}\hspace*{-10.5ex}& \smash{\begin{turn}{35}time(unred)\end{turn}} \hspace*{-9ex} & \smash{\begin{turn}{35}time(red)\end{turn}} \hspace*{-7ex}
\\
\hline
\hline
2 & 3 & 1.0478 & 0.2 & 0.1 & 2.2489 & 0.3 & 0.2 & 0.1230 & 0.8 & 0.5 & 0.6596 & 0.2 & 0.1\\
2 & 4 & 1.0478 & 0.2 & 0.1 & 3.2053 & 2.0 & 1.0 & 0.0635 & 9.0 & 5.2 & 0.6596 & 1.3 & 0.6\\
\hline
3 & 3 & 1.4425 & 1.0 & 0.6 & 4.6685 & 8.4 & 5.0 & 0.0689 & 177.9 & 123.6 & 0.6600 & 26.2 & 13.0 \\
3 & 4 & 1.4425 & 9.7 & 4.6 & 8.0294 & 117.4 & 67.2 & 0.0277 & 7{,}696.7 & 5{,}959.5 & 0.6600 & 1{,}537.2 & 862.4\\
\hline
4 & 2 & 1.8226 & 0.4 & 0.3 & 4.6032 & 2.4 & 1.6 & 0.1312 & 45.6 & 32.5 & 0.6601 & 5.6 & 3.9\\
4 & 3 & 1.8226 & 29.8 & 14.2 & 9.0300 & 232.8 & 130.8 & \multicolumn{3}{c|}{-- timeout (18 hours) --} & 0.6601 & 5{,}339.8 & 3{,}099.0\\
\hline
\end{tabular}}
}
\caption{Expected times and long-run averages for the polling system. (Time in seconds.)}
\label{tab:poll_job}
\end{table}
\subsubsection*{Polling system}
Second, we consider a polling system with two stations and one server,
similar to the one depicted in Figure~\ref{fig:MA} and inspired
by~\cite{Polling}. There are incoming requests of $N$ possible types,
each of them with a (possibly different) service rate. Additionally,
the stations each store these in a local queue of size $Q$. We vary the values
of~$Q$ and~$N$, analysing a total of six different settings.
Since---as for the previous case---analysis scales proportionally with
the error bound, we \mbox{keep this constant here}.
Table~\ref{tab:poll_job_tb} reports results for
time-bounded and time-interval bounded properties, and
Table~\ref{tab:poll_job} displays probabilities and runtime
results for expected times and long-run averages. For all analyses,
the goal set consists of all states for which both station queues are full.
\section{Conclusion}
\label{sec:conc}
This paper presented new algorithms for the quantitative analysis of Markov automata (MAs) and proved their correctness.
Three objectives have been considered: expected time, long-run average, and timed reachability.
The \toolname\ tool chain supports the modelling and reduction of MAs, and can analyse these three objectives.
It is also equipped with a prototypical tool to map GSPNs onto MAs.
The \toolname\ tool is accessible via its easy-to-use web interface that can be found at \mbox{\url{http://fmt.cs.utwente.nl/~timmer/mama}}.
Experimental results on a processor grid and a polling system give insight into the accuracy and scalability of the presented algorithms.
Future work will focus on efficiency improvements and reward extensions~\cite{atvapaper}.
\section*{Acknowledgements}
This work is funded by the EU FP7-projects SENSATION and MEALS, the STW project ArRangeer (grant 12238), the NWO project SYRUP (grant 612.063.817), and the DFG Sonderforschungsbereich AVACS.
\bibliographystyle{abbrv}
|
\section{Introduction}\label{sec:introduction}
The pressurized vessel of nuclear power plant reactors is generally an irreplaceable component, the integrity of which determines the lifetime of the installation. Prolonged exposure to irradiation causes hardening and embrittlement of the steel used to fabricate the vessel, which thereby might lose its capability to maintain integrity in case of an accident. It is therefore important to understand the origin of this hardening and embrittlement. It is known that changes in mechanical properties induced by irradiation can be understood in terms of changes in the micro- or nanostructure of the material.
In this context, iron-carbon (Fe-C) alloys are archetypal model materials for low-alloy ferritic steels, such as those used for reactor pressure vessels of existing light water nuclear reactors. In particular, the generally-agreed mechanistic framework within which RPV steel hardening and embrittlement are understood involves the contribution of three damage components: (i) matrix damage, (ii) precipitates, and (iii) grain boundary de-cohesion due to segregation \cite{odette1998recent}. It is generally assumed that the first component, dominated by point-defects created in the matrix (iron), can be studied by reference to the behaviour of Fe-C \cite{buswell1995irradiation}.
In this framework, this paper is a continuation of the work to develop a computer model for the nanostructure evolution in iron-carbon (Fe-C) alloys under irradiation, using the Object Kinetic Monte Carlo (OKMC) simulation technique. In our model, we study how the populations of SIA and vacancy clusters evolve over time and increasing damage accumulation in terms of displacements per atom (dpa). The clusters are described in the model by a set of parameters that define their stability, their diffusion properties, and their possibility of interacting with each other and with other defects and the pre-existing microstructure.
In \cite{jansson2013simulation,jansson2014okmc} we presented a set of mechanisms and parameters valid for irradiated Fe-C systems at low temperature, below 370 K. At this temperature, transmission electron microscope examinations reveal the formation under neutron irradiation of only prismatic dislocation loops with $1/2\langle111\rangle$ Burgers vector \cite{malerba2011review,zinkle2006microstructure}. The same low temperature Fe-C system has also recently been studied using cluster dynamics modelling \cite{abe2013effect}. However, at higher temperature a vast majority of loops with $\langle100\rangle$ loops is observed \cite{malerba2011review,hernandez2010transmission}. The stability and the diffusion properties of $\langle100\rangle$ loops are significantly different from those of $1/2\langle111\rangle$ loops \cite{dudarev2008effect,arakawa2007observation}. The model must therefore be adapted to take this into account, in order to be applied to higher temperature conditions. We extend here the model to
higher temperatures by taking as reference the irradiation experiment from the REVE campaign \cite{malerba2011review,hernandez2010transmission,lambrecht2008influence,lambrecht2009phd,lambrecht2009positron,bergner2010comparative}, where the irradiation temperature was 563--568 K. The ``pure'' Fe material in that experiment is estimated to contain less than 20 wt. ppm C. The average grain size is 250 $\mu$m and the dislocation density is reported to be $(7\pm2)\cdot10^{13}$ m$^{-2}$ The irradiation was performed in the in-pile section 2 of the CALLISTO loop of the BR2 reactor in SCK$\bullet$CEN in Mol, Belgium. The material was irradiated for 16 days until 0.19 dpa had accumulated. The fluence was $\sim$13.1$\cdot10^{19}$ n cm$^{-2}$, $E>1$ MeV \cite{malerba2011review}. This corresponds to a dpa rate of $1.37\cdot10^{-7}$ dpa/s.
After irradiation, the material was studied using positron annihilation spectroscopy (PAS) \cite{lambrecht2008influence,lambrecht2009phd,lambrecht2009positron,meslin2010characterization}, small-angle neutron scattering (SANS) \cite{bergner2010comparative} and transmission electron microscopy (TEM) \cite{hernandez2010transmission}. These different experimental techniques gives a thorough description of the nanostructure of the material in terms of defect densities and average vacancy and SIA cluster sizes.
\section{Computation method}\label{sec:methods}
We use the LAKIMOCA code for our OKMC simulations \cite{domain2004simulation}. The approach we use is explained in detail in \cite{jansson2013simulation} and \cite{jansson2014okmc}.
For convenience, we highlight here the fundamental ideas on the method.
OKMC is a stochastic simulation technique that considers the migration and interactions of objects in a pre-defined system with parameterized probabilities. The objects represent in our case vacancy and SIA clusters. Their shapes are usually spherical, except for large SIA clusters, that are represented by toroids. Reactions, such as clustering or annihilation, take place when the objects overlap geometrically. The probability for the objects to perform a migration jump are given in terms of Arrhenius frequencies for thermally activated events,
\begin{equation}
\Gamma_i = \nu_i\exp \left(\frac{-A_i}{k_B T}\right),
\end{equation}
where $\nu_i$ is the attempt frequency of event $i$, $A_i$ the activation energy, $k_B$ Boltzmann's constant and $T$ the temperature. Events are randomly chosen according to their probability, according to the Monte Carlo algorithm \cite{metropolis1953equation}. The simulated time is increased according to the residence time algorithm \cite{young1966monte} with
\begin{equation}\label{eq:residence}
\Delta t = \frac{1}{\sum_{i=1}^{N_{int}} \Gamma_i + \sum_{j=1}^{N_{ext}}P_j},
\end{equation}
where $N_{int}$ is the number of internal events such as defect jumps and $N_{ext}$ the number of external events, such as cascades or Frenckel pair creation, with $P_j$ being the probabilities for the external events. In the long term, Eq. \eqref{eq:residence} is indeed equivalent to
\begin{equation}
\Delta t' = -\ln{R}\Delta t,
\end{equation}
where $R$ is a uniform random number between 1 and 0 \cite{bortz1975new}.
Traps and sinks are immobile spherical objects that can either trap clusters with a certain binding energy, $E_t^\delta$, that may depend on the size of the trapped cluster, or remove them from the system. In LAKIMOCA, different traps have to be specified for different kinds of defects, such as vacancies and SIA clusters. We use traps to simulate the effect of carbon-vacancy clusters, that are able to trap SIA or vacancy clusters. Sinks are used to simulate the effect of dislocations. SIA clusters are observed by TEM to decorate dislocations \cite{hernandez2010transmission}, meaning they are not absorbed if large enough. We therefore only allow SIA clusters smaller than the core of the dislocations, \textit{i.e.} size 1--4, to be absorbed by the sinks. The number density and radius of the spherical sinks are defined to equal the sink strength of the dislocation density in the material (See \cite{nichols1978estimation}). The sink radius, $R_s^\delta$, is thus obtained as
\begin{equation}\label{eq:sink_radius}
R_s^\delta = \frac{\rho V }{4\pi N_s}Z^\delta-r_1^\delta
\end{equation}
where $\delta=v$ denote parameters for vacancies and $\delta=i$ parameters for SIA. $V$ is the volume of the simulation box, $N_s$ is the number of sinks. The defect radius $r^\delta$ has to be removed from the sink radius as the original sink strength expressions were derived for point-like defects. The radius $r_1^v = 4.32\cdot10^{-10}$ m is the capture radius of a single vacancy and $r_1^i = 5.17\cdot10^{-10}$ m the capture radius for a single SIA. The bias factor takes into account the strain field of defects. Since the strain field of SIA, compared to vacancies, is larger. $Z^v= 1.0$ and for SIA sinks, we tried different values.
Neutron or ion irradiation is simulated by introducing populations of vacancy and SIA clusters into the system with a certain rate per time and volume, corresponding to a certain dpa rate.
The dpa is calculated according to the NRT standard \cite{domain2004kinetic,norgett1975proposed}
\begin{equation}
dpa = \frac{0.8 E_{MD}}{2E_D},
\end{equation}
where $E_{MD}$ is the damage energy, the fraction of the kinetic energy of the primary knock-on atom (PKA) spectrum that is not absorbed by electronic excitation, and is well approximated by the energy of the cascades in the MD simulations. The displacement threshold energy is $E_D = 40$ eV. The cascade cluster populations are chosen randomly from a database with displacement cascade simulations \cite{stoller1996point,stoller1997primary,stoller2000statistical,
stoller2000evaluation,stoller2004secondary,stoller2000role}. The MD simulations were performed using the Finnis-Sinclair potential \cite{finnis1984simple} and using energies of
5 keV, 10 keV, 20 keV, 30 keV, 40 keV, 50 keV and 100 keV.
As anticipated in the introduction, in order to extend to higher temperatures the model that we presented in \cite{jansson2013simulation} for irradiation temperatures $<$370 K, it is necessary to allow for the presence of two types of loops. At low irradiation temperatures all SIA clusters could be assumed to be of $1/2\langle111\rangle$ type, because $\langle100\rangle$ loops are generally not observed after neutron irradiation in that temperature range \cite{zinkle2006microstructure,eyre1962direct,eyre1965electron,robertson1982low}, with only one exception for ultra-high pure Fe \cite{robertson1982low}, even though $\langle100\rangle$ loops can be produced under electron irradiation at temperatures as low as 140 K \cite{arakawa2006changes}.
At higher temperature (550--600 K), however, the fact that $\langle100\rangle$ loops are always observed obliges the existence of this kind of defect to be somehow considered in the model \cite{malerba2011review}. The origin of this dominance of $\langle100\rangle$ SIA clusters above a certain temperature, that depends among other things on the type or irradiation (neutrons or ions), even for a given temperature, is still debated and it is not unambiguously known whether these loops are created in this orientation already in cascades (this is never seen in molecular dynamics simulations of displacement cascades) or they are the result of a subsequent transformation undergone by the clusters produced in the cascade. Here we pragmatically made the implicit assumption that the SIA clusters below the threshold for visibility can be of both kinds, whereas all visible clusters are, by default, of $\langle100\rangle$ type. This means that the migration energy, $M^i$, for SIA and the trapping energy, $E_t^i$, as
functions of cluster size, will be different from the parameters used in \cite{jansson2013simulation}. Namely, below the visibility threshold, $N_{th}$, the energy will take an effective value, $M^i=0.2$ eV (see Sec. \ref{sec:small_sia}), that corresponds to a hypothetical weighted average between highly mobile $1/2\langle111\rangle$ loops and slower $\langle100\rangle$ loops, and similarly the trapping energy will be an effective average value; above the threshold, on the other hand, the value will be the one of $\langle100\rangle$ loops, $M^i= 0.9$ eV \cite{osetsky2013private}. SIA clusters of sizes between 1 and 5 are considered too small to have Burgers vectors and their migration energy are the same as in \cite{jansson2013simulation}. For vacancy clusters, only the trapping energy $E_t^v$, has been changed, as will be explained later. All other parameters are unchanged and fully described in \cite{jansson2013simulation}.
Another effect of temperature that must be taken into account is that the carbon-vacancy complexes stable at low and high temperature are not the same (this aspect was already discussed in \cite{jansson2013simulation} and applied for the simulation of the post-irradiation annealing).
Simulations of small C-V complexes in \cite{jansson2013simulation} revealed indeed that the two dominating C-V complexes at temperatures above 450 K are C and C$_2$V. Molecular dynamics (MD) studies shows that C binds with the edge of a $\langle100\rangle$-SIA loop with an energy of 1.1 eV, but is repulsed inside the glide prism of the loop. C$_2$V complexes were calculated to bind to a 61-$\langle100\rangle$ SIA with an energy of 0.6 eV \cite{anento2013unpublished}. As in \cite{jansson2013simulation}, we simulate the effect of these complexes using generic spherical traps for SIA and vacancies. The estimated amount of C in the the reference material used in the experiment was $\sim$65 appm. If we also include the amount of N atoms, that are assumed to have very similar properties to C atoms, we get a total amount of 134 appm. A priori, other interstitial impurities with similar effects might be present as well, \textit{e.g.} oxygen or hydrogen. Here, and in what follows, they will be formally all confused
with C, which is the main one. We then used two kinds of traps for SIA clusters, corresponding to C atoms and C$_2$V complexes. Since the latter complex contains twice as many C atoms as the former, the total density of traps will be less than the
amount of carbon in the system. We used 66 appm of C traps ($E_{t1}^v$) and 34 appm of C$_2$V traps ($E_{t2}^v$); together they correspond to 100 appm and therefore to 134 appm C. The 2:1 ratio of the densities of the two SIA traps were taken from the results of the annealing simulation with carbon explicit in \cite{jansson2013simulation}. We used consistently 100 appm of traps for vacancies ($E_t^v$). The values are listed in Table \ref{table:M_i}. In this table the only differences with respect to the parameters in \cite{jansson2013simulation} concerns clusters above size 6: up to the visibility threshold, migration and trapping energy are effective values that result from an iteration; above the threshold, the trapping energy for $\langle100\rangle$ loops is taken from \cite{anento2013unpublished}, while the migration energy was chosen based on indications from work \cite{osetsky2013private} performed using a recently developed on-the-fly KMC scheme \cite{xu2011simulating}. The vacancy trapping energy was
also the result of iterations, as is shown in the section on results. The capture radius was 5.0 Å for all traps, as was used in \cite{jansson2013simulation,jansson2014okmc}.
We used a simulation box size of $450\times520\times600a_0^3$, where $a_0 = 2.87\cdot10^{-10}$ m is the lattice parameter for $\alpha$-Fe. Periodic boundary conditions were applied in all three directions. Simulation temperature, grain size, dislocation density, and dpa rate were dictated by the reference experiments in the REVE campaign \cite{lambrecht2008influence,lambrecht2009phd,lambrecht2009positron,bergner2010comparative,hernandez2010transmission,meslin2010characterization} (See introduction).
\section{Results}\label{sec:results}
\subsection{The cluster density and mean size evolution compared to the reference experiment}
In this section we present the results obtained with the parameter choice that gives globally the best results as compared to experiments. For this we used $N_{th} = 90$, the visibility size threshold for SIA clusters; $E_t^v = 0.2$ eV, the trapping energy of vacancy clusters; and $Z^i = 3.0$, the bias between vacancy and SIA sink radii. The migration energy used is $M^i = 0.2$ eV for SIA clusters of size 6--$N_{th}$ and 0.9 eV above $N_{th}$. These values will be discussed in more detail in the sections to follow.
The evolution of the vacancy clusters can be compared with results from two different experimental techniques, PAS and SANS, in terms of density and mean cluster size. However, the statistical analysis has to take into account that the two techniques are sensitive to different parts of the vacancy cluster population. PAS is especially suitable for small vacancy clusters, being sensitive even to the presence of single vacancies. However, it provides correct information about the size only for vacancy clusters up to a diameter of $\sim$1 nm ($N^v\sim 50$), above which all clusters sizes are indistinguishable. SANS, on the other hand, can not detect clusters smaller than $\sim$1 nm in diameter.
In Fig. \ref{R0110-Z30-Ev04-60_298967.pdf} the number density of vacancy clusters of all sizes is shown as a function of dose. The density at 0.2 dpa is $10^{23}$ m$^{-3}$, which coincides with the experimental PAS value \cite{lambrecht2009phd,meslin2010characterization}. The vacancy cluster mean size versus dpa is shown in Fig. \ref{R20130417-35_464423_vac_mean_size_evolution.pdf} and the size distribution at 0.2 dpa is shown in Fig. \ref{monica_V_size_distribution.pdf}. The over-all mean size is $\sim$2 nm. If one considers that PAS does not distinguish a 50 vacancy cluster from bigger ones, the mean size is $\sim$0.7 nm, which is only slightly larger than the experimental PAS value of 0.6 nm
\cite{lambrecht2009phd,meslin2010characterization}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130423-n90_475680_vac_density_evolution.pdf}
\caption{Density of vacancy clusters of sizes observable by PAS and SANS, respectively, versus dpa. The data are compared with PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign.}
\label{R0110-Z30-Ev04-60_298967.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130423-n90_475680_vac_mean_size_evolution.pdf}
\caption{Vacancy cluster mean size versus dpa with PAS and SANS resolution. The experimental PAS data (triangles) are from \cite{lambrecht2009phd,meslin2010characterization}. The SANS data (bullets) shows the two peaks of the size distribution at 0.2 dpa from \cite{bergner2010comparative}. The lower SANS data point is the major peak.}
\label{R20130417-35_464423_vac_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130423_n90_V_size_distribution.pdf}
\caption{Vacancy cluster size distribution at 0.2 dpa. The bin size is 0.4 nm.}
\label{monica_V_size_distribution.pdf}
\end{figure}
If we now consider only the cluster population detectable by SANS, the mean size is $\sim$3 nm and the density $2.20\cdot10^{22}$ m$^{-2}$. However, in the actual experiment, SANS sees in fact a bimodal size distribution, with one peak at 1.9 nm (cluster diameter) and another one at 8 nm \cite{bergner2010comparative}. TEM studies have observed voids with a mean size of $12\pm0.4$ nm and an estimated density of $1.2\cdot10^{20}$ m$^{-3}$. The TEM peak is likely to be the same as the second peak seen by SANS. The density corresponding to this peak with large 8--12 nm clusters is, however, too low to be seen in our simulations. The mean size obtained for clusters larger than 1 nm by the simulations, 3.4 nm, thus corresponds well to the SANS and TEM data, being slightly higher than the main peak value, 1.9 nm, but still between the two peaks. The density of the clusters in the first peak of the SANS distribution is estimated to be $4\cdot10^{22}$ m$^{-3}$ \cite{bergner2010comparative}, slightly above our
simulated values.
The best cases for the visible SIA cluster evolution in terms of number density and cluster size are shown in Fig. \ref{R20130417-35_464423_visible_SIA.pdf} and Fig. \ref{R20130417-35_464423__best_case_SIA_mean_size_evolution.pdf}, respectively. The density shows good agreement with the experimental data from \cite{hernandez2010transmission}, whereas the mean cluster sizes are a bit underestimated by the model.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130423-n90_475680_visible_SIA.pdf}
\caption{Visible SIA cluster density evolution versus dpa. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line correspond to one cluster in the simulated volume.}
\label{R20130417-35_464423_visible_SIA.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130423-n90_475680_SIA_mean_size_evolution.pdf}
\caption{Visible SIA cluster mean cluster size evolution versus dpa. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{R20130417-35_464423__best_case_SIA_mean_size_evolution.pdf}
\end{figure}
\subsection{Effect of the choice of the visibility threshold}
For the SIA migration energy, $M^i$, for different sizes, as presented in Table \ref{table:M_i}, there are two values that may change and need to be established: the "effective" migration energy for the "mixture" of $\langle111\rangle$ and $\langle100\rangle$ clusters below the threshold for visibility and, to a certain extent, the latter threshold as well. Moreover, the value of 0.9 eV for $\langle100\rangle$ loops is based only on preliminary calculations. We therefore did trials with different values for $M^i$ below, as discussed in the next section, and above size $N_{th}$ (but above size $N^i>5$). The values 0.2 eV and 0.9 eV gave the best results. The effect of choosing different threshold sizes for visibility, $N_{th}$, was also studied. A few cases are shown in Fig. \ref{monica_Nth.pdf} for the visible SIA cluster density. Obviously, a higher $N_{th}$ gives
lower visible SIA cluster densities, while the SIA cluster growth seems not to be significantly affected. Using $N_{th}=90$, good agreement is obtained with the experimental data \cite{hernandez2010transmission} both in terms of density and mean SIA sizes, the latter being shown for different $N_{th}$ in Fig. \ref{R0110-Z30-Ev04-60_298967_SIA_mean_size.pdf}. No significant impact on the vacancy cluster densities and cluster mean sizes were observed when varying the $N_{th}$ parameter. This is likely because the vacancy trapping energy, $E_t^v$, which is the important parameter for the vacancy clusters, is not dependent on the $N_{th}$ parameter.
\begin{table}
\centering
\caption{Overview of the parameters for irradiation temperatures above 370 K, as used in this paper. All are a function of the cluster size, $N^\delta$, where $\delta$ = $i$ for SIA clusters and $v$ for vacancy clusters. $M^i$ is the migration energy for SIA clusters. The values for size 1 to 5 are the same as in \cite{jansson2013simulation}. $E_{t1}^i$ and $E_{t2}^i$ are the SIA trapping energies representing C and C$_2$V complexes, respectively. $E_t^v$ are the trapping energy for vacancy clusters. It is worth noting that the $E_{t2}^i$ energy is only valid up to $\sim$700 K, as the C$_2$V complexes will dissolve to C atoms and vacancies at this temperature.}
\label{table:M_i}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}} l l l l l}
\toprule
$N^\delta$ & $M^i$ & $E_{t1}^i$ & $E_{t2}^i$ & $E_t^v$\\
& [eV] & [eV] & [eV] & [eV] \\
\midrule
1 & 0.31 & 0.17 \cite{becquart2011p60} & 0.6 & 0.65 \cite{becquart2011p60} \\
2 & 0.42 & 0.28 \cite{becquart2011p60} & 0.6 & 1.01 \cite{becquart2011p60} \\
3 & 0.42 & 0.36 \cite{becquart2011p60} & 0.6 & 0.93 \cite{becquart2011p60} \\
4 & 0.80 & 0.34 \cite{becquart2011p60} & 0.6 & 0.96 \cite{becquart2011p60} \\
5 & 0.10 & 0.60 \cite{jansson2013simulation} & 1.2 & 1.23 \cite{becquart2011p60} \\
6 & 0.20 & 0.60 \cite{jansson2013simulation} & 1.2 & 1.20 \cite{becquart2011p60} \\
7--$N_{th}$ & 0.20 & 0.60 & 1.2 & 0.4 \\
$N_{th}<$ & 0.9 & 1.1 & 0.6 & 0.4 \\
\bottomrule
\end{tabular*}
\end{table}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Nth.pdf}
\caption{Visible SIA cluster density versus dpa for different values of the visibility threshold, $N_{th}$. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line corresponds to one cluster in the simulated volume.}
\label{monica_Nth.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Nth_SIA_mean_size_evolution.pdf}
\caption{The SIA mean cluster size versus dpa for different values of the threshold, $N_{th}$. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{R0110-Z30-Ev04-60_298967_SIA_mean_size.pdf}
\end{figure}
\subsection{Effect of varying the migration energy for invisible SIA clusters}\label{sec:small_sia}
Figs. \ref{monica_Ms.pdf} and \ref{monica_Ms_SIA_mean_size_evolution.pdf} shows the visible SIA cluster density and mean size, respectively, for different migration energy of SIA clusters of small sizes, $6\leq N^i < N_{th}$, \textit{i.e.} invisible cluster sizes. For sizes above $N_{th}$, $M^i=0.9$ eV, as described above and in Table \ref{table:M_i}. The cluster density does not vary significantly by varying the migration energy between 0.2 eV and 0.9 eV for invisible clusters. With 0.1 eV, no SIA clusters grow above the visibility threshold size, $N_{th}$. With a slightly higher migration energy, the SIA clusters are slowed down enough for nucleation of larger clusters to occur. Using $M^i=0.2$ eV, gives the largest mean SIA cluster size and also the best agreement with the experimental data. The effect of the $M^i$ for invisible SIA clusters on the vacancy density and mean cluster size is minimal. Over all, $M^i=0.2$ eV gives the best result.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Ms.pdf}
\caption{Visible SIA cluster density versus dpa for different migration energies $M^i_s$ for SIA clusters of size $6\leq N^i < N_{th}$. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line corresponds to one cluster in the simulated volume.}
\label{monica_Ms.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Ms_SIA_mean_size_evolution.pdf}
\caption{The SIA mean cluster size versus dpa for different migration energies $M^i_s$ for SIA clusters of size $6\leq N^i < N_{th}$. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{monica_Ms_SIA_mean_size_evolution.pdf}
\end{figure}
\subsection{Effect of varying the vacancy cluster trapping energy}
Vacancies are more mobile at 563 K than at 343 K, which was the irradiation temperature in our previous work \cite{jansson2013simulation}. Thus, vacancy traps play a more important role in this situation. In \cite{jansson2013simulation}, only V clusters of sizes $N^v \leq 6$ were trapped, essentially because precise data were available only up to that size and above it the migration energy was already so large that cluster migration would become a very rare event. In this study we tried different trapping energy for vacancy clusters above size $N^v = 6$. This trapping effect would correspond to the formation of complexes with C or other interstitial impurities. The results are shown in Fig. \ref{monica_Etvac.pdf} and \ref{monica_Etvac_SIA_mean_size_evolution.pdf} for the visible SIA cluster density and mean size evolution, respectively. For the vacancy clusters, the density and mean size evolutions are shown in Fig. \ref{monica_Etvac_vac_density_evolution.jp} and \ref{monica_Etvac_vac_mean_size_evolution.pdf}
, respectively. It can be seen that this parameter has a quite large effect, when varied. A higher $E_t^v$ increases both the cluster densities and the cluster mean sizes for both SIA and vacancy clusters. The exception to this trend is when considering only the vacancy clusters with the larger SANS resolution, whose mean size actually decreases with increased $E_t^v$. This effect is likely to be due to the fact that more strongly trapped vacancy clusters lead to a decrease of the number of recombinations and clustering of vacancy clusters, thereby decreasing the mean size of the large vacancy clusters visible by SANS, but increasing both size and density of the smaller vacancy clusters, seen by PAS. Less recombination leads to more SIA clusters and faster SIA cluster growth.
Taking all data for both SIA clusters and vacancies into account, the best fit is given by $E_t^v = 0.4$ eV, because it gives good agreement for the cluster densities and fair agreement for the cluster mean sizes. If higher values are chosen, the mean size is closer to experiment but the density of SIA clusters is totally off, because it builds up and saturates too early. The trapping energy for all vacancy cluster sizes are listed in Table \ref{table:M_i}. We confirmed by repeating the irradiation experiment at $T=343$ K, reported in \cite{jansson2013simulation}, that these $E_t^v$ values do not change the results at lower temperature.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Etvac.pdf}
\caption{Visible SIA cluster density versus dpa for different values of the trapping energy for vacancy clusters above size 6. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line corresponds to one cluster in the simulated system.}
\label{monica_Etvac.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Etvac_SIA_mean_size_evolution.pdf}
\caption{SIA cluster mean size versus dpa for different values of the trapping energy for vacancy clusters above size 6. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{monica_Etvac_SIA_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Etvac_vac_density_evolution.pdf}
\caption{Vacancy cluster densities versus dpa for different $E_t^v$ values. Higher densities are calculated with the PAS size resolution; lower densities with the same colour are calculated with the SANS size resolution. Results are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign.}
\label{monica_Etvac_vac_density_evolution.jp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Etvac_vac_mean_size_evolution.pdf}
\caption{Vacancy cluster mean sizes versus dpa for different $E_t^v$ values. Larger mean sizes are calculated using the SANS size resolution; smaller mean sizes are calculated using the PAS resolution. Results are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign. SANS points corresponds to the two peaks in the observed size distribution: the lower point is the major peak.}
\label{monica_Etvac_vac_mean_size_evolution.pdf}
\end{figure}
\subsection{Determination of the bias between SIA and vacancy sink radii}
The dislocation density plays a bigger role in this study than in \cite{jansson2013simulation,jansson2014okmc}, as the density here is larger, $\rho_d = (7\pm2)\cdot10^{13}$ m$^{-2}$. The effect of changing $Z^i$ is shown in Fig. \ref{monica_Z.pdf} and \ref{monica_Z_SIA_mean_size_evolution.pdf}for what concerns SIA clusters; the effect on the vacancy clusters is shown in Fig. \ref{monica_Z_vac_density_evolution.jp} and \ref{monica_Z_vac_mean_size_evolution.pdf}. Somehow, decreasing the value of $Z^i$ has similar effect to increasing the trapping energy for vacancies. A lower $Z^i$ increases the SIA visible cluster density lead to an earlier build up of SIA cluster density and also increases the growth of the SIA clusters. For vacancy clusters, the effect is reversed, so the density and cluster mean size are increased with increased $Z^i$, even though the effect is more significant for vacancy clusters visible by SANS than at the PAS size resolution. We get good agreement for the cluster density and the
vacancy cluster mean size, using $Z^i=3.0$. For the SIA cluster mean size evolution, the same $Z^i$ underestimates the cluster mean sizes slightly, but remains acceptable. The mean size would be closer to experiment with smaller values of $Z^i$, but then the density of SIA clusters would be too large and the build up would start too early and with early saturation (Fig. \ref{monica_Z.pdf}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Z.pdf}
\caption{Visible SIA cluster density versus dpa for two different values of the SIA bias factor for sinks, $Z^i$. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line corresponds to one cluster in the simulated volume.}
\label{monica_Z.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Z_SIA_mean_size_evolution.pdf}
\caption{SIA mean cluster size versus dpa. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{monica_Z_SIA_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Z_vac_density_evolution.pdf}
\caption{Vacancy cluster densities versus dpa for different $Z^i$ values. The higher densities are calculated with the PAS size resolution, whereas the lower densities with the same colour are calculated with the SANS size resolution. The data are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign.}
\label{monica_Z_vac_density_evolution.jp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_Z_vac_mean_size_evolution.pdf}
\caption{Vacancy cluster mean sizes versus dpa for different $Z^i$ values. Larger mean sizes are calculated using the SANS size resolution; smaller mean sizes are calculated using the PAS resolution. Results are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign. SANS points corresponds to the two peaks in the observed size distribution: the lower point is the major peak.}
\label{monica_Z_vac_mean_size_evolution.pdf}
\end{figure}
\subsection{Impact of the concentration of carbon in the matrix}
The impact of the C concentration was studied by trying different trap concentrations, as shown in Fig. \ref{monica_C.pdf}. Our reference experimental material is estimated to have a carbon concentration of 134 appm. Higher trap concentration gives a higher SIA cluster density, but even 200 appm of carbon still gives a fair agreement with the experimental data. Higher C concentrations also increase the mean SIA cluster sizes, as seen in Fig. \ref{monica_C_SIA_mean_size_evolution.pdf}.
The vacancy density increases slightly with increased C concentration from $8.6\cdot10^{22}$ m$^{-3}$ (at 0.2 dpa) with 50 appm C to $2.09\cdot10^{23}$ m$^{-3}$ with 300 appm C (\textit{Cf.} Fig. \ref{monica_C_vac_density_evolution.pdf}). If only considering the cluster sizes observable by SANS, these densities increases as well slightly from $1.5\cdot10^{22}$ m$^{-3}$ (at 0.2 dpa) with 50 appm to $2.7\cdot10^{22}$ m$^{-3}$ with 300 appm C (\textit{Cf.} \ref{monica_C_vac_mean_size_evolution.pdf}). The experimental SANS value at 0.2 dpa with 134 appm C is $4.0\cdot10^{22}$ m$^{-3}$, so a higher density of C concentration slightly improves the agreement with the experimental SANS data. The density and mean size evolutions are all very similar, with early saturations for all four C concentrations.
The average vacancy cluster size for clusters visible by PAS are very similar for the four different C concentrations; between 0.61 and 0.67 nm and thus in good agreement with the experimental PAS value, 0.6 nm. No clear trend can be observed. For clusters visible by SANS, the mean sizes at 0.2 dpa goes from 4.23 nm with 50 appm C to 2.83 nm with 300 appm C. A larger C density appears to lead to more nucleation and less recombination of vacancy clusters, which leads to higher vacancy cluster densities, but slower growth and thus smaller mean size in the SANS size category, at the dose considered here. Less recombinations also promotes the growth of SIA clusters. The main peak in the SANS bimodal distribution is at a size of 1.9 nm. The agreement of the mean cluster sizes with SANS data thus improves with higher concentration of C, suggesting that the actual content of interstitial impurities in the material might have been higher than reported. It is also quite possible that we have underestimated the number
of free C atoms, compared to the number of C$_2$V complexes. As C atoms are stronger traps than the C$_2$V complexes, less of the latter complexes gives more nucleation points.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_C.pdf}
\caption{Visible SIA cluster density versus dpa for different C concentrations. The experimental TEM data are from \cite{hernandez2010transmission}. The dotted line corresponds to one cluster in the simulated system.}
\label{monica_C.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_C_SIA_mean_size_evolution.pdf}
\caption{Visible SIA cluster mean size versus dpa for different C concentrations. The experimental TEM data are from \cite{hernandez2010transmission}.}
\label{monica_C_SIA_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_C_vac_density_evolution.pdf}
\caption{Vacancy cluster densities versus dpa for different C content. The higher densities are calculated with the PAS size resolution, whereas the lower densities with the same colour are calculated with the SANS size resolution. The data are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign.}
\label{monica_C_vac_density_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{monica_C_vac_mean_size_evolution.pdf}
\caption{Vacancy cluster mean sizes versus dpa for different C content. Larger mean sizes are calculated using the SANS size resolution; smaller mean sizes are calculated using the PAS resolution. Results are compared with experimental PAS \cite{lambrecht2009phd,meslin2010characterization} and SANS data \cite{bergner2010comparative} from the REVE campaign. SANS points corresponds to the two peaks in the observed size distribution: the lower point is the major peak.}
\label{monica_C_vac_mean_size_evolution.pdf}
\end{figure}
\section{Discussion}\label{sec:discussion}
The evolution of the vacancy and SIA clusters under irradiation are highly dependent on each other, making it delicate to fine-tune parameters of physical significance but unknown value to satisfy the experimental data for both kinds of defects. Changing parameters for vacancies, such as the trapping energy for vacancy clusters, $E_t^v$, will have a significant effect on the SIA cluster evolution, as we have shown above. For the final set of parameters of our model, we had to fine-tune three parameters: the migration energy of invisible SIA clusters, $M^i$, ($N<N_{th}$), the energy to trap vacancy clusters at impurity atoms or other features, $E_t^v$ and the bias factor, $Z^i$. The chosen values had to make sense physically and make the simulations satisfy all the experimental data. In the fine-tuning process, it was seen that it is not straightforward to get at the same time good agreement for density and size. A compromise had to be found as, especially in the case of SIA clusters, values of the parameters
that improve the mean size significantly increase the density, leading to too early build up and saturation. Although these differences are partially compensated by assuming that the content of interstitial species is higher than the nominal one, quite obviously the model has inherent limitations that we shall discuss in what follows.
One strong limitation is given by the fact that we consider in the model, for the sake of simplicity, a single population of loops, with different features depending on size. These features correspond to either the properties of $\langle100\rangle$ loops, of to the properties of a "mixture" of $\langle100\rangle$ and $1/2\langle111\rangle$ loops.
We do not know the proportions of $\langle100\rangle$ and $1/2\langle111\rangle$-SIA clusters, nor how the former clusters are created.
Are they the result of the interaction between $1/2\langle111\rangle$ loops created in cascades, as predicted by interatomic potentials \cite{stoller1996point,stoller1997primary,stoller2000statistical,
stoller2000evaluation,stoller2004secondary,stoller2000role,marian2002mechanism,terentyev2006study}? Or do they appear as the result of spontaneous transformation from $1/2\langle111\rangle$ loops \cite{arakawa2006changes}? Or is there an influence of temperature on their stability that makes them produced directly in cascades, this effect being unknown to the interatomic potentials \cite{dudarev2008effect}? Discussing the plausibility of these possibilities goes beyond the scope of the present work. Given this uncertainty, pragmatically we assumed that the type of loop is determined by size: visible loops are $\langle100\rangle$, invisible ones are of unknown nature. For these, we make use of an effective value of migration energy $M^i=0.2$ eV for $6\leq N^i \leq N_{th}$, where $N_{th}$ is the threshold for visibility. Clearly, a fully physical model should include a proper description of both classes of loops. However, given the uncertainties hinted at above, the development of such a model would imply
exploring one by one all possible mechanisms, with the use, unavoidably, of other equally uncertain assumptions. Before such a model can be reliably developed more insight into the physics of SIA clusters is needed, otherwise the increased complexity would not be necessarily rewarded by an increased reliability.
Another strong approximation in our model concerns the fact that we introduce traps that are conceptually associated with the presence of C and C-vacancy complexes trapping SIA clusters, as well as vacancy clusters, but the behaviour of which is only remotely connected with C atoms and the complexes that it forms with defects. To start with, the traps are introduced since the very beginning, while in reality they will first need to form. Secondly, the traps are rigorously immobile and unchanged in nature throughout the simulation, while C atoms are highly mobile at the temperature considered here: the stability of C-vacancy complexes is such that they might be continuously forming and dissociating, and it cannot be excluded that some mechanism of migration of these complexes might exist. The dynamic formation and disappearance of complexes is likely to affect the kinetics of the nanostructural evolution under irradiation and we should probably suppose that the traps we use are somehow more efficient than the
actual traps would be. However, again, the removal of this simplifying assumption and the introduction of a complete description of the physics of the traps implies the knowledge of a large number of parameters and mechanisms that we currently know very poorly. So, a more complete model should also eventually make use of assumptions, while being much more complex, thereby making it questionable whether the increased level of physical detail would be rewarded by higher reliability.
For the parameterization of the traps, we use \textit{ab initio} values for the interaction of single C atoms with SIA clusters, $E_{t1}^i$, for sizes up to $N^i=4$, and also with vacancy clusters, $E_t^v$, up to size 6 (\textit{Cf.} Table \ref{table:M_i}). Above these sizes the trapping energies were assumed, guessed or fitted. For the SIA traps, two populations of traps were used, representing the two dominating complexes that are able to trap SIA clusters: C and C$_2$V. Below size $N_{th}$, we can assume that the more mobile $1/2\langle111\rangle$-SIA clusters, assumed to be present in the "mixture", are more affected by the traps than the slowly moving $\langle100\rangle$ clusters, so the trapping energies are chosen to represent the interaction between the former SIA clusters and the C ($E_{t1}^i$) and C$_2$V ($E_{t2}^i$) complexes, respectively. A C atom binds to a $1/2\langle111\rangle$-SIA cluster with $\sim$0.6 eV \cite{terentyev2011interaction}. This value is thus used for $E_{t1}^i$ for $5\leq
N^i\leq N_{th}$. Above size $N_{th}$, the interaction is assumed to be exclusively with $\langle100\rangle$ and we used the MD result for the
binding energy between said cluster and the C atom, $E_{t1}^i=1.1$ eV \cite{anento2013unpublished}.
The second kind of SIA trap is supposed to represent C$_2$V complexes. We do not know the exact values for the binding between these complexes and small SIA clusters. We used for the smallest sizes (1--4) $E_{t2}^i=0.6$ eV. For $N^i\geq 5$, but $<N_{th}$, we fitted the trapping energy to be $E_{t2}^i=1.2$ eV, which can be seen as an intermediate value between the binding energy between a C$_2$V cluster and a $1/2\langle111\rangle$-SIA cluster, 1.5 eV, or a $\langle100\rangle$-SIA cluster, 0.6 eV. We assume SIA clusters larger than $N_{th}$ to be of $\langle100\rangle$ type, and $E_{t2}^i = 0.6$ eV corresponds to the binding energy between C$_2$V complexes and these clusters \cite{anento2013unpublished}.
For the simulations at lower temperature, 343 K, reported in \cite{jansson2013simulation, jansson2014okmc}, no trapping of vacancies above size 6 was used. In this study at 563 K, we found that a trapping energy of $E_t^v=0.4$ eV is necessary. The higher mobility of vacancy clusters at the higher temperature makes the traps more important. At low temperature, the vacancy clusters are so slowly moving that traps have no effect. The vacancy traps represent C and C$_2V$ complexes, even though we assume them both to have the same trapping energies for vacancies. We do not have any knowledge of calculations of the binding energy between vacancy clusters above size 6 and C atoms, making fitting the only option.
The bias factor for SIA sinks, $Z^i$, as defined in Eq. \eqref{eq:sink_radius}, proved to be an important parameter in our model. The bias factor takes into account the larger strain fields of SIA clusters, as compared to vacancy clusters. In rate theory calculations, the value chosen has usually been between 1.3 and 1.5. However, in order to lower the SIA density and still keep the general trend according to the experimental data, we found that a much higher value of $Z^i=3.0$ was needed.
It is hard to find a physical explanation for why such a high value is needed and perhaps it is merely a result of other approximations of the model. Possibly, the assumption of immobile traps is responsible here. Indeed, it is as if the C atoms, represented by traps, were immobile in our model. It is possible that several mobile C atoms could attach to the same vacancy or SIA cluster, thereby reducing the number of nucleation points for cluster growth. In our model, the number of traps is constant and given by the nominal C content in the material used in the experiment. We might thus be overestimating the number of nucleation points and, by increasing the sink strength for the SIA sinks by a large value of $Z^i$, we slow down the nucleation and clustering process at low dpa values. At higher dpa, the clusters are more dominating than the pre-existing sinks and new SIA defects are more likely to contribute to the cluster growth than to disappearing at the sinks. The main effect of a high $Z^i$ is thus a
delay of the visible SIA cluster growth. It can be likewise used to explore the effect of different dose rates or high fluence. These studies will be the subject of forthcoming papers.
\section{Summary and conclusions}\label{sec:conclusions}
We have developed a physical model that describes nanostructural evolution under irradiation in Fe-C alloys versus dose, at temperatures in the range of the operation conditions of reactor pressure vessel steels. The model reproduces fairly well experimental data, using only a few calibration parameters of clear physical meaning.
The main assumptions of the model, based on a combination of experimental and theoretical results, are:
\begin{itemize}
\item C atoms and complexes involving C atoms and vacancies, mainly C$_2$V, act as traps for SIA clusters and their effect can be described in first approximation as immobile traps to which a given trapping energy is associated that depends on the size of the trapped cluster.
\item While visible SIA loops have $\langle100\rangle$ Burgers vector, those invisible to the electronic microscope include loops of both $\langle100\rangle$ and $1/2\langle111\rangle$ type, and this fact can be taken into account by using effective migration parameters for invisible clusters.
\end{itemize}
This model can be used as a starting point to add, explicitly or effectively, the effect of substitutional solute atoms found in reactor pressure vessel steels and known to be responsible for their hardening and embrittlement, such as Cu, Ni and Mn.
\section*{Acknowledgement}
This work was carried out as part of the PERFORM60 project of the 7th Euratom
Framework Programme, partially supported by the European Commission, Grant
agreement number FP7-232612.
\bibliographystyle{model1a-num-names}
|
\section{Introduction}
\subsection{Aims and objectives}
The two fundamental concepts around which this article is orbiting are those of {\em weak equivalence} and {\em Morita equivalence}. Recall from, {\em e.g.}, \cite[\S5]{MoeMrc:LGSAC}
that two Lie groupoids ${\mathscr G}$ and ${\mathscr G}'$ are called weakly equivalent if there exist weak equivalences $\phi: {\mathscr H} \to {\mathscr G}$ and $\phi': {\mathscr H} \to {\mathscr G}'$ for
some third Lie groupoid ${\mathscr H}$ (see again {\em op.~cit.}~ for the precise definition of a weak equivalence $\phi$). For instance, the groupoids associated to two atlases of a manifold (or two transverse atlases of a foliated manifold) are weakly equivalent; each groupoid associated to a principal bundle of a Lie group $G$ and base manifold ${\mathcal M}$ is weakly equivalent to the unit Lie groupoid ${\mathscr U}({\mathcal M})$.
As a definition of Morita equivalence of two (Lie) groupoids might serve reversing the (classical) Morita theorem, that is, the requirement that their categories of representations (quasi-coherent ${\mathscr G}$-sheaves of $\Bbbk$-modules) are equivalent as symmetric monoidal categories. This leads to a quite general idea of equivalence which can be applied to any mathematical object that allows for the notion of ``representation'', or, more generally, (co)modules.
That the two notions of weak equivalence and Morita equivalence are essentially the same and also imply the presence of a principal bibundle (in an appropriate sense) is a well-known fact for (Lie) groupoids (in fact, the terminology varies and often coincides, which adds somewhat to the confusion), see \cite{MuhRenWil:EAIFGCSA, Hae:GDHEC, Mrc:FOTBATAHSM}. Note, however, that in the first of these references the respective concept of principal bundle slightly differs from the latter two.
Taking Lie groupoids as objects, one constructs, together with the isomorphism classes of principal bundles (as morphisms, sometimes called {\em Hilsum-Skandalis maps}) and equipped with the tensor product, a category, sometimes called the {\em Morita category}. Moreover, there is a functor from the category of Lie groupoids to this Morita category which transforms weak equivalences to isomorphisms that establishes a universal solution for functors having this property.
Roughly speaking, \emph{commutative} Hopf algebroids can be seen as presheaves of groupoids on affine schemes: the datum of a {\em flat} Hopf algebroid is equivalent to the datum of a certain stack with a specific presentation \cite{Naumann:07, FalCha:DOAV}.
In this perspective, one can establish an equivalence between (right) comodules over a Hopf algebroid and quasi-coherent sheaves with a groupoid action \cite[Thm.~2.2]{Hovey:02}.
Hopf algebroids
were introduced
in algebraic topology (see, {\em e.g.}, \cite{Rav:CCASHGOS}) as a {\em cogroupoid} kind of object,
which motivates the following definitions taken from \cite[Def.\ 6.1]{HovStr:CALEHT} resp.\ \cite{Hovey:02}.
For the necessary ingredients and notation used therein we refer to the main text.
\begin{definition}
\label{quellala1}
\label{quellala2} Let $(A,{\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids.
\begin{enumerate}
\item
A morphism $(A,{\mathcal H}) \to (B,{\mathcal K})$
is said to be a \emph{weak equivalence} if and only if
the respective induction functor
$\rcomod{{\mathcal H}} \to \rcomod{{\mathcal K}}$ establishes an equivalence of categories. The Hopf algebroids $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are said to be \emph{weakly equivalent} if there is a diagram
\vspace{-.1cm}
$$
\xymatrix@R=8pt{ & (C,{\mathcal J}) & \\ (A,{\mathcal H}) \ar@{->}
[ur] & & \ar@{->}[ul] (B,{\mathcal K}) }
\vspace{-.1cm}
$$
of weak equivalences of Hopf algebroids.
\item
Two flat Hopf algebroids are said to be {\em Morita equivalent} if their categories of (right) comodules are equivalent as symmetric monoidal categories.
\end{enumerate}
\end{definition}
For instance, the existence of a weak equivalence implies Morita equivalence since induction functors are always symmetric monoidal functors.
In the context of Hopf {\em algebras}, the second part in the above definition appeared in \cite[Def.~3.2.3]{Schau:HGABGE} baptised {\em monoidal Morita-Takeuchi} equivalence therein but also before in \cite[Def.~5.6]{Schau:HBGE}, where such a property was called {\em monoidal co-Morita} equivalence.
Let us also mention that a Morita theory for certain cocommutative Hopf algebroids (so-called {\em \'etale Hopf algebroids}) was developped in \cite{Mrc:THAOFOEGATPME} using a different notion of bundles (called {\em principal bimodules}). Furthermore, the idea of describing Morita theory in the language of bicategories was explained, for example, in \cite{Lan:BOOAAPM} for various contexts, such as rings, $C^*$-algebras, von Neumann algebras, Lie groupoids, symplectic groupoids, and Poisson manifolds.
\subsection{Main results}
Transferring the above statements from Lie groupoids to the case of commutative Hopf algebroids will be the main task (and result)
of this article, summarised as follows:
{\renewcommand{\thetheorem}{{A}}
\begin{thm}
\label{processocivile}
Let $(A,{\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids. The following are equivalent:
\begin{enumerate}[(1)]
\item $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are Morita equivalent.
\item There is a principal bibundle connecting $(A,{\mathcal H})$ and $(B,{\mathcal K})$.
\item $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are weakly equivalent.
\end{enumerate}
\end{thm}
}
One might be tempted to think that these results can be obtained by simply dualising the usual techniques in the groupoid case but things turn out to be more intricate: one of the main obstacles in mimicking the groupoid case is the construction of orbit spaces which correspond to quotients of affine schemes, which is a subtle concept with its own challenges.
In contrast to that, our arguments make large use of cotensor products of comodule algebras in correspondence to these quotients of affine schemes, which might seem technical at first sight but proves useful in this context.
The subsequent picture shows all implications between $(1), (2)$, and $(3)$ in the above theorem that we will explore in the main text:
\begin{small}
\begin{equation*}
\xymatrix{ & & (1) \ar@/_0.8pc/@{=>}_-{{\scriptstyle{\text{Proposition } \ref{prop:llegaras}}}}[lldd] & & \\ & & & & \\ (2) \ar@/_0.8pc/@{=>}_-{\scriptstyle{\text{Theorem } \ref{aromanflower}}}[rruu] \ar@{<=>}_-{\scriptstyle{\text{Proposition } \ref{prop:PHPKP}}}[rrrr] & & & & (3)
\ar@/_1pc/@{=>}_-{{\scriptstyle{\text{trivial }}}}[lluu] }
\end{equation*}
\begin{center}
\vspace{0.05cm}
\captionof{figure}{\footnotesize Paths in the proof of Theorem \ref{processocivile}}\label{ahaaha}
\end{center}
\end{small}
\vspace{-.5cm}
In particular, the step $(1) \Rightarrow (3)$ in the above Theorem \ref{processocivile} was conjectured in \cite[Conj.~6.3]{HovStr:CALEHT}: more precisely, Hovey and Strickland conjectured that in case the category of ${\mathcal H}$-comodules is equivalent to the one of comodules over ${\mathcal K}$, then the two Hopf algebroids $(A,{\mathcal H})$ and $(B, {\mathcal K})$
are connected by a chain of weak equivalences, and we show that this chain can be taken to be of length $2$.
By a chain of weak equivalences of length $n \geq 2$ we mean a zig-zag of weak equivalences in the sense of \cite[Def.~7.9.1]{Hirschhorn}, up to the equivalence transformations given in \cite[\S14.4]{Hirschhorn}. The key here is Proposition \ref{prop:1out2}, which shows that any zig-zag of weak equivalences of the form $\xymatrix{ \bullet & \ar@{->}^-{}[r] \bullet \ar@{->}^-{}[l] & \bullet }$
can be completed to a diagram of weak equivalences having the form
\begin{small}
$$
\xymatrix@R=15pt@C=15pt{ & \circ & \\ \bullet \ar@{-->}^-{}[ru] & & \ar@{-->}^-{}[lu] \bullet \\ & \ar@{->}^-{}[lu] \bullet \ar@{->}^-{}[ru] &
}
$$
\end{small}
which is commutative up to a $2$-isomorphism (a property dual to condition (BF3) in \cite[p.~254]{Pro:EASABOF}).
In this way, any chain of weak equivalences (in the above sense) between two flat Hopf algebroids $(A,{\mathcal H})$ and $(B,{\mathcal K})$ can be transformed
to one of the form
\begin{footnotesize}
$$
\mathfrak{Z}_{\scriptstyle{k+2}}:\;\;
\xymatrix@R=8pt@C=8pt{
& \scriptstyle{(D_1,\,{\mathcal I}_1)} & & \scriptstyle{(D_2,\,{\mathcal I}_2)} & & \scriptscriptstyle{\cdots\cdots\cdots\cdots\cdots\cdots} & & \scriptstyle{(D_{k},\,{\mathcal I}_{k})} & &\scriptstyle{(D_{k+1},\,{\mathcal I}_{k+1})} & \\
\scriptstyle{(A,{\mathcal H})} \ar@{->}^-{}[ru] & & \ar@{->}^-{}[ru] \scriptstyle{(C_1,\,{\mathcal J}_1)} \ar@{->}^-{}[lu] & & \ar@{->}^-{}[lu] \scriptstyle{(C_{2},\,{\mathcal J}_{2})} & \scriptscriptstyle{\cdots\cdots\cdots\cdots\cdots\cdots} & \scriptstyle{(C_{k-1},\,{\mathcal J}_{k-1})} \ar@{->}^-{}[ru] & & \ar@{->}^-{}[ru] \scriptstyle{(C_{k},\,{\mathcal J}_{k})} \ar@{->}^-{}[lu]& & \ar@{->}^-{}[lu] \scriptstyle{(B,{\mathcal K})}
}
$$
\end{footnotesize}
of length $2(k+1)$, which, in turn, can be completed to the following isosceles triangle
\begin{small}
$$
\xymatrix@R=10pt@C=12pt{ & & & & & \scriptstyle{(C_{k1},\,{\mathcal J}_{k1})} & & & & & \\
& & & & \ar@{->}^-{}[ru] \scriptstyle{(C_{(k-1)1},\,{\mathcal J}_{(k-1)1})} & & \ar@{->}^-{}[lu] \scriptstyle{(C_{(k-1)2},\,{\mathcal J}_{(k-1)2})} & & & & \\
& & & \ar@{.}^-{}[ru] \scriptstyle{(C_{11},\,{\mathcal J}_{11})} & & \ar@{.}^-{}[ru] \scriptscriptstyle{\!...\!} \ar@{.}^-{}[lu] & & \ar@{.}^-{}[lu] \scriptstyle{(C_{1k},\,{\mathcal J}_{1k})} & & & \\
& & \scriptstyle{(D_1,\,{\mathcal I}_1)} \ar@{->}^-{}[ru] & & \ar@{->}^-{}[lu]\scriptstyle{(D_2,\,{\mathcal I}_2)} \ar@{.}^-{}[ru] & & \ar@{.}^-{}[lu] \scriptstyle{(D_{k},\,{\mathcal I}_{k})} \ar@{->}^-{}[ru] & & \scriptstyle{(D_{k+1},\,{\mathcal I}_{k+1})} \ar@{->}^-{}[lu] & & \\
& \scriptstyle{(A,\,{\mathcal H})} \ar@{->}^-{}[ru] & & \ar@{->}^-{}[lu] \scriptstyle{(C_1,\,{\mathcal J}_1)} \ar@{->}^-{}[ru] & & \ar@{.}^-{}[lu] \scriptscriptstyle{\!...\!} \ar@{.}^-{}[ru] & & \ar@{->}^-{}[lu] \scriptstyle{(C_k,\,{\mathcal J}_k)} \ar@{->}^-{}[ru] & &\scriptstyle{(B,\,{\mathcal K})} \ar@{->}^-{}[lu] &
}
$$
\end{small}
of $(k+2)$ vertices on each side. Such a triangle is obtained by constructing ${k(k+1)}/{2}$ new flat Hopf algebroids being essentially two-sided translation Hopf algebroids built from trivial principal bundles.
The notion of {\em (quantum) principal bundle} that appears as a crucial ingredient in Theorem \ref{processocivile} is a relatively straightforward extension of the corresponding concept for Hopf algebras as introduced in \cite{BrzMaj:QGGTOQS}, see also \cite{Brz:TMIQPP}. In \cite[\S3.2.4]{Schau:HGABGE}, again in the realm of Hopf algebras, these objects were called {\em bi-Galois objects} and the corresponding implications $(1) \Leftrightarrow (2)$ of Theorem \ref{processocivile} were shown.
As a matter of fact, in many examples constructing bi-Galois objects or principal bundles has turned out to be a practicable way to establish monoidal equivalences between comodule categories; as a concrete illustration, see, for example, \cite{Mas:CDAGOFSCHAOFD, Bic:TRCOTQGOANDBF}. Analogous objects in sheaf theory
are known under the name of \emph{(bi)torsors}, see \cite{DemGab:GATIGAGGC}.
In fact, we gather flat Hopf algebroids and principal bundles along with their morphisms in a bicategory. More precisely, in Proposition \ref{piazzadellorologio} we prove that the data given by
\begin{center}
\begin{enumerate}[\quad \raisebox{1pt}{$\scriptstyle{\bullet}$}]
\item
flat Hopf algebroids (as $0$-cells),
\item
left principal bundles (as $1$-cells),
\item
as well as morphisms of left principal bundles (as $2$-cells)
\end{enumerate}
\end{center}
define a bicategory, denoted by $\mathsf{PB}^{\scriptscriptstyle{\ell}}$.
The bicategories of analogously constructed right resp.\ two-sided principal bundles (or bibundles) are denoted by $\mathsf{PB}^{\scriptscriptstyle{r}}$ and $\mathsf{PB}^{\scriptscriptstyle{b}}$, respectively. As in classical situations,
for two $0$-cells $(A,{\mathcal H})$ and $(B,{\mathcal K})$, the category $\lPB{{\mathcal H}}{{\mathcal K}}$ turns out to be a groupoid.
This leads to the structure of a \emph{bigroupoid} on the bicategory $\mathsf{PB}^{\scriptscriptstyle{b}}$, and hence to a \emph{categorical group} (or \emph{bigroup}) structure on each category $\bPB{{\mathcal H}}{{\mathcal H}}$, see, for instance, \cite{Noo:NOTGTGACM}.
Applying Theorem \ref{processocivile} above to a single flat Hopf algebroid yields the following result:
{\renewcommand{\thetheorem}{{B}}
\begin{theorem}
\label{thm:B}
Let $(A,{\mathcal H})$ be a flat Hopf algebroid and denote by $\mathscr{U}({\mathcal H})$ its associated principal unit bibundle.
Then the category $\big(\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big)$ of symmetric monoidal auto-equivalences of right ${\mathcal H}$-comodules with morphisms given by natural tensor transformations forms a categorical group, and the functors
\begin{equation*}
\begin{array}{rclrcl}
\big(\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big)
&\longrightarrow&
\big(\bPB{{\mathcal H}}{{\mathcal H}},\cotensor{{\mathcal H}}, \mathscr{U}({\mathcal H}) \big), &
{\mathscr F} &\longmapsto& {\mathscr F}({\mathcal H})
\\
\big(\bPB{{\mathcal H}}{{\mathcal H}},\cotensor{{\mathcal H}}, \mathscr{U}({\mathcal H}) \big) &\longrightarrow&
\big(\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big), & (P,\alpha,\beta) &\longmapsto& -\cotensor{{\mathcal H}}P
\end{array}
\end{equation*}
establish a monoidal equivalence of categorical groups.
\end{theorem}
}
Moreover, it turns out that there is a $2$-functor
$$
\mathscr{P}: 2\text{-}\mathsf{HAlgd} \longrightarrow \mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}
$$
from the $2$-category of flat Hopf algebroids to the conjugate of $\mathsf{PB}^{\scriptscriptstyle{\ell}}$, which sends any $1$-cell $\B{\phi}: (A,{\mathcal H}) \to (B, {\mathcal K})$ to its associated trivial left principal bundle $\mathscr{P}(\B{\phi})={\mathcal H}\tensor{\scriptscriptstyle{\phi}}B$, that is, the pull-back of the unit bundle $\mathscr{U}({\mathcal H})$.
A $1$-cell $\B{\phi}$ in $2\text{-}\mathsf{HAlgd}$ is a weak equivalence if and only if $\mathscr{P}(\B{\phi})$ is an invertible $1$-cell in $\mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}$, \emph{i.e.}, is part of an internal equivalence. We then present the pair $(\mathsf{PB}^{\scriptscriptstyle{\ell}}, \mathscr{P})$ as the universal solution with respect to this property:
{\renewcommand{\thetheorem}{{C}}
\begin{thm}\label{thm:Ramadan}
Let $\mathscr{F}: 2\text{-}\Sf{HAlgd} \to \mathscr{B}$ be a $2$-functor which sends weak equivalences to invertible $1$-cells. Then, up to isomorphism (of $2$-functors), there is a unique $2$-functor $\td{\mathscr{F}}$ such that the diagram
\begin{equation*}
\xymatrix@R=20pt{ 2\text{-}\Sf{HAlgd} \ar@{->}_-{\mathscr{F}}[rrd] \ar@{->}^-{\mathscr{P}}[rr] & & \mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}} \ar@{.>}^-{\td{\mathscr{F}}}[d] \\ & & \mathscr{B} }
\end{equation*}
commutes up to an isomorphism of $2$-functors.
\end{thm}
}
We finally want to mention that this universality leads to a kind of calculus of fractions in
the $2$-category $2\text{-}\Sf{HAlgd}$ with respect to weak equivalences in a sense ``dual" to the approach in \cite{Pro:EASABOF}.
\bigskip
\noindent \textbf{Acknowledgements.}
It is a pleasure to thank Alessandro Ardizzoni, Federica Galluzzi, and Fabio Gavarini for stimulating discussions and useful comments.
\section{Hopf algebroids and comodule algebras}\label{sec:HalgdComodalg}
All algebras are considered to be commutative $\Bbbk$-algebras, where $\Bbbk$ is a commutative ground ring. The $\Bbbk$-module of all algebra maps from $R$ to $C$ will be denoted by $R(C):={\rm Alg}_{\Bbbk}\big(R,C\big)$.
\subsection{Hopf algebroids}
\label{ssec:halgd}
Recall from, {\em e.g.}, \cite{Rav:CCASHGOS}
that a {\em commutative} Hopf algebroid is a pair $(A,{\mathcal H})$ of two commutative $\Bbbk$-algebras together with a diagram
$
\xymatrix@C=25pt{A\ar@<1ex>@{->}|-{\scriptscriptstyle{\sf{s}}}[r] \ar@<-1ex>@{->}|-{\scriptscriptstyle{\sf{t}}}[r] & \ar@{->}|-{ \scriptscriptstyle{\varepsilon}}[l]{\mathcal H}}
$
of algebra maps,
a structure $({}_{\scriptscriptstyle{\Sf{s}}}{\mathcal H}_{\scriptscriptstyle{\Sf{t}}}, \Delta, \varepsilon)$ of an $A$-coring with underlying $A$-bimodule ${}_{\scriptscriptstyle{A}}{\mathcal H}_{\scriptscriptstyle{A}}=
{}_{\scriptscriptstyle{\Sf{s}}}{\mathcal H}_{\scriptscriptstyle{\Sf{t}}}$, along with an isomorphism
$\mathscr{S}: {}_{\scriptscriptstyle{\Sf{s}}}H_{\scriptscriptstyle{\Sf{t}}} \to {}_{\scriptscriptstyle{\Sf{t}}}H_{\scriptscriptstyle{\Sf{s}}}$ of $A$-corings that fulfils $\mathscr{S}^2 = {\rm id}$,
where the codomain is the opposite $A$-coring of ${}_{\scriptscriptstyle{\Sf{s}}}H_{\scriptscriptstyle{\Sf{t}}}$. The map $\mathscr{S}$ is called the \emph{antipode} of ${\mathcal H}$. All the previous maps are asked to be compatible in the following way:
\begin{eqnarray}
\varepsilon \circ \Sf{s} \,=\, {\rm id}_{\scriptscriptstyle{A}}, & & \varepsilon \circ \Sf{t}\,=\, {\rm id}_{\scriptscriptstyle{A}}, \\
\Delta(1_{\scriptscriptstyle{{\mathcal H}}})\,=\, 1_{{\scriptscriptstyle{{\mathcal H}}}}\tensor{A} 1_{{\scriptscriptstyle{{\mathcal H}}}}, & & \varepsilon(1_{{\scriptscriptstyle{{\mathcal H}}}})\,=\, 1_{\scriptscriptstyle{A}},
\label{Eq:es}\\
\Delta(uv)\,=\, u_{(1)}v_{(1)}\tensor{A} u_{(2)}v_{(2)}, & & \varepsilon(uv)\,=\, \varepsilon(u)\varepsilon(v),\label{Eq:demul}\\
\label{maxxi}
\Sf{t}(\varepsilon(u))\,=\, \mathscr{S}(u_{(1)})u_{(2)}, & & \Sf{s}(\varepsilon(u))\,=\, u_{(1)}\mathscr{S}(u_{(2)}),\label{Eq:penultima}\\
\mathscr{S}(uv)\,=\, \mathscr{S}(u) \mathscr{S}(v), & & \mathscr{S}(1_{{\scriptscriptstyle{{\mathcal H}}}})\,=\, 1_{{\scriptscriptstyle{{\mathcal H}}}},\label{Eq:ultima}
\end{eqnarray}
for every $a \in A$, $u, v \in {\mathcal H}$, where we used Sweedler's notation for the comultiplication.
As all Hopf algebroids in this article are commutative and flat over the base ring, they are also faithfully flat since both the source and target are (left) split morphisms of modules over the base ring.
\smallskip
A \emph{morphism}
$\B{\phi}: (A,{\mathcal H}) \to (B,{\mathcal K})$ {\em of Hopf algebroids}
consists of a pair $\B{\phi}=(\phi_{\scriptscriptstyle{0}},\phi_{\scriptscriptstyle{1}})$
of algebra maps
$\phi_{\scriptscriptstyle{0}}: A \to B$ and $\phi_{\scriptscriptstyle{1}}: {\mathcal H} \to {\mathcal K}$ that are
compatible
with the structure maps of both ${\mathcal H}$ and ${\mathcal K}$ in a canonical way.
That is, the equalities
\begin{eqnarray}
\phi_{\scriptscriptstyle{1}} \circ \Sf{s} \,\,=\,\, \Sf{s} \circ \phi_{\scriptscriptstyle{0}}, & &
\phi_{\scriptscriptstyle{1}} \circ \Sf{t} \,\,=\,\, \Sf{t} \circ \phi_{\scriptscriptstyle{0}}, \\
\Delta \circ \phi_{\scriptscriptstyle{1}} \,\,=\,\, \chi \circ (\phi_{\scriptscriptstyle{1}} \tensor{A}\phi_{\scriptscriptstyle{1}}) \circ \Delta, && \varepsilon \circ \phi_{\scriptscriptstyle{1}}\,\,=\,\, \phi_{\scriptscriptstyle{0}} \circ \varepsilon, \\
\mathscr{S} \circ \phi_{\scriptscriptstyle{1}} \,\,=\,\, \phi_{\scriptscriptstyle{1}} \circ \mathscr{S},
\end{eqnarray}
hold, where $\chi$ is the obvious map $\chi: {\mathcal K}\tensor{A}{\mathcal K} \to {\mathcal K}\tensor{B}{\mathcal K}$, and where no distinction between the structure maps of ${\mathcal H}$ and ${\mathcal K}$ was made.
\begin{example}[{\em Scalar extension Hopf algebroid}]
For a Hopf algebroid $(A,{\mathcal H})$ and an algebra map $\phi_{\scriptscriptstyle{0}}: A \to B$, we can consider the so-called \emph{scalar extension} Hopf algebroid $(B, B\tensor{A}{\mathcal H}\tensor{A}B)$ in a canonical way such that
$(\phi_{\scriptscriptstyle{0}}, \phi_{\scriptscriptstyle{1}}): (A,{\mathcal H}) \to (B, B\tensor{A}{\mathcal H}\tensor{A}B)$, where $\phi_{\scriptscriptstyle{1}}(u)= 1_{\scriptscriptstyle{B}} \tensor{A} u \tensor{A} 1_{\scriptscriptstyle{B}}$, becomes a morphism of Hopf algebroids. In this way, any morphism
$\B{\phi}: (A,{\mathcal H}) \to (B,{\mathcal K})$ of Hopf algebroids factors through the following morphism
\begin{equation}\label{Eq:PHI}
{\Phi}: (B, B\tensor{A}{\mathcal H}\tensor{A}B) \to (B,{\mathcal K}), \quad b\tensor{A}u\tensor{A}b' \mapsto \Sf{s}(b)\phi_{\scriptscriptstyle{1}}(u)\Sf{t}(b')
\end{equation}
of Hopf algebroids.
\end{example}
\begin{rem}
Notice that the scalar extension Hopf algebroid $(B, B\tensor{A}{\mathcal H}\tensor{A}B)$ is not necessarily flat. This happens, for instance, if $\phi_{\scriptscriptstyle{0}}$ is a flat extension or if $B$ is {\em Landweber exact} over $(A,{\mathcal H})$ in the sense of \cite[Def.\ 2.1, Corollary 2.3]{HovStr:CALEHT}, which means that either the extension $A \to {\mathcal H}\tensor{A}B$, $a \mapsto \Sf{s}(a)\tensor{A}1_{\scriptscriptstyle{B}}$ or $A \to B\tensor{A} {\mathcal H}$, $a \mapsto 1_{\scriptscriptstyle{B}} \tensor{A} \Sf{t}(a)$ is flat, see also Remark \ref{rem:HS}. Another important situation is when ${\mathcal H}$ is assumed to be flat as an $A\tensor{}A$-module ({\em i.e.}, the extension $\Sf{s}\tensor{}\Sf{t}$ is flat). This happens, for instance, when ${\mathcal H}$ is geometrically transitive Hopf algebroid in the sense of Deligne and Brugui\`eres \cite{Deligne:1990, Bruguieres:1994}.
\end{rem}
\subsection{Comodules and bicomodules}
\label{ssec:bicomod}
This section gathers some standard material on comodules over commutative Hopf algebroids which will be needed in the sequel, see, {\em e.g.}, again \cite{Rav:CCASHGOS} for more information.
A right ${\mathcal H}$-comodule over a Hopf algebroid $(A, {\mathcal H})$ is a pair $(M,\rcoaction{{\mathcal H}}{M})$, where $M$ is an $A$-module and $\rcoaction{{\mathcal H}}{M}: M \to M\tensor{A} \shopf{H}$, $m \mapsto m_{(0)} \otimes_{\scriptscriptstyle{A}} m_{(1)}$ is an $A$-linear map, written in the usual Sweedler notation,
and which satisfies the usual coassociativity and counitary properties. Here, the $A$-module structure on $M\tensor{A} \shopf{H}$ with respect to which the coaction is $A$-linear is defined by $(m\tensor{A}u) \bract a := m\tensor{A} u\Sf{t}(a)$.
When the context is clear, we shall also drop sub- and superscripts on $\rcoaction{{\mathcal H}}{M}$ that are sometimes needed to distinguish various coactions.
{\em Morphisms of right ${\mathcal H}$-comodules} are defined in an obvious way, and the category of right ${\mathcal H}$-comodules will be denoted by $\rcomod{{\mathcal H}}$, whereas a morphism between two right ${\mathcal H}$-comodules $M$ and $N$ will be denoted as $\rcomod{{\mathcal H}}(M,N)$.
The category $\rcomod{{\mathcal H}}$ is symmetric monoidal, where the coaction on the tensor product is given by the codiagonal coaction, that is,
\begin{equation}
\label{Eq:tensor}
\rcoaction{{\mathcal H}}{M \tensor{A}N}: M\tensor{A}N \to (M\tensor{A}N)\tensor{A}\shopf{H},\quad m\tensor{A}n \mapsto (m_{(0)}\tensor{A}n_{(0)})\tensor{A} m_{(1)}n_{(1)}.
\end{equation}
The identity object is given by $(A,\Sf{t})$ and the symmetry is given by the natural transformation obtained from the tensor flip.
\begin{rem}
\label{mondragone}
There are situations where the tensor product $M\tensor{A}N$ of the underlying modules of two right ${\mathcal H}$-comodules can be endowed with more than one comodule structure. For distinction, we will from now on denote by
$M\otimes^{\scriptscriptstyle{A}} N$ the tensor product in $\rcomod{{\mathcal H}}$ endowed then with the coaction of equation \eqref{Eq:tensor}.
\end{rem}
To each right ${\mathcal H}$-comodule $(M,\rcoaction{}{})$ one can define the $\Bbbk$-vector space of \emph{coinvariants}:
$$
M^{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}=\big\{ m \in M \mid \rcoaction{}{}(m)=m\tensor{A}1_{\scriptscriptstyle{{\mathcal H}}} \big\}.
$$
This, in fact, establishes a functor which is naturally isomorphic to the functor $\rcomod{{\mathcal H}}\big(A,- \big)$, that is, we have a natural isomorphism of $\Bbbk$-vector spaces:
$$
\rcomod{{\mathcal H}}\big(A,M \big) \,\cong\, M^{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}.
$$
Analogously, one can define the category $\lcomod{{\mathcal H}}$ of left comodules, and both categories are isomorphic via the antipode. Explicitly, one can endow a left ${\mathcal H}$-comodule $(M,\lcoaction{{\mathcal H}}{M})$ with a right ${\mathcal H}$-comodule structure, denoted by $M^{\op}$,
\begin{equation}
\label{Eq:leftright}
\rcoaction{{\mathcal H}}{M^{\op}}: M^{\op} \to M^{\op}\tensor{A}\shopf{H},\quad m \mapsto m_{(0)}\tensor{A}\mathscr{S}(m_{(-1)}),
\end{equation}
and referred to as the \emph{opposite comodule} of $M$.
Since we always have $\mathscr{S}^2=id$ for commutative Hopf algebroids, this correspondence obviously establishes an isomorphism of symmetric monoidal categories.
For an arbitrary algebra $R$ and a right comodule $(N,\rcoaction{}{})$ whose underlying module is also an $(A,R)$-bimodule such that $\rcoaction{}{}$ is left $R$-linear, {\em i.e.}, $\rcoaction{{\mathcal H}}{M}(rn)=rn_ {(0)} \tensor{A} n_{(1)}$, for $r \in R$, $n \in N$, one can define a functor
\begin{equation}
\label{Eq:RM}
- \tensor{R}N : \rmod{R} \to \rcomod{{\mathcal H}}, \quad X \mapsto (X\tensor{R}N, X\tensor{R}\rcoaction{}{}).
\end{equation}
For two Hopf algebroids $(A,{\mathcal H})$ and $(B,{\mathcal K})$, the \emph{category of $({\mathcal H}, {\mathcal K})$-bicomodules} has
triples $(P, \lcoaction{{\mathcal H}}{P}, \rcoaction{{\mathcal K}}{P})$ as objects, where $P = \due P {\scriptscriptstyle{A}} {\scriptscriptstyle{B}}$ is an $(A,B)$-bimodule
such that $(P , \lcoaction{{\mathcal H}}{P})$ is a left comodule with a right $B$-linear coaction $\lcoaction{{\mathcal H}}{P}$, while $(P, \rcoaction{{\mathcal K}}{P})$ is right comodule with a left $A$-linear coaction $\rcoaction{{\mathcal K}}{P}$, and both coactions are compatible in the sense that
\begin{equation}\label{Eq:bicomod}
({\mathcal H}\tensor{A}\rcoaction{{\mathcal K}}{P}) \circ \lcoaction{{\mathcal H}}{P} \,\,=\,\,
(\lcoaction{{\mathcal H}}{P}\tensor{B}{\mathcal K}) \circ \rcoaction{{\mathcal K}}{P}.
\end{equation}
In other words, $\lcoaction{{\mathcal H}}{P}$ is a morphism of right ${\mathcal K}$-comodules, and $\rcoaction{{\mathcal K}}{P}$ of left ${\mathcal H}$-comodules, where the codomains of both maps are comodules according to the functor of equation \eqref{Eq:RM}. {\em Morphisms} of bicomodules are defined in a canonical way; denote by $\bicomod{{\mathcal H}}{{\mathcal K}}$ the category of $({\mathcal H},{\mathcal K})$-bicomodules.
Next, we recall the definition of the cotensor product. Let $(M, \rcoaction{}{})$ be a right ${\mathcal H}$-comodule and $(N,\lcoaction{}{})$ a left ${\mathcal H}$-comodule. The \emph{cotensor product bifunctor} is defined as the equaliser
$$
\xymatrix{ 0 \ar@{->}[r] & M\cotensor{{\mathcal H}}N \ar@{->}^-{}[r] & M\tensor{A}N
\ar@<0.5ex>@{->}^-{\scriptstyle{\rcoaction{}{}\tensor{A}N}}[rr] \ar@<-0.5ex>@{->}_-{\scriptstyle{M\tensor{A}\lcoaction{}{}}}[rr] & & M\tensor{A}{\mathcal H}\tensor{A} N, }
$$
which is a bifunctor from the product category $\rcomod{{\mathcal H}}\times \lcomod{{\mathcal H}}$ to $\rmod{A}$. If we further assume that $(N,\rcoaction{}{},\lcoaction{}{})$ is also an $({\mathcal H}, {\mathcal K})$-bicomodule,
the cotensor product lands in the category of right ${\mathcal K}$-comodules since our Hopf algebroids are flat.
This way, it is possible to define the bifunctor
\begin{equation}
\label{Eq:bicotensor}
-\cotensor{{\mathcal H}}-: \bicomod{{\mathcal J}}{{\mathcal H}}\times \bicomod{{\mathcal H}}{{\mathcal K}} \to \bicomod{{\mathcal J}}{{\mathcal K}}.
\end{equation}
One easily checks that ${\mathcal H}\cotensor{{\mathcal H}}N \cong N$ and $A\cotensor{{\mathcal H}}N \cong N^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} $ for every right ${\mathcal H}$-comodule $N$.
The associativity of the cotensor products is not always guaranteed unless one makes more assumptions on the comodules involved. For example, since all our Hopf algebroids are assumed to be flat, if $M$ is a flat $A$-module along with a flat $B$-module $N'$, one has
$$
M \cotensor{{\mathcal H}} (N \cotensor{{\mathcal K}} N') \simeq (M \cotensor{{\mathcal H}} N) \cotensor{{\mathcal K}} N'.
$$
Compare, for example, \cite[\S\S22.5--22.6]{BrzWis:CAC}
for more situations in which this associativity holds true.
\smallskip
Given a morphism
$\B{\phi}=(\phi_{\scriptscriptstyle{0}},\phi_{{1}}) :(A,{\mathcal H}) \to (B,{\mathcal K})$ of Hopf algebroids, there is a functor
\begin{equation}\label{Eq:indfunct}
\B{\phi}_{*}:= - \tensor{\scriptscriptstyle{\phi}}B: \rcomod{{\mathcal H}} \longrightarrow \rcomod{{\mathcal K}},
\end{equation}
called \emph{the induction functor}, which is defined on objects by sending any right comodule $(M,\rcoaction{{\mathcal H}}{M})$ to a right comodule $(M\tensor{\scriptscriptstyle{\phi}}B, \rcoaction{{\mathcal K}}{M\tensor{\scriptscriptstyle{\phi}}B})$ with underlying $B$-module $M\tensor{A}B$ and coaction
$$
\rcoaction{{\mathcal K}}{M\tensor{\scriptscriptstyle{\phi}}B}: M\tensor{\scriptscriptstyle{\phi}}B \to (M\tensor{\scriptscriptstyle{\phi}}B)\tensor{B} {\mathcal K}, \quad m\tensor{A}b \mapsto (m_{(0)}\tensor{A}1_{\scriptscriptstyle{B}}) \tensor{B} \phi_{\scriptscriptstyle{1}}(m_{(1)}) \Sf{t}(b).
$$
The image of ${\mathcal H}$ with the induction functor is, in fact, an $({\mathcal H},{\mathcal K})$-bicomodule.
In a similar way, we have the induction functor
$$
{}_{*}\B{\phi}:=B\tensor{\scriptscriptstyle{\phi}}-: \lcomod{{\mathcal H}} \to \lcomod{{\mathcal K}},
$$
between left comodules,
and $B\tensor{\scriptscriptstyle{\phi}}{\mathcal H}$ is now an $({\mathcal K},{\mathcal H})$-bicomodule.
The induction functor has a right adjoint
given by
\begin{equation}
\label{Eq:coind}
-\cotensor{{\mathcal K}}(B\tensor{\scriptscriptstyle{\phi}}{\mathcal H}): \rcomod{{\mathcal K}} \to \rcomod{{\mathcal H}},
\end{equation}
called the \emph{coinduction functor}.
\subsection{Comodule algebras}
\label{ssec:comoalg}
Recall that a \emph{left ${\mathscr G}$-action} of a groupoid ${\mathscr G}$ on a set $N$ consists of two maps $\alpha: N \to {\mathscr G}_{0}$ and ${\mathscr G}_{1}\, \due \times {\scriptscriptstyle{\Sf{s}}} {\, \scriptscriptstyle{\alpha}} N \to N, \ (g,n) \mapsto gn$, satisfying
$$
\alpha(gn)\,=\, \Sf{t}(g),\quad \iota_{\scriptscriptstyle{\alpha(n)}} n\,=\, n,\quad g'(gn)\,=\, (g'g)n.
$$
In this way, one can consider the \emph{left translation groupoid} ${\mathscr G} \lJoin N$ with ${\mathscr G}_{1}\, \due \times {\scriptscriptstyle{\Sf{s}}} {\, \scriptscriptstyle{\alpha}} N$ as set of arrows and $N$ as set of objects. This is the so-called \emph{semi-direct product groupoid}, see \cite[p.~163]{MoeMrc:LGSAC}.
Next, we want to give the analogue notion in the Hopf algebroids context. To this end, recall first that a \emph{left ${\mathcal H}$-comodule algebra} for a Hopf algebroid $(A,{\mathcal H})$ is a commutative monoid in the symmetric monoidal category $\lcomod{{\mathcal H}}$. That is, a pair $(R,\sigma)$ consisting of a commutative $A$-algebra $\sigma: A\to R$ which is also a left ${\mathcal H}$-comodule with coaction $\lcoaction{{\mathcal H}}{R}: R \to {\mathcal H}\tensor{A}R$,
satisfying for all $x, y \in R$
\begin{equation}\label{Eq:comodalg}
\lcoaction{{\mathcal H}}{R}(xy) \,=\, x_{(-1)}y_{(-1)} \tensor{A} x_{(0)}y_{(0)}
\quad \mbox{and} \quad \lcoaction{{\mathcal H}}{R}(1_{\scriptscriptstyle{R}})\,=\, 1_{{\scriptscriptstyle{{\mathcal H}}}}\tensor{A}1_{{\scriptscriptstyle{R}}}.
\end{equation}
In others words, the coaction $\lcoaction{{\mathcal H}}{R}$ is an $A$-algebra map, where ${\mathcal H}\tensor{A}R$ is seen as an $A$-algebra via $A \to {\mathcal H} \otimes_{\scriptscriptstyle{A}} R,\ a \mapsto \Sf{s}(a)\tensor{A} 1_{\scriptscriptstyle{R}}$.
A \emph{morphism of left ${\mathcal H}$-comodule algebras} is an $A$-algebra map which is also a left ${\mathcal H}$-comodule morphism.
{\em Right} ${\mathcal H}$-comodule algebras are analogously defined.
Note that for a left ${\mathcal H}$-comodule algebra $(R,\sigma)$ the $\Bbbk$-vector subspace
$$
R^{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}=\{x \in R \mid \lcoaction{{\mathcal H}}{R}(x)=1_{{\scriptscriptstyle{{\mathcal H}}}}\tensor{A}x \}
$$
of ${\mathcal H}$-coinvariant elements
is a $\Bbbk$-subalgebra of $R$ that does not necessarily contain the image $\sigma(A)$, unless one makes more assumptions; for instance, if the source and the target maps are equal.
A trivial example of a comodule algebra is the base algebra $A$ of a Hopf algebroid $(A, {\mathcal H})$ itself.
Assume now that $\gamma: B \to R$ is another algebra map such that $\lcoaction{{\mathcal H}}{R}$ is right $B$-linear, that is,
$$
\lcoaction{{\mathcal H}}{R}(x\,\gamma(b))\,=\, x_{(-1)}\tensor{A}x_{(0)} \gamma(b),
$$
for every $x \in R$ and $b \in B$. One can easily see that $\gamma(B) \subseteq R^{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}$.
In
this situation, the \emph{canonical map}
\begin{equation}
\label{Eq:canhr}
\can{{\mathcal H}}{R}: R\tensor{B}R \to {\mathcal H}\tensor{A}R, \quad x\tensor{B}y \mapsto x_{(-1)}\tensor{A}x_{(0)} y
\end{equation}
is a $B$-algebra map, where ${\mathcal H}\tensor{A}R$ is a $B$-algebra via $\gamma$ in the second factor.
The canonical map is also left ${\mathcal H}$-colinear, when $R\tensor{B}R$ is seen as a left comodule via the coaction $\lcoaction{{\mathcal H}}{R}\tensor{B}R$.
We have the following well-known properties:
\begin{lemma}
\label{lemma:comdalg}
Assume that $R$ carries a left ${\mathcal H}$-comodule algebra structure with underlying algebra map $\gs: A \to R$ and that $\gamma: B \to R$ is a morphism of algebras.
\begin{enumerate}
\item The pair $(R, {\mathcal H}\tensor{A}R)$ is a Hopf algebroid with the following structure maps:
\begin{equation*}
\begin{array}{rclrcl}
\mathsf{s} &:=& \lcoaction{{\mathcal H}}{R}, & \mathsf{t} &:=& 1_{{\scriptscriptstyle{{\mathcal H}}}}\tensor{A} -, \\
\varepsilon(u \otimes_{\scriptscriptstyle{A}} r) &:=& \varepsilon_{\scriptscriptstyle{{\mathcal H}}}(u)r, & \Delta(u \otimes_{\scriptscriptstyle{A}} r) &:=& (u_{(1)} \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{R}}) \otimes_{\scriptscriptstyle{R}} (u_{(2)} \otimes_{\scriptscriptstyle{A}} r), \\
&&& \mathscr{S} (u \otimes_{\scriptscriptstyle{A}} r) &:=& \mathscr{S}_{\scriptscriptstyle{{\mathcal H}}}(u)r_{(-1)}\tensor{A}r_{(0)}.
\end{array}
\end{equation*}
\item The map
$
(\sigma, -\tensor{A}1_{\scriptscriptstyle{R}}): (A,{\mathcal H}) \to (R,{\mathcal H}\tensor{A}R)
$
is a morphism of Hopf algebroids.
\item If $\lcoaction{{\mathcal H}}{R}$ is right $B$-linear, where $R$ is seen as an $(A,B)$-bimodule, then the canonical map of Eq.~\eqref{Eq:canhr} is a morphism of Hopf algebroids as well as a morphism of left ${\mathcal H}$-comodules.
\item If $R$ is an $({\mathcal H},{\mathcal K})$-bicomodule, then the canonical map
$$
\can{{\mathcal H}}{R}: (R {\otimes^{\scriptscriptstyle{B}}} R,\rcoaction{{\mathcal K}}{R{\otimes^{\scriptscriptstyle{B}}}R}) \to ({\mathcal H}\tensor{A}R,{\mathcal H}\tensor{A}\rcoaction{{\mathcal K}}{R})
$$
is also a morphism of right ${\mathcal K}$-comodules.
\end{enumerate}
\end{lemma}
\begin{proof}
These are routine computations.
\end{proof}
In analogy to groupoid terminology, the Hopf algebroid $(R,{\mathcal H}\tensor{A}R)$ of Lemma \ref{lemma:comdalg} is termed the \emph{ left translation Hopf algebroid of $(A,{\mathcal H})$ along $\sigma$}. Symmetrically, one can define a \emph{right translation Hopf algebroid of $(A,{\mathcal H})$} by employing {\em right} comodule algebras.
\begin{rem}
\label{remark:Orbits}
Recall that for a groupoid ${\mathscr G}$ one can define its set of orbits as follows: for any $x \in {\mathscr G}_{0}$, one considers either the set
$$
{\mathscr O}_{\scriptscriptstyle{x}} \,= \, \Sf{t}\big( \Sf{s}^{-1}(x)\big),
$$ or ${\mathscr O}_{\scriptscriptstyle{x}} = \Sf{s}\big( \Sf{t}^{-1}(x)\big)$.
An equivalence relation on ${\mathscr G}_{0}$ is now defined by setting $x \sim y $ if and only if ${\mathscr O}_{\scriptscriptstyle{x}}={\mathscr O}_{\scriptscriptstyle{y}}$. The \emph{set of orbits} of ${\mathscr G}$ is the quotient set ${\mathscr G}_{0}/\sim$, which is often denoted by ${\mathscr G}_{0}/{\mathscr G}$.
A more general situation arises when a groupoid acts on a set. Specifically,
the \emph{orbit set} of $N$, denoted by $N /{\mathscr G}$, is the orbit set of the left translation groupoid ${\mathscr G} \lJoin N $. An exercise shows that in case ${\mathscr G}$ is an action groupoid, then this set coincides with the classical set of orbits.
For a Hopf algebroid $(A,{\mathcal H})$ and any commutative algebra $C$, one can consider its underlying presheaf of groupoids,
canonically defined by $C \to ({\mathscr H}(C), A(C))=({ \rm Alg}_{\Bbbk}({\mathcal H},C), {\rm Alg}_{\Bbbk}(A,C))$ is the groupoid
$\xymatrix@C=25pt{{\mathcal H}(C)\ar@<0.5ex>@{->}[r] \ar@<-0.5ex>@{->}[r] & \ar@{->}[l] A(C)}$ defined by reversing the structure maps of $(A,{\mathcal H})$.
This leads then to the \emph{orbit presheaf} $C \mapsto {\mathscr O}(C):=A(C)/{\mathscr H}(C)$.
Clearly, there is a morphism ${\mathscr O} \to {\rm Alg}_{\scriptscriptstyle{\Bbbk}}(A^{{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}},-)$ of presheaves,
where $A^{{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}$ is the coinvariant subalgebra of $A$, that is, the set of elements $a \in A$ such that $\Sf{s}(a)=\Sf{t}(a)$. Thus, $A^{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}$ can be thought of as the coordinate ring of the orbit space.
In case of a general left ${\mathcal H}$-comodule algebra $(R,\alpha)$ and for any commutative algebra $C$,
the groupoid ${\mathscr H}(C)$ acts on $R(C)$ via $(g, x) \mapsto gx$ given by the algebra map
$$
gx: R \to C, \quad r \mapsto g(r_{(-1)})x(r_{(0)}).
$$
This determines the presheaf ${\mathscr O}_{\scriptscriptstyle{R}}: C \mapsto R(C)/{\mathscr H}(C)$ of orbits together with a morphism of presheaves ${\mathscr O}_{\scriptscriptstyle{R}} \to {\rm Alg}_{\scriptscriptstyle{\Bbbk}}\big(R^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}, - \big)$. So as before, $R^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}$ is the coordinate ring of the orbit space. On the other hand, one can easily check that $R^{{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} = R^{{\scriptscriptstyle{{\rm coinv}}_{({\mathcal H}\tensor{A}R)}}}$, where $(R,{\mathcal H}\tensor{A}R)$ is the left translation Hopf algebroid as above.
\end{rem}
\subsection{The coinvariant subalgebra for the tensor product of comodule algebras}
\label{ssec:pcomalg}
For any two left ${\mathcal H}$-comodule algebras $(R,\alpha)$ and $(S,\sigma)$, the comodule tensor product
$S{\tensor{A}}R$ is an $A$-algebra by means of the algebra map
$A \to S\tensor{A}R, \ a \mapsto \sigma(a)\tensor{A}1_{\scriptscriptstyle{R}}= 1_{\scriptscriptstyle{R}}\tensor{A} \alpha(a)$.
This algebra clearly admits the structure of a left ${\mathcal H}$-comodule algebra the coinvariant subalgebra of it can be described as follows:
\begin{lemma}
\label{lemma:prod}
For any two left ${\mathcal H}$-comodule algebras $(R,\alpha)$ and $(S,\sigma)$, we have an isomorphism
$$
(S{\tensor{A}}R)^{{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \,\cong\, S^{\scriptscriptstyle{o}}\cotensor{{\mathcal H}}R
$$
of algebras,
where $(S^{\scriptscriptstyle{o}},\sigma)$ is the opposite right ${\mathcal H}$-comodule algebra of $(S,\sigma)$.
\end{lemma}
\begin{proof}
For an element $s \tensor{A}r \in (S{\tensor{A}}R)^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}$, the equality
\begin{equation}
\label{Eq:sr}
1_{\scriptscriptstyle{{\mathcal H}}}\tensor{A}s\tensor{A}r\,\,=\,\, s_{(-1)}r_{(-1)}\tensor{A}s_{(0)}\tensor{A}
r_{(0)}
\end{equation}
holds in ${\mathcal H}\tensor{A}S\tensor{A}R$. Applying $({\rm id}_{\scriptscriptstyle{{\mathcal H}}} \otimes m_{\scriptscriptstyle{{\mathcal H}}} \otimes {\rm id}_{\scriptscriptstyle{R}}) \circ \tau_{12} \circ (\mathscr{S} \otimes {\rm id}_{\scriptscriptstyle{S}} \otimes \lcoaction{R}{{\mathcal H}})$ to both sides, where $\tau_{12}$ denotes the tensor flip and $m_{\scriptscriptstyle{{\mathcal H}}}$ the multiplication in ${\mathcal H}$, we obtain
\begin{equation*}
\begin{split}
s\tensor{A}r_{(-1)}\tensor{A}r_{(0)} &= s_{(0)} \tensor{A}\mathscr{S}(s_{(-1)}r_{{(-2)}})
r_{(-1)}\tensor{A}r_{(0)} \\
&= s_{(0)} \tensor{A}\mathscr{S}\big(s_{(-1)}\big) \Sf{t}\big(\varepsilon(r_{(-1)})\big)\tensor{A}r_{(0)}
\\
&=
s_{(0)} \tensor{A}\mathscr{S}\big(s_{(-1)}\big) \tensor{A}r,
\end{split}
\end{equation*}
which shows that $s\tensor{A}r \in S^{\scriptscriptstyle{o}}\cotensor{{\mathcal H}}R$. The converse is similarly deduced.
\end{proof}
\begin{rem}
\label{rem:ortitsCotensor}
For a left resp.\ right ${\mathscr G}$-set $(M,\alpha)$ and $(N, \beta)$ over a groupoid ${\mathscr G}$ as in Remark \ref{remark:Orbits},
the fibred product $N \times_{\scriptscriptstyle{{\mathscr G}_{0}}} M$
carries a left ${\mathscr G}$-action given by $g (n,m):= (ng^{-1}, gm)$, and one can consider its orbit space ({\em i.e.}, the orbit of the left translation groupoid ${\mathscr G} \lJoin (N \times_{\scriptscriptstyle{{\mathscr G}_{0}}} M)$), denoted by $N\tensor{{\mathscr G}}M$ in \cite[p.\ 166]{MoeMrc:LGSAC}. Taking Remark \ref{remark:Orbits} into account, Lemma \ref{lemma:prod} describes its analogue in the Hopf algebroid context, that is, the cotensor product of (left and right) ${\mathcal H}$-comodule algebras should be thought of as the orbit space of their tensor product.
\end{rem}
\subsection{Bicomodule algebras and two-sided translation Hopf algebroids}
\label{ssec:bicomodalg}
For two groupoids ${\mathscr G}$ and ${\mathscr H}$ and a set $N$ with compatible
left ${\mathscr G}$-action and right ${\mathscr H}$-action, one can consider the so-called \emph{two-sided semi-direct product groupoid} (or \emph{two-sided translation groupoid}), which is usually denoted by ${\mathscr G} \lJoin N \rJoin {\mathscr H}$. In what follows, we give the analogous construction for Hopf algebroids.
For two Hopf algebroids $(A,{\mathcal H})$ and $(B,{\mathcal K})$, consider an $({\mathcal H},{\mathcal K})$-bicomodule $P$ such that $(P,\alpha)$ is a left ${\mathcal H}$-comodule algebra and $(P,\beta)$ is a right ${\mathcal K}$-comodule algebra. We then say that the triple $(P,\alpha,\beta)$ is an \emph{$({\mathcal H},{\mathcal K})$-bicomodule algebra}. A \emph{morphism of $({\mathcal H},{\mathcal K})$-bicomodule algebras} is a map which is simultaneously a morphism of left ${\mathcal H}$-comodule algebras and right ${\mathcal K}$-comodule algebras.
\begin{lemdfn}
\label{lemma:bisemidirect}
Let $(P, \alpha, \beta)$ be an $({\mathcal H},{\mathcal K})$-bicomodule algebra. Then $(P, {\mathcal H}\tensor{A}P\tensor{B}{\mathcal K})$ with tensor product
defined by ${\mathcal H} \lJoin P \rJoin {\mathcal K}:=\shopf{{\mathcal H}}\tensor{A}P\tensor{B} \shopf{{\mathcal K}}$ carries a canonical structure of a flat Hopf algebroid the structure maps of which are given by:
\begin{enumerate}
\item the source and target are given by
\begin{equation}
\label{jugendohnegott}
\Sf{s}(p)\,:=\, 1_{\scriptscriptstyle{{\mathcal H}}} \tensor{A}p\tensor{B}1_{\scriptscriptstyle{{\mathcal K}}}, \quad \Sf{t}(p)\,:=\, \mathscr{S}(p_{(-1)}) \tensor{A}p_{(0)}\tensor{B}p_{(1)};
\end{equation}
\item the comultiplication and counit are as follows:
$$
\Delta(u\tensor{A}p\tensor{B}w) \,:= \, \big( u_{(1)}\tensor{A}p\tensor{B}w_{(1)}\big)\tensor{P}\big( u_{(2)}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)}\big) ,\quad \varepsilon(u\tensor{A}p\tensor{B}w)\,:=\, \alpha\big( \varepsilon(u)\big)p \beta\big( \varepsilon(w)\big);
$$
\item whereas the antipode is defined as:
$$
\mathscr{S}\big( u\tensor{A}p\tensor{B}w\big)\,:=\, \mathscr{S}(u p_{(-1)})\tensor{A}p_{(0)}
\tensor{B}p_{(1)}\mathscr{S}(w).$$
\end{enumerate}
Furthermore, there is a diagram
$$
\xymatrix@R=15pt{ & (P,{\mathcal H} \lJoin P \rJoin {\mathcal K}) & \\ (A,{\mathcal H}) \ar@{->}^-{{\B{\alpha}=(\alpha, \,\alpha_1)}}[ur] & & \ar@{->}_-{{\B{\beta}=(\beta, \,\beta_1)}}[ul] (B,{\mathcal K}) }
$$
of Hopf algebroids,
where $\alpha_{1}$ and $\gb_1$ are the maps $h \mapsto h \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}} \otimes_{\scriptscriptstyle{B}} 1_{\scriptscriptstyle{{\mathcal K}}}$ and
$k \mapsto 1_{\scriptscriptstyle{{\mathcal H}}} \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}} \otimes_{\scriptscriptstyle{B}} k$, respectively.
This Hopf algebroid will be termed \emph{two-sided translation Hopf algebroid}.
\end{lemdfn}
\begin{proof}
The fact that $\Sf{s}: P \to \shopf{{\mathcal H}}\tensor{A}P\tensor{B} \shopf{{\mathcal K}}$ is a flat extension is clear since $\shopf{{\mathcal H}}$ and $\shopf{{\mathcal K}}$ are flat; hence $\shopf{{\mathcal H}}\tensor{A}P\tensor{B} \shopf{{\mathcal K}}$ will give a flat Hopf algebroid over $P$.
Using the source map \rmref{jugendohnegott}, the comultiplication $\Delta$ and the counit $\varepsilon$ are obviously left $P$-linear; the
right $P$-linearity follows from
$$
\varepsilon\big( (u\tensor{A}p'\tensor{B}w) \Sf{t}(p)\big)\,=\, \varepsilon\big(u\mathscr{S}(p_{(-1)})\tensor{A}p'p_{(0)}
\tensor{B}p_{(1)}w\big) \,\overset{\eqref{Eq:penultima}}{=} \, \varepsilon(u\tensor{A}p' \tensor{B}w) p
$$
as well as
\begin{eqnarray*}
\Delta\big( (u\tensor{A}p'\tensor{B}w) \, \Sf{t}(p)\big) &=& \Delta\big(u\mathscr{S}(p_{(-1)})\tensor{A}p'p_{(0)}
\tensor{B}p_{(1)}w\big)
\\ &=& \big( u_{(1)}\mathscr{S}(p_{(-1)})\tensor{A}p'p_{(0)}\tensor{B}w_{(1)}p_{(1)}\big)\tensor{P}\big( u_{(2)}\mathscr{S}(p_{(-2)})\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)}p_{(2)}\big)
\\ &=& (u_{(1)}\tensor{A}p'\tensor{B}w_{(1)}) \, \Sf{t}(p_{(0)}) \tensor{P}\big( u_{(2)}\mathscr{S}(p_{(-1)})\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)}p_{(1)}\big)
\\ &=& \big( u_{(1)}\tensor{A}p'\tensor{B}w_{(1)}\big)\tensor{P}\big( u_{(2)}\mathscr{S}(p_{(-1)})\tensor{A}p_{(0)}\tensor{B}w_{(2)}p_{(1)}\big)
\\ &=& \big( u_{(1)}\tensor{A}p'\tensor{B}w_{(1)}\big)\tensor{P} (u_{(2)}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)}) \, \Sf{t}(p).
\end{eqnarray*}
In order to define a Hopf algebroid, we need these maps to satisfy Eqs.~\eqref{Eq:es}--\eqref{Eq:ultima}, which are either clear from definitions or follow by computations similar to the subsequent one proving \eqref{Eq:penultima}:
we have
\begin{eqnarray*}
\mathscr{S} (u_{(1)}\tensor{A}p\tensor{B}w_{(1)}) ( u_{(2)}\tensor{A} 1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)})
&=& \mathscr{S}(u_{(1)})\mathscr{S}(p_{(-1)}) u_{(2)} \tensor{A}p_{(0)}\tensor{B}p_{(1)}\mathscr{S}(w_{(1)}) w_{(2)}
\\ &=& \Sf{t}(\varepsilon(u))\mathscr{S}(p_{(-1)}) \tensor{A}p_{(0)}\tensor{B}p_{(1)}\Sf{t}(\varepsilon(w))
\\ &=& \mathscr{S}\big( \Sf{s}(\varepsilon(u)) p_{(-1)} \big) \tensor{A}p_{(0)}\tensor{B}p_{(1)} \Sf{t}(b)
\\ &=& \Sf{t}\big( \alpha(\varepsilon(u)) \, p \, \beta(\varepsilon(w))\big)
= \Sf{t}\big(\varepsilon( u\tensor{A}p\tensor{B}w) \big).
\end{eqnarray*}
The last statement is easily checked as well.
\end{proof}
Finally note that
for
a morphism
$\mathfrak{f}: (P,\alpha,\beta) \to (P',\alpha',\beta')$
of
$({\mathcal H},{\mathcal K})$-bicomodule algebras, Lemma \ref{lemma:bisemidirect} leads to a commutative diagram
\begin{equation}\label{Eq:qseramana}
\xymatrix@R=25pt@C=30pt{ & (P,{\mathcal H} \lJoin P \rJoin {\mathcal K}) \ar@{->}|-{{(\fk{f},\, {\mathcal H}\tensor{A}\fk{f}\tensor{B}{\mathcal K})}}[dd] & \\ (A,{\mathcal H}) \ar@{->}|-{{\B{\alpha}}}[ur] \ar@{->}|-{{\B{\alpha '}}}[dr] & & \ar@{->}|-{{\B{\beta}}}[ul] (B,{\mathcal K}) \ar@{->}|-{{\B{\beta '}}}[dl] \\ & (P',{\mathcal H} \lJoin P' \rJoin {\mathcal K}) & }
\end{equation}
of flat Hopf algebroids.
\begin{example}\label{exm:bhb}
Let $(A,{\mathcal H})$ be a Hopf algebroid, $C$ any algebra, and $h: {\mathcal H} \to C$ an algebra morphism. Using $\phiup := h \circ \Sf{s}: A \to C$ and $\psiup := h \circ \Sf{t}: A \to C$, construct the scalar extension Hopf algebroids $(C,{\mathcal H}_{\scriptscriptstyle{\phiup}}:=C\tensor{\phiup}{\mathcal H}\tensor{\phiup}C)$ resp.\ $(C,{\mathcal H}_{\scriptscriptstyle{\psiup}}:=C\tensor{\psiup}{\mathcal H}\tensor{\psiup}C)$,
where we used the notation $\tensor{\phiup}$ resp.\ $\tensor{\psiup}$ to distinguish between the two $A$-module structures on $C$ given by either $\phiup$ or $\psiup$. From \cite[Lemma 6.4]{HovStr:CALEHT} we deduce that $(C,{\mathcal H}_{\scriptscriptstyle{\phiup}}) \cong (C,{\mathcal H}_{\scriptscriptstyle{\psiup}})$ as Hopf algebroids; indeed, this isomorphism is explicitly given by:
$$
C\tensor{\phiup}{\mathcal H}\tensor{\phiup}C \to C\tensor{\psiup}{\mathcal H}\tensor{\psiup}C, \quad c\tensor{\phiup}u\tensor{\phiup}c' \mapsto c\, h(u_{(1)})\tensor{\psiup}u_{(2)}\tensor{\psiup}h\big(\mathscr{S}(u_{(3)})\big)c',
$$
with inverse
$d \tensor{\psiup} v \tensor{\psiup} d' \mapsto d \, h\big(\mathscr{S}(v_{(1)})\big) \tensor{\phiup} v_{(2)} \tensor{\phiup} h(v_{(3)}) d'$.
Now, assume that $C$ is of the form $C := B\tensor{\phi}{\mathcal H}\tensor{\psi}B'$ for some extensions $\xymatrix@C=15pt{B &\ar@{->}_-{\scriptscriptstyle{\phi}}[l] A \ar@{->}^-{\scriptscriptstyle{\psi}}[r] & B'}$ along with the obvious algebra map $h: {\mathcal H} \to C$ as well as $\phiup: A \to C$ and $\psiup:A \to C$. We can consider
$(C,\phiup,\psiup)$ as an $({\mathcal H}_{\scriptscriptstyle{\phi}},{\mathcal H}_{\scriptscriptstyle{\psi}})$-bicomodule algebra in a canonical way; this, in fact, is the bicomodule algebra arising from the cotensor product algebra $P^{\scriptscriptstyle{{\rm co}}}\cotensor{{\mathcal H}}P$ by considering, respectively, $P:={\mathcal H}\tensor{\phi}B$ and $P':={\mathcal H}\tensor{\psi}B'$ as $({\mathcal H},{\mathcal H}_{\scriptscriptstyle{\phi}})$- and $({\mathcal H},{\mathcal H}_{\scriptscriptstyle{\psi}})$-bicomodule algebras with obvious coactions.
Let $(C,{\mathcal H}_{\scriptscriptstyle{\phi}} \lJoin C \rJoin {\mathcal H}_{\scriptscriptstyle{\psi}})$ be the associated two-sided translation Hopf algebroid. Then one can show that there is an isomorphism
$$
(C, {\mathcal H}_{\scriptscriptstyle{\phiup}}) \, \cong\, (C,{\mathcal H}_{\scriptscriptstyle{\phi}} \lJoin C \rJoin {\mathcal H}_{\scriptscriptstyle{\psi}}) \, \cong\, (C, {\mathcal H}_{\scriptscriptstyle{\psiup}})
$$
of Hopf algebroids as can be seen by adapting the proof of Proposition \ref{prop:Cielitomissyoutoo} below.
\end{example}
\section{Principal bibundles in the Hopf algebroid context}
\label{sec:pb}
\subsection{General definitions}
\label{ssec:gpb}
In this section, we will introduce one of the main notions in this article. Similar concepts in the framework of Hopf algebras appeared under the name {\em quantum principal bundle} in \cite{BrzMaj:QGGTOQS, Brz:TMIQPP}
or {\em bi-Galois extension} in \cite{Schau:HBGE, Schau:HGABGE}.
\begin{definition}
\label{def:PB}
A \emph{left principal $({\mathcal H}, {\mathcal K})$-bundle} $(P, \ga, \gb)$
for two Hopf algebroids $(A, {\mathcal H})$ and $(B, {\mathcal K})$
is an $({\mathcal H},{\mathcal K})$-bicomodule algebra as in \S\ref{ssec:comoalg}, that is, $P$ is equipped with a left ${\mathcal H}$-comodule algebra and a right ${\mathcal K}$-comodule algebra structures with respect to the algebra maps $\alpha: A \to P$ resp.\ $\beta: B \to P$ such that
\begin{enumerate}
\item $\beta$ is a faithfully flat extension;
\item the canonical map
$$
\can{{\mathcal H}}{P}: P \otimes_{\scriptscriptstyle{B}} P \to {\mathcal H}\tensor{A}P, \quad p\tensor{B}p' \mapsto p_{(-1)} \tensor{A} p_{(0)} p'
$$
is bijective.
\end{enumerate}
\end{definition}
At times, when the context is clear and hence (we think that) no confusion can arise,
the subscripts in the notation $\mathsf{can}$ of the canonical map are dropped.
Maps between principal bundles are defined as follows:
\begin{definition}
A \emph{morphism of left principal $({\mathcal H}, {\mathcal K})$-bundles} $(P, \ga, \gb)$ and $(P', \ga', \gb')$ is a map $\fk{f}:P \to P'$ that is
a morphism of $({\mathcal H},{\mathcal K})$-bicomodule algebras, {\em i.e.}, simultaneously a morphism of $A$-algebras, $B$-algebras, and a morphism of $({\mathcal H}, {\mathcal K})$-bicomodules.
We will also call such a morphism an \emph{equivariant morphism}.
An \emph{isomorphism of left principal bundles} is a bijective morphism of left principal bundles.
The category of left principal $({\mathcal H},{\mathcal K})$-bundles will be denoted by $\lPB{{\mathcal H}}{{\mathcal K}}$.
\end{definition}
Let us denote the inverse of $\can{{\mathcal H}}{P}$ by a sort of Sweedler type notation,
$$
\can{{\mathcal H}}{P}^{-1}: {\mathcal H}\tensor{A}P \to P \tensor{B} P, \quad u \tensor{A} p \mapsto u_+\tensor{B} u_- p.
$$
where
\begin{equation}
\label{plusminus}
\tauup_{\scriptscriptstyle{P}}:= \can{{\mathcal H}}{P}^{-1}(- \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}}): {\mathcal H} \to P\tensor{B}P, \quad u \mapsto u_+ \tensor{B} u_-
\end{equation}
denotes the \emph{translation map}.
The following lemma summarises the properties of this map and its compatibility with the Hopf algebroid structure:
\begin{lemma}
Let $(P, \alpha,\beta)$ be a left principal $({\mathcal H},{\mathcal K})$-bundle.
One has for all $a, a' \in A$, $u, v \in {\mathcal H}$, and $p \in P$:
\begin{eqnarray}
\label{ceuta}
(uv)_+\tensor{B} (uv)_- &=& u_+ v_+ \tensor{B} v_- u_-, \\
\label{japan}
u_{+(-1)} \tensor{A} u_{+(0)} \tensor{B} u_{-} &=& u_{(1)} \tensor{A} u_{(2)+} \tensor{B} u_{(2)-},
\\
\label{Eq:eps+-}
u_+u_-&=& \alpha(\varepsilon(u)),
\\
\label{Eq:p-+}
p_{(-1)+} \tensor{B} p_{(-1)-} p_{(0)} &=& p \tensor{B} 1_{\scriptscriptstyle{P}},
\\
\label{Eq:u-+}
u_{+(-1)} \tensor{A} u_{+(0)} u_- &=& u \tensor{A} 1_{\scriptscriptstyle{P}},
\\
\label{bolognacentrale}
(\mathsf{s}(a)\mathsf{t}(a'))_+ \otimes_{\scriptscriptstyle{B}} (\mathsf{s}(a)\mathsf{t}(a'))_- &=& \ga(a) \otimes_{\scriptscriptstyle{B}} \ga(a').
\end{eqnarray}
Furthermore,
\begin{eqnarray}
\label{casadiaugusto1}
u_{+(0)} \otimes_{\scriptscriptstyle{B}} u_{-(0)} \otimes_{\scriptscriptstyle{B}} u_{+(1)} u_{-(1)} &=& u_{+} \otimes_{\scriptscriptstyle{B}} u_{-} \otimes_{\scriptscriptstyle{B}} 1_{\mathcal K} \quad \in P \otimes_{\scriptscriptstyle{B}} P \otimes_{\scriptscriptstyle{B}} {\mathcal K}, \\
\label{Eq:Su}
\mathscr{S}(u) \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}} &=& u_{{-(-1)}} \otimes_{\scriptscriptstyle{A}} u_{{-(0)}} u_{{+}}, \\
\label{zucchero}
\mathscr{S}(u)_+ \tensor{B} \mathscr{S}(u)_- &=& u_- \tensor{B} u_+, \\
\label{Eq:S+-}
u_{{(1)+}} \otimes_{\scriptscriptstyle{B}} u_{{(1)-}} \tensor{A} \mathscr{S}(u_{{(2)}}) &=& u_+ \otimes_{\scriptscriptstyle{B}} u_{{-(0)}} \tensor{A} u_{{-(-1)}}.
\end{eqnarray}
\end{lemma}
\begin{proof}
The first six equations are proved along the lines of the proof of \cite[Prop.~3.7]{Schau:DADOQGHA}, where the special case in which $P:={\mathcal H}$ is treated. Eq.~\rmref{casadiaugusto1} is obtained by the fact that the canonical map (and hence its inverse) is a morphism of right ${\mathcal K}$-comodules, as follows from Lemma \ref{lemma:comdalg} {\em (iv)}.
Eq.~\rmref{Eq:Su} is proven as follows: since $P$ is a left ${\mathcal H}$-comodule algebra and the coaction is $A$-linear, one has
\begin{equation*}
\begin{split}
\mathscr{S}(u) \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}} &= \mathscr{S}(u_{(1)})\mathsf{s}(\gve(u_{(2)})) \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{P}}
= \mathscr{S}(u_{(1)})\big(\ga(\gve(u_{(2)}))\big)_{(-1)}
\otimes_{\scriptscriptstyle{A}} \big(\ga(\gve(u_{(2)}))\big)_{(0)} \\
& \!\! \overset{\rmref{Eq:eps+-}}{=}
\mathscr{S}(u_{(1)})(u_{(2)+}u_{(2)-})_{(-1)}
\otimes_{\scriptscriptstyle{A}} (u_{(2)+}u_{(2)-})_{(0)} \overset{\rmref{japan}}{=}
\mathscr{S}(u_{(1)}) u_{(2)}u_{(3)-(-1)}
\otimes_{\scriptscriptstyle{A}} u_{(3)+}u_{(3)-(0)} \\
& \! \! \overset{\rmref{maxxi}}{=}
\mathsf{t}(\gve(u_{(1)})) u_{(2)-(-1)}
\otimes_{\scriptscriptstyle{A}} u_{(2)+}u_{(2)-(0)}
\overset{\rmref{bolognacentrale}}{=}
u_{-(-1)}
\otimes_{\scriptscriptstyle{A}} u_{-(0)} u_+.
\end{split}
\end{equation*}
Eq.~\rmref{zucchero} now follows by simply applying the inverse of the canonical map to both sides, using \rmref{Eq:p-+}.
Finally, Eq.~\rmref{Eq:S+-} is seen by applying \rmref{japan} to the element $\mathscr{S}(u)$, using \rmref{zucchero} and the fact that the antipode is an anti-coring morphism.
\end{proof}
Right principal bundles use the right ${\mathcal K}$-comodule algebra structure of $P$ and the canonical map:
$$
\can{P}{{\mathcal K}}: P \tensor{A} P \to P\tensor{B}{\mathcal K}, \quad p'\tensor{A}p \mapsto p'p_{(0)}\tensor{B}p_{(1)}.
$$
In this way, $P$ is said to be a \emph{right principal $({\mathcal H}, {\mathcal K})$-bundle} if $\alpha$ is a faithfully flat extension and the canonical map $\can{P}{{\mathcal K}}$ is bijective. The triple $(P,\alpha,\beta)$ is said to \emph{principal $({\mathcal H}, {\mathcal K})$-bibundle} provided $P$ is both left and right principal.
Since we will explicitly use principal bibundles, we also need the notation and the properties for the right translation map. The inverse of $\can{P}{{\mathcal K}}$ is denoted by
$$
P \tensor{B} {\mathcal K} \to P \otimes_{\scriptscriptstyle{A}} P, \quad p \otimes_{\scriptscriptstyle{B}} v \mapsto pv^- \otimes_{\scriptscriptstyle{A}} v^+,
$$
which fulfils the relations
\begin{eqnarray}
(vw)^+ \otimes_{\scriptscriptstyle{A}} (vw)^- &=& v^+ w^+ \otimes_{\scriptscriptstyle{A}} w^- v^-, \\
v^- v^+ &=& \gb(\varepsilon(v)), \\
p_{(0)} {p_{(1)}}^- \otimes_{\scriptscriptstyle{A}} {p_{(1)}}^+ &=& 1_{\scriptscriptstyle{P}} \otimes_{\scriptscriptstyle{A}} p, \label{Eq:colosseo} \\
v^- {v^{+}}_{(0)} \otimes_{\scriptscriptstyle{B}} {v^{+}}_{(1)} &=& 1_{\scriptscriptstyle{P}} \otimes_{\scriptscriptstyle{B}} v, \\
v^- \otimes_{\scriptscriptstyle{A}} {v^+}_{(0)} \otimes_{\scriptscriptstyle{A}} {v^+}_{(1)} &=& {v_{(1)}}^- \otimes_{\scriptscriptstyle{A}} {v_{(1)}}^+ \otimes_{\scriptscriptstyle{A}} v_{(2)}, \\
(\mathsf{s}(b)\mathsf{t}(b'))^- \otimes_{\scriptscriptstyle{A}} (\mathsf{s}(b)\mathsf{t}(b'))^+ &=& \gb(b) \otimes_{\scriptscriptstyle{A}} \gb(b').
\end{eqnarray}
With a similar argumentation that lead to \rmref{casadiaugusto1}, we have the identity
\begin{equation}
\label{casadiaugusto2}
{v^-}_{(-1)} {v^+}_{(-1)} \otimes_{\scriptscriptstyle{A}} {v^-}_{(0)} \otimes_{\scriptscriptstyle{A}} {v^+}_{(0)} = 1_{\mathcal H} \otimes_{\scriptscriptstyle{A}} {v^-} \otimes_{\scriptscriptstyle{A}} {v^+} \quad \in {\mathcal H} \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{A}} P.
\end{equation}
Analogously, one obtains
\begin{eqnarray*}
\mathscr{S}(v)^- \otimes_{\scriptscriptstyle{A}} \mathscr{S}(v)^+ &=& v^+ \otimes_{\scriptscriptstyle{A}} v^-, \\
{v_{(2)}}^- \otimes_{\scriptscriptstyle{A}} {v_{(2)}}^+ \otimes_{\scriptscriptstyle{B}} \mathscr{S}(v_{(1)}) &=&
v^- \otimes_{\scriptscriptstyle{A}} {v^+}_{(0)} \otimes_{\scriptscriptstyle{B}} {v^+}_{(1)} \quad \in P \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{B}} {\mathcal K}, \\
1_{\scriptscriptstyle{P}} \otimes_{\scriptscriptstyle{B}} \mathscr{S}(v) &=& v^- {v^+}_{(0)} \otimes_{\scriptscriptstyle{B}} {v^+}_{(1)}.
\end{eqnarray*}
In a similar way, one can define a morphism between right principal $({\mathcal H}, {\mathcal K})$-bundles. The obtained category will be denoted by $\rPB{{\mathcal H}}{{\mathcal K}}$.
Morphisms of principal bibundles are simultaneously morphisms of left and right principal bundles. The category obtained this way will be denoted by $\bPB{{\mathcal H}}{{\mathcal K}}$.
\pagebreak
\begin{rem}
\label{rema:PB} \noindent
\begin{enumerate}
\item For a morphism $\fk{f}:(P, \alpha, \gb) \to (P', \alpha', \gb')$ in $\lPB{{\mathcal H}}{{\mathcal K}}$, we have a commutative diagram:
\begin{equation}\label{Eq:barril}
\xymatrix@R=20pt{ {\mathcal H} \ar@{->}^-{\tauup_P}[rr] \ar@{->}_-{\tauup_{P'}}[rrd] & & P\tensor{B}P \ar@{->}^-{\fk{f}\tensor{B}\fk{f}}[d] \\ & & P'\tensor{B}P', }
\end{equation}
where $\tauup$ is the corresponding translation map.
\item
The definition above is left-right symmetric: if ${}_{{\scriptscriptstyle{{\mathcal H}}}}P_{{\scriptscriptstyle{{\mathcal K}}}}$ is a left principal $({\mathcal H}, {\mathcal K})$-bundle, then the opposite bicomodule ${}_{{\scriptscriptstyle{{\mathcal K}}}}P^{\scriptscriptstyle{{\rm co}}}{}_{{\scriptscriptstyle{{\mathcal H}}}}$ is a right principal $({\mathcal K}, {\mathcal H})$-bundle with respect to the canonical map
$$
P^{\scriptscriptstyle{{\rm co}}} \otimes_{\scriptscriptstyle{B}} P^{\scriptscriptstyle{{\rm co}}} \to P^{\scriptscriptstyle{{\rm co}}} \otimes_{\scriptscriptstyle{A}} {\mathcal H}, \quad p' \otimes_{\scriptscriptstyle{B}} p \mapsto p'p_{(0)} \otimes_{\scriptscriptstyle{A}} \mathscr{S}(p_{(-1)}).
$$
Using \rmref{zucchero}, one immediately verifies that
$$
P^{\scriptscriptstyle{{\rm co}}} \otimes_{\scriptscriptstyle{A}} {\mathcal H} \to P^{\scriptscriptstyle{{\rm co}}} \otimes_{\scriptscriptstyle{B}} P^{\scriptscriptstyle{{\rm co}}}, \quad p \otimes_{\scriptscriptstyle{A}} h \mapsto ph_+ \otimes_{\scriptscriptstyle{B}} h_-
$$
defines the inverse of this map. If we denote by $\alpha^{{\scriptscriptstyle{{\rm co}}}}: A \to P^{\scriptscriptstyle{{\rm co}}}$ and $\beta^{{\scriptscriptstyle{{\rm co}}}}: B \to P^{{\scriptscriptstyle{{\rm co}}}}$, respectively, the corresponding algebra maps, then the correspondence $(P,\alpha,\beta) \to (P^{{\scriptscriptstyle{{\rm co}}}}, \beta^{{\scriptscriptstyle{{\rm co}}}}, \alpha^{{\scriptscriptstyle{{\rm co}}}})$ establishes an isomorphism of categories between $\lPB{{\mathcal H}}{{\mathcal K}}$ and $\rPB{{\mathcal K}}{{\mathcal H}}$. The bundle $(P^{{\scriptscriptstyle{{\rm co}}}}, \beta^{{\scriptscriptstyle{{\rm co}}}}, \alpha^{{\scriptscriptstyle{{\rm co}}}})$ so constructed is called the \emph{opposite bundle} of $(P,\alpha, \beta)$.
\item
Since $P_{\scriptscriptstyle{B}}$ is faithfully flat, we know
by the faithfully flat descent theory (see, for instance \cite[Theorem 3.10]{Kaoutit/Gomez:2003a})
that the subalgebra of ${\mathcal H}$-coinvariants is $P^{{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}= \beta(B)$ as $\beta$ is injective.
Moreover, since $\alpha: A\to P$ is a right ${\mathcal H}$-colinear map, we have the following commutative diagram
$$
\xymatrix{ A^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \ar@{->}^-{\alpha^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}}[rr] \ar@{_{(}->}^-{}[d] & & P^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \cong B \ar@{^{(}->}^-{\beta}[d] \\ A \ar@{->}^-{\alpha}[rr] && P}
$$
of algebras.
On the other hand, the category of relative left comodules, that is, the category of left $({\mathcal H}\tensor{A}P)$-comodule is (monoidally)
equivalent to the category of $B$-modules, where $(P,{\mathcal H}\tensor{A}P)$ is the translation Hopf algebroid along $\alpha$.
Conversely, given an $({\mathcal H},{\mathcal K})$-bicomodule algebra $(P, \alpha, \beta)$ such that the functor $-\tensor{B}P: \rmod{B} \to \rcomod{{\mathcal H}\tensor{A}P}$ establishes an equivalence of categories, $(P, \alpha, \beta)$ carries the structure of a left principal $({\mathcal H}, {\mathcal K})$-bundle.
\item
For the trivial Hopf algebroid $(B,{\mathcal K}):=(B,B)$, a left
principal $({\mathcal H}, B)$-bundle is a left ${\mathcal H}$-comodule algebra $(P,\alpha)$ with a faithfully flat extension $\beta: B \to P$
whose ${\mathcal H}$-coaction is a $B$-linear map and where $\can{{\mathcal H}}{P}: P\tensor{B} P \to {\mathcal H}\tensor{A}P$ is bijective.
\end{enumerate}
\end{rem}
\begin{example}[{\em Unit bundles}]
\label{Exam:H}
The underlying ${\mathcal H}$-bicomodule of any flat Hopf algebroid $(A,{\mathcal H})$ is a left
principal
$({\mathcal H}, {\mathcal H})$-bundle. More precisely, ${\mathcal H}$ is an ${\mathcal H}$-bicomodule via the algebra maps $\Sf{s}, \Sf{t}: A \to {\mathcal H}$ and both ring extensions are faithfully flat by assumption.
So, we only need to check {\em (ii)} in Definition \ref{def:PB}. In this case we have
$$
\can{{\mathcal H}}{{\mathcal H}}: {\mathcal H} \tensor{A} {\mathcal H} \to {\mathcal H}\tensor{A}{\mathcal H}, \quad u\tensor{A}v \mapsto u_{(1)}\tensor{A} u_{(2)}v,
$$
where the domain tensor product is defined by $\thopf{{\mathcal H}}$ in both factors, while the codomain tensor product is the standard one from the coproduct of ${{\mathcal H}}$.
The inverse of $\can{{\mathcal H}}{{\mathcal H}}$ is, as for Hopf algebras,
$$
\can{{\mathcal H}}{{\mathcal H}}^{-1}: {\mathcal H}\tensor{A}{\mathcal H} \to {\mathcal H} \tensor{A} {\mathcal H}, \quad u\tensor{A}v \mapsto u_{(1)}\tensor{A} \mathscr{S}(u_{(2)})v.
$$
This bundle is refereed to as the \emph{unit principal bundle} and will be denoted by $\mathscr{U}({\mathcal H})$. Note that $\mathscr{U}({\mathcal H})$ is both a left and a right
principal $({\mathcal H}, {\mathcal H})$-bundle, and therefore a principal bibundle.
\end{example}
\begin{example}[{\em Induced} or {\em pull-back bundles}]
\label{Exam:Ind}
For a morphism $\boldsymbol{\psi}=(\psi_{\scriptscriptstyle{0}},\psi_{\scriptscriptstyle{1}}): (B,{\mathcal K})\to (C, {\mathcal J})$ of Hopf algebroids and a left principal $({\mathcal H}, {\mathcal K})$-bundle $(P,\alpha,\beta)$,
consider $P\tensor{B}C$ with the obvious algebra extensions $\td{\alpha}: A \to P\tensor{B}C$ and $\td{\beta}: C \to P\tensor{B}C$. It is clear that $\td{\beta}$ is a faithfully flat extension and that $P\tensor{B}C$ is an $({\mathcal H},{\mathcal J})$-bicomodule:
its left coaction is $\lcoaction{{\mathcal H}}{P\tensor{B}C}:=\lcoaction{{\mathcal H}}{P}\tensor{B}C$ and its right coaction is defined by the composition
$$
\xymatrix@C=40pt{\rcoaction{{\mathcal J}}{P\tensor{B}C}: P\tensor{B}C\ar@{->}^-{\scriptscriptstyle{\rcoaction{{\mathcal K}}{P}\tensor{A}B}}[r] & P\tensor{B}{\mathcal K}\tensor{B}C \ar@{->}^-{\scriptscriptstyle{P\tensor{B}\psi_{\scriptscriptstyle{1}}\tensor{B}C}}[r] & P\tensor{B}{\mathcal J}\tensor{B}C \ar@{->}^-{\scriptscriptstyle{P\tensor{B}\xiup_{{\mathcal J}}}}[r] & (P\tensor{B}C)\tensor{C}{\mathcal J},}
$$
where $\xiup_{{\mathcal J}}: {\mathcal J}\tensor{B}C \to C\tensor{C}{\mathcal J}, \ w\tensor{B}c \mapsto 1_C\tensor{C}w\Sf{t}(c)$. Explicitly, one obtains
$$
\rcoaction{{\mathcal J}}{P\tensor{B}C}(p\tensor{B}c)\,=\, (p_{(0)}\tensor{B}1_C)\tensor{C} \psi_{\scriptscriptstyle{1}}(p_{(1)})\Sf{t}(c),
$$
and both coactions are algebra maps. Thus, $P\tensor{B}C$ is both a left ${\mathcal H}$-comodule algebra and a right ${\mathcal J}$-comodule algebra.
The canonical map $\can{{\mathcal H}}{P\tensor{B}C}$ is bijective since, up to canonical isomorphisms, it is of the form $\can{{\mathcal H}}{P}\tensor{B}C$.
Hence, $(P\tensor{B}C, \td{\alpha},\td{\beta})$ is a left
principal
$({\mathcal H}, {\mathcal J})$-bundle, called the \emph{induced bundle} of $P$ or \emph{pull-back bundle} of $P$, and denoted $\psi^*(P)$ or $\psi^*\big((P,\alpha,\beta)\big)$.
Of course, this establishes a functor $\lPB{{\mathcal H}}{{\mathcal K}} \to \lPB{{\mathcal H}}{{\mathcal J}}$.
\end{example}
\begin{example}[{\em Restricted principal bundles}]
\label{exam:ResPB}
For a left
principal $({\mathcal H}, {\mathcal K})$-bundle $(P,\alpha, \beta)$ and an algebra map $\tauup: B \to R$, consider the scalar extension Hopf algebroid $(R,{\mathcal K}_{\scriptscriptstyle{R}}) :=(R,R\tensor{B}{\mathcal K}\tensor{B}R)$, along with the obvious algebra maps
$\alpha_{\scriptscriptstyle{R}}: A \to P \to P_{\scriptscriptstyle{R}}$ and $\beta_{\scriptscriptstyle{R}}: R \to P_{\scriptscriptstyle{R}}$, where $P_{\scriptscriptstyle{R}}:=P\tensor{B}R$.
It is clear that $P_{\scriptscriptstyle{R}}$ admits the structure of an $({\mathcal H},{\mathcal K}_{\scriptscriptstyle{R}})$-bicomodule with coactions, up to natural isomorphisms, defined by $\lcoaction{{\mathcal H}}{P_{\scriptscriptstyle{R}}} := \lcoaction{{\mathcal H}}{P}\tensor{B}R$ and
$\rcoaction{{\mathcal K}_{\scriptscriptstyle{R}}}{P_{\scriptscriptstyle{R}}} := \rcoaction{{\mathcal K}}{P}\tensor{B}R$. These are clearly algebra maps which convert $(P_{\scriptscriptstyle{R}}, \lcoaction{{\mathcal H}}{P_{\scriptscriptstyle{R}}})$ and $(P_{\scriptscriptstyle{R}}, \rcoaction{{\mathcal K}_{\scriptscriptstyle{R}}}{P_{\scriptscriptstyle{R}}})$ into comodule algebras. The canonical maps are, up to natural isomorphism, given by
$$
\can{{\mathcal H}}{P_{\scriptscriptstyle{R}}}\, := \, \can{{\mathcal H}}{P}\tensor{B}R,\qquad
\can{P_{\scriptscriptstyle{R}}}{{\mathcal K}_{\scriptscriptstyle{R}}}\, := \, R\tensor{B}\can{P}{{\mathcal K}}\tensor{B}R.
$$
Obviously, $\beta_{\scriptscriptstyle{R}}$ is a faithfully flat extension, hence $(P_{\scriptscriptstyle{R}},\alpha_{\scriptscriptstyle{R}}, \beta_{\scriptscriptstyle{R}})$ is again a left
principal $({\mathcal H},{\mathcal K}_{\scriptscriptstyle{R}})$-bundle, and we have that $(P_{\scriptscriptstyle{R}})^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \simeq R$. We refer to this construction as the \emph{restricted} principal bundle of $(P,\alpha,\beta)$ with respect to $\tauup$.
Again, this yields a functor
$\lPB{{\mathcal H}}{{\mathcal K}} \to \lPB{{\mathcal H}}{{\mathcal K}_{\scriptscriptstyle{R}}}$.
\end{example}
\begin{rem}
\label{rem:ir} \noindent
\begin{enumerate}
\item
If we assume that $(P, \alpha,\beta)$ in Example \ref{exam:ResPB} is only an $({\mathcal H}, {\mathcal K})$-bicomodule algebra, then it is possible to compute
the coinvariant subalgebra $(P_{\scriptscriptstyle{R}})^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}$ of the restricted $({\mathcal H}, {\mathcal K}_{\scriptscriptstyle{R}})$-bicomodule algebra $(P_{\scriptscriptstyle{R}}, \alpha_{\scriptscriptstyle{R}}, \beta_{\scriptscriptstyle{R}})$ by means of the coinvariant subalgebra $P^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}$ provided that $\tauup$ is a flat extension. One then has the following chain of algebra isomorphisms:
$$
(P_{\scriptscriptstyle{R}})^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \,\,\cong \,\, A\cotensor{{\mathcal H}}( P\tensor{B}R)\, \, \cong \,\, ( A\cotensor{{\mathcal H}}P) \tensor{B}R \,\, \cong \,\, P^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}} \tensor{B}R.
$$
\item
For a left principal $({\mathcal H}, {\mathcal K})$-bundle $(P,\alpha,\beta)$ and a morphism $\B{\psi}=( \psi_{\scriptscriptstyle{0}},\psi_{{1}}): (B,{\mathcal K}) \to (C,{\mathcal J})$
of Hopf algebroids, one can consider the induced left
principal $({\mathcal H},{\mathcal J})$-bundle $\psi^*((P,\alpha,\beta))$ on the one hand, and the restricted left
principal $(C,{\mathcal K}_{\scriptscriptstyle{C}})$-bundle $(P_{\scriptscriptstyle{C}}, \alpha_{\scriptscriptstyle{C}}, \beta_{\scriptscriptstyle{C}})$ on the other hand.
However, using the canonical morphism $\B{\Psi}$ of Hopf algebroids associated to $\B{\psi}$ as defined in Eq.~\eqref{Eq:PHI}, the bundle $(P_{\scriptscriptstyle{C}}, \alpha_{\scriptscriptstyle{C}}, \beta_{\scriptscriptstyle{C}})$ induced by $\B{\Psi}$ coincides with $\psi^*(P)$, {\em i.e.},
$$
\psi^*\big((P,\alpha,\beta)\big)\,\,=\,\, \B{\Psi}^*\big((P_{\scriptscriptstyle{C}}, \alpha_{\scriptscriptstyle{C}}, \beta_{\scriptscriptstyle{C}}) \big).
$$
\end{enumerate}
\end{rem}
\begin{example}[{\em Trivial Bundles}]
\label{exam:HopfMorph}
An example of an induced principal bundle is the following, which although rather basic will reveal important in subsequent sections; \emph{cf}.~ also Example \ref{exm:bhb}. For any morphism $(\phi_{\scriptscriptstyle{0}}, \phi_{\scriptscriptstyle{1}}): (A,{\mathcal H}) \to (B,{\mathcal K})$ of Hopf algebroids, consider
\begin{equation}
\label{klaipeda}
P := {\mathcal H}\tensor{\phi}B := {\mathcal H} \tensor{A} B = {\mathcal H} \tensor{\Bbbk} B/{\rm span}\{\mathsf{t}(a) u \otimes b - u \otimes \phi_{\scriptscriptstyle{0}}(a)b \mid u \in {\mathcal H}, b \in B, a \in A \},
\end{equation}
as a left principal $({\mathcal H}, {\mathcal K})$-bundle by pulling back the unit bundle $\mathscr{U}({\mathcal H})$. More precisely, consider the following algebra maps:
$$
\ga: A \to P, \quad a \mapsto \mathsf{s}(a) \otimes_{\scriptscriptstyle{A}} 1_{\scriptscriptstyle{B}},
\qquad \mbox{and} \qquad
\gb: B \to P, \quad b \mapsto 1_{\scriptscriptstyle{{\mathcal H}}} \otimes_{\scriptscriptstyle{A}} b.
$$
Obviously, $P_{\scriptscriptstyle{B}}$ is a faithfully flat module, that is, $\beta$ is a faithfully flat extension. The algebra $P$ is an $({\mathcal H},{\mathcal K})$-bicomodule with left coaction $\lcoaction{{\mathcal H}}{P}:=\Delta_{{\mathcal H}}\tensor{A}B$ along with the right coaction
$$
\rcoaction{{\mathcal K}}{P}: P \to P\tensor{B}{\mathcal K},
\quad u\tensor{A}b \mapsto (u_{(1)}\tensor{A}1_B)\tensor{B} \phi_{\scriptscriptstyle{1}}(u_{(2)})\Sf{t}(b).
$$
Both left and right coactions are easily seen to be morphisms of algebras.
The canonical map is defined as
$$
\can{{\mathcal H}}{P}: P\tensor{B}P \to {\mathcal H}\tensor{A}P, \quad (u\tensor{A}b)\tensor{B}(v\tensor{A}b') \mapsto u_{(1)}\tensor{A}(u_{(2)}v\tensor{A}bb'),
$$
which by Example \ref{Exam:H} is clearly bijective, and the corresponding translation map reads:
$$
\tauup_{P}: H \to P\tensor{B}P,\qquad u \mapsto (u_{\scriptscriptstyle{(1)}} \tensor{A}1_{\scriptscriptstyle{B}}) \tensor{B} (\mathscr{S}(u_{\scriptscriptstyle{(2)}}) \tensor{A}1_{\scriptscriptstyle{B}}).
$$
The fact that the subalgebra of ${\mathcal H}$-coinvariant elements is isomorphic to $B$, see Remark \ref{rema:PB} {\em (ii)}, can be deduced directly in this case: from the isomorphisms
$$
A\cotensor{{\mathcal H}}({\mathcal H}\tensor{A}B) \cong (A\cotensor{{\mathcal H}}{\mathcal H})\tensor{A}B\cong B
$$
one obtains that $P^{\scriptscriptstyle{\scriptscriptstyle{{\rm coinv}}_{{\mathcal H}}}}\cong A{\kasten{6pt}}_{{\mathcal H}}P \cong B$ via $\beta$.
The second canonical map is in this case given by
\begin{equation}
\label{Eq:canK}
\can{P}{{\mathcal K}}: P\tensor{A}P \to P\tensor{B}{\mathcal K}, \quad (u\tensor{A}b)\tensor{A}(v\tensor{A}b') \mapsto (uv_{(1)}\tensor{A}b)\tensor{B} \phi_{\scriptscriptstyle{1}}(v_{(2)})\Sf{t}(b').
\end{equation}
\end{example}
This example motivates the following definition.
\begin{definition}\label{def:bibundle}
We say that a left principal $({\mathcal H}, {\mathcal K})$-bundle $P$ is \emph{trivial} if it is isomorphic to an induced bundle of the unit bundle $\mathscr{U}({\mathcal H})$ as defined in Example \ref{Exam:H}, {\em i.e.}, if there is an isomorphism
$$
P \,\, \cong\,\, \phi^*(\mathscr{U}({\mathcal H})):={\mathcal H}\tensor{\phi}B
$$
of principal bundles with respect to some Hopf algebroid morphism $\B{\phi}: (A, {\mathcal H}) \to (B, {\mathcal K})$.
\end{definition}
Sufficient and necessary conditions under which a left principal bundle is trivial are given in the subsequent proposition.
\begin{proposition}
\label{prop:indPB}
Let $(P,\alpha,\beta)$ be a left principal $({\mathcal H}, {\mathcal K})$-bundle. The following are equivalent:
\begin{enumerate}
\item $(P,\alpha, \beta)$ is a trivial principal bundle;
\item $\beta$ splits as an algebra map, that is, there is an algebra map $\gamma: P \to B$ such that $\gamma \circ \beta ={\rm id}_{\scriptscriptstyle{B}}$.
\end{enumerate}
\end{proposition}
\begin{proof}
Proving $(i) \Rightarrow (ii)$ is immediate from the definitions.
To prove $(ii) \Rightarrow (i)$, we first need
to construct a Hopf algebroid morphism $(\phi_{\scriptscriptstyle{1}}, \phi_{\scriptscriptstyle{0}}):(A,{\mathcal H}) \to (B, {\mathcal K})$.
Here, the algebra map $\phi_{\scriptscriptstyle{0}}: A \to B$ will be defined as the composition $\phi_{\scriptscriptstyle{0}}= \gamma \circ \alpha$, whereas $\phi_{\scriptscriptstyle{1}}$ is given by
$$
\phi_{\scriptscriptstyle{1}}: {\mathcal H} \to {\mathcal K}, \quad u \mapsto \Sf{s}(\gamma(u_{+(0)})) u_{+(1)} \Sf{t}(\gamma(u_-)),
$$
using the notation in \rmref{plusminus} for the translation map;
a routine computation shows that $\boldsymbol{\phi}=(\phi_{\scriptscriptstyle{0}}, \phi_{\scriptscriptstyle{1}})$ is a morphism of Hopf algebroids, indeed.
Consider then the trivial left principal $({\mathcal H}, {\mathcal K})$-bundle ${\mathcal H}\tensor{\scriptscriptstyle{\phi}}B = {\mathcal H} \tensor{A} B$ as in \rmref{klaipeda}.
Let us check that
$$
f: {\mathcal H}\tensor{A}B \to P, \quad u\tensor{A}b \mapsto u_+ \beta(\gamma(u_-)) \beta(b),
$$
is a bijection
whose inverse will be
$$
g: P \to {\mathcal H}\tensor{A}B, \quad p \mapsto p_{(-1)} \tensor{A} \gamma(p_{(0)}).
$$
For any $p \in P$, we have
\begin{eqnarray*}
f(g(p)) &=& f\big(p_{(-1)} \tensor{A}\gamma(p_{(0)})\big) \\
&=& p_{(-1)+} \beta(\gamma(p_{(-1)-})) \beta(\gamma(p_{(0)}))\\
&=& p_{(-1)+} \beta\big(\gamma(p_{(-1)-} p_{(0)})\big) \\ &\overset{\scriptscriptstyle{\eqref{Eq:p-+}}}{=}& p \beta(\gamma(1_{\scriptscriptstyle{P}})) \,\,=\,\, p.
\end{eqnarray*}
On the other hand, for any $u\tensor{A}b \in {\mathcal H}\tensor{A}B$, one computes
\begin{eqnarray*}
g(f(u\tensor{A}b)) &=& g\big(u_+ \beta\big(\gamma(u_-)\beta(b)\big) \\
&=& u_{+(-1)} \tensor{A} \gamma(u_{+(0)}) \gamma(u_-) b \\
&=& u_{+(-1)} \tensor{A}\gamma\big(u_{+(0)} u_- \big)b \\ &\overset{\scriptscriptstyle{\eqref{Eq:u-+}}}{=}& u \tensor{A} \gamma(1_{\scriptscriptstyle{P}})b \,\,=\,\, u\tensor{A}b.
\end{eqnarray*} Thus, $f$ and $g$ are mutually inverse. It is also clear that $g$ is both an $A$-algebra and $B$-algebra map, as well as an $({\mathcal H},{\mathcal K})$-bicomodule map. Therefore, $g$ is an isomorphism of left principal $({\mathcal H}, {\mathcal K})$-bundles.
\end{proof}
The following lemma is an analogue of the respective statement for Lie groupoids in \cite[p.~165]{MoeMrc:LGSAC}. However, the proof given in this context here is direct and does not rely on local triviality of bundles.
\begin{lemma}
\label{lemma:ISO}
Any morphism between left principal $({\mathcal H}, {\mathcal K})$-bundles is an isomorphism. In particular, the category of left principal bundles $\lPB{{\mathcal H}}{{\mathcal K}}$ is a groupoid.
\end{lemma}
\begin{proof}
Let $\fk{f}:(P,\alpha,\beta) \to (P',\alpha',\beta')$ be a morphism between two left principal $({\mathcal H}, {\mathcal K})$-bundles. By definition both $\beta$ and $\beta'$ are faithfully flat extensions; hence, it suffices to check that either $\fk{f}\tensor{B}P'$ or $P\tensor{B}\fk{f}$ is an isomorphism as $\fk{f}$ is an $A$-algebra and $B$-algebra map. To this end, consider the following chain
$$
\xymatrix@C=30pt{P\tensor{B}P' \ar@{->}^-{\cong}[r] & (P\tensor{B}P)\tensor{P} P' \ar@{->}^-{{\sf{can}}\tensor{P}P'}[r] & ({\mathcal H}\tensor{A}P)\tensor{P}P' \ar@{->}^-{\cong}[r] & {\mathcal H}\tensor{A}P' \ar@{->}^-{\sf{can}^{-1}}[r] & P'\tensor{B}P'}
$$
of isomorphisms,
where we have used the fact that $\can{{\mathcal H}}{P}$ is right $P$-linear, is explicitly given by
\begin{equation*}
p\tensor{B}p' \longmapsto (p\tensor{B}1)\tensor{P}p' \longmapsto p_{(-1)} \tensor{A} p_{(0)} \tensor{P} p' \longmapsto p_{(-1)} \tensor{A} \fk{f}(p_{(0)}) p' \longmapsto p_{(-1) +} \tensor{B} p_{(-1) -} \fk{f}(p_{(0)}) p'
\end{equation*}
which by equation \eqref {Eq:p-+} is exactly the map $p\tensor{B}p' \mapsto \fk{f}(p)\tensor{B}p'$ as $\fk{f}$ is a comodule morphism. Therefore, $\fk{f}\tensor{B}P'$ is an isomorphism and so is $\fk{f}$.
\end{proof}
\subsection{Comments on local triviality of principal bundles}
\label{ssec:ltpb}
In the Lie groupoid context, it is well-known that any left principal bundle is locally trivial \cite[p.~165]{MoeMrc:LGSAC}.
Thus, the study of principal bundles in this context can be done locally.
In the Hopf algebroid framework, the notion of ``local triviality'' is not so clear.
The perhaps right way to treat local triviality in this context might be to consider the site of all affine schemes over $\Spectre{\Bbbk}$ with a certain Grothendieck topology $\tau$, and say that a left principal bundle $(P,\alpha,\beta)$ is locally trivial if there is a $\tau$-cover $\Spectre{B'} \to \Spectre{B}$ such that the pull-back bundle $P\tensor{B}B'$ is a trivial left principal $({\mathcal H},{\mathcal K}_{\scriptscriptstyle{B'}})$-bundle. However, as we will see below, when $\tau$ is the Zariski topology, any locally trivial left principal bundle is also globally trivial.
Also, the local triviality for the \emph{fpqc} (faithfully flat quasi-compact)
topology is tautologically true since for any left principal bundle $(P,\alpha,\beta)$, the map $\beta: B \to P$ is by definition a faithfully flat extension.
Moreover, the naive approach to local triviality by localisation apparently does not yield anything new:
let $(P,\alpha,\beta)$ be a left principal $({\mathcal H},{\mathcal K})$-bundle. Denote by $\mathscr{Y}:=\Spectre{B}$ the underlying topological space of the locally ringed space associated to $B$, and by $\Omega(B)$ its subspace of maximal ideals.
Take a prime ideal $y \in \mathscr{Y}$ and consider the localisation $B_{\scriptscriptstyle{y}}$ at this point (the stalk) with $\tauup_{\scriptscriptstyle{y}}: B \to B_{\scriptscriptstyle{y}}$ as the canonical localisation algebra map. Using the notation $\beta_{\scriptscriptstyle{y}}: B_{\scriptscriptstyle{y}} \to P_{\scriptscriptstyle{y}} := P\tensor{B}B_{\scriptscriptstyle{y}}$ and $\alpha_{\scriptscriptstyle{y}}: A \to P \to P_{\scriptscriptstyle{y}}$, we obtain the restricted left
principal
$({\mathcal H},{\mathcal K}_{\scriptscriptstyle{y}})$-bundle $(P_{\scriptscriptstyle{y}},\alpha_{\scriptscriptstyle{y}},\beta_{\scriptscriptstyle{y}})$ with respect to $\tauup_{\scriptscriptstyle{y}}$ as defined in Example \ref{exam:ResPB}.
In this way, any left principal $({\mathcal H}, {\mathcal K})$-bundle $(P,\alpha, \beta)$ can be restricted to a ``local principal bundle'' $(P_{\scriptscriptstyle{y}},\alpha_{\scriptscriptstyle{y}},\beta_{\scriptscriptstyle{y}})$ for every $y \in \mathscr{Y}$. One can say that $(P, \alpha,\beta)$ is locally trivial if and only if $(P_{\scriptscriptstyle{y}},\alpha_{\scriptscriptstyle{y}},\beta_{\scriptscriptstyle{y}})$ is trivial for every $y \in \mathscr{Y}$. Hence, by Proposition \ref{prop:indPB}, this happens if and only if $\beta_{\scriptscriptstyle{y}}: B_{\scriptscriptstyle{y}} \to P_{\scriptscriptstyle{y}}$ splits as an algebra map for every $y \in \mathscr{Y}$; if and only if $\beta_{\scriptscriptstyle{\fk{m}}}: B_{\scriptscriptstyle{\fk{m}}} \to P_{\scriptscriptstyle{\fk{m}}}$ splits as an algebra map for every $\fk{m} \in \Omega(B)$; if and only if $\beta: B \to P$ splits as an algebra map, see \cite[p.~111{\it f}.]{Bou:AC12}. In this sense, $P$ would be locally trivial if and only if it is globally so.
In a different direction, assume that there exists for any $y \in \mathscr{Y}$
an element $f \notin y$ such that $\beta_{\scriptscriptstyle{f}}: B_{\scriptscriptstyle{f}} \to P_{\scriptscriptstyle{f}}$ splits as an algebra map, which by Proposition \ref{prop:indPB} means that the restricted left principal bundle $(P_{\scriptscriptstyle{f}},\alpha_{\scriptscriptstyle{f}},\beta_{\scriptscriptstyle{f}})$ is trivial on the open neighbourhood $\mathscr{Y}_{\scriptscriptstyle{f}}:=\Spectre{B_{\scriptscriptstyle{f}}}$ of $y$ in $\mathscr{Y}$: there is a section $\sigma_{\scriptscriptstyle{f}}: \mathscr{Y}_{\scriptscriptstyle{f}} \to \Spectre{P_{\scriptscriptstyle{f}}} \to \Spectre{P}$, that is, ${}^a\beta_{\scriptscriptstyle{f}} \circ \sigma_{\scriptscriptstyle{f}} ={\rm id}_{\mathscr{Y}_{\scriptscriptstyle{f}}}$, where ${}^a\beta_{\scriptscriptstyle{f}}: \Spectre{P_{\scriptscriptstyle{f}}} \to \mathscr{Y}_{\scriptscriptstyle{f}}$ is the associate continuous map of $\beta_{\scriptscriptstyle{f}}: B_{\scriptscriptstyle{f}} \to P_{\scriptscriptstyle{f}}$.
Again,
one sees that a left bundle $(P,\alpha,\beta)$ with this assumption is in fact a (globally) trivial bundle. Indeed, take a maximal ideal $\fk{m} \in \Omega(B)$: under the assumptions made, there is an $h \notin \fk{m}$ such that $\beta_{\scriptscriptstyle{h}}: B_{\scriptscriptstyle{h}} \to P_{\scriptscriptstyle{h}}$ splits as an algebra map; write $\sigma_{\scriptscriptstyle{h}}: P_{\scriptscriptstyle{h}} \to B_{\scriptscriptstyle{h}}$ for this splitting. Then one can easily check that $$P_{\scriptscriptstyle{\fk{m}}}=P\tensor{B}B_{\scriptscriptstyle{\fk{m}}} \,\cong\, P\tensor{B}B_{\scriptscriptstyle{h}} \tensor{B_{\scriptscriptstyle{h}}}B_{\scriptscriptstyle{\fk{m}}}\,=\, \xymatrix@C=20pt{P_{\scriptscriptstyle{h}}\tensor{B_{\scriptscriptstyle{h}}} B_{\scriptscriptstyle{\fk{m}}} \ar@{->}^-{\sigma_{\scriptscriptstyle{h}}\tensor{B}B_{\scriptscriptstyle{h}}}[rr]& & B_{\scriptscriptstyle{h}}\tensor{B_{\scriptscriptstyle{h}}}B_{\scriptscriptstyle{\fk{m}}}\, \cong \, B_{\scriptscriptstyle{\fk{m}}} }
$$
is an algebra map which splits $\beta_{\scriptscriptstyle{\fk{m}}}$. Thus, $\beta_{\scriptscriptstyle{\fk{m}}}$ splits for every $\fk{m} \in \Omega(B)$, and so does $\beta$. Therefore, $(P,\alpha,\beta)$ is a trivial bundle.
Now assume that the topology $\tau$ is the Zariski one. Then, for a locally trivial left principal bundle $(P,\alpha,\beta)$ there exists an extension $B \to B'\,:=\, \prod_{\scriptstyle{1\leq i \leq n}} B_{\scriptscriptstyle{f_i}}$ for some set $\{f_i\}_{\scriptstyle{1\leq i \leq n}}$ of elements in $B$ such that $B=\sum_{\scriptstyle{1\leq i \leq n}} Bf_i$ and such that $P\tensor{B}B'$ is a trivial bundle. For any maximal ideal $\fk{m} \in \Omega(B)$, there must be some $f_j \notin \fk{m}$ for which the bundle $(P_{\scriptscriptstyle{f_j}},\alpha_{\scriptscriptstyle{f_j}},\beta_{\scriptscriptstyle{f_j}})$ is trivial. We then conclude, as above, that $(P,\alpha,\beta)$ must be also trivial.
On the other hand, it seems that the local triviality property of a given left principal $({\mathcal H},{\mathcal K})$-bundle $(P,\alpha,\beta)$ is already contained in our condition of faithfully flatness of $\beta$. More specifically, since $\beta$ is a flat extension, $\beta_{\scriptscriptstyle{y}}$ is also a flat extension for every $y \in \mathscr{Y}$. Therefore, also $B_{\scriptscriptstyle{y}} \to P_{\scriptscriptstyle{z}}$ is a flat extension
for every $y \in \mathscr{Y}$ and $ z \in ({}^{\scriptscriptstyle{a}}\beta)^{-1}(y)$,
where ${}^{\scriptscriptstyle{a}}\beta: \Spectre{P} =: \mathscr{X} \to \Spectre{B}=\mathscr{Y}$ is the associated continuous map of $\beta$.
In other words, $\mathscr{Y}$ is flat over $\mathscr{X}$ \cite[p.~254]{Hartshorne}; hence, as mentioned in \cite[Def.\ 1.2]{Pfl:ADTATWQORM}, this appears to be a good substitute for local triviality, see \cite[Sec.\ 3]{Palamodov} for a deeper discussion of this point.
\subsection{Natural comodule transformations}
Let $(P, \ga, \gb)$ be a left principal $({\mathcal H}, {\mathcal K})$-bundle. As mentioned before, one can define a functor
$$
- \cotensor{{\mathcal H}} P: \rcomod{{\mathcal H}} \to \rcomod{{\mathcal K}}
$$
since our Hopf algebroids are all assumed to be flat.
We will give some natural transformations involving this functor, which will be useful in the sequel.
\begin{lemma}
\label{levissima}
One has the following natural transformations:
\begin{enumerate}
\item
for any right ${\mathcal H}$-comodule $M$, the map
\begin{equation}
\label{giotto1}
\zeta_{\scriptscriptstyle{M}}: (M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} P \to M \otimes_{\scriptscriptstyle{A}} P,
\quad (m \cotensor{{\mathcal H}} p) \otimes_{\scriptscriptstyle{B}} p' \mapsto m \otimes_{\scriptscriptstyle{A}} pp'
\end{equation}
is an isomorphism of right ${\mathcal K}$-comodules, where the coaction of the left hand side is the codiagonal one. The inverse of $\zeta_{\scriptscriptstyle{M}}$ is given by
$$
\bar\zeta_{\scriptscriptstyle{M}}: m \otimes_{\scriptscriptstyle{A}} p \mapsto (m_{\scriptscriptstyle{(0)}} \cotensor{{\mathcal H}} m_{\scriptscriptstyle{(1)+}}) \otimes_{\scriptscriptstyle{B}} m_{\scriptscriptstyle{(1)-}}p;
$$
\item
for any right ${\mathcal H}$-comodule $M$, the map
\begin{equation}
\label{giotto2}
\eta_{\scriptscriptstyle{M}}: M \to (M \cotensor{{\mathcal H}} P) \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}},
\quad m \mapsto (m_{\scriptscriptstyle{(0)}} \cotensor{{\mathcal H}} m_{\scriptscriptstyle{(1)+}}) \cotensor{{\mathcal K}} m_{\scriptscriptstyle{(1)-}}
\end{equation}
defines a morphism of right ${\mathcal H}$-comodules.
\end{enumerate}
\end{lemma}
\begin{proof}
To prove {\em (i)}, we proceed as follows:
that \rmref{giotto1} is a morphism of comodules follows from the fact that $P$ is a comodule algebra. Moreover, from \rmref{japan} one deduces that the inverse is well-defined and using the flatness of $P$ over $B$ along with \rmref{Eq:eps+-} and \rmref{Eq:p-+}, one checks that the given maps are mutually inverse:
for example,
$$
\bar\zeta_{\scriptscriptstyle{M}} \circ \zeta_{\scriptscriptstyle{M}}\big((m \cotensor{{\mathcal H}}p) \otimes_{\scriptscriptstyle{B}} p'\big)
= (m_{\scriptscriptstyle{(0)}} \cotensor{{\mathcal H}} m_{\scriptscriptstyle{(1)+}}) \otimes_{\scriptscriptstyle{B}} m_{\scriptscriptstyle{(1)-}}pp'
= (m \cotensor{{\mathcal H}} p_{\scriptscriptstyle{(-1)+}}) \otimes_{\scriptscriptstyle{B}} p_{\scriptscriptstyle{(-1)-}}p_{\scriptscriptstyle{(0)}} p'
\overset{\scriptscriptstyle{\rmref{Eq:p-+}}}{=} (m \cotensor{{\mathcal H}}p) \otimes_{\scriptscriptstyle{B}} p',
$$
where in the second step we used that $m \cotensor{{\mathcal H}}p$ lies in $M \cotensor{{\mathcal H}} P$.
As for {\em (ii)},
since $P$ is flat over $B$, the inclusion $(M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} P \hookrightarrow M \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{B}} P$ is the kernel of the map $M \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{B}} P \to M \otimes_{\scriptscriptstyle{A}} {\mathcal H} \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{B}} {\mathcal K} \otimes_{\scriptscriptstyle{B}} P$ given by
\begin{eqnarray*}
m \otimes_{\scriptscriptstyle{A}} p \otimes_{\scriptscriptstyle{B}} q &\longmapsto& m_{(0)} \otimes_{\scriptscriptstyle{A}} m_{(1)} \otimes_{\scriptscriptstyle{A}} p_{(0)} \otimes_{\scriptscriptstyle{B}} p_{(1)} \otimes_{\scriptscriptstyle{B}} q - m_{(0)} \otimes_{\scriptscriptstyle{A}} m_{(1)} \otimes_{\scriptscriptstyle{A}} p \otimes_{\scriptscriptstyle{B}} \mathscr{S}(q_{(1)}) \otimes_{\scriptscriptstyle{B}} q_{(0)} \\
&& - m \otimes_{\scriptscriptstyle{A}} p_{(-1)} \otimes_{\scriptscriptstyle{A}} p_{(0)} \otimes_{\scriptscriptstyle{B}} p_{(1)} \otimes_{\scriptscriptstyle{B}} q
+ m \otimes_{\scriptscriptstyle{A}} p_{(-1)} \otimes_{\scriptscriptstyle{A}} p_{(0)} \otimes_{\scriptscriptstyle{B}} \mathscr{S}(q_{(1)}) \otimes_{\scriptscriptstyle{B}} q_{(0)}.
\end{eqnarray*}
Composing this map with
$
M \to M \otimes_{\scriptscriptstyle{A}} P \otimes_{\scriptscriptstyle{B}} P, \ m \mapsto m_{(0)} \otimes_{\scriptscriptstyle{A}} m_{(1)+} \otimes_{\scriptscriptstyle{B}} m_{(1)-},
$
and applying \rmref{japan} shows that \rmref{giotto2} is well-defined on the given cotensor products; that it is
also a morphism of comodules follows from \rmref{Eq:S+-}.
\end{proof}
\section{Principal bibundles versus weak equivalences}\label{sec:pbwe}
As recalled in Definition \ref{quellala1},
a morphism $\B{\phi}: (A,{\mathcal H}) \to (B,{\mathcal K})$ of flat Hopf algebroids is said to be a \emph{weak equivalence} if and only if the induced functor $\B{\phi}_{*}: \rcomod{{\mathcal H}} \to \rcomod{{\mathcal K}}$ of Eq.~\eqref{Eq:indfunct} establishes an equivalence of categories (which is, in fact, a monoidal symmetric equivalence).
Let us consider the trivial bundle $P={\mathcal H}\tensor{\scriptscriptstyle{\phi}}B$ associated to a given morphism $\B{\phi}$.
One can easily check that the opposite bundle is $P^{\scriptscriptstyle{{\rm co}}} = B \tensor{\scriptscriptstyle{\phi}} {\mathcal H}$ as defined in Remark \ref{rema:PB} {\em (ii)}. The associated functors are, up to natural isomorphisms,
$$
\B{\phi}_{*} \cong -\cotensor{{\mathcal H}}P\quad \text{ and }\quad {}_{*}\B{\phi} \cong - \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}.
$$
Moreover, as mentioned before, $- \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}$ is a right adjoint to $-\cotensor{{\mathcal H}}P$.
\subsection{The case of trivial principal bibundles}
\label{ssec:tpb-we}
Part of the following proposition was shown in \cite[Theorem 6.2]{HovStr:CALEHT} by using a different approach,
see also \cite[Theorem D \& 5.5]{Hovey:02}. In Theorem \ref{aromanflower} below we give a more general result.
\begin{proposition}
\label{lemma:trivial bibundle}
Let $\boldsymbol{\phi}=(\phi_{\scriptscriptstyle{0}}, \phi_{\scriptscriptstyle{1}}): (A,{\mathcal H}) \to (B,{\mathcal K})$ be a morphism of flat Hopf algebroids, and consider the associated trivial bundle $P={\mathcal H}\tensor{\phi}B$. The following are equivalent:
\begin{enumerate}
\item $P$ is a principal $({\mathcal H}, {\mathcal K})$-bibundle.
\item The canonical morphism
$$
\Phi: B\tensor{A}{\mathcal H}\tensor{A}B \to {\mathcal K}, \quad b\tensor{A}u\tensor{A}b' \mapsto \Sf{s}(b) \phi_{\scriptscriptstyle{1}}(u)\Sf{t}(b')
$$
of Hopf $B$-algebroids
is an isomorphism, and $\alpha$ is a faithfully flat extension.
\item The morphism $\boldsymbol{\phi}$ is a weak equivalence.
\end{enumerate}
\end{proposition}
\begin{proof}
To prove $(i) \Rightarrow (ii)$, we only need to check that $\Phi$ is bijective.
By assumption, $\can{P}{{\mathcal K}}$ is bijective, and denote the translation map here as
$$
\tau: {\mathcal K} \to P\tensor{A}P, \quad k \mapsto (u^k\tensor{\phi}b^k)\tensor{A}(v^k\tensor{\phi}c^k),
$$
which means that for every $k \in {\mathcal K}$
$$
1_{\scriptscriptstyle{P}} \tensor{B}k \, = \, (1_{\scriptscriptstyle{{\mathcal H}}}\tensor{\phi}1_{\scriptscriptstyle{B}})\tensor{A}k\,=\, \big(u^kv^k_{(1)}\tensor{A}b^k\big)\tensor{B}\phi_{\scriptscriptstyle{1}}(v^k_{(2)})\Sf{t}(c^k),
$$
Applying the counit of ${\mathcal H}$ we obtain
$$
k \,=\, \Sf{s}(b^k)\phi_{\scriptscriptstyle{1}}\big(\Sf{s}(\varepsilon(u^k))v^k\big)\Sf{t}(c^k).
$$
Define now the map
$$
\Lambda: {\mathcal K} \to B \tensor{A} {\mathcal H} \tensor{A} B, \quad k \mapsto \phi_{\scriptscriptstyle{0}}(\varepsilon(u^k))b^k \tensor{A}v^k\tensor{A}c^k.
$$
Using the previous equality, we easily get that $\Phi\circ \Lambda = {\rm id}$. In the opposite direction, we have
$$
\Lambda \circ \Phi(b\tensor{A}u\tensor{A}b') \,=\, b\tensor{A}u\tensor{A}b'
$$
since $k=\Sf{s}(b)\phi_{\scriptscriptstyle{1}}(u)\Sf{t}(b')$ is uniquely determined by the equation
$$
1_{\scriptscriptstyle{P}} \tensor{B}k\,=\, \big(1_{\scriptscriptstyle{{\mathcal H}}}\tensor{\phi}b\big) \tensor{B} \phi_{\scriptscriptstyle{1}}(u)\Sf{t}(b').
$$
In order to prove $(ii) \Rightarrow (iii)$, we already know by definition that $\B{\phi}_{*}=-\cotensor{{\mathcal H}}P$ is a symmetric monoidal functor. We need to establish natural isomorphisms
\begin{equation}\label{Eq:natisos}
(-\cotensor{{\mathcal H}}P) \circ(-\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}}) \cong {\rm id}_{\rcomod{{\mathcal K}}}, \quad (-\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}}) \circ (-\cotensor{{\mathcal H}}P) \cong {\rm id}_{\rcomod{{\mathcal H}}}.
\end{equation}
First recall that we have a commutative diagram
$$
\xymatrix@C=50pt{ 0 \ar@{->}^-{}[r] & \ar@{->}^-{\cong}[d] P^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal H}} P \ar@{->}^-{}[r] & P^{\scriptscriptstyle{{\rm co}}} \tensor{A} P \ar@{=}^-{}[d] \\ 0 \ar@{->}^-{}[r] & B\tensor{A}{\mathcal H}\tensor{A}B \ar@{->}^-{B\tensor{A}\Delta\tensor{A}B}[r] & B\tensor{A}{\mathcal H}\tensor{A}{\mathcal H}\tensor{A}B. }
$$
Hence, the canonical injection $ P^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal H}} P \hookrightarrow P^{\scriptscriptstyle{{\rm co}}} \tensor{A} P$ splits in the category of $B$-bimodules.
For a right ${\mathcal K}$-comodule $N$, we then have a chain of isomorphisms
$$
\big(N\cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}} \big) \cotensor{{\mathcal H}}P \,\cong \,\, N\cotensor{{\mathcal K}}\big(P^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal H}}P\big) \,\, \cong \,\, N \cotensor{{\mathcal K}}{\mathcal K} \,\, \cong \,\, N
$$
of right ${\mathcal K}$-comodules,
where we used the fact that $\Phi$ is an isomorphism of ${\mathcal K}$-bicomodules.
Clearly, the resulting isomorphism is natural and this gives the first natural isomorphism in \eqref{Eq:natisos}. To establish the second one, we will use the faithfully flatness of $P_{\scriptscriptstyle{A}}$, that is, of $\alpha$.
For a right ${\mathcal H}$-comodule $M$ define by means of Eq.~\eqref{giotto2}
the following morphism
$$
\theta_{M}: M \to \big( M\cotensor{{\mathcal H}}P\big)\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}}, \quad m \mapsto \big(m_{\scriptscriptstyle{(0)}}\cotensor{{\mathcal H}}(m_{\scriptscriptstyle{(1)}}\tensor{\phi}1_{\scriptscriptstyle{B}}) \big) \cotensor{{\mathcal K}} (1_{\scriptscriptstyle{B}} \tensor{\phi} \mathscr{S}(m_{\scriptscriptstyle{(2)}}))
$$
of right ${\mathcal H}$-comodules.
Using the natural isomorphisms $\zeta$ of \eqref{giotto1}, one can show that $\theta_{M}\tensor{A}P$ is an isomorphism, and hence that $\theta$ is a natural isomorphism. Therefore, $\B{\phi}_{*}$ is an equivalence of categories.
The step $(iii) \Rightarrow (i)$ is seen as follows: by Example \ref{exam:HopfMorph}, $P$ is a left principal $({\mathcal H}, {\mathcal K})$-bundle. To check that $P$ is also a right principal $({\mathcal H}, {\mathcal K})$-bundle, we need to verify that the canonical map $\can{P}{{\mathcal K}}$ of Eq.~\eqref{Eq:canK} is bijective as well as that $\alpha$ is a faithfully flat extension.
Since $\boldsymbol{\phi}_{*}$ is an equivalence of categories, there is a natural isomorphism
$$
-\tensor{A} \boldsymbol{\phi}_{*}({\mathcal H}) \,\cong \, \boldsymbol{\phi}_{*}\circ ( -\tensor{A}{\mathcal H}),
$$
where $-\tensor{A}{\mathcal H}: \rcomod{{\mathcal H}} \to \rcomod{{\mathcal H}}$ is the composition of the forgetful functor with the functor defined as in \eqref{Eq:RM}, and where $P=\boldsymbol{\phi}_{*}({\mathcal H})$ is an $A$-module via the algebra map
$\alpha: A \to P, \ a \mapsto \Sf{s}(a)\tensor{A}1_{\scriptscriptstyle{B}}$. Hence, such a natural isomorphism directly implies
that $\alpha$ is a faithfully flat extension.
Let us then prove that $\can{P}{{\mathcal K}}$ is bijective. Since the counit of the adjunction $\phi_{*} \dashv \big(-\cotensor{{\mathcal K}}\,{}_{*}\phi({\mathcal H})\big)$ is a natural isomorphism (see \S\ref{ssec:bicomod}), we denote by
$$
\B{\xiup}_{\scriptscriptstyle{{\mathcal K}}}: {\mathcal K} \to B\tensor{A}{\mathcal H}\tensor{A}B, \quad k \mapsto b^k\tensor{A}u^k\tensor{A}c^k
$$
its inverse at ${\mathcal K}$, with the help of which we can write
$$
k=\Sf{s}(b^k) \phi_{\scriptscriptstyle{1}}(u^k)\Sf{t}(c^k)
$$
for every $k \in {\mathcal K}$. Define moreover
$$
\Psi: P\tensor{B}{\mathcal K} \cong {\mathcal H}\tensor{A}{\mathcal K} \to P\tensor{A}P, \quad u\tensor{A}k \mapsto \big(u\mathscr{S}(v^k_{(1)})\tensor{\phi}b^k\big)\tensor{A}\big(v^k_{(2)}\tensor{\phi}c^k\big)
$$
and compute
\begin{eqnarray*}
\Psi\circ \can{P}{{\mathcal K}}\big( (u\tensor{A}b)\tensor{A}(v\tensor{A}b')\big)
&=& \Psi\big((uv_{(1)}\tensor{A}b)\tensor{B} \phi_{\scriptscriptstyle{1}}(v_{(2)})\Sf{t}(b')\big)\\
&=& \Psi\big(uv_{(1)}\tensor{A}\mathsf{s}(b) \phi_{\scriptscriptstyle{1}}(v_{(2)})\mathsf{t}(b')\big)\\
&=& \Psi\big(uv_{(1)}\tensor{A}\Phi\big(b \tensor{A}v_{(2)}\tensor{A}b'\big)\big) \\
&=& uv_{(1)}\mathscr{S}(v_{(2)})\tensor{A}b \tensor{A}v_{(3)}\tensor{A}b' \\
&=& \big(u\Sf{s}(\varepsilon(v_{(1)}))\tensor{A}b\big) \tensor{A}\big(v_{(2)}\tensor{A}b'\big) \\
&=& \big(u\tensor{A}b\big) \tensor{A}\big(v\tensor{A}b'\big),
\end{eqnarray*}
which shows that $\Psi \circ \can{P}{{\mathcal K}}={\rm id}$. The opposite direction is verified as follows:
\begin{eqnarray*}
\can{P}{{\mathcal K}}\circ \Psi(u\tensor{A}k) &=& \can{P}{{\mathcal K}}\Big(\big(u\mathscr{S}(v^k_{(1)})\tensor{A}b^k\big)\tensor{A}\big( v^k_{(2)}\tensor{A}c^k\big)\Big)\\
&=& \big(u\mathscr{S}(v^k_{(1)})v^k_{(2)}\tensor{A}b^k\big)\tensor{B}\big(\phi_{\scriptscriptstyle{1}}(v^k_{(3)})\Sf{t}(c^k)\big) \\
&=& \big(u\tensor{A}\phi_{\scriptscriptstyle{0}}(\varepsilon(v^k_{(1)})b^k\big)\tensor{B}\big( \phi_{\scriptscriptstyle{1}}(v^k_{(2)})\Sf{t}(c^k)\big) \\
&=& u\tensor{A}\big(\Sf{s}\big(\phi_{\scriptscriptstyle{0}}(\varepsilon(v^k_{(1)})b^k\big) \phi_{\scriptscriptstyle{1}}(v^k_{(2)})\Sf{t}(c^k)\big) \\
&=& u\tensor{A}\big(\Sf{s}(b^k) \phi_{\scriptscriptstyle{1}}(v^k)\Sf{t}(c^k)\big) \\
& =& u\tensor{A}k,
\end{eqnarray*}
which gives the desired equality.
\end{proof}
\begin{rem}
\label{rem:HS}
The statement that $\alpha$ is a flat extension is equivalent to saying that $B$ is Landweber exact over $(A,{\mathcal H})$ in the sense of \cite[Def.\ 2.1]{HovStr:CALEHT}, see Lemma 2.2 in {\em op.~cit}. This, as mentioned before, implies in particular that $(B,B\tensor{A}{\mathcal H}\tensor{A}B)$ is a flat Hopf algebroid.
\end{rem}
\subsection{The case of general principal bibundles}
\label{ssec:lpb-we}
Let now $(P, \alpha, \beta)$ be an $({\mathcal H},{\mathcal K})$-bicomodule algebra. Consider the two-sided translation Hopf algebroid $(P,{\mathcal H}\lJoin P\rJoin {\mathcal K})$ as in Lemma \ref{lemma:bisemidirect}.
Recall that the tensor product ${\mathcal H}\tensor{A}P\tensor{B}{\mathcal K}$ is defined by using the module structures $\shopf{{\mathcal H}}$, $\due P {\scriptscriptstyle{A}} {\scriptscriptstyle{B}}$, and $\shopf{{\mathcal K}}$, and also that there is a diagram of Hopf algebroids
$\B{\alpha}: (A,{\mathcal H}) \rightarrow (P ,{\mathcal H}\lJoin P\rJoin {\mathcal K}) \leftarrow (B,{\mathcal K}): \B{\beta}$, where $\B{\beta}$ and $\B{\alpha}$ are the maps as in Lemma \ref{lemma:bisemidirect}.
On the other hand, one can consider the extended Hopf algebroids
$(P, P\tensor{A}{\mathcal H}\tensor{A}P)$ and $(P, P\tensor{A}{\mathcal K}\tensor{A}P)$, together with the morphisms of Hopf algebroids:
\begin{small}
\begin{equation}
\label{Eq:MSB}
P\tensor{B}{\mathcal K}\tensor{B}P \to {\mathcal H}\lJoin P\rJoin {\mathcal K},\;\; p'\tensor{B}w\tensor{B}p \mapsto \Sf{s}(p')\beta_{{1}}(w) \Sf{t}(p) = \mathscr{S}\big(p_{{(-1)}} \big) \tensor{A}p_{{(0)}}p'\tensor{B}p_{{(1)}}w,
\end{equation}
\begin{equation}
\label{Eq:LSB}
P\tensor{A}{\mathcal H}\tensor{A}P \to {\mathcal H}\lJoin P\rJoin {\mathcal K},\;\; p'\tensor{A}u\tensor{A}p \mapsto \Sf{s}(p')\alpha_{{1}}(u) \Sf{t}(p) = u\mathscr{S}\big(p_{{(-1)}} \big) \tensor{A}p_{{(0)}}p'\tensor{B}p_{{(1)}},
\end{equation}
\end{small}
where $\sf{s}$ and $\sf{t}$ are the source and the target maps of ${\mathcal H}\lJoin P\rJoin {\mathcal K}$ as given in Lemma \ref{lemma:bisemidirect}.
The following proposition shows that principal bundles lead to weak equivalences.
\begin{prop}
\label{prop:Cielitomissyoutoo}
We have the following implications:
\begin{enumerate}
\item If $(P,\alpha,\beta)$ is a left principal $({\mathcal H},{\mathcal K})$-bundle, then $\B{\beta}$ is a weak equivalence.
\item If $(P,\alpha,\beta)$ is a right principal $({\mathcal H},{\mathcal K})$-bundle, then $\B{\alpha}$ is a weak equivalence.
\item If $(P,\alpha,\beta)$ is a principal $({\mathcal H},{\mathcal K})$-bibundle, then $\B{\beta}$ and $\B{\alpha}$ are weak equivalences. In this case, $(A, {\mathcal H})$ and $(B,{\mathcal K})$ are weakly equivalent, see Definition \ref{quellala1}.
\end{enumerate}
\end{prop}
\begin{proof}
Part {\em (iii)} is clearly derived from {\em (i)} and {\em (ii)}. We only prove {\em (i)} since
{\em (ii)} is obtained {\em mutatis mutandum}.
Using Proposition \ref{lemma:trivial bibundle}, we need
to check that the map $B \to {\mathcal K}\tensor{B}P$ is faithfully flat, which is clear from the assumptions, and that the map in
Eq.~\eqref{Eq:MSB} is bijective. Denote this map by $\tilde{\boldsymbol{\gb}}$
and by
$\tilde{\boldsymbol{\gb}}'$
what is going to be its inverse,
given by
$$
\tilde{\boldsymbol{\gb}}': {\mathcal H}\lJoin P\rJoin {\mathcal K} \to P\tensor{B}{\mathcal K}\tensor{B}P,
\quad
u\tensor{A}p\tensor{B}w \mapsto pu_{{+}} \tensor{B} \mathscr{S}(u_{{-(1)}}) w\tensor{B}u_{{-(0)}}.
$$
We compute from one hand
\begin{eqnarray*}
\td{\B{\beta}} \circ \tilde{\boldsymbol{\gb}}' (u\tensor{A}p\tensor{B}w)
&=& \td{\B{\beta}}( pu_{{+}} \tensor{B} \mathscr{S}(u_{{-(1)}}) w\tensor{B}u_{{-(0)}}) \\
&=& \mathscr{S}(u_{{-(-1)}}) \tensor{A} u_{{-(0)}} u_{{+}}p \tensor{B} u_{{-(1)}} \mathscr{S}(u_{{-(2)}}) w \\
&=& \mathscr{S}(u_{{-(-1)}}) \tensor{A} u_{{-(0)}} u_{{+}}p \tensor{B} w \\
&\overset{\eqref{Eq:Su}}{=}& u\tensor{A}p\tensor{B}w.
\end{eqnarray*}
From the other hand, to check that also $\td{\B{\beta}}' \! \circ \td{\B{\beta}} = {\rm id}$,
we first deduce from Eq.~\rmref{Eq:p-+}
\begin{equation}
\label{waitingagain}
p_{(0)} \otimes_{\scriptscriptstyle{B}} p_{(1)} \otimes_{\scriptscriptstyle{B}} p_{(2)} \otimes_{\scriptscriptstyle{B}} 1_{\scriptscriptstyle{B}} =
p_{(-1)+(0)} \otimes_{\scriptscriptstyle{B}} p_{(-1)+(1)} \otimes_{\scriptscriptstyle{B}} p_{(1)} \otimes_{\scriptscriptstyle{B}} p_{(-1)-} p_{(0)},
\end{equation}
which we use to see that
\begin{eqnarray*}
\td{\B{\beta}}' \! \circ \td{\B{\beta}} ( p'\tensor{B}w\tensor{B}p)
&=& \td{\B{\beta}}'(\mathscr{S}(p_{{(-1)}}) \tensor{B} p_{{(0)}}p' \tensor{B}p_{{(1)}} w) \\
&\overset{\eqref{zucchero}}{=}&
p_{(0)} p_{(-1)-} p' \tensor{B} \mathscr{S}(p_{(-1)+(1)}) p_{(1)} w \tensor{B} p_{(-1)+(0)} \\
&\overset{\eqref{waitingagain}}{=}&
p' \tensor{B} \mathsf{t}(\gve(p_{(1)})) w \tensor{B} p_{(0)} \\
&\overset{\eqref{maxxi}}{=}&
p'\tensor{B}w\tensor{B}p,
\end{eqnarray*}
and this concludes the proof.
\end{proof}
\begin{corollary}
\label{coro:mpb}
Let $\fk{f}:(P,\alpha,\beta) \to (P',\alpha',\beta')$ be a morphism in $\mathsf{PB}^\ell({\mathcal H},{\mathcal K})$. Then the associated morphism
$$
(\fk{f}, {\mathcal H}\tensor{A}\fk{f}\tensor{B}{\mathcal K}): (P,{\mathcal H}\lJoin P\rJoin {\mathcal K}) \to (P',{\mathcal H}\lJoin P'\rJoin {\mathcal K})
$$
between the two-sided translation Hopf algebroids (see \S\ref{ssec:bicomodalg})
is an isomorphism of Hopf algebroids and therefore a weak equivalence.
\end{corollary}
\begin{proof}
This directly follows from Lemma \ref{lemma:ISO}. That this morphism is a weak equivalence can also be deduced from Proposition \ref{prop:Cielitomissyoutoo} {\em (i)} and the commutative diagram \eqref{Eq:qseramana}.
\end{proof}
\begin{rem}
As mentioned in \S\ref{ssec:ltpb}, in the Lie groupoid context it is well-known that any morphism between principal bundles is an isomorphism \cite[p.~165]{MoeMrc:LGSAC}, and hence induces an isomorphism between the associated two-sided translation groupoids.
Corollary \ref{coro:mpb} states an analogous result for the associated two-sided Hopf algebroids attached to flat Hopf algebroids.
As a consequence, any two-stage zigzag of weak equivalences, as described in the isosceles triangle in the Introduction, is unique up to an isomorphism.
\end{rem}
\section{The bicategory of principal bundles as a universal solution}
\label{sec:bicat}
In this section, we introduce the cotensor product of two principal bundles in the Hopf algebroid context, which is the analogue of the tensor product of principal bundles in the framework of Lie groupoids \cite[p.~166]{MoeMrc:LGSAC}, where it is defined as the orbit space of the fibred product of the underlying bundles. In the case of Hopf algebroids, the cotensor product leads to the orbit space (which is the coinvariant subalgebra as mentioned in \S\ref{ssec:pcomalg}) of the tensor product of the underlying comodule algebras.
With this product, principal bundles can be shown to form a bicategory. It turns out that trivial bundles constitute a $2$-functor from the canonical $2$-category of flat Hopf algebroids to this bicategory, which yields a certain universal solution (or a calculus of fractions with respect to weak equivalences).
\vspace{-0,3cm}
\subsection{The cotensor product of principal bundles}
\label{ssec:ppb}
Consider three flat Hopf algebroids $(A, {\mathcal H})$, $(B, {\mathcal K})$, and $(C, {\mathcal J})$, and let
$(P, \alpha, \beta)$ be a left principal $({\mathcal H}, {\mathcal K})$-bundle and $(Q, \sigma, \theta)$ a left
principal $({\mathcal K},{\mathcal J})$-bundle.
Recall from \rmref{Eq:bicotensor} that $P\cotensor{{\mathcal K}}Q$ carries the structure of an $({\mathcal H}, {\mathcal J})$-bicomodule.
Moreover, it is clear from the definition of a comodule algebra
that this is simultaneously an $A$-algebra and $C$-algebra via the following commutative diagram
\begin{equation}\label{Eq:earthquake}
\xymatrix@R=20pt@C=30pt{ A \ar@{->}^-{\alpha}[r] \ar@/_2pc/@{-->}_-{\td{\alpha}}[rrd] & P \ar@{->}^-{- \tensor{B} 1_ {\scriptscriptstyle{Q}}}[rr] & & P\tensor{B}Q \\ & & P\cotensor{{\mathcal K}}Q \ar@{->}^-{}[ru] & \\ & 0 \ar@{->}^-{}[ru] & & Q \ar@{->}_-{1_ {\scriptscriptstyle{P}} \tensor{B} -}[uu] \\ & & & C. \ar@/^2pc/@{-->}^-{\td{\theta}}[luu] \ar@{->}_-{\theta}[u] }
\end{equation}
This structure converts the triple $(P\cotensor{{\mathcal K}}Q,\td{\alpha}, \td{\theta})$ into an $({\mathcal H},{\mathcal J})$-bicomodule algebra. In the subsequent lemma we show that this gives in particular a left principal bundle:
\begin{lemma}
\label{lemma:cotensorpb}\noindent
\begin{enumerate}
\item The correspondence
\begin{eqnarray*}
\mathsf{PB}^\ell({\mathcal H},{\mathcal K}) \times \mathsf{PB}^\ell({\mathcal K},{\mathcal J}) & \longrightarrow & \mathsf{PB}^\ell({\mathcal H},{\mathcal J}), \\
\big( (P, \alpha, \beta), (Q, \sigma, \theta) \big) & \longmapsto & (P\cotensor{{\mathcal K}}Q,\td{\alpha}, \td{\theta}), \\
(F,G) & \longmapsto & F \cotensor{{\mathcal K}} G
\end{eqnarray*}
gives a well-defined functor.
\item
The canonical algebra extension $P\cotensor{{\mathcal K}}Q \hookrightarrow P\tensor{B}Q$ is faithfully flat.
\end{enumerate}
\end{lemma}
\begin{proof}
Part {\em (i)}:
as we have seen before, the obvious algebra map $\theta': C \to Q \to P\tensor{B}Q$ factors through
\begin{equation*}
\label{Eq:triangle}
\xymatrix{ C \ar@{->}^-{\td{\theta}}[r] \ar@{->}_-{{\theta'}}[rd] & P\cotensor{{\mathcal K}}Q \ar@{_{(}->}[d] \\ & P\tensor{B}Q,}
\end{equation*}
and $\theta'$ is a faithfully flat extension since $\beta$ and $\theta$ are so. The faithfully flatness of the map $\td{\theta}: C \to P \cotensor{{\mathcal K}} Q$ is seen as follows: one has a chain of $C$-module isomorphisms
\begin{equation}
\label{Eq:Paciencia}
\xymatrix@C=40pt@R=0pt{(P\cotensor{{\mathcal K}}Q)\tensor{C}Q \ar@{->}^-{\cong}[r] & P\cotensor{{\mathcal K}} (Q\tensor{C}Q) \ar@{->}_-{\cong}^-{P\cotensor{{\mathcal K}}\Sf{can}}[r] &
P\cotensor{{\mathcal K}}({\mathcal K}\tensor{B}Q) \ar@{->}^-{\cong}[r] & P\tensor{B}Q \\
(p\cotensor{{\mathcal K}}q)\tensor{C}q' \ar@{|->}[r] & p\cotensor{{\mathcal K}}(q\tensor{C}q')\ar@{|->}[r] & p\cotensor{{\mathcal K}}(q_{\scriptscriptstyle{(-1)}}\tensor{B}q_{\scriptscriptstyle{(0)}}q') \ar@{|->}[r] & p\tensor{B} qq', }
\end{equation}
hence $ (P \cotensor{{\mathcal K}} Q)\otimes_{\scriptscriptstyle{C}} Q $ is also faithfully flat over $C$, and since by assumption $Q$ is so over $C$, we deduce that
$P \cotensor{{\mathcal K}} Q$ is faithfully flat over $C$.
For better distinction, let us denote the involved translation maps as
$$
\tauup_{\scriptscriptstyle{P}}: {\mathcal H} \to P\tensor{B}P,\quad u \mapsto u_{+}\tensor{B}u_{-},
\qquad \tauup_{\scriptscriptstyle{Q}}: {\mathcal K} \to Q\tensor{C}Q, \quad w \mapsto w_{[+]} \tensor{C} w_{[-]}.
$$
The canonical map that turns the cotensor product into a bundle is given as
$$
\mathsf{can}:
(P\cotensor{{\mathcal K}}Q ) \tensor{C} (P\cotensor{{\mathcal K}}Q) \to {\mathcal H}\tensor{A} (P\cotensor{{\mathcal K}}Q),\quad (p\cotensor{{\mathcal K}}q) \tensor{C}(p'\cotensor{{\mathcal K}}q') \mapsto p_{(-1)}\tensor{A} (p_{(0)}p'\cotensor{{\mathcal K}} qq'),
$$
and what is going to be its inverse is defined by
$$
\tilde{\mathsf{can}}:
{\mathcal H}\tensor{A} ( P\cotensor{{\mathcal K}}Q ) \to (P\cotensor{{\mathcal K}}Q ) \tensor{C} (P\cotensor{{\mathcal K}}Q),
\quad u\tensor{A}(p\cotensor{{\mathcal K}} q) \mapsto
(u_{+(0)} \cotensor{{\mathcal K}} u_{+ (1)[+]} ) \tensor{C} (p u_{-} \cotensor{{\mathcal K}} q u_{+ (1)[-]} ),
$$
which are well-defined maps by the $A$-linearity of the coaction as well as using \rmref{bolognacentrale}.
We then compute
\begin{equation*}
\begin{split}
(\tilde{\mathsf{can}} \circ \mathsf{can})\big( (p\cotensor{{\mathcal K}}q) \tensor{C}(p'\cotensor{{\mathcal K}}q')\big) &= \tilde{\mathsf{can}}\big(p_{(-1)}\tensor{A} (p_{(0)}p'\cotensor{{\mathcal K}} qq' ) \big) \\
&= (p_{(-1) + (0)} \cotensor{{\mathcal K}} p_{(-1) + (1)[+]} ) \tensor{C} (p_{(0)} p_{(-1)-} p' \cotensor{{\mathcal K}} qq' p_{(-1) + (1)[-]} ) \\
&\overset{\eqref{Eq:p-+} }{=}
(p_{(0)}\cotensor{{\mathcal K}} p_{(1)[+]} ) \tensor{C} (p' \cotensor{{\mathcal K}} qq' p_{(1)[-]}) \\
&= (p\cotensor{{\mathcal K}} q_{(-1)[+]} ) \tensor{C} ( p' \cotensor{{\mathcal K}} q' q_{(0)} q_{(-1)[-]} ) \\
&\overset{\eqref{Eq:p-+}}{=} (p\cotensor{{\mathcal K}} q) \otimes_{\scriptscriptstyle{C}} (p' \cotensor{{\mathcal K}} q'), \\
\end{split}
\end{equation*}
where we used the definition of the cotensor product in the fourth step. The opposite verification is left to the reader. To prove part {\em (ii)},
consider the isomorphism of Eq.~\eqref{Eq:Paciencia}. It is clear that this is an isomorphism of left $P\cotensor{{\mathcal K}}Q$-modules; since $Q$ is a faithfully flat $C$-module, $(P\cotensor{{\mathcal K}}Q)\tensor{C}Q \cong P\tensor{B}Q$ is a faithfully flat $P\cotensor{{\mathcal K}}Q$-module as well.
\end{proof}
\begin{rem}
\label{jumpingwindow}
Of course, the construction of the functor
in Lemma \ref{lemma:cotensorpb} can be adapted
{\em mutatis mutandum} for right principal bundles as well as for principal bibundles.
\end{rem}
An example of the cotensor product construction above arises from the following proposition,
\begin{proposition}\label{prop:1out2}
Let $(A,{\mathcal H})$ and $(C_i,{\mathcal J}_i)$, $i=1,2$, be flat Hopf algebroids. Then any diagram of weak equivalences
\begin{small}
$$
\xymatrix@R=15pt{ (C_1,{\mathcal J}_1) & & (C_2,{\mathcal J}_2) \\ & \ar@{->}^-{\B{\theta}_1}[lu] (A,{\mathcal H}) \ar@{->}_-{\B{\theta}_2}[ru] &}
$$
\end{small}
can be completed to the following diagram
\begin{small}
\begin{equation}
\label{kontoauszug}
\xymatrix@R=18pt{ & \big(P_1^{\scriptscriptstyle{{\rm co}}}\cotensor{{\mathcal H}}P_2, {\mathcal J}_1\lJoin \big(P_1^{\scriptscriptstyle{{\rm co}}}\cotensor{{\mathcal H}}P_2\big) \rJoin {\mathcal J}_2\big) & \\ (C_1,{\mathcal J}_1) \ar@{->}^-{\B{\zeta}_1}[ru] & & (C_2,{\mathcal J}_2), \ar@{->}_-{\B{\zeta}_2}[lu] \\ & \ar@{->}^-{\B{\theta}_1}[lu] (A,{\mathcal H}) \ar@{->}_-{\B{\theta}_2}[ru] &}
\end{equation}
\end{small}
of weak equivalences,
where $P_i={\mathcal H}\tensor{\theta_i}C_i$, $i=1,2$, are the respective associated trivial bundles.
\end{proposition}
\begin{proof}
Since $\theta_i$ is a weak equivalence,
$P_i$ is a principal $({\mathcal H},{\mathcal J}_i)$-bibundle by Proposition \ref{lemma:trivial bibundle}.
Therefore, by Lemma \ref{lemma:cotensorpb} (and its right hand side version, see Remark \ref{jumpingwindow}),
the cotensor product $P^{\scriptscriptstyle{{\rm co}}}_1\cotensor{{\mathcal H}}P_2$ is a principal $({\mathcal J}_1,{\mathcal J}_2)$-bibundle as well and the proof is completed using Proposition \ref{prop:Cielitomissyoutoo} {\em (iii)}.
\end{proof}
\begin{example}\label{exm:HS}
A particular situation of Proposition \ref{prop:1out2} is the one considered in Example \ref{exm:bhb}: let $\phi: B \rightarrow A \leftarrow B': \psi$ be a diagram of commutative algebras. Assume that $\alpha:A \to P:={\mathcal H}\tensor{\phi}B$, $a \mapsto \Sf{s}(a)\tensor{A}1_{\scriptscriptstyle{B}}$, and $\alpha':A \to P':={\mathcal H}\tensor{\psi}B'$ are faithfully flat extensions. This, in particular, means that $B$ and $B'$ are Landweber exact.
Consider the algebra $C:=B\tensor{A}{\mathcal H}\tensor{A}B'$ along with the scalar extension Hopf algebroids $(C,{\mathcal H}_{\scriptscriptstyle{\phiup}})$ and $(C,{\mathcal H}_{\scriptscriptstyle{\psiup}})$, where $\phiup: A \rightarrow C \leftarrow A: \psiup$ are the obvious maps constructed from $\phi$ resp.\ $\psi$ as in Example \ref{exm:bhb}. Now
$(B,{\mathcal H}_{\scriptscriptstyle{\phi}}) \overset{\scriptscriptstyle{\B{\alpha}}}{\longleftarrow} (A,{\mathcal H}) \overset{\scriptscriptstyle{\B{\alpha'}}}{\longrightarrow} (B',{\mathcal H}_{\scriptscriptstyle{\psi}})$ is a diagram of weak equivalences by Proposition \ref{lemma:trivial bibundle}. Applying
Proposition \ref{prop:1out2}, we get a diagram
$$
(B,{\mathcal H}_{\scriptscriptstyle{\phi}}) \longrightarrow (C,{\mathcal H}{\scriptscriptstyle{\phiup}}) \,\cong\, (C,{\mathcal H}{\scriptscriptstyle{\phi}} \lJoin C \rJoin {\mathcal H}{\scriptscriptstyle{\psi}}) \,\cong\, (C,{\mathcal H}{\scriptscriptstyle{\phiup}}) \longleftarrow (B',{\mathcal H}_{\scriptscriptstyle{\psi}})\
$$
of weak equivalences, where the middle isomorphisms are as in Example \ref{exm:bhb}.
This, in fact, is part of the proof given in \cite[Theorem 6.5]{HovStr:CALEHT}.
\end{example}
\subsection{The bicategory of principal bundles}
\label{ssec:bicat}
In particular, the constructions in the preceding subsection allow for the main observation in this section:
\begin{prop}
\label{piazzadellorologio}
The data given by
\begin{center}
\begin{enumerate}[\quad \raisebox{1pt}{$\scriptstyle{\bullet}$}]
\item
flat Hopf algebroids (as $0$-cells),
\item
left principal bundles (as $1$-cells),
\item
as well as morphisms of left principal bundles (as $2$-cells)
\end{enumerate}
\end{center}
define a bicategory.
\end{prop}
\begin{proof}
The unit $0$-cells in this bicategory are the unit bundles of the form $\mathscr{U}({\mathcal H})$ as in Example \ref{Exam:H}.
The multiplication of two principal bundles ({\em i.e.}, their cotensor product) and of their morphisms is given as in Lemma \ref{lemma:cotensorpb}.
The associativity of the cotensor product is not obvious in this case as it does not follow directly from the flatness of the involved Hopf algebroids: let $(A, {\mathcal H})$, $(B, {\mathcal K})$, $(C, {\mathcal J})$, and $(D, {\mathcal I})$ be flat Hopf algebroids, as well as $(P, \ga, \gb)$, $(Q, \gs, \theta)$, and $(S, \gamma, \gd)$ be left principal
$({\mathcal H}, {\mathcal K})$-, $({\mathcal K}, {\mathcal J})$-, resp.\ $({\mathcal J}, {\mathcal I})$-bundles. First of all,
we have the following diagram
\begin{small}
$$
\xymatrix@R=10pt{ P\cotensor{{\mathcal K}} (Q \cotensor{{\mathcal J}} S) \ar@{^{(}->}^-{}[r] & P\tensor{B} (Q \cotensor{{\mathcal J}} S) \; \ar@{^{(}->}^-{}[rd] & \\ & & P\tensor{B}Q\tensor{C}S, \\
(P \cotensor{{\mathcal K}} Q) \cotensor{{\mathcal J}} S \ar@{^{(}->}^-{}[r] & (P \cotensor{{\mathcal K}} Q) \tensor{C} S\; \ar@{^{(}->}^-{}[ru] &
}
$$
\end{small}
where the upper injections result from definitions and the flatness of $P$ over $B$. The second map of the lower injections
follows from the fact that, as in Lemma \ref{lemma:cotensorpb} {\em (ii)}, the injection $P \cotensor{{\mathcal K}} Q \hookrightarrow P \otimes_{\scriptscriptstyle{B}} Q$ is faithfully flat. Using the universal property of kernels, we deduce the desired natural isomorphism
$$
(P \cotensor{{\mathcal K}} Q) \cotensor{{\mathcal J}} S \overset{\simeq}{\lra} P \cotensor{{\mathcal K}} (Q \cotensor{{\mathcal J}} S).
$$
The remaining axioms to be verified in a bicategory are left to the reader.
\end{proof}
We denote this bicategory by ${\mathsf{PB}}^\ell$ and refer to it as the \emph{bicategory of (left) principal bundles}. The category of $1$- and $2$-cells from $(A, {\mathcal H})$ to $(B, {\mathcal K})$ then is the category $\lPB{{\mathcal H}}{{\mathcal K}}$, see \S\ref{ssec:gpb}.
Similarly, we can introduce the \emph{bicategory of right principal bundles} $\mathsf{PB}^{\scriptscriptstyle{r}}$ and also the \emph{bicategory of principal bibundles} $\mathsf{PB}^{\scriptscriptstyle{b}}$
as mentioned in Remark \ref{jumpingwindow}. On the other hand,
by Remark \ref{rema:PB} {\em (ii)} there is an isomorphism
${\mathsf{PB}}^\ell \cong ({\mathsf{PB}}^{\scriptscriptstyle{r}})^\op$ of bicategories, using B\'enabou's terminology \cite[\S3]{Benabou}: for a bicategory $\mathscr{B}$, denote by $\mathscr{B}^\op$ its \emph{transpose bicategory}, obtained from $\mathscr{B}$ by reversing $1$-cells. On the other hand, its \emph{conjugate bicategory} $\mathscr{B}^{\scriptscriptstyle{{\rm co}}}$ is obtained by reversing $2$-cells. We will call a morphism between two bicategories in the sense of \cite[\S4]{Benabou} a \emph{$2$-functor}.
\subsection{Invertible $1$-cells}\label{ssec:1cells}
Recall that an \emph{internal equivalence} between two $0$-cells $(A,{\mathcal H})$ and $(B,{\mathcal K})$ in ${\mathsf{PB}}^\ell$ is given by two $1$-cells $(P,\alpha,\beta)$ and $(Q,\sigma,\theta)$ in $\lPB{{\mathcal H}}{{\mathcal K}}$ resp.\ $\lPB{{\mathcal K}}{{\mathcal H}}$, such that
$$
P\cotensor{{\mathcal K}}Q \cong \mathscr{U}({\mathcal H}),\quad Q\cotensor{{\mathcal H}}P \cong \mathscr{U}({\mathcal K}),
$$
holds
as $1$-cells, respectively, in $\bPB{{\mathcal H}}{{\mathcal H}}$ and $\bPB{{\mathcal K}}{{\mathcal K}}$.
Here we are implicitly assuming the triangle property, that is,
we assume the following diagrams
\begin{small}
\begin{equation}\label{Eq:trianglI}
\xymatrix@R=13pt{ & Q \cotensor{{\mathcal H}} {\mathcal H} \ar@{->}^-{\cong}[r] & Q\cotensor{{\mathcal H}}\big(P\cotensor{{\mathcal K}}Q\big)\;\; \ar@{^{(}->}^-{}[rd] \ar@{->}^-{\cong}[dd] & \\ Q \ar@{->}^-{\cong}[ru] \ar@{->}_-{\cong}[rd] & & & Q\tensor{A}P\tensor{B}Q \\ & {\mathcal K} \cotensor{{\mathcal K}} Q \ar@{->}^-{\cong}[r] & \big(Q\cotensor{{\mathcal H}}P\big)\cotensor{{\mathcal K}}Q \;\; \ar@{^{(}->}^-{}[ru] & }
\end{equation}
\end{small}
and
\begin{small}
\begin{equation}\label{Eq:trianglII}
\xymatrix@R=13pt{ & P \cotensor{{\mathcal K}} {\mathcal K} \ar@{->}^-{\cong}[r] & P\cotensor{{\mathcal K}}\big(Q\cotensor{{\mathcal H}}P\big) \;\; \ar@{^{(}->}^-{}[rd] \ar@{->}^-{\cong}[dd] & \\ P \ar@{->}^-{\cong}[ru] \ar@{->}_-{\cong}[rd] & & & P\tensor{B}Q\tensor{A}P \\ & {\mathcal H} \cotensor{{\mathcal H}} P \ar@{->}^-{\cong}[r] & \big(P\cotensor{{\mathcal K}}Q\big)\cotensor{{\mathcal H}}P \;\; \ar@{^{(}->}^-{}[ru] & }
\end{equation}
\end{small}
to be commutative. In this case, we also say that $(A, {\mathcal H})$ and $(B, {\mathcal K})$ are \emph{internally equivalent in $\mathsf{PB}^\ell$}. Internal equivalences are, up to $2$-isomorphisms, uniquely determined. More precisely, given a $1$-cell $P$ in $\mathsf{PB}^\ell$, if we assume that there exists $Q$ and $Q'$ in $\mathsf{PB}^\ell$ such that
$$
Q\cotensor{{\mathcal H}}P \cong \mathscr{U}({\mathcal K}),\quad P\cotensor{{\mathcal K}}Q \cong \mathscr{U}({\mathcal H}),
$$ and
$$
Q'\cotensor{{\mathcal H}}P \cong \mathscr{U}({\mathcal K}),\quad P\cotensor{{\mathcal K}}Q' \cong \mathscr{U}({\mathcal H}),
$$ then we have $Q\cong Q'$ as $1$-cells. As in the general case, this is an easy consequence of the associativity of the cotensor product in $\mathsf{PB}^\ell$. Such a $P$ is called an \emph{invertible left principal bundle}.
Examples of invertible left principal bundles are typically obtained by bibundles:
\begin{prop}
\label{fororomano}
Let $(P,\alpha, \beta)$ be a left principal $({\mathcal H}, {\mathcal K})$-bundle and let $(Q, \sigma, \gamma)$ be a right principal $({\mathcal K}, {\mathcal H})$-bundle.
\begin{enumerate}
\item
The translation map
$\tauup: {\mathcal H} \to P\tensor{B}P$
factors through the map
$$
\tauup': {\mathcal H} \to P\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}}.
$$
Analogously, the translation map
$\nuup: {\mathcal K} \to Q\tensor{A}Q$ factors through
$$
\nuup': {\mathcal K} \to Q \cotensor{{\mathcal H}} Q^{\scriptscriptstyle{{\rm co}}}.
$$
\item
Assume moreover that $(P,\alpha, \beta)$ is a
principal $({\mathcal H},{\mathcal K})$-bibundle. Then $(P^{\scriptscriptstyle{{\rm co}}}, \beta,\alpha)$ is a principal $({\mathcal K}, {\mathcal H})$-bibundle and the translation maps induce isomorphisms
$$
\mathscr{U}({\mathcal H}) \overset{\simeq}{\longrightarrow} P\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}},
\quad \mathscr{U}({\mathcal K}) \overset{\simeq}{\longrightarrow} P^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal H}} P
$$
of
principal $({\mathcal H}, {\mathcal H})$-bibundles resp.\ of principal $({\mathcal K}, {\mathcal K})$-bibundles. Furthermore, $(P,\alpha, \beta)$ is an invertible $1$-cell in $\mathsf{PB}^\ell({\mathcal H},{\mathcal K})$.
\end{enumerate}
\end{prop}
\begin{proof}
Part {\em (i)}: to show that the image of the map $\tauup: u \mapsto u_+ \otimes_{\scriptscriptstyle{B}} u_-$ lands for every $u \in {\mathcal H}$ in the cotensor product $P\cotensor{{\mathcal K}}P^{\scriptscriptstyle{{\rm co}}}$, we need to show that
$$
u_{+(0)} \otimes_{\scriptscriptstyle{B}} u_{+(1)} \otimes_{\scriptscriptstyle{B}} u_- = u_{+} \otimes_{\scriptscriptstyle{B}} \mathscr{S}(u_{-(1)}) \otimes_{\scriptscriptstyle{B}} u_{-(0)} \quad \in P \otimes_{\scriptscriptstyle{B}} {\mathcal K} \otimes_{\scriptscriptstyle{B}} P,
$$
where we used the coopposite comodule structure given in \rmref{Eq:leftright}. This is done by applying the map
$$
P \otimes_{\scriptscriptstyle{B}} P \otimes_{\scriptscriptstyle{B}} {\mathcal K} \to P \otimes_{\scriptscriptstyle{B}} {\mathcal K} \otimes_{\scriptscriptstyle{B}} P, \quad p' \otimes_{\scriptscriptstyle{B}} p \otimes_{\scriptscriptstyle{B}} w \mapsto p' \otimes_{\scriptscriptstyle{B}} w\mathscr{S}(p_{(1)}) \otimes_{\scriptscriptstyle{B}} p_{(0)}
$$
to both sides of Eq.~\rmref{casadiaugusto1}. The situation for right bundles is proven {\em mutatis mutandum}.
Part {\em (ii)}: by Lemma \ref{lemma:cotensorpb} the cotensor product carries the structure of a principal bundle. It is furthermore clear that $\tauup'$ is compatible with the source and target maps of ${\mathcal H}$. The fact that $\tauup'$ is left ${\mathcal H}$-colinear follows directly from \rmref{japan}. To show that this map is also right ${\mathcal H}$-colinear one uses \rmref{Eq:leftright} along with \rmref{Eq:S+-}.
To prove that $\tauup'$ is an isomorphism then follows from Lemma \ref{lemma:ISO} as it is, by Eq.~\rmref{ceuta}, a morphism of principal (bi)bundles.
To check the last statement, one only needs to show the triangle property \eqref{Eq:trianglI} (notice that here there is, in fact, only one diagram). Using the notation of \S\ref{ssec:gpb}, the commutativity of \eqref{Eq:trianglI} reads in this case:
$$
p_{(0)} \tensor{B} p_{(1)}{}^{-} \tensor{A} p_{(1)}{}^{+} \, = \,
p_{(-1) +} \tensor{B} p_{(-1) -} \tensor{A} p_{(0)} \quad \in P \tensor{B} P \tensor{A} P,
$$
for every $p \in P$. To verify this, one first applies the map $P\tensor{B}\can{P}{{\mathcal K}}^{-1}$ to both terms and then uses Eq.~\eqref{Eq:p-+} in order to obtain the same element $p_{(0)} \tensor{B} 1_{\scriptscriptstyle{P}} \tensor{B} p_{(1)}$ in $P\tensor{B}P\tensor{B}{\mathcal K}$.
\end{proof}
\begin{proposition}
\label{prop:el parte}\noindent
\begin{enumerate}
\item
Let $(P, \alpha, \beta)$ be a left principal $({\mathcal H},{\mathcal K})$-bundle. Assume moreover that $(P, \alpha, \beta)$ is an invertible $1$-cell in ${\mathsf{PB}}^\ell$ with inverse $(Q, \theta,\sigma) \in {\mathsf{PB}}^\ell({\mathcal K},{\mathcal H})$. Then $(P, \alpha, \beta)$ is a principal $({\mathcal H},{\mathcal K})$-bibundle and $(Q, \theta, \sigma)$ is a principal $({\mathcal K},{\mathcal H})$-bibundle. Furthermore, we have an isomorphism
$$
Q \cong P^{\scriptscriptstyle{{\rm co}}}
$$
of principal bundles.
\item
Let $\B{\phi}: (A, {\mathcal H}) \to (B, {\mathcal K})$ be a morphism of flat Hopf algebroids. Then $\B{\phi}$ is a weak equivalence if and only if the trivial bundle $P={\mathcal H} \otimes_{\scriptscriptstyle{\phi}} B$ is an invertible $1$-cell in $\lPB{{\mathcal H}}{{\mathcal K}}$.
\end{enumerate}
\end{proposition}
\begin{proof}
For better orientation, we recall here that the algebra diagrams defining $P$ and $Q$ are
$$
\xymatrix{A\ar@{->}^-{\alpha}[r] & P & B \ar@{->}_-{\beta}[l]}, \qquad \xymatrix{A\ar@{->}^-{\sigma}[r] & Q & B, \ar@{->}_-{\theta}[l]}
$$
where $\beta$ and $\sigma$ are faithfully flat, and also that the canonical maps $\can{{\mathcal H}}{P}$ and $\can{{\mathcal K}}{Q}$ are bijective.
Part {\em (i)}: by assumption, we have the following $2$-isomorphisms
$$
\chi: {\mathcal H} \overset{\simeq}{\lra} P \cotensor{{\mathcal K}} Q, \quad
u \longmapsto p^{\scriptscriptstyle{u}} \cotensor{{\mathcal K}} q^{\scriptscriptstyle{u}}, \qquad \mbox{and} \qquad
\zeta: {\mathcal K} \overset{\simeq}{\lra} Q \cotensor{{\mathcal H}} P, \quad
w \longmapsto q^{\scriptscriptstyle{w}} \cotensor{{\mathcal H}} p^{\scriptscriptstyle{w}},
$$
where $\chi$ is, in particular, a morphism of ${\mathcal H}$-bicomodules and $\zeta$ is a morphism of ${\mathcal K}$-bicomodules.
The triangle properties then say that we have, up to a canonical isomorphism,
\begin{equation}
\label{Eq:chizeta}
\begin{array}{rcll}
\chi(p_{{(-1)}}) \cotensor{{\mathcal H}} p_{(0)} &=& p_{(0)} \cotensor{{\mathcal K}} \zeta(p_{{(1)}}) & \in P\tensor{B}Q\tensor{A}P, \\
\zeta(q_{{(-1)}}) \cotensor{{\mathcal K}} q_{(0)} &=& q_{(0)} \cotensor{{\mathcal H}} \chi(q_{(1)}) & \in Q\tensor{B}P\tensor{A}Q,
\end{array}
\end{equation}
for all $p \in P$, $q \in Q$.
On the other hand, we also have an isomorphism
\begin{equation}\label{Eq:Gamma}
\xymatrix@C=35pt{ P\tensor{B}Q \ar@{->}^-{\cong}[r] & (P\cotensor{{\mathcal K}}Q)\tensor{A}Q \ar@{->}^-{\chi^{-1}\tensor{A}Q}[r] &
{\mathcal H}\tensor{A}Q}
\end{equation}
of $({\mathcal H},{\mathcal K})$-bicomodules,
where the first isomorphism is the natural transformation of Eq.~\eqref{giotto1}.
Using this isomorphism, we can easily check that $\alpha$ is a faithful extension.
Indeed, take a morphism $f$ such that $f\tensor{A}P=0$;
then $f\tensor{A}{\mathcal H} \tensor{A} Q = 0$ which yields $f=0$ since $\due {\mathcal H} {\scriptscriptstyle{A}} {}$ and $\due Q {\scriptscriptstyle{A}} {}$ are faithfully flat.
Now for a monomorphism $i: X \to X'$ of $A$-modules, we obtain, using again the isomorphism \eqref{Eq:Gamma}, that $\ker(i\tensor{A}P)\tensor{B}Q = 0$, which by the bijectivity of the canonical map $\can{{\mathcal K}}{Q}$ implies that $\ker(i\tensor{A}P)=0$
since $\due {\mathcal K} {\scriptscriptstyle{B}} {}$ and $\due Q {\scriptscriptstyle{A}} {}$ are faithfully flat.
This shows that $\alpha$ is a faithfully flat extension.
We still need to check that the canonical map $\mathsf{can} : P\tensor{A}P \to P\tensor{B}{\mathcal K}$ is bijective. To this end, we define what is going to be its inverse as
$$
\td{\mathsf{can}}: P\tensor{B}{\mathcal K} \to P\tensor{A}P, \quad
p\tensor{B}w \mapsto p \mathsf{g}(q^{\scriptscriptstyle{w}})\tensor{A}p^{\scriptscriptstyle{w}},
$$
where $\mathsf{g}$ is simultaneously the $A$-algebra and $B$-algebra map given explicitly by
$$
\mathsf{g}:Q \to P, \quad q \mapsto \beta \big(\varepsilon\big( \zeta^{-1}(q_{{(0)}}\cotensor{{\mathcal H}}q_{(1)+})\big)\big) q_{(1)-}.
$$
This map satisfies
\begin{equation}
\label{Eq:G}
p^{{u}} \mathsf{g}(q^{{u}}) = \alpha\big(\varepsilon(u) \big), \qquad
\mathsf{g}(q^{\scriptscriptstyle{w}}) p^{\scriptscriptstyle{w}} = \beta\big(\varepsilon(w) \big),
\end{equation}
for every $u \in {\mathcal H}, w \in {\mathcal K}$, which is seen as follows:
as for the second one, we have for $w \in {\mathcal K}$
\begin{eqnarray*}
\mathsf{g}(q^{\scriptscriptstyle{w}}) p^{\scriptscriptstyle{w}} &=& \beta \big(\varepsilon\big( \zeta^{-1}(q^{\scriptscriptstyle{w}}{}_{{(0)}}\cotensor{{\mathcal H}}q^{\scriptscriptstyle{w}}{}_{(1)+})\big) \big) q^{\scriptscriptstyle{w}}{}_{(1)-} p^{\scriptscriptstyle{w}} \\
&{=}& \beta\big(\varepsilon\big( \zeta^{-1}(q^{\scriptscriptstyle{w}}\cotensor{{\mathcal H}}p^{\scriptscriptstyle{w}}{}_{(-1)+})\big) \big) p^{\scriptscriptstyle{w}}{}_{(-1)-} p^{\scriptscriptstyle{w}}{}_{(0)} \\
&\overset{\eqref{Eq:u-+}}{=}& \beta\big(\varepsilon\big( \zeta^{-1}(q^{\scriptscriptstyle{w}}\cotensor{{\mathcal H}}p^{\scriptscriptstyle{w}})\big) \big) \\
&=& \beta\big(\varepsilon(w) \big).
\end{eqnarray*}
As for the first equation in \eqref{Eq:G}, by the right ${\mathcal H}$-colinearity of $\chi$ and Eq.~\eqref{japan}
$$
p^{\scriptscriptstyle{u}} \tensor{B} (q^{\scriptscriptstyle{u}}{}_{(0)}\cotensor{{\mathcal H}} q^{\scriptscriptstyle{u}}{}_{(1)+})
\tensor{B}q^{\scriptscriptstyle{u}}{}_{(1)-}
\,=\, p^{\scriptscriptstyle{u_{+ (-1)}}} \tensor{B} (q^{\scriptscriptstyle{u_{+ (-1)}}}\cotensor{{\mathcal H}}u_{+(0)}) \tensor{B} u_{-} \quad \in P\tensor{B}\big(Q\cotensor{{\mathcal H}}P\big)\tensor{B}P,
$$
holds for any $u \in {\mathcal H}$, an equation which can be seen in $P\tensor{B}Q\tensor{A}P\tensor{B}P$ since $P_{\scriptscriptstyle{B}}$ is flat. Therefore,
$$
p^{\scriptscriptstyle{u}} \mathsf{g}(q^{\scriptscriptstyle{u}}) \,=\,
p^{\scriptscriptstyle{u_{+ (-1)}}} \beta\big( \varepsilon \zeta^{-1}\big( q^{\scriptscriptstyle{u_{+ (-1)}}}\cotensor{{\mathcal H}}u_{+(0)}\big) \big) u_{-}.
$$
On the other hand, by the first equality of Eq.~\eqref{Eq:chizeta},
$$
(p^{\scriptscriptstyle{u_{+ (-1)}}} \cotensor{{\mathcal K}} q^{\scriptscriptstyle{u_{+ (-1)}}} ) \cotensor{{\mathcal H}} u_{+(0)}\tensor{B} u_{-}
\, = \, \big( \chi(u_{+ (-1)})\cotensor{{\mathcal H}}u_{+(0)}\big)\tensor{B} u_{-} \,=\, \big( u_{+ (0)}\cotensor{{\mathcal K}}\zeta(u_{+(1)})\big)\tensor{B} u_{-},
$$
which implies that
$$
p^{\scriptscriptstyle{u_{+ (-1)}}} \tensor{B} ( q^{\scriptscriptstyle{u_{+ (-1)}}}\cotensor{{\mathcal H}}u_{+(0)} ) \tensor{B} u_{-} \, = \, u_{+(0)}\tensor{B}\zeta(u_{+ (1)})\tensor{B} u_{-},
$$
from which, in turn, we obtain that
$$
p^{\scriptscriptstyle{u}} \mathsf{g}(q^{\scriptscriptstyle{u}}) \, =\,
p^{\scriptscriptstyle{u_{+\,(-1)}}} \beta\big( \varepsilon \zeta^{-1}\big( q^{\scriptscriptstyle{u_{+(-1)}}}\cotensor{{\mathcal H}}u_{+(0)}\big) \big) u_{-}
\, = \, u_{+(0)}\beta\big(\varepsilon(u_{+(1)})\big) u_{-}\, = \, u_{+}u_{-} \, \overset{\eqref{Eq:eps+-}}{=} \, \alpha(\varepsilon(u)),
$$
as claimed. Using Eqs.~\eqref{Eq:G}, we now compute from one hand,
\begin{eqnarray*}
\mathsf{can} \circ \td{\mathsf{can}} (p\tensor{B}w) & = & \mathsf{can} ( p\,\mathsf{g}(q^{\scriptscriptstyle{w}})\tensor{A}p^{\scriptscriptstyle{w}}) \\
& = & p\,\mathsf{g}(q^{\scriptscriptstyle{w}})p^{\scriptscriptstyle{w}}{}_{{(0)}}\tensor{B} p^{\scriptscriptstyle{w}}{}_{{(1)}} \\
& = & p\,\mathsf{g}(q^{{w_{(1)}}})p^{{w_{(1)}}}\tensor{B} w_{{(2)}} \\
& = & p\,\beta(\varepsilon(w_{(1)}))\tensor{B} w_{(2)} \\ &=& p\tensor{B}w,
\end{eqnarray*}
and from the other side,
\begin{eqnarray*}
\td{\mathsf{can}} \circ \mathsf{can}(p'\tensor{A}p) &=& \td{\mathsf{can}}(p'p_{{(0)}}\tensor{B}p_{{(1)}}) \\
& =& p'p_{{(0)}}
\mathsf{g}(q^{\scriptscriptstyle{p_{(1)}}}) \tensor{A}p^{\scriptscriptstyle{p_{(1)}}} \\
&\overset{\eqref{Eq:chizeta}}{ =} & p'p^{\scriptscriptstyle{p_{(-1)}}}
\mathsf{g}(q^{\scriptscriptstyle{p_{(-1)}}}) \tensor{A}p_{{(0)}} \\
& =& p'\alpha\big( \varepsilon( p_{{(-1)}})\big)
\tensor{A}p_{{(0)}} \\ &=& p'\tensor{A}p,
\end{eqnarray*}
which gives the desired bijection, and so $(P,\alpha,\beta)$ is a principal bibundle. Similarly, one checks that $(Q,\theta,\sigma)$ is so as well.
To complete the proof of the first part, we also need to check that $Q$ is the opposite bundle of $P$. For this, we use the following chain of isomorphisms of $\Bbbk$-modules
$$
P\tensor{A}P \cong {\mathcal K}\tensor{B}P \cong (Q\cotensor{{\mathcal H}}P) \tensor{B}P \cong Q\tensor{A}P,
$$ where the last isomorphism is given by Eq.~\eqref{giotto1},
which leads to an isomorphism $P\cong Q$ of $A$-modules since $P$ is faithfully flat over $A$
(alternatively, one can try to check that $\mathsf{g}: P \to Q$ is a bundle map and hence an isomorphism by Lemma \ref{lemma:ISO}).
In the same way, using the faithfully flatness of $P$ over $B$, one shows
that this is also an isomorphism of $B$-modules, and thus that $Q$ is the opposite bundle of $P$.
To prove {\em (ii)}, assume first that $\B{\phi}$ is a weak equivalence.
Then $P$ is a right principal $({\mathcal H},{\mathcal K})$-bundle by Proposition \ref{lemma:trivial bibundle}, along with the fact that $- \cotensor{{\mathcal H}} P$
defines an equivalence of categories with inverse
$
- \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}.
$
From this it is clear that
$
P \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}} \simeq \mathscr{U}({\mathcal H})
$
and
$P^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal H}} P\simeq \mathscr{U}({\mathcal K})$, see Example \ref{Exam:H} for notation.
To prove the converse, using Proposition \ref{lemma:trivial bibundle} again, we only have to show that $P = {\mathcal H} \otimes_{\scriptscriptstyle{\phi}} B$ is a bibundle, which is a direct consequence of {\em (i)}.
\end{proof}
Recall that a {\em bigroupoid} (see, {\em e.g.}, \cite{Noo:NOTGTGACM})
is a bicategory in which every $1$-cell and every $2$-cell has an inverse (not necessarily in the strict sense for $1$-cells).
\begin{corollary}\label{coro:twoJapanese}
For two $0$-cells $(A,{\mathcal H})$ and $(B,{\mathcal K})$ (that is, flat Hopf algebroids), the full subcategory of invertible $1$-cells in $\mathsf{PB}^\ell({\mathcal H},{\mathcal K})$ coincides with the full subcategory $\mathsf{PB}^{\scriptscriptstyle{b}}({\mathcal H},{\mathcal K})$ of principal bibundles. In particular, the bicategory $\mathsf{PB}^\ell$ is a bigroupoid.
\end{corollary}
The last statement follows from Lemma \ref{lemma:ISO}
\subsection{The $2$-functor $\mathscr{P}$ and principal bundles as universal solution}
\label{ssec:2F}
It is well-known that groupoids, functors, and natural transformations form a $2$-category.
Adapting this to Hopf algebroids, one can construct a $2$-category as observed in \cite[\S3.1]{Naumann:07}.
Here, $0$-cells are Hopf algebroids (or even flat ones), $1$-cells are morphisms of Hopf algebroids, and for
two $1$-cells $(\zeta_{{0}},\zeta_{{1}}), (\theta_{{0}},\theta_{{1}}): (A, {\mathcal H}) \to (B, {\mathcal K})$, a $2$-cell $\fk{c}: (\zeta_{{0}},\zeta_{{1}}) \to (\theta_{{0}},\theta_{{1}})$ is defined to be an algebra map $\fk{c}: {\mathcal H} \to B$
that makes the diagrams
\begin{equation}\label{Eq:2cells}
\xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}}[r] & B \\ A \ar@{->}_-{\zeta_{\scriptscriptstyle{0}}}[ru] \ar@{->}^-{\sf{s}}[u] & }\qquad \xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}}[r] & B \\ A \ar@{->}_-{\theta_{\scriptscriptstyle{0}}}[ru] \ar@{->}^-{\sf{t}}[u] & } \qquad \xymatrix{ {\mathcal H} \ar@{->}^-{\Delta}[rr] \ar@{->}_-{\Delta}[d] & & {\mathcal H}\tensor{A}{\mathcal H} \ar@{->}^-{m_{\scriptscriptstyle{{\mathcal K}}} (\zeta_{\scriptscriptstyle{1}}\tensor{A}\sf{t}\fk{c})}[d] \\ {\mathcal H}\tensor{A}{\mathcal H} \ar@{->}_-{m_{\scriptscriptstyle{{\mathcal K}}}(\sf{s}{\fk{c}}\,\tensor{A}\theta_{\scriptscriptstyle{1}})}[rr] & & {\mathcal K} }
\end{equation}
commutative, where $m_{\scriptscriptstyle{{\mathcal K}}}$ denotes the multiplication in ${\mathcal K}$.
The identity $2$-cell for $(\zeta_{{0}},\zeta_{{1}})$ is given by $1_{\scriptscriptstyle{\zeta}}:= \zeta_{{0}} \circ \varepsilon$.
The tensor product (or vertical composition) of $2$-cells
is given as
$$
\xymatrix{ \fk{c}' \circ \fk{c}: (\zeta_{{0}},\zeta_{{1}}) \ar@{->}^-{\fk{c}}[r] & (\theta_{{0}},\theta_{{1}}) \ar@{->}^-{\fk{c}'}[r] & (\xi_{{0}},\xi_{{1}}),}
$$
which yields a map
\begin{equation}
\label{Eq:verticalComp}
\fk{c}' \B{\circ} \fk{c}: {\mathcal H} \to B, \quad u \mapsto \fk{c}(u_{{(1)}}) \fk{c}'(u_{{(2)}}).
\end{equation}
We denote by $2\text{-}\mathsf{HAlgd}$
the $2$-category whose $0$-cells are flat Hopf algebroids.
Examples of $2$-cells in this $2$-category are described by the following lemma:
\begin{lemma}\label{lemma:volveranotro}
Let $\B{\phi}: (A,{\mathcal H}) \to (B,{\mathcal K})$ be a morphism of flat Hopf algebroids. As in Example \ref{exam:HopfMorph}, consider its associated trivial left principal $({\mathcal H},{\mathcal K})$-bundle $(P :={\mathcal H}\tensor{\scriptscriptstyle{\phi}}B,\alpha,\beta)$ together with the diagram
$$
\xymatrix@R=15pt{ & (P,{\mathcal H}\lJoin P\rJoin{\mathcal K}) & \\ (A,{\mathcal H}) \ar@{->}^-{{\B{\alpha}=(\alpha, \,\alpha_1)}\;\;}[ur] \ar@{->}_-{\B{\phi}}[rr]& & \ar@{->}_-{\;\;{\B{\beta}=(\beta, \,\beta_1)}}[ul] (B,{\mathcal K})}
$$
of Hopf algebroids,
where the top is the two-sided translation Hopf algebroid defined in Lemma \ref{lemma:bisemidirect}.
Then there is a $2$-isomorphism
$\B{\alpha} \cong \B{\beta}\circ \B{\phi}$, that is, the above diagram is commutative up to an isomorphism.
\end{lemma}
\begin{proof}
Consider the following two algebra maps
$$
\fk{c}: {\mathcal H} \to P, \quad u \mapsto u\tensor{\scriptscriptstyle{\phi}}1_{ \scriptscriptstyle{B}}, \quad \mbox{and} \quad \fk{c}': {\mathcal H} \to P, \quad u \mapsto \mathscr{S}(u)\tensor{\scriptscriptstyle{\phi}}1_{ \scriptscriptstyle{B}}.
$$
Let us check that $\fk{c}: \B{\alpha} \to \B{\beta} \circ \B{\phi}$ and $\fk{c'}: \B{\beta} \circ \B{\phi} \to \B{\alpha}$ are $2$-cells in $2\text{-}\mathsf{HAlgd}$. To this end, we need to show the commutativity of the diagrams in Eq.~\eqref{Eq:2cells}, corresponding to $\fk{c}$ and $\fk{c'}$.
By definition, it is clear that the triangles
$$
\xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}}[r] & P \\ A \ar@{->}_-{\alpha}[ru] \ar@{->}^-{\sf{s}}[u] & }\qquad \xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}}[r] & P \\ A \ar@{->}_-{\beta\phi_{\scriptscriptstyle{0}}}[ru] \ar@{->}^-{\sf{t}}[u] & } \qquad \qquad
\xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}'}[r] & P \\ A \ar@{->}_-{\beta\phi_{\scriptscriptstyle{0}}}[ru] \ar@{->}^-{\sf{s}}[u] & }\qquad \xymatrix{ {\mathcal H} \ar@{->}^-{\fk{c}'}[r] & P \\ A \ar@{->}_-{\alpha}[ru] \ar@{->}^-{\sf{t}}[u] & }
$$
commute.
We only show the rectangle in \eqref{Eq:2cells} for $\fk{c'}$ since an analogous proof works for $\fk{c}$.
Thus, we want to show that $m_{\scriptscriptstyle{{\mathcal H}\lJoin P\rJoin{\mathcal K}}}\circ \big( (\beta_1 \circ \phi_{\scriptscriptstyle{1}})\tensor{A}(\Sf{t} \circ \fk{c'})\big) \circ \Delta \,=\, m_{\scriptscriptstyle{{\mathcal H}\lJoin P\rJoin{\mathcal K}}} \circ \big( (\Sf{s} \circ \fk{c'})\tensor{A}\alpha_{\scriptscriptstyle{1}}\big) \circ \Delta$, where the target and source $\Sf{t}, \Sf{s}$ are those of ${\mathcal H}\lJoin P\rJoin{\mathcal K}$.
Taking into account the structure maps of Lemma \ref{lemma:bisemidirect}, we compute for $u \in {\mathcal H}$
\begin{eqnarray*}
m_{\scriptscriptstyle{{\mathcal H}\lJoin P\rJoin{\mathcal K}}}\circ \big( (\beta_1 \circ \phi_{\scriptscriptstyle{1}})\tensor{A}(\Sf{t} \circ \fk{c'})\big) \circ \Delta(u) &=& \big( 1_{ \scriptscriptstyle{{\mathcal H}}}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B} \phi_{\scriptscriptstyle{1}}(u_{(1)})\big) \Sf{t}\big(\mathscr{S}(u_{(2)})\tensor{\scriptscriptstyle{\phi}}1_{\scriptscriptstyle{B}} \big) \\
&=& \big( 1_{ \scriptscriptstyle{{\mathcal H}}}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B} \phi_{\scriptscriptstyle{1}}(u_{(1)})\big) \big(u_{(4)} \tensor{A} (\mathscr{S}(u_{(3)}) \tensor{\scriptscriptstyle{\phi}}1_{\scriptscriptstyle{B}}) \tensor{B} \phi_{\scriptscriptstyle{1}}(\mathscr{S}(u_{(2)})) \big) \\
&=& u_{(4)} \tensor{A} (\mathscr{S}(u_{(3)})\tensor{\scriptscriptstyle{\phi}}1_{\scriptscriptstyle{B}}) \tensor{B} \phi_{\scriptscriptstyle{1}}(u_{(1)})\phi_{\scriptscriptstyle{1}}(\mathscr{S}(u_{(2)})) \\
&=& u_{(3)}\tensor{A} (\mathscr{S}(u_{(2)})\tensor{\scriptscriptstyle{\phi}}1_{\scriptscriptstyle{B}}) \tensor{B} \Sf{s}(\phi_{\scriptscriptstyle{0}}(\varepsilon(u_{(1)}))) \\
&=& u_{(3)}\tensor{A} \big( \mathscr{S}(u_{(2)})\tensor{\scriptscriptstyle{\phi}}\phi_{\scriptscriptstyle{0}}(\varepsilon(u_{(1)})) \big)\tensor{B} 1_{ \scriptscriptstyle{{\mathcal K}}} \\
&=& u_{(2)}\tensor{A} \big( \mathscr{S}(u_{(1)})\tensor{\scriptscriptstyle{\phi}} 1_{ \scriptscriptstyle{B}} \big)\tensor{B} 1_{ \scriptscriptstyle{{\mathcal K}}} \\
&=& m_{\scriptscriptstyle{{\mathcal H}\lJoin P\rJoin{\mathcal K}}}\circ \big( (\Sf{s}\circ \fk{c'})\tensor{A}\alpha_1\big) \circ \Delta(u).
\end{eqnarray*}
Finally, using the vertical composition as defined in \eqref{Eq:verticalComp}, one can easily check that $\fk{c} \B{\circ} \fk{c'}= (\beta \phi_{\scriptscriptstyle{0}}) \circ \varepsilon$ and that $\fk{c'} \B{\circ} \fk{c}= \alpha \circ \varepsilon$. Therefore $\fk{c} \B{\circ} \fk{c'}=1_{\scriptscriptstyle{\B{\beta}\circ \B{\phi}}}$ and $\fk{c'} \B{\circ} \fk{c} =1_{\scriptscriptstyle{\alpha}}$, and this completes the proof.
\end{proof}
For a non necessarily trivial bundle, one has the following property:
\begin{lemma}\label{lemma:paciencia}
Let $(P,\alpha, \beta)$ be a $1$-cell in $\mathsf{PB}^\ell({\mathcal H},{\mathcal K})$, and denote by $(P,{\mathcal H}\lJoin P\rJoin {\mathcal K})$ the two-sided translation Hopf algebroid, together with the diagram
$$
\xymatrix@R=15pt{ & (P,{\mathcal H}\lJoin P\rJoin {\mathcal K}) & \\ (A,{\mathcal H}) \ar@{->}^-{{\B{\alpha}=(\alpha, \,\alpha_1)}\;\;}[ur] & & \ar@{->}_-{\;\;{\B{\beta}=(\beta, \,\beta_1)}}[ul] (B,{\mathcal K}) }
$$
of flat Hopf algebroids.
Consider the trivial bundles $\B{\alpha}^*\big( \mathscr{U}({\mathcal H})\big)={\mathcal H}\tensor{\B{\alpha}}P$ and $\B{\beta}^*\big( \mathscr{U}({\mathcal K})\big)={\mathcal K}\tensor{\B{\beta}}P$.
Then the map
$$
\fk{h}:(P,\alpha,\beta) \longrightarrow \Big(\B{\alpha}^*\big( \mathscr{U}({\mathcal H})\big) \cotensor{{\mathcal H}\lJoin P\rJoin {\mathcal K}} \B{\beta}^*\big( \mathscr{U}({\mathcal K})\big)^{\scriptscriptstyle{{\rm co}}},\td{\alpha}, \td{\beta}\Big), \quad p \longmapsto \big(p_{(-1)}\tensor{\B{\alpha}}p_{(0)} \big) \, \cotensor{{\mathcal H}\lJoin P\rJoin {\mathcal K}}\, \big(1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(1)} \big)
$$
defines an isomorphism of left principal $({\mathcal H},{\mathcal K})$-bundles.
\end{lemma}
\begin{proof}
Recall that a generic element of the form $(u\tensor{\B{\alpha}}p) \tensor{P} (p'\tensor{\B{\beta}}w) \in \B{\alpha}^*\big( \mathscr{U}({\mathcal H})\big)\, \tensor{P}\, \B{\beta}^*\big( \mathscr{U}({\mathcal K})\big)^{\scriptscriptstyle{{\rm co}}}$ belongs to the cotensor product $\B{\alpha}^*\big( \mathscr{U}({\mathcal H})\big)\, \cotensor{{\mathcal H}\lJoin P\rJoin {\mathcal K}}\, \B{\beta}^*\big( \mathscr{U}({\mathcal K})\big)^{\scriptscriptstyle{{\rm co}}}$ if and only if
\begin{equation}
\label{firenzesmn}
(u_{(1)}\tensor{\B{\alpha}} p) \tensor{P} \big( u_{(2)}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B} 1_{\scriptscriptstyle{{\mathcal K}}}\big) \tensor{P} (p'\tensor{\B{\beta}}w) = (u\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( 1_{\scriptscriptstyle{{\mathcal H}}}\tensor{A}pp'\tensor{B} w_{(1)}\big) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}w_{(2)})
\end{equation}
holds true in $\B{\alpha}^*\big( \mathscr{U}({\mathcal H})\big)\tensor{P} ({\mathcal H}\lJoin P\rJoin {\mathcal K}) \tensor{P} \B{\beta}^*\big( \mathscr{U}({\mathcal K})\big)^{\scriptscriptstyle{{\rm co}}}$. Hence, in order to check that $\fk{h}$ is well-defined, one needs to show this equality for $\fk{h}(p)$, for all $ p \in P$. The left hand side in \rmref{firenzesmn} for $\fk{h}(p)$ reads as
$$
(p_{(-2)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( p_{(-1)}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B} 1_{\scriptscriptstyle{{\mathcal K}}}\big) \tensor{P} ( p_{(0)}\tensor{\B{\beta}}p_{(1)}),
$$
while the right hand side becomes
$$
(p_{(-1)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( 1_{\scriptscriptstyle{{\mathcal H}}}\tensor{A}p_{(0)}\tensor{B} p_{(1)}\big) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(2)}).
$$
Using the expression of the target map of ${\mathcal H}\lJoin P\rJoin {\mathcal K}$ given in Lemma \ref{lemma:bisemidirect}, we have that
\begin{eqnarray*}
& & (p_{(-2)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( p_{(-1)}\tensor{A}1_{\scriptscriptstyle{P}}\tensor{B} 1_{\scriptscriptstyle{{\mathcal K}}}\big) \tensor{P} ( p_{(0)}\tensor{\B{\beta}}p_{(1)}) \\ &= & (p_{(-2)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( p_{(-1)}\tensor{A} 1_{\scriptscriptstyle{P}} \tensor{B} 1_{\scriptscriptstyle{{\mathcal K}}} \big) \Sf{t}(p_{(0)}) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(1)})\\ &= & (p_{(-3)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( p_{(-2)}\mathscr{S}(p_{(-1)})\tensor{A}p_{(0)}\tensor{B} p_{(1)}\big) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(2)}) \\ &\overset{\eqref{Eq:penultima}}{=} & (p_{(-2)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( \Sf{s}(\varepsilon(p_{(-1)}))\tensor{A}p_{(0)}\tensor{B} p_{(1)}\big) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(2)}) \\ & = & (p_{(-1)}\tensor{\B{\alpha}}1_{\scriptscriptstyle{P}}) \tensor{P} \big( 1_{\scriptscriptstyle{{\mathcal H}}}\tensor{A}p_{(0)}\tensor{B} p_{(1)}\big) \tensor{P} ( 1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}p_{(2)}),
\end{eqnarray*} which shows that $\fk{h}$ is a well-defined map.
Recall now that the algebra maps $\td{\alpha}$ and $\td{\beta}$ are given by
$$
\td{\alpha}(a)= (\Sf{s}(a) \tensor{\B{\alpha}}1_{\scriptscriptstyle{P}})\cotensor{{\mathcal H}\lJoin P\rJoin {\mathcal K}} (1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}1_{\scriptscriptstyle{{\mathcal K}}}); \qquad \td{\beta}(b)= (1_{\scriptscriptstyle{{\mathcal H}}} \tensor{\B{\alpha}}1_{\scriptscriptstyle{P}})\cotensor{{\mathcal H}\lJoin P\rJoin {\mathcal K}} (1_{\scriptscriptstyle{P}}\tensor{\B{\beta}}\Sf{t}(b)).
$$
Clearly, $\fk{h}$ is simultaneously an $A$-algebra and a $B$-algebra map, and the fact that $\fk{h}$ is an $({\mathcal H}, {\mathcal K})$-bicomodule map is also clear from the definitions.
Thus, $\fk{h}$ is a morphism of left principal bundles,
and so an isomorphism by Lemma \ref{lemma:ISO}.
\end{proof}
Next we give a further property of the Diagram \rmref{kontoauszug} that appeared in Proposition \ref{prop:1out2}.
\begin{lemma}\label{lema:1out2}
Let $\theta_i:(A,{\mathcal H}) \to (C_i,{\mathcal J}_i)$, $i=1,2$, be two weak equivalences. Then the
diagram of weak equivalences \rmref{kontoauszug} constructed in Proposition \ref{prop:1out2} is commutative up to a $2$-isomorphism.
\end{lemma}
\begin{proof}
Denote by $P_i := {\mathcal H}\tensor{\theta_i}C_i$, $i=1,2$ the respective associated trivial bibundles of $\theta_i$. Up to a canonical isomorphism, the bundle $Q := P_1^{\scriptscriptstyle{{\rm co}}}\cotensor{{\mathcal H}}P_2$ is of the form $Q=C_1\tensor{A}{\mathcal H}\tensor{A}C_2$. So, considering the obvious algebra map $\fk{c}: {\mathcal H} \to Q$, $u \mapsto 1\tensor{A}u \tensor{A}1$ and writing $\B{\phi}:= \B{\zeta}_1 \circ \B{\theta}_1$ and $\B{\psi}:=\B{\zeta}_2 \circ \B{\theta}_2$, one can use the definition of the maps $\B{\zeta}_i$ in Lemma \ref{lemma:bisemidirect} to show that the diagrams in \eqref{Eq:2cells} are commutative, and that hence $\fk{c}: \B{\phi} \to \B{\psi}$ is a $1$-cell in $2\text{-}\mathsf{HAlgd}$. Its inverse is $\fk{c}^{-1}: {\mathcal H} \to Q$ which sends $u \mapsto 1\tensor{A}\mathscr{S}(u)\tensor{A}1$.
\end{proof}
Denote by $ \mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}$ the conjugate bicategory of $\mathsf{PB}^\ell$, defined by reversing $2$-cells.
\begin{proposition}\label{prop:P}
There is a $2$-functor
$$
\mathscr{P}: 2\text{-}\mathsf{HAlgd} \longrightarrow \mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}},
$$
which sends any $1$-cell $\B{\phi}: (A,{\mathcal H}) \to (B, {\mathcal K})$ to its associated trivial left principal bundle $P={\mathcal H}\tensor{\scriptscriptstyle{\phi}}B$. Moreover, a $1$-cell $\B{\phi}$ in $2\text{-}\mathsf{HAlgd}$ is a weak equivalence if and only if $\mathscr{P}(\B{\phi})$ is an invertible $1$-cell in $\mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}$.
\end{proposition}
\begin{proof}
Let $\fk{c}: \B{\phi} \to \B{\psi}$ be a $2$-cell in $2\text{-}\mathsf{HAlgd}$. Then its image by $\mathscr{P}$ is given by
$$
\mathscr{P}(\fk{c}): {\mathcal H}\tensor{\psi}B \to {\mathcal H}\tensor{\phi}B, \quad u\tensor{\psi}b \mapsto u_{(1)}\tensor{\phi}\fk{c}(u_{(2)})b,
$$
which is easily shown to be a morphism of left principal bundles. The remaining axioms which $\mathscr{P}$ is required to fulfil are also easily shown and therefore left to the reader. Nevertheless, notice that for two composable $1$-cells $\B{\phi}:(A,{\mathcal H}) \to (B, {\mathcal K})$ and $\B{\phi}': (B, {\mathcal K}) \to (C,{\mathcal J})$ one has
$$
\mathscr{P}(\B{\phi}' \circ \B{\phi})\,\, \cong\,\, \mathscr{P}(\B{\phi})\cotensor{{\mathcal K}} \mathscr{P}(\B{\phi}'),
$$
that is, $\mathscr{P}$ is contravariant. The last statement is a direct consequence of Proposition \ref{prop:el parte} {\em (ii)}.
\end{proof}
The following theorem is Theorem \ref{thm:Ramadan} in the Introduction and is our second main result:
\begin{thm}\label{thm:elevenyearsandsevenmonths}
Let $\mathscr{F}: 2\text{-}\Sf{HAlgd} \to \mathscr{B}$ be a 2-functor which sends weak equivalences to invertible $1$-cells. Then, up to isomorphism (of $2$-functors), there is a unique $2$-functor $\td{\mathscr{F}}$ such that the following diagram
\begin{equation}
\label{tiburtina}
\xymatrix@R=20pt{ 2\text{-}\Sf{HAlgd} \ar@{->}_-{\mathscr{F}}[rrd] \ar@{->}^-{\mathscr{P}}[rr] & & \mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}} \ar@{.>}^-{\td{\mathscr{F}}}[d] \\ & & \mathscr{B} }
\end{equation}
commutes up to an isomorphism of $2$-functors.
\end{thm}
\begin{proof}
For two $0$-cells $(A,{\mathcal H})$ and $(B,{\mathcal K})$ and a $1$-cell $(P,\alpha,\beta)$ in
$\mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}({\mathcal H},{\mathcal K})$, from Proposition \ref{prop:Cielitomissyoutoo} one obtains that $\B{\beta}:(B,{\mathcal K}) \to (P,{\mathcal H} \lJoin P \rJoin {\mathcal K})$ is a weak equivalence. Then, by assumption, $\mathscr{F}(\B{\beta})$ is an invertible $1$-cell in $\mathscr{B}\big(\mathscr{F}(A,{\mathcal H}), \mathscr{F}(B,{\mathcal K})\big)$; denote by $\mathscr{F}(\B{\beta})^{-1} \in \mathscr{B}\big(\mathscr{F}(B,{\mathcal K}), \mathscr{F}(A,{\mathcal H})\big)$ its inverse. Define furthermore
$$
\td{\mathscr{F}}(P,\alpha,\beta) := \mathscr{F}(\B{\beta})^{-1} \circ \mathscr{F}(\B{\alpha}),
$$
which gives a $1$-cell in $\mathscr{B}\big(\mathscr{F}(A,{\mathcal H}), \mathscr{F}(B,{\mathcal K})\big)$. In particular, the image of the unit bundle $(\mathscr{U}({\mathcal H}), \bf{\Sf{s}}, \bf{\Sf{t}})$ then is, by using Lemma \ref{lemma:volveranotro}, of the form
$$
\td{\mathscr{F}}\big( \mathscr{U}({\mathcal H})\big) \cong \mathscr{F}({\rm id}_{\scriptscriptstyle{(A,\, {\mathcal H})}}) = \bf{1}_{\scriptscriptstyle{\mathscr{F}(A,\, {\mathcal H})}},
$$
the identity $1$-cell of the monoidal category $\mathscr{B}\big(\mathscr{F}(A,{\mathcal H}), \mathscr{F}(A,{\mathcal H}) \big)$. Now, the image of a $2$-cell $\fk{f}: (P',\alpha',\beta') \to (P,\alpha,\beta)$ in $\mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}({\mathcal H},{\mathcal K})$ by $ \td{\mathscr{F}}$ is going to be a $2$-isomorphism: define
$$
\td{\mathscr{F}}(\fk{f}): \td{\mathscr{F}}(P',\alpha',\beta') = \mathscr{F}(\B{\beta'})^{-1} \circ \mathscr{F}(\B{\alpha'}) \longrightarrow \mathscr{F}(\B{\beta})^{-1} \circ \mathscr{F}(\B{\alpha}) = \td{\mathscr{F}}(P,\alpha,\beta)
$$
as the unique isomorphism in $\mathscr{B}\big(\mathscr{F}(A,{\mathcal H}), \mathscr{F}(B,{\mathcal K})\big)$
satisfying
$$
\mathscr{F}(\B{\beta'}) \circ \td{\mathscr{F}}(\fk{f}) \,=\,
1_{\scriptscriptstyle{\mathscr{F}(\B{\alpha'})}} \,=\, 1_{\scriptscriptstyle{\mathscr{F}(\fk{f}) \circ \mathscr{F}(\B{\alpha})}}
$$
since from Diagram \eqref{Eq:qseramana} follows that $\fk{f} \circ \B{\alpha} = \B{\alpha'}$ and $\fk{f} \circ \B{\beta}=\B{\beta'}$ as $2$-cells in $2\text{-}\Sf{HAlgd}$, where, by abuse of notation, we did not distinguish between the vertical and horizontal composition in $\mathscr{B}$.
The fact that $\td{\mathscr{F}}$ is compatible with both vertical and horizontal compositions of $\mathsf{PB}^{\ell \, \scriptscriptstyle{{\rm co}}}$ is shown as follows:
first, as seen above, $\td{\mathscr{F}}\big( \mathscr{U}({\mathcal H})\big) \cong \bf{1}_{\scriptscriptstyle{\mathscr{F}(A,\, {\mathcal H})}}$ for every $0$-cell $(A,{\mathcal H})$.
Second, for $(P, \alpha, \beta) \in \lPB{{\mathcal H}}{{\mathcal K}}$ and $(Q,\sigma,\theta) \in \lPB{{\mathcal K}}{{\mathcal J}}$ consider their product $$(P\cotensor{{\mathcal K}}Q, \td{\alpha},\td{\theta}) \in \lPB{{\mathcal H} \lJoin P \rJoin {\mathcal K}}{{\mathcal K} \lJoin Q \rJoin {\mathcal J}},$$ where $\td{\alpha}$ and $\td{\theta}$ are as in Diagram \eqref{Eq:earthquake}.
Consider the morphism $\B{\sigma}:(B,{\mathcal K}) \to (Q,{\mathcal K}\lJoin Q \rJoin {\mathcal J})$ of Hopf algebroids as in Lemma \ref{lemma:bisemidirect}.
From the trivial bundles $\B{\sigma}^*(\mathscr{U}({\mathcal K})) \in \lPB{{\mathcal K}}{{\mathcal K} \lJoin Q\rJoin {\mathcal J}}$ and $\B{\beta}^*(\mathscr{U}({\mathcal K})) \in \lPB{{\mathcal K}}{{\mathcal H} \lJoin P\rJoin {\mathcal K}}$ we can construct their product
$\B{\beta}^*(\mathscr{U}({\mathcal K}))^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal K}} \B{\sigma}^*(\mathscr{U}({\mathcal K}))$, which belongs to $\lPB{{\mathcal H} \lJoin P\rJoin {\mathcal K}}{{\mathcal K}\lJoin Q\rJoin {\mathcal J}}$. On the other hand, an easy verification shows that $(P\tensor{B}{\mathcal K}\tensor{B}Q, \gamma, \delta)$
is also a principal bundle in $\lPB{{\mathcal H}\lJoin P\rJoin {\mathcal K}}{{\mathcal K}\lJoin Q\rJoin {\mathcal J}}$, where
$$
\gamma: P \to P\tensor{B}{\mathcal K} \tensor{B}Q, \quad p \mapsto p\tensor{B}1_{\scriptscriptstyle{{\mathcal K}}}\tensor{B}1_{\scriptscriptstyle{Q}}; \qquad \delta: Q \to P\tensor{B}{\mathcal K} \tensor{B}Q, \quad q \mapsto 1_{\scriptscriptstyle{P}}\tensor{B}1_{\scriptscriptstyle{{\mathcal K}}} \tensor{B}q,
$$
and using the canonical bicomodule structure given by the coaction
\begin{small}
$$
P\tensor{B}{\mathcal K}\tensor{B}Q \to ({\mathcal H}\lJoin P\rJoin {\mathcal K}) \tensor{P}(P\tensor{B}{\mathcal K}\tensor{B}Q), \;\; p\tensor{B}w \tensor{B}q \mapsto (1_{\scriptscriptstyle{{\mathcal H}}}\tensor{A}p\tensor{B}w_{(1)})\tensor{P}(1_{\scriptscriptstyle{P}}\tensor{B}w_{(2)}\tensor{B}q)
$$
\end{small}
as well as
\begin{small}
$$
P\tensor{B}{\mathcal K}\tensor{B}Q \to (P\tensor{B}{\mathcal K}\tensor{B}Q) \tensor{Q} ({\mathcal K}\lJoin Q\rJoin {\mathcal J}), \;\; p\tensor{B}w \tensor{B}q \mapsto (p\tensor{B}w_{(1)} \tensor{B}1_{\scriptscriptstyle{Q}})\tensor{Q}(w_{(2)}\mathscr{S}(q_{(-1)})\tensor{B}q_{(0)}\tensor{C}q_{(1)}).
$$
\end{small}
Taking into account the canonical isomorphism
$$
\B{\beta}^*(\mathscr{U}({\mathcal K}))^{\scriptscriptstyle{{\rm co}}} \cotensor{{\mathcal K}} \B{\sigma}^*(\mathscr{U}({\mathcal K})) \,=\, \big( P\tensor{\scriptscriptstyle{\beta}}{\mathcal K}\big) \cotensor{{\mathcal K}} \big({\mathcal K} \tensor{\scriptscriptstyle{\sigma}}Q \big) \cong P\tensor{B}{\mathcal K}\tensor{B}Q
$$
of bicomodule algebras, we can then identify both principal bundles. The two-sided translation Hopf algebroids associated to $(P\cotensor{{\mathcal K}}Q, \td{\alpha},\td{\theta})$ resp.\ $(P\tensor{B}{\mathcal K}\tensor{B}Q, \gamma, \delta)$ are now related via the morphism
$$
\B{\muup}: \Big(P\cotensor{{\mathcal K}}Q,\, {\mathcal H}\lJoin (P\cotensor{{\mathcal K}}Q)\rJoin {\mathcal J} \Big) \to \Big(P\tensor{B}{\mathcal K}\tensor{B}Q,\, ({\mathcal H} \lJoin P\rJoin {\mathcal K}) \lJoin(P\tensor{B}{\mathcal K}\tensor{B}Q)\rJoin ( {\mathcal K}\lJoin Q \rJoin {\mathcal J})\Big)
$$
of Hopf algebroids,
sending
$$
(p'\cotensor{{\mathcal K}}q', u\tensor{A}(p\cotensor{{\mathcal K}}q) \tensor{B} j) \mapsto \big( p'_{(0)}\tensor{B}p'_{(1)}\tensor{B}q',\, \B{\td{\alpha}}(u) \tensor{P} (p_{(0)}\tensor{B}p_{(1)}\tensor{B}q) \tensor{Q} \B{\td{\theta}}(j) \big),
$$
where $\B{\td{\alpha}}$ and $\B{\td{\theta}}$ are the associated maps to $\td{\alpha}$ and $\td{\theta}$ as in Lemma \ref{lemma:bisemidirect}, and from which we deduce the following commutative diagram:
\begin{small}
$$
\xymatrix{ & & \scriptstyle{\Big(P\cotensor{{\mathcal K}}Q,\; {\mathcal H} \lJoin (P\cotensor{{\mathcal K}}Q) \rJoin {\mathcal J} \Big)} \ar@{->}^-{\scriptstyle{\B{\muup}}}[d] & & \\ & & \scriptstyle{\Big(P\tensor{B}{\mathcal K}\tensor{B}Q,\; ({\mathcal H}\lJoin P\rJoin {\mathcal K}) \lJoin (P\tensor{B}{\mathcal K}\tensor{B}Q) \rJoin ({\mathcal K}\lJoin Q\rJoin {\mathcal J})\Big)} & & \\ & \scriptstyle{(P,{\mathcal H}\lJoin P\rJoin {\mathcal K})} \ar@{->}^-{\scriptstyle{\B{\gamma}}}[ur] & & \scriptstyle{(Q, {\mathcal K} \lJoin Q\rJoin {\mathcal J})} \ar@{->}_-{\scriptstyle{\B{\delta}}}[ul] & \\ \scriptstyle{(A,{\mathcal H})} \ar@/^4pc/^-{\scriptstyle{\B{\td{\alpha}}}}[uuurr] \ar@{->}_-{\scriptstyle{\B{\alpha}}}[ur] & & \ar@{->}^-{\scriptstyle{\B{\beta}}}[ul] \scriptstyle{(B,{\mathcal K})} \ar@{->}_-{\scriptstyle{\B{\sigma}}}[ur] & & \scriptstyle{(C,{\mathcal J})}. \ar@{->}^-{\scriptstyle{\B{\theta}}}[ul] \ar@/_4pc/_-{\scriptstyle{\B{\td{\theta}}}}[uuull] }
$$
\end{small}
Applying the functor $\mathscr{F}$ to this diagram and taking into account that $\B{\beta}, \B{\delta}, \B{\theta}$, and $\B{\td{\theta}}$ are weak equivalences by Proposition \ref{prop:Cielitomissyoutoo} (while $\B{\alpha}$ and $\B{\sigma}$ are not necessarily so since $P$ and $Q$ are just left bundles), we obtain the equality
$$
\mathscr{F}(\B{\td{\theta}})^{-1} \circ \mathscr{F}(\B{\td{\alpha}}) \,=\, \mathscr{F}(\B{\theta})^{-1} \circ \mathscr{F}(\B{\sigma}) \circ \mathscr{F}(\B{\beta})^{-1} \circ \mathscr{F}(\B{\alpha}),
$$
which means that
$$
\td{\mathscr{F}}(Q,\sigma,\theta) \circ \td{\mathscr{F}}(P,\alpha,\beta) \,=\, \td{\mathscr{F}}(P\cotensor{{\mathcal K}}Q, \td{\alpha},\td{\theta}),
$$
that is, $\td{\mathscr{F}}$ is contravariant (in the proof of Proposition \ref{prop:P} we saw that $\mathscr{P}$ is also contravariant, hence $\td{\mathscr{F}} \circ \mathscr{P}$ is covariant). To show that $\td{\mathscr{F}}$ is unique up to isomorphism, one uses Lemma
\ref{lemma:paciencia}. Finally, to check that the Diagram \rmref{tiburtina} is commutative up to $2$-isomorphism, one makes use of Lemma \ref{lemma:volveranotro}.
\end{proof}
\section{Principal bibundles and Morita equivalences}
In this section, which contains one of our main results (Theorem \ref{processocivile} in the Introduction), we explore the relationship between bibundles and Morita theory. We remind the reader that, as in Definition \ref{quellala2}, two flat Hopf algebroids are said to be \emph{Morita equivalent} if their categories of (right) comodules are equivalent as symmetric monoidal categories.
\subsection{Principal bibundles versus monoidal equivalence}
The result we want to prove first and which will be part of the main theorem reads as follows:
\begin{thm}
\label{aromanflower}
Let $(A, {\mathcal H})$ and $(B, {\mathcal K})$ be two flat Hopf algebroids and
$(P, \ga, \gb)$ be a principal $({\mathcal H}, {\mathcal K})$-bibundle. Then the functor
$$
- \cotensor{{\mathcal H}} P: \rcomod{{\mathcal H}} \lra \rcomod{{\mathcal K}}
$$
induces a symmetric monoidal equivalence of categories.
\end{thm}
\begin{proof}
Let us first check that the functor is symmetric monoidal:
by Remark \ref{rema:PB} {\em (ii)}, there is an algebra isomorphism
$$
A \cotensor{{\mathcal H}} P \simeq P^{\scriptscriptstyle{{\rm coinv}}} \simeq B
$$
as $\gb$ is injective. Second, for two right ${\mathcal H}$-comodules $M$ and $N$ define the map
$$
\gd: (M \cotensor{{\mathcal H}} P) \otimes^{\scriptscriptstyle{B}} (N \cotensor{{\mathcal H}} P) \to (M \otimes^{\scriptscriptstyle{A}} N)\cotensor{{\mathcal H}} P, \quad
(m \cotensor{{\mathcal H}} p) \otimes^{\scriptscriptstyle{B}} (n \cotensor{{\mathcal H}} p') \mapsto (m \otimes^{\scriptscriptstyle{A}} n) \cotensor{{\mathcal H}} pp',
$$
which is a morphism of right ${\mathcal K}$-comodules, where the tensor products are those of comodules as explained in Remark \ref{mondragone}.
In order to show that $\gd$ is an isomorphism, we proceed similarly as before
and show that $\gd \otimes_{\scriptscriptstyle{B}} {\rm id}_{\scriptscriptstyle{P}}$ is an isomorphism since $P$ is faithfully flat over $B$. Now a straightforward verification proves that the composition
$$
\zeta_{\scriptscriptstyle{M \otimes_{\scriptscriptstyle{A}} N}} \circ (\gd \otimes_{\scriptscriptstyle{B}} {\rm id}_{\scriptscriptstyle{P}}): \big((M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} (N \cotensor{{\mathcal H}} P)\big)\otimes_{\scriptscriptstyle{B}} P \to (M \otimes_{\scriptscriptstyle{A}} N) \otimes_{\scriptscriptstyle{A}} P,
$$
using the natural transformation $\zeta$ from \rmref{giotto1}, coincides with the following chain
\begin{equation*}
\begin{split}
\big((M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} (N \cotensor{{\mathcal H}} P)\big) \otimes_{\scriptscriptstyle{B}} P &\overset{{\rm id} \otimes_{\scriptscriptstyle{B}} \zeta_{\scriptscriptstyle{N}}}{ \relbar\joinrel\relbar\joinrel\longrightarrow} (M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} (N \otimes_{\scriptscriptstyle{A}} P) \overset{\simeq}{\lra} \big((M \cotensor{{\mathcal H}} P) \otimes_{\scriptscriptstyle{B}} N\big) \otimes_{\scriptscriptstyle{A}} P \\
&\overset{\zeta_{\scriptscriptstyle{N}} \otimes_{\scriptscriptstyle{A}} {\rm id}_{\scriptscriptstyle{N}}}{\relbar\joinrel\relbar\joinrel\longrightarrow} (M \otimes_{\scriptscriptstyle{A}} P) \otimes_{\scriptscriptstyle{A}} N \overset{\simeq}{\lra} (M \otimes_{\scriptscriptstyle{A}} N) \otimes_{\scriptscriptstyle{A}} P
\end{split}
\end{equation*}
of isomorphisms,
where the last step simply uses the tensor flip and the associativity of the tensor product. Clearly, $\gd$ is a natural transformation and compatible with the symmetry of the tensor product of comodules.
Now we check that $- \cotensor{{\mathcal H}} P$ is an equivalence of categories, using the natural transformation
$$
\eta_{\scriptscriptstyle{M}}: M \to (M \cotensor{{\mathcal H}} P) \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}},
\quad m \mapsto (m_{(0)} \cotensor{{\mathcal H}} m_{(1)+}) \cotensor{{\mathcal K}} m_{(1)-}
$$
for any right ${\mathcal H}$-comodule $M$ from \rmref{giotto2}. As above, one shows that $\eta_{\scriptscriptstyle{M}} \otimes_{\scriptscriptstyle{A}} P$ is an isomorphism by using the natural transformation $\zeta_{-}$ from \rmref{giotto1}. Explicitly, the inverse of $\eta_{\scriptscriptstyle{M}} \otimes_{\scriptscriptstyle{A}} P$ is given by
$$
\xymatrix@C=30pt{\big((M \cotensor{{\mathcal H}} P) \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}\big)\tensor{A}P \ar@{->}^-{\zeta_{\scriptscriptstyle{M \cotensor{{\mathcal H}} P}}}[r] & \ar@{->}^-{\zeta_{\scriptscriptstyle{M }}}[r] (M \cotensor{{\mathcal H}} P) \tensor{B} P & M\tensor{A}P, }
$$ where the first $\zeta$ corresponds to the left principal bundle $P^{\scriptscriptstyle{{\rm co}}}$ while the second one corresponds to $P$.
One therefore has a natural isomorphism
$$
( - \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}) \circ ( - \cotensor{{\mathcal H}} P) \overset{\simeq}{\lra} {\rm id}_{\rcomod{{\mathcal H}}}.
$$
Analogously, one obtains a natural isomorphism
$
( - \cotensor{{\mathcal H}} P) \circ ( - \cotensor{{\mathcal K}} P^{\scriptscriptstyle{{\rm co}}}) \to {\rm id}_{\rcomod{{\mathcal K}}},
$
which concludes the proof.
\end{proof}
The converse of Theorem \ref{aromanflower} will be investigated in the next section; however, we give here a partial answer when two Hopf algebroids are weakly equivalent.
\begin{proposition}
\label{prop:PHPKP}
Two flat Hopf algebroids $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are weakly equivalent if and only if there is a principal bibundle connecting them.
\end{proposition}
\begin{proof}
The implication $(\Leftarrow)$ directly follows from part {\em (iii)} of Proposition \ref{prop:Cielitomissyoutoo}. As for the opposite direction $(\Rightarrow)$, assume that there is a diagram
$$
\xymatrix@R=15pt{ & (C,{\mathcal J}) & \\ (A,{\mathcal H}) \ar@{->}^-{\B{\varphiup}}[ur] & & \ar@{->}_-{\B{\omegaup}}[ul] (B,{\mathcal K}) }
$$
of flat Hopf algebroids,
where $\B{\varphi}$ and $\B{\omegaup}$ are weak equivalences. Denote the associated trivial bundles
by $P:={\mathcal K}\tensor{\omega}C$ and $Q:=C\tensor{\varphi}{\mathcal H}$.
As shown in Proposition \ref{lemma:trivial bibundle} and explained in Remark \ref{jumpingwindow},
$P \in \bPB{{\mathcal K}}{{\mathcal J}}$ and $Q \in \bPB{{\mathcal J}}{{\mathcal H}}$ are trivial bibundles, and we can form the bundle $P\cotensor{{\mathcal J}}Q$, which is an object in $\bPB{{\mathcal K}}{{\mathcal H}}$, or equivalently $(P\cotensor{{\mathcal J}}Q)^{\scriptscriptstyle{{\rm co}}} \in \bPB{{\mathcal H}}{{\mathcal K}}$, and this finishes the proof.
\end{proof}
\subsection{Symmetric monoidal equivalence versus principal bibundles}\label{sec:tesoro} Starting with two Morita equivalent flat Hopf algebroids, the aim of this subsection is to extract from these data a principal bibundle.
To this end, let us first recall some basic facts on monoidal functors, restricting ourselves to the case of monoidal categories of comodules over flat Hopf algebroids.
Let $(A, {\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids, and assume
that there is a symmetric monoidal equivalence
$$
\mathcal{F}: \rcomod{{\mathcal H}} \longrightarrow \rcomod{{\mathcal K}}
$$
with inverse $\mathcal{G}$ in what follows. In particular, this means that there is a natural isomorphism
\begin{equation}\label{Eq:Phi}
\phiup^{\scriptscriptstyle{1}}_{-,-}: \mathcal{F}( -{\comdtensor{A}}-)\, \longrightarrow\, \mathcal{F}(-){\comdtensor{B}} \mathcal{F}(-),
\qquad \phiup^{\scriptscriptstyle{0}}: B \overset{\cong}{\longrightarrow} \mathcal{F}(A),
\end{equation}
where the latter is an algebra isomorphism. Both $\phiup^{\scriptscriptstyle{1}}$ and $\phiup^{\scriptscriptstyle{0}}$ should be compatible in a coherent way with the associativity, the commutativity ({\em i.e.}, the symmetries), and the unitary property of the tensor products of both $\rcomod{{\mathcal H}}$ and $\rcomod{{\mathcal K}}$. Notice that, in this case, there also exists a symmetric monoidal equivalence between {\em left} comodules.
The inverse natural transformation of $\phiup$ will be denoted by $\psiup$. It is known that the functor $\mathcal{G}$ is also a symmetric monoidal functor; its associated natural isomorphism can be computed from that of $\mathcal{F}$ by using the natural transformation defining the equivalence.
Now, let $M \in \bicomod{T}{{\mathcal H}}$, where $T$ is any commutative algebra, {\em i.e.}, $M$ is a $(T,A)$-bimodule and right ${\mathcal H}$-comodule with left $T$-linear coaction.
Then, we have an algebra map
$$
\lambda_{\Sf{l}}: T \to {\rcomod{{\mathcal H}}}(M, M), \quad t \mapsto \{ m \mapsto tm \},
$$
which is used to get a new algebra map
$$
\xymatrix{A \ar@{->}^-{\lambda_{\Sf{l}}}[r] & {\rcomod{{\mathcal H}}}(M, M) \ar@{->}^-{\mathcal{F}}[r] & {\rcomod{{\mathcal K}}}\big(\mathcal{F}(M), \mathcal{F}(M)\big), }
$$
from which we obtain that $\mathcal{F}(M)$ is a $(T,B)$-bimodule and that its right coaction $\rcoaction{{\mathcal K}}{\mathcal{F}(M)}$ is left $T$-linear, that is, $\mathscr{F}(M) \in \bicomod{T}{{\mathcal K}}$. Moreover, $\mathscr{F}$ is restricted to the functor $$\mathscr{F}: \bicomod{T}{{\mathcal H}} \to \bicomod{T}{{\mathcal K}}.$$
Following \cite[\S23 \& \S39.3]{BrzWis:CAC}, since $\mathcal{F} $ is right exact and commutes with inductive limits, there is a natural isomorphism over (right) modules $\Sf{Mod}_{T}$
\begin{equation}\label{Eq:Upsilon}
\Upsilon_{-,M}: \mathcal{F}(-\tensor{T}M) \longrightarrow -\tensor{T}\mathcal{F}(M),
\end{equation}
which is natural on $M$ as well, and where the functor $-\tensor{T}M: \rmod{T} \to \rcomod{{\mathcal H}}$ is defined as in \eqref{Eq:RM}. Furthermore, $\Upsilon$ defines morphisms of right ${\mathcal K}$-comodules. Notice that $\Upsilon_{\scriptscriptstyle{T,\, M}}: \mathcal{F}(M) \to T\tensor{T}\mathcal{F}(M)$ is just the canonical map sending $x \mapsto 1_{\scriptscriptstyle{T}}\tensor{T} x$.
For instance, in case $M:={\mathcal H}$ with left $A$-action given by the source $\Sf{s}$, we obtain an algebra map
$$
\lambda_{\Sf{s}}: A \to {\rcomod{{\mathcal H}}}({\mathcal H}, {\mathcal H}), \quad a \mapsto \{ u \mapsto \Sf{s}(a)u \}.
$$
The composition
$$
\xymatrix{A \ar@{->}^-{\lambda_{\Sf{s}}}[r] & {\rcomod{{\mathcal H}}}({\mathcal H}, {\mathcal H}) \ar@{->}^-{\mathcal{F}}[r] & {\rcomod{{\mathcal K}}}\big(\mathcal{F}({\mathcal H}), \mathcal{F}({\mathcal H})\big) }
$$
induces on $\mathcal{F}({\mathcal H})$ an $(A,B)$-bimodule structure with a left $A$-linear right coaction $\rcoaction{{\mathcal K}}{\mathcal{F}({\mathcal H})}$. In fact, ${\mathcal F}({\mathcal H})$ becomes an $({\mathcal H},{\mathcal K})$-bicomodule with these actions as follows.
The structure of a left ${\mathcal H}$-comodule
is given by
\begin{equation}\label{Eq:lcoact}
\xymatrix{ \lcoaction{{\mathcal H}}{\mathcal{F}({\mathcal H})}: \mathcal{F}({\mathcal H}) \ar@{->}^-{\mathcal{F}(\Delta)}[r] & \mathcal{F}({\mathcal H}\tensor{A}{\mathcal H}) \ar@{->}_-{\cong}^-{\Upsilon}[r] & {\mathcal H} \tensor{A}\mathcal{F}({\mathcal H}), }
\end{equation}
using the natural isomorphism of Eq.~\eqref{Eq:Upsilon}, which can be shown to be a morphism of right ${\mathcal K}$-comodules. Similar arguments hold true for ${\mathcal G}$. Furthermore, we have natural isomorphisms
\begin{equation}\label{Eq:FG}
{\mathcal F}\,\cong \, -\cotensor{{\mathcal H}}{\mathcal F}({\mathcal H}), \quad {\mathcal G} \,\cong \, -\cotensor{{\mathcal K}}{\mathcal G}({\mathcal K}).
\end{equation}
Since ${\mathcal H}$ is a monoid in $\Sf{Comod}_{{\mathcal H}}$, it follows that $\mathcal{F}({\mathcal H})$ is a monoid in $\Sf{Comod}_{{\mathcal K}}$. Thus, $\mathcal{F}({\mathcal H})$ is a right ${\mathcal K}$-comodule algebra with respect to the underlying algebra map
\begin{equation}\label{Eq:betaFH}
\beta: B \overset{\phiup^{\scriptscriptstyle{0}}}{\cong} \mathcal{F}(A) \overset{\mathcal{F}(\Sf{t})}{\longrightarrow} \mathcal{F}({\mathcal H}).
\end{equation}
Explicitly, the multiplication in $\mathcal{F}({\mathcal H})$ is given by
\begin{equation}\label{Eq:mFH}
\xymatrix@C=30pt{ \textsl{m}_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}: \mathcal{F}({\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H}) \ar@{->}^-{\psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}}}[r] & \mathcal{F}({\mathcal H} \comdtensor{A}{\mathcal H}) \ar@{->}^-{\mathcal{F}(\textsl{m}_{\scriptscriptstyle{{\mathcal H}}})}[r] & \mathcal{F}({\mathcal H}). }
\end{equation}
Note that $\mathcal{F}({\mathcal H})$ is commutative since $\phiup$ is so (preserves the symmetries) as well as ${\mathcal H}$.
Next, we want to endow $\mathscr{F}({\mathcal H})$ with the structure of a left ${\mathcal H}$-comodule algebra using the left comodule structure of Eq.~\eqref{Eq:lcoact}.
The $A$-algebra structure on $\mathcal{F}({\mathcal H})$ is given by the linear map
\begin{equation}\label{Eq:alpha}
\alpha: A \to \mathcal{F}({\mathcal H}), \quad a \mapsto \mathcal{F}(\lambda_{\Sf{s}}(a)) (1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}})=a. 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}},
\end{equation}
where $1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$ is just the identity element of the right ${\mathcal K}$-comodule algebra $\mathcal{F}({\mathcal H})$, which can be identified with $\mathcal{F}(\Sf{t}) \circ \phiup^{\scriptscriptstyle{0}}(1_{\scriptscriptstyle{B}})= \mathcal{F}(\Sf{t})(1_{\scriptscriptstyle{\mathcal{F}(A)}})$.
We have:
\begin{lemma}\label{lemma:mAFH}
The map $\alpha$ of Eq.~\eqref{Eq:alpha} is an algebra map. That is, there exists a map which makes the diagram
$$
\xymatrix{ \scriptstyle{\mathcal{F}({\mathcal H})\tensor{}\mathcal{F}({\mathcal H})} \ar@{->}[r] \ar@{->}[d] & \scriptstyle{\mathcal{F}({\mathcal H}) \tensor{A} \mathcal{F}({\mathcal H})} \ar@{-->}[rr] & & \scriptstyle{\mathcal{F}({\mathcal H})} \\ \scriptstyle{\mathcal{F}({\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H})} \ar@{->}^-{\scriptstyle{\psiup}}[r] & \scriptstyle{\mathcal{F}({\mathcal H} \comdtensor{A} {\mathcal H})} \ar@{->}_-{\scriptstyle{\textsl{m}_{\scriptscriptstyle{{\mathcal H}}}}}[rru] & & }
$$
commutative.
\end{lemma}
\begin{proof}
It is clear that $\alpha(1_{\scriptscriptstyle{A}}) =1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$ since $\mathcal{F}(\lambda_{\Sf{s}}(1_{\scriptscriptstyle{A}}))=id_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$. Now, for $a, a' \in A$ compute
\begin{eqnarray*}
\textsl{m}_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}\big( \alpha(a)\tensor{B}\alpha(a') \big) &=& \mathcal{F}(\textsl{m}_{\scriptscriptstyle{{\mathcal H}}}) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}}\, \Big( \mathcal{F}(\lambda_{\Sf{s}}(a)) (1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}) \tensor{B} \mathcal{F}(\lambda_{\Sf{s}}(a')) (1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}) \Big)
\\ &=& \mathcal{F}(\textsl{m}_{\scriptscriptstyle{{\mathcal H}}}) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}}\circ \big( \mathcal{F}(\lambda_{\Sf{s}}(a)) \comdtensor{B} \mathcal{F}(\lambda_{\Sf{s}}(a')) \big) \, \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}(\textsl{m}_{\scriptscriptstyle{{\mathcal H}}}) \circ \mathcal{F}\big(\lambda_{\Sf{s}}(a) \comdtensor{A} \lambda_{\Sf{s}}(a')\big) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}} \, \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}\big( \textsl{m}_{\scriptscriptstyle{{\mathcal H}}} \circ (\lambda_{\Sf{s}}(a) \comdtensor{A} \lambda_{\Sf{s}}(a')) \big) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}} \, \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}\big( \lambda_{\Sf{s}}(aa') \circ \textsl{m}_{\scriptscriptstyle{{\mathcal H}}} \big) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}} \, \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}( \lambda_{\Sf{s}}(aa')) \circ \mathcal{F}( \textsl{m}_{\scriptscriptstyle{{\mathcal H}}}) \circ \psiup_{\scriptscriptstyle{{\mathcal H},\, {\mathcal H}}} \, \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}( \lambda_{\Sf{s}}(aa')) \circ \textsl{m}_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \tensor{B} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big)
\\ &=& \mathcal{F}( \lambda_{\Sf{s}}(aa')) \big( 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}} \big) \,\,=\,\, \alpha(aa').
\end{eqnarray*}
As the last statement is obvious, this finishes the proof.
\end{proof}
In order to show that the coaction \eqref{Eq:lcoact} is an algebra map with respect to $\alpha$, we need to introduce the following natural transformations:
\begin{equation}\label{Eq:peleroja}
\begin{array}{rcl}
\B{\Omega}_{X,Y}: \big(X\tensor{A}\mathcal{F}({\mathcal H})\big) \comdtensor{B}\big(Y\tensor{A}\mathcal{F}({\mathcal H})\big) & \longrightarrow & (X\tensor{A}Y)\tensor{A}\mathcal{F}({\mathcal H}),\quad \\ (x\tensor{A}p) \comdtensor{B}(y\tensor{A}q) &\longmapsto& (x\tensor{A}y) \tensor{A} pq,
\end{array}
\end{equation}
\begin{equation}\label{Eq:pelerojasera}
\begin{array}{rcl}
\B{\nabla}_{X,Y}: \big(X\tensor{A}{\mathcal H}\big) \comdtensor{A}\big(Y\tensor{A}{\mathcal H}\big) & \longrightarrow& (X\tensor{A}Y)\tensor{A}{\mathcal H}, \quad \\ (x\tensor{A}u) \comdtensor{A}(y\tensor{A}v) &\longmapsto& (x\tensor{A}y) \tensor{A} uv,
\end{array}
\end{equation}
where $X$ and $Y$ are $A$-modules and where we used the multiplication in $\mathcal{F}({\mathcal H})$.
Using a functor similar to the one in \eqref{Eq:RM}, one sees that $\B{\Omega}$ defines morphisms of right ${\mathcal K}$-comodules
since the right ${\mathcal K}$-coaction of $\mathcal{F}({\mathcal H})$ is left $A$-linear (with respect to the $A$-action given by $\alpha$). Analogously, $\B{\nabla}$ defines morphisms of right ${\mathcal H}$-comodules. These natural transformations are compatible in the following way:
\begin{proposition}\label{prop:Dstar}
The diagram
\begin{small}
$$
\xymatrix@R=18pt@C=18pt{ & \scriptstyle{\mathcal{F}\big((X\tensor{A}{\mathcal H}) \comdtensor{A} (Y\tensor{A}{\mathcal H}) \big)} \ar@{->}^-{\scriptstyle{\mathcal{F}(\B{\nabla})}}[rr]
& & \scriptstyle{\mathcal{F}((X\tensor{A}Y)\tensor{A}{\mathcal H})} \ar@{->}^-{\scriptstyle{\Upsilon}}[dd] \\ \scriptstyle{\mathcal{F}(X\tensor{A}{\mathcal H}) \comdtensor{B} \mathcal{F}(Y\tensor{A}{\mathcal H})} \ar@{->}_-{ \scriptstyle{\Upsilon\comdtensor{B} \Upsilon}}[dr] \ar@{->}^-{\scriptstyle{\psiup}}[ru] & & & \\ & \scriptstyle{\big(X\tensor{A}\mathcal{F}({\mathcal H})\big) \comdtensor{B}\big(Y\tensor{A}\mathcal{F}({\mathcal H})\big)} \ar@{->}^-{\scriptstyle{\B{\Omega}}}[rr] & & \scriptstyle{(X\tensor{A}Y)\tensor{A}\mathcal{F}({\mathcal H})} }
$$
\end{small}
of right ${\mathcal K}$-comodules commutes.
\end{proposition}
\begin{proof}
First, notice that both $\Upsilon \circ \mathcal{F}(\B{\nabla}) \circ \psiup$ and $\B{\Omega} \circ (\Upsilon\comdtensor{B} \Upsilon)$ are natural transformations on $(X,Y)$. Now, up to the canonical isomorphisms $A\tensor{A}{\mathcal H} \cong {\mathcal H}$ and $A\tensor{A}\mathcal{F}({\mathcal H}) \cong \mathcal{F}({\mathcal H}) $, we see that the diagram commutes for $X:=A$ and $Y:=A$ as this is just the definition of the multiplication $\textsl{m}_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$ defined in \eqref{Eq:mFH}. Using the naturality of both paths in the diagram, one can also show that the diagram commutes when $X$ and $Y$ are free $A$-modules of finite rank. Since the
involved functors commute with direct sums, the same holds true when $X$ and $Y$ are free $A$-modules. Lastly, since all involved functors
are right exact, one can use free representations of any $A$-module to complete the proof.
\end{proof}
\begin{proposition}\label{prop:FHcomdalg}
The pair $(\mathcal{F}({\mathcal H}), \alpha)$ is a left ${\mathcal H}$-comodule algebra with respect to the coaction \eqref{Eq:lcoact}.
\end{proposition}
\begin{proof}
We need to check that the map
$\B{\lambda}= \Upsilon \circ \mathcal{F}(\Delta)$
in \eqref{Eq:lcoact}
is an algebra map. First, we prove unitality, that is, $\B{\lambda}(1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}})= \B{\lambda}(\alpha(1_{\scriptscriptstyle{A}})) =1_{\scriptscriptstyle{{\mathcal H}}} \tensor{A} 1_{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$: this follows from the commutative diagram
\begin{small}
$$
\xymatrix{ \scriptstyle{\mathcal{F}(A)} \ar@{->}^-{\mathcal{F}(\Sf{t})}[rr] \ar@{->}_-{\mathcal{F}(\Sf{t})}[d] & & \scriptstyle{\mathcal{F}(A)} \ar@{->}^-{\mathcal{F}(\Delta)}[d] \\ \scriptstyle{\mathcal{F}({\mathcal H})} \ar@{->}^-{\mathcal{F}(\Sf{t}\tensor{A}{\mathcal H})}[rr] \ar@{->}_-{\Upsilon}[d] & & \scriptstyle{\mathcal{F}({\mathcal H}\tensor{A}{\mathcal H})} \ar@{->}^-{\Upsilon}[d]
\\ \scriptstyle{A\tensor{A} \mathcal{F}({\mathcal H})} \ar@{->}^-{\Sf{t}\tensor{A}\mathcal{F}({\mathcal H})}[rr] & & \scriptstyle{{\mathcal H}\tensor{A}\mathcal{F}({\mathcal H})}, }
$$
\end{small}
where the left hand side $\Upsilon$ is just the canonical map $y \mapsto 1_{\scriptscriptstyle{A}}\tensor{A}y$.
Now we proceed to check that $\B{\lambda}$ is multiplicative. To this end, we show that the diagram
\begin{small}
$$
\xymatrix@C=45pt{ \scriptstyle{\mathcal{F}({\mathcal H})\tensor{} \mathcal{F}({\mathcal H})} \ar@{->}_-{\scriptstyle{}}[d] \ar@{->}^-{\scriptstyle{}}[r] & \scriptstyle{\mathcal{F}({\mathcal H})\comdtensor{A}\mathcal{F}({\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\textsl{m}}}[rr] & & \scriptstyle{\mathcal{F}({\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\mathcal{F}(\Delta)}}[dd] \\ \scriptstyle{\mathcal{F}({\mathcal H})\comdtensor{B}\mathcal{F}({\mathcal H})} \ar@{->}|-{\scriptscriptstyle{\mathcal{F}(\Delta) \comdtensor{B}\mathcal{F}(\Delta) }}[dd] \ar@{->}^-{\scriptscriptstyle{\psiup}}[r] & \scriptstyle{\mathcal{F}({\mathcal H}\comdtensor{A}{\mathcal H})} \ar@{->}|-{\scriptscriptstyle{\mathcal{F}(\Delta\comdtensor{A}\Delta) }}[dd] \ar@{->}|-{\scriptscriptstyle{\mathcal{F}(\B{\rho})}}[rd] \ar@{->}|-{\scriptscriptstyle{\mathcal{F}(\textsl{m})}}[rru] & & \\ & & \scriptstyle{\mathcal{F}\big( ({\mathcal H}\comdtensor{A}{\mathcal H}) \tensor{A}{\mathcal H}\big)} \ar@{->}^-{\scriptscriptstyle{\Upsilon}}[ddd] \ar@{->}^-{\scriptscriptstyle{\mathcal{F}(\textsl{m}\,\tensor{A}{\mathcal H}) }}[r] & \scriptstyle{\mathcal{F}({\mathcal H}\comdtensor{A}{\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\Upsilon}}[ddd] \\ \scriptstyle{\mathcal{F}({\mathcal H}\tensor{A}{\mathcal H}) \comdtensor{B} \mathcal{F}({\mathcal H}\tensor{A}{\mathcal H})} \ar@{->}|-{\scriptscriptstyle{\Upsilon \comdtensor{B}\Upsilon }}[dd] \ar@{->}^-{\scriptscriptstyle{\psiup}}[r] & \scriptstyle{ \mathcal{F}\big( ({\mathcal H}\tensor{A}{\mathcal H}) \comdtensor{A} ({\mathcal H}\tensor{A}{\mathcal H})\big) } \ar@{->}|-{\scriptscriptstyle{\mathcal{F}(\B{\nabla})}}[ru] & & \\ & & & \\ \scriptstyle{\big({\mathcal H}\tensor{A}\mathcal{F}({\mathcal H})\big) \comdtensor{B}\big({\mathcal H}\tensor{A}\mathcal{F}({\mathcal H})\big) } \ar@{->}^-{\scriptscriptstyle{\B{\Omega}}}[rr] & & \scriptstyle{ ({\mathcal H}\comdtensor{A}{\mathcal H})\tensor{A}\mathcal{F}({\mathcal H}) } \ar@{->}^-{\scriptscriptstyle{\textsl{m}\,\tensor{A}\mathcal{F}({\mathcal H}) }}[r] & \scriptstyle{{\mathcal H}\tensor{A}\mathcal{F}({\mathcal H})}}
$$
\end{small}
is commutative, which follows from Lemma \ref{lemma:mAFH}, Proposition \ref{prop:Dstar}, as well as from the very definitions of all involved maps and natural transformations.
\end{proof}
Our next aim is to show that $\mathcal{F}({\mathcal H})$ is a principal left $({\mathcal H}, {\mathcal K})$-bundle with respect to $\alpha$ and $\beta$. As a start, the subsequent lemma concerns the faithfully flatness.
\begin{lemma}\label{lemma:fplano}
Assume that there is a symmetric monoidal equivalence $$\mathcal{F}: \Sf{Comod}_{{\mathcal H}} \to \Sf{Comod}_{{\mathcal K}}$$ with inverse $\mathcal{G}$. Then, for every right ${\mathcal H}$-comodule $M$ whose underlying $A$-module is faithfully flat,
$\mathcal{F}(M)$ is a faithfully flat $B$-module.
\end{lemma}
\begin{proof}
One can easily check that there is a natural isomorphism
$$
\mathcal{O}_{\scriptscriptstyle{{\mathcal K}}}(-) \tensor{B} \mathcal{F}(M) \overset{\cong}{\longrightarrow} \mathcal{F}\big( \mathcal{G}(-)\otimes^{\scriptscriptstyle{A}} M \big),
$$
where $\mathcal{O}_{\scriptscriptstyle{{\mathcal K}}}: \rcomod{{\mathcal K}} \to \rmod{B}$ denotes the forgetful functor.
Hence, $\mathcal{O}_{\scriptscriptstyle{{\mathcal K}}}(-) \tensor{B} \mathcal{F}(M)$ is a faithful and exact functor. Using the fact that $\mathcal{F}(M)$ carries the structure of a left ${\mathcal K}$-comodule (in fact its opposite comodule), we see that $-\tensor{B}\mathcal{F}(M)$ is a faithful and exact functor.
\end{proof}
With the help of this lemma we can state:
\begin{proposition}\label{prop:FH}
The triple $(\mathcal{F}({\mathcal H}), \alpha,\beta)$ forms a left principal $({\mathcal H},{\mathcal K})$-bundle.
\end{proposition}
\begin{proof}
From Proposition \ref{prop:FHcomdalg} follows that $(\mathcal{F}({\mathcal H}), \alpha)$ is a left ${\mathcal H}$-comodule algebra. Therefore, $(\mathcal{F}({\mathcal H}), \alpha,\beta)$ is an $({\mathcal H},{\mathcal K})$-bicomodule algebra since $(\mathcal{F}({\mathcal H}),\beta)$ is a right ${\mathcal K}$-comodule algebra.
As $\thopf{{\mathcal H}}$ is faithfully flat, $\mathcal{F}({\mathcal H})_B$ is, using Lemma \ref{lemma:fplano}, also faithfully flat and therefore $\beta$ is a faithfully flat extension.
To complete the proof, we need to check that the canonical map
$$
\xymatrix@C=40pt{\can{{\mathcal H}}{\mathcal{F}({\mathcal H})}: \mathcal{F}({\mathcal H})\comdtensor{B}\mathcal{F}({\mathcal H}) \ar@{->}^-{\scriptstyle{\B{\lambda}\comdtensor{B}\mathcal{F}({\mathcal H})}}[r] & {\mathcal H} \tensor{A} \mathcal{F}({\mathcal H}) \comdtensor{B} \mathcal{F}({\mathcal H}) \ar@{->}^-{\scriptstyle{{\mathcal H} \tensor{A} \, \textsl{m}}}[r] & {\mathcal H} \tensor{A} \mathcal{F}({\mathcal H}) }
$$
is bijective. To this end, using Eqs.~ \eqref{Eq:lcoact} and \eqref{Eq:mFH} to express the coaction and the multiplication in ${\mathcal F}({\mathcal H})$, we write down the map $\can{\scriptscriptstyle{{\mathcal H}}}{\scriptscriptstyle{\mathcal{F}({\mathcal H})}}$ in the diagram
$$
\xymatrix@R=30pt@C=40pt{ \scriptstyle{\mathcal{F}({\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H})} \ar@{->}_-{\scriptscriptstyle{\psiup}}[d] \ar@{->}^-{\scriptscriptstyle{\mathcal{F}(\Delta)\comdtensor{B} \mathcal{F}({\mathcal H})}}[rr] & & \scriptstyle{\mathcal{F}({\mathcal H}\tensor{A}{\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\Upsilon\comdtensor{B} \mathcal{F}({\mathcal H})}}[rr] \ar@{->}_-{\scriptscriptstyle{\psiup}}[d] & & \scriptstyle{{\mathcal H}\tensor{A}\mathcal{F}({\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H})} \ar@{->}_-{\scriptscriptstyle{{\mathcal H}\tensor{A}\psiup}}[d] \\ \scriptstyle{\mathcal{F}({\mathcal H}\comdtensor{A} {\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\mathcal{F}(\Delta\tensor{A}{\mathcal H})}}[rr] \ar@/_1pc/_-{\scriptscriptstyle{\mathcal{F}(\can{{\mathcal H}}{{\mathcal H}})}}[rrd] & & \scriptstyle{\mathcal{F}\big( ({\mathcal H}\tensor{A}{\mathcal H})\comdtensor{A} {\mathcal H}\big) = \mathcal{F}\big( {\mathcal H}\tensor{A}({\mathcal H}\comdtensor{A} {\mathcal H})\big) } \ar@{->}^-{\scriptscriptstyle{\Upsilon}}[rr] \ar@{->}_-{\scriptscriptstyle{\mathcal{F}({\mathcal H}\tensor{A}\,\textsl{m})}}[d] & & \scriptstyle{{\mathcal H}\tensor{A}\mathcal{F}({\mathcal H}\comdtensor{A} {\mathcal H})} \ar@{->}_-{\scriptscriptstyle{{\mathcal H}\tensor{A}\mathcal{F}(\textsl{m})}}[d] \\ & & \scriptstyle{\mathcal{F}({\mathcal H}\tensor{A} {\mathcal H})} \ar@{->}^-{\scriptscriptstyle{\Upsilon}}[rr] & & \scriptstyle{{\mathcal H}\tensor{A} \mathcal{F}({\mathcal H})}. }
$$
Once shown that this diagram is commutative, it follows that the canonical map for ${\mathcal F}({\mathcal H})$ is bijective
as $\can{{\mathcal H}}{\mathcal{F}({\mathcal H})} = \Upsilon \circ \mathcal{F}(\can{{\mathcal H}}{{\mathcal H}}) \circ \psiup$, where $\can{{\mathcal H}}{{\mathcal H}}$ is bijective being
the canonical map of the unit bundle $\mathscr{U}({\mathcal H})$.
To check that the above diagram is commutative, one only needs to show the commutativity of the rectangle in the upper right. This, in fact, forms part of the well-known properties of the natural transformation $\Upsilon$; for the sake of completeness, we explain how this works: to start with, denote by $\mathcal{T}, \mathcal{S}: \rmod{A} \to \rcomod{{\mathcal K}}$ the functors
$$
\mathcal{T}(X)\,=\, \mathcal{F}(X\tensor{A}{\mathcal H})\comdtensor{B} \mathcal{F}({\mathcal H}),\quad \mathcal{S}(X) \,=\, \mathcal{F}\big( X\tensor{A}({\mathcal H}\comdtensor{A}{\mathcal H})\big).
$$
Clearly, $\psiup_{(-\tensor{A}{\mathcal H}),\,{\mathcal H}}: \mathcal{T} \to \mathcal{S}$ is a natural transformation. Since $\mathcal{T}$ and $\mathcal{S}$ commute with direct limits,
we have for every $A$-module $X$:
$$ (X\tensor{A}\psiup_{(A\tensor{A}{\mathcal H}),\,{\mathcal H}}) \circ \Upsilon^{\scriptscriptstyle{\mathcal{T}}}_X \,\, =\,\, \Upsilon^{\scriptscriptstyle{\mathcal{S}}}_X \circ \psiup_{(X\tensor{A}{\mathcal H}),\,{\mathcal H}}.
$$
Using this equality for $X:=A$, we deduce the claim since $\Upsilon^{\scriptscriptstyle{\mathcal{T}}}_X=
\Upsilon^{\scriptscriptstyle{\mathcal{F}}}_X\comdtensor{B}\mathcal{F}({\mathcal H})$ holds.
\end{proof}
\begin{corollary}
\label{coro:FHviaPo}
Let $(D, {\mathcal I})$ be another flat Hopf algebroid. Then the functor $\mathcal{F}$ restricts to a functor
$$
\mathcal{F}: \lPB{{\mathcal I}}{{\mathcal H}} \longrightarrow \lPB{{\mathcal I}}{{\mathcal K}}.
$$
\end{corollary}
\begin{proof}
By Proposition \ref{prop:FH}, the triple $({\mathcal F}({\mathcal H}), \alpha,\beta)$
defines a principal left $({\mathcal H}, {\mathcal K})$-bundle; the cotensor product
$(R\cotensor{{\mathcal H}}\mathcal{F}({\mathcal H}), \td{\delta}, \td{\beta})$, where $(R, \delta, \omega)$ is a principal left $({\mathcal I}, {\mathcal H})$-bundle, yields as in Lemma \ref{lemma:cotensorpb} a principal left $({\mathcal I}, {\mathcal K})$-bundle.
Then, the first natural isomorphism of Eq.~\eqref{Eq:FG} leads to $R\cotensor{{\mathcal H}}\mathcal{F}({\mathcal H}) \cong \mathcal{F}(R)$, which is an isomorphism of $({\mathcal I}, {\mathcal K})$-bicomodules, and this proves the claim.
\end{proof}
The following proposition (mentioned in Figure \ref{ahaaha}
in the Introduction) shows that two Morita equivalent Hopf algebroids are connected by a principal bibundle.
\begin{proposition}
\label{prop:llegaras}
Let $(A,{\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids.
Assume that there is a symmetric monoidal equivalence of categories $\mathcal{F}: \Sf{Comod}_{{\mathcal H}} \to \Sf{Comod}_{{\mathcal K}}$ with inverse $\mathcal{G}$. Then $(\mathcal{F}({\mathcal H}),\alpha,\beta)$ is a principal $({\mathcal H},{\mathcal K})$-bibundle whose opposite bundle is $\mathcal{G}({\mathcal K})$.
\end{proposition}
\begin{proof}
Set $P:=\mathcal{F}({\mathcal H})$ and $Q:=\mathcal{G}({\mathcal K})$. From Proposition \ref{prop:FH} follows that $(P,\alpha, \beta)$ is a left principal $({\mathcal H},{\mathcal K})$-bundle. Interchanging $\mathcal{F}$ with $\mathcal{G}$, we also obtain that $(\mathcal{G}, \sigma, \theta)$ is
a left principal $({\mathcal K},{\mathcal H})$-bundle, where $\theta: A\cong \mathcal{G}(B) \to \mathcal{G}({\mathcal K})$, and $\sigma$ is constructed in the same way as was $\alpha$.
On the other hand, using the equivalences $\mathcal{F}$ and $\mathcal{G}$ together with the natural transformations
$$
\mathcal{F} \, \cong\, -\cotensor{{\mathcal H}}P, \quad
\mathcal{G} \, \cong\, -\cotensor{{\mathcal K}}Q,
$$
of Eq.~\eqref{Eq:FG}, we obtain the isomorphisms
$$
P\cotensor{{\mathcal K}} Q \cong \mathscr{U}({\mathcal H}),\qquad Q\cotensor{{\mathcal H}} P \cong \mathscr{U}({\mathcal K})
$$
of ${\mathcal H}$ and ${\mathcal K}$-bicomodules, respectively, which fulfil the triangle properties \eqref{Eq:trianglI} and \eqref{Eq:trianglII}.
This implies that $(P,\alpha,\beta)$ is an invertible $1$-cell in the category $\mathsf{PB}^\ell({\mathcal H},{\mathcal K})$ of principal left bundles.
Now, conclude the proof by making use of Proposition \ref{prop:el parte} {\em (i)}.
\end{proof}
In a summarising manner, we can state the main theorem of this article:
\begin{theorem}
\label{thm:MAIN}
Let $(A,{\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids. The following are equivalent:
\begin{enumerate}[(a)]
\item $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are Morita equivalent.
\item There is a principal bibundle connecting $(A,{\mathcal H})$ and $(B,{\mathcal K})$.
\item $(A,{\mathcal H})$ and $(B,{\mathcal K})$ are weakly equivalent.
\end{enumerate}
\end{theorem}
\begin{proof}
The implication $(a) \Rightarrow (b)$ is Proposition \ref{prop:llegaras}, whereas the implication $(b) \Rightarrow (c)$ is contained in Proposition \ref{prop:PHPKP}. Finally, the step $(c) \Rightarrow (a)$ is obvious from the very definitions.
\end{proof}
\begin{rem}
As mentioned in Figure \ref{ahaaha}
in the Introduction, Theorem \ref{aromanflower} also states the implication $(b) \Rightarrow (a)$, whereas Proposition \ref{prop:PHPKP} moreover yields $(c) \Rightarrow (b)$.
\end{rem}
\subsection{The categorical group of monoidal symmetric auto-equivalences}
\label{ssec:autoequi}
In this subsection,
we combine the results of Theorems \ref{thm:MAIN} and \ref{aromanflower} by taking a single flat Hopf algebroid. More precisely, we show that all symmetric monoidal auto-equivalences of the category of right ${\mathcal H}$-comodules form a categorical group with morphisms given by natural tensor transformations,
and conclude that
this group
is equivalent to the categorical group of principal bibundles.
Denote by $\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H})$ the category of monoidal symmetric auto-equivalences of
the category of (right) comodules $\rcomod{{\mathcal H}}$ over a flat Hopf algebroid $(A,{\mathcal H})$.
Morphisms in this category are \emph{natural tensor transformations}, that is, natural transformations $\theta: {\mathscr F} \to {\mathscr F}'$ such that the diagrams
\begin{equation}
\label{Eq:BlackandWhite}
\xymatrix@C=45pt{ {\mathscr F}\big(X\comdtensor{A} Y\big) \ar@{->}^-{\Theta_{\Sscript{X\comdtensor{A} Y}}}[r] \ar@{->}^-{\cong}_-{\phiup^{\Sscript{1,\, {\mathscr F}}}}[d] & {\mathscr F}'\big(X\comdtensor{A}Y\big) \ar@{->}^-{\phiup^{\Sscript{1,\, {\mathscr F}'}}}_-{\cong}[d] \\
{\mathscr F}(X)\comdtensor{B} {\mathscr F}(Y) \ar@{->}^-{\Theta_{\Sscript{X}} \comdtensor{B} \Theta_{\Sscript{Y}}}[r] & {\mathscr F}'(X)\comdtensor{B}{\mathscr F}'(Y) } \qquad \xymatrix@C=25pt{ {\mathscr F}(A) \ar@{->}^-{\Theta_{\Sscript{A}}}[rr] & & {\mathscr F}'(A) \\
& B \ar@{->}^-{\cong}_-{\phiup^{\Sscript{0,\, {\mathscr F}}}}[ur] \ar@{->}^-{\phiup^{\Sscript{0,\,{\mathscr F}'}}}_-{\cong}[ul] & }
\end{equation}
commute.
Note that this gives a sets-category (in the sense that homomorphisms between two objects form a set) as $\rcomod{{\mathcal H}}$ is a Grothendieck category and the involved functors preserve inductive limits. The category $\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H})$ is itself
a monoidal category with multiplication given by
the composition of functors and identity object given
by the identity equivalence ${\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}$.
On the other hand, as in Subsection \ref{ssec:ppb}, we are interested in the monoidal category $\big(\bPB{{\mathcal H}}{{\mathcal H}},\cotensor{{\mathcal H}}, \mathscr{U}({\mathcal H}) \big)$. Both categories are in fact \emph{categorical groups} (more precisely, a \emph{$2$-group} and a \emph{bigroup})
and are equivalent as such.
\begin{proposition}
\label{prop:pbmaps}
Let $(A,{\mathcal H})$ and $(B,{\mathcal K})$ be two flat Hopf algebroids, ${\mathscr F},{\mathscr F}': \rcomod{{\mathcal H}} \to \rcomod{{\mathcal K}}$ two symmetric monoidal equivalences, and
$\Theta: {\mathscr F} \to {\mathscr F}'$ a natural tensor transformation.
Then $\Theta_{{\mathcal H}}: {\mathscr F}({\mathcal H}) \to {\mathscr F}'({\mathcal H})$ is a morphism of principal $({\mathcal H},{\mathcal K})$-bibundles. In particular, $\Theta$ is a natural isomorphism and consequently $\big( \Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big)$ is a categorical group.
\end{proposition}
\begin{proof}
By definition, $\Theta_{{\mathcal H}}$ is a morphism of right ${\mathcal K}$-comodule algebras. Let us check that it is also a morphism of left ${\mathcal H}$-comodule algebras. Recall that the respective left comodule algebra structure of both ${\mathscr F}({\mathcal H})$ and ${\mathscr F}'({\mathcal H})$ is given as in Proposition \ref{prop:FHcomdalg}.
That $\Theta_{{\mathcal H}}$ is left ${\mathcal H}$-colinear follows from the following diagram:
\begin{small}
$$
\xymatrix@C=40pt@R=15pt{ & {\mathscr F}\big({\mathcal H}\tensor{A} {\mathcal H}\big) \ar@{->}^-{{\mathscr F}(\Delta)}[rr] \ar@{->}^-{\Theta_{\Sscript{{\mathcal H}\tensor{A}{\mathcal H}}}}[ddd] & & {\mathscr F}({\mathcal H}) \ar@{->}^-{\Theta_{\Sscript{{\mathcal H}}}}[ddd] \\ {\mathcal H}\tensor{A}{\mathscr F}({\mathcal H}) \ar@/_1pc/@{->}^-{\B{\lambda}}[rrru] \ar@{->}_-{\cong}^-{\Upsilon^{\Sscript{{\mathscr F}}}}[ru] \ar@{->}_-{{\mathcal H}\tensor{A}\Theta_{\Sscript{{\mathcal H}}}}[ddd] & & & & \\ & & & & \\
& {\mathscr F}'({\mathcal H}\tensor{A}{\mathcal H}) \ar@{->}^-{{\mathscr F}'(\Delta)}[rr] & & {\mathscr F}'({\mathcal H}) \\ {\mathcal H}\tensor{A}{\mathscr F}'({\mathcal H}) \ar@/_1pc/@{->}^-{\B{\lambda}}[rrru] \ar@{->}_-{\cong}^-{\Upsilon^{\Sscript{{\mathscr F}'}}}[ru] & & & & }
$$
\end{small}
where the left hand square is commutative by the universal property of the natural isomorphism $\Upsilon$. The $A$-algebra structure of ${\mathscr F}({\mathcal H})$ is given by the algebra map $\alpha^{\Sscript{{\mathscr F}({\mathcal H})}}: A \to {\mathscr F}({\mathcal H})$, $a \mapsto {\mathscr F}(\lambda_{\Sscript{\Sf{s}}}(a))(1_{\Sscript{{\mathscr F}({\mathcal H})}})$, and similarly for ${\mathscr F}'({\mathcal H})$, see Eq.~\eqref{Eq:alpha}.
Thus, for any $a \in A$, we have
$$
\Theta_{\Sscript{{\mathcal H}}} \circ \alpha^{\Sscript{{\mathscr F}({\mathcal H})}}(a)\,=\, \Theta_{\Sscript{{\mathcal H}}} \circ {\mathscr F}(\lambda_{\Sscript{\Sf{s}}}(a))(1_{\Sscript{{\mathscr F}({\mathcal H})}}) \,=\, {\mathscr F}'(\lambda_{\Sscript{\Sf{s}}}(a)) \circ \Theta_{\Sscript{{\mathcal H}}}(1_{\Sscript{{\mathscr F}({\mathcal H})}}) \,=\, {\mathscr F}'(\lambda_{\Sscript{\Sf{s}}}(a))(1_{\Sscript{{\mathscr F}'({\mathcal H})}})\,=\, \alpha^{\Sscript{{\mathscr F}'({\mathcal H})}}(a)
$$
since $\Theta_{\Sscript{{\mathcal H}}}$ is a $B$-algebra map.
Therefore, $\Theta_{\Sscript{{\mathcal H}}}$ is an $A$-algebra map as it is multiplicative, and this finishes the proof of the first statement.
Now, by Lemma \ref{lemma:ISO}, $\Theta_{\Sscript{{\mathcal H}}}$ is an isomorphism and this suffices to show that $\Theta$ is a natural isomorphism:
using the natural isomorphisms given in Eqs.~\eqref {Eq:Upsilon} and \eqref{Eq:FG}, one can see that the diagram
$$
\xymatrix@C=40pt{ {\mathscr F} \ar@{->}^-{\Theta}[rr] \ar@{->}_-{\cong}[d] & & {\mathscr F}' \ar@{->}^-{\cong}[d] \\ -\cotensor{{\mathcal H}}{\mathscr F}({\mathcal H}) \ar@{->}_-{-\cotensor{{\mathcal H}}\Theta_{\Sscript{{\mathcal H}}}}[rr] & & -\cotensor{{\mathcal H}}{\mathscr F}'({\mathcal H}) }
$$
of natural transformations
commutes, which means that $\Theta$ is a natural isomorphism.
\end{proof}
The following is Theorem \ref{thm:B} in the Introduction:
\begin{theorem}
\label{them:catggrp}
The functors
\begin{equation*}
\begin{array}{rclrcl}
\big(\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big)
&\longrightarrow&
\big(\bPB{{\mathcal H}}{{\mathcal H}},\cotensor{{\mathcal H}}, \mathscr{U}({\mathcal H}) \big), &
{\mathscr F} &\longmapsto& {\mathscr F}({\mathcal H})
\\
\big(\bPB{{\mathcal H}}{{\mathcal H}},\cotensor{{\mathcal H}}, \mathscr{U}({\mathcal H}) \big) &\longrightarrow&
\big(\Sf{Aut}^{\scriptscriptstyle{\otimes}}(A,{\mathcal H}), \B{\circ},{\rm id}_{\scriptscriptstyle{\rcomod{{\mathcal H}}}}\big), & (P,\alpha,\beta) &\longmapsto& -\cotensor{{\mathcal H}}P
\end{array}
\end{equation*}
establish a monoidal equivalence of categorical groups.
\end{theorem}
\begin{proof}
This essentially follows from Proposition \ref{prop:pbmaps}, Theorems \ref{thm:MAIN} and \ref{aromanflower}, in combination with Corollary \ref{coro:twoJapanese}.
\end{proof}
|
\section{Introduction}
The Alternating Direction Method of Multipliers (ADM or ADMM) is a very popular algorithm with wide applications in signal and image processing, machine learning, statistics, compressive sensing, and operations
research. Combined with problem reformulation tricks, the method can reduce a complicated problem into much simpler subproblems.
The vanilla ADM applies to a linearly-constrained problem with separable convex objective functions in the following ``ADM-ready" form:
\begin{align}\tag{P1}\label{for:P3-ADMM}
\left\{\begin{array}{ll}\Min\limits_{\vx,\vy} \quad& f(\vx)+g(\vy)\\
\st \quad& \vA\vx+\vB\vy=\vb,\end{array}\right.
\end{align}
where functions $f,g$ are proper, closed, convex but not necessarily differentiable. ADM reduces~\eqref{for:P3-ADMM} into two simpler subproblems and then iteratively updates $\vx$, $\vy$, as well as a multiplier (dual) variable $\vz$.
Given $(\vx^k,\vy^k,\vz^k)$, ADM generates $(\vx^{k+1},\vy^{k+1},\vz^{k+1})$ as follows
\begin{enumerate}
\item
$\vy^{k+1} \in \argmin\limits_\vy g(\vy) +(2\lambda)^{-1}\|\vA\vx^k+\vB\vy-\vb+\lambda\vz^k\|_2^2,$\\[-15pt]
\item $\vx^{k+1} \in \argmin\limits_\vx f(\vx)+(2\lambda)^{-1}\|\vA\vx+\vB\vy^{k+1}-\vb+\lambda\vz^k\|_2^2,$\\[-15pt]
\item $\vz^{k+1} = \vz^k+\lambda^{-1}(\vA\vx^{k+1}+\vB\vy^{k+1}-\vb),$
\end{enumerate}
where $\lambda>0$ is a fixed parameter.
We use ``$\in$'' since the subproblems do not necessarily have unique solutions.
Since $\{f,\vA,\vx\}$ and $\{g,\vB,\vy\}$ are in symmetric positions in \eqref{for:P3-ADMM}, swapping them does not change the problem. This corresponds to switching the order that $\vx$ and $\vy$ are updated in each iteration. But, since the variable updated first is used in the updating of the other variable, this swap leads to a different sequence of variables and thus a different algorithm.
Note that the order switch does not change the per-iteration cost of ADM. Also note that one, however, cannot mix the two update orders at different iterations because it will generally cause divergence, even when the primal-dual solution to \eqref{for:P3-ADMM} is unique.
\subsection{ADM works in many different ways}\label{sec:11}
In spite of its popularity and vast literature, there are still simple unanswered questions about ADM: how many ways can ADM be applied? and which ways work better? Before answering these questions, let us examine the following problem, to which we can find \textbf{twelve different ways to apply ADM:} \beq\label{uv}\Min_{\vx}~ u(\vx) + v(\vC\vx),
\eeq
where $u$ and $v$ are proper, closed, convex functions and $\vC$ is a linear mapping. Problem \eqref{uv} generalizes a large number of signal and image processing, inverse problem, and machine learning models.
We shall reformulate
\eqref{uv} into the form of \eqref{for:P3-ADMM}. By introducing dummy variables in two different ways, we obtain two ADM-ready formulations of problem \eqref{uv}:
\begin{align}\label{uveq}
\left\{\begin{array}{ll}\Min\limits_{\vx,\vy} \quad& u(\vx)+v(\vy)\\
\st \quad& \vC\vx-\vy=0\end{array} \right.\quad\mbox{and}
&\qquad\left\{\begin{array}{ll}\Min\limits_{\vx,\bar{\vy}} \quad& u(\vx)+v(\vC\bar{\vy})\\
\st \quad& \vx-\bar{\vy}=0.\end{array} \right.
\end{align}
In addition, we can derive the dual problem of \eqref{uv}:
\beq\label{uvdual}
\Min_\vv~ u^*(-\vC^* \vv) + v^*(\vv),
\eeq
where $u^*,v^*$ are the convex conjugates (i.e., Legendre transforms) of functions $u,v$, respectively, $\vC^*$ is the adjoint of $\vC$, and $\vv$ is the dual variable. (The steps to derive \eqref{uvdual} from \eqref{uv} are standard and thus omitted.) Then, we also reformulate \eqref{uvdual} into two ADM-ready forms, which use different dummy variables:
\begin{align}\label{uvdualeq}
\left\{\begin{array}{ll}\Min\limits_{\vu,\vv} \quad& u^*(\vu)+v^*(\vv)\\
\st \quad& \vu+\vC^* \vv=0\end{array} \right.\quad\mbox{and}
&\qquad\left\{\begin{array}{ll}\Min\limits_{\bar\vu,\vv} \quad& u^*(\vC^*\bar{\vu})+v^*(\vv)\\
\st \quad& \bar{\vu}+\vv=0.\end{array} \right.
\end{align}
Clearly, ADM can be applied to all of the four formulations in \eqref{uveq} and \eqref{uvdualeq}, and including the update order swaps, there are \emph{eight different ways} to apply ADM.
Under some technical conditions such as the existence of saddle-point solutions, all the eight ADM will converge to a saddle-point solution or solutions for problem \eqref{uv}. In short, they all work.
It is worth noting that by the Moreau identity, the subproblems involving $u^*$ and $v^*$ can be easily reduced to subproblems involving $u$ and $v$, respectively. No significant computing is required.
The two formulations in \eqref{uveq}, however, lead to significantly different ADM subproblems. In the ADM applied to the left formulation, $u$ and $\vC$ will appear in one subproblem and $v$ in the other subproblem. To the right formulation, $u$ will be alone while $v$ and $\vC$ will appear in the same subproblem. This difference applies to the two formulations in \eqref{uvdualeq} as well. It depends on the structures of $u,v,\vC$ to determine the better choices. Therefore, out of the eight, four will have (more) difficult subproblems than the rest.
There are \emph{another four ways} to apply ADM to problem \eqref{uv}. Every one of them will have three subproblems that separately involve $u,v,\vC$, so they are all different from the above eight.
To get the first two, let us take the left formulation in \eqref{uveq} and introduce a dummy variable $\vs$, obtaining a new equivalent formulation
\beq\label{xyz} \left\{\begin{array}{ll}\Min\limits_{\vx,\vy,\vs} \quad& u(\vs)+v(\vy)\\
\st \quad& \vC\vx-\vy=0,\\
& \hspace{8pt}\vx-\vs\,=0. \end{array} \right.
\eeq
It turns out that the same ``dummy variable'' trick applied to the right formulation in \eqref{uveq} also gives \eqref{xyz}, up to a change of variable names. Although there are three variables, we can group $(\vy,\vs)$ and treat $\vx$ and $(\vy,\vs)$ as the two variables. Then problem \eqref{xyz} has the form \eqref{for:P3-ADMM}. Hence, we have two ways to apply ADM to \eqref{xyz} with two different update orders. Note that $\vy$ and $\vs$ do not appear together in any equation or function, so the ADM subproblem that updates $(\vy,\vs)$ will further decouple to two separable subproblems of $\vy$ and $\vs$; in other words, the resulting ADM has three subproblems involving $\{\vx,\vC\}$ ,$\{\vy,v\}$, $\{\vs,u\}$ separately.
The other two ways are results of the same ``dummy variable'' trick applied to the either formulation in \eqref{uvdualeq}. Again, since now $\vC$ has its own subproblem, these four ways are distinct from the previous eight ways.
As demonstrated through an example, there are quite many ways to formulate the same optimization problem into ``ADM-ready" forms and obtain different ADM algorithms.
While most ADM users choose just one way without paying much attention to the other choices, some show preferences toward a specific formulation. For example, some prefer \eqref{xyz} over those in \eqref{uveq} and \eqref{uvdualeq} since $\vC$, $u$, $v$ all end up in separate subproblems.
When applying ADM to certain $\ell_1$ minimization problems, the authors of~\cite{Yall1,yang2013dual} emphasize on the dual formulations, and later the authors of~\cite{Xiao2013Primal} show a preference over the primal formulations. When ADM was proposed to solve a traffic equilibrium problem, it was first applied to the dual formulation in~\cite{Gabay1983} and, years later, to the primal formulation in~\cite{Fukushima1996The}. Regarding which one of the two variables should be updated first in ADM, neither a rule nor an equivalence claim is found in the literature.
Other than giving preferences to ADM with simpler subproblems, there is no results that compare the different formulations.
\subsection{Contributions}
This chapter shows that, applied to certain pairs of different formulations of the same problem, ADM will generate equivalent sequences of variables that can be mapped exactly from one to another at every iteration. Specifically, between the sequence of an ADM algorithm on a primal formulation and that on the corresponding dual formulation, such maps exist.
We also show that whenever at least one of $f$ and $g$ is a quadratic function (including affine function as a special case), possibly subject to an affine constraint, the sequence of an ADM algorithm can be mapped to that of the ADM algorithm using the opposite order for updating their variables.
Abusing the word ``equivalence'', we say that ADM has ``primal-dual equivalence'' and ``update-order equivalence (with a quadratic objective function).'' Equivalent ADM algorithms take the same number of iterations to reach the same accuracy. (However, it is possible that one algorithm is slightly better than the other in terms of numerical stability, for example, against round-off errors.)
Equipped with these equivalence results, the first eight ways to apply ADM to problem~\eqref{uv} that were discussed in section \ref{sec:11} are reduced to four ways in light of primal-dual equivalence, and the four will further reduce to two whenever $u$ or $v$, or both, is a quadratic function.
The last four ways to apply ADM on problem~\eqref{uv} discussed in section \ref{sec:11}, which yield three subproblems that separately involve $u$, $v$, and $\vC$, are all equivalent and reduce to just one due to primal-dual equivalence and one variable in them is associated with 0 objective (for example, variable $\vx$ has 0 objective in problem \eqref{xyz}).
Take the $\ell_p$-regularization problem, $p\in [1,\infty]$,
\beq\label{lassoprb}
\Min_{\vx}~ \|\vx\|_p + f(\vC\vx)
\eeq
as an example, which is special case of problem~\eqref{uv} with a quadratic function $u$ when $p=2$. We list its three different formulations, whose ADM algorithms are truly different, as follows. When $p\not=2$ \emph{and} $f$ is non-quadratic, each of the first two formulations leads to a pair of different ADM algorithms with different orders of variable update; otherwise, each pair of algorithms is equivalent.
\begin{enumerate}
\item Left formulation of \eqref{uveq}:
$$\left\{\begin{array}{ll}\Min\limits_{\vx,\vy} \quad& \|\vx\|_p+f(\vy)\\
\st \quad& \vC\vx-\vy=0.\end{array} \right.$$
The subproblem for $\vx$ involves $\ell_p$-norm and $\vC$. The other one for $\vy$ involves $f$.
\item Right formulation of \eqref{uveq}:
$$\left\{\begin{array}{ll}\Min\limits_{\vx,\vy} \quad& \|\vx\|_p+f(\vC\vy)\\
\st \quad& \vx-\vy=0.\end{array} \right.$$
The subproblem for $\vx$ involves $\ell_p$-norm and, for $p=1$ and $2$, has a closed-form solution. The other subproblem for $\vy$ involves $f(\vC\cdot)$.
\item Formulation \eqref{xyz}: for any $\mu>0$,
$$ \left\{\begin{array}{ll}\Min\limits_{\vx,\vy,\vs} \quad& \|\vs\|_p+f(\vy)\\
\st \quad& \vC\vx-\vy=0,\\
& \hspace{0pt}\mu(\vx-\vs)=0. \end{array} \right.$$
The subproblem for $\vx$ is quadratic program involving $\vC^*\vC+\mu\vI$. The subproblem for $\vs$ involves $\ell_p$-norm. The subproblem for $\vy$ involves $f$. The subproblems for $\vs$ and $\vy$ are independent.
\end{enumerate}
The best choice depends on which has the simplest subproblems.
The result of ADM's primal-dual equivalence is surprising for three reasons. Firstly, ADM iteration updates \emph{two} primal variable, $\vx^k$ and $\vy^k$ in \eqref{for:P3-ADMM} and \emph{one} dual variable, all in different manners. The updates to the primal variables are done in a Gauss-Seidel manner and involve minimizing functions $f$ and $g$, but the update to the dual variable is explicit and linear. Surprisingly, ADM actually treats one of the two primal variables and the dual variable equally as we will later show. Secondly, most literature describes ADM as an inexact version of the Augmented Lagrangian Method (ALM) \cite{ALM}, which updates $(\vx,\vy)$ together rather than one after another. Although ALM maintains the primal variables, under the hood ALM is the dual-only proximal-point algorithm that iterates the dual variable. It is commonly believed that ADM is an inexact dual algorithm. Thirdly, primal and dual problems typically have different sizes and regularity properties, causing the same algorithm, even if it is applicable to both, to exhibit different performance. For example, the primal and dual variables may have different dimensions. If the primal function $f$ is Lipschitz differentiable, the dual function $f^*$ is strongly convex but can be non-differentiable, and vice versa. Such primal-dual differences often mean that it is numerically advantageous to solve one rather than the other, yet our result means that there is no such primal-dual difference on ADM.
Our maps between equivalent ADM sequences have very simple forms, as the reader will see below. Besides the technical proofs that establish the maps, it is interesting to mention the operator-theoretic perspective of our results. It is shown in \cite{Gabay1983} that the dual-variable sequence of ADM coincides with a sequence of the Douglas-Rachford splitting (DRS) algorithm~\cite{Douglas1956on,Lions1979spliting}. Our ADM's primal-dual equivalence can be obtained through the above ADM--DRS relation and the Moreau identity: $\prox_h + \prox_{h^*}=\vI$, applied to the proximal maps of $f$ and $f^*$ and those of $g$ and $g^*$. The details are omitted in this chapter. Here, $\prox_{h}(x) := \argmin_s h(s)+\frac{1}{2}\|s - x\|^2$.
Our results of primal-dual equivalence for ADM extends to the Peaceman-Rachford splitting (PRS) algorithm. Let the PRS operator~\cite{Peaceman1955the} be denoted as $\vT_{\mathrm{PRS}} = (2\prox_{f}-\vI)\circ(2\prox_{g}-\vI)$. The DRS operator is the average of the identity map and the PRS operator: $\vT_{\mathrm{DRS}}=\frac{1}{2}\vI+\frac{1}{2}\vT_{\mathrm{PRS}}$, and the Relaxed PRS (RPRS) operator is a weighted-average: $\vT_{\mathrm{RPRS}}=(1-\alpha)\vI+\alpha \vT_{\mathrm{PRS}}$, where $\alpha\in(0,1]$. The DRS and PRS algorithms that iteratively apply their operators to find a fixed point were originally proposed for evolving PDEs with two spatial dimensions in the 1950s and then extended to finding a root of the sum of two maximal monotone (set-valued) mappings by Lions and Mercier \cite{Lions1979spliting}. Eckstein showed, in~\cite[Chapter~3.5]{eckstein1989splitting}, that DRS/PRS applied to the primal problem~\eqref{uv} is equivalent to DRS/PRS applied to the dual problem~\eqref{uvdualeq} when $\vC=\vI$. We will show that RPRS applied to~\eqref{uv} is equivalent to RPRS applied to~\eqref{uvdual} for all $\vC$.
In addition to the aforementioned primal-dual and update-order equivalence, we obtain a primal-dual algorithm for the saddle-point formulation of \eqref{for:P3-ADMM} that is also equivalent to the ADM. This primal-dual algorithm is generally \emph{different} from the primal-dual algorithm proposed by Chambolle and Pock~\cite{chambolle2011first}, while they become the same in a special case. The connection between these two algorithms will be explained.
Even when using the same number of dummy variables, truly different ADM algorithms can have different iteration complexities (do not confuse them with the difficulties of their subproblems). The convergence analysis of ADM, such as conditions for sublinear or linear convergence, involves many different scenarios \cite{DengYin2012a,DavisYin2014,DavisYin2014b}. The discussion of convergence rates of ADM algorithms is beyond the scope of this chapter. Our focus is on the equivalence.
\subsection{Organization}
This chapter is organized as follows. Section~\ref{sec:notations} specifies our notation, definitions, and basic assumptions. The three equivalence results for ADM are shown in sections~\ref{sec:ADM},~\ref{sec:ADM-PD}, and~\ref{sec:ADM-equal2}: The primal-dual equivalence of ADM is discussed in sections~\ref{sec:ADM}; ADM is shown to be equivalent to a primal-dual algorithm applied to the saddle-point formulation in section~\ref{sec:ADM-PD}; In section~\ref{sec:ADM-equal2}, we show
the update-order equivalence of ADM if $f$ or $g$ is a quadratic function, possibly subject to an affine constraint. The primal-dual equivalence of RPRS is shown in section~\ref{sec:DRS}. We conclude this chapter with the application of our results on total variation image denoising in section~\ref{sec:app1}.
\section{Notation, definitions, and assumptions}\label{sec:notations}
Let $\cH_1$, $\cH_2$, and $\cG$ be (possibly infinite dimensional) Hilbert spaces. Bold lowercase letters such as $\vx$, $\vy$, $\vu$, and $\vv$ are used for points in the Hilbert spaces. In the example of \eqref{for:P3-ADMM}, we have $\vx\in\cH_1$, $\vy\in\cH_2$, and $\vb\in\cG$. When the Hilbert space a point belongs to is clear from the context, we do not specify it for the sake of simplicity. The inner product between points $\vx$ and $\vy$ is denoted by $\langle \vx,\vy\rangle$, and $\|\vx\|_2:=\sqrt{\langle \vx,\vx\rangle}$ is the corresponding norm. $\|\cdot\|_1$ and $\|\cdot\|_\infty$ denote the $\ell_1$ and $\ell_\infty$ norms, respectively. Bold uppercase letters such as $\vA$ and $\vB$ are used for both continuous linear mappings and matrices. $\vA^*$ denotes the adjoint of $\vA$. $\vI$ denotes the identity mapping.
If $\cC$ is a convex and nonempty set, the indicator function $\iota_\cC$ is defined as follows:
\begin{align*}
\iota_\cC(\vx) = \left\{\begin{array}{ll}0, & \mbox{ if }\vx\in \cC,\\ \infty, &\mbox{ if } \vx\notin \cC. \end{array}\right.
\end{align*}
Both lower and upper case letters such as $f$, $g$, $F$, and $G$ are used for functions. Let $\partial f(\vx)$ be the subdifferential of function $f$ at $\vx$. The proximal operator $\prox_{f(\cdot)}$ of function $f$ is defined as
\begin{align*}
\prox_{f(\cdot)}(\vx) = \argmin_\vy f(\vy) +\frac{1}{2}\|\vy-\vx\|_2^2,
\end{align*}
where the minimization has the unique solution.
The convex conjugate $f^*$ of function $f$ is defined as
\begin{align*}
f^*(\vv) = \sup_\vx \{\langle \vv,\vx \rangle - f(\vx)\}.
\end{align*}
Let $\vL:\cH\rightarrow \cG$, the \emph{infimal postcomposition}~\cite[Def. 12.33]{bauschke2011convex} of $f:\cH\rightarrow(-\infty,+\infty]$ by $\vL$ is given by
\begin{align*}
\vL\triangleright f: \vs\mapsto \inf f(\vL^{-1}(\vs)) = \inf_{\vx:\vL\vx=\vs}f(\vx),
\end{align*}
with $\dom(\vL\triangleright f)=\vL(\dom(f))$.
\begin{lemma}\label{lemma:conjuga_infimal}If $f$ is convex and $\vL$ is affine and expressed as $\vL(\cdot)=\vA\cdot+\vb$, then $\vL\triangleright f$ is convex and the convex conjugate of $\vL\triangleright f$ can be found as follows:
\begin{align*} (\vL\triangleright f)^*(\cdot) = f^*(\vA^*\cdot)+\langle\cdot,\vb\rangle. \end{align*}
\end{lemma}
\begin{proof}Following from the definitions of convex conjugate and infimal postcomposition, we have
\begin{align*}
(\vL\triangleright f)^*(\vv)&= \sup_{\vy}\langle \vv,\vy\rangle -\vL\triangleright f(\vy)=\sup_{\vx}\langle \vv,\vA\vx+\vb\rangle -f(\vx)\\
&=\sup_{\vx}\langle \vA^*\vv,\vx\rangle -f(\vx)+\langle \vv,\vb\rangle=f^*(\vA^*\vv)+\langle \vv,\vb\rangle.
\end{align*}
\end{proof}
\begin{definition}An algorithm on one problem is \emph{equivalent to} another algorithm on the same or another equivalent problem means that the steps in one algorithm can be recovered from the steps in another algorithm, with proper initial conditions and parameters.
\end{definition}
\begin{definition}An optimization algorithm is called \emph{primal-dual equivalent} if this algorithm applied to the primal formulation is equivalent to the same algorithm applied to its Lagrange dual.
\end{definition}
It is important to note that most algorithms are not primal-dual equivalent. ALM applied to the primal problem is equivalent to proximal point method applied to the dual problem~\cite{rockafellar1973dual}, but both algorithms are not primal-dual equivalent. In this chapter, we will show that ADM and RPRS are primal-dual equivalent.
We make the following assumptions throughout the chapter:
\begin{assumption}Functions in this chapter are assumed to be proper, closed, and convex.
\end{assumption}
\begin{assumption}The saddle-point solutions to all the optimization problems in this chapter are assumed to exist.
\end{assumption}
\section{Equivalent problems}\label{sec:problems}
A \emph{primal formulation} equivalent to ~\eqref{for:P3-ADMM} is
\begin{align}\tag{P2}\label{for:P4-ADMM}
\left\{\begin{array}{ll}
\Min\limits_{\vs,\vt} \quad& F(\vs)+G(\vt)\\
\st \quad& \vs+\vt=\vzero,\end{array}\right.
\end{align}
where $\vs,\vt\in \cG$ and
\begin{subequations}\label{for:subprobs}
\begin{align}
F(\vs) &:= \min\limits_{\vx}f(\vx)+\iota_{\{\vx:\vA\vx=\vs\}}(\vx),\\
G(\vt) &:=\min\limits_\vy g(\vy)+\iota_{\{\vy:\vB\vy-\vb=\vt\}}(\vy).\label{for:subprobG}
\end{align}
\end{subequations}
\begin{remark}If we define $\vL_f$ and $\vL_g$ as $\vL_f(\vx)=\vA\vx$ and $\vL_g(\vy)=\vB\vy-\vb$, respectively, then
$$F=\vL_f\triangleright f,\qquad G=\vL_g\triangleright g.$$
\end{remark}
The Lagrange dual of~\eqref{for:P3-ADMM} is
\begin{align}\label{for:dualtoP1}
\Min_\vv \quad f^*(-\vA^*\vv) + g^*(-\vB^*\vv)+\langle\vv,\vb\rangle,
\end{align}
which can be derived from $\Min\limits_\vv\left(-\min\limits_{\vx,\vy}L(\vx,\vy,\vv)\right)$ with the Lagrangian defined as follows:
\begin{align*
L(\vx,\vy,\vv) = f(\vx)+g(\vy) + \langle \vv,\vA\vx+\vB\vy-\vb \rangle.
\end{align*}
An \emph{ADM-ready} formulation of ~\eqref{for:dualtoP1} is\begin{align}\tag{D1}\label{for:D3-ADMM}
\left\{\begin{array}{ll}
\Min\limits_{\vu,\vv} \quad& f^*(-\vA^*\vu)+g^*(-\vB^*\vv)+\langle \vv,\vb\rangle\\
\st \quad& \vu-\vv=\vzero.\end{array}\right.
\end{align}
When ADM is applied to an ADM-ready formulation of the Lagrange dual problem, we call it \emph{Dual ADM}. The original ADM is called \emph{Primal ADM}.
Following similar steps, the ADM ready formulation of the Lagrange dual of~\eqref{for:P4-ADMM} is
\begin{align}\tag{D2}\label{for:D4-ADMM}
\left\{\begin{array}{ll}
\Min\limits_{\vu,\vv} \quad& F^*(-\vu)+G^*(-\vv)\\
\st \quad& \vu-\vv=\vzero.\end{array}\right.
\end{align}
The equivalence between~\eqref{for:D3-ADMM} and~\eqref{for:D4-ADMM} is trivial since
\begin{align*}
F^*(\vu)&=f^*(\vA^*\vu),\\
G^*(\vv)&=g^*(\vB^*\vv)-\langle \vv,\vb\rangle,
\end{align*}
which follows from Lemma~\ref{lemma:conjuga_infimal}.
Although there can be multiple equivalent formulations of the same problem (e.g.,~\eqref{for:P3-ADMM},~\eqref{for:P4-ADMM},~\eqref{for:dualtoP1}, and ~\eqref{for:D3-ADMM}/\eqref{for:D4-ADMM} are equivalent), an algorithm may or may not be applicable to some of them. Even when they are, on different formulations, their behaviors such as convergence and speed of convergence are different. In particular, most algorithms have different behaviors on primal and dual formulations of the same problem. An algorithm applied to a primal formulation does not dictate the behavior of the same algorithm applied to the related dual formulation. The simplex method in linear programming has different performance when applied to both the primal and dual problems, i.e., the primal simplex method starts with a primal basic feasible solution (dual infeasible) until the dual feasibility conditions are satisfied, while the dual simplex method starts with a dual basic feasible solution (primal infeasible) until the primal feasibility conditions are satisfied. The ALM also has different performance when applied to the primal and dual problems, i.e., ALM applied to the primal problem is equivalent to proximal point method applied to the related dual problem, and proximal point method is, in general, different from ALM on the same problem.
\section{Primal-dual equivalence of ADM}\label{sec:ADM}
In this section we show the primal-dual equivalence of ADM. Algorithms~\ref{alg:ADM-1}-\ref{alg:ADM-3}~describe how ADM is applied to~\eqref{for:P3-ADMM},~\eqref{for:P4-ADMM}, and~\eqref{for:D3-ADMM}/~\eqref{for:D4-ADMM}\cite{ADM1,ADM2}.
\begin{algorithm}[H]
\caption{ADM on~\eqref{for:P3-ADMM}}
\label{alg:ADM-1}
\begin{algorithmic}
\STATE initialize $\vx_1^0$, $\vz_1^0$, $\lambda>0$
\FOR {$k =0,1, \cdots$}
\STATE $\vy_1^{k+1} \in \argmin\limits_\vy g(\vy) +(2\lambda)^{-1}\|\vA\vx_1^k+\vB\vy-\vb+\lambda\vz_1^k\|_2^2$
\STATE $\vx_1^{k+1} \in \argmin\limits_\vx f(\vx)+(2\lambda)^{-1}\|\vA\vx+\vB\vy_1^{k+1}-\vb+\lambda\vz_1^k\|_2^2$
\STATE $\vz_1^{k+1} = \vz_1^k+\lambda^{-1}(\vA\vx_1^{k+1}+\vB\vy_1^{k+1}-\vb)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\caption{ADM on~\eqref{for:P4-ADMM}}
\label{alg:ADM-2}
\begin{algorithmic}
\STATE initialize $\vs_2^0$, $\vz_2^0$, $\lambda>0$
\FOR {$k =0,1, \cdots$}
\STATE $\vt_2^{k+1} = \argmin\limits_\vt G(\vt)+(2\lambda)^{-1}\|\vs_2^k+\vt+\lambda\vz_2^k\|_2^2$
\STATE $\vs_2^{k+1} = \argmin\limits_\vs F(\vs)+(2\lambda)^{-1}\|\vs+\vt_2^{\vk+1}+\lambda\vz_2^k\|_2^2$
\STATE $\vz_2^{k+1} = \vz_2^k+\lambda^{-1}(\vs_2^{k+1}+\vt_2^{k+1})$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\caption{ADM on~\eqref{for:D3-ADMM}/\eqref{for:D4-ADMM}}
\label{alg:ADM-3}
\begin{algorithmic}
\STATE initialize $\vu_3^0$, $\vz_3^0$, $\lambda>0$
\FOR {$k =0,1, \cdots$}
\STATE $\vv_3^{k+1} = \argmin\limits_\vv G^*(-\vv)+{\lambda\over 2}\|\vu_3^k-\vv+\lambda^{-1}\vz_3^k\|_2^2$
\STATE $\vu_3^{k+1} = \argmin\limits_\vu F^*(-\vu)+{\lambda\over2}\|\vu-\vv_3^{\vk+1}+\lambda^{-1}\vz_3^k\|_2^2$
\STATE $\vz_3^{k+1} = \vz_3^k+\lambda(\vu_3^{k+1}-\vv_3^{k+1})$
\ENDFOR
\end{algorithmic}
\end{algorithm}
The $\vy_1^{k}$ and $\vx_1^{k}$ in Algorithm~\ref{alg:ADM-1} may not be unique because of the matrices $\vA$ and $\vB$, while $\vA\vx^k_1$ and $\vB\vy^k_1$ are unique. In addition, $\vA\vx^k_1$ and $\vB\vy^k_1$ are calculated for twice and thus stored in the implementation of Algorithm~\ref{alg:ADM-1} to save the second calculation. Following the equivalence of Algorithms~\ref{alg:ADM-1} and~\ref{alg:ADM-2} in Part 1 of the following theorem~\ref{thm:ADMM-equal}, we can view problem \eqref{for:P4-ADMM} as the \emph{master problem} of \eqref{for:P3-ADMM}. We can say that ADM is essentially an algorithm applied only to the master problem \eqref{for:P4-ADMM}, which is Algorithm ~~\ref{alg:ADM-2}; this fact has been obscured by the often-seen Algorithm ~\ref{alg:ADM-1}, which integrates ADM on the master problem with the independent subproblems in~\eqref{for:subprobs}.
\begin{theorem}[Equivalence of Algorithms~\ref{alg:ADM-1}-\ref{alg:ADM-3}] \label{thm:ADMM-equal}
Suppose $\vA\vx_1^0=\vs_2^0=\vz_3^0$ and $\vz_1^0=\vz_2^0=\vu_3^0$ and that the same parameter $\lambda$ is used in Algorithms~\ref{alg:ADM-1}-\ref{alg:ADM-3}. Then, their equivalence can be established as follows:
\begin{enumerate}
\item
From $\vx_1^k$, $\vy_1^k$, $\vz_1^k$ of Algorithm~\ref{alg:ADM-1}, we obtain $\vt_2^k$, $\vs_2^k$, $\vz_2^k$ of Algorithm~\ref{alg:ADM-2} through:
\begin{subequations}
\begin{align}
\vt_2^k&=\vB\vy_1^k-\vb,\label{p1vt}\\
\vs_2^k&=\vA\vx_1^k,\label{p1vs}\\
\vz_2^k&=\vz_1^k.\label{p1vz}
\end{align}
\end{subequations}
From $\vt_2^k$, $\vs_2^k$, $\vz_2^k$ of Algorithm~\ref{alg:ADM-2}, we obtain $\vy_1^k$, $\vx_1^k$, $\vz_1^k$ of Algorithm~\ref{alg:ADM-1} through:
\begin{subequations}
\begin{align}
\vy_1^k &= \argmin_\vy\{g(\vy): \vB\vy-\vb=\vt_2^k\},\label{p2vy}\\
\vx_1^k &= \argmin_\vx\{f(\vx): \vA\vx=\vs_2^k\},\label{p2vx}\\
\vz_1^k&=\vz_2^k.\label{p2vz}
\end{align}
\end{subequations}
\item
We can recover the iterates of Algorithm~\ref{alg:ADM-2} and ~\ref{alg:ADM-3} from each other through
\begin{align}\label{p3uz}
\vu_3^k=\vz_2^k,\qquad \vz_3^k=\vs_2^k.
\end{align}
\end{enumerate}
\end{theorem}
\begin{proof}
\emph{Part 1.} Proof by induction.\\
We argue that under \eqref{p1vs} and \eqref{p1vz}, Algorithms 1 and 2 have \emph{essentially identical} subproblems in their \emph{first} steps at the $k$th iteration. Consider the following problem, which is obtained by plugging the definition of $G(\cdot)$ into the $\vt_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2}:
\begin{align}\label{for:ytsub}
(\vy_1^{k+1},\vt_2^{k+1})=\argmin_{\vy,\vt} g(\vy)+\iota_{\{(\vy,\vt):\vB\vy-\vb=\vt\}}(\vy,\vt)+(2\lambda)^{-1}\|\vs_2^k+\vt+\lambda\vz_2^k\|_2^2.
\end{align}
If one minimizes over $\vy$ first while keeping $\vt$ as a variable, one eliminates $\vy$ and recovers the $\vt_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2}. If one minimizes over $\vt$ first while keeping $\vy$ as a variable, then after plugging in \eqref{p1vs} and \eqref{p1vz}, problem \eqref{for:ytsub} reduces to the $\vy_1^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-1}.
In addition, $(\vy_1^{k+1},\vt_2^{k+1})$ obeys
\begin{align}\vt_2^{k+1}=\vB\vy_1^{k+1}-\vb, \label{for:tbBy}
\end{align}
which is \eqref{p1vt} at ${k+1}$. Plugging $\vt = \vt_2^{k+1}$ into \eqref{for:ytsub} yields problem \eqref{p2vy} for $\vy_1^{k+1}$, which must be equivalent to the $\vy_1^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2}. Therefore, the $\vy_1^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-1} and the $\vt_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2} are equivalent through \eqref{p1vt} and \eqref{p2vy} at $k+1$, respectively.
Similarly, under \eqref{for:tbBy} and \eqref{p1vz}, we can show that the $\vx_1^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-1} and the $\vs_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2} are equivalent through the formulas for \eqref{p1vs} and \eqref{p2vx} at $k+1$, respectively.
Finally, under \eqref{p1vt} and \eqref{p1vs} at $k+1$ and $\vz_2^k=\vz_1^k$, the formulas for $\vz_1^{k+1}$ and $\vz_2^{k+1}$ in Algorithms 1 and 2 are identical, and they return $\vz_1^{k+1}=\vz_2^{k+1}$, which is \eqref{p1vz} and \eqref{p2vz} at $k+1$.
\emph{Part 2.}
Proof by induction. Suppose that \eqref{p3uz} holds. We shall show that \eqref{p3uz} holds at $k+1$. Starting from the optimality condition of the $\vt_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2}, we derive
\begin{align*}
& \vzero\in \partial G(\vt_2^{k+1})+\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k)\\
\Longleftrightarrow~ & \vt_2^{k+1}\in\partial G^*(-\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k))\\
\Longleftrightarrow~ & \lambda\left[\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k)\right]-(\lambda\vz_2^k+\vs_2^k)\in\partial G^*(-\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k))\\
\Longleftrightarrow~ & -\lambda\left[\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k)\right]+(\lambda\vu_3^k+\vz_3^k)\in-\partial G^*(-\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k))\\
\Longleftrightarrow~ & \vzero\in -\partial G^*(-\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k))-\lambda\left[\vu_3^k-\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k)+\lambda^{-1}\vz_3^k\right]\\
\Longleftrightarrow~ & \vv_3^{k+1}=\lambda^{-1} (\vs_2^k+\vt_2^{k+1}+\lambda\vz_2^k)=\lambda^{-1} (\vz_3^k+\vt_2^{k+1}+\lambda\vz_2^k),
\end{align*}
where the last equivalence follows from the optimality condition for the $\vv_3^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-3}.
Starting from the optimality condition of the $\vs_2^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-2}, and applying the update, $\vz_2^{k+1}=\vz_2^k+\lambda^{-1}(\vs_2^{k+1}+\vt_2^{k+1})$, in Algorithm~\ref{alg:ADM-2} and the identity of $\vt_2^{k+1}$ obtained above, we derive \begin{align*}
&\vzero\in \partial F(\vs_2^{k+1})+\lambda^{-1}(\vs_2^{k+1}+\vt_2^{k+1}+\lambda\vz_2^k)\\
\Longleftrightarrow~ &\vzero\in \partial F(\vs_2^{k+1})+\vz_2^{k+1}\\
\Longleftrightarrow~ &\vzero\in \vs_2^{k+1}-\partial F^*(-\vz_2^{k+1})\\
\Longleftrightarrow~ &\vzero\in \lambda(\vz_2^{k+1}-\vz_2^k)-\vt_2^{k+1}-\partial F^*(-\vz_2^{k+1})\\
\Longleftrightarrow~ &\vzero\in \lambda(\vz_2^{k+1}-\vz_2^k)+\vz_3^k+\lambda (\vz_2^k-\vv_3^{k+1})-\partial F^*(-\vz_2^{k+1})\\
\Longleftrightarrow~ &\vzero\in -\partial F^*(-\vz_2^{k+1})+\lambda(\vz_2^{k+1}-\vv_3^{k+1}+\lambda^{-1}\vz_3^k)\\
\Longleftrightarrow~ & \vz_2^{k+1} = \vu_3^{k+1}.
\end{align*}
where the last equivalence follows from the optimality condition for the $\vu_3^{k+1}$-subproblem of Algorithm~\ref{alg:ADM-3}. Finally, combining the update formulas of $\vz_2^{k+1}$ and $\vz_3^{k+1}$ in Algorithm~~\ref{alg:ADM-2} and \ref{alg:ADM-3}, respectively, as well as the identities for $\vu_3^{k+1}$ and $\vv_3^{k+1}$ obtained above, we obtain
\begin{align*}
\vz_3^{k+1}&=\vz_3^{k}+\lambda(\vu_3^{k+1}-\vv_3^{k+1})=\vs^k+\lambda (\vz_2^{k+1}-\vz_2^k-\lambda^{-1}(\vs_2^k+\vt_2^{k+1}))\\
&=\lambda (\vz_2^{k+1}-\vz_2^k)-\vt_2^{k+1} = \vs_2^{k+1}.
\end{align*}
\end{proof}
\begin{remark} Part 2 of the theorem (ADM's primal-dual equivalence) can also be derived by combining the following two equivalence results:\ (i) the equivalence between ADM on the primal problem and the Douglas-Rachford splitting (DRS) algorithm~\cite{Douglas1956on,Lions1979spliting} on the dual problem~\cite{Gabay1983}, and (ii) the equivalence result between DRS algorithms applied to the master problem~\eqref{for:P4-ADMM} and its dual problem (cf.~\cite[Chapter~3.5]{eckstein1989splitting}\cite{eckstein1994some}). In this chapter, however, we provide an elementary algebraic proof in order to derive the formulas in theorem~\ref{thm:ADMM-equal} that recover the iterates of one algorithm from another.
\end{remark}
Part 2 of the theorem shows that ADM is a symmetric primal-dual algorithm. The reciprocal positions of parameter $\lambda$ indicates its function to ``balance'' the primal and dual progresses
Part 2 of the theorem also shows that Algorithms~\ref{alg:ADM-2} and~\ref{alg:ADM-3} have no difference, in terms of per-iteration complexity and the number of iterations needed to reach an accuracy. However, Algorithms~\ref{alg:ADM-1} and~\ref{alg:ADM-2} have difference in terms of per-iteration complexity. In fact, Algorithm~\ref{alg:ADM-2} is implemented for Algorithm~\ref{alg:ADM-1} because Algorithm~\ref{alg:ADM-2} has smaller complexity than Algorithm~\ref{alg:ADM-1}. See the examples in sections~\ref{sec:examples41} and~\ref{sec:examples42}.
\subsection{Primal-dual equivalence of ADM on~\eqref{uv} with three subproblems}
In section~\ref{sec:11}, we introduced four different ways to apply ADM on~\eqref{uv} with three subproblems. The ADM-ready formulation for the primal problem is~\eqref{xyz}, and the ADM applied to this formulation is
\begin{subequations}\label{alg:xyz_primal}
\begin{align}
\vx^{k+1} &= \argmin\limits_\vx \|\vx-\vs^k+\lambda\vz_\vs^k\|_2^2+\|\vC\vx-\vy^k+\lambda\vz_\vy^k\|_2^2,\\
\vs^{k+1} &= \argmin\limits_\vs u(\vs) +(2\lambda)^{-1}\|\vx^{k+1}-\vs+\lambda\vz_\vs^k\|_2^2,\\
\vy^{k+1} &= \argmin\limits_\vy v(\vy)+(2\lambda)^{-1}\|\vC\vx^{k+1}-\vy+\lambda\vz_\vy^k\|_2^2,\\
\vz_\vs^{k+1} &= \vz_\vs^k+\lambda^{-1}(\vx^{k+1}-\vs^{k+1}),\\
\vz_\vy^{k+1} &= \vz_\vy^k+\lambda^{-1}(\vC\vx^{k+1}-\vy^{k+1}).
\end{align}
\end{subequations}
Similarly, we can introduce a dummy variable $\vt$ into the left formulation in~\eqref{uvdualeq} and obtain a new equivalent formulation
\beq\label{xyz2}
\left\{\begin{array}{ll}\Min\limits_{\vu,\vv,\vt} \quad& u^*(\vu)+v^*(\vt)\\
\st \quad& \vC^*\vv+\vu=0, ~\vv-\vt=0. \end{array} \right.
\eeq
The ADM applied to~\eqref{xyz2} is
\begin{subequations}\label{alg:xyz_dual}
\begin{align}
\vv^{k+1} &= \argmin\limits_\vv \|\vC^*\vv+\vu^k+\lambda^{-1}\vz_\vu^k\|_2^2+\|\vv-\vt^k+\lambda^{-1}\vz_\vt^k\|_2^2,\\
\vu^{k+1} &= \argmin\limits_\vu u^*(\vu) +{\lambda\over2}\|\vC^*\vv^{k+1}+\vu+\lambda^{-1}\vz_\vu^k\|_2^2,\\
\vt^{k+1} &= \argmin\limits_\vt v^*(\vt) +{\lambda\over2}\|\vv^{k+1}-\vt+\lambda^{-1}\vz_\vt^k\|_2^2,\\
\vz_\vu^{k+1} &= \vz_\vu^k+\lambda(\vC^*\vv^{k+1}+\vu^{k+1}),\\
\vz_\vt^{k+1} &= \vz_\vt^k+\lambda(\vv^{k+1}-\vt^{k+1}).
\end{align}
\end{subequations}
Interestingly, as shown in the following corollary, ADM algorithms~\eqref{alg:xyz_primal} and~\eqref{alg:xyz_dual} applied to~\eqref{xyz} and ~\eqref{xyz2} are equivalent.
\begin{corollary}
If the initialization for algorithms~\eqref{alg:xyz_primal} and~\eqref{alg:xyz_dual} satisfies $\vz_\vy^0 = \vt^0$, $\vz_\vs^0=\vu^0$, $\vs^0=-\vz_\vu^0$, and $\vy^0=\vz_\vt^0$. Then for $k\geq1$, we have the following equivalence results between the iterations of the two algorithms:
\begin{align*}
\vz_\vy^k = \vt^k,\quad \vz_\vs^k=\vu^k,\quad \vs^k=-\vz_\vu^k,\quad \vy^k=\vz_\vt^k.
\end{align*}
\end{corollary}
The proof is similar to the proof of Theorem~\ref{thm:ADMM-equal} and is omitted here.
\subsection{Example: basis pursuit}\label{sec:examples41}
The basis pursuit problem seeks for the minimal $\ell_1$ solution to a set of linear equations:
\begin{align}\label{for:BP}
\Min_\vu \|\vu\|_1\quad\st~\vA\vu=\vb.
\end{align}
Its Lagrange dual is
\begin{align}\label{for:BPDual}
\Min_\vx -\vb^T\vx\quad\st~\|\vA^*\vx\|_\infty \leq1.
\end{align}
The YALL1 algorithms~\cite{Yall1} implement ADMs on a set of primal and dual formulations for basis pursuit and LASSO, yet ADM for~\eqref{for:BP} is not given (however, a linearized ADM is given for \eqref{for:BP}). Although seemingly awkward, problem ~\eqref{for:BP} can be turned equivalently into the ADM-ready form
\begin{align}\label{for:BPSplt}
\Min_{\vu,\vv} \|\vv\|_1 +\iota_{\{\vu:\vA\vu=\vb\}}(\vu)\quad\st~\vu-\vv=\vzero.
\end{align}
Similarly, problem ~\eqref{for:BPDual} can be turned equivalently into the ADM-ready form
\begin{align}\label{for:BPDualSplt}
\Min_{\vx,\vy} -\vb^T\vx +\iota_{B_1^\infty}(\vy)\quad\st~\vA^*\vx-\vy=\vzero,
\end{align}
where $B_1^\infty=\{\vy:\|\vy\|_\infty\leq1\}$.
For simplicity, let us suppose that $\vA$ has full row rank so the inverse of $\vA\vA^*$ exists. (Otherwise, $\vA\vu=\vb$ are redundant whenever they are consistent; and $(\vA\vA^*)^{-1}$ shall be replaced by the pseudo-inverse below.) ADM for problem \eqref{for:BPSplt} can be simplified to the iteration:
\begin{subequations}\label{eqn:BP-P}
\begin{align}
\label{for:BP-PADM-u1}\vv_3^{k+1} = &\argmin_{\vv} \|\vv\|_1+{\lambda\over 2}\|\vu_3^k-\vv+{1\over \lambda}\vz_3^k\|_2^2,\\
\label{for:BP-PADM-v1}\vu_3^{k+1} = &{\vv_3^{k+1}-{1\over \lambda}\vz_3^k}-\vA^*(\vA\vA^*)^{-1}(\vA(\vv_3^{k+1}-{1\over \lambda}\vz_3^k)-\vb),\\
\vz_3^{k+1}=&\vz_3^k+\lambda(\vu_3^{k+1}-\vv_3^{k+1}).
\end{align}
\end{subequations}
And ADM for problem \eqref{for:BPDualSplt} can be simplified to the iteration:
\begin{subequations}\label{eqn:BP-D}
\begin{align}
\label{for:BP-DADM-y1}\vy_1^{k+1} = &\cP_{B_1^\infty}(\vA^*\vx_1^k+\lambda\vz_1^k),\\
\label{for:BP-DADM-x1}\vx_1^{k+1} = &(\vA\vA^*)^{-1}(\vA\vy_1^{k+1}-\lambda(\vA\vz_1^k-\vb)),\\
\vz_1^{k+1}=&\vz_1^k+\lambda^{-1}(\vA^*\vx_1^{k+1}-\vy_1^{k+1}),
\end{align}
\end{subequations}
where $\cP_{B_1^\infty}$ is the projection onto $B_1^\infty$.
Looking into the iteration in~\eqref{eqn:BP-D}, we can find that $\vA^*\vx_1^k$ is used in both the $k$th and $k+1$st iterations. To save the computation, we can store $\vA^*\vx_1^k$ as $\vs_2^k$. In addition, let $\vt_2^k=\vy_1^k$ and $\vz_2^k=\vz_1^k$, we have
\begin{subequations}\label{eqn:BP-D2}
\begin{align}
\label{for:BP-DADM-t1}\vt_2^{k+1} = &\cP_{B_1^\infty}(\vs_2^k+\lambda\vz_2^k),\\
\label{for:BP-DADM-s1}\vs_2^{k+1} = &\vA^*(\vA\vA^*)^{-1}(\vA(\vt_2^{k+1}-\lambda\vA\vz_2^k)+\lambda\vb)),\\
\vz_2^{k+1}=&\vz_2^k+\lambda^{-1}(\vs_2^{k+1}-\vt_2^{k+1}),
\end{align}
\end{subequations}
which is exactly Algorithm~\ref{alg:ADM-2} for~\eqref{for:BPDualSplt}. Thus, Algorithm~\ref{alg:ADM-2} has smaller complexity than Algorithm~\ref{alg:ADM-1}, i.e., one matrix vector multiplication $\vA^*\vx_1^k$ is saved from Algorithm~\ref{alg:ADM-2}.
The corollary below follows directly from Theorem~\ref{thm:ADMM-equal} by associating \eqref{for:BPDualSplt} and \eqref{for:BPSplt} as \eqref{for:P3-ADMM} and \eqref{for:D4-ADMM}, and \eqref{eqn:BP-D} and \eqref{eqn:BP-P} with the iterations of Algorithms~\ref{alg:ADM-1} and \ref{alg:ADM-3}, respectively.
\begin{corollary}\label{thm:BP-equal}Suppose that $\vA\vu=\vb$ are consistent. Consider ADM iterations \eqref{eqn:BP-P} and \eqref{eqn:BP-D}. Let $\vu_3^0=\vz_1^0$ and $\vz_3^0=\vA^*\vx_1^0$. Then, for $k\geq1$, iterations \eqref{eqn:BP-P} and \eqref{eqn:BP-D} are equivalent. In particular,
\begin{itemize}
\item From $\vx_1^k$, $\vz_1^k$ in~\eqref{eqn:BP-D}, we obtain $\vu_3^k$, $\vz_3^k$ in~\eqref{eqn:BP-P} through:
\begin{align*}
\vu_3^k=\vz_1^k,\qquad \vz_3^k=\vA^*\vx_1^k.
\end{align*}
\item From $\vu_3^k$, $\vz_3^k$ in~\eqref{eqn:BP-P}, we obtain $\vx_1^k$, $\vz_1^k$ in~\eqref{eqn:BP-D} through:
\begin{align*}
\vx_1^k =(\vA\vA^*)^{-1}\vA\vz_3^k,\qquad
\vz_1^k=\vu_3^k.
\end{align*}
\end{itemize}
\end{corollary}
\subsection{Example: basis pursuit denoising}\label{sec:examples42}
The basis pursuit denoising problem is
\begin{align}\label{eq:BPD}
\Min_\vu \|\vu\|_1 +{1\over 2\alpha}\|\vA\vu-\vb\|_2^2
\end{align}
and its Lagrange dual, in the ADM-ready form, is
\begin{align}\label{eq:BPDDual}
\Min_{\vx,\vy} -\langle\vb,\vx\rangle +{\alpha\over2}\|\vx\|_2^2 +\iota_{B_1^\infty}(\vy)\quad\st~ \vA^*\vx-\vy=\vzero.
\end{align}
The iteration of ADM for \eqref{eq:BPDDual} is
\begin{subequations}\label{for:BPDN-DADM}
\begin{align}
\label{for:BPDN-DADM-y1}\vy_1^{k+1} = &\cP_{B_1^\infty}(\vA^*\vx_1^k+\lambda\vz_1^k),\\
\label{for:BPDN-DADM-x1}\vx_1^{k+1} = &(\vA\vA^*+\alpha\lambda\vI)^{-1}(\vA\vy_1^{k+1}-\lambda(\vA\vz_1^k-\vb)),\\
\vz_1^{k+1}=&\vz_1^k+\lambda^{-1}(\vA^*\vx_1^{k+1}-\vy_1^{k+1}).
\end{align}
\end{subequations}
Looking into the iteration in~\eqref{for:BPDN-DADM}, we can find that $\vA^*\vx_1^k$ is used in both the $k$th and $k+1$st iterations. To save the computation, we can store $\vA^*\vx_1^k$ as $\vs_2^k$. In addition, let $\vt_2^k=\vy_1^k$ and $\vz_2^k=\vz_1^k$, we have
\begin{subequations}\label{for:BPDN-DADM2}
\begin{align}
\label{for:BPDN-DADM-t1}\vt_2^{k+1} = &\cP_{B_1^\infty}(\vs_2^k+\lambda\vz_2^k),\\
\label{for:BPDN-DADM-s1}\vs_2^{k+1} = &\vA^*(\vA\vA^*+\alpha\lambda\vI)^{-1}(\vA(\vt_2^{k+1}-\lambda\vz_2^k)+\lambda\vb)),\\
\vz_2^{k+1}=&\vz_2^k+\lambda^{-1}(\vs_2^{k+1}-\vt_2^{k+1}),
\end{align}
\end{subequations}
which is exactly Algorithm~\ref{alg:ADM-2} for~\eqref{eq:BPDDual}. Thus, Algorithm~\ref{alg:ADM-2} has smaller complexity than Algorithm~\ref{alg:ADM-1}, i.e., one matrix vector multiplication $\vA^*\vx_1^k$ is saved from Algorithm~\ref{alg:ADM-2}. In addition, if $\vA^*\vA=\vI$,~\eqref{for:BPDN-DADM-s1} becomes
\begin{align}
\vs_2^{k+1} = &(\alpha\lambda+1)^{-1}(\vt_2^{k+1}-\lambda\vz_2^k+\lambda\vA^*\vb),
\end{align}
and no matrix vector multiplications is needed during the iteration because $\lambda\vA^*\vb$ can be precalculated.
The ADM-ready form of the original problem \eqref{eq:BPD} is
\begin{align}
\Min_{\vu,\vv} \|\vv\|_1 +{1\over2\alpha}\|\vA\vu-\vb\|_2^2 \quad\st~\vu-\vv=\vzero,
\end{align}
whose ADM iteration is
\begin{subequations}\label{for:BPDN-PADM}
\begin{align}
\label{for:BPDN-PADM-u1}\vv_3^{k+1} = &\argmin_{\vv} \|\vv\|_1+{\lambda\over 2}\|\vu_3^k-\vv+{1\over \lambda}\vz_3^k\|_2^2,\\
\label{for:BPDN-PADM-v1}\vu_3^{k+1} = &(\vA^*\vA+\alpha\lambda\vI)^{-1}(\vA^*\vb+\alpha\lambda\vv_3^{k+1}-\alpha\vz_3^k),\\
\vz_3^{k+1}=&\vz_3^k+\lambda(\vu_3^{k+1}-\vv_3^{k+1}).
\end{align}
\end{subequations}
The corollary below follows directly from Theorem~\ref{thm:ADMM-equal}.
\begin{corollary}Consider ADM iterations \eqref{for:BPDN-DADM} and \eqref{for:BPDN-PADM}. Let $\vu_3^0=\vz_1^0$ and $\vz_3^0=\vA^*\vx_1^0$. For $k\geq1$, ADM on the dual and primal problems~\eqref{for:BPDN-DADM} and~\eqref{for:BPDN-PADM} are equivalent in the following way:
\begin{itemize}
\item From $\vx_1^k$, $\vz_1^k$ in~\eqref{for:BPDN-DADM}, we recover $\vu_3^k$, $\vz_3^k$ in~\eqref{for:BPDN-PADM} through:
\begin{align*}
\vu_3^k=\vz_1^k, \qquad
\vz_3^k=\vA^*\vx_1^k.
\end{align*}
\item From $\vu_3^k$, $\vz_3^k$ in~\eqref{for:BPDN-PADM}, we recover $\vx_1^k$, $\vz_1^k$ in~\eqref{for:BPDN-DADM} through:
\begin{align*}
\vx_1^k=-(\vA\vu_3^{k}-\vb)/\alpha,\qquad
\vz_1^k=\vu_3^k.\\
\end{align*}
\end{itemize}
\end{corollary}
\begin{remark} Iteration \eqref{for:BPDN-PADM} is different from that of ADM for another ADM-ready form of \eqref{eq:BPD}
\begin{align}
\Min_{\vu,\vv} \|\vu\|_1 +{1\over2\alpha}\|\vv\|_2^2\quad\st~\vA\vu-\vv=\vb,
\end{align}
which is used in~\cite{Yall1}. In general, there are different ADM-ready forms and their ADM algorithms yield different iterates. ADM on one ADM-ready form is equivalent to it on the corresponding dual ADM-ready form.
\end{remark}
\section{ADM as a primal-dual algorithm on the saddle-point problem}\label{sec:ADM-PD}
As shown in section~\ref{sec:ADM}, ADM on a pair of convex primal and dual problems are equivalent, and there is a connection between $\vz_1^k$ in Algorithm~\ref{alg:ADM-1} and dual variable $\vu_3^k$ in Algorithm~\ref{alg:ADM-3}. This primal-dual equivalence naturally suggests that ADM is also equivalent to a primal-dual algorithm involving both primal and dual variables.
We derive problem~\eqref{for:P3-ADMM} into an equivalent primal-dual saddle-point problem \eqref{for:saddle-point} as follows:
\begin{align}
&\min_{\vy,\vx} g(\vy)+f(\vx)+\iota_{\{(\vx,\vy):\vA\vx=\vb-\vB\vy\}}(\vx,\vy)\nonumber\\
=&\min_{\vy} g(\vy)+F(\vb-\vB\vy)\nonumber\\
=&\min_\vy\max_\vu g(\vy)+\langle -\vu, \vb-\vB\vy\rangle -F^*(-\vu)\label{for:saddle-pointb}\\
=&\min_\vy\max_\vu g(\vy)+\langle \vu, \vB\vy-\vb\rangle -f^*(-\vA^*\vu).\label{for:saddle-point}
\end{align}
A primal-dual algorithm for solving \eqref{for:saddle-point} is described in Algorithm~\ref{alg:primal-dual}. Theorem~\ref{thm:ADM-PD} establishes the equivalence between Algorithms~\ref{alg:ADM-1} and \ref{alg:primal-dual}.
\begin{algorithm}[H]
\caption{Primal-dual formulation of ADM on Problem~\eqref{for:saddle-point}}
\label{alg:primal-dual}
\begin{algorithmic}
\STATE initialize $\vu_4^0$, $\vu_4^{-1}$, $\vy_4^0$, $\lambda>0$
\FOR {$k =0,1, \cdots$}
\STATE $\bar{\vu}_4^k=2\vu_4^{k}-\vu_4^{k-1}$
\STATE $\vy_4^{k+1}=\argmin\limits_\vy g(\vy)+(2\lambda)^{-1}\|\vB\vy-\vB\vy_4^k+\lambda\bar\vu_4^k\|_2^2$
\STATE $\vu_4^{k+1} =\argmin\limits_\vu f^*(-\vA^*\vu)-\langle\vu,\vB\vy_4^{k+1}-\vb\rangle+\lambda/2\|\vu-\vu_4^k\|_2^2$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{remark}
Paper~\cite{chambolle2011first} proposed a primal-dual algorithm for~\eqref{for:saddle-pointb} and obtained the connection between ADM and that primal-dual algorithm~\cite{esser2010a}: When $\vB=\vI$, ADM is equivalent to the primal-dual algorithm in~\cite{chambolle2011first}; When $\vB\neq\vI$, the primal-dual algorithm is a preconditioned ADM as an additional proximal term $\delta/2\|\vy-\vy_4^k\|_2^2-(2\lambda)^{-1}\|\vB\vy-\vB\vy_4^k\|_2^2$ is added to the subproblem for $\vy_4^{k+1}$. This is also a special case of inexact ADM in~\cite{DengYin2012a}. Our Algorithm \ref{alg:primal-dual} is a primal-dual algorithm that is equivalent to ADM in the general case.
\end{remark}
\begin{theorem}[Equivalence between Algorithms~\ref{alg:ADM-1} and \ref{alg:primal-dual}] \label{thm:ADM-PD}
Suppose that $\vA\vx_1^0=\lambda(\vu_4^0-\vu_4^{-1})+\vb-\vB\vy_4^0$ and $\vz_1^0=\vu_4^0$. Then, Algorithms~\ref{alg:ADM-1} and~\ref{alg:primal-dual} are equivalent with the identities:
\begin{align}\label{for:Axuub}
\vA\vx_1^k=\lambda(\vu_4^k-\vu_4^{k-1})+\vb-\vB\vy_4^k,\qquad \vz_1^k=\vu_4^k,
\end{align} for all $k>0$.
\end{theorem}
\begin{proof}By assumption, \eqref{for:Axuub} holds at iteration $k=0$. \\
Proof by induction. Suppose that \eqref{for:Axuub} holds at iteration $k\ge 0$. We shall establish \eqref{for:Axuub} at iteration $k+1$. From the first step of Algorithm~\ref{alg:ADM-1}, we have
\begin{align*}
\vy_1^{k+1}=&\argmin_\vy g(\vy)+(2\lambda)^{-1}\|\vA\vx_1^k+\vB\vy-\vb+\lambda\vz_1^k\|_2^2\\
=&\argmin_\vy g(\vy)+(2\lambda)^{-1}\|\lambda(\vu_4^k-\vu_4^{k-1})+\vB\vy-\vB\vy_4^{k}+\lambda\vu_4^k\|_2^2,
\end{align*}
which is the same as the first step in Algorithm~\ref{alg:primal-dual}. Thus we have $\vy_1^{k+1}=\vy_4^{k+1}$.
Combing the second and third steps of Algorithm~\ref{alg:ADM-1}, we have
\begin{align*}
\vzero \in \partial f(\vx_1^{k+1})+\lambda^{-1}\vA^*(\vA\vx_1^{k+1}+\vB\vy_1^{k+1}-\vb+\lambda\vz_1^k)=\partial f(\vx_1^{k+1}) + \vA^*\vz_1^{k+1}.
\end{align*}
Therefore,
\begin{align*}
&\vx_1^{k+1}\in\partial f^*(-\vA^*\vz_1^{k+1}) \\
\Longrightarrow~& \vA\vx_1^{k+1}\in\partial F^*(-\vz_1^{k+1})\\
\Longleftrightarrow~& \lambda (\vz_1^{k+1}-\vz_1^{k})+\vb-\vB\vy_1^{k+1}\in\partial F^*(-\vz_1^{k+1})\\
\Longleftrightarrow~&\vz_1^{k+1} =\argmin_{\vz} F^*(-\vz)-\langle\vz,\vB\vy_1^{k+1}-\vb\rangle+\lambda/2\|\vz-\vz_1^k\|_2^2\\
\Longleftrightarrow~&\vz_1^{k+1} =\argmin_{\vz} f^*(-\vA^*\vz)-\langle\vz,\vB\vy_4^{k+1}-\vb\rangle+\lambda/2\|\vz-\vu_4^k\|_2^2,
\end{align*}
where the last line is the second step of Algorithm~\ref{alg:primal-dual}. Therefore, we have $\vz_1^{k+1}=\vu_4^{k+1}$ and $\vA\vx_1^{k+1}=\lambda (\vz_1^{k+1}-\vz_1^{k})+\vb-\vB\vy_1^{k+1}=\lambda (\vu_4^{k+1}-\vu_4^{k})+\vb-\vB\vy_4^{k+1}$.
\end{proof}
\section{Equivalence of ADM for different orders}\label{sec:ADM-equal2}
In both problem \eqref{for:P3-ADMM} and Algorithm ~\ref{alg:ADM-1}, we can swap $\vx$ and $\vy$ and obtain Algorithm \ref{alg:ADM-4}, which is still an algorithm of ADM. In general, the two algorithms are different. In this section, we show that for a certain type of functions $f$ (or $g$), Algorithms~\ref{alg:ADM-1} and \ref{alg:ADM-4} become equivalent.
\begin{algorithm}[H]
\caption{ADM2 on~\eqref{for:P3-ADMM}}
\label{alg:ADM-4}
\begin{algorithmic}
\STATE initialize $\vy_5^0$, $\vz_5^0$, $\lambda>0$
\FOR {$k =0,1, \cdots$}
\STATE $\vx_5^{k+1} = \argmin\limits_\vx f(\vx)+(2\lambda)^{-1}\|\vA\vx+\vB\vy_5^{k}-\vb+\lambda\vz_5^k\|_2^2$
\STATE $\vy_5^{k+1} = \argmin\limits_\vy g(\vy)+(2\lambda)^{-1}\|\vA\vx_5^{k+1}+\vB\vy-\vb+\lambda\vz_5^k\|_2^2$
\STATE $\vz_5^{k+1} = \vz_5^k+\lambda^{-1}(\vA\vx_5^{k+1}+\vB\vy_5^{k+1}-\vb)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
The assumption that we need is that either $\prox_{ F(\cdot)}$ or $\prox_{ G(\cdot)}$ is affine (cf. \eqref{for:subprobs} for the definitions of $F$ and $G$).
\begin{definition} A mapping $T$ is affine if, for any $\vr_1$ and $\vr_2$,
\begin{align*}
T\left(\frac{1}{2}\vr_1+\frac{1}{2}\vr_2\right)=\frac{1}{2}T\vr_1+\frac{1}{2}T\vr_2.
\end{align*}
\end{definition}
\begin{proposition} Let $\lambda>0$. The following statements are equivalent:
\begin{enumerate}
\item $\prox_{G(\cdot)}$ is affine;
\item $\prox_{\lambda G(\cdot)}$ is affine;
\item $a\prox_{G(\cdot)}\circ b\vI+c\vI$ is affine for any scalars $a$, $b$ and $c$;
\item $\prox_{G^*(\cdot)}$ is affine;
\item {$G$ is convex quadratic (or, affine or constant) and its domain $\dom(G)$ is either $\cG$ or the intersection of hyperplanes in $\cG$.}
\end{enumerate}
{In addition, if function $g$ is convex quadratic and its domain is the intersection of hyperplanes, then function $G$ defined in \eqref{for:subprobG} satisfies Part 5 above.}
\end{proposition}
\begin{proposition}If $\prox_{G(\cdot)}$ is affine, then the following holds for any $\vr_1$ and $\vr_2$:
\begin{align}\label{for:prxg}
\prox_{G(\cdot)}(2\vr_1-\vr_2)=2\prox_{G(\cdot)}\vr_1-\prox_{G(\cdot)}\vr_2.
\end{align}
\end{proposition}
\begin{proof}
Equation \eqref{for:prxg} is obtained by defining $\bar{\vr}_1=2\vr_1-\vr_2$ and $\bar{\vr}_2:=\vr_2$ and rearranging
\begin{align*}\prox_{G(\cdot)}\left(\frac{1}{2}\bar{\vr}_1 +\frac{1}{2}\bar{\vr}_2\right) =\frac{1}{2}\prox_{G(\cdot)}\bar{\vr}_1+\frac{1}{2}\prox_{G(\cdot)}\bar{\vr}_2. \end{align*}
\end{proof}
\begin{theorem}[Equivalence of Algorithms~\ref{alg:ADM-1} and~\ref{alg:ADM-4}] \label{thm:ADMM-equal2}
\ \\
\begin{enumerate}
\item \label{swaporder_equiv_part1}Assume that $\prox_{\lambda G(\cdot)}$ is affine. Given the sequences $\vy_5^{k}$, $\vz_5^{k}$, and $\vx_5^{k}$ of Algorithm~\ref{alg:ADM-4}, if $\vy_5^0$ and $\vz_5^0$ satisfy $-\vz_5^0\in \partial G(\vB\vy_5^0-\vb)$, then
we can initialize Algorithm~\ref{alg:ADM-1} with $\vx_1^0=\vx_5^1$ and $\vz_1^0=\vz_5^0+\lambda^{-1}(\vA\vx_5^{1}+\vB\vy_5^{0}-\vb)$, and recover the sequences $\vx_1^k$ and $\vz_1^k$ of Algorithm~\ref{alg:ADM-1} through
\begin{subequations}\label{for:ADM-order}
\begin{align}
\vx_1^k&=\vx_5^{k+1},\\
\vz_1^k&=\vz_5^{k}+\lambda^{-1}(\vA\vx_5^{k+1}+\vB\vy_5^{k}-\vb).
\end{align}
\end{subequations}
\item Assume that $\prox_{\lambda F(\cdot)}$ is affine. Given the sequences $\vx_1^{k}$, $\vz_1^{k}$, and $\vy_1^{k}$ of Algorithm~\ref{alg:ADM-1}, if $\vx_1^0$ and $\vz_1^0$ satisfy $-\vz_1^0\in \partial F(\vA\vx_1^0)$, then we can initialize Algorithm~\ref{alg:ADM-4} with $\vy_5^0=\vy_1^1$ and $\vz_5^0=\vz_1^0+\lambda^{-1}(\vA\vx_1^{0}+\vB\vy_1^{1}-\vb)$, and recover the sequences $\vy_5^k$ and $\vz_5^k$ of Algorithm~\ref{alg:ADM-4} through
\begin{subequations}\label{for:ADM-order2}
\begin{align}
\vy_5^k&=\vy_1^{k+1},\\
\vz_5^k&=\vz_1^{k}+\lambda^{-1}(\vA\vx_1^{k}+\vB\vy_1^{k+1}-\vb).
\end{align}
\end{subequations}
\end{enumerate}
\end{theorem}
\begin{proof} We prove Part \ref{swaporder_equiv_part1} only by induction. (The proof for the other part is similar.)
The initialization of Algorithm~\ref{alg:ADM-1} clearly follows~\eqref{for:ADM-order} at $k=0$. Suppose that~\eqref{for:ADM-order} holds at $k\ge 0$. We shall show that~\eqref{for:ADM-order} holds at $k+1$. We first show from the affine property of $\prox_{\lambda G(\cdot)}$ that \begin{align}\label{for:thm2-y}
\vB\vy_1^{k+1}=2\vB\vy_5^{k+1}-\vB\vy_5^{k}.
\end{align}
The optimization subproblems for $\vy_1$ and $\vy_5$ in Algorithms~\ref{alg:ADM-1} and~\ref{alg:ADM-4}, respectively, are as follows:
\begin{align*}
\vy_1^{k+1} &= \argmin\limits_\vy g(\vy)+(2\lambda)^{-1}\|\vA\vx_1^{k}+\vB\vy-\vb+\lambda \vz_1^k\|_2^2,\\
\vy_5^{k+1} &= \argmin\limits_\vy g(\vy) +(2\lambda)^{-1}\|\vA\vx_5^{k+1}+\vB\vy-\vb+\lambda \vz_5^k\|_2^2.
\end{align*}
Following the definition of $G$ in \eqref{for:subprobs}, we have
\begin{subequations}\label{for:bBy}
\begin{align}
\vB\vy_1^{k+1}&-\vb = \prox_{\lambda G(\cdot)}(-\vA\vx_1^{k}-\lambda \vz_1^k),\\
\vB\vy_5^{k+1}&-\vb = \prox_{\lambda G(\cdot)}(-\vA\vx_5^{k+1}-\lambda \vz_5^k),\\
\vB\vy_5^{k}\quad&-\vb = \prox_{\lambda G(\cdot)}(-\vA\vx_5^{k}-\lambda \vz_5^{k-1}).\label{for:bByc}
\end{align}
\end{subequations}
The third step of Algorithm~\ref{alg:ADM-4} is
\begin{align}\label{for:alg4step3}\vz_5^{k}= \vz_5^{k-1}+\lambda^{-1}(\vA\vx^k_5+\vB\vy^k_5-\vb).\end{align}
(Note that for $k=0$, the assumption $-\vz_5^0\in \partial G(\vB\vy_5^0-\vb)$ ensures the existence of $\vz_5^{-1}$ in \eqref{for:bByc} and \eqref{for:alg4step3}.)
Then, \eqref{for:ADM-order} and \eqref{for:alg4step3} give us
\begin{align*}
\vA\vx_1^{k}+\lambda\vz_1^{k}&\overset{\eqref{for:ADM-order}}{=}\vA\vx_5^{k+1}+\lambda\vz_5^{k}+\vA\vx_5^{k+1}+\vB\vy_5^{k}-\vb\\
&\ =\ 2(\vA\vx_5^{k+1}+\lambda\vz_5^{k})-(\lambda\vz_5^{k}-\vB\vy_5^{k}+\vb)\\
&\overset{\eqref{for:alg4step3}}{=}\ 2(\vA\vx_5^{k+1}+\lambda\vz_5^{k})-(\vA\vx_5^{k}+\lambda\vz_5^{k-1}).
\end{align*}
Since $\prox_{\lambda G(\cdot)}$ is affine, we have \eqref{for:prxg}. Once we plug in \eqref{for:prxg}: $\vr_1=-\vA\vx_5^{k+1}-\lambda\vz_5^{k}$, $\vr_2=-\vA\vx_5^{k}-\lambda\vz_5^{k-1}$, and $2\vr_1-\vr_2 = -\vA\vx_1^{k}-\lambda\vz_1^k$ and then apply \eqref{for:bBy}, we obtain~\eqref{for:thm2-y}.
Next, the third step of Algorithm~\ref{alg:ADM-4} and \eqref{for:thm2-y} give us
\begin{align*}
\vB\vy_1^{k+1}-\vb+\lambda \vz_1^k &\overset{\eqref{for:thm2-y}}{=} 2(\vB\vy_5^{k+1}-\vb)-(\vB\vy_5^{k}-\vb)+\lambda\vz_5^{k}+(\vA\vx_5^{k+1}+\vB\vy_5^{k}-\vb)\\
&\ =\ (\vB\vy_5^{k+1}-\vb)+\lambda\vz_5^{k}+(\vA\vx_5^{k+1}+\vB\vy_5^{k+1}-\vb)\\
&\ =\ (\vB\vy_5^{k+1}-\vb)+\lambda\vz_5^{k+1}.
\end{align*}
This identity shows that the updates of $\vx_1^{k+1}$ and $\vx_5^{k+2}$ in Algorithms~\ref{alg:ADM-1} and~\ref{alg:ADM-4}, respectively, have identical data, and therefore, we recover $\vx_1^{k+1}=\vx_5^{k+2}$.
Lastly, from the third step of Algorithm~\ref{alg:ADM-1} and the identities above, it follows that
\begin{align*}
\vz_1^{k+1}&=\vz_1^k+\lambda^{-1}(\vA\vx_1^{k+1}+\vB\vy_1^{k+1}-\vb)\\
&=\vz_1^k+\lambda^{-1}\left(\vA\vx_5^{k+2}+(\vB\vy_5^{k+1}-\vb+\lambda\vz_5^{k+1}-\lambda\vz_1^k)\right)\\
&=\vz_5^{k+1}+\lambda^{-1}(\vA\vx_5^{k+2}+\vB\vy_5^{k+1}-\vb).
\end{align*}
Therefore, we obtain \eqref{for:ADM-order} at $k+1$.
\end{proof}
\begin{remark} We can avoid the technical condition $-\vz_5^0\in \partial G(\vB\vy_5^0-\vb)$ on Algorithm~\ref{alg:ADM-4} in Part~\ref{swaporder_equiv_part1} of Theorem \ref{thm:ADMM-equal2}. When it does not hold, we can use the always-true relation $-\vz_5^1\in \partial G(\vB\vy_5^1-\vb)$ instead; correspondingly, we shall add 1 iteration to the iterates of Algorithm~\ref{alg:ADM-4}, namely, initialize Algorithm~\ref{alg:ADM-1} with $\vx_1^0=\vx_5^2$ and $\vz_1^0=\vz_5^1+\lambda^{-1}(\vA\vx_5^{2}+\vB\vy_5^{1}-\vb)$ and recover the sequences $\vx_1^k$ and $\vz_1^k$ of Algorithm~\ref{alg:ADM-1} through
\begin{subequations}\label{for:ADM-order_offset}
\begin{align}
\vx_1^k&=\vx_5^{k+2},\\
\vz_1^k&=\vz_5^{k+1}+\lambda^{-1}(\vA\vx_5^{k+2}+\vB\vy_5^{k+1}-\vb).
\end{align}
\end{subequations}
Similar arguments apply to the other part of Theorem \ref{thm:ADMM-equal2}.
\end{remark}
\section{Primal-dual equivalence of RPRS}
\label{sec:DRS}
In this section, we consider the following convex problem:
\begin{align}\tag{P3}\label{for:P1-DRS}
\Min_\vx \quad f(\vx)+g(\vA\vx),
\end{align}
and its corresponding Lagrangian dual
\begin{align}\tag{D3}\label{for:D1-DRS}
\Min_\vv \quad f^*(\vA^*\vv) + g^*(-\vv).
\end{align}
In addition, we introduce another primal-dual pair equivalent to~\eqref{for:P1-DRS}-\eqref{for:D1-DRS}:
\begin{align}\tag{P4}\label{for:P2-DRS}
\Min_\vy &\quad (f^*\circ \vA^*)^*(\vy)+g(\vy),\\
\tag{D4}\label{for:D2-DRS}
\Min_\vu &\quad f^*(\vu) + (g\circ \vA)^*(-\vu).
\end{align}
Lemma~\ref{lemma:Equivalency} below will establish the equivalence between the two primal-dual pairs.
\begin{remark}When $\vA=\vI$, we have $(f^*\circ \vA^*)^*=f$, and problem~\eqref{for:P1-DRS} is exactly the same as problem~\eqref{for:P2-DRS}. Similarly, problem~\eqref{for:D1-DRS} is exactly the same as problem~\eqref{for:D2-DRS}.
\end{remark}
\begin{lemma}\label{lemma:Equivalency}
Problems~\eqref{for:P1-DRS} and~\eqref{for:P2-DRS} are equivalent in the following sense:
\begin{itemize}
\item Given any solution $\vx^*$ to~\eqref{for:P1-DRS}, $\vy^*=\vA\vx^*$ is a solution to~\eqref{for:P2-DRS},
\item Given any solution $\vy^*$ to~\eqref{for:P2-DRS}, $\vx^*\in\argmin\limits_{\vx:\vA\vx=\vy^*} f(\vx)$ is a solution to~\eqref{for:P1-DRS}.
\end{itemize}
The equivalence between problems~\eqref{for:D1-DRS} and~\eqref{for:D2-DRS} is similar:
\begin{itemize}
\item Given any solution $\vv^*$ to~\eqref{for:D1-DRS}, $\vA^*\vv^*$ is a solution to~\eqref{for:D2-DRS},
\item Given any solution $\vu^*$ to~\eqref{for:D2-DRS}, $\vv^*\in\argmin\limits_{\vv:\vA^*\vv=\vu^*} g^*(-\vv)$ is a solution to~\eqref{for:D1-DRS}.
\end{itemize}
\end{lemma}
\begin{proof}
We prove only the equivalence of~\eqref{for:P1-DRS} and~\eqref{for:P2-DRS}, the proof for the equivalence of~\eqref{for:D1-DRS} and~\eqref{for:D2-DRS} is similar.
Part 1: If $\vx^*$ is a solution to~\eqref{for:P1-DRS}, we have $\vzero\in\partial f(\vx^*)+\vA^*\partial g(\vA\vx^*)$. Assume that there exists $\vq$ such that $-\vq\in\partial g(\vA\vx^*)$ and $\vA^*\vq\in\partial f(\vx^*)$. Then we have
\begin{align*}
\vA^*\vq\in\partial f(\vx^*) \Longleftrightarrow & \vx^*\in\partial f^*(\vA^*\vq)\\
\Longrightarrow & \vA\vx^* \in \vA\partial f^*(\vA^*\vq)= \partial (f^*\circ \vA^*)(\vq)\\
\Longleftrightarrow & \vq \in \partial(f^*\circ \vA^*)^*(\vA\vx^*).
\end{align*}
Therefore,
\begin{align*}
\vzero \in \partial(f^*\circ \vA^*)^*(\vA\vx^*) +\partial g(\vA\vx^*)
\end{align*}
and $\vA\vx^*$ is a solution to~\eqref{for:P2-DRS}.
Part 2: If $\vy^*$ is a solution to~\eqref{for:P2-DRS}, the optimality condition gives us
\begin{align*}
\vzero \in \partial(f^*\circ \vA^*)^*(\vy^*) +\partial g(\vy^*).
\end{align*}
Assume that there exists $\vq$ such that $-\vq\in\partial g(\vy^*)$ and $\vq\in \partial(f^*\circ \vA^*)^*(\vy^*)$. Then we have
\begin{align}\label{for:PD-equal-1}
\vq \in \partial(f^*\circ \vA^*)^*(\vy^*)\Longleftrightarrow &\vy^* \in \partial (f^*\circ \vA^*)(\vq).
\end{align}
Consider the following optimization problem for finding $\vx^*$ from $\vy^*$
\begin{align*}
\Min_\vx f(\vx)\qquad\mbox{subject to }\vA\vx=\vy^*,
\end{align*}
and the corresponding dual problem
\begin{align*}
\Max_\vv -f^*(\vA^*\vv)+\langle\vv,\vy^*\rangle.
\end{align*}
It is easy to obtain from~\eqref{for:PD-equal-1} that $\vq$ is a solution of the dual problem. The optimal duality gap is zero and the strong duality gives us
\begin{align}
f(\vx^*)=f(\vx^*)-\langle \vq,\vA\vx^*-\vy^*\rangle = -f^*(\vA^*\vq)+\langle \vq,\vy^*\rangle.
\end{align}
Thus $\vx^*$ is a solution of $\Min\limits_\vx f(\vx)-\langle \vA^*\vq, \vx\rangle$ and
\begin{align}
\vA^*\vq \in \partial f(\vx^*)\Longleftrightarrow \vzero\in \partial f(\vx^*)-\vA^*\vq.
\end{align}
Because $-\vq\in\partial g(\vy^*)=\partial g(\vA\vx^*)$,
\begin{align}
\vzero\in \partial f(\vx^*)+\vA^*\partial\vg(\vA\vx^*)= \partial f(\vx^*)+\partial(\vg\circ \vA)(\vx^*)
\end{align}
Therefore $\vx^*$ is a solution of~\eqref{for:P1-DRS}.
\end{proof}
Next we will show the equivalence between the RPRS to the primal and dual problems:
\begin{align*}
\boxed{\mbox{RPRS on~\eqref{for:P1-DRS}}}\Longleftrightarrow \boxed{\mbox{RPRS on~\eqref{for:D2-DRS}}}\\
\boxed{\mbox{RPRS on~\eqref{for:P2-DRS}}}\Longleftrightarrow \boxed{\mbox{RPRS on~\eqref{for:D1-DRS}}}
\end{align*}
We describe the RPRS on~\eqref{for:P1-DRS} in Algorithm~\ref{alg:DRS-P4}, and the RPRS on other problems can be obtained in the same way.
\begin{algorithm}[H]
\caption{RPRS on~\eqref{for:P1-DRS}}
\label{alg:DRS-P4}
\begin{algorithmic}
\STATE initialize $\vw^0$, $\lambda>0$, $0<\alpha\leq1$.
\FOR {$k =0,1, \cdots$}
\STATE $\vx^{k+1} = \prox_{\lambda f(\cdot)} \vw^k$
\STATE $\vw^{k+1}=(1-\alpha) \vw^k+\alpha(2\prox_{\lambda g\circ \vA(\cdot)}-\vI)(2\vx^{k+1}-\vw^k)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{theorem}\label{thm:DRS-equal}[Primal-dual equivalence of RPRS] RPRS on~\eqref{for:P1-DRS} is equivalent to RPRS on~\eqref{for:D2-DRS}. RPRS on~\eqref{for:P2-DRS} is equivalent to RPRS on~\eqref{for:D1-DRS}.
\end{theorem}
Before proving this theorem, we introduce a lemma, which was also given in~\cite[Proposition~3.34]{eckstein1989splitting}. Here, we prove it in a different way using the generalized Moreau decomposition.
\begin{lemma}\label{lemma:2}
For $\lambda>0$, we have
\begin{align}\label{for:PD-lemma}
\lambda^{-1}(2\prox_{\lambda F(\cdot)}-\vI)\vw = (\vI-2\prox_{\lambda^{-1}F^*(\cdot)})(\vw/\lambda)= (2\prox_{\lambda^{-1}F^*(-\cdot)}-\vI)(-\vw/\lambda).
\end{align}
\end{lemma}
\begin{proof}
We prove it using the generalized Moreau decomposition~\cite[Theorem 2.3.1]{esser_primal_2010}
\begin{align}\label{for:Moreau-decomp}
\vw = \prox_{\lambda F(\cdot)}(\vw) + \lambda \prox_{\lambda^{-1}F^*(\cdot)}(\vw/\lambda).
\end{align}
Using the generalized Moreau decomposition, we have
\begin{align*}
\lambda^{-1}(2\prox_{\lambda F(\cdot)}-\vI)\vw =&2\lambda^{-1}\prox_{\lambda F(\cdot)}(\vw)-\vw/\lambda\overset{\eqref{for:Moreau-decomp}}{=}2\lambda^{-1}(\vw-\lambda\prox_{\lambda^{-1}F^*(\cdot)}(\vw/\lambda))-\vw/\lambda\\
=&\vw/\lambda-2\prox_{\lambda^{-1}F^*(\cdot)}(\vw/\lambda) = (\vI-2\prox_{\lambda^{-1}F^*(\cdot)})(\vw/\lambda).
\end{align*}
The last equality of~\eqref{for:PD-lemma} comes from
\begin{align*}
\prox_{\lambda^{-1}F^*(-\cdot)}(-\vw/\lambda) = -\prox_{\lambda^{-1}F^*(\cdot)}(\vw/\lambda).
\end{align*}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:DRS-equal}]
We will prove only the equivalence of RPRS on~\eqref{for:P1-DRS} and~\eqref{for:D2-DRS}. The proof for the other equivalence is the same. The RPRS on~\eqref{for:P1-DRS} and~\eqref{for:D2-DRS} can be formulated as
\begin{align}\label{for:P-DRS}
\vw_1^{k+1}&=(1-\alpha) \vw_1^k+\alpha(2\prox_{\lambda g\circ \vA(\cdot)}-\vI)(2\prox_{\lambda f(\cdot)}-\vI)\vw_1^k,
\end{align}
and
\begin{align}\label{for:D-DRS}
\vw_2^{k+1}&=(1-\alpha) \vw_2^k+\alpha(2\prox_{\lambda^{-1}(g\circ \vA)^*(-\cdot)}-\vI)(2\prox_{\lambda^{-1}f^*(\cdot)}-\vI)\vw_2^k,
\end{align}
respectively. In addition, we can recover the variables $\vx^k$ (or $\vv^k$) from $\vw_1^k$ (or $\vw_2^k$) using the following forms:
\begin{align}
\vx^{k+1} & = \prox_{\lambda f(\cdot)} \vw_1^{k},\\
\vv^{k+1} & = \prox_{\lambda^{-1}f^*(\cdot)} \vw_2^{k}.
\end{align}
Proof by induction. Suppose $\vw_2^k = \vw_1^k/\lambda$ holds. We next show that $\vw_2^{k+1} = \vw_1^{k+1}/\lambda$.
\begin{align*}
\vw_2^{k+1}=&(1-\alpha) \vw_1^k/\lambda+\alpha(2\prox_{\lambda^{-1}(g\circ \vA)^*(-\cdot)}-\vI)(2\prox_{\lambda^{-1}f^*(\cdot)}-\vI)(\vw_1^k/\lambda)\\
\overset{\eqref{for:PD-lemma}}{=}&(1-\alpha) \vw_1^k/\lambda+\alpha(2\prox_{\lambda^{-1}(g\circ \vA)^*(-\cdot)}-\vI)(-\lambda^{-1}(2\prox_{\lambda f(\cdot)}-\vI)\vw_1^k)\\
\overset{\eqref{for:PD-lemma}}{=}&(1-\alpha) \vw_1^k/\lambda+\alpha\lambda^{-1}(2\prox_{\lambda(g\circ \vA)(-\cdot)}-\vI)(2\prox_{\lambda f(\cdot)}-\vI)\vw_1^k\\
=&\lambda^{-1}[(1-\alpha) \vw_1^k+\alpha(2\prox_{\lambda(g\circ \vA)(\cdot)}-\vI)(2\prox_{\lambda f(\cdot)}-\vI)\vw_1^k]\\
=&\vw_1^{k+1}/\lambda.
\end{align*}
In addition we have
\begin{align*}
\vx^{k+1} +\lambda\vv^{k+1}=&\prox_{\lambda f(\cdot)} \vw_1^{k}+\lambda\prox_{\lambda^{-1}f^*(\cdot)} \vw_2^{k}\\
=&\prox_{\lambda f(\cdot)} \vw_1^{k}+\lambda\prox_{\lambda^{-1}f^*(\cdot)} (\vw_1^{k}/\lambda)=\vw_1^{k}.
\end{align*}
\end{proof}
\begin{remark}Eckstein showed in~\cite[Chapter~3.5]{eckstein1989splitting} that DRS/PRS on~\eqref{for:P1-DRS} is equivalent to DRS/PRS on~\eqref{for:D1-DRS} when $\vA=\vI$. This special case can be obtained from this theorem immediately because when $\vA=\vI$, \eqref{for:D1-DRS} is exactly the same as~\eqref{for:D2-DRS} and we have
\begin{align*}
\boxed{\textnormal{DRS/PRS on~\eqref{for:P1-DRS}}}\Longleftrightarrow \boxed{\textnormal{DRS/PRS on~\eqref{for:D2-DRS}}}\Longleftrightarrow \boxed{\textnormal{DRS/PRS on~\eqref{for:D1-DRS}}}\Longleftrightarrow\boxed{\textnormal{DRS/PRS on~\eqref{for:P2-DRS}}}.
\end{align*}
\end{remark}
\begin{remark}In order to make sure that RPRS on the primal and dual problems are equivalent, the initial conditions and parameters have to satisfy conditions described in the proof of Theorem~\ref{thm:DRS-equal}. We need the initial condition to satisfy $\vw_2^0=\vw_1^0/\lambda$ and the parameter for RPRS on the dual problem has to be chosen as $\lambda^{-1}$, see the differences in~\eqref{for:P-DRS} and~\eqref{for:D-DRS}.
\end{remark}
\section{Application: total variation image denoising}
\label{sec:app1}
ADM (or split Bregman~\cite{SplitBregman}) has been applied on many image processing applications, and we apply the previous equivalence results of ADM to derive several equivalent algorithms for total variation denoising.
The total variation (ROF model~\cite{ROF}) applied on image denoising is
\begin{align*}
\Min_{x\in BV(\Omega)} \int_\Omega |D x|+ {\alpha\over2}\| x-b\|_2^2
\end{align*}
where $x$ stands for an image, and $BV(\Omega)$ is the set of all bounded variation functions on $\Omega$. The first term is known as the total variation of $x$, minimizing which tends to yield a piece-wise constant solution. The discrete version is as follows:
\begin{align*}
\Min_{\vx} \|\nabla \vx\|_{2,1}+ {\alpha\over2}\| \vx-\vb\|_2^2.
\end{align*}
Without loss of generality, we consider the two-dimensional image $\vx$, and the discrete total variation $\|\nabla\vx\|_{2,1}$ of image $\vx$ is defined as
\begin{align*}
\|\nabla\vx\|_{2,1}=\sum_{ij}|(\nabla \vx)_{ij}|,
\end{align*}
where $|\cdot|$ is the 2-norm of a vector. The equivalent ADM-ready form~\cite[Equation (3.1)]{SplitBregman} is
\begin{align}\label{TVD-P}
\Min_{\vx,\vy} \ \|\vy\|_{2,1}+ {\alpha\over2}\| \vx-\vb\|_2^2\qquad\st \ \vy-\nabla \vx=\vzero,
\end{align}
and its dual problem in ADM-ready form~\cite[Equation (8)]{Chambolle} is
\begin{align}\label{TVD-D}
\Min_{\vv,\vu} \ {1\over2\alpha}\| \textnormal{div }\vu+\alpha\vb\|_2^2+\iota_{\{\vv:{\|\vv\|_{2,\infty}\leq1}\}}(\vv)\qquad \st \ \vu-\vv=\vzero,
\end{align}
where $\|\vv\|_{2,\infty}=\max\limits_{ij}|(\vv)_{ij}|$. In addition, the equivalent saddle-point problem is
\begin{align}\label{TVD-PD}
\Min_{\vx}\Max_{\vv}\quad {1\over2\alpha}\|\vx-\vb\|_2^2+\langle \vv,\nabla \vx\rangle-\iota_{\{\vv:{\|\vv\|_{2,\infty}\leq1}\}}(\vv).
\end{align}
We list the following equivalent algorithms for solving the total variation image denoising problem. The equivalence result stated in Corollary~\ref{cor:TV} can be obtained from theorems~\ref{thm:ADMM-equal}-\ref{thm:ADMM-equal2}.
\begin{enumerate}
\item Algorithm~\ref{alg:ADM-1} (primal ADM) on~\eqref{TVD-P} is
\begin{subequations}\label{eqn:TV-P}
\begin{align}
\vx_1^{k+1} = &\argmin_{\vx}{\alpha\over2}\|\vx-\vb\|_2^2+(2\lambda)^{-1}\|\nabla\vx-\vy_1^k+\lambda\vz_1^k\|_2^2,\\
\vy_1^{k+1} = &\argmin_{\vy}\|\vy\|_{2,1}+(2\lambda)^{-1}\|\nabla\vx_1^{k+1}-\vy+\lambda\vz_1^k\|_2^2,\\
\vz_1^{k+1}=&\vz_1^k+\lambda^{-1}(\nabla\vx_1^{k+1}-\vy_1^{k+1}).
\end{align}
\end{subequations}
\item Algorithm~\ref{alg:ADM-3} (dual ADM) on~\eqref{TVD-D} is
\begin{subequations}\label{eqn:TV-D}
\begin{align}
\vu_3^{k+1} = &\argmin_{\vu}{{1\over2\alpha}\|\textnormal{div }\vu+\alpha\vb\|_2^2+{\lambda\over2}\|\vv_3^k-\vu+\lambda^{-1}\vz_3^k\|_2^2},\\
\vv_3^{k+1} = &\argmin_{\vv}\iota_{\{\vv:\|\vv\|_{2,\infty}\leq1\}}(\vv)+{\lambda\over 2}\|\vv-\vu_3^{k+1}+\lambda^{-1}\vz_3^k\|_2^2,\\
\vz_3^{k+1} = & \vz_3^k+\lambda(\vv_3^{k+1}-\vu_3^{k+1}).
\end{align}
\end{subequations}
\item Algorithm~\ref{alg:primal-dual} (primal-dual) on~\eqref{TVD-PD} is
\begin{subequations}\label{eqn:TV-PD}
\begin{align}
\bar\vv_4^{k} = & 2\vv_4^{k}-\vv_4^{k-1},\\
\vx_4^{k+1} = &\argmin_{\vx}{\alpha\over2}\|\vx-\vb\|_2^2+(2\lambda)^{-1}\|\nabla\vx-\nabla \vx_4^k+\lambda\bar\vv_4^k\|_2^2,\\
\vv_4^{k+1} = &\argmin_{\vv}\iota_{\{\vv:\|\vv\|_{2,\infty}\leq1\}}(\vv)-\langle\vv,\nabla\vx_4^{k+1}\rangle+{\lambda\over 2}\|\vv-\vv^k\|_2^2.
\end{align}
\end{subequations}
\item Algorithm~\ref{alg:ADM-4} (primal ADM with order swapped) on~\eqref{TVD-P} is
\begin{subequations}\label{eqn:TV-P2}
\begin{align}
\vy_5^{k+1} = &\argmin_{\vy}\|\vy\|_{2,1}+(2\lambda)^{-1}\|\nabla\vx_5^{k}-\vy+\lambda\vz_5^k\|_2^2,\\
\vx_5^{k+1} = &\argmin_{\vx}{\alpha\over2}\|\vx-\vb\|_2^2+(2\lambda)^{-1}\|\nabla\vx-\vy_5^{k+1}+\lambda\vz_5^k\|_2^2,\\
\vz_5^{k+1}=&\vz_5^k+\lambda^{-1}(\nabla\vx_5^{k+1}-\vy_5^{k+1}).
\end{align}
\end{subequations}
\end{enumerate}
\begin{corollary}\label{cor:TV}
Let $\vx_5^0=\vb+\alpha^{-1}\textnormal{div }\vz_5^0$. If the initialization for all algorithms~\eqref{eqn:TV-P}-\eqref{eqn:TV-P2} satisfies $\vy_1^0 = - \vz_3^0 =\nabla \vx_4^0-\lambda (\vv_4^0-\vv_4^{-1}) =\vy_5^{1}$ and $\vz_1^0 = \vv_3^0 =\vv_4^0 = \vz_5^0+\lambda^{-1}(\nabla \vx_5^0-\vy_5^{1})$. Then for $k\geq1$, we have the following equivalence results between the iterations of the four algorithms:
\begin{align*}
\begin{array}{llll}
\vy_1^k &= - \vz_3^k &=\nabla \vx_4^k-\lambda (\vv_4^k-\vv_4^{k-1}) &=\vy_5^{k+1},\\
\vz_1^k &= \vv_3^k &=\vv_4^k &= \vz_5^k+\lambda^{-1}(\nabla \vx_5^k-\vy_5^{k+1}).
\end{array}
\end{align*}
\end{corollary}
\begin{remark} In any of the four algorithms, the $\nabla$ or $\textnormal{div}$ operator is separated in a different subproblem from the term $\|\cdot\|_{2,1}$ or its dual norm $\|\cdot\|_{2,\infty}$. The $\nabla$ or $\textnormal{div }$ operator is translation invariant so their subproblems can be solved by a diagonalization trick~\cite{FTVd}. The subproblems involving the term $\|\cdot\|_{2,1}$ or the indicator function $\iota_{\{\vv:\|\vv\|_{2,\infty}\leq1\}}$ have closed-form solutions. Therefore, in addition to the equivalence results, all the four algorithms have essentially the same per-iteration costs.
\end{remark}
\section*{Acknowledgments}
This work is supported by NSF Grants DMS-1349855 and DMS-1317602 and ARO MURI Grant W911NF-09-1-0383. We thank Jonathan Eckstein for bringing his early work~\cite[Chapter~3.5]{eckstein1989splitting} and~\cite{eckstein1994some} to our attention.
\bibliographystyle{siam}
|
\section{Introduction\label{sec1}}
The observations of neutrino oscillations (see Refs.~\cite{neu-b1,neu-b2,neu-b3,neu-b4,neu-data}) imply that neutrinos have tiny masses and are mixed, which have demonstrated that lepton flavor in neutrino sector is not conserved. Nevertheless, in the Standard Model (SM) with massive neutrinos, the expected rates for the charged lepton flavor violating (LFV) processes are
very tiny, for instance ${\rm{Br}}(\mu\rightarrow e\gamma)<10^{-54}$~\cite{uer-SM1,uer-SM2} and ${\rm{Br}}(Z\rightarrow l_i^\pm l_j^\mp)\sim10^{-54}$~\cite{Zll-vSM1,Zll-vSM2,Zll-vSM3,Zll-vSM4}, which are far from the experimental reach. In Table~\ref{tab1}, we show the present experimental limits and future sensitivities for some LFV processes~\cite{Z-exp1,Z-exp2,Z-exp3,uer-exp1,uer-exp2,t-exp1,t-exp2,u3e-exp1,u3e-exp2,t-exp3,ueN-exp1,ueN-exp2}. Thus, any signal of LFV in charged lepton sector would be a hint of new physics.
\begin{table*}[htbp]
\caption{Present experimental limits and future sensitivities for some LFV processes.}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}llllll@{}}
\hline
LFV process & Present limit & Future sensitivity \\
\hline
$Z\rightarrow e\mu$ & $<1.7\times10^{-6}$ \cite{Z-exp1} & $\sim2.0\times10^{-9}$ \cite{Z-exp3} \\
$Z\rightarrow e\tau$ & $<9.8\times10^{-6}$ \cite{Z-exp1} & $\sim(1.3-6.5)\times10^{-8}$ \cite{Z-exp3} \\
$Z\rightarrow \mu\tau$ & $<1.2\times10^{-5}$ \cite{Z-exp2} & $\sim(0.44-2.2)\times10^{-8}$ \cite{Z-exp3} \\
$\mu\rightarrow e\gamma$ & $<5.7\times10^{-13}$ \cite{uer-exp1} & $\sim6\times10^{-14}$ \cite{uer-exp2} \\
$\tau\rightarrow e\gamma$ & $<3.3\times10^{-8}$ \cite{t-exp1} & $\sim10^{-8}-10^{-9}$ \cite{t-exp2} \\
$\tau\rightarrow \mu\gamma$ & $<4.4\times10^{-8}$ \cite{t-exp1} & $\sim10^{-8}-10^{-9}$ \cite{t-exp2} \\
$\mu\rightarrow 3e$ & $<1.0\times10^{-12}$ \cite{u3e-exp1} & $\sim10^{-16}$ \cite{u3e-exp2} \\
$\tau\rightarrow 3e$ & $<2.7\times10^{-8}$ \cite{t-exp3} & $\sim10^{-9}-10^{-10}$ \cite{t-exp2} \\
$\tau\rightarrow 3\mu$ & $<2.1\times10^{-8}$ \cite{t-exp3} & $\sim10^{-9}-10^{-10}$ \cite{t-exp2} \\
$\mu\rightarrow e:\rm{Ti}$ & $<6.1\times10^{-13}$ \cite{ueN-exp1} & $\sim10^{-18}$ \cite{ueN-exp2} \\
\hline
\end{tabular*}
\label{tab1}
\end{table*}
Several predictions for the LFV processes $Z\rightarrow l_i^\pm l_j^\mp$ have been obtained in the framework of various SM extensions~\cite{Zll-R1,Zll-R2,Zll-R3,Zll-R4,Zll-R5,Zll-R6,Zll-R7,Zll-R8,Zll-R9,Zll-R10,Zll-R11,Zll-R12,Zll-R13,Zll-R14,Zll-R15,Zll-R16,Zll-R17}. In this work, we investigate the processes $Z\rightarrow l_i^\pm l_j^\mp$ in the ``$\mu$ from $\nu$ Supersymmetric Standard Model'' ($\mu\nu$SSM)~\cite{mnSSM,mnSSM1,mnSSM2}. Within the $\mu\nu$SSM, nonzero vacuum expectative values (VEVs) of sneutrinos lead to R-parity and lepton number violations, and generate three tiny massive Majorana neutrinos at the tree level through the seesaw machanism~\cite{mnSSM,mnSSM1,mnSSM2,meu-m,meu-m1,meu-m2,meu-m3,neu-zhang1}. Especially, the $\mu$ problem~\cite{m-problem} of the Minimal Supersymmetric Standard Model (MSSM)~\cite{MSSM,MSSM1,MSSM2,MSSM3,MSSM4} had been solved in the $\mu\nu$SSM, via the R-parity breaking couplings ${\lambda _i}\hat \nu _i^c\hat H_d^a\hat H_u^b$ in the superpotential. The $\mu$ term is generated spontaneously through the nonzero VEVs of right-handed sneutrinos, $\mu = {\lambda _i}\left\langle {\tilde \nu _i^c} \right\rangle$, when the electroweak symmetry is broken (EWSB).
In our previous work, we had analyzed some LFV processes $l_j^-\rightarrow l_i^-\gamma$, $l_j^- \rightarrow l_i^- l_i^- l_i^+$ and muon conversion to electron in nuclei within the $\mu\nu$SSM, under minimal flavor violation assumptions~\cite{ref-zhang,ref-zhang1}. The numerical results show that the expected rates for the LFV processes under minimal flavor violation assumptions, still remain orders of magnitude below the future experimental sensitivities. In this work, we continue to analyze the LFV processes with slepton flavor mixing, including $Z\rightarrow l_i^\pm l_j^\mp$.
The paper has the following structure. In Section~\ref{sec2}, we present the $\mu\nu$SSM briefly, including its superpotential and the general soft SUSY-breaking terms. Section~\ref{sec3} contains the analytical expressions of the LFV processes $Z\rightarrow l_i^\pm l_j^\mp$. In Section~\ref{sec4}, we give the numerical analysis, under some assumptions and constraints on parameter space. The summary is given in Section~\ref{sec5}. The couplings are collected in Appendix~\ref{app-coupling}.
\section{the $\mu\nu$SSM\label{sec2}}
Besides the superfields of the MSSM, the $\mu\nu$SSM introduces three singlet right-handed neutrino superfields $\hat{\nu}_i^c\;(i=1,\;2,\;3)$. The corresponding superpotential of the $\mu\nu$SSM is given by~\cite{mnSSM}
\begin{eqnarray}
W =&&{\epsilon _{ab}} \Big( {Y_{u_{ij}}}\hat H_u^b\hat Q_i^a\hat u_j^c + {Y_{d_{ij}}}\hat H_d^a\hat Q_i^b\hat d_j^c
+ {Y_{e_{ij}}}\hat H_d^a\hat L_i^b\hat e_j^c + {Y_{\nu _{ij}}}\hat H_u^b\hat L_i^a\hat \nu _j^c \Big) \nonumber\\
&&- {\epsilon _{ab}}{\lambda _i}\hat \nu _i^c\hat H_d^a\hat H_u^b + \frac{1}{3}{\kappa _{ijk}}\hat \nu _i^c\hat \nu _j^c\hat \nu _k^c \, ,
\end{eqnarray}
where $\hat H_d^T = \Big( {\hat H_d^0,\hat H_d^ - } \Big)$, $\hat H_u^T = \Big( {\hat H_u^ + ,\hat H_u^0} \Big)$, $\hat Q_i^T = \Big( {{{\hat u}_i},{{\hat d}_i}} \Big)$, $\hat L_i^T = \Big( {{{\hat \nu}_i},{{\hat e}_i}} \Big)$ are $SU(2)$ doublet superfields, and $\hat d_i^c$, $\hat u_i^c$ and $\hat e_i^c$ represent the singlet down-type quark, up-type quark
and charged lepton superfields, respectively. In addition, $Y_{u,d,\nu,e}$, $\lambda$, $\kappa$ respectively are dimensionless matrices, a vector, a totally symmetric tensor. And $a,b$ are SU(2) indices with antisymmetric tensor $\epsilon_{12}=-\epsilon_{21}=1$. In this paper, the summation convention is implied on repeated indices.
In the superpotential, the first three terms are the same as the MSSM. Next two terms can generate the effective bilinear terms $\epsilon _{ab} \varepsilon_i \hat H_u^b\hat L_i^a$, $\epsilon _{ab} \mu \hat H_d^a\hat H_u^b$, and $\varepsilon_i= Y_{\nu _{ij}} \left\langle {\tilde \nu _j^c} \right\rangle$, $\mu = {\lambda _i}\left\langle {\tilde \nu _i^c} \right\rangle$, once the electroweak symmetry is broken. The last term generates the effective Majorana masses for neutrinos at the electroweak scale. And the last two terms explicitly violate lepton number and R-parity.
In the framework of supergravity mediated supersymmetry breaking, the general soft SUSY-breaking terms in the $\mu\nu$SSM are given as
\begin{eqnarray}
- \mathcal{L}_{soft}=&&m_{{{\tilde Q}_{ij}}}^{\rm{2}}\tilde Q{_i^{a\ast}}\tilde Q_j^a
+ m_{\tilde u_{ij}^c}^{\rm{2}}\tilde u{_i^{c\ast}}\tilde u_j^c + m_{\tilde d_{ij}^c}^2\tilde d{_i^{c\ast}}\tilde d_j^c
+ m_{{{\tilde L}_{ij}}}^2\tilde L_i^{a\ast}\tilde L_j^a \nonumber\\
&&+ \; m_{\tilde e_{ij}^c}^2\tilde e{_i^{c\ast}}\tilde e_j^c + m_{{H_d}}^{\rm{2}} H_d^{a\ast} H_d^a
+ m_{{H_u}}^2H{_u^{a\ast}}H_u^a + m_{\tilde \nu_{ij}^c}^2\tilde \nu{_i^{c\ast}}\tilde \nu_j^c \nonumber\\
&&+ \; \epsilon_{ab} \Big[{{({A_u}{Y_u})}_{ij}}H_u^b\tilde Q_i^a\tilde u_j^c
+ {{({A_d}{Y_d})}_{ij}}H_d^a\tilde Q_i^b\tilde d_j^c + {{({A_e}{Y_e})}_{ij}}H_d^a\tilde L_i^b\tilde e_j^c + \textrm{H.c.} \Big] \nonumber\\
&&+ \; \Big[ {\epsilon _{ab}}{{({A_\nu}{Y_\nu})}_{ij}}H_u^b\tilde L_i^a\tilde \nu_j^c
- {\epsilon _{ab}}{{({A_\lambda }\lambda )}_i}\tilde \nu_i^c H_d^a H_u^b
+ \frac{1}{3}{{({A_\kappa }\kappa )}_{ijk}}\tilde \nu_i^c\tilde \nu_j^c\tilde \nu_k^c + \textrm{H.c.} \Big] \nonumber\\
&&- \; \frac{1}{2}\Big({M_3}{{\tilde \lambda }_3}{{\tilde \lambda }_3}
+ {M_2}{{\tilde \lambda }_2}{{\tilde \lambda }_2} + {M_1}{{\tilde \lambda }_1}{{\tilde \lambda }_1} + \textrm{H.c.} \Big)\,.
\end{eqnarray}
Here, the first two lines consist of mass squared terms of squarks, sleptons and Higgses. The next two lines contain the trilinear scalar couplings. In the last lines, $M_3$, $M_2$ and $M_1$ denote Majorana masses corresponding to $SU(3)$, $SU(2)$ and $U(1)$ gauginos $\hat{\lambda}_3$, $\hat{\lambda}_2$ and $\hat{\lambda}_1$, respectively.
In addition to the terms from $\mathcal{L}_{soft}$, the tree-level scalar potential receives the usual D and F term contributions~\cite{mnSSM1}.
Once the electroweak symmetry is spontaneously broken, the neutral scalars develop in general the following VEVs:
\begin{eqnarray}
\langle H_d^0 \rangle = \upsilon_d \,, \qquad \langle H_u^0 \rangle = \upsilon_u \,, \qquad
\langle \tilde \nu_i \rangle = \upsilon_{\nu_i} \,, \qquad \langle \tilde \nu_i^c \rangle = \upsilon_{\nu_i^c} \,.
\end{eqnarray}
Thus one can define neutral scalars as usual
\begin{eqnarray}
&&H_d^0=\frac{h_d + i P_d}{\sqrt{2}} + \upsilon_d, \qquad\; \tilde \nu_i = \frac{(\tilde \nu_i)^\Re + i (\tilde \nu_i)^\Im}{\sqrt{2}} + \upsilon_{\nu_i}, \nonumber\\
&&H_u^0=\frac{h_u + i P_u}{\sqrt{2}} + \upsilon_u, \qquad \tilde \nu_i^c = \frac{(\tilde \nu_i^c)^\Re + i (\tilde \nu_i^c)^\Im}{\sqrt{2}} + \upsilon_{\nu_i^c}.
\end{eqnarray}
And one can define
\begin{eqnarray}
\tan\beta=\frac{\upsilon_u}{\sqrt{\upsilon_d^2+\upsilon_{\nu_i}\upsilon_{\nu_i}}}.
\end{eqnarray}
For simplicity, we will assume that all parameters in the potential are real in the model. The CP-odd neutral scalar mass matrix $M_P^2$ and charged scalar mass matrix $M_{S^{\pm}}^2$ can respectively isolate massless unphysical Goldstone bosons $G^0$ and $G^{\pm}$, which can be written as~\cite{ref-zhang,ref-zhang1,ref-zhang2}
\begin{eqnarray}
&&G^0 = {1 \over \sqrt{\upsilon_d^2+\upsilon_u^2+\upsilon_{\nu_i} \upsilon_{\nu_i}}} \Big(\upsilon_d {P_d}-\upsilon_u{P_u}+\upsilon_{\nu_i}{(\tilde \nu_i)^\Im}\Big),\\
&&G^{\pm} = {1 \over \sqrt{\upsilon_d^2+\upsilon_u^2+\upsilon_{\nu_i} \upsilon_{\nu_i}}} \Big(\upsilon_d H_d^{\pm} - \upsilon_u {H_u^{\pm}}+\upsilon_{\nu_i}\tilde e_{L_i}^{\pm}\Big).
\end{eqnarray}
In the physical gauge, the Goldstone bosons $G^0$ and $G^{\pm}$ are eaten by $Z$-boson and $W$-boson, respectively, and disappear from the Lagrangian.
The masses of neutral and charged gauge bosons can be given by
\begin{eqnarray}
&&m_{Z}={e\over {\sqrt{2}s_{_W} c_{_W}}}\sqrt{\upsilon_u^2+\upsilon_d^2+\upsilon_{\nu_i} \upsilon_{\nu_i}}, \\
&&m_{W}={e\over\sqrt{2}s_{_W}}\sqrt{\upsilon_u^2+\upsilon_d^2+\upsilon_{\nu_i} \upsilon_{\nu_i}},
\end{eqnarray}
where $e$ denotes the electromagnetic coupling constant, $s_{_W}=\sin\theta_{_W}$ and $c_{_W}=\cos\theta_{_W}$ with the Weinberg angle $\theta_{_W}$, respectively.
\section{The LFV decays $Z\rightarrow l_i^\pm l_j^\mp$\label{sec3}}
\begin{figure}[htbp]
\setlength{\unitlength}{1mm}
\centering
\includegraphics[width=2.8in]{fig-1a.eps}%
\includegraphics[width=2.8in]{fig-1b.eps}
\caption[]{Feynman diagrams for $Z\rightarrow l_i^\pm l_j^\mp$ in the $\mu\nu$SSM. (a) represents the contributions from neutral fermions $\chi_\eta^0$ and charged scalars $S_{\alpha,\rho}^-$ loops, while (b) represents the contributions from charged fermions $\chi_{\beta,\zeta}$ and neutral scalars $N_\alpha$ ($N=S,P$) loops.}
\label{fig1}
\end{figure}
The Feynman diagrams for $Z\rightarrow l_i^\pm l_j^\mp$ in the $\mu\nu$SSM are depicted by Fig.~\ref{fig1}. And the corresponding effective amplitude for $Z\rightarrow l_i^\pm l_j^\mp$ can be written as~\cite{Zll-R9}
\begin{eqnarray}
\mathcal{M}_{\mu}= e{\bar l_i}{\gamma _\mu }({F_L^{ij}}{P_L} + {F_R^{ij}}{P_R}){l_j},
\end{eqnarray}
with
\begin{eqnarray}
{F_{L,R}^{ij}} = F_{L,R}^{(n)ij} + F_{L,R}^{(c)ij},
\end{eqnarray}
where $F_{L,R}^{(n)ij}$ denote the contributions from the virtual neutral fermion loops, and $F_{L,R}^{(c)ij}$ stand for the contributions from the virtual charged fermion loops, respectively. After integrating the heavy freedoms out, we formulate those coefficients as follows:
\begin{eqnarray}
&&F_L^{(n)ij} = \sum\limits_{N=S,P} \Big[ \frac{{m_{{\chi _\zeta }}}{m_{{\chi _\beta }}}}
{{e}{m_W^2}}C_R^{{N_\alpha }{\chi _\zeta }{{\bar \chi }_{2+i}}}C_L^{Z{\chi _\beta }{{\bar \chi }_\zeta }}
C_L^{{N_\alpha }{\chi _{2+j}}{{\bar \chi }_\beta }}{G_1}({x_{{N_\alpha }}},{x_{{\chi _\zeta }}},{x_{{\chi _\beta }}})\nonumber\\
&&\qquad\qquad\qquad\quad - \: \frac{1}{2{e}} C_R^{{N_\alpha }{\chi _\zeta }{{\bar \chi }_{2+i}}}C_R^{Z{\chi _\beta }
{{\bar \chi }_\zeta }}C_L^{{N_\alpha }{\chi _{2+j}}{{\bar \chi }_\beta }}{G_2}({x_{{N_\alpha }}},
{x_{{\chi _\zeta }}},{x_{{\chi _\beta }}}) \Big],\\
&&F_L^{(c)ij} = \, \frac{1}{2{e}}C_R^{S_\rho ^ - \chi _\eta ^0{{\bar \chi }_{2+i}}}
C_R^{ZS_\alpha ^ - S_\rho ^ {-\ast} }C_L^{S_\alpha ^{-\ast} {\chi _{2+j}}\bar \chi _\eta ^0}{G_2}
({x_{\chi _\eta ^0}},{x_{S_\alpha ^ - }},{x_{S_\rho ^ - }}) ,\\
&&F_R^{(n,c)ij} = \left. {F_L^{(n,c)ij}} \right|{ _{L \leftrightarrow R}} .
\end{eqnarray}
Here, the concrete expressions for coupling coefficients $C_{L,R}$ can be found in Appendix~\ref{app-coupling}, $x= {m^2}/{m_W^2}$ and $m$ is the mass for the corresponding particle. And the form factors $G_{i} $ are given by
\begin{eqnarray}
&&{G_1}(\textit{x}_1 , x_2 , x_3) = \frac{1}{{16{\pi ^2}}}\Big[ \frac{{{x_1}\ln {x_1}}}{{({x_1} - {x_2})({x_1} - {x_3})}} + \frac{{{x_2}\ln {x_2}}}{{({x_2} - {x_1})({x_2} - {x_3})}} \nonumber\\
&&\qquad\qquad\qquad\qquad\quad\;\; + \: \frac{{{x_3}\ln {x_3}}}{{({x_3} - {x_1})({x_3} - {x_2})}}\Big], \\
&&{G_2}(\textit{x}_1 , x_2 , x_3) = \frac{1}{{16{\pi ^2}}}\Big[ \frac{{x_1^2\ln {x_1}}}{{({x_1} - {x_2})({x_1} - {x_3})}} + \frac{{x_2^2\ln {x_2}}}{{({x_2} - {x_1})({x_2} - {x_3})}}\nonumber\\
&&\qquad\qquad\qquad\qquad\quad\;\; + \: \frac{{x_3^2\ln {x_3}}}{{({x_3} - {x_1})({x_3} - {x_2})}} \Big].
\end{eqnarray}
Then, we can obtain the branching ratio of $Z\rightarrow l_i^\pm l_j^\mp$
\begin{eqnarray}
{\rm{Br}}(Z\rightarrow l_i^\pm l_j^\mp)= \frac{e^2}{12\pi}\frac{m_Z}{{{\Gamma}}_Z}\Big({\left| {F_L^{ij}} \right|^2} + {\left| {F_R^{ij}} \right|^2}\Big),
\end{eqnarray}
where ${{\Gamma}}_Z$ denotes the total decay width of $Z$-boson. In the numerical calculation, we choose ${{\Gamma}}_Z \simeq 2.4952\:{\rm{GeV}}$~\cite{PDG}.
\section{Numerical analysis\label{sec4}}
In order to obtain a more transparent numerical results, we take the minimal flavor violation (MFV) assumptions for some parameters in the $\mu\nu{\rm SSM}$, which assume
\begin{eqnarray}
&&\;\,{\kappa _{ijk}} = \kappa {\delta _{ij}}{\delta _{jk}}, \;\;
{({A_\kappa }\kappa )_{ijk}} =
{A_\kappa }\kappa {\delta _{ij}}{\delta _{jk}}, \quad\;
\lambda _i = \lambda , \qquad\;\;\;\;
{{\rm{(}}{A_\lambda }\lambda {\rm{)}}_i} = {A_\lambda }\lambda,\nonumber\\
&&\;\,{Y_{{u _{ij}}}} = {Y_{{u _i}}}{\delta _{ij}},\quad
(A_u Y_u)_{ij}={A_{u_i}}{Y_{{u_i}}}{\delta _{ij}},\;\;\:
{Y_{{\nu _{ij}}}} = {Y_{{\nu _i}}}{\delta _{ij}},\;\;\,
(A_\nu Y_\nu)_{ij}={a_{{\nu_i}}}{\delta _{ij}},\nonumber\\
&&\;\,{Y_{{d_{ij}}}} = {Y_{{d_i}}}{\delta _{ij}},\quad\:
(A_d Y_d)_{ij}={A_{d_i}}{Y_{{d_i}}}{\delta _{ij}},\quad
{Y_{{e_{ij}}}} = {Y_{{e_i}}}{\delta _{ij}},\qquad\;\;\;\,
\upsilon_{\nu_i^c}=\upsilon_{\nu^c},\nonumber\\
&&m_{\tilde Q_{ij}}^2 = m_{{{\tilde Q_i}}}^2{\delta _{ij}}, \qquad\:\,
m_{\tilde u_{ij}^c}^2 = m_{{{\tilde u_i}^c}}^2{\delta _{ij}}, \quad\;\:\:
m_{\tilde d_{ij}^c}^2 = m_{{{\tilde d_i}^c}}^2{\delta _{ij}}, \quad\;\;\;\;
m_{\tilde \nu_{ij}^c}^2 = m_{{{\tilde \nu_i}^c}}^2{\delta _{ij}},
\label{MFV}
\end{eqnarray}
where $i,\;j,\;k =1,\;2,\;3 $. Restrained by the quark and lepton masses, we could have
\begin{eqnarray}
{Y_{{u_i}}} = \frac{{{m_{{u_i}}}}}{{{\upsilon_u}}},\qquad
{Y_{{d_i}}} = \frac{{{m_{{d_i}}}}}{{{\upsilon_d}}},\qquad
{Y_{{e_i}}} = \frac{{{m_{{l_i}}}}}{{{\upsilon_d}}},
\end{eqnarray}
where $m_{u_i}$, $m_{d_i}$ and $m_{l_i}$ are the up-quark, down-quark and charged lepton masses, respectively, and we choose the values from Ref.~\cite{PDG}.
For soft breaking slepton mass matrices $m_{{{\tilde L},{\tilde e^c}}}^2$ and trilinear coupling matrix $({A_e}{Y_e})$, we will introduce the slepton flavor mixings, which take into account the off-diagonal terms for the matrices and are defined as~\cite{sl-mix,sl-mix1,sl-mix2,sl-mix3,sl-mix4,neu-zhang2}
\begin{eqnarray}
&&\quad\;\,{m_{\tilde L}^2} = \left( {\begin{array}{*{20}{c}}
1 & \delta_{12}^{LL} & \delta_{13}^{LL} \\
\delta_{12}^{LL} & 1 & \delta_{23}^{LL} \\
\delta_{13}^{LL} & \delta_{23}^{LL} & 1 \\
\end{array}} \right){m_{L}^2},\\
&&\quad\:{m_{\tilde e^c}^2} = \left( {\begin{array}{*{20}{c}}
1 & \delta_{12}^{RR} & \delta_{13}^{RR} \\
\delta_{12}^{RR} & 1 & \delta_{23}^{RR} \\
\delta_{13}^{RR} & \delta_{23}^{RR} & 1 \\
\end{array}} \right){m_{E}^2},\\
&&({A_e}{Y_e}) = \left( {\begin{array}{*{20}{c}}
m_{l_1}{A_e} & \delta_{12}^{LR}{m_{L}}{m_{E}} & \delta_{13}^{LR}{m_{L}}{m_{E}} \\
\delta_{12}^{LR}{m_{L}}{m_{E}} & m_{l_2}{A_e} & \delta_{23}^{LR}{m_{L}}{m_{E}} \\
\delta_{13}^{LR}{m_{L}}{m_{E}} & \delta_{23}^{LR}{m_{L}}{m_{E}} & m_{l_3}{A_e} \\
\end{array}} \right){1\over {\upsilon_d}}.
\end{eqnarray}
For simplicity, we will choose the slepton mixing parameters
\begin{eqnarray}
&&\delta_{12}^{LL}=\delta_{12}^{RR}=\delta_{12}^{LR}\equiv \delta_{12}^X, \nonumber\\ &&\delta_{13}^{LL}=\delta_{13}^{RR}=\delta_{13}^{LR}\equiv \delta_{13}^X, \nonumber\\ &&\delta_{23}^{LL}=\delta_{23}^{RR}=\delta_{23}^{LR}\equiv \delta_{23}^X.
\end{eqnarray}
At the EW scale, the soft masses $m_{\tilde H_d}^2$, $m_{\tilde H_u}^2$ and $m_{\tilde \nu_i^c}^2$ can be derived from the minimization conditions of the tree-level neutral scalar potential, which are given in Refs.~\cite{mnSSM1,ref-zhang}. Ignoring the terms of the second order in $Y_{\nu}$ and assuming $(\upsilon_{\nu_i}^2+\upsilon_d^2-\upsilon_u^2)\approx (\upsilon_d^2-\upsilon_u^2)$, one can have the minimization conditions of the tree-level neutral scalar potential with respect to $\upsilon_{\nu_i}\:(i=1,2,3)$ below~\cite{neu-zhang2}
\begin{eqnarray}
m_{\tilde L_{ij}}^2 \upsilon_{\nu_j}+{G^2\over 4} (\upsilon_d^2-\upsilon_u^2)\upsilon_{\nu_i}=\Big[\lambda \upsilon_d (\upsilon_u^2+\upsilon_{\nu^c}^2) - \kappa \upsilon_u \upsilon_{\nu^c}^2\Big] Y_{\nu_i} -\upsilon_u \upsilon_{\nu^c}a_{\nu_i},
\label{eq-min}
\end{eqnarray}
where $G^2=g_1^2+g_2^2$ and $g_1 c_{_W} =g_2 s_{_W}=e$. Solving Eq.~(\ref{eq-min}), we can gain the left-handed sneutrino VEVs
\begin{eqnarray}
\upsilon_{\nu_i}=\frac{{\rm{det}}\: T_i}{{\rm{det}}\: T},\qquad (i=1,2,3),
\label{eq-vi}
\end{eqnarray}
where
\begin{eqnarray}
T = \left( {\begin{array}{*{20}{c}}
m_{\tilde L_{11}}^2 +{G^2\over 4}(\upsilon_d^2-\upsilon_u^2) & m_{\tilde L_{12}}^2 & m_{\tilde L_{13}}^2 \\ [6pt]
m_{\tilde L_{21}}^2 & m_{\tilde L_{22}}^2 +{G^2\over 4}(\upsilon_d^2-\upsilon_u^2) & m_{\tilde L_{23}}^2 \\ [6pt]
m_{\tilde L_{31}}^2 & m_{\tilde L_{32}}^2 & m_{\tilde L_{33}}^2 +{G^2\over 4}(\upsilon_d^2-\upsilon_u^2) \\ [6pt]
\end{array}} \right),
\end{eqnarray}
and $T_i$ can be acquired from $T$ by replacing the $i$-th column with
\begin{eqnarray}
\left( {\begin{array}{*{20}{c}}
\Big[\lambda \upsilon_d (\upsilon_u^2+\upsilon_{\nu^c}^2) - \kappa \upsilon_u \upsilon_{\nu^c}^2\Big] Y_{\nu_1} -\upsilon_u \upsilon_{\nu^c}a_{\nu_1} \\ [6pt]
\Big[\lambda \upsilon_d (\upsilon_u^2+\upsilon_{\nu^c}^2) - \kappa \upsilon_u \upsilon_{\nu^c}^2\Big] Y_{\nu_2} -\upsilon_u \upsilon_{\nu^c}a_{\nu_2} \\ [6pt]
\Big[\lambda \upsilon_d (\upsilon_u^2+\upsilon_{\nu^c}^2) - \kappa \upsilon_u \upsilon_{\nu^c}^2\Big] Y_{\nu_3} -\upsilon_u \upsilon_{\nu^c}a_{\nu_3} \\ [6pt]
\end{array}} \right).
\end{eqnarray}
In the $\mu\nu{\rm SSM}$, the sneutrino sector may appear the tachyons, which masses squared are negative. So, we need analyse the masses of the sneutrinos. The masses squared of left-handed sneutrinos are basically determined by soft breaking slepton mass matrix $m_{\tilde L}^2$. And the CP-even and CP-odd right-handed sneutrino masses squared can be approximately written as
\begin{eqnarray}
&&m_{S_{5+i}}^2\approx (A_\kappa+4\kappa\upsilon_{\nu^c})\kappa\upsilon_{\nu^c} +A_\lambda \lambda \upsilon_d \upsilon_u/\upsilon_{\nu^c}-2\lambda^2(\upsilon_d^2+\upsilon_u^2),\\
&&m_{P_{5+i}}^2\approx -3A_\kappa \kappa\upsilon_{\nu^c} +(A_\lambda/\upsilon_{\nu^c}+4\kappa)\lambda \upsilon_d \upsilon_u-2\lambda^2(\upsilon_d^2+\upsilon_u^2).
\end{eqnarray}
Here, the main contribution for the mass squared is the first term as $\kappa$ is large, in the limit of $\upsilon_{\nu^c} \gg \upsilon_{u,d}$. Therefore, we could use the approximate relation
\begin{eqnarray}
-4\kappa\upsilon_{\nu^c}\lesssim A_\kappa \lesssim 0,
\label{tachyon}
\end{eqnarray}
to avoid the tachyons.
Before calculation, the constraints on the parameters of the $\mu\nu{\rm SSM}$ from neutrino experiments should be considered at first. Three flavor neutrinos $\nu_{e,\mu,\tau}$ could mix into three massive neutrinos $\nu_{1,2,3}$ during their flight, and the mixings are described by the Pontecorvo-Maki-Nakagawa-Sakata unitary matrix $U_{_{PMNS}}$ \cite{ref-PMNS1,ref-PMNS2}. The experimental observations of the parameters in $U_{_{PMNS}}$ for the normal mass hierarchy show that \cite{neu-data}
\begin{eqnarray}
&&\sin^2\theta_{12} =0.302_{-0.012}^{+0.013},\qquad \Delta m_{21}^2 =7.50_{-0.19}^{+0.18}\times 10^{-5} {\rm eV}^2, \nonumber\\
&&\sin^2\theta_{23}=0.413_{-0.025}^{+0.037},\qquad \Delta m_{31}^2 =2.473_{-0.067}^{+0.070}\times 10^{-3} {\rm eV}^2, \nonumber\\
&&\sin^2 \theta_{13} =0.0227_{-0.0024}^{+0.0023}.
\label{neutrino-oscillation}
\end{eqnarray}
In the $\mu\nu{\rm SSM}$, the three tiny neutrino masses are obtained through TeV scale seesaw mechanism~\cite{mnSSM,mnSSM1,mnSSM2,meu-m,meu-m1,meu-m2,meu-m3,neu-zhang1}. Assumed that the charged lepton mass matrix in the flavor basis is in the diagonal form, we parameterize the unitary matrix which diagonalizes the effective light neutrino mass matrix $m_{eff}$ (see Ref.~\cite{ref-zhang}) as \cite{ref-Uv1,ref-Uv2}
\begin{eqnarray}
{U_\nu} = &&\left( {\begin{array}{*{20}{c}}
{{c_{12}}{c_{13}}} & {{s_{12}}{c_{13}}} & {{s_{13}}{e^{ - i\delta }}} \\
{ - {s_{12}}{c_{23}} - {c_{12}}{s_{23}}{s_{13}}{e^{i\delta }}} & {{c_{12}}{c_{23}} - {s_{12}}
{s_{23}}{s_{13}}{e^{i\delta }}} & {{s_{23}}{c_{13}}} \\
{{s_{12}}{s_{23}} - {c_{12}}{c_{23}}{s_{13}}{e^{i\delta }}} & { - {c_{12}}{s_{23}} - {s_{12}}
{c_{23}}{s_{13}}{e^{i\delta }}} & {{c_{23}}{c_{13}}} \\
\end{array}} \right) \nonumber\\
&&\times \: diag(1,{e^{i\frac{{{\alpha _{21}}}}{2}}},{e^{i\frac{{{\alpha _{31}}}}{2}}})\:,
\label{PMNS-matrix}
\end{eqnarray}
where ${c_{_{ij}}} = \cos {\theta _{ij}}$, ${s_{_{ij}}} = \sin {\theta _{ij}}$. In our calculation, the values of $\theta_{ij}$ are obtained from the experimental data in Eq.~(\ref{neutrino-oscillation}), and all CP violating phases $\delta$, $\alpha _{21}$ and $\alpha _{31}$ are set to zero. $U_\nu$ diagonalizes $m_{eff}$ in the following way:
\begin{eqnarray}
U_\nu ^T m_{eff}^T{m_{eff}}{U_\nu} = diag({m_{\nu _1}^2},{m_{\nu _2}^2},{m_{\nu _3}^2}).
\label{eff}
\end{eqnarray}
For the neutrino mass spectrum, we assume it to be normal hierarchical, i.e., ${m_{\nu_1}}{\rm{ < }}{m_{\nu_2}}{\rm{ < }}{m_{\nu_3}}$, and we choose the neutrino mass $m_{\nu_1}=10^{-2}\:{\rm{eV}}$ as input in our numerical analysis, limited on neutrino masses from neutrinoless double-$\beta$ decay~\cite{neu-m-limit} and cosmology~\cite{neu-m-limit1}. The other two neutrino masses $m_{\nu_{2,3}}$ can be obtained through the experimental data on the differences of neutrino mass squared in Eq.~(\ref{neutrino-oscillation}). Then, we can numerically derive $Y_{\nu_i} \sim \mathcal{O}(10^{-7})$ and $a_{\nu_i} \sim \mathcal{O}(-10^{-4}{\rm{GeV}})$ from Eq.~(\ref{eff}). Accordingly, $\upsilon_{\nu_i} \sim \mathcal{O}(10^{-4}{\rm{GeV}})$ through Eq.~(\ref{eq-min}). Due to $\upsilon_{\nu_i}\ll\upsilon_{u,d}$, we can have
\begin{eqnarray}
\tan\beta\simeq \frac{\upsilon_u}{\upsilon_d}.
\end{eqnarray}
Recently, a neutral Higgs with mass around $125\;{\rm GeV}$ reported by ATLAS~\cite{ATLAS} and CMS~\cite{CMS} also contributes a strict constraint on relevant parameter space of the model. The global fit to the ATLAS and CMS Higgs data gives~\cite{mh-AC}:
\begin{eqnarray}
m_{{h}}=125.66\pm0.34\;{\rm GeV}.
\label{M-h}
\end{eqnarray}
In the $\mu\nu$SSM, the loop effects of right-handed neutrino/sneutrino on the SM-like Higgs mass can be neglected, due to small neutrino Yukawa couplings $Y_{\nu_i} \sim \mathcal{O}(10^{-7})$ and left-handed neutrino superfield VEVs $\upsilon_{\nu_i} \sim \mathcal{O}(10^{-4}{\rm{GeV}})$. Through the numerical computation in Ref.~\cite{ref-zhang}, we also can numerically ignore the radiative corrections from $b$ quark, $\tau$ lepton and their supersymmetric partners on the SM-like Higgs mass. Then, the main radiative corrections on the SM-like Higgs mass in the $\mu\nu$SSM come from the top quark and its supersymmetric partners, similarly to the MSSM. However when $\tan\beta$ is large enough, we also need to consider the radiative corrections from $b$ quark and its supersymmetric partners. Due to the introduction of some new couplings in the superpotential, the SM-like Higgs mass in the $\mu\nu$SSM gets additional contribution at tree-level~\cite{mnSSM1}. Therefore, the SM-like Higgs in the $\mu\nu{\rm SSM}$ can easily account for the mass around $125\,{\rm GeV}$, especially for small $\tan\beta$. For moderate $\tan\beta$ and large mass of the pseudoscalar $M_A$, the SM-like Higgs mass in the $\mu\nu{\rm SSM}$ is approximately given by
\begin{eqnarray}
m_h^2 \simeq m_Z^2 \cos^2 2\beta + \frac{6 \lambda^2 s_{_W}^2 c_{_W}^2}{ e^2} m_Z^2 \sin^2 2\beta+\bigtriangleup m_h^2,
\label{eq-mh}
\end{eqnarray}
with the main radiative corrections~\cite{ref-mh-rad,ref-mh-rad1}
\begin{eqnarray}
\bigtriangleup m_h^2 = \frac{3m_t^4}{4\pi^2\upsilon^2}\ln \frac{M_S^2}{m_t^2} + \frac{3m_t^4}{4\pi^2\upsilon^2} \frac{X_t^2}{M_S^2} (1-\frac{X_t^2}{12M_S^2}),
\end{eqnarray}
where $\upsilon=174$ GeV, $M_S =\sqrt{m_{{\tilde t}_1}m_{{\tilde t}_2}}$ with $m_{{\tilde t}_{1,2}}$ being the stop masses, $X_t \equiv A_t-\mu\cot\beta$ with $A_t=A_{u_3}$ denoting the trilinear Higgs-stop coupling and $\mu=3\lambda \upsilon_{\nu^c}$ being the Higgsino mass parameter.
Through the analysis of the parameter space in Ref.~\cite{mnSSM1}, we could choose the reasonable values for some parameters as $\kappa=0.4$, $\lambda=0.1$, $A_\lambda=500\;{\rm GeV}$, $\upsilon_{\nu^c}=1\;{\rm TeV}$ and $m_L=m_E={A_{e}}=1\;{\rm TeV}$ for simplicity in the following numerical calculation. Through Eq.~(\ref{tachyon}), we could choose ${A_{\kappa}}=-300\;{\rm GeV}$ to avoid the tachyons. For the Majorana masses of the gauginos, we will imply the approximate GUT relation $M_1=\frac{\alpha_1^2}{\alpha_2^2}M_2\approx 0.5 M_2$ and $M_3=\frac{\alpha_3^2}{\alpha_2^2}M_2\approx 2.7 M_2$. The gluino mass, $m_{{\tilde g}}\approx M_3$, is larger than about $1.2$ TeV from the ATLAS and CMS experimental data~\cite{ATLAS-sg1,ATLAS-sg2,CMS-sg1,CMS-sg2}. So, we conservatively choose $M_2=1\;{\rm TeV}$. And the first two generations of squarks are strongly constrained by direct searches at the LHC~\cite{ATLAS-sq1,CMS-sq1}. Therefore, we take $m_{{\tilde Q}_{1,2}}=m_{{\tilde u}^c_{1,2}}=m_{{\tilde d}^c_{1,2}}=2\;{\rm TeV}$. The third generation squark masses are not constrained by the LHC as strongly as the first two generations, and affect the SM-like Higgs mass. So, we could adopt $m_{{\tilde Q}_3}=m_{{\tilde u}^c_3}=m_{{\tilde d}^c_3}=1\;{\rm TeV}$. For simplicity, we take $A_{d_{1,2,3}}=A_{u_{1,2}}=1\;{\rm TeV}$. Then, through Eq.~(\ref{eq-mh}), we can choose $\tan \beta=3$ and $A_{u_3}=1.13\;{\rm TeV}$, to keep the SM-like Higgs mass $m_h\simeq 125.7\,{\rm GeV}$.
\begin{figure}
\setlength{\unitlength}{1mm}
\centering
\includegraphics[width=3.1in]{fig-2a.eps}%
\includegraphics[width=3.1in]{fig-2b.eps}
\includegraphics[width=3.1in]{fig-2c.eps}%
\includegraphics[width=3.1in]{fig-2d.eps}
\caption[]{LFV rates for $\mu-e$ transitions versus slepton mixing parameter $\delta_{12}^X$, where the dashed lines denote the present limits and the dotted lines denote the future sensitivities.}
\label{fig2}
\end{figure}
It's well known that the LFV processes are flavor dependent. The LFV rates for $l_j-l_i$ transitions depend respectively on the slepton mixing parameters $\delta_{ij}^X$, which can be numerically confirmed by Fig.~\ref{fig2} and Fig.~\ref{fig3}. In Fig.~\ref{fig2}, we plot the LFV rates for $\mu-e$ transitions versus slepton mixing parameter $\delta_{12}^X$, where the dashed lines denote the present limits and the dotted lines denote the future sensitivities. Considered that the LFV rates for $\mu-e$ transitions don't depend on $\delta_{13}^X$ and $\delta_{23}^X$, we have chosen $\delta_{13}^X=\delta_{23}^X=0$. The numerical results in Fig.~\ref{fig2} show that the LFV rates for $\mu-e$ transitions are increasing, along with increasing of slepton mixing parameter $\delta_{12}^X$. The branching ratio of $\mu\rightarrow e\gamma$ easily reach the present experimental bound, and constrains $\delta_{12}^X \lesssim \mathcal{O}(10^{-6})$. Under the constraint of the present experimental limit for ${\rm{Br}}(\mu\rightarrow e\gamma)$, the expected rate for $Z\rightarrow e\mu$ still remains orders of magnitude below the future experimental sensitivity, and the expected rates for $\mu\rightarrow 3e$ and $\mu-e$ conversion in nuclei don't reach the present experimental bounds. However, the high future experimental sensitivities still keep a hope to detect $\mu\rightarrow 3e$ and $\mu-e$ conversion in nuclei.
In addition, the dominance of the $\gamma$-mediated channel in the decays $l_j^- \rightarrow l_i^- l_i^- l_i^+$ allows us to derive the simplified relation~\cite{Hisano}
\begin{eqnarray}
\frac{{\rm{Br}}(l_j^- \rightarrow l_i^- l_i^- l_i^+)}{{\rm{Br}}(l_j^- \rightarrow l_i^- \gamma)} \simeq \frac{e^2}{32\pi^2}(\frac{{16}}{3}\ln \frac{{{m_{{l_j}}}}}{{2{m_{{l_i}}}}} - \frac{{14}}{9}),
\end{eqnarray}
which is in agree with the numerical result.
\begin{figure}
\setlength{\unitlength}{1mm}
\centering
\includegraphics[width=3.1in]{fig-3a.eps}%
\includegraphics[width=3.1in]{fig-3d.eps}
\includegraphics[width=3.1in]{fig-3b.eps}%
\includegraphics[width=3.1in]{fig-3e.eps}
\includegraphics[width=3.1in]{fig-3c.eps}%
\includegraphics[width=3.1in]{fig-3f.eps}
\caption[]{(a--c) LFV rates for $\tau-e$ transitions versus slepton mixing parameter $\delta_{13}^X$, and (d--f) LFV rates for $\tau-\mu$ transitions versus slepton mixing parameter $\delta_{23}^X$, where the dashed lines denote the present limits and the dotted lines denote the future sensitivities.}
\label{fig3}
\end{figure}
In Fig.~\ref{fig3}(a--c), we picture the LFV rates for $\tau-e$ transitions versus slepton mixing parameter $\delta_{13}^X$, for $\delta_{12}^X=\delta_{23}^X=0$. And we plot the LFV rates for $\tau-\mu$ transitions versus slepton mixing parameter $\delta_{23}^X$ in Fig.~\ref{fig3}(d--f), as $\delta_{12}^X=\delta_{13}^X=0$. The numerical results still show that the LFV rates are increasing, along with increasing of slepton mixing parameters. The present experimental bounds of ${\rm{Br}}(\tau\rightarrow e\gamma)$ and ${\rm{Br}}(\tau\rightarrow \mu\gamma)$ respectively constrain $\delta_{13}^X \lesssim \mathcal{O}(10^{-2})$ and $\delta_{23}^X \lesssim \mathcal{O}(10^{-2})$. Under the constraints of the present experimental limits for ${\rm{Br}}(\tau\rightarrow e\gamma)$ and ${\rm{Br}}(\tau\rightarrow \mu\gamma)$, the expected rates for $Z\rightarrow e\tau$ and $Z\rightarrow \mu\tau$ still remain orders of magnitude below the future experimental sensitivities, and the expected rates for $\tau\rightarrow 3e$ and $\tau\rightarrow 3\mu$ don't reach the present experimental bounds. However, the high future experimental sensitivities still keep a hope to detect $\tau\rightarrow 3e$ and $\tau\rightarrow 3\mu$.
\section{Summary\label{sec5}}
In this paper, we study the LFV processes $Z\rightarrow l_i^\pm l_j^\mp$, $l_j^-\rightarrow l_i^-\gamma$, $l_j^- \rightarrow l_i^- l_i^- l_i^+$ and muon conversion to electron in nuclei with slepton flavor mixing, within framwork of the $\mu\nu$SSM. The numerical results show that the LFV rates for $l_j-l_i$ transitions depend respectively on the slepton mixing parameters $\delta_{ij}^X$, because the LFV processes are flavor dependent. In the $\mu\nu$SSM, the branching ratio of $l_j^-\rightarrow l_i^-\gamma$ can easily reach the present experimental bounds. So, it's a high hope to detect $l_j^-\rightarrow l_i^-\gamma$ in the future. And the present experimental limits of ${\rm{Br}}(l_j^-\rightarrow l_i^-\gamma)$ constrain $\delta_{12}^X \lesssim \mathcal{O}(10^{-6})$, $\delta_{13}^X \lesssim \mathcal{O}(10^{-2})$ and $\delta_{23}^X \lesssim \mathcal{O}(10^{-2})$. Under the constraints of the present experimental limits for ${\rm{Br}}(l_j^-\rightarrow l_i^-\gamma)$, the expected rates for $Z\rightarrow l_i^\pm l_j^\mp$ still remain orders of magnitude below the future experimental sensitivities, and the expected rates for $l_j^- \rightarrow l_i^- l_i^- l_i^+$ and $\mu-e$ conversion in nuclei don't reach the present experimental bounds. However, the high future experimental sensitivities still keep a hope to detect $l_j^- \rightarrow l_i^- l_i^- l_i^+$ and $\mu-e$ conversion in nuclei within the $\mu\nu$SSM.
\begin{acknowledgments}
The work has been supported by the National Natural Science Foundation of China (NNSFC)
with Grant No. 11275036, No. 11047002, the open project of State
Key Laboratory of Mathematics-Mechanization with Grant No. Y3KF311CJ1, the Natural
Science Foundation of Hebei province with Grant No. A2013201277, and Natural Science Fund of Hebei University with Grant No. 2011JQ05, No. 2012-242.
\end{acknowledgments}
|
\section{Introduction\label{SEC:INTRO}}
The utility of calibrating the power-law correlation between galaxy
rotation rates and their luminosities \citep{Tully:77:661} in the
mid-infrared (MIR) has been clearly demonstrated by \citet{Sorce:13:94}.
Their use of the 3.6 $\mu$m {\it SPITZER} Infrared Array Camera (IRAC)
photometry provided a calibration of the Tully-Fisher relation (TFR) with a
scatter comparable to that seen in the I-band \citep{Tully:12:78}. Having
space-based photometry in the MIR mitigates the effects of dust and removes
possible systematics when attempting to constrain the motions of galaxies
across the entire sky.
While the IRAC 3.6 $\mu$m calibration is useful, there are a limited number
of galaxies that have been observed through the camera's 4-arcminute
field-of-view. The W1 band of the {\it Wide-field Infrared Survey
Explorer} \citep[{\it WISE},][]{Wright:10:1868} is similar in wavelength
coverage ($\lambda_{eff} = 3.4\ \mu$m), is also space-based and thus enjoys
all the benefits of the IRAC calibration. In addition, the {\it WISE}
mission has covered the entire sky to a depth similar to the IRAC coverage
of selected nearby galaxies. The number of calibrator galaxies with {\it
WISE} imaging is 310, an increase of 46\% over the sample available to
\citet{Sorce:13:94}, although the number of calibrator galaxies observed in
the IRAC [3.6] band continues to increase. This additional utility of
all-sky coverage motivated this work.
The {\it WISE} imaging represents the opportunity for providing
high-quality 3.4 $\mu$m (W1) and 4.6 $\mu$m (W2) photometry over the entire
sky, however, the automated catalog photometry available from the mission
has not been optimized for extended galaxies. Corrections can be made to
the catalog photometry, however when applied to the TFR the resulting
scatter is significantly larger \citep[0.69 mag,][]{Lagattuta:13:88} than
for the I-band calibration \citep[0.41 mag,][]{Tully:12:78}. We have instigated a separate project to provide
high-quality surface photometry of all {\it WISE} galaxies larger than 0.8
arcminutes on the sky. The {\it WISE} Nearby Galaxy Atlas
\citep[WNGA,][]{Seibert:14b} will provide photometry that is quality
controlled for over 20,000 galaxies. This photometry, optimized for
extended sources, significantly reduces the resulting scatter in the TFR
calibration, and thus improves the resulting distances. Having an accurate
calibration of the TFR for these two WISE passbands will allow the use of
this large sample to explore the structure and dynamics of local galaxy
bulk flows. This calibration has been completed and is presented herein.
The focus of this paper is the calibration of the TFR using photometry in
the {\it WISE} W1 and W2 bands, however, we take the occasion to update the
I-band calibration and present it in \S\ref{SEC:ICALIBRATION}. We
introduce a significant number of new calibration candidates by considering
all galaxies associated with the calibrating clusters (see
\S\ref{SEC:CALIBRATORS}) contained in the 2MASS redshift survey complete to
K=11.75 \citep{Huchra:12:26}.
\citet{Sorce:13:94} found a reduction in the scatter of the TFR when
applying a correction to the IRAC photometry based on the optical-MIR color
with the optical measure being provided by I-band photometry.
Unfortunately, there is no space-based all-sky survey in the I-band. Thus,
the uncorrected calibration for W1 and W2 may be useful for those wishing
to extend their catalogs to as many galaxies as possible (those without
I-band photometry), even though the scatter will be slightly larger. While
the TFR in the optical has proven to be a straight power-law, there is
evidence that in the MIR there is curvature in the relation. We
investigate that possibility and present our results in \S\ref{SEC:CURVE}.
Once an accurate calibration is derived, we can use the distances derived
thereby for calculating the Hubble constant, $H_0$ \citep{Courtois:12:174,
Sorce:12:L12}. We do this in \S\ref{SEC:H0_CLUSTERS} using the subset of
clusters from the calibration set that have recession velocities that place
them in the Hubble flow \citep[$> 4000$\,km~s$^{-1}$,][]{Tully:12:78,
Sorce:13:94}. For a more robust measure of $H_0$ that extends well into
the Hubble-flow (to $z>1$) we use TFR distances to re-normalize the
distance scale for the UNION2 SN~Ia sample \citep{Amanullah:10:712} and
calculate $H_0$ directly from the re-normalization in
\S\ref{SEC:H0_SUPERNOVAE}.
\section{Data\label{SEC:DATA}}
\subsection{Calibrators\label{SEC:CALIBRATORS}}
We adopt the galaxy cluster technique for deriving the calibration
described by \citet{Tully:12:78} (see also \S\ref{SEC:RELATIVEDISTANCES}).
We take advantage of the fact that the galaxies within a given cluster are
at the same distance and that the galaxy masses, and hence HI linewidths,
span a range large enough to determine the slope of the correlation for
each cluster. We then shift each cluster along the luminosity axis such
that their data appear to be from a single cluster. We iteratively combine
the galaxy data derived from a set of thirteen nearby clusters to derive a
universal slope, and then set the zero-point of the relation using the
universal slope applied to nearby galaxies with accurate distance
measurements derived from independent techniques. To minimize the effect
of the Malmquist bias, the slopes are derived from fitting the inverse
Tully-Fisher relation \citep[ITFR,][]{Willick:94:1}. Details on the method
and the calibrator and cluster sample can be found in \citet{Tully:12:78}
and \citet{Sorce:13:94}. The appendix in \citet{Tully:12:78} discusses
issues specific to each of the 13 clusters.
In order to avoid excessive noise in the calibrations we apply several cuts
to the input sample \citep{Tully:12:78}. Because we must de-project the HI
linewidths based on the observed inclination, we exclude galaxies more
face-on than $45^{\circ}$, the limit where typical errors in the
de-projections begin to exceed 8\%. Morphological types earlier than Sa
greatly increase the scatter in the relation, most likely due to the mass
of the bulge not contributing to the HI linewidth, and are excluded.
Systems with insufficient or confused HI, and galaxies that appear
disrupted are also excluded. Following \citet{Sorce:13:94}, the
\citet{Tully:12:78} sample for Abell 2634 has been extended to include the
adjacent Abell 2666 which is, within measurement uncertainties, at the same
distance.
\tabletypesize{\tiny}
\setlength{\tabcolsep}{0.01cm}
\begin{deluxetable*}{cccccccccccccccl}
\tablecolumns{16}
\tablecaption{Calibrator Data\\
(full table in online version)\label{tab_data}}
\tablehead{
\colhead{PGC\tablenotemark{1}} &
\colhead{Name\tablenotemark{2}} &
\colhead{$I_T$\tablenotemark{3}} &
\colhead{$I_T^{b,i,k}$\tablenotemark{4}} &
\colhead{$W1_T$\tablenotemark{5}} &
\colhead{$W1_T^{b,i,k,a}$\tablenotemark{6}} &
\colhead{$W2_T$\tablenotemark{7}} &
\colhead{$W2_T^{b,i,k,a}$\tablenotemark{8}} &
\colhead{$C_{I-W1}$\tablenotemark{9}} &
\colhead{$C_{I-W2}$\tablenotemark{10}} &
\colhead{b/a\tablenotemark{11}} &
\colhead{Inc\tablenotemark{12}} &
\colhead{$W_{mx}$\tablenotemark{13}} &
\colhead{$W^i_{mx}$\tablenotemark{14}} &
\colhead{$\log$($W^i_{mx}$)\tablenotemark{15}} &
\colhead{Sam\tablenotemark{16}}
}
\startdata
40095 & NGC4312 & 10.650 $\pm$ 0.152 & 10.230 & 11.252 $\pm$ 0.001 & 11.233 & 11.893 $\pm$ 0.001 & 11.899 & -0.661 & -1.327 & 0.27 & 79 & 217 & 221 & 2.344 $\pm$ 0.036 & Virgo \\
40105 & NGC4313 & 10.537 $\pm$ 0.184 & 9.970 & 11.056 $\pm$ 0.001 & 11.028 & 11.653 $\pm$ 0.001 & 11.657 & -0.716 & -1.345 & 0.22 & 85 & 257 & 258 & 2.412 $\pm$ 0.028 & Virgo \\
40201 & NGC4330 & 11.429 $\pm$ 0.189 & 10.810 & 11.936 $\pm$ 0.001 & 11.902 & 12.503 $\pm$ 0.001 & 12.503 & -0.750 & -1.351 & 0.17 & 90 & 251 & 251 & 2.400 $\pm$ 0.026 & Virgo \\
40507 & NGC4380 & 10.109 $\pm$ 0.104 & 9.820 & 11.038 $\pm$ 0.001 & 11.043 & 11.651 $\pm$ 0.001 & 11.675 & -0.881 & -1.513 & 0.52 & 61 & 265 & 304 & 2.483 $\pm$ 0.042 & Virgo \\
\nodata \\
\enddata
\tablenotetext{1}{Principal Galaxies Catalog Number, $^2$Common name,}
\tablenotetext{3}{I-band mag (Vega), $^4$I-band mag with $A^I_{b,i,k}$ applied,}
\tablenotetext{5}{{\it WISE} W1 mag (AB), $^6$W1 mag with $A^{W1}_{b,i,k,a}$ applied,}
\tablenotetext{7}{{\it WISE} W2 mag (AB), $^8$W2 mag with $A^{W2}_{b,i,k,a}$ applied,}
\tablenotetext{9}{$I_T^{b,i,k} - W1_T^{b,i,k,a}$ color (AB), $^{10}$$I_T^{b,i,k} - W2_T^{b,i,k,a}$ color (AB),}
\tablenotetext{11}{Axial ratio b/a, $^{12}$Inclination in degrees,}
\tablenotetext{13}{Uncorrected linewidth, $^{14}$Inclination-corrected linewidth,}
\tablenotetext{15}{Logarithm of the inclination-corrected HI linewidth,}
\tablenotetext{16}{Sample name}
\end{deluxetable*}
{\it WISE} photometry is available for all targets (see
Table~\ref{tab_data}). The Spitzer photometry \citep{Sorce:13:94} was
acquired from pointed observations with the consequence that a significant
fraction of calibration candidates remained unobserved (although the number
observed continues to increase). Likewise, the I-band photometry, acquired
by pointed observations, remains incomplete. With the current tally, there
are 310 cluster calibrators with {\it WISE} W1 and W2 photometry, compared
with 213 available to \citet{Sorce:13:94} for the Spitzer calibration, and
291 of the 310 {\it WISE} calibrators have I-band photometry, compared with
the 267 available to \citet{Tully:12:78} for the previous I-band
calibration.
A minor update with the current work is the conversion of Galactic
obscuration values to \citet{Schlafly:11:103} from \citet{Schlegel:98:525}.
This change has negligible impact on the {\it WISE} W1 magnitudes (and even
less on W2) and only a $1-2\%$ impact at I-band.
A small number of ambiguous cases are being rejected from the cluster
calibration sample. One case is now not ambiguous: PGC~42081 was flagged
in the earlier calibrations as possibly foreground to the Virgo Cluster.
Recent Hubble Space Telescope observations provide a distance of 9.5 Mpc
from a tip of the red giant branch measurement \citep{Karachentsev:14:4},
confirming that this galaxy is in the foreground. Further in the case of
the Virgo Cluster, the galaxies PGC~41531 and 43601 are considered probable
background galaxies. Their velocities (1626 and 1783 km~s$^{-1}$\ respectively)
and distances are consistent with membership in the structure including the
Virgo W Cluster and M Cloud at roughly twice the Virgo distance.
Similarly, PGC~30498, which was considered as a candidate for the Antlia
Cluster because of proximity on the sky, is now considered an outlying
associate of the more distant Hydra Cluster. The velocity range of the two
clusters overlap. PGC~30498, located between the two clusters, $3^{\circ}$
from Antlia, has a velocity and distance compatible with Hydra. It is
$5^{\circ}$ from Hydra; too removed to be taken into the Hydra sample. See
the appendix of \citet{Tully:12:78} for discussions of the environments of
these clusters.
The same inclination, morphology, and HI quality criteria described above
are applied to our zero-point calibrator sample along with the additional
constraint that each zero-point galaxy have a well-known distance derived
either from Cepheid or TRGB measurements. To set the Cepheid distance
scale, we use the recently updated LMC distance modulus of $18.48\pm0.04$
\citep{Scowcroft:11:76, Scowcroft:12:84, Monson:12:146, Freedman:12:24}.
We use a TRGB calibration that has been demonstrated to be consistent with
the Cepheid scale by \citet{Rizzi:07:815} and \citet{Tully:08:184}.
\subsection{HI Line Widths\label{SEC:HILINEWIDTHS}}
We use HI linewidth measurements from the Cosmic Flows project
\citep{Tully:13:86} that contains over 14,000 galaxies with measurements of
$W_{m50}$, the width at 50\% of the mean flux within the velocity range in
the HI line that encompasses 90\% of the total line flux. These data are
available at the Extragalactic Distance Database (EDD)
website\footnote{\url{http://edd.ifa.hawaii.edu}; catalog ``All Digital
HI''}. This observed parameter is de-projected and corrected to a
measurement of the intrinsic maximum rotation velocity width, $W^i_{mx}$.
This is accomplished using a method that accounts for galaxy inclination,
relativistic broadening and finite spectral resolution as described in
\citet{Courtois:09:1938, Courtois:11:2005} and reviewed in
\citet{Tully:12:78}. The error in $W^i_{mx}$ is derived from the signal at
the 50\% level divided by the noise measured outside the line in regions of
no signal. An error threshold of 20 km~s$^{-1}$\ is applied to remove noisy
measurements. Retained profiles meet a minimum per-channel signal-to-noise
requirement of $S/N \geq 2$ and are also visually inspected to remove
pathological cases.
In subsequent plots, it becomes obvious that the errors in the HI
line-widths dominate the observational errors. Slow rotators exhibit a
higher fractional error because of their small line widths. Lower
inclination systems are also prone to higher errors motivating our
inclination threshold of 45$^{\circ}$, below which a 5$^{\circ}$ error in
inclination results in a $> 8$\% error in linewidth.
\subsection{W1 and W2 Data and Photometry\label{SEC:W1PHOTOMETRY}}
Thanks to the {\it WISE} public data release, available from the NASA/IPAC
infrared science archive
(IRSA)\footnote{\url{http://irsa.ipac.caltech.edu/Missions/wise.html}}, all
of the galaxies in our sample have imaging in the {\it WISE} W1 and W2
bands. Image cutouts combining the level 1b (single) image products were
drizzled using version 3.8.3 of the Image Co-addition with Optional
Resolution Enhancement (ICORE) software \citep{Masci:09:67,Masci:13:02010}.
To minimize background problems, we selected the 1b images with moon angles
greater than 25$^{\circ}$, and with epochs at least 2000 seconds from an
annealing event. We combined the resulting image set on an output scale of
1.0 arcseconds per pixel.
Photometry of the calibrator galaxies was performed using the photometry
routines developed for the WNGA \citep{Seibert:14b}. This method uses
elliptical apertures with fixed shapes, orientations, and centers but
varying major axes in steps approximately equal to a resolution element in
the W1 band (6 arcseconds) to measure the flux of the galaxy within each
annulus from the center to the edge of the galaxy. Foreground stars and
contaminating neighbor galaxies are masked prior to measurement and this
masking is accounted for in computing the flux within each annulus. The
influence of partially resolved and unresolved background galaxies is
mitigated by allowing our sky value to contain flux from these objects.
This is achieved by setting a masking limit in the sky annulus fainter
than which objects are not masked. This produces an accurate sky that
accounts for these fainter galaxies that will be present in the measurement
annulus, but are very difficult to detect and mask. Without this observed sky
value, these faint, barely resolved galaxies prevent the surface photometry
growth curve from converging.
The default axial ratios for the measurement ellipses for the WNGA are
those given by HyperLEDA\footnote{\url{http://leda.univ-lyon1.fr}}
\citep{Paturel:03:45}. However, since the dominant source of error in
calibrating the TFR arises from errors in the HI linewidth inclination
correction, much effort has gone into determining accurate axial ratios and
from them inclinations. For this paper, we chose to use the axial ratios
that were used to determine the correction to the HI linewidths. These are
derived from optical imaging, mostly I-band \citep[see
\S\ref{SEC:IPHOTOMETRY} and \S4.4 in][]{Courtois:11:1935}. We found that,
in the mean, the difference in the W1 photometry between using the default
axial ratios and using the HI linewidth correction axial ratios was on the
order of 4 milli-magnitudes, well below our photometric error threshold.
In order to derive the total magnitudes of the galaxy in the {\it WISE}
bands, $W1_T$ and $W2_T$, the radial photometric profile is analyzed and
two versions of ``total magnitude'' are derived: (1) an asymptotic total
magnitude that is the integration of the galaxy radial profile up to the
point where the profile curve of growth has mathematically converged within
the errors and (2) a procedure that starts with the isophotal magnitude
within 25.5 mag arcsec$^{-2}$ then adds a small extrapolation derived by
extending an exponential disk fit to infinity \citep{Tully:96:2471}. The
extrapolation is given by the formula
\begin{equation}
\Delta m_{ext} = 2.5\ {\rm log}[1-(1+\Delta n) e^{-\Delta n}]
\label{delm}
\end{equation}
where $\Delta n = (\mu_{25.5}-\mu_0)/1.086$ is the number of disk
exponential scalelengths between the central surface brightness $\mu_0$ and
the limiting isophotal surface brightness $\mu_{25.5}$. The exponential
disk central surface brightness $\mu_0$ excludes the bulge by defining the
exponential disk fit over the range from the effective radius (enclosing
half the light of the galaxy) to the $\mu_{25.5}$ isophotal radius. If the
disk central surface brightness is brighter than $\mu_0= 20$ then the
correction $\Delta m_{ext}$ is less than $0.03$ mag.
The ensemble difference between these two types of magnitudes is
characterized by a mean offset of 0.0003 magnitudes and a standard
deviation of 0.0234 magnitudes. To check for a systematic trend with
magnitude we fit the differences as a function of asymptotic magnitude and
derived a line with a slope of 0.0013 $\pm$ 0.0006 and a zero-point of
0.0109 $\pm$ 0.0073. As a further check, we used both of these magnitudes
to carry out the calibration and the resulting set of coefficients were
statistically identical. We have chosen to use the asymptotic magnitudes
for the calibration presented herein because they are a standard output
product of the WNGA and thus require no extra processing beyond our
photometry pipeline. In addition, the extrapolated disk magnitudes are
only appropriate for disk galaxies, while the asymptotic magnitudes are
consistent regardless of galaxy type.
We convert our W1 and W2 magnitudes from the Vega to the AB system using
the Vega-AB offsets of 2.699 mag for W1 and 3.339 mag for W2 from Table~3
of Section~IV.4.h of the Explanatory Supplement to the {\it WISE} All-Sky
Data Release
Products\footnote{\url{http://wise2.ipac.caltech.edu/docs/release/allsky/expsup/ sec4\_4h.html}}.
Uncertainties in the observed W1 magnitudes are similar to or smaller than
those measured for the IRAC [3.6] magnitudes
\citep[$\pm0.05$,][]{Sorce:12:133}. The smaller uncertainties arise for
galaxies that have a large number of individual images from the {\it WISE}
survey and thus when coadded are deeper than the IRAC [3.6] images. This
variable depth coverage in the {\it WISE} survey is due to the fact that
the scans were conducted as great circles intersecting at the ecliptic
poles \citep{Wright:10:1868}, thus the frame coverage density increases
from a minimum at the ecliptic plane to a maximum at the ecliptic poles.
We apply the following corrections to our measured total magnitudes:
\begin{enumerate}
\item $A^{[W1,2]}_b$, a Milky Way extinction correction\citep{Schlafly:11:103,Fitzpatrick:99:63}
\item $A^{[W1,2]}_i$, an internal extinction correction\citep{Giovanelli:95:1059,Giovanelli:97:22,Tully:98:2264},
\item $A^{[W1,2]}_k$, a Doppler shift or k-correction \citep{Oke:68:21,Huang:07:840}.
\item $A^{[W1,2]}_a$, a total flux aperture correction from Table 5 of Section IV.4.c of the {\it WISE} Explanatory Supplement\footnote{\url{http://wise2.ipac.caltech.edu/docs/release/allsky/expsup/ sec4\_4c.html}}.
\end{enumerate}
In these and subsequent equations the notation $W1,2$ means the values for
the {\it WISE} W1 and W2 bands. All these corrections are discussed in
detail in \citet{Sorce:12:133}. The internal extinction correction is
described by the formula $A_i^{[W1,2]} = \gamma_{W1,2}\log(a/b)$
\citep{Tully:98:2264}, where $a/b$ is the major to minor axial ratio and
$\gamma_{W1}$ has the form
\begin{equation}
\gamma_{W1} = 0.12 + 0.21 (\log W^i_{mx} - 2.5).
\end{equation}
The factor $\gamma_{W2}$ can be obtained by multiplying $\gamma_{W1}$ by
the ratio of the reddening coefficients $R_{W2}/R_{W1} = 0.661$
\citep{Fitzpatrick:99:63}. The k-corrections for $W1$ and $W2$ are very
small and roughly the same over the redshift range of interest. The
correction is based on Figure~6 in \citet{Huang:07:840} and has the form
$A^{[W1,2]}_k = -2.27z$. The {\it WISE} aperture correction,
$A^{[W1,2]}_a$, arises because the photometric calibration of {\it WISE} is
conducted with point sources within a fixed aperture that misses some of
the scattered light that is picked up in the extended apertures required to
measure galaxies. The fixed apertures used for the {\it WISE} W1 and W2
photometric calibrations are 8.25 arcseconds in radius and therefore much
smaller than any of the galaxies used in this paper, thus each galaxy has
fixed corrections of $A^{W1}_a = -0.034$ mag and $A^{W2}_a = -0.041$ mag
applied. The fully corrected {\it WISE} magnitude is then
\begin{equation}
\begin{split}
W1,2^{b,i,k,a}_T = W1,2_T &-A^{[W1,2]}_b - A^{[W1,2]}_i \\
&-A^{[W1,2]}_k - A^{[W1,2]}_a.
\end{split}
\end{equation}
\subsection{I-band Photometry\label{SEC:IPHOTOMETRY}}
The sources of the I-band photometry were discussed in \citet{Tully:12:78}.
There are contributions from \citet{Courtois:11:1935} and from the
literature. The present calibration is augmented with 24 new galaxies, an
increase of 9\%. Photometric corrections and analysis procedures are the
same as in the previous publication save for the small shift in reddening
due to our Galaxy in going from \citet{Schlegel:98:525} to
\citet{Schlafly:11:103} and the small shift in distance scale zero point
implicit in the shift of the LMC modulus from 18.50 to 18.48
\citep{Scowcroft:11:76, Scowcroft:12:84, Monson:12:146, Freedman:12:24}.
The main interest of the current paper is the calibration of the {\it WISE}
W1 and W2 band TFR, but an I-band re-calibration is worth presenting. We
collect I-band magnitudes because, as will be discussed in
\S\ref{SEC:COLOR_TERM}, we can couple the I-band and {\it WISE} magnitudes
and recover the I-band scatter through an optical - {\it WISE} color
correction. For determining the {\it WISE} color terms, we convert the
I-band Vega magnitudes to the AB system using the offset from
\citet{Frei:94:1476} of 0.342 magnitudes. This publication provides an
opportunity to update the I-band calibration to assure consistency between
optical and MIR distance measurements. For the I-band TFR re-calibration,
the native Vega system is used.
The resulting input data for calibrating the Tully-Fisher relation are
presented in Table~\ref{tab_data}. This table gives the input total W1, W2
AB photometry, $W1_T$ and $W2_T$, and the input total I-band Vega
photometry, $I_T$, and the corrected magnitudes, $W1_T^{b,i,k,a}$,
$W2_T^{b,i,k,a}$, and $I_T^{b,i,k}$ for each calibrator galaxy. Also
presented are the optical to MIR AB colors, along with the axial ratios and
inclinations and input and corrected HI linewidths and the sample (ZeroPt
or cluster) each calibrator resides in.
\section{The W1 and W2 Calibration\label{SEC:W1CALIBRATION}}
The similarity of {\it WISE} W1 and W2 bands allows us to use identical
procedures for both bands. Thus we will describe both calibrations and
present both sets of results together.
It has been shown that the Malmquist bias incurred by fitting the direct
TFR can be mitigated by fitting the inverse relation \citep{Willick:94:1}.
The major effect of the bias in fitting the direct relation is to flatten
the slope since fainter galaxies with the same linewidth are excluded by a
photometric or signal-to-noise cut. Even with fitting the inverse relation
a residual bias due to scatter in the sample remains. This is addressed in
section~\ref{SEC:BIAS}.
We use a linear regression fitting technique that uses the linewidth errors
as the input measurement error. For the {\it WISE} data, this is sensible
since the formal measurement errors on the magnitudes are very small
compared to the linewidth errors. The I-band magnitude errors are larger
and so we will make an adjustment to the linewidth errors that will account
for these larger photometric errors (see \S\ref{SEC:ICALIBRATION}).
\subsection{Relative Distances and TFR Slope\label{SEC:RELATIVEDISTANCES}}
\begin{figure}
\includegraphics[width=\linewidth]{f1a}
\includegraphics[width=\linewidth]{f1b}
\caption{Linear Tully-Fisher relation (TFR) in the {\it WISE} W1
(top) and W2 (bottom) bands for the Virgo Cluster. The solid line
is the inverse fit of the universal template correlation. The
dashed line is the fit to Virgo alone.}
\label{fig_virgo}
\end{figure}
The TFR posits a universal slope in luminosity versus HI linewidth. Our
first step in deriving this universal slope is to fit each cluster
individually. The results of these fits are shown in
Figures~\ref{fig_virgo} and \ref{fig_others}. Examining the dashed lines
in these figures shows how similar the individual slopes are. In addition,
we see no significant trend in the slope with distance, a benefit of using
the ITFR which mitigates the Malmquist bias. The slope values for the
individual clusters are given in column three of
Tables~\ref{tab_clusters_w1} and \ref{tab_clusters_w2}.
\begin{figure}
\includegraphics[width=\linewidth]{f2a}
\includegraphics[width=\linewidth]{f2b}
\caption{Linear TFR in the {\it WISE} W1 (top) and W2 (bottom)
bands for Ursa Major, Fornax, Centaurus, Antlia, Pegasus, Hydra,
Cancer, Pisces, Abell 400, Coma, Abell 1367, and Abell 2634/2666.
Solid lines are the inverse fit of the universal template, while
dashed lines are the fits for each cluster.}
\label{fig_others}
\end{figure}
\tabletypesize{\scriptsize}
\tabcolsep=0.11cm
\begin{deluxetable*}{lrrrrrrrrr}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablecaption{W1 Cluster Fit Properties\label{tab_clusters_w1}}
\tablehead{
\colhead{Cluster\tablenotemark{1}} &
\colhead{N\tablenotemark{2}} &
\colhead{Slope\tablenotemark{3}} &
\colhead{ZP\tablenotemark{4}} &
\colhead{rms\tablenotemark{5}} &
\colhead{ZP$_{cur}$\tablenotemark{6}} &
\colhead{rms$_{cur}$\tablenotemark{7}} &
\colhead{N$_{cc}$\tablenotemark{8}} &
\colhead{ZP$_{cc}$\tablenotemark{9}} &
\colhead{rms$_{cc}$\tablenotemark{10}}
}
\startdata
Virgo & 30 & -9.16 $\pm$ 0.38 & 10.66 $\pm$ 0.11 & 0.60 & 10.67 $\pm$ 0.06 & 0.55 & 30 & 10.83 $\pm$ 0.09 & 0.52 \\
U Ma & 35 & -9.81 $\pm$ 0.37 & 10.80 $\pm$ 0.11 & 0.63 & 10.73 $\pm$ 0.05 & 0.70 & 34 & 10.95 $\pm$ 0.10 & 0.58 \\
Fornax & 15 & -9.52 $\pm$ 0.56 & 10.85 $\pm$ 0.13 & 0.50 & 10.81 $\pm$ 0.09 & 0.53 & 15 & 10.99 $\pm$ 0.11 & 0.44 \\
Antlia & 21 & -12.40 $\pm$ 1.30 & 12.54 $\pm$ 0.11 & 0.51 & 12.46 $\pm$ 0.05 & 0.50 & 16 & 12.70 $\pm$ 0.09 & 0.36 \\
Centaurus & 15 & -14.16 $\pm$ 1.43 & 12.53 $\pm$ 0.15 & 0.59 & 12.49 $\pm$ 0.07 & 0.62 & 13 & 12.71 $\pm$ 0.14 & 0.52 \\
Pegasus & 18 & -9.10 $\pm$ 0.81 & 12.81 $\pm$ 0.14 & 0.58 & 12.90 $\pm$ 0.09 & 0.57 & 17 & 13.00 $\pm$ 0.10 & 0.39 \\
Hydra & 25 & -9.12 $\pm$ 0.43 & 13.59 $\pm$ 0.13 & 0.67 & 13.40 $\pm$ 0.04 & 0.57 & 19 & 13.74 $\pm$ 0.12 & 0.54 \\
Pisces & 61 & -9.16 $\pm$ 0.28 & 13.91 $\pm$ 0.07 & 0.53 & 13.73 $\pm$ 0.03 & 0.50 & 59 & 13.93 $\pm$ 0.06 & 0.47 \\
Cancer & 13 & -10.35 $\pm$ 0.63 & 13.73 $\pm$ 0.12 & 0.41 & 13.59 $\pm$ 0.06 & 0.41 & 13 & 13.80 $\pm$ 0.10 & 0.35 \\
A400 & 9 & -9.84 $\pm$ 2.67 & 14.54 $\pm$ 0.13 & 0.40 & 14.41 $\pm$ 0.07 & 0.34 & 8 & 14.67 $\pm$ 0.10 & 0.28 \\
A1367 & 23 & -9.46 $\pm$ 0.68 & 14.63 $\pm$ 0.09 & 0.46 & 14.43 $\pm$ 0.04 & 0.39 & 22 & 14.66 $\pm$ 0.09 & 0.42 \\
Coma & 24 & -7.62 $\pm$ 0.40 & 14.58 $\pm$ 0.09 & 0.47 & 14.37 $\pm$ 0.04 & 0.36 & 24 & 14.60 $\pm$ 0.08 & 0.41 \\
A2634/66 & 21 & -9.91 $\pm$ 0.71 & 14.96 $\pm$ 0.10 & 0.48 & 14.73 $\pm$ 0.04 & 0.44 & 21 & 15.05 $\pm$ 0.09 & 0.43 \\
\enddata
\tablenotetext{1}{\ Cluster name}
\tablenotetext{2}{\ Number of galaxies measured in cluster}
\tablenotetext{3}{\ Slope of the fit to individual clusters}
\tablenotetext{4}{\ Zero-point with universal slope, no color correction (mag)}
\tablenotetext{5}{\ Scatter about universal slope, no color correction (mag)}
\tablenotetext{6}{\ Zero-point with universal curve, no color correction (mag)}
\tablenotetext{7}{\ Scatter about universal curve, no color correction (mag)}
\tablenotetext{8}{\ Number of color-corrected galaxies measured in cluster}
\tablenotetext{9}{\ Zero-point with universal slope after color correction (mag)}
\tablenotetext{10}{\ Scatter about universal slope after color correction (mag)}
\end{deluxetable*}
\begin{deluxetable*}{lrrrrrrrrr}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablecaption{W2 Cluster Fit Properties\label{tab_clusters_w2}}
\tablehead{
\colhead{Cluster\tablenotemark{1}} &
\colhead{N\tablenotemark{2}} &
\colhead{Slope\tablenotemark{3}} &
\colhead{ZP\tablenotemark{4}} &
\colhead{rms\tablenotemark{5}} &
\colhead{ZP$_{cur}$\tablenotemark{6}} &
\colhead{rms$_{cur}$\tablenotemark{7}} &
\colhead{N$_{cc}$\tablenotemark{8}} &
\colhead{ZP$_{cc}$\tablenotemark{9}} &
\colhead{rms$_{cc}$\tablenotemark{10}}
}
\startdata
Virgo & 30 & -9.33 $\pm$ 0.39 & 11.21 $\pm$ 0.12 & 0.64 & 11.21 $\pm$ 0.06 & 0.58 & 30 & 11.42 $\pm$ 0.10 & 0.52 \\
U Ma & 35 & -9.90 $\pm$ 0.37 & 11.34 $\pm$ 0.11 & 0.63 & 11.27 $\pm$ 0.05 & 0.71 & 34 & 11.54 $\pm$ 0.10 & 0.57 \\
Fornax & 15 & -9.85 $\pm$ 0.58 & 11.40 $\pm$ 0.14 & 0.56 & 11.34 $\pm$ 0.09 & 0.60 & 15 & 11.58 $\pm$ 0.12 & 0.45 \\
Antlia & 21 & -12.13 $\pm$ 1.25 & 13.09 $\pm$ 0.11 & 0.51 & 13.00 $\pm$ 0.05 & 0.49 & 16 & 13.28 $\pm$ 0.09 & 0.35 \\
Centaurus & 15 & -14.53 $\pm$ 1.49 & 13.08 $\pm$ 0.16 & 0.62 & 13.04 $\pm$ 0.07 & 0.65 & 13 & 13.30 $\pm$ 0.14 & 0.52 \\
Pegasus & 18 & -9.70 $\pm$ 0.86 & 13.42 $\pm$ 0.14 & 0.61 & 13.50 $\pm$ 0.09 & 0.60 & 17 & 13.61 $\pm$ 0.09 & 0.38 \\
Hydra & 25 & -9.13 $\pm$ 0.44 & 14.19 $\pm$ 0.14 & 0.70 & 13.97 $\pm$ 0.04 & 0.59 & 19 & 14.33 $\pm$ 0.12 & 0.55 \\
Pisces & 61 & -9.28 $\pm$ 0.28 & 14.52 $\pm$ 0.07 & 0.54 & 14.32 $\pm$ 0.03 & 0.51 & 59 & 14.53 $\pm$ 0.06 & 0.47 \\
Cancer & 13 & -10.84 $\pm$ 0.67 & 14.32 $\pm$ 0.13 & 0.48 & 14.16 $\pm$ 0.06 & 0.47 & 13 & 14.39 $\pm$ 0.10 & 0.37 \\
A400 & 9 & -11.14 $\pm$ 3.29 & 15.17 $\pm$ 0.14 & 0.43 & 15.03 $\pm$ 0.07 & 0.36 & 8 & 15.27 $\pm$ 0.10 & 0.29 \\
A1367 & 23 & -9.73 $\pm$ 0.72 & 15.22 $\pm$ 0.11 & 0.51 & 14.99 $\pm$ 0.04 & 0.43 & 22 & 15.24 $\pm$ 0.09 & 0.43 \\
Coma & 24 & -7.61 $\pm$ 0.41 & 15.20 $\pm$ 0.11 & 0.52 & 14.97 $\pm$ 0.04 & 0.39 & 24 & 15.20 $\pm$ 0.09 & 0.42 \\
A2634/66 & 21 & -10.23 $\pm$ 0.76 & 15.55 $\pm$ 0.11 & 0.50 & 15.30 $\pm$ 0.04 & 0.45 & 21 & 15.64 $\pm$ 0.09 & 0.43 \\
\enddata
\tablenotetext{1}{\ Cluster name}
\tablenotetext{2}{\ Number of galaxies measured in cluster}
\tablenotetext{3}{\ Slope of the fit to individual clusters}
\tablenotetext{4}{\ Zero-point with universal slope, no color correction (mag)}
\tablenotetext{5}{\ Scatter about universal slope, no color correction (mag)}
\tablenotetext{6}{\ Zero-point with universal curve, no color correction (mag)}
\tablenotetext{7}{\ Scatter about universal curve, no color correction (mag)}
\tablenotetext{8}{\ Number of color-corrected galaxies measured in cluster}
\tablenotetext{9}{\ Zero-point with universal slope after color correction (mag)}
\tablenotetext{10}{\ Scatter about universal slope after color correction (mag)}
\end{deluxetable*}
In order to find the universal TFR, we must combine all 13 clusters by
shifting the data along the magnitude axis, in effect moving each cluster
to the same distance. Virgo is nearest and most complete and offers a
natural choice for the reference cluster. The individual fits to each
cluster provide an estimate of the relative distances from Virgo through
comparing the TFR zero points (column four of Tables~\ref{tab_clusters_w1}
and \ref{tab_clusters_w2}).
These zero points recommend the following groups. The first group is
comprised of Virgo, Ursa Major, and Fornax, a set that we consider the most
complete because they are all nearby. This group is followed by the
Centaurus--Antlia--Pegasus group, then the Hydra--Cancer--Pisces group, and
finally the group comprised of Coma and the three Abell clusters, A0400,
A1367, and A2634/66. As discussed in \citet{Tully:12:78} and
\citet{Sorce:13:94}, we adopt an iterative procedure for combining the
clusters. Starting with the nearest group, we use the zero points for
Fornax and Ursa Major to shift the galaxy magnitudes within those clusters
to align with Virgo. A least-squares fit to the ITFR is then made to this
aligned group. The resulting ensemble slope is then assumed in fitting all
the individual clusters with only the zero points allowed to vary. Using
these new zero points, we then shift the next group to align with Virgo and
add it to the ensemble fit. This procedure is repeated, adding each of the
groups in turn until we have a final ensemble fit for all 13 clusters.
This procedure has been proven to work \citep{Sorce:13:94,Tully:12:78}
because the slope of the TFR is independent of the magnitude cutoff of each
cluster.
Our resulting universal slopes are -9.56 $\pm$ 0.12 (W1) and -9.74 $\pm$
0.12 (W2). The universal slopes and the shifted cluster ensembles are
shown in Figure~\ref{fig_all} which is annotated with the zero point
offsets relative to Virgo for each cluster. The agreement in these offsets
between the W1 and W2 data are quite good. The universal slope is also
shown as the solid lines in Figures~\ref{fig_virgo} and \ref{fig_others}.
\begin{figure}
\includegraphics[width=\linewidth]{f3a}
\includegraphics[width=\linewidth]{f3b}
\caption{Linear TFR in the {\it WISE} W1 (top) and W2 (bottom)
bands obtained from the galaxies in 13 clusters. Offsets given
with respect to the Virgo Cluster represent distance modulus
differences between each cluster and Virgo. The solid line is the
least-squares fit to all of the offset shifted galaxies with errors
entirely in linewidths, the TFR. These relations have an rms
scatter of 0.54 mag for W1 and 0.56 mag for W2.}
\label{fig_all}
\end{figure}
\subsection{Zero Point and Absolute Distances\label{SEC:ZEROPOINT}}
There are 37 nearby galaxies in our zero point sample (see
Table~\ref{tab_data}) that pass our selection criteria and for which there
are good, independent distances from either the Cepheid period-luminosity
method or the TRGB method. The distance moduli used are from
\citet[Table~2]{Tully:12:78}. Since WISE is an all-sky data set, we are
able to measure the total asymptotic W1 and W2 magnitudes for all of them
and calibrate our distances in an absolute sense. We use the universal
slope and the independent absolute magnitudes as input to our least-squares
fit and allow only the zero point to vary. The resultant fits are shown in
Figure~\ref{fig_zp}.
\begin{figure}
\includegraphics[width=\linewidth]{f4a}
\includegraphics[width=\linewidth]{f4b}
\caption{Linear TFR for the 37 galaxies with distances established
by observations of Cepheid variables or the TRGB for the W1 (top)
and the W2 (bottom). The solid black line is the least-squares fit
with the slope established by the 13 cluster template. The zero
point of the TFR is set at the value of this fit at log$W^i_{mx} =
2.5$, as indicated by the solid (red) vertical and horzontal lines.
The zero-point fits have an rms scatter of 0.45 mag for W1 and 0.49
mag for W2.}
\label{fig_zp}
\end{figure}
The measured zero points are -20.35 $\pm$ 0.07 for W1 and -19.76 $\pm$ 0.08
for W2. As was pointed out in \citet{Sorce:13:94}, NGC2841 is the fastest
rotator and the biggest outlier. There is still no good reason to exclude
this galaxy from the zero point sample, so it is included here.
These zero points allow us to put the {\it WISE} TFR on an absolute scale.
Since we have already calculated the cluster distances relative to Virgo,
we need only calculate the offset between the constrained zero points in
Figure~\ref{fig_all} and the absolute zero points in Figure~\ref{fig_zp}
and apply these offsets to our W1 and W2 ensembles and combine them with
the zero point calibrators, which we do in Figure~\ref{fig_zpall}. The
zero point calibration allows us to express the TFR as
\begin{subequations}
\label{eq_itfr}
\begin{align}
\begin{split}
\mathcal{M}^{b,i,k,a}_{W1} = & - (20.35 \pm 0.07) \\
& - (9.56 \pm 0.12)(\log W^i_{mx} - 2.5),
\label{eq_itfr_w1}
\end{split} \\
\begin{split}
\mathcal{M}^{b,i,k,a}_{W2} = & - (19.76 \pm 0.08) \\
& - (9.74 \pm 0.12)(\log W^i_{mx} - 2.5).
\label{eq_itfr_w2}
\end{split}
\end{align}
\end{subequations}
We adopt a convention here and throughout the paper, that TFR predicted
values are given in script, hence our TFR predicted absolute magnitudes are
given as $\mathcal{M}^{b,i,k,a}_{W1,2}$. To derive the distance modulus
for a given galaxy based on pure W1 or W2 photometry, we subtract the
appropriate predicted TFR absolute magnitude from Equation~\ref{eq_itfr}
from the input corrected total magnitude:
\begin{equation}
\mu_{W1,2} = W1,2^{b,i,k,a}_T - \mathcal{M}^{b,i,k,a}_{W1,2}.
\label{eq_tfmu_wise}
\end{equation}
\begin{figure}
\includegraphics[width=\linewidth]{f5a}
\includegraphics[width=\linewidth]{f5b}
\caption{Linear TFR with slope fit to the galaxies in 13 clusters
and the absolute magnitude scale set by 37 zero-point calibrators
for the W1 (top) and the W2 (bottom).}
\label{fig_zpall}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{f6a}
\includegraphics[width=\linewidth]{f6b}
\caption{Linear TFR for I-band (Vega) using galaxies shifted to the
apparent distance of Virgo (top) and on the absolute magnitude
scale set by 36 zero-point calibrators (bottom). This relation has
an rms scatter of 0.46 mag and a zero-point rms scatter of 0.40
mag.}
\label{fig_icalibration}
\end{figure}
The rms scatter about the mean TFR will allow us to assess the usefulness
of this relation for distance measurement. In order to do this, we use the
zero point and the Virgo offset for each cluster to shift each galaxy
magnitude in a given cluster onto the absolute magnitude scale. We compare
this ensemble absolute magnitude, $M^{ens}_{W1,2}$, with the predicted
magnitudes from Equation~\ref{eq_itfr} to derive the residual for every
galaxy in the sample as follows:
\begin{equation}
\Delta M_{W1,2} = M^{ens}_{W1,2} - \mathcal{M}^{b,i,k,a}_{W1,2}.
\label{eq_resid}
\end{equation}
We calculate the rms scatter of the resulting ensemble of residuals. We
define the scatter in this case to be the standard deviation of the
residuals, or the square root of the second moment of the residuals. The
distribution of the residuals is approximately Gaussian, therefore, one can
take these values as a $1\sigma$ error, i.e., 68\% of the galaxies fall
within this $1\sigma$ envelope. The W1 calibration has a scatter of 0.54
magnitudes, while the W2 calibration has a scatter of 0.56 magnitudes
representing distance errors of 27\% and 28\%, respectively. The scatter
in the zero point fits are slightly better at 0.45 mag (W1) and 0.49 mag
(W2). We point out that the formal errors on the zeropoint values are much
smaller at 0.07 mag (W1) and 0.08 mag (W2).
We expect the scatter in the W1 and W2 passbands to exceed that in the I-band.
Since, at a given linewidth, red and blue galaxies separate in magnitude in different passbands, the TFR rms scatter must change with passband.
We expect the scatter in the TFR to reach a minimum where metallicity
and young population effects are minimized. The empirical evidence suggests
the minimum is near the peak of the stellar light for disk galaxies around
$1~\mu$m. The I-band is much closer to $1~\mu$m than are the W1 and W2
bands.
Nonetheless, these are the pure {\it WISE}
W1 and W2 TFRs requiring no other photometry to derive distance moduli to
any galaxy. For a sample that may not have complete I-band coverage, one
may decide that the statistical benefit of a larger sample outweighs the
larger scatter of these pure {\it WISE} TFRs. \citet{Sorce:13:94} discuss
the sources of this scatter in the MIR TFR and conclude that the most
significant arises due to a color term in the TFR. We explore the
analogous color terms for the {\it WISE} W1 and W2 data in
\S\ref{SEC:COLOR_TERM}. In addition, when comparing cluster distances
derived from the pure {\it WISE} and the I-band TFRs, there is evidence for
a systematic offset that may be the result of curvature in the pure MIR TFR
relation. We discuss this in \S\ref{SEC:CURVE}, but first we derive a new
I-band TFR.
\section{I-band Calibration\label{SEC:ICALIBRATION}}
We use the identical procedure to calibrate the I-band TFR as we did for
the {\it WISE} calibration except we adjust the linewidth errors to account
for the larger I-band photometric errors. This adjustment is carried out
as follows. We use a preliminary TFR derived with the original linewidth
errors to project the I-band photometric errors onto the linewidth axis.
This generates a linewidth error due only to the photometric errors. This
photometric linewidth error is then added in quadrature with the original
linewidth errors and the TFR is re-generated. The result of this final
fitting is shown in Figure~\ref{fig_icalibration}. The individual cluster
fits and results for each of the calibration clusters are shown in
Table~\ref{tab_clusters_i}. The error adjustment flattens the TFR slightly
from a slope of -8.97 to -8.95. The final TFR calibration from the I-band
data can be expressed as:
\begin{equation}
\begin{split}
\mathcal{M}^{b,i,k}_{I} = &- (21.34 \pm 0.07) \\
&- (8.95 \pm 0.14)(\log W^i_{mx} - 2.5).
\end{split}
\label{eq_itfr_i}
\end{equation}
The scatter is calculated in the same way as for the {\it WISE} calibration
and results in a value of 0.46 mag rms, smaller than for the {\it WISE} bands
as expected. The formula for the distance modulus using the I-band TFR is:
\begin{equation}
\mu_I = I^{b,i,k}_T - \mathcal{M}^{b,i,k}_I.
\label{eq_tfmu_i}
\end{equation}
\begin{deluxetable}{lrrrr}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablecaption{I-band Cluster Fit Properties\label{tab_clusters_i}}
\tablehead{\colhead{Cluster\tablenotemark{1}} &
\colhead{N\tablenotemark{2}} &
\colhead{Slope\tablenotemark{3}} &
\colhead{ZP\tablenotemark{4}} &
\colhead{rms\tablenotemark{5}}
}
\startdata
Virgo & 30 & -8.75 $\pm$ 0.39 & 9.77 $\pm$ 0.09 & 0.51 \\
U Ma & 34 & -8.46 $\pm$ 0.36 & 9.88 $\pm$ 0.10 & 0.57 \\
Fornax & 15 & -8.64 $\pm$ 0.54 & 9.88 $\pm$ 0.11 & 0.42 \\
Antlia & 16 & -11.26 $\pm$ 1.38 & 11.52 $\pm$ 0.10 & 0.40 \\
Centaurus & 13 & -11.09 $\pm$ 1.22 & 11.60 $\pm$ 0.13 & 0.48 \\
Pegasus & 17 & -7.54 $\pm$ 0.70 & 11.88 $\pm$ 0.10 & 0.43 \\
Hydra & 19 & -9.03 $\pm$ 0.62 & 12.51 $\pm$ 0.13 & 0.55 \\
Pisces & 59 & -9.63 $\pm$ 0.46 & 12.76 $\pm$ 0.06 & 0.46 \\
Cancer & 13 & -8.85 $\pm$ 0.69 & 12.73 $\pm$ 0.09 & 0.32 \\
A400 & 8 & -9.58 $\pm$ 3.21 & 13.56 $\pm$ 0.10 & 0.28 \\
A1367 & 22 & -9.70 $\pm$ 0.93 & 13.45 $\pm$ 0.09 & 0.41 \\
Coma & 24 & -7.00 $\pm$ 0.49 & 13.39 $\pm$ 0.08 & 0.39 \\
A2634/66 & 21 & -9.26 $\pm$ 0.83 & 13.94 $\pm$ 0.10 & 0.44 \\
\enddata
\tablenotetext{1}{\ Cluster name}
\tablenotetext{2}{\ Number of galaxies measured in cluster}
\tablenotetext{3}{\ Slope of the fit to individual clusters}
\tablenotetext{4}{\ Zero-point with universal slope (mag)}
\tablenotetext{5}{\ Scatter about universal slope (mag)}
\end{deluxetable}
The calibration cluster distances derived from the single-band uncorrected
TFRs in the {\it WISE} bands and the I-band are listed in columns five
through seven of Table~\ref{tab_tfr_distances}. The cluster distance
offsets for the {\it WISE} linear TFR, relative to the I-band TFR, are
illustrated by the (red) open diamonds in Figure~\ref{fig_wise_curlin}.
The particular cluster is indicated with the code listed in column two of
Table~\ref{tab_tfr_distances}. We note that, compared to our I-band
distances, the linear {\it WISE} TFR predicts distances that are lower for
nearby clusters and higher for more distant clusters. If there is
curvature in the MIR TFR, it could manifest itself in just this fashion
\citep[see \S V.a.i in][]{Aaronson:86:536}. Since we must use more distant
clusters to estimate $H_0$ from the TFR, such a systematic deviation from a
linear relation would bias the distances larger and produce a smaller
$H_0$. Thus it behooves us to consider this possible curvature in more
detail.
\begin{figure}
\includegraphics[width=\linewidth]{f7a}
\includegraphics[width=\linewidth]{f7b}
\caption{Distance offsets in Mpc relative to the I-band linear TFR
of the calibration clusters for W1 (top) and W2 (bottom) using the
linear TFR, shown by (red) diamonds and the curved TFR, shown by
(blue) squares. The dashed (blue) lines show the average offset
for clusters beyond 50 Mpc ($>4000$ km~s$^{-1}$) for the curved TFRs,
while the dash-dot (red) lines show the average offset for the
linear TFRs.}
\label{fig_wise_curlin}
\end{figure}
\begin{deluxetable*}{llccccccccc}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablecaption{TFR Cluster Distance Comparison\tablenotemark{1}\label{tab_tfr_distances}}
\tablehead{
& & & & \multicolumn{7}{c}{This work} \\
\cline{5-11} \\
\colhead{Cluster} &
\colhead{Code\tablenotemark{2}} &
\colhead{\citet{Tully:12:78}} &
\colhead{\citet{Sorce:13:94}} &
\colhead{I-band} &
\colhead{W1} &
\colhead{W2} &
\colhead{W1cur} &
\colhead{W2cur} &
\colhead{W1cc} &
\colhead{W2cc} \\
}
\startdata
Virgo & V & $15.9 \pm 0.8$ & $14.7 \pm 0.9$ &
$16.6 \pm 0.8$ &
$15.9 \pm 0.9$ & $15.6 \pm 1.0$ &
$16.9 \pm 0.6$ & $16.8 \pm 0.6$ & $16.2 \pm 0.8$ & $16.2 \pm 0.9$ \\
U Ma & U & $17.4 \pm 0.9$ & $18.0 \pm 0.9$ &
$17.5 \pm 0.9$ &
$16.9 \pm 1.0$ & $16.6 \pm 1.0$ &
$17.4 \pm 0.6$ & $17.2 \pm 0.6$ & $17.2 \pm 0.9$ & $17.1 \pm 0.9$ \\
Fornax & F & $17.3 \pm 1.0$ & $17.4 \pm 1.2$ &
$17.5 \pm 1.0$ &
$17.4 \pm 1.2$ & $17.1 \pm 1.2$ &
$18.1 \pm 0.8$ & $17.8 \pm 0.8$ & $17.5 \pm 1.0$ & $17.4 \pm 1.0$ \\
Antlia & An & $37 \pm 2$ & $37 \pm 2$ &
$38 \pm 2$ &
$39 \pm 2$ & $38 \pm 2$ &
$40 \pm 1$ & $39 \pm 1$ & $39 \pm 2$ & $39 \pm 2$ \\
Centaurus & Ce & $38 \pm 3$ & $39 \pm 4$ &
$39 \pm 3$ &
$38 \pm 3$ & $37 \pm 3$ &
$39 \pm 2$ & $39 \pm 2$ & $39 \pm 3$ & $38 \pm 3$ \\
Pegasus & Pe & $43 \pm 3$ & $45 \pm 3$ &
$44 \pm 2$ &
$43 \pm 3$ & $43 \pm 3$ &
$47 \pm 2$ & $48 \pm 2$ & $44 \pm 2$ & $44 \pm 2$ \\
Hydra & H & $59 \pm 4$ & $56 \pm 4$ &
$59 \pm 4$ &
$62 \pm 4$ & $62 \pm 4$ &
$60 \pm 2$ & $60 \pm 2$ & $62 \pm 4$ & $62 \pm 4$ \\
Pisces & Pi & $64 \pm 2$ & $65 \pm 3$ &
$67 \pm 3$ &
$72 \pm 3$ & $73 \pm 3$ &
$71 \pm 2$ & $71 \pm 2$ & $68 \pm 3$ & $68 \pm 3$ \\
Cancer & Ca & $65 \pm 3$ & $67 \pm 4$ &
$66 \pm 3$ &
$66 \pm 4$ & $67 \pm 5$ &
$66 \pm 2$ & $66 \pm 2$ & $64 \pm 3$ & $64 \pm 4$ \\
A400 & A4 & $94 \pm 5$ & $97 \pm 5$ &
$100 \pm 5$ &
$102 \pm 7$ & $105 \pm 8$ &
$103 \pm 4$ & $105 \pm 4$ & $100 \pm 5$ & $100 \pm 6$ \\
A1367 & A1 & $94 \pm 5$ & $96 \pm 6$ &
$94 \pm 5$ &
$104 \pm 6$ & $105 \pm 6$ &
$101 \pm 3$ & $101 \pm 3$ & $98 \pm 5$ & $98 \pm 5$ \\
Coma & Co & $90 \pm 4$ & $95 \pm 6$ &
$90 \pm 4$ &
$99 \pm 5$ & $101 \pm 6$ &
$96 \pm 3$ & $97 \pm 3$ & $94 \pm 5$ & $94 \pm 5$ \\
A2634/66 & A2 & $121 \pm 7$ & $112 \pm 7$ &
$117 \pm 6$ &
$121 \pm 7$ & $121 \pm 7$ &
$116 \pm 3$ & $116 \pm 3$ & $117 \pm 6$ & $117 \pm 6$ \\
\enddata
\tablenotetext{1}{\ all distances in Mpc}
\tablenotetext{2}{\ see Figure\ref{fig_bias}}
\end{deluxetable*}
\section{Curvature in the MIR TFR\label{SEC:CURVE}}
Curvature in the near-IR TFR has been seen before using H-band luminosities, as
described in \citet[and references therein]{Aaronson:86:536}. We adopt the
same strategy for dealing with the curvature, namely we take an empirical
approach rather than attempt to correct the magnitudes or linewidths.
Quadratic fits are also used in \citet[Appendix C]{Sakai:00:698} for the
$BVRIH_{-0.5}$ bands which show an increase in the curvature term with
wavelength.
We treat the curvature of the MIR TFR as an additional bias, or
perturbation term on top of the linear relation seen in optical TFRs. By
adding a curvature term, we are fitting the relation with a quadratic and
as such the curvature of a quadratic requires that we fit with the
dependent variable in magnitudes. We attempted to fit an inverted
quadratic, but the curvature does not follow the data well, thus we are
forced to fit the direct TFR. We minimize the Malmquist bias by using the
ensemble of cluster galaxies shifted to the distance of Virgo with the
linear TFR as the input for the fit. Fitting the direct relation with a
least-squares fitter means that our errors will be on the magnitude axis,
however we have already stated that the linewidth errors dominate,
especially compared to the {\it WISE} photometry. We therefore use the
linear TFR to project the linewidth errors onto the magnitude axis and use
these projected magnitude errors in the fitting.
We fit the same ensemble created from the linear fit to derive a universal
curve. The results of these fits for both W1 and W2 are shown by the solid
(green) lines in Figure~\ref{fig_ens_curve}, while the linear fits are
shown by the dashed (red) lines \citep[compare these to Figure 5
in][]{Aaronson:86:536}. We notice that the curved fits are close to the
linear fits, especially at the faint end of the relations. The curved fits
reduce the rms scatter from 0.54 to 0.52 mag in W1 and from 0.56 to 0.55
mag in W2. This brings the distance errors down to 26\% in W1 and down to
27\% in W2. In addition, the curved fits improve $\chi^2_{\nu}$, which
goes from 3.1 to 2.5 in W1 and from 3.4 to 2.6 in W2.
We present the annotated ensemble in Figure~\ref{fig_all_curve} for W1 in
the top panel and for W2 in the bottom panel. Both of the fits have
similar slope terms of $-8.36 \pm 0.11$ for W1 and $-8.40 \pm 0.12$ for W2.
The curvature terms are $3.60 \pm 0.50$ for W1 and $4.32 \pm 0.51$ for W2.
We could compare these curvatures to the one found in
\citet{Aaronson:86:536}, for the H-band, however, they use a different
velocity measure for their TFR fitting. The distance modulus offsets
from Virgo listed on
the annotation for the figure are in reasonable agreement with those for
the I-band shown in Figure~\ref{fig_icalibration}. The cluster distances
shown in column five of Table~\ref{tab_tfr_distances} for the I-band are
also in agreement with the distances for the curved {\it WISE} TFRs shown
in columns eight and nine. The distance offsets relative to the I-band TFR
distances are shown graphically in Figure~\ref{fig_wise_curlin} by the
(blue) open squares. It is clear that the curved fits reduce the
systematic relative to the I-band.
Now that we have a universal curve, we can use the same procedure that was
used for the linear TFR to find the zero-point of the curved relation. We
present the results of the zero-point fitting in Figure~\ref{fig_zp_curve}.
The formal errors on the zero-point values are smaller relative to the
linear fits as is the rms scatter which is reduced from 0.45 to 0.39 mag in
W1 and from 0.49 to 0.43 mag in W2. We point out that NGC2841 is no longer
such a large outlier as it was with the linear TFR. The final curved TFR
is presented for both W1 and W2 in Figure~\ref{fig_zpall_curve}. We
express the curved {\it WISE} TFR as
\begin{subequations}
\label{eq_itfr_curve}
\begin{align}
\begin{split}
\mathcal{M}^{b,i,k,a}_{W1} = &- (20.48 \pm 0.05) \\
&- (8.36 \pm 0.11)(\log W^i_{mx} - 2.5) \\
&+ (3.60 \pm 0.50)(\log W^i_{mx} - 2.5)^2,
\label{eq_itfr_w1_curve}
\end{split} \\
\begin{split}
\mathcal{M}^{b,i,k,a}_{W2} = &- (19.91 \pm 0.05) \\
&- (8.40 \pm 0.12)(\log W^i_{mx} - 2.5) \\
&+ (4.32 \pm 0.51)(\log W^i_{mx} - 2.5)^2.
\label{eq_itfr_w2_curve}
\end{split}
\end{align}
\end{subequations}
The distance modulus can then be calculated as before with
Equation~\ref{eq_tfmu_wise}.
\begin{figure}
\includegraphics[width=\linewidth]{f8a}
\includegraphics[width=\linewidth]{f8b}
\caption{Fits to the ensemble of 13 clusters shifted to the
distance of Virgo for the {\it WISE} W1 (top) and W2 (bottom). The
solid (green) line is the quadratic error-weighted fit to the
direct TFR with errors entirely in linewidth, projected onto the
magnitude axis using the linear TFR. The dashed (red) line is the
linear fit to the inverse TFR. The two fits are very similar,
especially at the faint end. The annotations are for the curved
direct fit and show an improvement in both rms scatter and
$\chi^2_{\nu}$ over the linear fit.}
\label{fig_ens_curve}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{f9a}
\includegraphics[width=\linewidth]{f9b}
\caption{Curved TFR in the {\it WISE} W1 (top) and W2 (bottom)
bands obtained from the galaxies in 13 clusters. Offsets given
with respect to the Virgo Cluster represent distance modulus
differences between each cluster and Virgo. The solid line is the
least-squares fit to all of the offset shifted galaxies with errors
entirely in linewidths, projected onto the magnitude axis using the
linear TFR. the relations have an rms scatter of 0.52 mag for W1
and 0.55 mag for W2.}
\label{fig_all_curve}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{f10a}
\includegraphics[width=\linewidth]{f10b}
\caption{Curved TFR for the {\it WISE} W1 (top) and W2 (bottom)
band using the 37 galaxies with distances established by
observations of Cepheid variables or the TRGB. The solid line is
the least-squares fit with the coefficients established by the 13
cluster template. The zero point of the TFR is set at the value of
this fit at log$W^i_{mx} = 2.5$ as indicated by the solid (red)
vertical and horzontal lines. The zero-point fits have a scatter
of 0.39 mag in W1 and 0.43 mag in W2.}
\label{fig_zp_curve}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{f11a}
\includegraphics[width=\linewidth]{f11b}
\caption{Curved TFR with the curve fit to the galaxies in 13
clusters and the absolute magnitude scale set by 37 zero-point
calibrators for the W1 (top) and the W2 (bottom).}
\label{fig_zpall_curve}
\end{figure}
\section{Optical - MIR Color Term\label{SEC:COLOR_TERM}}
We do not repeat the color term discussion from \citet{Sorce:13:94},
however, we remind the reader that there is good reason to suspect that a
color term might exist because the TFR relation steepens with wavelength.
Indeed, such a color term was detected in \citet{Sorce:13:94} (see their
\S3.3 and their Figures~6 through 8) and was used to reduce their scatter
from 0.49 to 0.44 magnitudes. The {\it WISE} data also show correlations
between the optical to MIR color and the mean linear TFR residuals as shown
in Figure~\ref{fig_color_term}. Figure~\ref{fig_color_term_i} shows an
attempt to find a similar trend in the I-band residuals, but the slope of
our fit is consistent with an insignificant ($1.2\sigma$) correlation. We
note that in this section we are using the linear, not the curved, {\it
WISE} TFR.
\begin{figure*}
\includegraphics[width=0.5\linewidth]{f12a}
\includegraphics[width=0.5\linewidth]{f12b}
\caption{Deviations from the mean linear {\it WISE} TFRs as a
function of $I - W1$ (left) and $I - W2$ (right) color.}
\label{fig_color_term}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{f13}
\end{center}
\caption{Deviations from the mean I-band TFR as a function of $I -
W1$ color (note that the slope is significant only at $1.2\sigma$
and the zero point is consistent with zero).}
\label{fig_color_term_i}
\end{figure}
We use the I-band minus {\it WISE} band color to correct the magnitudes
and improve the scatter of the fits following \citet{Sorce:13:94}. We
use the residuals calculated in Equation~\ref{eq_resid} and then we fit
the correlation between the I-band to {\it WISE} band color and the
residuals as shown in Figure~\ref{fig_color_term}. Thus we derive the
correction to the magnitude that will produce an absolute magnitude from
the TFR with the least scatter:
\begin{subequations}
\label{eq_cc}
\begin{align}
\Delta W1^{color} & = -0.470 - 0.561(I^{b,i,k}_{T} -
W1^{b,i,k,a}_{T}), \label{eq_cc_w1} \\
\Delta W2^{color} & = -0.874 - 0.617(I^{b,i,k}_{T} -
W2^{b,i,k,a}_{T}). \label{eq_cc_w2}
\end{align}
\end{subequations}
These are then used to adjust the input magnitudes as follows:
\begin{equation}
C_{W1,2} = W1,2^{b,i,k,a}_{T} - \Delta W1,2^{color}.
\label{eq_corrected}
\end{equation}
We repeat the entire fitting process using $C_{W1}$ and $C_{W2}$ instead of
$W1^{b,i,k,a}_{T}$ and $W2^{b,i,k,a}_{T}$. Using these pseudo-magnitudes
reduces the ensemble scatter from 0.54 for W1 and 0.56 for W2 to an
ensemble scatter of 0.46 magnitudes for both bands, which compares well
with the scatter of 0.44 magnitudes after color term correction found in
\citet{Sorce:13:94}. The individual cluster zero-points and scatters for
the color-corrected pseudo-magnitudes are shown in columns seven and eight
of Tables~\ref{tab_clusters_w1} and \ref{tab_clusters_w2}. The value of
0.46 mag for the color-corrected ensemble scatter corresponds to a distance
error of 23\% in both W1 and W2. In addition, the universal slopes for W1
and W2 are now nearly identical with a value of -9.12 for W1 and -9.11 for
W2, whereas prior to color correction they were -9.56 for W1 and -9.74 for
W2. The scatter in the zero point sample was also reduced from 0.45 mag in
W1 and 0.49 mag in W2 to 0.41 mag for W1 and 0.42 mag W2. Compare these
with a color term corrected scatter of 0.37 mag for the zero point sample
in \citet{Sorce:13:94}. Figures~\ref{fig_w1_cc} and \ref{fig_w2_cc} show
the result of fitting these pseudo magnitudes. Since the color correction
requires I-band photometry, the sample used for the color correction is
reduced from 310 to 291 galaxies. The number of color-corrected galaxies
in each cluster is listed in column six of Tables~\ref{tab_clusters_w1} and
\ref{tab_clusters_w2}.
\begin{figure}
\includegraphics[width=\linewidth]{f14a}
\includegraphics[width=\linewidth]{f14b}
\caption{Linear TFR for W1 after adjustments for the color term
with galaxies shifted to the apparent distance of Virgo (top) and
on the absolute magnitude scale set by 37 zero-point calibrator
galaxies (bottom).}
\label{fig_w1_cc}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{f15a}
\includegraphics[width=\linewidth]{f15b}
\caption{Linear TFR for W2 after adjustments for the color term
with galaxies shifted to the apparent distance of Virgo (top) and
on the absolute magnitude scale set by 37 zero-point calibrator
galaxies (bottom).}
\label{fig_w2_cc}
\end{figure}
This pseudo-magnitude calibration can now be expressed as
\begin{subequations}
\label{eq_itfr_cc}
\begin{align}
\begin{split}
\mathcal{M}_{C_{W1}} = &- (20.22 \pm 0.07) \\
&- (9.12 \pm 0.12)(\log W^i_{mx} - 2.5),
\label{eq_itfr_cc_w1}
\end{split} \\
\begin{split}
\mathcal{M}_{C_{W2}} = &- (19.63 \pm 0.07) \\
&- (9.11 \pm 0.12)(\log W^i_{mx} - 2.5).
\label{eq_itfr_cc_w2}
\end{split}
\end{align}
\end{subequations}
In order to derive the distance modulus for a given galaxy, we subtract
Equation~\ref{eq_itfr_cc} from Equation~\ref{eq_corrected}:
\begin{equation}
\mu_{C_{W1,2}} = C_{W1,2} - \mathcal{M}_{C_{W1,2}}.
\label{eq_tfmu}
\end{equation}
As a check for a color term in the I-band TFR, we plot the residuals with
respect to the mean TFR, $\Delta M_I = M^{ens}_I - \mathcal{M}^{b,i,k}_I$
(analogous to Equation~\ref{eq_resid}), as a function of the I-band minus
W1 color in Figure~\ref{fig_color_term_i}. The formal error on the zero
point is larger than zero point itself. The slope has a value which is
insignificant at $1.2\sigma$. Thus, we conclude that an I-band color
correction would have little or no effect.
We point out that this color-correction has the effect of linearizing the
{\it WISE} TFR and thereby removing the curvature we found in the pure {\it
WISE} linear TFR. This color-corrected TFR has the advantage of lower
scatter, while the curved TFR has the advantage that it does not rely on
any other source of photometry. We refer to these color-corrected
magnitudes as W1cc and W2cc in plots and tables to distinguish them from
the pure {\it WISE} magnitudes W1 and W2.
\section{Comparison with Previous Calibrations\label{SEC:TFR_COMPARE}}
\begin{deluxetable*}{lrrlcccrlc}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablecaption{TFR Parameter Comparison\label{tab_tfr_params}}
\tablehead{
& &
\multicolumn{4}{c}{Universal Slope/Curve} & &
\multicolumn{3}{c}{Zero Point} \\
\cline{3-6} \cline{8-10} \\
\colhead{Reference} &
\colhead{Photometry} &
\colhead{Ngal} &
\colhead{Slope} &
\colhead{Curve} &
\colhead{rms} &
&
\colhead{Ngal} &
\colhead{Mag} &
\colhead{rms}
}
\startdata
\citet{Tully:12:78} & I-band (Vega) & 267 & $-8.81 \pm 0.16$ & \ldots & 0.41 & & 36 & $-21.39 \pm 0.07$ & 0.36 \\
This work & I-band (Vega) & 291 & $-8.95 \pm 0.14$ & \ldots & 0.46 & & 36 & $-21.34 \pm 0.07$ & 0.40 \\
\hline
\citet{Sorce:13:94} & IRAC [3.6] (AB) & 213 & $-9.74 \pm 0.22$ & \ldots & 0.49 & & 26 & $-20.34 \pm 0.10$ & 0.44 \\
This work & W1 (AB) & 310 & $-9.56 \pm 0.12$ & \ldots & 0.54 & & 37 & $-20.35 \pm 0.07$ & 0.45 \\
This work & curved W1 (AB) & 310 & $-8.36 \pm 0.11$ & $3.60 \pm 0.50$ & 0.52 & & 37 & $-20.48 \pm 0.05$ & 0.39 \\
\hline
\citet{Sorce:13:94} & $M_{C3.6\mu{m}}$ (AB) & 213 & $-9.13 \pm 0.22$ & \ldots & 0.44 & & 26 & $-20.34 \pm 0.08$ & 0.37 \\
This work & $M_{CW1}$ (AB)& 291 & $-9.12 \pm 0.12$ & \ldots & 0.46 & & 36 & $-20.22 \pm 0.07$ & 0.41 \\
\citet{Lagattuta:13:88} & $M_{corr}$ (AB) & 568 & $-10.05$ & \ldots & 0.69 & & \ldots & $-19.54$ & \ldots \\
\hline
This work & W2 (AB) & 310 & $-9.74 \pm 0.12$ & \ldots & 0.56 & & 37 & $-19.76 \pm 0.08$ & 0.49 \\
This work & $M_{CW2}$ (AB)& 291 & $-9.11 \pm 0.12$ & \ldots & 0.46 & & 36 & $-19.63 \pm 0.07$ & 0.42 \\
This work & curved W2 (AB) & 310 & $-8.40 \pm 0.12$ & $4.32 \pm 0.51$ & 0.55 & & 37 & $-19.91 \pm 0.05$ & 0.43 \\
\enddata
\end{deluxetable*}
We compare our results with previous TFR calibrations in
Table~\ref{tab_tfr_params}. In the I-band, the new calibration agrees with
that from \citet{Tully:12:78} to well within the formal errors on the
parameters. Our scatter is a little higher perhaps due to adding fainter
galaxies. Comparing our W1 calibration to the IRAC [3.6] result in
\citet{Sorce:13:94} shows consistency, both in the pure linear calibration
parameters and in the color corrected parameters. When we restrict our
sample to the same galaxies used in \citet{Sorce:13:94}, we obtain the
exact same scatter for the pure linear W1 TFR (0.49 mag). The only
deviation of note is the color-corrected zero point which is fainter for
the W1. The zero-point samples are not identical and it is possible the
color corrections couple with the I-band in a different way due to
differences in filter responses between the IRAC [3.6] and W1 bandpasses.
The curved {\it WISE} TFR offers an improvement over the linear {\it WISE}
TFR, although the rms scatter is still not as good as the color-corrected
linear {\it WISE} TFR. We point out that the formal errors on the
zero-points for both the curved W1 and W2 TFRs are the lowest of all the
fits and the scatter on the zero-point curved W1 calibration is marginally
lower than for the color-corrected linear W1 TFR.
Lastly, we compare our results to the calibration in
\citet{Lagattuta:13:88}. This calibration was derived from {\it WISE}
catalog photometry and not derived by the authors from their own photometry
of the W1 images, as we have done here. The extended photometry is based
on 2MASS apertures with a correction applied to account for the shallowness
of the 2MASS survey as compared to the {\it WISE} survey. No errors on the
individual luminosity-linewidth correlation parameters are given, so we can
only compare the scatter which is greater by 50\% than the calibration
presented here. The zero-point from that paper has been converted from
Vega to AB magnitudes in Table~\ref{tab_tfr_params}.
\section{The Hubble Constant, $H_0$\label{SEC:HUBBLE}}
We can now use the TFR relation to derive distances and, with cosmological
model-corrected recession velocities, estimate the local Hubble constant
($H_0$). Before we do this, we must consider any residual bias in our
distance estimates due to our sample.
\subsection{Distance Bias\label{SEC:BIAS}}
The residual bias pointed out by \citet{Willick:94:1} and discussed in
detail in \citet{Sorce:13:94} is mitigated to some extent here since the
current sample was selected using the 2MASS redshift survey complete to
K=11.75 \citep{Huchra:12:26}, effectively bringing the sample selection
wavelength much closer to the {\it WISE} bands than the original sample,
which was selected in the B-band. However, we still must account for the
fact that with a faint-end limit, more faint galaxies will be scattered
into the sample than bright galaxies out of the sample.
The bias analysis carried out in \citet{Sorce:13:94} and
\citet{Tully:12:78} is repeated here, but using a \citet{Schechter:76:297}
function with $\alpha = -1.0$ instead of $-0.9$. This value of $\alpha$
was arrived at by fitting the {\it WISE} W1 luminosity function of the
combined nearest three calibration clusters: Virgo-Fornax-UMa \citep[see
their \S3.1,]{Tully:12:78}. The bright end characteristic magnitude for
the {\it WISE} is the same as was used for the IRAC [3.6] magnitudes:
$M^{\star}_{W1} = -22$. Augmenting our sample selection with the 2MASS
redshift survey allows us to assume a flat cutoff in the magnitudes as a
function of linewidth. The driving factor in calculating the bias is the
observed scatter in the TFR. Since the scatter in the color-corrected {\it
WISE} TFR (for both W1 and W2) is the same as the I-band TFR (0.46 mag), we
can use a simulation with this scatter to characterize the bias for all
three TFRs. For the pure {\it WISE} linear and curved TFRs we use a
scatter of 0.54 mag. A simulated TFR having the appropriate scatter is
generated from these parameters and randomly sampled at a range of cutoff
magnitude, $M^{lim}$, which slides to brighter limits linearly as distance
increases. The bias $\langle M \rangle_{measured}$ is determined at
intervals of $M^{lim}$ corresponding to increasing distance. At each
cutoff limit, a random set of galaxies brighter than $M^{lim}$ is drawn
from the simulated TFR and used to calculate a new TFR. The average
deviation from the input (true) TFR is the bias $\langle M
\rangle_{measured}$. This bias is plotted in Figure~\ref{fig_bias} for
both the pure (open blue triangles) and color-corrected (solid red circles)
{\it WISE} TFR. The solid and dotted curves are normalized to zero at a
distance modulus of $\mu = 31$ (Virgo) where we are assumed to be complete.
These curves are described by the formulae
\begin{subequations}
\label{eq_bias}
\begin{align}
b_{pure} = 0.006(\mu - 31)^{2.3}
\label{eq_bias_wise} \\
b_{cc} = 0.004(\mu - 31)^{2.3}
\label{eq_bias_cc_wise}
\end{align}
\end{subequations}
where $\mu$ is the distance modulus to the object (galaxy or cluster)
derived using one of Equation~\ref{eq_tfmu_wise}, \ref{eq_tfmu_i} or
\ref{eq_tfmu}. This bias function is slightly steeper than that seen in
\citet{Sorce:13:94} having an exponent of 2.3 instead of 2 due to an
increase in the assumed scatter from 0.40 to 0.46 magnitudes. The letter
codes in Figure~\ref{fig_bias} show the cutoff magnitudes for the
calibration clusters (see column two of Table~\ref{tab_tfr_distances}) by
their horizontal placement and the resulting bias by the vertical
intersection with the solid line. For a galaxy in the field, the corrected
distance modulus is thus
\begin{subequations}
\label{eq_dmod}
\begin{align}
\begin{split}
\mu^c_{pure} =&\ (W1,2_T^{b,i,k,a} - \mathcal{M}_{W1,2}^{b,i,k,a})\ + \\
&\ 0.006[(W1,2_T^{b,i,k,a} - \mathcal{M}_{W1,2}^{b,i,k,a}) - 31]^{2.3},
\end{split} \\
\begin{split}
\mu^c_{cc} =&\ (C_{W1,2} - \mathcal{M}_{C_{W1,2}})\ + \\
&\ 0.004[(C_{W1,2} - \mathcal{M}_{C_{W1,2}}) - 31]^{2.3}.
\end{split}
\end{align}
\end{subequations}
\begin{figure}
\includegraphics[width=\linewidth]{f16}
\caption{Bias $\langle M \rangle_{measured}$ for the pure {\it
WISE} TFR (blue open triangles) and color-corrected {\it WISE} TFR
(solid red circles) as a function of absolute magnitude limit which
increases with distance. The solid curve is the empirical bias fit
to the color-corrected {\it WISE} points which has the form $b =
0.004(\mu - 31)^{2.3}$. The dotted curve is the empirical bias fit
to the pure {\it WISE} data which has the form $b = 0.006(\mu -
31)^{2.3}$. Letters at the bottom are codes for the 13 calibrating
clusters (see column two of Table~\ref{tab_tfr_distances}). Their
horizontal positions indicate sample limits and the vertical
intercepts with the solid curve give the corresponding biases.}
\label{fig_bias}
\end{figure}
We list the biases for the pure {\it WISE} and the color-corrected {\it
WISE} magnitudes in column three of Table~\ref{tab_h0_data}. Now we turn
to using our calibrating clusters to estimate $H_0$.
\begin{deluxetable*}{lrrrrrl}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\tablecaption{Cluster Distances and $H_0$ Data\label{tab_h0_data}}
\tablehead{
\colhead{Cluster\tablenotemark{1}} &
\colhead{$V_{mod}$\tablenotemark{2}} &
\colhead{Bias\tablenotemark{3}} &
\colhead{DM\tablenotemark{4}} &
\colhead{$D_{Mpc}$\tablenotemark{5}} &
\colhead{$V_{mod}/D_{Mpc}$\tablenotemark{6}} &
\colhead{TFR Band\tablenotemark{7}}
}
\startdata
\multirow{5}{*}{Virgo} & \multirow{5}{*}{1495. $\pm$ 37.} & 0.000 & 31.14 $\pm$ 0.08 & 16.93 $\pm$ 0.59 & 88.32 $\pm$ 5.47 & W1cur \\
& & 0.000 & 31.12 $\pm$ 0.08 & 16.78 $\pm$ 0.59 & 89.09 $\pm$ 5.52 & W2cur \\
& & 0.000 & 31.05 $\pm$ 0.12 & 16.21 $\pm$ 0.85 & 92.22 $\pm$ 7.51 & W1cc \\
& & 0.000 & 31.04 $\pm$ 0.12 & 16.16 $\pm$ 0.86 & 92.52 $\pm$ 7.59 & W2cc \\
& & 0.000 & 31.10 $\pm$ 0.11 & 16.61 $\pm$ 0.85 & 90.00 $\pm$ 7.23 & I-band \\
\hline
\multirow{5}{*}{U Ma} & \multirow{5}{*}{1079. $\pm$ 14.} & 0.000 & 31.21 $\pm$ 0.07 & 17.45 $\pm$ 0.56 & 61.83 $\pm$ 2.88 & W1cur \\
& & 0.000 & 31.18 $\pm$ 0.07 & 17.22 $\pm$ 0.55 & 62.66 $\pm$ 2.91 & W2cur \\
& & 0.000 & 31.17 $\pm$ 0.12 & 17.16 $\pm$ 0.92 & 62.90 $\pm$ 4.44 & W1cc \\
& & 0.000 & 31.16 $\pm$ 0.12 & 17.09 $\pm$ 0.92 & 63.13 $\pm$ 4.45 & W2cc \\
& & 0.000 & 31.21 $\pm$ 0.12 & 17.47 $\pm$ 0.93 & 61.78 $\pm$ 4.32 & I-band \\
\hline
\multirow{5}{*}{Fornax} & \multirow{5}{*}{1358. $\pm$ 45.} & 0.000 & 31.28 $\pm$ 0.10 & 18.06 $\pm$ 0.83 & 75.21 $\pm$ 6.21 & W1cur \\
& & 0.000 & 31.25 $\pm$ 0.10 & 17.79 $\pm$ 0.81 & 76.33 $\pm$ 6.31 & W2cur \\
& & 0.000 & 31.21 $\pm$ 0.13 & 17.48 $\pm$ 1.04 & 77.68 $\pm$ 7.63 & W1cc \\
& & 0.000 & 31.21 $\pm$ 0.13 & 17.43 $\pm$ 1.05 & 77.93 $\pm$ 7.75 & W2cc \\
& & 0.000 & 31.22 $\pm$ 0.13 & 17.54 $\pm$ 1.00 & 77.43 $\pm$ 7.42 & I-band \\
\hline
\multirow{5}{*}{Antlia} & \multirow{5}{*}{3198. $\pm$ 74.} & 0.060 & 33.00 $\pm$ 0.07 & 39.81 $\pm$ 1.31 & 80.33 $\pm$ 4.67 & W1cur \\
& & 0.060 & 32.97 $\pm$ 0.07 & 39.26 $\pm$ 1.30 & 81.45 $\pm$ 4.73 & W2cur \\
& & 0.040 & 32.96 $\pm$ 0.11 & 38.99 $\pm$ 1.97 & 82.01 $\pm$ 6.37 & W1cc \\
& & 0.040 & 32.94 $\pm$ 0.11 & 38.78 $\pm$ 1.96 & 82.47 $\pm$ 6.40 & W2cc \\
& & 0.040 & 32.89 $\pm$ 0.12 & 37.91 $\pm$ 2.04 & 84.35 $\pm$ 6.87 & I-band \\
\hline
\multirow{5}{*}{Centaurus} & \multirow{5}{*}{3823. $\pm$ 82.} & 0.000 & 32.97 $\pm$ 0.09 & 39.30 $\pm$ 1.56 & 97.28 $\pm$ 6.19 & W1cur \\
& & 0.000 & 32.95 $\pm$ 0.09 & 38.94 $\pm$ 1.56 & 98.18 $\pm$ 6.29 & W2cur \\
& & 0.000 & 32.93 $\pm$ 0.16 & 38.55 $\pm$ 2.71 & 99.18 $\pm$ 9.79 & W1cc \\
& & 0.000 & 32.92 $\pm$ 0.16 & 38.42 $\pm$ 2.72 & 99.50 $\pm$ 9.89 & W2cc \\
& & 0.000 & 32.93 $\pm$ 0.15 & 38.62 $\pm$ 2.54 & 98.99 $\pm$ 9.26 & I-band \\
\hline
\multirow{5}{*}{Pegasus} & \multirow{5}{*}{3062. $\pm$ 78.} & 0.000 & 33.37 $\pm$ 0.10 & 47.27 $\pm$ 2.09 & 64.77 $\pm$ 4.72 & W1cur \\
& & 0.000 & 33.42 $\pm$ 0.10 & 48.19 $\pm$ 2.15 & 63.53 $\pm$ 4.66 & W2cur \\
& & 0.000 & 33.22 $\pm$ 0.12 & 44.04 $\pm$ 2.32 & 69.54 $\pm$ 5.74 & W1cc \\
& & 0.000 & 33.23 $\pm$ 0.12 & 44.34 $\pm$ 2.29 & 69.06 $\pm$ 5.61 & W2cc \\
& & 0.000 & 33.21 $\pm$ 0.12 & 43.93 $\pm$ 2.45 & 69.70 $\pm$ 5.99 & I-band \\
\hline
\multirow{5}{*}{Hydra} & \multirow{5}{*}{4088. $\pm$ 72.} & 0.015 & 33.89 $\pm$ 0.07 & 59.95 $\pm$ 1.77 & 68.19 $\pm$ 3.32 & W1cur \\
& & 0.015 & 33.90 $\pm$ 0.06 & 60.26 $\pm$ 1.76 & 67.84 $\pm$ 3.28 & W2cur \\
& & 0.010 & 33.97 $\pm$ 0.14 & 62.09 $\pm$ 3.91 & 65.84 $\pm$ 5.67 & W1cc \\
& & 0.010 & 33.96 $\pm$ 0.14 & 62.06 $\pm$ 3.95 & 65.87 $\pm$ 5.72 & W2cc \\
& & 0.010 & 33.85 $\pm$ 0.14 & 58.94 $\pm$ 3.75 & 69.36 $\pm$ 6.02 & I-band \\
\hline
\multirow{5}{*}{Pisces} & \multirow{5}{*}{4759. $\pm$ 39.} & 0.030 & 34.23 $\pm$ 0.06 & 70.21 $\pm$ 1.77 & 67.78 $\pm$ 2.32 & W1cur \\
& & 0.030 & 34.26 $\pm$ 0.06 & 71.19 $\pm$ 1.80 & 66.85 $\pm$ 2.29 & W2cur \\
& & 0.020 & 34.17 $\pm$ 0.09 & 68.30 $\pm$ 2.83 & 69.68 $\pm$ 3.61 & W1cc \\
& & 0.020 & 34.17 $\pm$ 0.09 & 68.36 $\pm$ 2.84 & 69.62 $\pm$ 3.61 & W2cc \\
& & 0.020 & 34.12 $\pm$ 0.09 & 66.65 $\pm$ 2.70 & 71.40 $\pm$ 3.63 & I-band \\
\hline
\multirow{5}{*}{Cancer} & \multirow{5}{*}{5059. $\pm$ 82.} & 0.030 & 34.10 $\pm$ 0.08 & 66.10 $\pm$ 2.43 & 76.54 $\pm$ 4.21 & W1cur \\
& & 0.030 & 34.11 $\pm$ 0.08 & 66.28 $\pm$ 2.42 & 76.32 $\pm$ 4.17 & W2cur \\
& & 0.020 & 34.04 $\pm$ 0.12 & 64.24 $\pm$ 3.41 & 78.75 $\pm$ 5.76 & W1cc \\
& & 0.020 & 34.04 $\pm$ 0.12 & 64.24 $\pm$ 3.54 & 78.75 $\pm$ 5.95 & W2cc \\
& & 0.020 & 34.09 $\pm$ 0.11 & 65.68 $\pm$ 3.28 & 77.03 $\pm$ 5.37 & I-band \\
\hline
\multirow{5}{*}{A400} & \multirow{5}{*}{7228. $\pm$ 97.} & 0.165 & 35.05 $\pm$ 0.08 & 102.52 $\pm$ 3.92 & 70.50 $\pm$ 3.79 & W1cur \\
& & 0.165 & 35.11 $\pm$ 0.08 & 105.10 $\pm$ 3.98 & 68.77 $\pm$ 3.67 & W2cur \\
& & 0.110 & 35.00 $\pm$ 0.12 & 100.00 $\pm$ 5.38 & 72.28 $\pm$ 5.14 & W1cc \\
& & 0.110 & 35.01 $\pm$ 0.12 & 100.37 $\pm$ 5.53 & 72.01 $\pm$ 5.22 & W2cc \\
& & 0.110 & 35.01 $\pm$ 0.12 & 100.46 $\pm$ 5.45 & 71.95 $\pm$ 5.15 & I-band \\
\hline
\multirow{5}{*}{A1367} & \multirow{5}{*}{6969. $\pm$ 93.} & 0.120 & 35.02 $\pm$ 0.06 & 101.06 $\pm$ 2.93 & 68.96 $\pm$ 3.01 & W1cur \\
& & 0.120 & 35.02 $\pm$ 0.06 & 100.93 $\pm$ 2.90 & 69.05 $\pm$ 2.99 & W2cur \\
& & 0.080 & 34.96 $\pm$ 0.11 & 98.13 $\pm$ 4.93 & 71.02 $\pm$ 4.76 & W1cc \\
& & 0.080 & 34.95 $\pm$ 0.11 & 97.54 $\pm$ 5.00 & 71.44 $\pm$ 4.86 & W2cc \\
& & 0.080 & 34.86 $\pm$ 0.11 & 93.89 $\pm$ 4.63 & 74.23 $\pm$ 4.89 & I-band \\
\hline
\multirow{5}{*}{Coma} & \multirow{5}{*}{7370. $\pm$ 76.} & 0.060 & 34.91 $\pm$ 0.06 & 95.81 $\pm$ 2.72 & 76.92 $\pm$ 3.07 & W1cur \\
& & 0.060 & 34.94 $\pm$ 0.06 & 97.41 $\pm$ 2.74 & 75.66 $\pm$ 2.99 & W2cur \\
& & 0.040 & 34.86 $\pm$ 0.11 & 93.93 $\pm$ 4.53 & 78.46 $\pm$ 4.82 & W1cc \\
& & 0.040 & 34.87 $\pm$ 0.11 & 94.06 $\pm$ 4.62 & 78.36 $\pm$ 4.90 & W2cc \\
& & 0.040 & 34.77 $\pm$ 0.10 & 89.91 $\pm$ 4.22 & 81.97 $\pm$ 4.92 & I-band \\
\hline
\multirow{5}{*}{A2634/66} & \multirow{5}{*}{8938. $\pm$ 164.} & 0.105 & 35.32 $\pm$ 0.06 & 115.72 $\pm$ 3.36 & 77.24 $\pm$ 3.77 & W1cur \\
& & 0.105 & 35.32 $\pm$ 0.06 & 115.66 $\pm$ 3.32 & 77.28 $\pm$ 3.74 & W2cur \\
& & 0.070 & 35.34 $\pm$ 0.12 & 116.90 $\pm$ 6.12 & 76.46 $\pm$ 5.71 & W1cc \\
& & 0.070 & 35.33 $\pm$ 0.12 & 116.63 $\pm$ 6.10 & 76.64 $\pm$ 5.71 & W2cc \\
& & 0.070 & 35.35 $\pm$ 0.12 & 117.44 $\pm$ 6.16 & 76.11 $\pm$ 5.69 & I-band \\
\enddata
\tablenotetext{1}{\ Cluster name}
\tablenotetext{2}{\ Mean cluster cosmology-corrected velocity in CMB frame (km~s$^{-1}$)}
\tablenotetext{3}{\ Bias, $b$ (mag)}
\tablenotetext{4}{\ Bias-corrected distance modulus (mag)}
\tablenotetext{5}{\ Cluster distance (Mpc)}
\tablenotetext{6}{\ Hubble parameter km~s$^{-1}$\ Mpc$^{-1}$}
\tablenotetext{7}{\ Source photometry}
\end{deluxetable*}
\subsection{$H_0$ From Clusters\label{SEC:H0_CLUSTERS}}
As we have already pointed out, we expect a systematic problem with using the
linear, uncorrected {\it WISE} TFR (see \S\S\ref{SEC:ICALIBRATION} and
\ref{SEC:CURVE}) to calculate $H_0$. We present these values to illustrate
this systematic, but we concentrate on the curved {\it WISE} TFR or the
color-corrected {\it WISE} TFR for calculating distances used to derive
$H_0$. With distance moduli and hence distances for each cluster derived
from the ensemble of galaxies used to calibrate the cluster (see
Table~\ref{tab_h0_data}), we can use the ensemble velocity to calculate a
Hubble constant, $H_0$, for each cluster. We use the bi-weight method
described in \citet{Beers:90:32} to derive a robust ensemble velocity for
each cluster. These velocities are then shifted to the cosmic microwave
background (CMB) frame and adjusted based on a cosmological model which
assumes $\Omega_{m} = 0.27$ and $\Omega_{\Lambda} = 0.73$. The
cosmological adjustments are admittedly small, but not insignificant.
These velocities and associated errors are listed in
Table~\ref{tab_h0_data} in column two labeled $V_{mod}$
\citep[see][equation 14]{Tully:13:86} to indicate the adjustment for the
cosmological model specified previously. We calculate $H_0 =
V_{mod}/D_{Mpc}$ for each cluster as shown in column seven of the
aforementioned table and plotted for {\it WISE} and the I-band in
Figures~\ref{fig_h0_wise} and \ref{fig_h0_i}.
\begin{figure*}
\includegraphics[width=0.5\linewidth]{f17a}
\includegraphics[width=0.5\linewidth]{f17b}
\includegraphics[width=0.5\linewidth]{f17c}
\includegraphics[width=0.5\linewidth]{f17d}
\caption{Hubble parameter as a function of distance for the {\it
WISE} W1 curved TFR (top left), W2 curved TFR (top right), W1
color-corrected TFR (W1cc, bottom left), and W2 color-corrected TFR
(W2cc, bottom right).}
\label{fig_h0_wise}
\end{figure*}
\begin{figure}
\includegraphics[width=\linewidth]{f18}
\caption{Hubble parameter as a function of distance for the I-band.}
\label{fig_h0_i}
\end{figure}
Examining Figures~\ref{fig_h0_wise} and \ref{fig_h0_i}, we see that the
nearer clusters have a large scatter while those beyond 50 Mpc ($V_{mod} >
4000$ km~s$^{-1}$) have a smaller scatter. This is simply the result of the
peculiar motions induced by local structures in our supercluster complex
(the transformation of velocities to the cosmic microwave background frame
gives all nearby galaxies large peculiar velocities). We plot an error
envelope of 200 km~s$^{-1}$\ as a dotted line in each figure to show the effect of
peculiar velocities on $H_0$ as a function of distance. In order to derive
an estimate of the universal Hubble constant, we consider only clusters
beyond 50 Mpc and we average the log of their resulting $H_0$ values since
the errors are predominantly in the distance and symmetric about the
distance modulus. We find an error-weighted, logarithmic averaged Hubble
constant of $H_0 = 73.1 \pm 1.8$ km~s$^{-1}$\ Mpc$^{-1}$ for both W1cc and W2cc.
For the curved pure {\it WISE} TFR, we get $H_0 = 72.2 \pm 1.7$ km~s$^{-1}$\
Mpc$^{-1}$ for W1 and $H_0 = 71.6 \pm 1.7$ km~s$^{-1}$\ Mpc$^{-1}$ for W2. For
the I-band we get a larger value of $H_0 = 74.5 \pm 1.6$ km~s$^{-1}$\ Mpc$^{-1}$.
This amounts to a range of $\sim\pm$1.5\% from a logarithmic average of
72.9 km~s$^{-1}$\ Mpc$^{-1}$ derived from all five of these cluster TFR $H_0$
values.
Using the linear, pure {\it WISE} TFR, we derive values of $H_0 = 70.6 \pm
1.6$ km~s$^{-1}$\ Mpc$^{-1}$ for W1 and $H_0 = 69.8 \pm 1.6$ km~s$^{-1}$\ Mpc$^{-1}$ for W2.
These values are low compared with either the curved or the color-corrected
TFR values, as expected from a systematic that biases distant clusters toward
larger distances.
All the errorbars listed above are formal statistical errorbars and do not
account for systematics. The calibration clusters may still be strongly
influenced by local large-scale structures and thus may not provide the
most robust estimate of $H_0$. In addition, there are only seven clusters
beyond 50 Mpc and small number statistics may play a role. Systematic
errors are discussed in detail in the next subsection when we extend our
reach well beyond local structures and use a larger sample of supernova
host galaxies to estimate $H_0$.
\subsection{$H_0$ From Supernovae\label{SEC:H0_SUPERNOVAE}}
The precision of distances derived from Type Ia supernovae (SNe Ia) offers
a better avenue for determining $H_0$ free from the small number statistics
that influence the determination of $H_0$ from seven nearby galaxy
clusters. In order to exploit the reach of SNe Ia, which is well beyond
the local velocity perturbations we see in our cluster $H_0$ estimations,
we must tie the SN Ia distance scale to the distance scale established by
the TFR. Even though there are few SN Ia that have been detected in nearby
galaxies, there are 56 SNe Ia that have been detected in host galaxies
within the Cosmic Flows sample (see \S\ref{SEC:HILINEWIDTHS}). These
galaxies also have I-band photometry allowing the color-corrected MIR TFR
to be used in addition to the curved pure {\it WISE} TFR. This permits an
accurate determination of the offset between the SN Ia and TFR distance
moduli.
We use the UNION2 sample of SNe Ia \citep{Amanullah:10:712} for this
comparison. This sample encompasses distances out to beyond $z \sim 1$,
and includes all the SNe Ia hosts from the Cosmic Flows galaxy sample. We
can improve our statistical error in the situation where there are multiple
SNe Ia within a cluster. Of the 13 clusters used to calibrate the TFR,
eight have had one or more SN Ia erupt within one or more member galaxies
\citep[see Table~2 of][]{Sorce:12:L12}. We use the same bi-weight method
from \citet{Beers:90:32} to derive robust averages for the group velocities
and distance moduli based on the SNe Ia and based on the TFR. We also use
SN Ia hosts not in clusters. These individual hosts will have lower weight
by virtue of their higher statistical error, however the ensemble will help
to constrain the offset.
Figure~\ref{fig_mucomp_wise} presents the comparison of the TFR distance
moduli derived from the {\it WISE} passbands and the SN Ia distance moduli
for the eight clusters and 56 individual galaxies, while
Figure~\ref{fig_mucomp_i} shows the same comparison for the I-band. The
distance modulus offsets are derived from error-weighted fits with the
slope fixed at a value of 1. The clusters have the largest influence on
these offsets due to their low statistical errors, yet the resultant fits
appear to bisect the distributions for the individual galaxies as well.
The resulting distance modulus offsets are identical for the curved TFR for
W1 and W2: $0.57 \pm 0.02$ mag. The rms values are calculated only from
the individual galaxy residuals and are 0.45 mag for W1 and 0.48 mag for
W2. For the color-corrected TFR, we find an offset of $0.53 \pm 0.03$ mag
for W1cc and $0.52 \pm 0.03$ mag for W2cc and rms values of 0.53 mag for
W1cc and 0.54 mag for W2cc. For the I-band the offset is $0.51 \pm 0.03$
mag with a scatter of 0.55 mag which is very close to the values shown in
the top panel of Figure~2 from \citet{Courtois:12:174}, which is also
derived only from the SN Ia - TFR offset.
\begin{figure*}
\includegraphics[width=0.5\linewidth]{f19a}
\includegraphics[width=0.5\linewidth]{f19b}
\includegraphics[width=0.5\linewidth]{f19c}
\includegraphics[width=0.5\linewidth]{f19d}
\caption{The distance modulus offsets derived from the SN Ia sample
and from the TF relation for the {\it WISE} W1 curved TFR (top
left), and W2 curved TFR (top right), W1 color-corrected TFR
(bottom left) and W2 color-corrected TFR (bottom right). Open
squares indicate the ensemble robust averages for the eight
clusters that have had one or more SN Ia erupt within member
galaxies. Solid circles indicate individual hosts within which a
SN Ia from the UNION2 sample has erupted. Note the small scatters
and low formal errors using the curved pure {\it WISE} TFRs.}
\label{fig_mucomp_wise}
\end{figure*}
As was pointed out by \citet{Courtois:12:174}, the scatters in the linear
TFR offset data are $\sim 10$\% larger than expected from the combination
of the individual scatters in the linear TFRs (0.46 mag for the
color-corrected W1,2 and the I-band) and the SN Ia (0.20 mag) distance
scales. We don't present the offset data for the pure {\it WISE} linear
TFR, but these scatters are even larger at 0.56 mag for W1 and 0.58 mag for
W2. For the curved pure {\it WISE} TFRs, the SN Ia - TFR distance modulus
scatter is actually less than expected when adding the TFR scatter (0.52
mag for curved W1 and 0.55 mag for curved W2) in quadrature with the SN Ia
distance modulus scatter. In fact, the scatter in the curved W1 is 20\%
less than the scatter in the I-band. The scatter in all of these offsets
have been calculated with the exact same sample and using the exact same
method. With a sample of 56 galaxies, it is hard to explain this away with
small number statistics. It is possible that this results from a better
alignment between the clusters and the individual galaxies using the curved
TFRs, although this is not obvious from Figure~\ref{fig_mucomp_wise}. As
a test of the TFR, this lower scatter is strong evidence in favor of using
the curved pure {\it WISE} TFR for deriving distances.
\begin{figure}
\includegraphics[width=\linewidth]{f20}
\caption{The distance modulus offset derived from the SN Ia sample
and from the TF relation for the I-band. Symbols have the same
meaning as in Figure~\ref{fig_mucomp_wise}.}
\label{fig_mucomp_i}
\end{figure}
The formal errors on the curved TFR W1 and W2 distance modulus offsets
correspond to an error in $H_0$ of 0.7 km~s$^{-1}$\ Mpc$^{-1}$, while the formal
errors on the color-corrected and I-band offsets correspond to an error in
$H_0$ of 1.1 km~s$^{-1}$\ Mpc$^{-1}$. Using the curved TFR W1 and W2 distance
modulus offsets represents a 30\% reduction in the $H_0$ error budget.
Figure~\ref{fig_h0sne} shows the calculation of the normalization of the
Hubble constant, $H_{Norm}$, using only the SNe Ia from the UNION2 sample
that overlap with the Cosmic Flows sample. The overlap zero point is not,
in fact, $H_{Norm} = 100$ km~s$^{-1}$\ Mpc$^{-1}$, but slightly less: $H_{Norm} =
96.8 \pm 2.3$ km~s$^{-1}$\ Mpc$^{-1}$, an offset of 3.2 km~s$^{-1}$\ Mpc$^{-1}$. A
similar offset was found by \citet{Courtois:12:174} when setting the
distance zero-point using the I-band TFR. Once we apply this normalization
offset, we derive values of $H_0 = 73.7 \pm 2.4$ km~s$^{-1}$\ Mpc$^{-1}$ for W1
and W2, and $H_0 = 75.2 \pm 2.5$ km~s$^{-1}$\ Mpc$^{-1}$ for W1cc and $H_0 = 75.5
\pm 2.5$ km~s$^{-1}$\ Mpc$^{-1}$ for W2cc, and $H_0 = 75.9 \pm 2.5$ km~s$^{-1}$\
Mpc$^{-1}$ for the I-band. These errors are the combination in quadrature
of the SN Ia -- TFR offset $H_0$ error (stated in the previous paragraph)
and the UNION2 normalization offset uncertainty shown in
Figure~\ref{fig_h0sne}. Since the two curved pure {\it WISE} values and
the I-band value are independent from one another, we are allowed to
perform a log average of these three values. This gives our best estimate
of the Hubble constant of $H_0 = 74.4 \pm 1.4$ km~s$^{-1}$\ Mpc$^{-1}$. Here the
error is the statistical error in the mean value.
\begin{figure}
\includegraphics[width=\linewidth]{f21}
\caption{Hubble parameter normalization as a function of recession
velocity using only the UNION2 SN Ia distance moduli
\citep{Amanullah:10:712} for galaxies in common with the Cosmic
Flows sample. Here we are only verifying the normalization of this
sample at 100 km~s$^{-1}$\ Mpc$^{-1}$ for our sub-sample of SN Ia host
galaxies. This plot shows that, for our sub-sample, the
normalization is less than the nominal value and thus we need to
apply an offset of $-3.2 \pm 2.3$ km~s$^{-1}$\ Mpc$^{-1}$ when we
re-normalize the SN Ia hosts with the TFR distances.}
\label{fig_h0sne}
\end{figure}
For completeness we present the values derived with the linear pure {\it WISE}
TFR. Using the same SN Ia hosts, we derive $H_0 = 73.0 \pm 2.7$ km~s$^{-1}$\
Mpc$^{-1}$ for W1 and $H_0 = 72.7 \pm 2.7$ km~s$^{-1}$\ Mpc$^{-1}$ for W2. These
values are low as expected from the systematic bias caused by using a
linear fit on a curved TFR.
Systematic errors need to be accounted for in the $H_0$ calculations. By
using IR photometry we reduce the uncertainties due to dust significantly.
Since we have used three bandpasses for these calculations, we can use the
differences to estimate the systematic errors between bands. Based on the
range of $H_0$ we derive for all three bandpasses, we estimate the
inter-band systematic to be $\pm1.1$ km~s$^{-1}$\ Mpc$^{-1}$. The other source of
systematic error is the error on the distance to the LMC, which forms the
basis for our TFR distance scale and has a systematic of $\pm0.033$ mag
\citep{Freedman:12:24}, which also corresponds to an error on $H_0$ of
$\pm1.1$ km~s$^{-1}$\ Mpc$^{-1}$. Another error is the formal TFR zero-point
error (see Table~\ref{tab_tfr_params}, column eight) which is 0.05 mag,
which corresponds to an error on $H_0$ of $\pm1.8$ km~s$^{-1}$\ Mpc$^{-1}$.
Adding these in quadrature gives a systematic error on $H_0$ of $\pm2.4$
km~s$^{-1}$\ Mpc$^{-1}$. Thus our best value is $H_0 = 74.4 \pm1.4$(stat)
$\pm\ 2.4$(sys) km~s$^{-1}$\ Mpc$^{-1}$. We can add the statistical and
systematic errors in quadrature to give a total error of $\pm2.8$ km~s$^{-1}$\
Mpc$^{-1}$, which amounts to a percentage error of $\sim4$\%.
\subsection{Comparison with Previous $H_0$ Results\label{SEC:H0_COMPARE}}
\begin{deluxetable}{llcc}
\tablewidth{0in}
\tabletypesize{\scriptsize}
\tablecaption{Hubble Constant Comparison\label{tab_h0_compare}}
\tablehead{
\colhead{Reference} &
\colhead{TFR band\tablenotemark{1}} &
\colhead{Clusters\tablenotemark{2,3}} &
\colhead{SNe Ia\tablenotemark{3,4}}
}
\startdata
\citet{Tully:12:78} & I-band & $75.1 \pm 1.0$ & \ldots \\
\citet{Courtois:12:174} & I-band & \ldots & $75.9 \pm 3.8$ \\
This work & I-band & $74.5 \pm 1.6$ & $75.9 \pm 2.5$ \\
\citet{Sorce:13:94} & [3.6]cc & $74 \pm 4$ & \ldots \\
\citet{Sorce:12:L12} & [3.6]cc & \ldots & $75.2 \pm 3.0$ \\
This work & W1lin & $71 \pm 2$ & $73.0 \pm 2.7$ \\
This work & W2lin & $70 \pm 2$ & $72.7 \pm 2.7$ \\
This work & W1cc & $73 \pm 2$ & $75.1 \pm 2.5$ \\
This work & W2cc & $73 \pm 2$ & $75.1 \pm 2.5$ \\
This work & W1cur & $72 \pm 2$ & $73.7 \pm 2.4$ \\
This work & W2cur & $72 \pm 2$ & $73.7 \pm 2.4$ \\
\hline
This work & $<$W12cur,I$>$ & $73 \pm 1$ & $74.4 \pm 2.8$\tablenotemark{5} \\
\enddata
\tablenotetext{1}{\ `lin' indicates linear TFR, `cc' indicates optical - MIR color-corrected photometry, `cur' indicates curved TFR}
\tablenotetext{2}{\ seven clusters with $V_{mod} > 4000$ km~s$^{-1}$}
\tablenotetext{3}{\ km~s$^{-1}$\ Mpc$^{-1}$}
\tablenotetext{4}{\ offsets applied to UNION2 SN Ia sample \citep{Amanullah:10:712}}
\tablenotetext{5}{\ includes statistical and systematic errors}
\end{deluxetable}
Our cluster $H_0$ values compare well with previous determinations for
clusters calibrated with the TFR as shown in column three of
Table~\ref{tab_h0_compare}. Here we are listing statistical errors only.
For the I-band, \citet{Tully:12:78} find $H_0
= 75.1 \pm 1.0$ km~s$^{-1}$\ Mpc$^{-1}$, which agrees to within 1\% of our value
of $74.5 \pm 1.6$ km~s$^{-1}$\ Mpc$^{-1}$. \citet{Sorce:13:94} used the
color-corrected IRAC [3.6] band ([3.6]cc) to derive $H_0 = 73.8 \pm 1.1$
km~s$^{-1}$\ Mpc$^{-1}$, which is well within the statistical errorbars (and also
within 1\%) of our color-corrected W1 (W1cc) value of $73.1 \pm 1.8$ km~s$^{-1}$\
Mpc$^{-1}$. Our pure {\it WISE} curved TFR cluster $H_0$ values are low
although still in statistical agreement with the other values.
Comparing the $H_0$ values derived by bringing the UNION2 sample onto the
TFR distance scale also shows good consistency, as can be seen in column
four of Table~\ref{tab_h0_compare}. The I band value from
\citet{Courtois:12:174} of $H_0 = 75.9 \pm 3.8$ km~s$^{-1}$\ Mpc$^{-1}$ is
identical to our value, although their errorbar includes systematic errors
and so appears larger than ours.
The color-corrected IRAC [3.6] value presented in
\citet{Sorce:12:L12} of $H_0 = 75.2 \pm 3.0$ km~s$^{-1}$\ Mpc$^{-1}$ is less than
two-tenths of a percent different from our value of $75.1 \pm 2.5$ km~s$^{-1}$\
Mpc$^{-1}$ derived from color-corrected W1.
What is new is using the uncorrected curved TFR to derive values of $H_0$.
While these values are low for the seven clusters used to calibrate the
TFR, when used to re-normalize the UNION2 SN Ia sample, the values agree
well with current best estimates of $H_0$ (see below). The other
advantage of using the uncorrected curved TFR for {\it WISE} is that it is
truly independent of the I-band, unlike the color-corrected TFR values,
and thus we can average all three bands (in the logarithm) to derive a
more robust value of $H_0$. This is presented in the last column of the
last row of Table~\ref{tab_h0_compare}, and the error includes both
statistical and systematic errors.
It is interesting to note that \citet{Sakai:00:698} give a value of
$H_0 = 71 \pm 4$ (stat) $\pm 7$ (sys) km~s$^{-1}$\ Mpc$^{-1}$ using a
weighted average of their four band ($BVIH_{-0.5}$) TFRs fit with
linear relations. At the end of their \S5.2.1, they give a value of
$H_0 = 73 \pm 2$ (stat) km~s$^{-1}$\ Mpc$^{-1}$ for a curved I-band TFR.
This is higher than the $H_0$ derived from linear TFRs and closer to
current estimates, including our own.
A current independent estimate for $H_0$ that is useful for comparison is
that presented in \citet{Freedman:12:24}: $H_0 = 74.3 \pm 1.5$(stat) $\pm\
2.1$(sys) km~s$^{-1}$ Mpc$^{-1}$ which has a percentage systematic error of 2.8\%.
All of the SN Ia derived $H_0$ values we present here agree with this value
within the errors. Another value to compare with is that derived from the
{\it PLANCK} mission and presented in \citet{Collaboration:13:5076}: $H_0 =
67.3 \pm 1.2$ km~s$^{-1}$\ Mpc$^{-1}$. Our lowest value of $H_0$ is the one
derived from seven clusters in the uncorrected curved W2 band ($71.6 \pm
1.7$ km~s$^{-1}$\ Mpc$^{-1}$). This value is 3.6$\sigma$ high using their
errorbar and 2.5$\sigma$ high using our errorbar. Our best result of $H_0 =
74.4 \pm 2.8$(stat and sys) km~s$^{-1}$\ Mpc$^{-1}$ is 5.9$\sigma$ high using
their errorbar and 2.5$\sigma$ high using our error estimate. Our data do
not favor such a low value of $H_0$. We can also compare with another CMD
$H_0$ value from \citet{Hinshaw:13:19} who quote $H_0 = 69.32 \pm 0.80$
(stat) km~s$^{-1}$\ Mpc$^{-1}$ in their Table~4. This value is closer to our value
but a tension still exists. Relativistic corrections for foreground lensing
in the CMB analyses may resolve this tension \citep{Clarkson:14:7860}.
\section{Conclusions\label{SEC:CONCLUSIONS}}
We have derived a calibration of the absolute magnitude-linewidth relation
for the {\it WISE} W1 and W2 filters. The raw, linear calibration, using
only {\it WISE} photometry that is aperture corrected, k-corrected, and has
been corrected for internal and external extinction gives:
\begin{subequations}
\begin{align}
\begin{split}
M^{b,i,k,a}_{W1} = &- (20.35 \pm 0.07) \\
&- (9.56 \pm 0.12)(\log W^i_{mx} - 2.5),
\end{split} \\
\begin{split}
M^{b,i,k,a}_{W2} = &- (19.76 \pm 0.08) \\
&- (9.74 \pm 0.12)(\log W^i_{mx} - 2.5).
\end{split}
\end{align}
\end{subequations}
These calibrations show a scatter of 0.54 magnitudes in W1 and 0.56
magnitudes in W2.
The I-band sample grew by 24 galaxies (9\%) compared to the previous
calibration and so we updated it to:
\begin{equation}
\begin{split}
M^{b,i,k,e}_{I} = &- (21.34 \pm 0.07) \\
&- (8.95 \pm 0.14)(\log W^i_{mx} - 2.5).
\end{split}
\end{equation}
This calibration has a scatter of 0.46 magnitudes.
We find evidence for curvature in the MIR TFR based on a comparison between
calibration cluster distances generated using linear TFRs in the I-band and
in the {\it WISE} W1 and W2 bands. We use the ensemble of cluster galaxies
shifted to have an apparent distance of Virgo to fit this curved TFR and
find the following curved TFRs for W1 and W2:
\begin{subequations}
\begin{align}
\begin{split}
\mathcal{M}^{b,i,k,a}_{W1} = &- (20.48 \pm 0.05) \\
&- (8.36 \pm 0.11)(\log W^i_{mx} - 2.5) \\
&+ (3.60 \pm 0.50)(\log W^i_{mx} - 2.5)^2,
\end{split} \\
\begin{split}
\mathcal{M}^{b,i,k,a}_{W2} = &- (19.91 \pm 0.05) \\
&- (8.40 \pm 0.12)(\log W^i_{mx} - 2.5) \\
&+ (4.32 \pm 0.51)(\log W^i_{mx} - 2.5)^2.
\end{split}
\end{align}
\end{subequations}
These calibrations have a scatter of 0.52 mag for W1 and 0.55 mag for W2,
an improvement over the pure linear TFRs. The formal errors on the
zero-point calibration are the smallest of all the calibrations derived
here.
Following previous work on calibrating the TFR in the MIR
\citep{Sorce:13:94}, we apply an optical - MIR color correction to our raw
W1 and W2 magnitudes in order to reduce the scatter. The corrections have
the form:
\begin{subequations}
\begin{align}
\Delta W1^{color} & = -0.470 - 0.561(I^{b,i,k}_{T} -
W1^{b,i,k,a}_{T}), \\
\Delta W2^{color} & = -0.874 - 0.617(I^{b,i,k}_{T} -
W2^{b,i,k,a}_{T}).
\end{align}
\end{subequations}
Where $I^{b,i,k}_T$ values are derived from I-band imaging. These are then
used to adjust the input magnitudes as follows:
\begin{equation}
C_{W1,2} = W1,2^{b,i,k,a}_{T} - \Delta W1,2^{color}.
\end{equation}
We used these pseudo-magnitudes to generate color-corrected linear
calibrations of the form:
\begin{subequations}
\begin{align}
\begin{split}
M_{C_{W1}} = &- (20.22 \pm 0.07) \\
&- (9.12 \pm 0.12)(\log W^i_{mx} - 2.5),
\end{split} \\
\begin{split}
M_{C_{W2}} = &- (19.63 \pm 0.07) \\
&- (9.11 \pm 0.12)(\log W^i_{mx} - 2.5).
\end{split}
\end{align}
\end{subequations}
These both show a scatter of 0.46 magnitudes, identical to the I-band
scatter. These equations represent the most accurate calibration of the
luminosity-linewidth relation available for {\it WISE} data at this time.
We investigate a residual bias in the TFRs resulting from a flat magnitude
cutoff that varies with distance and produces more of a bias as the cutoff
samples the sparser upper end of the luminosity function. We determine two
bias functions. One for the pure {\it WISE} TFRs both curved and linear,
and one for the I-band and the color-corrected {\it WISE} TFRs:
\begin{subequations}
\begin{align}
b_{pure} &= 0.006(\mu - 31)^{2.3} \\
b_{cc} &= 0.004(\mu - 31)^{2.3},
\end{align}
\end{subequations}
where $\mu$ represents the distance modulus of a field galaxy.
From the calibrations we generate bias-corrected distances to the
calibrating clusters and derive a Hubble constant from the clusters far
enough away to be in the Hubble flow (D $> 50$ Mpc). We derive $H_0 = 72.2
\pm 1.7$ km~s$^{-1}$\ Mpc$^{-1}$ for the curved pure W1 TFR and $H_0 = 71.6 \pm
1.7$ km~s$^{-1}$\ Mpc$^{-1}$ for the curved pure W2 TFR. The color-corrected W1
and W2 TFRs give the same value of $H_0 = 73.1 \pm 1.8$ km~s$^{-1}$\ Mpc$^{-1}$,
and we get $H_0 = 74.5 \pm 1.6$ km~s$^{-1}$\ Mpc$^{-1}$ using the I-band TFR.
To leverage the redshift reach of SNe Ia, we measure the zero-point offset of
the UNION2 SN Ia sample by comparing the distances in 56 SN Ia hosts galaxies
in common with the Cosmic Flows 2 sample. The measured offsets give $H_0 =
73.7 \pm 2.4$ using the curved W1 and W2 TFRs and $H_0 = 75.9 \pm 2.5$ using
the I-band linear TFR. Taking the log average of these values gives a Hubble
constant of $H_0 = 74.4 \pm 1.4$ km~s$^{-1}$\ Mpc$^{-1}$. The total systematic error
on our measure of $H_0$ includes the systematic error in the calibration, the
zero-point error, the SN Ia distance error, and a band-to-band systematic
measured using the I-band and W1 and W2, and amounts to $\pm\ 2.4$ km~s$^{-1}$\
Mpc$^{-1}$. Thus our best value is $H_0 = 74.4 \pm1.4$(stat) $\pm\ 2.4$(sys)
km~s$^{-1}$\ Mpc$^{-1}$. Our estimates of $H_0$ do not favor the low values of $H_0$
presented in \citet{Collaboration:13:5076} and \citet{Hinshaw:13:19}, although
relativistic corrections may resolve this tension as suggested in
\citet{Clarkson:14:7860}.
\acknowledgments
We acknowledge useful conversations with the following people: Wendy
Freedman, Barry Madore, Eric Persson, and Andrew Monson.
JDN and MS acknowledge support from the NASA Astrophysical Data Analysis
Program under grant NNX12AE19G for the {\it WISE} Nearby Galaxies Atlas.
HC and JS acknowledge support from the Lyon Institute of Origins under
grant ANR-10-LABX-66 and from CNRS under PICS-06233. RBT acknowledges
support from the US National Science Foundation award AST09-08846. We
acknowledge the use of the HyperLeda database (http://leda.univ-lyon1.fr).
This publication makes use of data products from the {\it Wide-field
Infrared Survey Explorer (WISE)}, which is a joint project of the
University of California, Los Angeles, and the Jet Propulsion
Laboratory/California Institute of Technology, funded by the National
Aeronautics and Space Administration.
This research has made use of the NASA/IPAC Extragalactic Database (NED)
which is operated by the Jet Propulsion Laboratory, California Institute of
Technology, under contract with the National Aeronautics and Space
Administration.
\bibliographystyle{apj}
|
\section{Introduction}
Even though theories of strong interactions have been with us for about forty years there are still many outstanding questions that need to be answered. For instance we still lack a complete understanding of such phenomena as confinement, chiral symmetry breaking and infrared fixed points with associated conformal behavior. Solving strongly interacting theories would not only yield a complete understanding of Quantum Chromo Dynamics (QCD) but would also provide invaluable assistance to the search for possible physics beyond the Standard Model. For a review on these aspects see for instance \cite{Sannino:2009za,Hill:2002ap}.
Strong interactions also have the potential to exhibit another interesting dynamical phenomena as first noted by Dashen \cite{Dashen}. It is quite amusing to note that even before the discovery of QCD he had observed that under special circumstances the combination of charge conjugation and parity (CP) could potentially be spontaneously broken \cite{Dashen}. It was discovered that by appropriate choices of the masses of the quarks the vacuum could align in a direction in which CP was broken. Specifically for three quark flavors if the strange quark mass become sufficiently negative the vacuum will pick up nontrivial complex phases and CP will be broken. A CP violating phase can also materialize for only two quark flavors as emphasized by M. Creutz and collaborators \cite{Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs}. Here it is found that once the mass of the down quark become equal in magnitude but opposite in sign to the mass of the up quark a vacuum degeneracy appears signaling that the Nambu-Goldstone bosons have become massless. At this point it has been suggested that the neutral Nambu-Goldstone boson (the neutral pion), being odd under CP, might condense and spontaneously break CP \cite{Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs}. For both two and three quark flavors the spontaneous CP violation is a consequence of the alignment of the vacuum by an (external) perturbation, i.e. the mass term.
It is only natural to ask whether this persists if we slowly change the various parameters of QCD. First, since vector-like gauge theories with a set of massless fermions in an arbitrary complex representation of the gauge group have the same global chiral symmetry all such systems must share the phenomena of spontaneous CP violation in a manner identical to QCD. Second, it was recently shown that in the four flavor case there is a region similar to the three flavor case in which CP is violated \cite{Ryttov:2014kua}. In fact for any number of flavors $N_f$ one can show that for appropriately tuned external mass perturbations CP will break spontaneously \cite{Ryttov:2014kua}.
In this work we intend to change yet another set of parameters and search for CP violation. Instead of assuming that the fermions belong to a complex representation of the gauge group we shall turn our attention to two Dirac flavors in a \emph{real} and \emph{pseudoreal} representation. This has the immediate effect of enhancing the global symmetry of the theory and therefore it could potentially lead to a different dynamical behavior. Loosely speaking we will find that once we add a mass perturbation for the two Dirac fermions and drive the vacuum in different directions the system always has the freedom (due to the \emph{enhanced} global symmetry) to choose a vacuum in which CP is preserved. The behavior is quite novel since the system makes a discontinuous jump from one vacuum to another as we dial the masses from positive/negative to negative/positive values. In other words if we try to force the system into a CP violating phase it responds by jumping into a vacuum in which CP is preserved and hence where a fewer number of symmetries are broken.
Examples of theories that fall under our general considerations are $SU(2)$ gauge theories with $N_f=2$ massless fermions in either the fundamental (pseudoreal) representation or the three dimensional adjoint (real) representation. Besides being playgrounds for studying modifications of QCD these theories are also the basis for a great number of attempts to build realistic technicolor models able to explain electroweak symmetry breaking in a natural way \cite{Weinberg:1975gm,Weinberg:1979bn,Sannino:2004qp,Dietrich:2005jn,Dietrich:2005wk,Foadi:2007ue,Ryttov:2008xe,Belyaev:2008yj,Dietrich:2009af,Dietrich:2009ix,Dietrich:2008ni,Gudnason:2006mk}. This is yet another strong motivation for our work.
Situations where the alignment of the vacuum is produced by other perturbations than mass perturbations have also been investigated in the context of technicolor theories. Specifically the question as to how the gauging of the electroweak symmetry, which explicitly breaks the global symmetry, aligns the vacuum \cite{Preskill:1980mz}. In addition some years ago it was suggested that the observed CP violation in the Standard Model could be accounted for in (extended) technicolor theories \cite{Eichten:1980du,Lane:2002wv,Martin:2004ec,Lane:2005we,Lane:2005vp,Lane:2000es,Rador:2009sy}. Here quarks and techniquarks are coupled via four fermion interactions and in this way alignment in the techniquark sector would determine the alignment in the quark sector.
The paper is organized as follows: In Section \ref{sec:VA} we discuss vacuum alignment in general terms. We also discuss the discrete symmetries C and P for fermions in a real or pseudoreal representation of the gauge group. In Section \ref{sec:real} and \ref{sec:pseudoreal} we present our results for the theory with two Dirac fermions in a real representation and pseudoreal of the gauge group. We then conclude in Section \ref{sec:conclusion}.
\section{Symmetries and Vacuum Alignment}\label{sec:VA}
As our starting point we consider vector-like gauge theories with a set of massless fermions belonging to some irreducible representation of the gauge group. The specific situation where the gauge group is $SU(3)$ and the representation is the fundamental representation corresponds to massless QCD with some number of flavors.
In general such gauge theories with a fermionic matter sector have a global continuous symmetry that depends on the representation to which the fermions belong. If the representation is complex the global symmetry is $G=SU(N_f)_L \times SU(N_f)_R\times U(1)_B$ and if the representation is real or pseudoreal the global symmetry is enhanced to $G=SU(2N_f)$ \cite{Peskin:1980gc}. Here $N_f$ counts the number of Dirac fermions $\Psi_D = (\psi_L, \psi_R)^T$ where $\psi_L$ and $\psi_R$ are respectively a left-handed and right-handed Weyl fermion. The abelian $U(1)_B$ is the standard baryon number and is anomaly free.
The reason for the appearance of the enhanced symmetry when the representation is real or pseudoreal is due to the fact that one can always write a right-handed Weyl fermion in a given representation as a left-handed Weyl fermion in the conjugated representation. Then if the representation is real or pseudoreal the conjugated representation is equivalent to the representation itself. Hence if we have $N_f$ Dirac fermions we can consider them as $2N_f$ left-handed Weyl fermions $\lambda_L^f,\ f=1,\ldots, 2N_f$ all in the same real or pseudoreal representation. From now on we shall omit the subscript $L$ and always take the Weyl fermions $\lambda^f$ to be left-handed.
Gauge theories may also posses the discrete symmetries charge conjugation C, parity P and time reversal T or any combination hereof. Charge conjugation exchanges a particle with its antiparticle or vice versa while parity changes the chirality of a particle. For Dirac fermions in a complex representation of the gauge group these discrete symmetries are implemented in standard fashion and can be found for instance in the appendix of \cite{Ryttov:2014kua}.
For a set of fermions $\lambda^f$ belonging to a real or pseudoreal representation of the gauge group one might first expect them to be Majorana, i.e. that they are equal to their own antiparticles. This follows from the fact that charge conjugation involves a complex conjugation of the fermion fields and the reality or pseudoreality of the fermion representation. However they also transform as a fundamental under global $SU(2N_f)$ and for $N_f>1$ this is a complex representation. Hence under charge conjugation, in which complex conjugation occurs, these fermions cannot be transformed into them selves. Instead under charge conjugation and parity we must consider them in the Dirac basis in which the implementation of all the discrete symmetries are known.
Next consider theories that are asymptotically free and assume that as the energy scale is lowered the gauge interactions become strong such that a bilinear condensate is formed spontaneously breaking the continuous global symmetry $G$ to some subgroup $G'$. Even though the Hamiltonian $\mathcal{H}$ of the system is invariant under all transformations of $G$ the vacuum is only invariant under transformations of $G'$. There is a degeneracy of vacua all parameterized by the elements that belong to $G$ but not to $G'$. As is well known this degeneracy is manifested in the appearance of a set of Nambu-Goldstone bosons.
If we add a small perturbation $\mathcal{H}'$ to the system explicitly breaking the global symmetry $G$ the vacuum degeneracy will be lifted and the vacuum will become \emph{aligned} with the explicit symmetry breaking perturbation. In order to find the aligned (correct) vacuum that is left over we must minimize the total energy $\langle \Omega |g^{\dagger} \left( \mathcal{H} + \mathcal{H}' \right) g| \Omega \rangle $ over all vacua $g|\Omega \rangle$ where $g\in G/G'$ and $| \Omega \rangle$ is some standard reference vacuum. It should be clear that the contribution $\langle \Omega | g^{\dagger} \mathcal{H} g | \Omega \rangle = \langle \Omega | \mathcal{H} | \Omega \rangle$ is just a constant since the unperturbed Hamiltonian is symmetric under all transformations. Hence we can simplify the procedure of finding the aligned vacuum by considering the minimization of only the perturbation
\begin{eqnarray}
E &=& \langle \Omega | g^{\dagger} \mathcal{H}' g | \Omega \rangle
\end{eqnarray}
It should be noted that we can view the minimization in two different ways: Either we find the rotated vacuum $g | \Omega \rangle$ that minimizes the energy or we find the rotated Hamiltonian $g^{\dagger} \mathcal{H}' g$ that is compatible with the reference vacuum $| \Omega \rangle$. The two views are equivalent.
\section{Real Representations}\label{sec:real}
As one of our main examples we will consider two massless Dirac fermions denoted as $(U_L,U_R)^T$ (up) and $(D_L,D_R)^T$ (down) transforming according to a real representation of the gauge group. From the above discussion this is equivalent to four Weyl fermions all transforming according to the same real representation
\begin{eqnarray}\label{eq:lambda}
\lambda &=& \left(
\begin{array}{c}
U_L \\
D_L \\
- i \sigma^2 U_R^* \\
- i \sigma^2 D_R^*
\end{array}
\right)
\end{eqnarray}
The Weyl fermions $\lambda^f,\ f=1,\dots,4$ are all treated as left-handed fields. We have also suppressed the gauge index. The theory enjoys an $SU(4)$ global symmetry in which the four Weyl fields are rotated among each other. Assume now that the gauge interactions become strong such that the following condensate is formed \cite{Holdom:1984sk,Yamawaki:1985zg,Appelquist:1986an,Appelquist:1986tr,Appelquist:1987fc,Appelquist:1988yc,Appelquist:1997dc,Brodsky:2008be,Ryttov:2007sr,Ryttov:2007cx,Ryttov:2009yw,Ryttov:2010hs,Ryttov:2010iz,Ryttov:2012qu,Ryttov:2013hka,Ryttov:2013ura,Mojaza:2012zd,Pica:2010mt,Pica:2010xq}
\begin{eqnarray}
\langle \Omega | \lambda^f \lambda^{f'} | \Omega \rangle &=& - \frac{1}{2} \Delta E^{ff'} \ , \qquad E=
\left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right)
\end{eqnarray}
where for brevity we have defined $\lambda \lambda = - \lambda^T i \sigma^2 \lambda$. In terms of the up and down fermion fields we can also write the condensate in the standard form $\langle \Omega | \lambda^f \lambda^{f'} | \Omega \rangle E_{ff'}= 2 \langle \Omega | \overline{U}_R U_L + \overline{D}_R D_L | \Omega \rangle $ by contracting the flavor indices with $E_{ff'}$. This justifies our choice of the specific reference vacuum $E$ since here the condensate is particularly simple. Also the notation is such that $\overline{U}_R = U_R^*$ and $\overline{D}_R=D_R^*$.
Since the condensate is symmetric in its indices it is invariant only under $SO(4)$ transformations and therefore spontaneously breaks the $SU(4)$ global symmetry. This spontaneous symmetry breaking leads to the well known appearance of $15-6=9$ Nambu-Goldstone bosons.
Let us try to force the vacuum in a different direction by adding a small mass term. Such a perturbation explicitly breaks the $SU(4)$ global symmetry, lifts the vacuum degeneracy and provides the Nambu-Goldstone bosons with a mass. To be concrete assume that a Dirac mass term is added to the theory
\begin{eqnarray}
\mathcal{H}' &=& \frac{1}{2} \lambda M^{\dagger} \lambda + \text{h.c.} \ , \qquad \qquad M = M^{\dagger} =
\left(
\begin{array}{cccc}
0 & 0 & m_U & 0 \\
0 & 0 & 0 & m_D \\
m_U & 0 & 0 & 0 \\
0 & m_D & 0 & 0
\end{array}
\right)
\end{eqnarray}
Here the Dirac mass matrix is symmetric and each entry $m_U$ and $m_D$ are taken to be real such that it preserves CP. In order to find the correct aligned vacuum we need to follow the above steps of minimizing the energy
\begin{eqnarray}
E(V) &=& \langle \Omega | \frac{1}{2}(g \lambda ) M (g\lambda) | \Omega \rangle + \text{h.c.} = - \frac{\Delta}{4} \text{Tr} \left[ M^{\dagger} V+MV^{\dagger} \right]
\end{eqnarray}
where the rotated vacuum is $V=gEg^T$ and $g\in G/G'$. We must find the rotated (aligned) vacuum $V$ that minimizes the energy. To this end we first note that $V$ must be of the same form as the Dirac mass matrix since the energy involves a trace. Second we note that the vacuum must be symmetric since the reference vacuum $E$ is symmetric. In addition it must be unitary and can therefore only depend on two phases $e^{i\theta_U}$ and $e^{i \theta_D}$. Lastly since the vacuum must have unit determinant the two angles must up to a multiple of $\pi$. Putting it all together we find that the aligned vacuum must be of the form
\begin{eqnarray}
V &=&
\left(
\begin{array}{cccc}
0 & 0 & e^{i \theta_U} & 0 \\
0 & 0 & 0 & e^{i \theta_D} \\
e^{i \theta_U} & 0 & 0 & 0 \\
0 & e^{i\theta_D} & 0 & 0
\end{array}
\right) \ , \qquad \theta_U + \theta_D = \pi n
\end{eqnarray}
where $n$ is some integer number. It is important to observe that the angles should add up to \emph{any} multiple of $\pi$. This is in contrast to the case where the fermions belong to a complex representation of the gauge group \cite{Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs,Ryttov:2014kua}. Here the vacuum angles must add up to an \emph{even} multiple of $\pi$. As we shall see below this has a major impact on the behavior of the system as we vary the fermion masses aligning the vacuum in different directions. This differece is a direct consequence of the enhanced global symmetry.
It should be clear that without loss of generality we can take $n=0$ or $n=1$. In this way we can speak about two different branches of the system in which the energy is
\begin{eqnarray}
E(V) &=& \left\{
\begin{array}{l}
-\Delta \left( m_U + m_D \right) \cos \theta_U \ , \qquad \theta_D = - \theta_U \ , \qquad \quad n=0 \\
- \Delta \left( m_U - m_D \right) \cos \theta_U \ , \qquad \theta_D = \pi - \theta_U \ , \qquad n=1
\end{array}
\right.
\end{eqnarray}
We want to minimize the energy over the vacuum angle $\theta_U$. The location of the minima depends on the sign of the up and down fermion masses. If both fermion masses have the same sign the minima are located on the $n=0$ branch. If they are both positive they are found at $\theta_U$ being equal to an even multiple of $\pi$ and if they are both negative they are found at $\theta_U$ being equal to an odd multiple of $\pi$. On the other hand the fermion masses might also differ in sign. In that case the minima are located on the $n=1$ branch. If the up fermion mass is positive and the down fermion mass is negative then the minima are found at $\theta_U$ being equal to an even multiple of $\pi$ while in the opposite case where the up fermion mass is negative and the down fermion mass is positive the minima are found at $\theta_U$ being equal to an odd multiple of $\pi$. Summarizing we therefore find that the aligned vacuum is
\begin{eqnarray}
V_{++} &=& \left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right) \ , \qquad \qquad \quad E(V_{++}) = - \Delta \left( | m_U| + |m_D | \right) \ , \qquad m_U,m_D > 0 \\
V_{--} &=&
\left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -1 \\
-1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right) \ , \qquad \ E(V_{--}) = - \Delta \left( | m_U| + |m_D | \right) \ , \qquad m_U,m_D < 0 \\
V_{+-} &=& \left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & -1\\
1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right) \ , \qquad \qquad E(V_{+-}) = - \Delta \left( |m_U|+|m_D| \right) \ , \qquad m_U>0,m_D<0 \\
V_{-+} &=& \left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & 1 \\
-1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right) \ , \qquad \qquad E(V_{--}) = - \Delta \left( |m_U|+|m_D| \right) \ , \qquad m_U<0,m_D>0
\end{eqnarray}
We are now in a position to study the response of the system to the external mass perturbation. First imagine that both up and down explicit fermion mass terms $m_U$ and $m_D$ are positive. Then the system must be found to reside in the single unique vacuum $V_{++}$ which is on the $n=0$ branch. This is for arbitrary positive values $m_U,m_D>0$. Let us now fix $m_U$ and slowly dial the down fermion mass $m_D$ towards zero. As the down fermion mass passes through the critical value $m_D=0$ the system becomes discretely degenerate and a new additional vacuum $V_{+-}$ appears on the $n=1$ branch. If we continue to decrease the down fermion mass to negative values the discrete degeneracy disappears and the vacuum of the system uniquely becomes $V_{+-}$. Nothing is then expected to further happen for any negative value $m_D<0$.
What has happened is quite novel since the system must have gone through a discontinuous jump at the critical value where the down fermion mass vanishes. As we cross the critical value the system must respond by jumping from the $n=0$ branch to the $n=1$ branch. It should be clear that the above discontinuous behavior occurs for an arbitrary magnitude and sign of $m_U$ and $m_D$. It does not matter whether $m_U$ and $m_D$ initially are positive or negative. If we fix one of them and dial the other then as the second mass passes through zero the system must perform a discontinuous jump from the $n=0$ ($n=1$) branch to the $n=1$ ($n=0$) branch. In Fig. \ref{RealPlot} we plot the vacuum structure in terms of the up and down explicit fermion mass.
\begin{figure}[bt]
\centering
\includegraphics[width=0.3\columnwidth]{RealPlot}
\caption{A plot of the vacuum structure (with $V_{++}=-V_{--}$ and $V_{+-} = -V_{-+}$) as function of the up and down mass $m_U$ and $m_D$ with two Dirac flavors in a real representation of the gauge group. When crossing from one vacuum into another the system makes a discontinuous jump. Blue corresponds to the $n=0$ branch while red corresponds to the $n=1$ branch. }
\label{RealPlot}
\end{figure}
Irrespective of which of the four vacua the system resides in the continuous global symmetry $SU(4)$ is broken to $SO(4)$.\footnote{The fact that these four vacua preserve an $SO(4)$ symmetry can be checked by using the generators found \cite{Foadi:2007ue}. } The embedding (or direction) of the $SO(4)$ subgroup of course depends on the specific vacuum. Note also that in all four vacua there are no nontrivial complex phases. In this way discrete CP symmetry is preserved for arbitrary mass perturbations, i.e. we cannot drive the vacuum into a direction in which CP is spontaneously broken.
Let us contrast this to the case where the fermions are in a complex representation of the gauge group and for which the continuous global symmetry is $SU(N_f)_L \times SU(N_f)_R \times U(1)_B$. We will discuss the two flavor case since this is the situation that mimics our example above \cite{Dashen,Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs}. However as we shall see there are crucial differences.
Consider therefore two Dirac flavors, up and down, in a complex representation of the gauge group and take their mass perturbations to be positive. The vacuum is then aligned in the direction $\text{diag}(1,1)$. But if we fix the up fermion mass and slowly dial the down fermion mass towards zero no new vacua appear \cite{Dashen,Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs}. One might naively have expected that the system would choose the vacuum $\text{diag}(1,-1)$ but this does not have unit determinant so it is not a possibility. Instead the vacuum continues to be $\text{diag}(1,1)$ until the critical point where the up and down fermion masses are equal in magnitude but opposite in sign is encountered. The vacuum is here forced to take complex values $\text{diag}(e^{i\theta},e^{-i\theta})$ where $\theta$ is some arbitrary angle. Again it cannot be driven in the direction $\text{diag}(1,-1)$ since it does not have unit determinant. Instead a continuous vacuum degeneracy appears indicating that the Nambu-Goldstone bosons have once again become massless. At this critical point it has been suggested that the neutral Nambu-Goldstone boson (the neutral pion), being odd under CP, develops a VEV spontaneously breaking CP \cite{Dashen,Creutz:1995wf,Creutz:2003xu,Creutz:2003xc,Creutz:2010ts,Creutz:2013xfa,Aoki:2014moa,Creutz:2000bs}. This is the Dashen phase. For completeness we note that as the down fermion mass passes the critical point and decreases even further the vacuum is aligned in the direction $\text{diag}(-1,-1)$ and the vacuum degeneracy disappears.
This differs in essential aspects from the case where the fermions belong to a real representation of the gauge group and the system has an enhanced global symmetry. Here we find no evidence for the existence of a Dashen type phase. Technically it is due to the fact that the system has two different branches $n=0$ and $n=1$ and that as we dial the fermion masses the system can undergo discontinuous jumps from one branch to another. If we for the moment imagine that the system did not have the $n=1$ branch then the behavior would be just as in the case where the fermion representation is complex. As we pass the down fermion mass through zero nothing is expected to happen. Then again at the point where the down and up fermion masses are equal in magnitude but opposite in sign a vacuum degeneracy appears since the energy of the system vanishes identically for any value of the vacuum angles. Here the Nambu-Goldstone bosons become massless and via a mechanism similar to the case for complex representations where the neutral one develops a VEV could break CP spontaneously. However this is not observed.
\section{Pseudoreal Representations}\label{sec:pseudoreal}
In this section we complete our study by considering a similar setup as above but with fermions belonging to a pseudoreal representation of the gauge group. We again take two Dirac flavors and arrange them into a set of left-handed transforming fermions $\lambda^f,\ f=1,\ldots,4$. The theory enjoys a continuous global symmetry $G=SU(4)$ and a set of discrete symmetries. As above we imagine that at some energy scale the gauge interactions become strong and the following condensate is formed
\begin{eqnarray}
\langle \Omega | \lambda^f \lambda^{f'} | \Omega \rangle &=& - \frac{1}{2} \Delta E^{ff'} \ , \qquad E=
\left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
-1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right)
\end{eqnarray}
Since the condensate is antisymmetric in its indices it leaves invariant only a $G'=Sp(4)$ subgroup. Hence it breaks the continuous global symmetry spontaneously and is the source for the appearance of $15-10=5$ Nambu-Goldstone bosons. Note again that if we write $\lambda^f$ in terms of the two Dirac fermions as in Eq. \ref{eq:lambda} the condensate can be written as $\langle \Omega | \lambda^f \lambda^{f'} | \Omega \rangle E_{ff'}= 2 \langle \Omega | \overline{U}_R U_L + \overline{D}_R D_L | \Omega \rangle$ justifying our choice of the reference vacuum $E$ since the condensate is here particularly simple. Adding to the theory the following explicit Dirac mass term
\begin{eqnarray}
\mathcal{H}' &=&\frac{1}{2} \lambda M^{\dagger} \lambda + \text{h.c.} \ , \qquad M=-M^{\dagger} =
\left(
\begin{array}{cccc}
0 & 0 & m_U & 0 \\
0 & 0 & 0 & m_D \\
-m_U & 0 & 0 & 0 \\
0 & -m_D & 0 & 0
\end{array}
\right)
\end{eqnarray}
we want to find the rotated vacuum $V=g E g^T,\ g \in G/G'$ that minimizes the energy
\begin{eqnarray}
E(V) &=& \langle \Omega | \frac{1}{2} (g \lambda ) M^{\dagger} (g \lambda) | \Omega \rangle + \text{h.c.} = - \frac{\Delta}{4} \text{Tr} \left[ M^{\dagger} V + M V^{\dagger} \right]
\end{eqnarray}
Both the up and down fermion mass terms $m_U$ and $m_D$ are taken to be real. Via a similar line of arguments as in the case with a real fermion representation we find that the rotated vacuum depends on two angles $\theta_U$ and $\theta_D$ that adds to a multiple of $\pi$ and be of the form
\begin{eqnarray}
V &=& \left(
\begin{array}{cccc}
0 & 0 & e^{i \theta_U} & 0 \\
0 & 0 & 0 & e^{i \theta_D} \\
- e^{i\theta_U} & 0 & 0 & 0 \\
0 & - e^{i \theta_D} & 0 & 0
\end{array}
\right) \ , \qquad \theta_U + \theta_D = \pi n
\end{eqnarray}
Again there are two different branches corresponding to $n=0$ and $n=1$. All other branches are equivalent to one of these. Hence we want to minimize
\begin{eqnarray}
E(V) &=& \left\{
\begin{array}{l}
- \Delta \left( m_U + m_D \right) \cos \theta_U \ , \qquad \theta_D = -\theta_U \ , \qquad\quad n=0 \\
- \Delta \left( m_U - m_D \right) \cos \theta_U \ , \qquad \theta_D = \pi -\theta_U \ , \qquad n=1
\end{array}
\right.
\end{eqnarray}
over the vacuum angle $\theta_U$. Depending on the sign of the up and down fermion masses the minima are located at the same values of the vacuum angles as in the case with a real fermion representation
\begin{eqnarray}
V_{++} &=& \left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1\\
-1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right) \ , \qquad E(V_{++})= - \Delta \left( |m_U|+|m_D| \right)\ , \qquad m_U,m_D>0 \\
V_{--} &=& \left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -1\\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right) \ , \qquad E(V_{--})= - \Delta \left( |m_U|+|m_D| \right)\ , \qquad m_U,m_D<0 \\
V_{+-} &=& \left(
\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & -1\\
-1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right) \ , \qquad E(V_{+-})= - \Delta \left( |m_U|+|m_D| \right)\ , \qquad m_U>0,m_D<0 \\
V_{-+} &=& \left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & 1\\
1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right) \ , \qquad E(V_{-+})= - \Delta \left( |m_U|+|m_D| \right)\ , \qquad m_U<0,m_D>0
\end{eqnarray}
The behavior is identical to the case of a real fermion representation. If we fix the mass of one of the fermions and dial the other one then at the critical point where it vanishes the system makes a discontinuous jump from one branch to the other. There is no critical point where all the Nambu-Goldstone bosons become massless with a vacuum degeneracy reappearing. Hence it is not possible with a Dashen type phase where CP is spontaneously broken when the fermions belong to a pseudoreal representation of the gauge group.
\section{Conclusion}\label{sec:conclusion}
We investigated the vacuum structure of fermionic gauge theories with two Dirac favors by adding a small explicit mass perturbation. We picked the fermions to transform according to a real or pseudoreal representation of the gauge group. By dialing the explicit masses we studied how the vacuum aligned in different directions. We were motivated by the fact that when the fermions are in a complex representation of the gauge group the system has the potential to develop a so called Dashen phase where CP is spontaneously broken. We found that this is not an option if one considers theories with enhanced global symmetries. In some sense the global symmetry is so large that if one attempts to force the system in a CP violating direction it has the freedom to align in a direction in which CP is preserved. In other words the system prefers to break the least number of symmetries.
\acknowledgments
The author would like to thank F. Sannino for discussions and careful reading of the manuscript. The author would also like to thank the CERN Theory Division for kind hospitality while this work was completed. The CP$^3$-Origins centre is partially funded by the Danish National Research Foundation, grant number DNRF90.
|
\section{Introduction.}
Last years an increasing number of papers was devoted to the investigations of operators related to
a conjugation in a Hilbert space, see, e.g. ~\cite{cit_100_GP}, \cite{cit_200_GP2}, \cite{cit_300_Z}, \cite{cit_400_LZ} and references therein.
A conjugation $J$ in a Hilbert space $H$ is an {\it antilinear} operator on $H$ such that $J^2 x = x$, $x\in H$,
and $(Jx,Jy)_H = (y,x)_H$, $x,y\in H$.
The conjugation $J$ generates the following bilinear form:
$$ [x,y]_J := (x,Jy)_H,\qquad x,y\in H. $$
A linear operator $A$ in $H$ is said to be $J$-symmetric ($J$-skew-symmetric) if
\begin{equation}
\label{f1_2}
[Ax,y]_J = [x,Ay]_J,\qquad x,y\in D(A),
\end{equation}
or, respectively,
\begin{equation}
\label{f1_3}
[Ax,y]_J = -[x,Ay]_J,\qquad x,y\in D(A).
\end{equation}
A linear operator $A$ in $H$ is said to be $J$-isometric if
\begin{equation}
\label{f1_3_1}
[Ax,Ay]_J = [x,y]_J,\qquad x,y\in D(A).
\end{equation}
If $\overline{D(A)} = H$, then conditions~(\ref{f1_2}), (\ref{f1_3}) and (\ref{f1_3_1}) are equivalent to the following conditions:
\begin{equation}
\label{f1_3_2}
JAJ \subseteq A^*,
\end{equation}
\begin{equation}
\label{f1_3_3}
JAJ \subseteq -A^*,
\end{equation}
and
\begin{equation}
\label{f1_3_4}
JA^{-1}J \subseteq A^*,
\end{equation}
respectively.
A linear operator $A$ in $H$ is called $J$-self-adjoint ($J$-skew-self-adjoint, or $J$-unitary) if
\begin{equation}
\label{f1_4}
JAJ = A^*,
\end{equation}
\begin{equation}
\label{f1_5}
JAJ = -A^*,
\end{equation}
or
\begin{equation}
\label{f1_6}
JA^{-1}J = A^*,
\end{equation}
respectively.
We shall prove that each densely defined $J$-skew-symmetric operator
(each $J$-isometric operator with $\overline{D(A)}=\overline{R(A)}=H$) in a Hilbert space $H$ has a
$J$-skew-self-adjoint (respectively $J$-unitary) extension in a Hilbert space $\widetilde H\supseteq H$.
We shall follow the ideas of Galindo in~\cite{cit_450_G} with necessary modifications.
In particular, Lemma in~\cite{cit_450_G} can not be applied in our case, since its assumptions can never be satisfied with
$T$: $T^2 = I$. In fact, in this case $T$ would be a conjugation in $H$. Choosing an element $f\in H$ of an orthonormal basis
in $H$ which corresponds to $T$,
we would get $(f,Tf)= (f,f) = 1\not= 0$. Moreover, an exit out of the original space can appear in our case.
\noindent
We notice that under stronger assumptions on a $J$-skew-symmetric operator the existence of
a $J$-skew-self-adjoint extension was proved by Kalinina in~\cite{cit_550_K}.
\noindent
\textbf{Notations.}
As usual, we denote by $\mathbb{R}, \mathbb{C}, \mathbb{N}, \mathbb{Z}, \mathbb{Z}_+$,
the sets of real numbers, complex numbers, positive integers, integers and non-negative integers,
respectively. Set $\overline{0,d} = \{ 0,1,...,d \}$, if $d\in \mathbb{N}$; $\overline{0,\infty} = \mathbb{Z}_+$.
If H is a Hilbert space then $(\cdot,\cdot)_H$ and $\| \cdot \|_H$ mean
the scalar product and the norm in $H$, respectively.
Indices may be omitted in obvious cases.
For a linear operator $A$ in $H$, we denote by $D(A)$
its domain, by $R(A)$ its range, and $A^*$ means the adjoint operator
if it exists. If $A$ is invertible then $A^{-1}$ means its
inverse.
For a set $M\subseteq H$
we denote by $\overline{M}$ the closure of $M$ in the norm of $H$.
By $\mathop{\rm Lin}\nolimits M$ we denote
the set of all linear combinations of elements of $M$,
and $\mathop{\rm span}\nolimits M:= \overline{ \mathop{\rm Lin}\nolimits M }$.
By $E_H$ we denote the identity operator in $H$, i.e. $E_H x = x$,
$x\in H$. In obvious cases we may omit the index $H$.
All appearing Hilbert spaces are assumed to be separable.
\section{Extensions of $J$-skew-symmetric and $J$-isometric operators.}
We shall make use of the following lemma.
\begin{lem}
\label{l2_1}
Let $H$ be a Hilbert space with a positive even or infinite dimension, and $J$ be a conjugation on $H$.
Then there exists a subspace $M$ in $H$ such that
$$ M \oplus JM = H. $$
\end{lem}
\textbf{Proof. }
Let $\{ f_n \}_{n=0}^{2d+1}$ be an orthonormal basis in $H$ corresponding to $J$, i.e. such that $Jf_n = f_n$, $0\leq n\leq d$;
$d\in Z_+\cup\{ +\infty \}$ ($2d+2=\dim H$).
Set
$$ f_{2k,2k+1}^+ = \frac{1}{\sqrt{2}}( f_{2k} + if_{2k+1} ),\quad f_{2k,2k+1}^- = \frac{1}{\sqrt{2}}( f_{2k} - if_{2k+1} ),\qquad
k\in \overline{0,d}. $$
It is easy to see that $\{ f_{2k,2k+1}^+, f_{2k,2k+1}^- \}_{k=0}^d$ is an orthonormal basis in $H$.
Set $M := \mathop{\rm span}\nolimits \{ f_{2k,2k+1}^+ \}_{k=0}^d$. It remains to notice that
$JM = \mathop{\rm span}\nolimits \{ f_{2k,2k+1}^- \}_{k=0}^d$.
$\Box$
\begin{thm}
\label{t2_1}
Let $H$ be a Hilbert space and $J$ be a conjugation on $H$. Let $A$ be a $J$-skew-symmetric ($J$-isometric) operator
in $H$. Suppose that $\overline{D(A)}=H$ (respectively $\overline{D(A)}=\overline{R(A)}=H$).
Then there exists a $J$-skew-self-adjoint (respectively $J$-unitary) extension of $A$ in a Hilbert space $\widetilde H\supseteq H$
(with an extension of $J$ to a conjugation on $\widetilde H$).
\end{thm}
\textbf{Proof.} Let $A$ be such an operator as that in the statement of the theorem.
The operator $A$ admits the closure which is $J$-skew-symmetric (respectively $J$-isometric) (see, e.g.~\cite[p. 18]{cit_300_Z}).
Thus, without loss of generality we shall assume that $A$ is closed.
In what follows, in the case of a $J$-skew-symmetric ($J$-isometric) $A$, we shall say about
case~(a) (respectively case~(b)).
Set $H_2 = H\oplus H$, and consider the following transformations on $H_2$:
$$ J_2 \{ x, y \} = \{ Jx, Jy \},\ V \{ x, y \} = \{ y, -x \},\ U \{ x, y \} = \{ y, x \},\quad \forall \{ x,y \}\in H_2, $$
and $R := UJ_2 = J_2U$, $K := VR$. Observe that $R$ and $K$ are conjugations on $H_2$.
The graph of an arbitrary linear operator $C$ in the Hilbert space $H$ will be denoted by $G_C$ ($\subseteq H_2$).
Observe that
\begin{equation}
\label{f2_5}
J_2 G_C = G_{JCJ},\quad RG_C = UG_{JCJ}.
\end{equation}
If $\overline{D(C)}=H$, then
\begin{equation}
\label{f2_6}
G_{C^*} = H_2\ominus VG_C.
\end{equation}
In the case~(a) we may write:
$$ (\{ x,Ax \}, \{ JAJ y,y \}) = (x,JAJy) + (Ax,y) = 0,\quad \forall x\in D(A), y\in D(JAJ). $$
Then
\begin{equation}
\label{f2_7}
G_A \perp RG_A.
\end{equation}
In the case~(b), we have
$$ (\{ x,Ax \}, \{ JA^{-1}J y,-y \}) = 0,\quad \forall x\in D(A), y\in D(JA^{-1}J), $$
and therefore
\begin{equation}
\label{f2_9}
G_A \perp KG_A.
\end{equation}
Set $D = \left\{ \begin{array}{cc} H_2\ominus [G_A\oplus RG_A] & \mbox{in the case (a)} \\
H_2\ominus [G_A\oplus KG_A] & \mbox{in the case (b)}\end{array}\right.$.
If $D=\{0\}$ then it means that $A$ is $J$-skew-self-adjoint (respectively $J$-unitary), see considerations for the operator $B$ below.
In the opposite case, we have $RD=D$ (respectively $KD=D$).
At first, suppose that $D$ has a positive even or infinite dimension.
By~Lemma~\ref{l2_1} we obtain that there exists a subspace $X\subseteq D$ such that
$X\oplus RX = D$ (respectively $X\oplus KX = D$).
Since each element of $X$ is orthogonal to $RG_A = VG_{-JAJ}$ ($KG_A = VG_{JA^{-1}J}$), by~(\ref{f2_6}) it follows that
\begin{equation}
\label{f2_11}
X\subseteq G_{-JA^*J} \quad (\mbox{respectively }X\subseteq G_{J(A^{-1})^*J}).
\end{equation}
Set $G' = G_A\oplus X$. Suppose that $\{ 0,y \}\in G'$. Then there exist $\{ x,Ax \}\in G_A$ such that $\{0,y\} -
\{ x,Ax \} = \{ -x,y-Ax\}\in X$. By~(\ref{f2_11}) we get
$y-Ax = JA^*Jx$ (respectively $y-Ax = -J(A^{-1})^*Jx$), and therefore $y=0$.
Thus, $G'$ is a graph $G_B$ of a densely defined linear operator $B$. Moreover, we have
$$ G_B\oplus RG_B = H_2 \quad (\mbox{respectively } G_B\oplus KG_B = H_2). $$
In the case~(a) we get
$$ UG_B\oplus URG_B = H_2; $$
$$ G_{ (-B)^*} = H_2 \ominus VG_{-B} = H_2 \ominus UG_B = URG_B = J_2 G_B = G_{JBJ}. $$
\noindent
In the case~(b) we get
$$ VG_B\oplus VKG_B = H_2; $$
$$ G_{B^*} = H_2 \ominus VG_B = VKG_B = - RG_B = G_{JB^{-1}J}. $$
Suppose now that $D$ has a positive odd dimension.
In this case we consider a linear operator $\mathcal{A} = A\oplus A$, with $D(\mathcal{A}) = D(A)\oplus D(A)$, in a Hilbert space
$\mathcal{H} = H\oplus H$ with a conjugation $\mathcal{J} = J\oplus J$.
Observe that $\mathcal{A}$ is a closed $\mathcal{J}$-skew-symmetric ($\mathcal{J}$-isometric) operator with
$\overline{D(\mathcal{A})} = \mathcal{H}$ (respectively $\overline{D(\mathcal{A})} = \overline{R(\mathcal{A})} = \mathcal{H}$).
Its graph $G_{\mathcal{A}}$ in a Hilbert space $\mathcal{H}_2 = \mathcal{H}\oplus \mathcal{H}$ may be identified
with $G_A\oplus G_A$ in $H_2\oplus H_2$:
$$ G_{\mathcal{A}} = \{ \{ (f,Af), (g,Ag) \},\ f,g\in D(A) \}. $$
Let $\mathcal{R}$, $\mathcal{K}$ be constructed for $\mathcal{A}$ as $R$ and $K$ for $A$.
In the case (a) we see that
$$ \mathcal{H}_2\ominus [G_\mathcal{A}\oplus \mathcal{R} G_\mathcal{A}] =
(H_2\ominus [G_A\oplus RG_A]) \oplus (H_2\ominus [G_A\oplus RG_A]), $$
has a positive even dimension.
In the case (b), $\mathcal{H}_2\ominus [G_\mathcal{A}\oplus \mathcal{K} G_\mathcal{A}]$
has a positive even dimension.
Thus, we may apply the above construction with $\mathcal{A}$ instead of $A$.
\noindent
$\Box$
|
\section{Introduction}
The question of determining the model that best describes the universe is
the ultimate goal of cosmology. The energy-momentum content of the present
universe seems to be a perfect fluid mainly consisting of a dark sector
(possibly consisting of a dark energy and a dark matter component),
baryonic matter, and radiation \cite{cosmology}. In the standard model of cosmology
(namely, $\Lambda$CDM) dark matter \cite{DM} and baryonic matter are
considered to
be dust, dark energy \cite{DE} is taken to be the Einstein's
cosmological
constant,
and radiation is described by the usual energy-momentum term for
radiation. Although the standard model seems to be
compatible with
observations yet it has some problems. The magnitudes of potential
theoretical contributions to cosmological constant (CC)
are
extremely higher than the value of CC deduced from
the energy density of the universe \cite{CCP}.
There are many
attempts to solve this problem, namely, the CC problem. Nevertheless none
is wholly satisfactory.
The best option seems to employ a symmetry such as metric reversal
symmetry \cite{MRS} to cancel CC and then attribute the dark energy to
something
else e.g. to modified gravity \cite{modified-grav}, or to some scalar
field such as
quintessence \cite{quintessence}. Cold dark matter (i.e. dust-like dark
matter with
no or negligible interaction with itself and with baryonic matter and
photons) scenario of
$\Lambda$CDM as well suffers from some problems such as rotation curves of
spiral-like galaxies i.e. cuspy halo problem, missing satellite galaxies
problem \cite{cuspy,satellite}. There are many alternatives to cold dark
matter (CDM) scenario
including warm-dark matter \cite{warm-DM}, Bose-Einstein condensate dark
matter \cite{condense-DM,scalar-condensate}, and scalar field dark matter
\cite{scalar-DM}.
The above considerations essentially hold for the time from the
radiation dominated era till the present era. The standard paradigm for
the era before the radiation dominated era is an inflationary era (that
serves to solve the problems of the standard cosmology
such as horizon, flatness, absence of monopoles problems)
\cite{inflation}.
Usually the inflationary era and the epoch after this era are studied
separately. This is not only due to the need to concentrate on each of
these and to try to understand each epoch better before a
possible unification. In fact the most serious problem in the direction of
the unification
\footnote{I mean a true unification i.e. description of
the whole cosmological evolution by a single scale factor in
the metric.}
of the whole cosmic history is the difficulty of merging
these two epochs because of the form of the dependence of the energy
density of dust and radiation on scale factor (i.e. on redshift).
In $\Lambda$CDM the energy density of
radiation dominates over that of inflaton
if one goes back to sufficiently large redshifts.
This is due to the fact that the energy density of inflaton is essentially
constant during
inflationary era while the energy density of radiation scales like
$\frac{1}{a^4}$ where $a$ is the scale factor. In other words, to have a
true unification, the creation of radiation and
matter after the inflationary era must be taken into account in the
scale factor without destroying the standard cosmology before and after
the inflation, and this is not an easy task. The models in
literature that unify all eras of cosmological evolution in a single model
\cite{all-unified,Erdem}
are not wholly realistic since they do not include baryonic matter
although they are able to produce eras of cosmological
evolution with correct equations of state in the corresponding eras, and
some have graceful exit from
inflationary era. The matter in these
models must be identified with dark matter since the
energy densities of these models do not contain energy components that
scale
proportional to $\frac{1}{a^3}$ for all times (or
at least for a sufficiently long time).
The models in
\cite{all-unified} use the energy densities expressed in
terms of simple functions of Hubble parameter
and/or scale parameter as the starting point rather than starting
from the scale factor.
Although one may, in principle, determine the scale factor from this
information the form of scale factor may be rather complicated in
some cases. On the
other hand a relatively
simple scale factor may result in a rather complicated and unmanageable
functional form for the energy density when expressed in
terms of the scale factor or the Hubble parameter. Therefore in some cases
it
may be more suitable to consider a specific ansatz for scale factor
such as in this study and in \cite{Erdem}. The same approach is adopted in
this study.
Moreover the present study introduces
a general prescription to include dust and radiation into unification.
In this study, first, in Section II, I introduce a scheme to unify the
cosmological evolution
before and after the radiation dominated era. Then I give a
concrete realization of this scheme in Section III. In Section IV I
discuss the
observational compatibility of this scheme in the context of the model
introduced in Section III. Finally I conclude in Section V. The scale
factor in this model is a sum of two terms. The
first term is a pure dark
energy contribution. The second term is responsible for the
baryonic matter and radiation terms and additional terms that may be
mainly identified with dark matter. There is also an additional term due
to coupling
between these terms, and this term gives another contribution to
the dark energy and dark matter.
Some of the ideas employed here have been already studied
in literature.
In this study I do not make a sharp distinction between dark energy and
dark matter because the dark energy and dark matter terms are coupled
and the equation of state (EoS) of some terms
e.g. EoS of the coupling term between dark matter and dark energy terms
evolve with time.
The superficiality of
a distinction between dark energy and dark matter is considered in
many studies in literature, either explicitly or implicitly
\cite{Matos1, Matos2, Arbey,unified,Aviles}.
This option is quite possible
since dark energy and dark matter are not observed directly. What we see
observationally is a missing element in the energy-momentum tensor of the
Einstein equations, other than baryonic matter and radiation, and this
missing quantity may be described by two components; dark energy and dark
matter. It is, in principle, equally possible that this quantity is
composed of a single component, say, dark fluid.
In \cite{Erdem} I had introduced a universe composed of a
dark fluid (that may be written in terms of two scalar fields). In fact
the scale factor in that study is essentially $a_1(t)$ in Eq.(\ref{a2}) of
this paper.
The present study, somewhat, may be considered
as an extension of \cite{Erdem} where baryonic matter and radiation are
included. However there are important differences as well. The main aim
of this
study is to introduce a scheme to merge the cosmological evolution of the
time before and after the production of radiation into a single scale
factor with the baryonic matter and the usual radiation terms included.
The modified form of $a_1(t)$ in \cite{Erdem} only serves
as a realization of
this scheme. Furthermore I do not discuss the
scalar field identification of the energy density due the part of
the scale factor similar to $a_1(t)$ of \cite{Erdem}
(although it can be easily done), and I do not consider the
cosmological perturbations of
these quantities, and the inflationary era in this study because these
points would cause divergence of the main goal of the paper
and would increase the volume of this study drastically. I leave
these points to future studies.
\section{Outline of the Model}
Consider the Robertson-Walker metric
\begin{equation}
ds^2\;=\;g_{\mu\nu}dx^\mu\,dx^\nu\,=\,-dt^2\,+\,a^2(t)\tilde{g}_{ij}dx^idx^j
\label{a1}
\end{equation}
I take the 3-dimensional space be flat, i.e.
$\tilde{g}_{ij}=\delta_{ij}$ for the sake of simplicity, which is an
assumption consistent with cosmological observations \cite{PDG,Planck}.
I let the form of the scale factor be
\begin{eqnarray}
a(t)&=&c_0\left(a_1(t)\,-\,a_2(t)\right) \label{a2} \\
&&
c_0=\frac{1}{A_1-A_2}~,~~A_1=a_1(t_0)~,~~A_2=a_2(t_0) \nonumber
\end{eqnarray}
where $t_0$ denotes the present time. We will see that $a_1(t)$ is the
part of
the scale factor responsible for
dark energy and dark matter, and $a_2(t)$ is the one
mainly responsible for dust and radiation and additional contribution to
dark matter-energy, and we shall see later that a
mixing between the sectors due to $a_1$ and $a_2$ act as an additional
source of dark energy. We assume that $a_1(t)$ and $a_2(t)$ are chosen in
such a way that $a(t)\,>\,0$ for all $t$. In general one may
identify the dust by a mixture of baryonic matter
and dust-like dark matter. The best fit values that I could find
by trial and error for the specific toy model considered in this
study for
implementation of the present scheme
seem to prefer
the case where the dust
term is wholly or almost wholly due to baryonic matter.
We first focus on the
$a_2(t)$ term and specify it as
\begin{equation}
a_2(t)\,=\,x(t)\,a(t) \label{a3}
\end{equation}
where $x(t)$ is some function that its form will be specified later.
Eqs. (\ref{a2}) and (\ref{a3}) may be used to relate
$a(t)$ and $a_1(t)$, $a_2(t)$ in a more applicable way, and to derive the
corresponding Hubble parameter.
We observe that
\begin{eqnarray}
a_2&=&
ax\,=\,c_0(a_1-a_2)x~~~\Rightarrow~~
a_2=\frac{c_0x}{1+c_0x}a_1~,~~
a\,=\,\frac{c_0}{1+c_0x}a_1\,=\,\frac{1}{x}a_2
\label{a4a}
\end{eqnarray}
In a similar way the Hubble parameter is found to be
\begin{eqnarray}
&&\frac{\dot{a}_2}{a}\,=\,\frac{\dot{a}}{a}\,+\,\dot{x}~~~\Rightarrow~~~
H\,=\,\frac{\dot{a}}{a}\,=\,
\frac{\dot{a}_{1n}}{a_{1n}}
-\frac{c_0\dot{x}}{1+c_0x}
\label{a4b}
\end{eqnarray}
where we have used
\begin{eqnarray}
\frac{\dot{a}_1}{a}
&=&\left(\frac{1+c_0x}{c_0}\right)\frac{\dot{a}_1}{a_1}~,~~
\frac{\dot{a}_{1}}{a_{1}}\,=\,
\frac{\dot{a}_{1n}}{a_{1n}}
\label{a5a}\\
&&a_1\,=\,A_1\,a_{1n}~,~~a_{1n}(t_0)=1 \nonumber
\label{a5b}
\end{eqnarray}
Note that $a(t_0)=1$
by convention.
We let
\begin{eqnarray}
&&x(t)\,=\,A_2\frac{1}{c_1-c_2}x_3(t)\left(c_1x_1(t)-c_2x_2(t)\right)
\label{a3a} \\
&&x_1(t)\,=\,\exp{\int_{t_0}^t\tilde{H}_2^{(1)}dt}~,~~x_2(t)\,=\,
\exp{\int_{t_0}^t\tilde{H}_2^{(2)}dt} \label{a3b} \\
&&x_3(t)\,=\, \exp{\int_{t_0}^t\tilde{H}_2^{(3)}dt} \label{a3c}
\end{eqnarray}
where $c_1$, $c_2$ are some constant coefficients, and
\begin{eqnarray}
&&\tilde{H}_2^{(1)}
\,=\,\alpha_{o1}\left[-\frac{\alpha_b}{a^{\frac{3}{2}}}-
\frac{\alpha_r}{a^2}+\frac{\alpha_x}{a^3}\right]\,,~
\tilde{H}_2^{(2)}
\,=\,\alpha_{o2}\left[\frac{\alpha_b}{a^{\frac{3}{2}}}+
\frac{\alpha_r}{a^2}-\frac{\alpha_K}{a}\right]\,,~
\tilde{H}_2^{(3)}
\,=\,\alpha_c\left(\frac{1}{a^3}-\frac{1}{a}\right) \nonumber \\
\label{a10a} \end{eqnarray}
where $\alpha_{o1}$, $\alpha_{o2}$, $\alpha_b$, $\alpha_r$, $\alpha_x$,
$\alpha_K$ are some other constant coefficients.
In fact,
in (\ref{a3a}) we could take the simpler form where $\alpha_c=0$,
$\alpha_{o1}=\alpha_{o2}=1$, $c_1=1$,
$c_2=0$. This
would be enough as long as we are concerned only with merging of the eras
before and after the radiation domination, and the resulting model would
be compatible with Union2 data set at an order of magnitude level. The
more involved form in (\ref{a3a}) is used to make the model
phenomenologically more viable.
This point will
be discussed when we discuss the phenomenological viability of the model
in Section IV. One may determine $\dot{x}$ in Eq.(\ref{a4b}) by using Eq.(\ref{a3a}),
\begin{equation}
\dot{x}(t)
\,=\,A_2\frac{1}{c_1-c_2}x_3\left[-B(t)\left(\frac{\alpha_b}{a^\frac{3}{2}}+
\frac{\alpha_r}{a^2}\right)\,+\,S(t)\frac{\alpha_x}{a^3}\,+\,K(t)
\frac{\alpha_K}{a}\right]
\end{equation}
where
\begin{eqnarray}
B(t)&=&\frac{A_2}{c_1-c_2}x_3
\left(c_1\alpha_{o1}x_1+c_2\alpha_{o2}x_2\right) \label{c1} \\
S(t)&=&\frac{A_2}{c_1-c_2}x_3\left[c_1\alpha_{o1}\alpha_xx_1+
\alpha_c\left(c_1x_1-c_2x_2\right)\right]
\label{c2} \\
K(t)&=&\frac{A_2}{c_1-c_2}x_3\left[c_2\alpha_{o2}\alpha_Kx_2-
\alpha_c\left(c_1x_1-c_2x_2\right)\right] \label{c3}
\end{eqnarray}
Hence
one may express
(\ref{a4b}) as
\begin{equation}
H\,=\,H_{1n}\,+\,A(t)\,\tilde{H}_2\,+\,H_\Delta
\label{a3ac}
\end{equation}
where
\begin{eqnarray}
H_\Delta&=&-\Xi(t)\,\frac{1}{a^3}\,-\,\psi(t)\,\frac{1}{a} \label{a3da}\\
&&A(t)\,=\,
\frac{c_0B(t)}{1+c_0x(t)}~,~~
\tilde{H}_2\,=\,
\frac{\alpha_b}{a^{\frac{3}{2}}}\,+\,
\frac{\alpha_r}{a^2} \label{a3db} \\
&&\Xi(t)\,=\,
\frac{c_0S(t)}{1+c_0x(t)}~,~~
\psi(t)\,=\,
\frac{c_0K(t)}{1+c_0x(t)} \label{a3dc}
\end{eqnarray}
We let
\begin{equation}
H_{1n0}=\tilde{\Omega}_1^\frac{1}{2}H_0~,~~
A_0\alpha_b=\tilde{\Omega}_b^{\frac{1}{2}}H_0~,~~
\,A_0\alpha_r=\tilde{\Omega}_r^{\frac{1}{2}}H_0~,~~
\,\Xi_0=\tilde{\Omega}_x^{\frac{1}{2}}H_0~,~~
\,\psi_0=\tilde{\Omega}_K^{\frac{1}{2}}H_0
\label{a10b}
\end{equation}
where $H_{1n0}=H_{1n}(t_0)$, $H_0=H(t_0)$,
$A_0=A(t_0)$, $\Xi_0=\Xi(t_0)$, $\psi_0=\psi(t_0)$. Because the three dimensional part
of metric is taken to be flat the present energy density is equal to the
critical energy density the above equations imply that
\begin{equation}
\tilde{\Omega}_1^\frac{1}{2}\,+\,
\tilde{\Omega}_b^{\frac{1}{2}}\,+\,
\tilde{\Omega}_r^{\frac{1}{2}}\,-\,
\tilde{\Omega}_x^{\frac{1}{2}}\,-\,
\tilde{\Omega}_K^{\frac{1}{2}}\,=\,1
\label{a10ba}
\end{equation}
Note that, at this point
$\tilde{\Omega}_1$,
$\tilde{\Omega}_b$,
$\tilde{\Omega}_r$,
$\tilde{\Omega}_x$,
$\tilde{\Omega}_K$ cannot be identified as density parameters since
density parameters should satisfy
$\Omega_1$+$\Omega_b$+$\Omega_r$+$\Omega_x$+$\Omega_K$=1. In Chapter IV we
will see that this condition is not satisfied for the phenomenologically viable
sets of parameters, so
$\tilde{\Omega}_1$,
$\tilde{\Omega}_x$,
$\tilde{\Omega}_K$ cannot be identified as density parameters separately,
instead one must define the total density parameter for dark sector by
$\Omega_D^\frac{1}{2}$=$\tilde{\Omega}_1-
\tilde{\Omega}_x-\tilde{\Omega}_K$ rather than the separate contribution
due to $H_{1n}$ and $H_\Delta$ while I identify
$\tilde{\Omega}_b$, $\tilde{\Omega}_r$ as the density
parameters corresponding dust and radiation.
Therefore to retain the
physical content of this paper more evident I will not make a distinction
between
$\tilde{\Omega}_b$, $\tilde{\Omega}_r$
and the density parameters
for baryonic matter, radiation; $\Omega_b$, $\Omega_r$
while I keep this distinction for the others i.e. for the ones due
to $H_{1n}$ and $H_\Delta$ terms.
The $\frac{\alpha_b}{a^\frac{3}{2}}$ and
$\frac{\alpha_r}{a^2}$ terms result in energy densities that are
identified as the energy densities for baryonic matter and radiation. In
principle, there may be also contributions due to the
$\Xi\frac{1}{a^3}$ and $\psi\frac{1}{a}$.
The sign of the $\Xi\frac{1}{a^3}$ term is negative of
the usual stiff matter. It may be identified as stiff matter under
pressure so that it has a negative deceleration parameter. The main
function of
this term is to damp the
energy densities of baryonic matter
and radiation in the time before the radiation dominated era. The function
of the $\frac{1}{a}$ term is similar. It ensures the behavior of the
energy density in late times be well-behaved (i.e. preventing the energy
density to grow too
fast (through the $\frac{1}{a}$ term in $x_2(t)$ and $x_3(t)$)).
Although the $\psi\frac{1}{a}$ term is similar to that of a negative
curvature 3-space it is different from such a term since its origin is
the Hubble parameter $H$ while a usual 3-curvature term arises from the
3-curvature part of metric. Note that this term arises even in a flat
3-space
in this construction. Therefore I identify the $\frac{\Xi}{a^3}$ and
$\frac{\psi}{a}$ terms in $H$ as additional contributions to dark sector.
Another point worth to mention is; It is evident that the
square of (\ref{a4b}) (in conjunction with
(\ref{a10a})) results in an $A^2\tilde{H}_{2}^2$
term containing $A^2\frac{\alpha_b^2}{a^3}$ and
$A^2\frac{\alpha_r^2}{a^4}$ terms that may be identified with
the standard baryonic matter and radiation terms, respectively
if $A$ is taken to be constant while it depends on time in this scheme as
it is evident from (\ref{a3a}). In fact variation of $A$ with time makes
it possible to go to zero before the radiation dominated era as desired.
Therefore, given the considerable success of the standard model at least
in the observed relatively small redshifts, the variation in $A$ after the
matter - radiation decoupling time should be small so that this scheme
mimics the standard model at relatively small redshifts where
observational data is available. If one takes
$\left(\frac{dA}{dt}\right)_{t\simeq\,t_0}$ sufficiently small one may
guarantee an almost constant value for $A$ for a sufficiently long
time (e.g. from the present time till the beginning of the radiation
dominated era). We will see in Section IV
that there exist such values of $A$ with reasonable phenomenological
viability.
Another term
arising from $\tilde{H}_2^2$ is the cross term,
$A^2\frac{\alpha_b\alpha_r}{a^\frac{7}{2}}$. This term may be identified
as the energy density term due to the transitory time where massive
particles that act as radiation at high energies turn into more dust-like
at intermediate energies. Another term in $H^2$ is $H_{1n}^2$. This term
will be considered as a pure dark sector term. Finally the cross term
$2H_{1n}\tilde{H}_2$ gives an additional contribution to the dark sector
for the phenomenologically viable values of the parameters.
It may be easily shown that
this term does not necessarily imply strong interaction between the
dark fluid and
radiation and baryonic matter as its form may suggest
if the parameters of the underlying physics at microscopic
scale satisfy some restrictions. Otherwise one may use
screening mechanisms such as
\cite{chameleon,symmetron,Pospelov} to explain the
unobservablity of dark matter-energy.
Next we derive the general form of
the equation of state for this
model. We derive the explicit form of the equation of state after (EOS)
after we give the explicit form of $a_1(t)$ in the section.
However giving the general form of EOS in this scheme provides us a more
model independent formula and may be useful for other choices of $a_1(t)$
in future.
After
using Eqs.(\ref{a3ac},\ref{a3da},\ref{a3db},\ref{a3dc})
one obtains EOS, $\omega$ as
\begin{eqnarray}
\omega&=&\frac{p}{\rho}\,=\,\frac{\frac{G_{11}}{g_{11}}}{G_{00}}\,=\,
-\frac{2\dot{H}+3H^2}{3H^2} \nonumber \\
&=&
-\frac{2\dot{H}_{1n}\,+\,3H_{1n}^2}{3H^2}
\,-\,\frac{AH_{1n}\left(\frac{3\alpha_b}{a^\frac{3}{2}}+\frac{2\alpha_r}{a^2}
\right)}{3H^2}\,+\,\frac{A^2\left(\frac{\alpha_r^2}{a^4}
+\frac{\alpha_b\alpha_r}{a^\frac{7}{2}}\right)}{3H^2} \nonumber \\
&&\,-\,\frac{2\dot{A}\tilde{H}_2}{3H^2}
\,+\,
\frac{
A\left(\frac{\Xi}{a^3}+\frac{\psi}{a}\right)
\left(\frac{3\alpha_b}{a^\frac{3}{2}}+\frac{2\alpha_r}{a^2}\right)}{3H^2}
\,+\,
\frac{6H_{1n}\left(\frac{\Xi}{a^3}+\frac{\psi}{a}\right)}{3H^2}\nonumber
\\
&&-\,
\frac{2H\left(\frac{3\Xi}{a^3}+\frac{\psi}{a}\right)\,+\,3
\left(\frac{\Xi^2}{a^6}+\frac{\psi^2}{a^2}+\frac{2\Xi\psi}{a^4}\right)}
{3H^2}
\label{a11a}
\end{eqnarray}
The terms inside the first parenthesis in the second line correspond to
the contribution of
the dark sector term $H_{1n}$. The other
terms in the same line correspond to the
contributions of dust and radiation and their coupling with
dark sector term $H_{1n}$. The remaining terms
are the term corresponding to variation of $A$,
the term corresponding to coupling of
curvature-like term and the stiff matter under
negative pressure with dust and radiation, the
term corresponding to coupling of
curvature-like term and the stiff matter under negative
pressure with $H_{1n}$, the term corresponding to coupling of
curvature-like term and the stiff matter under negative pressure
with the other terms, and the contribution of the curvature-like
term and the stiff matter under negative pressure, respectively.
It is evident from (\ref{a11a})
that the pressure for baryonic matter is zero as should be, and the
pressure for radiation is $\frac{1}{3}$ as
expected.
A point worth to mention at this point
is; The coupling term between baryonic matter and radiation
in Eq.(\ref{a11a}) has an equation of state $\frac{1}{6}$ (that may be
seen by considering the ratio of the
$\frac{\alpha_b\alpha_r}{a^\frac{7}{2}}$ in $p$ by the corresponding term
in $\rho$ i.e. $2\frac{\alpha_b\alpha_r}{a^\frac{7}{2}}$). The redshift
dependence of this term is between that of baryonic matter and
radiation. This time dependence is more natural than the standard picture
where there is no such term.
Massive particles at high energies act as radiation
and at lower turns into dust. The coupling term accounts for the
transitory time when massive particles pass from radiation to dust state.
In order to obtain the evolution of $\omega$ as a
function redshift or time explicitly,
$H_{1n}$ must be specified. This will be done in the next section.
However I give a $\omega$ versus redshift graph in Fig.\ref{fig:1}
for $a_{1n}$ introduced in the next section for a phenomenologically
viable set of parameters (i.e. those with small $\chi^2$ values and
with energy densities for recombination and nucleosynthesis as
discussed in Section IV) to have an idea about the evolution of
$\omega$ with redshift. To draw this graph I have converted time,
t to redshift, z (for Union2.1 data) through
the relation $z=\frac{1}{a}-1$, and then used Mathematica to use
this relation to make the calculations (although the original
quantities are expressed in terms of time). This procedure is
applicable for small redshifts. However, in general,
it becomes inapplicable
due to highly non-linear form of scale factor and Hubble parameter
since it requires huge RAM and CPU for computation, if
it can be done at all,
and hence requires a separate computational physics project by
itself. Therefore I have used equation of state versus and energy
density versus time graphs (instead of redshift) in Section IV.
In fact, even that option
required a long time of order of months to make the
necessary computations.
\section{An Explicit Realization of the Model}
Now we focus on the $a_1(t)$ term. We take
\begin{eqnarray}
a_1(t)&=&A_1a_{1n}(t)\label{a13a} \\
a_{1n}(t)&=&[p_1\,+\,p_2b_2t]^r\,\exp{[-b_1(b_2t)^{-1/s}]} \label{a13b}
\end{eqnarray}
where $A_1\,<\,1$,
$p_1$, $p_2$, $b_2$, $b_1$ are some constants that to be fixed or
bounded by consistency arguments or cosmological observations. This scale
factor is a generalization of the scale factor in
\cite{Erdem} where
$r=1$, $s=6$. A similar scale factor is considered in \cite{Sami} as
well. One of the shortcomings of \cite{Erdem}
is that the present value of the equation of state parameter in that model
(for phenomenologically relevant choices of parameters where the model mimics
$\Lambda$CDM)
is $\sim\,-0.4$ while observations imply that it
should be $\simeq\,(-0.68)$ --- $(-0.74)$ \cite{PDG,Planck}.
In the present study there
is an
additional contribution due to mixing of the terms due to $a_1$ and $a_2$
and hence there is less need to modify the scale factor in \cite{Erdem}.
However
I prefer to adopt the more general form in (\ref{a13b}) to seek a greater
parameter space and to insure the correct equation of state parameter.
We had shown in Eq.(\ref{a4b}) that the Hubble parameter may be expressed
as
$H\,=\,H_{1n}\,+\,A(t)\,\tilde{H}_2\,+\,H_\Delta$.
Now we concentrate on the $H_{1n}=\frac{\dot{a}_{1n}}{a_{1n}}$ part of the
Hubble parameter. In fact this amounts to specifying the model wholly
since the other terms, as well, depend on $a_{1n}$ as we have seen.
The corresponding $H_{1n}$ is given by
\begin{equation}
H_{1n}(t)\,=\,
\frac{\dot{a}_{1n}}{a_{1n}}
\,=\,\frac{rp_2b_2}{(p_1+p_2b_2t)}
+\frac{1}{s}b_2b_1(b_2t)^{-\left(1+\frac{1}{s}\right)}
\label{a14}
\end{equation}
We let
\begin{eqnarray}
&&1=a_{1n0}=a_{1n}(t_0)\,=\,
(p_1+p_2b_2t_0)^r
\exp{[-b_1(b_2t_0)^{-\frac{1}{s}}]}\label{a15}
\end{eqnarray}
and
\begin{eqnarray}
&&
H_{0}t_0\,=\,\xi,~~~
(p_1+p_2b_2t_0)^r \,=\,
\exp{[b_1(b_2t_0)^{-\frac{1}{s}}]}\,=\, \beta\,>\,1
\label{a16a} \\
H_{1n0}t_0
&&=\,H_{1n}(t_0)t_0\,=\,\frac{rp_2b_2t_0}{(p_1+p_2b_2t_0)}
+\frac{1}{s}b_1(b_2t_0)^{-\frac{1}{s}}
\,=\,\xi\xi_1\,\nonumber \\
&&\Rightarrow~~~~
rp_2b_2t_0\,=\,\beta^{\frac{1}{r}}(\xi\xi_1-\frac{1}{s}\ln{\beta})~,~~~
p_1\,=\,
\beta^{\frac{1}{r}}[1-\frac{1}{r}(\xi\xi_1-\frac{1}{6}\ln{\beta})]
\label{a16b}
\\
&&\Rightarrow~~~
H_1(t)\,=\,H_{1n}(t)\,=\,H_{1n}(\gamma)
\,=\,\frac{1}{t_0}\{\frac{\xi\xi_1-\frac{1}{s}\ln{\beta}}
{[1+\frac{\gamma-1}{r}(\xi\xi_1-\frac{1}{s}\ln{\beta})]}
+\frac{1}{s}\gamma^{-\frac{s+1}{s}}\ln{\beta}\}
\label{a16c} \\
&&\mbox{here}~~~~\gamma=\frac{t}{t_0} \nonumber
\end{eqnarray}
where $t_0$ is the present age of the universe. One observes from
(\ref{a10b}) and the above expression
that
\begin{equation}
H_{1n0}=\tilde{\Omega}_1^\frac{1}{2}H_0=\frac{1}{t_0}\xi\xi_1
~~\Rightarrow~~\tilde{\Omega}_1^\frac{1}{2}\,=\,\xi_1
\end{equation}
We will see in the next section that
$\tilde{\Omega}_1$
cannot be identified as the density parameter
corresponding to $H_{1n}$. Instead one must define an overall density
parameter for the dark sector by
$\Omega_D^\frac{1}{2}$=
$\tilde{\Omega}_1^\frac{1}{3}-
\tilde{\Omega}_x^\frac{1}{3}-
\tilde{\Omega}_K^\frac{1}{3}$.
Observational values
of $H_0=\left(\frac{\dot{a}}{a}\right)_{t=t_0}$ and $\frac{1}{t_0}$ are
almost the same. Therefore $\xi^2\xi_1^2\,\simeq\,\xi_1^2$.
After determining the $H_{1n}$ we are almost ready to find the explicit
values of the energy density and the equation of state. The only missing
element for calculation of these quantities is to find $A$, $\Xi$, $\psi$
in (\ref{a3ac}, \ref{a3da}, \ref{a3db}, \ref{a3dc}). Another point to
be addressed is to show that there exist sets of A whose variation with
time are small for low redshifts so that the terms that are proportional
to $\frac{1}{a^\frac{3}{2}}$
and $\frac{1}{a^2}$ in $A\,\tilde{H}_2$ term may be identified with dust and
usual radiation terms, respectively.
In order to determine $A$, $\Xi$, $\psi$ (and to determine the rate of
variation of $A$ with time) one should derive an approximation scheme
for the evaluation of these quantities because these quantities depend on
$x_1(t)$,
$x_2(t)$,
$x_3(t)$ (that are defined in (\ref{a3b}) and (\ref{a3c})), and these
quantities, in turn, are defined in a recursive way since
$x_i(t)$=$\exp{(\int_{t_0}^t\tilde{H}_2^{(i)}dt)}$ ($i=1,2,3$) and
$\tilde{H}_2^{(i)}$ depend on $a(t)$, and $a(t)$, in turn, depends
on $x_i(t)$ through Eq.(\ref{a4a}). In other words, in order to determine
the approximate values of $x_i(t)$ one must identify the zeroth order
approximation and a method how to obtain the higher order approximations
in an iterative way. One may use the following observations to obtain the
zeroth order approximation;
$A_1c_0(1+c_0x(0))^{-1}=A_1\frac{1}{A_1-A2}(1+\frac{A_2}{A_1-A_2})^{-1}=1$
and $\dot{A}\sim\,0$ $\Leftrightarrow$
$\dot{x}_i\sim\,0$ ($i=1,2,3$) i.e. $\dot{x}\sim\,0$, $x(t)\simeq\,x(0)=1$ for
small redshifts. This implies that the zeroth order approximation for the scale
factor $a(t)$ should be taken as $a^{(0)}(t)\,=\,a_{1n}(t)$
Hence for phenomenologically
viable cases (where $\dot{A}\sim\,0$ for small redshifts) one may take
the zeroth order approximations as
\begin{equation}
x_i^{(0)}(t)\,=\,\exp{\int_{t_0}^t\tilde{H}_2^{(i0)}dt}~~~i=1,2,3
\label{a3ba}
\end{equation}
where
$\tilde{H}_2^{(10)}$, $\tilde{H}_2^{(20)}$, $\tilde{H}_2^{(30)}$ is
obtained from $\tilde{H}_2^{(1)}$, $\tilde{H}_2^{(2)}$,
$\tilde{H}_2^{(3)}$ by replacing $a(t)$ by $a_{1n}(t)$ in those
expression, for example,
\begin{equation} \tilde{H}^{(10)}_2
\,=\,\alpha_{o1}\alpha_b\int_{t_0}^t\{-\frac{1}{a_{1n}^{\frac{3}{2}}}-
\frac{\frac{\alpha_r}{\alpha_b}}{a_{1n}^2}+
\frac{\frac{\alpha_x}{\alpha_b}}{a_{1n}^3}\} \label{a17a} \end{equation}
Then
\begin{eqnarray}
&&A^{(0)}(t)\,=\,
\frac{c_0B^{(0)}(t)}{1+c_0x^{(0)}(t)} \label{d1} \\ &&x^{(0)}
(t)\,=\,\frac{A_2}{c_1-c_2}x_3^{(0)}
\left(c_1x_1^{(0)}-c_2x_2^{(0)}\right) \label{d2} ~~etc.
\label{d2} \end{eqnarray}
One may get the next order approximation by using
\begin{equation}
a(t)\simeq\,a^{(1)}(t)= c_0A_1(1+c_0x^{(0)})^{-1} a_{1n}
\end{equation}
The next order quantities $A^{(1)}$, $x^{(1)}$ may be obtained from (\ref{d1}) and (\ref{d2}) by replacing the
superindices $(0)$ by $(1)$
where
\begin{equation}
x_i^{(1)}(t)\,=\,\exp{\int_{t_0}^t\tilde{H}_2^{(i1)}dt}
\label{a3bc}
\end{equation}
Here $\tilde{H}_2^{(i1)}$ is obtained from
$\tilde{H}_2^{(i)}$ by replacing $a(t)$ by
$a^{(1)}(t)= c_0A_1(1+c_0x^{(0)})^{-1} a_{1n}$.
For k'th approximation we replace $a(t)$ by
$a^{(k)}(t)= c_0A_1(1+c_0x^{(k-1)})^{-1} a_{1n}$.
In principle this may be done up to arbitrarily higher order approximations
but it is quite difficult to calculate even $A^{(1)}$ even with the help
of computers. In fact I have divided the interval $t\,-\,t_0$ in to
coarser subintervals to decrease the CPU time and have used the
approximate numerical values in the i'th interval (by assuming $A^{(0)}$
to be almost constant in those intervals) by using the formula
\begin{equation} A^{(0)}(t_i)=\frac{A^{(0)}(t_{i-1})\,+\,
A^{(0)}(t_{i+1})}{2} \label{ax}
\end{equation}
to find $A^{(1)}$. I have seen (by trial and error) that it is
possible to find almost constant
$A^{(0)}$ and $A^{(1)}$ values for many relevant (i.e. of
small $\chi^2$ values considered in the next section)
choices of parameters,
$\alpha_b$, $r$, $s$, $\xi_1$, $\xi_1$, $A_1$, $A_2$, $c_1$, $c_2$,
$\alpha_r$, $\alpha_c$, $\alpha_{o1}$, $\alpha_{o2}$, $\alpha_x$,
$\alpha_K$. For example the variations of
$A^{(0)}$ and $A^{(1)}$ with time for one of the phenemonologically
viable sets in Table \ref{table-9} is given in Table \ref{table-1}.
\section{Compatibility with Observations}
Now we check the phenomenological viability of the model. The
observational analysis of the model for all possible values of the
parameters, $\beta$, $r$, $s$, $\xi$, $\xi_1$, etc. is an extremely
difficult job (if not impossible at all) because expressing the Hubble
parameter, deceleration parameter etc. in terms of the scale factor is
quite difficult since these quantities are highly nonlinear
functions of the scale factor in this model. Therefore I adopt
some guidelines to seek the phenomenologically viable sets of
parameters. These guidelines are: \\
1- I take the model mimic the standard model i.e. the $\Lambda$CDM
model, at least from the time of decoupling of matter and radiation up to
the present time. Therefore I take the present time values of
the equation of state of the whole universe and the density parameter of
the baryonic matter and radiation to be the same as
$\Lambda$CDM. \\
2- In searching for the phenomenologically viable parameter space I start
from the values of the parameters in \cite{Erdem} i.e. $r=1$, $s=6$,
$\xi=1$, and $\beta\sim\,O(1)$ since the universe studied in \cite{Erdem}
mimics the true universe roughly. \\
3- Due to the highly non-linear relation between the Hubble parameter and
the scale factor I seek the relevant parameter space usually by trial and
error rather than a continuous scan of the parameter space. Therefore the
optimum values obtained here most probably may not correspond to the best
possible optimization. Rather they hopefully correspond to a good
approximation to the best optimal values.
\subsection{Compatibility with Union2.1 Data}
In this subsection we use the Union2.1 compilation data set to
find the optimal values of $\beta$, $r$, $s$ starting from
$\beta=3$, $r=1$, $s=6$. We find the theoretical
values of distance moduli, $\mu$ for the redshift values of
Union2.1 and calculate the corresponding $\chi^2$ value
by using the measured values of $\mu$ and their errors.
The expression for distance modulus is
\begin{equation}
\mu\,=\,5\,Log_{10}\left(\frac{d_L}{1\,Mpc}\right)\,+\,25
\end{equation}
where
\begin{equation}
d_L\,=\,\frac{c\,a_0}{a(t)}\int_t^{t_0}\frac{dt^\prime}{a(t^\prime)}\,=\,
\frac{c\,a_0}{\frac{A_1c_0}{1+c_0x(t)}a_{1n}(t)}
\int_t^{t_0}\frac{dt}{\frac{A_1c_0}{(1+c_0x(t^\prime))}a_{1n}(t^\prime)}
\label{mu1a}
\end{equation}
where for small redshifts reduces to
\begin{equation}
d_L\,
\,\simeq\,
\frac{c}{a_{1n}(t)}\int_t^{t_0}\frac{dt^\prime}{a_{1n}(t^\prime)}
\label{mu1b}
\end{equation}
where we have used the requirement that
$\frac{A_1c_0}{1+c_0x}\simeq\,1$ at small redshifts as
discussed in the preceding section (see Table \ref{table-7}), and
$a_0=a(0)=1$. In $\Lambda$CDM
$\int\frac{dt}{a(t)}$ is usually expressed in terms of redshift, $z$ and
Hubble
parameter $H$, and then the
results for different $z$'s are compared with the data directly. This
is not possible in this model because $H$ cannot be expressed in terms of
$a(t)$ in a simple way.
Therefore in this study first we convert redshift values of Union2 to time
values by using $z=
\frac{1}{a(\gamma)}-1\simeq\,
\frac{1}{a_{1n}(\gamma)}-1$
and then solve it for $\gamma$. The corresponding expression for the
theoretical value of the luminosity distance $d_L$ in this case (i.e. in
terms of $\gamma$) is
\begin{equation}
d_L\,\simeq\,
\frac{c\,t_0\,
\beta^{-1+\gamma^{-\frac{1}{s}}}}{[1+\frac{\gamma-1}{r}\left(\xi\xi_1
-\frac{1}{s}\ln{\beta}\right)]^r}
\int_\gamma^1
d\gamma\frac{\beta^{-1+\gamma^{-\frac{1}{s}}}}
{[1+\frac{\gamma-1}{r}\left(\xi\xi_1-\frac{1}{s}\ln{\beta}\right)]^r}
\label{mu2}
\end{equation}
where $a_{1n}(t)$ is expressed in terms
of $\beta$, $r$, $s$, $\gamma=\frac{t}{t_0}$ by using the parameterization
given in the preceding section. Eq.(\ref{mu2}) may be written in more
standard form
in terms of $H_0$ by using $H_0t_0=\xi$.
Then we find (\ref{mu1b}) numerically for
each of the $\gamma$ corresponding to observational redshifts. Finally we
find the corresponding $\chi_0^2$
values by using the formula
\begin{equation}
\chi_0^2\,=\,\sum_{i=1}^{i=580}\{\frac{(\left(
\mu^{th(0)}(\gamma(i),r,s,\beta,\xi\xi_1,t_0)
-\mu^{obs}_i\right)^2}{\left(\sigma_i\right)^2}\} \label{mu3}
\end{equation}
where the subscript $0$ in $\chi_0$ and the superscript $(0)$ in
$\mu^{th(0)}$ stands for the fact that $a(t)$ is
approximated by its zero'th order approximation i.e. by $a_{1n}$, the
superindices $th$ and $obs$ stand for the theoretical and
observational values of $\mu$, and the subindices $i$ denote the values
of the corresponding quantity for the $i$'th data point in
Union2 data set.
One may
try a better approximation by replacing $a_{1n}(t)$ in (\ref{mu1b}) by a
better approximation of
$a(t)$ i.e. by
$\frac{c_0A_1}{1+c_0x^{(0)}(t)}a_{1n}(t)$ where $x^{(0)}(t)$ is defined by
Eq.(\ref{a17a}).
In principle, then, one may evaluate the integral (\ref{mu1a}) after
replacing $a_{1n}(t)$ by
$\frac{c_0A_1}{1+c_0x^{(0)}(t)}a_{1n}(t)$.
However this seems to be
inapplicable for
standard computers because of the complicated form of the integral. One
needs a separate computational physics project for this aim. Instead one
may try a rough approximation (hopefully better than $a_{1n}$); we take
the $\frac{1+c_0x^{(0)}}{c_0A_1}$ term in the integral to outside of the
integral
with its $\gamma$ value being the bound of the integral. This
approximation is a good approximation provided that
$\frac{c_0A_1}{1+c_0x(t)}$ does not vary much in the time interval between
$t_0$ and the time corresponding to the given redshift value. Otherwise
the higher order approximation may worsen the approximation rather than
improving.
The corresponding formulas (in the first order approximation) become
\begin{equation}
d_L\,\simeq\,
\left(\frac{1+c_0x^{(0)}}{c_0A_1}\right)^2\frac{c\,t_0\,
\beta^{-1+\gamma^{-\frac{1}{s}}}}{[1+\frac{\gamma-1}{r}\left(\xi\xi_1
-\frac{1}{s}\ln{\beta}\right)]^r}
\int_\gamma^1
d\gamma\frac{\beta^{-1+\gamma^{-\frac{1}{s}}}}{[1+\frac{\gamma-1}{r}\left(\xi_1
-\frac{1}{s}\ln{\beta}\right)]^r}
\label{mu4}
\end{equation}
\begin{equation}
\chi^2\,=\,\sum_{i=1}^{i=580}\{\frac{(\left(
\mu^{th}(\gamma(i),r,s,\beta,\xi\xi_1,t_0)
-\mu^{obs}_i\right)^2}{\left(\sigma_i\right)^2}\} \label{mu5}
\end{equation}
After trial and error I have found many sets of parameters with relatively
small $\chi_0^2$, $\chi^2$ values. For example the $\chi_0^2$,
$\chi^2$ values for two phenemonologically viable sets of parameters are
given in Table \ref{table-9} where the reduced $\chi_0^2$,
$\chi_{red\,0}^2=\frac{\chi_0^2}{580-5}$, and the reduced $\chi^2$ values
$\chi_{red}^2=\frac{\chi}{580-12}$ are in the order of 1 (where 580 is
the number of data points, and 5, 12 are the number of free parameters $r$,
$s$, $\beta$ etc. to be adjusted).
The sets of parameters (that I could by trial and error) with relatively small
$\chi^2$ values satisfy
$c_1\simeq\,c_2\,\simeq\,1$, $\alpha_c\ll\,1$. By using this information
one may check the validity of (\ref{a10ba}) and determine if
one may identify $\tilde{\Omega}_1$,
$\tilde{\Omega}$, $\tilde{\Omega}_K$ by the corresponding density
parameters; $\Omega_1$, $\Omega_x$, $\Omega_K$
for the phenomenologically relevant parameters
by using Eq.(\ref{a10b}) and Eq.(\ref{a3db}) and
Eq.(\ref{a3dc}). We observe that
$x(0)$=$A_2$, $x_1(0)$=$x_2(0)$=$x_3(0)$=1, $c_1=1$ and
for relevant values of the parameters
Hence, after using (\ref{a10b}), we obtain
\begin{equation}
\frac{\tilde{\Omega}_x^\frac{1}{2}}{\Omega_b^\frac{1}{2}}\,\simeq\,
\left(\frac{\alpha_{o1}}{\alpha_{o1}+\alpha_{o2}}\right)
\left(\frac{\alpha_x}{\alpha_b}\right)~,~~~
\frac{\tilde{\Omega}_K^\frac{1}{2}}{\Omega_b^\frac{1}{2}}\,\simeq\,
\left(\frac{\alpha_{o2}}{\alpha_{o1}+\alpha_{o2}}\right)
\left(\frac{\alpha_K}{\alpha_b}\right)\label{mu7}
\end{equation}
We observe that for phenemonologically viable sets of parameters, for example,
for those in
Table \ref{table-9} we have
$\tilde{\Omega}_x^\frac{1}{2}\,\sim\,\tilde{\Omega}_K^\frac{1}{2}\,\sim\,
\frac{1}{2}\Omega_b^\frac{1}{2}$ and
(\ref{a10ba}) may be satisfied since
$\tilde{\Omega}_1^\frac{1}{2}$=
$\xi_1\,\simeq\,0.98$$\sim\,1$.
We notice that
$(\tilde{\Omega}_1^\frac{1}{2}
+\Omega_b^\frac{1}{2}+\Omega_r^\frac{1}{2}
+\tilde{\Omega}_x^\frac{1}{2}
+\tilde{\Omega}_K^\frac{1}{2})^2
\,\neq\,1$. However one may
define a total density parameter for the dark sector by
\begin{equation}
\Omega_D^\frac{1}{2}=
\tilde{\Omega}_1^\frac{1}{2}
-\tilde{\Omega}_x^\frac{1}{2}
-\tilde{\Omega}_K^\frac{1}{2}
\end{equation}
Then the density parameters satisfies the necessary condition,
$(\Omega_D^\frac{1}{2}+\Omega_b^\frac{1}{2}+\Omega_r^\frac{1}{2})^2=1$.
In other words, $H_{1n}$ and $H_\Delta$
terms can not be identified as separate contributions to dark sector,
rather they must be considered as just a single
object in order not to introduce an ambiguity in their identification.
\subsection{Compatibility with Recombination and Nucleosynthesis}
In this subsection we investigate if this model is compatible with
the
cosmological depiction of the recombination and nucleosynthesis, at
least, at the order of magnitude level. In a
similar vein as the preceding subsection we require this model mimic the
standard model, $\Lambda$CDM, as much as possible. We assume that the
radiation
and the baryonic matter are in thermal equilibrium in the eras of
recombination and nucleosynthesis since we adopt the same equations of
thermal equilibrium as $\Lambda$CDM. Therefore, in the following, first we
drive the condition for thermal equilibrium for this model.
Then we find the sets of parameters with least $\chi^2$ values that may
produce
successful recombination and nucleosynthesis eras. The correct choices
should have sufficient radiation energy densities in these eras. In other
words the redshift at the recombination time, $z_{re}$ should be
in the order of $(1+z_{re})^4\,>\,
(1+z_*)^4\,\simeq\,(1100)^4$ where
* denotes time of
last scattering surface; and in the nucleosynthesis era
the energy density of neutrinos should reach energy densities of the order
of $(1\,MeV)^4$. We seek an approximate, rough agreement with $\Lambda$CDM
since the search of the parameter space is done by trial and error rather
than a systematic search of the whole parameter space. Therefore a
detailed, thorough analysis and compatibility survey would be too
ambitious especially considering this is a toy model.
Before checking if there exist a set of parameters compatible with
recombination and nucleosynthesis we should check if the thermal
equilibrium is maintained in these eras in for the given set of parameters
because we adopt the standard analysis in $\Lambda$CDM, and that analysis
assumes existence of thermal equilibrium. As is well known, if there is
thermal equilibrium then we should have $\Gamma\,>\,H$
where $\Gamma$ is the rate of the interaction between radiation and
the matter and $H$ is the Hubble parameter.
However the implementation of this condition in this model is not
exactly the same
as in $\Lambda$CDM.
In the case of recombination era the
implementation of this condition does not give exactly the same result as
$\Lambda$CDM since , in $\Lambda$CDM the recombination takes place in
radiation dominated era and the total energy density is almost wholly
due to radiation
while, in this model, the total energy density of the universe at
this era is not almost wholly due to radiation although the equation
of state parameter for phenomenologically relevant cases is
similar that of radiation dominated universe at the time of
recombination and
we require the radiation
energy density to be the same or almost the same as $\Lambda$CDM. In
the case of nucleosynthesis, even the equation state parameter in
this model does not mimic that of a radiation dominated universe.
Therefore we should derive the corresponding conditions for thermal
equilibrium for this model.
The condition for thermal equilibrium in the recombination era
is
\begin{eqnarray}
&&\Gamma\,>\,H~\Rightarrow\,~1.97\times\,10^{-19}s^{-1}\times\,0.0227\,
(\frac{T}{T_{ph\,0}})\,>\,\alpha_1\,H_0\Omega_{ph}^\frac{1}{2}
\nonumber \\
&&\Rightarrow~
\left(\frac{\frac{\rho}{\rho_0}}{\frac{\rho_{ph}}{\rho_{ph0}}}
\right)^\frac{1}{2}\,<\,
0.19\,\left(\frac{T}{T_{ph\,0}}\right) =0.19(1+z_{re})\nonumber \\
\label{rn5}
\end{eqnarray}
Here we have used the identities,
\begin{equation}
\alpha_1^2=\frac{\rho}{\rho_{ph}}
=\frac{1}{\Omega_{ph}}\left(\frac{\frac{\rho}{\rho_0}}
{\frac{\rho_{ph}}{\rho_{ph0}}}\right) ~,~~
\frac{\rho_r}{\rho_{r0}}=\frac{\rho_{ph}}{\rho_{ph0}}=a^{-4}
\label{rn5a}
\end{equation}
where $\alpha_1\simeq\,1$ is the $\Lambda$CDM value and $\alpha_1^2\leq\,1$
at the time of recombination and is not constant in this model,
and I have used the PDG values,
$H_0=72\,km\,Mpc^{-1}\,s^{-1}$,
$\Omega_{ph}=4.8\times\,10^{-5}$. Note that $\Gamma$ in (\ref{rn5}) is the same
as the $\Lambda$CDM value while $H$ is different from
the $\Lambda$CDM value.
Next consider the condition on thermal equilibrium at and before the time
of nucleosynthesis.
In thermal equilibrium we have
\begin{eqnarray}
&&\frac{\Gamma_\nu}{H}\,\approx\,\frac{1}{\alpha_2}\frac{\sqrt{45}G_{wk}^2}
{64\pi^3}\sqrt{\frac{hc^5}{G}}
(k_BT)^3
\,\simeq\,\frac{1}{\alpha_2}\left(\frac{T}{10^{10}K}\right)^3\,>\,1
\nonumber \\
&&\Rightarrow\,
~~
\left(\frac{\frac{\rho}{\rho_0}}{\frac{\rho_r}{\rho_{r0}}}
\right)^\frac{1}{2}\,<\,
\Omega_r^\frac{1}{2}\left(\frac{T}{10^{10}K}\right)^3
\label{rn7}
\end{eqnarray}
where we have used the identity similar to (\ref{rn5a}), where $\alpha_1$
and the subindex $ph$ are replaced by $\alpha_2$ and $r$, respectively and the
ratio is evaluated at the time of nucleosynthesis. In this case, as well,
$\Gamma_\nu$ is the same as its $\Lambda$CDM value while the expression for
$H$ in terms of temperature is different since $\alpha_2\leq\,1$ and is
not a constant (i.e. it gives a different value when evaluated at different
time during nucleosynthesis) in this
model while $\alpha_2=1$ in $\Lambda$CDM.
During
thermal equilibrium the ratio of neutrinos to all nucleons, $X_n$ is given
by
\begin{equation}
X_n\,=\,\frac{1}{1+\exp{(\frac{Q}{k_BT})}} \label{rn8})
\end{equation}
where $Q$ is the rest mass energy difference between a neutron and
a proton, $Q=m_n-m_p=1.239\, MeV$. After the thermal equilibrium between
the neutrinos and the nucleons are lost i.e. after decoupling the value of
$X_n$ further decreases due to decay of free neutrons as
\begin{equation}
X_n\,=\,X_{n0}\exp{[-(\frac{t}{\tau_0})]} \label{rn9}
\end{equation}
where $X_{n0}$ is the $X_n$ of Eq.(\ref{rn8}) at the time of decoupling,
and $\tau_0=885.7$ seconds is the lifetime of a free neutron.
Therefore the effect of this model is to change the value of
$X_{n0}$ (that depends on $\alpha_2$) and probably the
value of $X_n$ as well.
Now we are ready to check the viability of this model.
I could give only four graphs
and three tables that partially
summarize the results of my calculations
related to this and the next paragraphs in order
not to expand the size of the paper too much.
Otherwise the size of the manuscript would be
almost doubled. First we check the viability of the
model for recombination and nucleosynthesis eras.
To this end I have used the equations
(\ref{a11a},\ref{a16c},\ref{a3db},\ref{a3dc}) in the
zeroth order approximation where
$a(t)\simeq\,a_{1n}(t)$ (as discussed before Eq.(\ref{a3ba}))
to draw $\omega$, $\frac{\rho_r}{\rho_{r0}}$,
versus time graphs by using a Mathematica code
that I have prepared for this aim for the sets of the parameters,
$r$,$s$,$\beta$,$\xi\xi_1$, $A_1$, $A_2$, $c_1$, $c_2$,
$\frac{\alpha_r}{\alpha_b}$,$\frac{\alpha_x}{\alpha_b}$,
$\frac{\alpha_K}{\alpha_b}$, $\alpha_c$,$\alpha_{o1}$,
$\alpha_{o2}$, $t_0$ $\Omega_b^\frac{1}{2}$, that correspond
to some relatively small $\chi^2$
values obtained in preceding subsection.
Then I have tried to find at least one set of parameters with
phenomenologically viable $\omega_0$, $\frac{\rho_r}{\rho_{r0}}$,
$\frac{\rho}{\rho_0}$ values i.e. $\omega_0$, in the range
$-0.68~$ \textemdash $~-0.74$;
$\frac{\rho_r}{\rho_{r0}}\,>\,(1100)^2\simeq\,10^{12}$ (in the range
of redshifts $z\sim\,800~-~3000$) at the time of
recombination, and
$\frac{\rho_r}{\rho_{r0}}\,>\,\frac{(1\,MeV)^4}
{5\times\,10^{-5}\,(2.5\times\,10^{-3}\,eV)^4}\,>\,10^{38}$
at the time of nucleosynthesis where I
have approximated $x_1(t)$, $x_2(t)$, $x_3(t)$
by $x_1^{(0)}(t)$, $x_2^{(0)}(t)$, $x_3^{(0)}(t)$
(that are defined in Eq.(\ref{a3ba})), and $a(t)$
by $a_{1n}(t)$ (that is defined in (\ref{a13b})) as
discussed in the preceding section. I have found two sets
of parameters given in Table \ref{table-9} that
satisfy these conditions.
A comment is in order at this point.
The zeroth order
approximation is reliable only for small redshifts. However
this approximation is reliable at any redshift if one is
only interested in the energy density - redshift relation.
This may be seen as follows:
Assume that the energy density $\rho$ is related to redshift
$z$ by $\rho\,=\,f(z)$ in the zeroth order approximation
(where $f(z)$ is an arbitrary function), and
in an approximation
better than the zeroth order we have
$\frac{c_0A_1}{A_1-A_2}=\frac{1}{x}$ i.e. $a(t)=\frac{1}{x}a_{1n}(t)$.
Then the energy density after the correction
is $\rho^\prime=f(z^{\prime})$.
If one rescales $z^\prime$ as $\frac{1}{x}z^\prime=z$ then one
obtains the same redshift and energy density values. In other words the
redshift - energy density relation is invariant under such corrections.
However this is not true for the redshift - time relation. If
the approximation is not a good approximation to the true value then
the redshift - time relation will be distorted. This, in turn, may cause
the distortion of the value of the equation of state and the distortion
of the variation of the energy densities with time in an amount
depending on the reliability of zeroth order approximation. Keeping these
observations in mind I content to use zeroth order approximation for the times
of recombination and nucleosynthesis because even employing zeroth order
approximation needs a lot of computer CPU and RAM, and in many cases
the use of first order approximation as well does not improve the situation.
We will come back to these points when discuss the times of recombination and
nucleosynthesis.
Next I have checked if thermal equilibrium is maintained at the times of
recombination and nucleosynthesis and if recombination and nucleosynthesis are
realized in this model.
One may have an idea on thermal equilibrium at the time of recombination by
using the values of Table \ref{table-9} at $z\simeq\,1100$ and
Eq.(\ref{rn5}). However a more rigorous way is to draw
$\left(
\frac{\frac{\rho}{\rho_0}}{\frac{\rho_r}{\rho_{r0}}}\right)^\frac{1}{2}
\left(\frac{1}{0.19
\frac{T}{T_{ph0}}}\right)$ (that may be obtained from
Eq.(\ref{rn5})) versus time graphs to determine the time intervals (and
then the corresponding redshift intervals) where
$\left(\frac{\frac{\rho}{\rho_0}}{\frac{\rho_r}{\rho_{r0}}}\right)^\frac{1}{2}
\left(\frac{1}{0.19\frac{T}{T_{ph0}}}\right)\leq\,1$ for each of the sets
A and B. In fact I have used $\frac{T}{T_{ph0}}\,=\,1+z$ for the
relevant redshifts. The resulting intervals are the intervals where thermal
equilibrium is maintained as shown in Figure \ref{fig:15} for the set B
in Table \ref{table-9}. The smallest redshifts where the thermal
equilibrium is lost are $z=2317$ ($\gamma=1.615\times\,10^{-10}$
with $\frac{\rho_r}{\rho_{r0}}\,\simeq\,5.52\times\,10^{13}$) and
$z=1625$ ($\gamma=2.82\times\,10^{-10}$
$\frac{\rho_r}{\rho_{r0}}\,\simeq\,3\times\,10^{13}$) for the sets A and B,
respectively. This implies that the photon electron decoupling
takes place before the time of last scattering at an energy of
$\simeq\,2317\times\,6\times\,10^{-4}$ eV$\,\simeq\,1.4$ eV and
$\simeq\,1625\times\,6\times\,10^{-4}$ eV$\,\simeq\,0.98$ eV for the sets
A and B, respectively (assuming the transition being directly to the
ground state of hydrogen atom) to be compared to the value of photon energy
of about $\simeq\,1100\times6\times\,10^{-4}$ eV$\,\simeq\,0.66$ eV for
$\Lambda$CDM at the time of last scattering.
This, in turn, implies that photon electron decoupling
in this model for the sets of
parameters A and B is at a smaller redshift than
$\Lambda$CDM where thermal equilibrium is maintained until decoupling.
(Thermal equilibrium would be maintained till $z\simeq\,2.4$ in
$\Lambda$CDM if recombination of electrons and protons to form neutral
atoms had not taken place as may be seen from (\ref{rn5}) by setting
$\alpha_1=1$). In fact the corresponding times for decoupling are already
smaller than that of $\Lambda$CDM by five orders of magnitude. A detailed
comprehensive separate study is need to see if these imply some
interesting phenomenologically viable alternatives or just an artifact of the toy
model and/or the sets of parameters considered. This may also be due to the limitation
of the applicability of zeroth order approximation that we have discussed above.
$a(t)\simeq\,a_{1n}(t)$ is not violated badly at the
time of recombination for the most of the relevant sets of
parameters. For example for the sets of parameters given in
Table \ref{table-9} the first order approximation results in
$a(t)\simeq\,0.4\,a_{1n}(t)$ i.e.
$\frac{c_0A_1}{A_1-A_2}\simeq\,0.4$ and does not vary much at
the time of recombination. Therefore it seems that the effect of the
limitation of the applicability of zeroth order approximation to the time of recombination
must be limited.
However this shift does not
introduce a major problem since the redshift values, hence
the photon energy density at recombination remains almost the same
and thermal equilibrium is maintained.
Next I have checked if thermal equilibrium is maintained
at the peaks in Table \ref{table-9} where the energy densities
are sufficient for nucleosynthesis.
I have used Eq.(\ref{rn7}) to find the range of
temperatures where thermal equilibrium is maintained.
I have found that this condition is
satisfied for $T\,>\,3\times\,10^{10}\,K$ (provided that
$\Omega_r\simeq\,5\times\,10^{-5}$) for the second peaks. This value gives
us $X_{n0}$ in (\ref{rn9}) by using (\ref{rn8}) as $X_{n0}\simeq\,0.39$
which is quite large compared to the $\Lambda$CDM value of
$\simeq\,0.25$. The time that takes $3\times\,10^{10}K\simeq\,1\,MeV$
drop to 0.07 MeV (that is when
$\frac{\rho_r}{\rho_{r0}}\,\sim\,10^{32}$) in this model is something like
$\sim\,2\times\,10^{-16}\times\,t_0\simeq\,90$ seconds. Therefore
$X_{n0}$ does not drop significantly through Eq.(\ref{rn9}). In other
words the final result $X_{n}\simeq\,0.35$ is much larger than the
$\Lambda$CDM value $\simeq\,0.13$ (which agrees well with observations).
Probably the main source of this discrepancy is inapplicability of zeroth order
approximations to redhifts and energy densities to this era to obtain
correct energy density - time relations.
The variations of $\frac{c_0A_1}{A_1-A_2}$ and $A$ are quite
large and their values are quite different
than those at $z\sim\,0$
at the time of nucleosynthesis that makes the
applicability of the zeroth order approximation extremely difficult to
obtain correct energy density - time relation. In other words the
main source of the discrepancy may be due to the fact that
the real time of free decay
may be of the order of $\sim\,1000$ seconds in this model
instead of 90 seconds.
The use of first order approximation does not improve the situation
because in the calculation of first order approximations to
$\frac{c_0A_1}{A_1-A_2}$ and $A(t)$
one uses the zeroth order approximation $a(t)\,\simeq\,a_{1n}(t)$ in
the integrals for $x_i$, $i=1,2,3$. This, in turn, results in
over contribution of large redshifts and hence
larger and more varying $x_i$ with respect to their true values since
$x_i\,<1$ and they get smaller i.e.
$\frac{c_0A_1}{A_1-A_2}$ gets larger
at larger redshifts.
Therefore the energy density versus time graphs in
the figures \ref{fig:9}, \ref{fig:10} and \ref{fig:15} must
be considered with some care: The time values in those graphs
should be taken with utmost care
especially in the case of nucleosynthesis
while the magnitudes of energy
densities and the corresponding redshifts are expected to
be the same as the exact values.
All these points must be studied in more detail in future studies.
However I have been able to show that this scheme can produce a model
that mimics the standard model:
There is a current accelerated epoch whose present equation of
state (for the whole universe) is ~-0.7 (that is, at least,
roughly in agreement with observations e.g. see the value in Table \ref{table-9}
for a phenomenologically relevant set of parameters).
Before this epoch $\omega$ changes sign
and the time near this sign change may be considered as the matter
dominated era. Although the sign change of $\omega$ occurs at a later time
in this model compared to $\Lambda$CDM the time and the
redshift of onset of the
accelerated era (i.e. $\omega\simeq\,-\frac{1}{3}$) are comparable with
those of $\Lambda$CDM.
There is an epoch before the matter dominated era
where $\omega$ is on average close to
$\frac{1}{3}$, and may be identified by radiation dominated era,
and the time of the maximum value of $\omega$ may be considered as the time when
the universe was like stiff matter or denser (as
in the cores of stars). Then $\omega$ changes sign
again reaches to two minima peaks as mentioned before and eventually
approaches to -1 as time goes to zero
(due to the $H_{1n}$, in particular the first part of
it) and
this epoch probably may be considered as the inflationary era.
Moreover the model is able to give relatively small reduced
$\chi_0^2$ and $\chi^2$ values for Union2.1 data set, and it
can, at least roughly, account for
recombination and nucleosynthesis times. I think
this is a sufficiently well starting point for a toy model whose
main aim is to embody the creation of matter
and radiation in the scale
factor of Robertson-Walker metric. However
there is a great deal of points to be clarified and addressed in future
studies such as checking the whole parameter space of this model by using
a more elaborate software and to use more powerful computers that may
give scan the whole parameter space in a better approximation than the one
given here, and considering a more detailed analysis of the recombination
and nucleosynthesis epochs, studying the evolution of cosmological
perturbations in this model, and considering
possible extensions of this model towards a more realistic model.
\section{Conclusion}
In this study a scheme for obtaining a scale factor (in Robertson-Walker
metric) that may account for the times before, during, and after the
radiation dominated eras is introduced. The prescription to
obtain the scale factor in this model is quite simple; First one
introduces a scale factor for the pure dark sector, and then the full
scale factor is obtained by a relation between these two scale factors.
The
result is a scheme to produce the scale factor for the whole universe,
including baryonic matter, radiation, and dark energy-matter (i.e. dark
sector) in such a way that the times before, during, and after radiation
dominated era are expressed by a single scale factor in Robertson-Walker
metric. Different choices of the pure dark sector scale factor (denoted
by $a_1$ in this paper) and different choices of the relation between
$a_1(t)$ and the scale factor of the full universe, $a(t)$ give different
models. As an illustration of this scheme a model with a specific scale
factor for the
pure dark sector and a specific relation between $a_1(t)$ and $a(t)$ is
considered. The phenomenological viability of this model is checked
through its
compatibility with Union2.1 data set, and with recombination and
nucleosynthesis by using trial error and Mathematica
software for almost randomly chosen sets of parameters. Two sets of
parameters with relatively small $\chi^2$ values for Union2.1 data set,
and that are compatible with successful recombination and nucleosynthesis
at an order of
magnitude level are found. These results are encouraging in view of the
fact that only a tiny portion of the whole parameter space could be
considered in this way. A separate, detailed, and comprehensive
computational
project with more advanced software codes and/or powerful
computing facilities that may scan the full parameter space and may employ
better approximation schemes is needed to reach a definite view on the
observational viability of this scheme and/or this model. Moreover the
effect of this model on cosmological perturbations should be considered
and possible implications and extensions of this
scheme to inflationary era should be studied in future. Furthermore
different pure dark sector scale factors and different options to relate
pure dark sector and the full universe scale factors may be
considered in future to see the full range of possibilities that this
scheme may offer.
\begin{acknowledgments}
I would like to thank Professor Joan Sol$\grave{a}$ for reading the manuscript
and for his valuable comments. I would also like to thank Professor
Gregory Gabadadze for reading the manuscript, and to
Dr. A. Aviles for his help in the initial
phase of writing the Mathematica code used in this study.
\end{acknowledgments}
|
\section{Introduction}
In this paper we study a model problem of homogenization for a discrete crystalline flow. The analysis will be
carried over by using the minimizing-movement scheme of Almgren, Taylor and Wang \cite{ATW83}
(later thus renamed by De Giorgi, see e.g.~\cite{AGS}). This consists in introducing a time scale $\tau$, iteratively defining
a sequence of sets $E^\tau_k$ as minimizers of
\begin{equation}
\min \Bigl\{ P(E)+{1\over \tau} D(E, E^\tau_{k-1})\Bigr\} ,
\end{equation}
where $P$ is a perimeter energy and $D$ is a (suitably defined) distance-type energy between sets, and $E^\tau_0$
is a given initial datum, and subsequently computing a time-continuous limit $E(t)$ of $\{E^\tau_k\}$ as $\tau\to0$, which defines the desired geometric motion related to the energy~$P$.
The study of geometric motions in inhomogeneous
environments has a very large literature (see e.g.~\cite{BC, DY, Gl, LS}). The ones in a discrete setting can be
considered a somewhat extreme case, in that the corresponding energies possess a large number of local minimizers
(actually, by the discrete nature of the problem all states are local minimizers), while on the contrary their continuum limits
(see e.g.~\cite{ABC,BraPia2} for a rigorous definition) possess no local (non global) minimizer. As a consequence, gradient flows tend to be ``pinned''
(i.e., the resulting limit $E(t)$ is constant), in contrast with the formal evolution of their limit continuous energies,
to which the Almgren-Taylor-Wang approach can be used to obtain a non trivial evolution (for the case of $P$ the crystalline perimeter in two dimensions see \cite{AT95}).
In a recent paper by Braides, Gelli and Novaga \cite{BGN} the Almgren-Taylor-Wang approach has been used coupled to a homogenization procedure. In this case the perimeters (and the distances) depend on a small parameter $\epsilon$, and consequently, after introducing a time scale $\tau$, the time-discrete motions are the $E^{\tau,\epsilon}_k$ defined iteratively by
\begin{equation}
E^{\tau,\epsilon}_k \hbox{ is a minimizer of } \min \Bigl\{ P_\epsilon(E)+{1\over \tau} D(E, E^{\tau,\epsilon}_{k-1})\Bigr\}.
\end{equation}
The time-continuous limit $E(t)$ of $\{E^{\tau,\epsilon}_k\}$ then may depend how mutually $\epsilon$ and $\tau$ tend to $0$.
This type of problems can be cast in the general framework of {\em minimizing movements along a $\Gamma$-converging sequence} (see \cite{Bra13}). In particular, if we have a large number of local minimizers then the limit motion will be pinned if $\tau<\!\!<\epsilon$ suitably fast (in a sense, we can pass to the limit in $\tau$ first, and then apply the Almgren-Taylor-Wang approach, which clearly gives pinning when the initial data are local minimizers). On the contrary, if $\epsilon<\!\!<\tau$ fast enough
and $P_\epsilon$ $\Gamma$-converge to a limit perimeter $P$ (which is always the case by compactness), then the limit $E$ will be
the evolution related to the limit $P$ (again, in a sense, in this case we can pass to the limit in $\epsilon$ first).
In \cite{BGN} the energies $P_\epsilon$ are {\em ferromagnetic energies} defined on subsets $E\subset \epsilon\mathbb{Z}^2$, of the form
$$
P_\epsilon(E)= {\alpha}\,\epsilon\, \#\{(i,j)\in\epsilon\mathbb{Z}^2\times\epsilon\mathbb{Z}^2: i\in E, j\not\in E, \ |i-j|=\epsilon\}
$$
($\alpha>0$ a positive parameter). The continuum limit of these energies can be proved to be the crystalline perimeter
$$
P(E)=\alpha\int_{\partial E}\|\nu\|_1d\H^1,
$$
where $\nu$ is the normal to $\partial E$ and $\|(\nu_1,\nu_2)\|_1=|\nu_1|+|\nu_2|$ (see \cite{ABC}). The {\em flat flow} of this perimeter is the motion by crystalline curvature described by Taylor \cite{Ta}. In the case of initial datum a coordinate rectangle, the evolution is a rectangle with the same centre and sides of lengths $L_1, L_2$ governed by the system of ordinary differential equations
$$
\begin{cases}\displaystyle \dot L_1= -{4\alpha\over L_2}\cr\cr
\displaystyle \dot L_2= -{4\alpha \over L_1}.\end{cases}
$$
In \cite{BGN} all possible evolutions have been characterized as $\epsilon, \tau\to 0$, showing that the relevant mutual scale is when $\tau/\epsilon\to\gamma\in(0,+\infty)$. In the case of initial datum a coordinate rectangle the resulting evolution is still a rectangle. In the case of a unique evolution, the side-lengths $L_1(t), L_2(t)$ of this rectangle are governed by a system of `degenerate' ordinary differential equations
$$
\begin{cases}\displaystyle \dot L_1= -{2\over \gamma}\Bigl\lfloor{2\gamma\alpha\over L_2}\Bigr\rfloor\\ \cr
\displaystyle \dot L_2= -{2\over \gamma}\Bigl\lfloor{2\gamma\alpha\over L_1}\Bigr\rfloor.\end{cases}
$$
Note that the right-hand sides are discontinuous; however existence (and uniqueness, except for some special cases) of solution can be easily proved by a direct argument. This characterization highlights the effect of the periodicity through the scaling $\gamma$ and that the motion is obtained by overcoming some energy barriers in a `quantized' manner by the presence of a discontinuous right-hand side. In particular, we have {\em pinning of large rectangles}: if both initial side-lengths are above the {\em pinning threshold} $\widetilde L=2\gamma\alpha$ then the right-hand sides are zero and the motion is pinned. The limit cases (total pinning and continuous crystalline flow) correspond to the limit values $\gamma=0$ and $\gamma=+\infty$. This analysis shows that the ``correct scaling'' for this problem is $\epsilon\sim\tau$, which gives the most information about all the limit evolutions.
The analysis described above exhibits a limit evolution in which we may read the effect of the $\Gamma$-limit energy (through the crystalline form of the evolution and the coefficient $\alpha$) and of the interplay between the time and space scales through the scaling $\gamma$. Scope of this work is to show that in general the situation can be more complex, and the periodic microstructure can affect the limit evolution without changing the $\Gamma$-limit. To this end we will introduce a further inhomogeneity in the perimeters $P_\epsilon$ by considering
$$
P_\epsilon(E)= {1\over 2}\epsilon\,\sum \{c_{ij}:i,j\in \mathbb{Z}^2,\epsilon i\in E, \epsilon j\not\in E, \ |i-j|=1\},
$$
(we use the notation $\sum \{ x_a: a\in A\}=\sum_{a\in A} x_a$)
where the coefficients $c_{ij}$ equal $\alpha$ except for some well-separated periodic square
inclusions where $c_{ij}=\beta>\alpha$.
These inclusions are not energetically favorable and they can be neglected in the computation of the $\Gamma$-limit, which is still the perimeter $P$ above, with the same coefficient $\alpha$. They can be considered as ``obstacles'' that can be bypassed when computing minimizers of $P_\epsilon$; however their presence is felt in the minimizing-movement procedure since they may influence the choice of $E^{\tau,\epsilon}_k$ through the interplay between the distance and perimeter terms. As a result, the motion can be either decelerated or accelerated with respect to the homogeneous case.
As already remarked in \cite{BGN} the relevant case for the description of the motion is that of initial data coordinate
rectangles, since all other cases can be reduced to the study of this one. We will then restrict our analysis to that case.
This (apparently) simple situation already contains all the relevant features of the evolution and highlights the differences with respect to \cite{BGN}. We will show that the limit motion can still be described through a system of degenerate ordinary differential equations of the form
$$
\begin{cases}\displaystyle \dot L_1= -{2\over \gamma}\,f\Bigl({\gamma\over L_2}\Bigr)\cr \cr
\displaystyle \dot L_2= -{2\over \gamma}\,f\Bigl({\gamma\over L_1}\Bigr)\end{cases}
$$
with $f$ a locally constant function on compact subsets of $(0,+\infty)$ which depends on $\alpha$, the period and size of the inclusions but not on $\gamma$ (neither on the value $\beta$). The {\em effective velocity} $f$ is obtained by a homogenization formula which optimizes the motion of the sides of the rectangle, resulting in an oscillation around a linear motion with velocity ${1\over\gamma} f(\gamma/L)$ (which is locally constant as noted above). Note that, in the case of no inclusion, the system is of the same form with $f(Y)=\lfloor 2\alpha Y\rfloor$.
The dependence on the inclusions gives a new pinning threshold
$$
\overline L={4\gamma\alpha\over 2+N_\beta}
$$
depending on the size of the inclusion $N_\beta$.
The reason for this new pinning threshold is that, in order that a side may move, it needs to be able to overcome a barrier of $N_\beta$ inclusions. Note that, if the initial data have side-lengths $\overline L< L<\widetilde L$, then we may have a microscopic motion which stops after a finite number of time steps, and is not eventually detected in the limit. It should be remarked that the presence of the inclusions may indeed accelerate the motion, so that $f(Y)>\lfloor 2\alpha Y\rfloor$ for some $Y$.
\bigskip
The paper is organized as follows. In Section~\ref{setting} we define all the energies that we will consider. We then formulate the discrete-in-time scheme analogous to the Almgren, Taylor and Wang approach. Section~\ref{rectangular} contains the proof of the convergence of the discrete scheme in the case of a rectangular initial set. Contrary to the case in \cite{BGN} it is not trivial to show that the minimizers of this scheme are actually rectangles. This is a technical result contained in Proposition~\ref{rectangleprop}. Subsection~\ref{newpinning} contains the computation of the new pinning threshold, showing that it depends on the percentage $N_\beta$ of defects in the lattice. Subsection~\ref{newvelocity} deals with the new definition of the effective velocity of a side by means of a homogenization formula resulting from a one-dimensional `oscillation-optimization' problem. This velocity can be expressed uniquely (up possibly to a discrete set of values), as a function the ratio of $\gamma$ and the side-length (Definition~\ref{effvel}). The description of the homogenized limit motion is contained in Subsection~\ref{limitmotion}. In the last Section 4 we explicitly compute the velocity function by means of algebraic formulas in some simple cases, showing a nontrivial comparison with the case with no inclusions.
\section{Setting of the problem}\label{setting}
If $x=(x_1,x_2)\in \mathbb{R}^2$ we set $\|x\|_1=|x_1|+|x_2|$ and $\|x\|_\infty=\max\{|x_1|,|x_2|\}$. If $A$ is a Le\-bes\-gue\hbox{-}measurable set we denote by $|A|$ its two\hbox{-}dimensional Lebesgue measure. The symmetric difference of $A$ and $B$ is denoted by $A\triangle B$, their Hausdorff distance by $\text{d}_\mathcal{H}(A,B)$.
If $E$ is a set of finite perimeter then $\partial^*E$ is its reduced boundary (see, for example \cite{Bra98}). The
measure-theoretical inner normal to $E$ at a point $x$ in $\partial^*E$ is denoted by $\nu=\nu_E(x)$ .
\subsection{Inhomogeneous ferromagnetic energies}
\defN_{\alpha\beta}{N_{\alpha\beta}}
The energies we consider are interfacial energies defined in an inhomogeneous environment as follows: let $0<\alpha<\beta<+\infty$, $N_\alpha,N_\beta\geq1$ and set $N_{\alpha\beta}=N_\alpha+N_\beta$. We consider the $N_{\alpha\beta}$-periodic coefficients $c_{ij}$ indexed on {\em nearest-neighbours} of $\mathbb{Z}^2$ (i.e., $i,j\in \mathbb{Z}^2$ with $|i-j|=1$) defined for $i,j$ such that
$$
0\le {i_1+j_1\over 2}, {i_2+j_2\over 2}<N_{\alpha\beta}
$$
by
\begin{equation}
c_{ij}= \begin{cases} \beta &\hbox{if $\displaystyle0\le {i_1+j_1\over 2}, {i_2+j_2\over 2}\le N_\beta$}\\
\alpha &\hbox{otherwise.}\end{cases}
\end{equation}
These coefficients label the bonds between points in $\mathbb{Z}^2$, so that they describe a matrix of $\alpha$-bonds with $N_{\alpha\beta}$-periodic inclusions of $\beta$-bonds grouped in squares of side-length $N_\beta$. The periodicity cell is pictured in Fig.~\ref{fig:0}.
\begin{figure}[htbp]
\centering
\def190pt{100pt}
\input{fig0.pdf_tex}
\caption{Periodicity cell. Continuous lines represent $\beta$\hbox{-}bonds, dashed lines $\alpha$\hbox{-}bonds.}\label{fig:0}
\end{figure}
Correspondingly, to these coefficients we associate the energy defined on subsets $\mathcal{I}$ of $\mathbb{Z}^2$ by
\begin{equation}
\P^{\alpha,\beta}(\mathcal{I})=\sum\Bigl\{c_{ij}: |i-j|=1, i\in \mathcal{I}, j\in \mathbb{Z}^2\setminus \mathcal{I}\Bigr\}.
\end{equation}
As recalled in the Introduction we use the notation $\sum \{ x_a: a\in A\}=\sum_{a\in A} x_a$.
In order to examine the overall properties of $\P^{\alpha,\beta}$ we introduce the family of scaled energies defined on subsets $\mathcal{I}$ of $\epsilon\mathbb{Z}^2$ by
\begin{equation}\label{pabe}
\P^{\alpha,\beta}_\epsilon(\mathcal{I})=\sum\Bigl\{\epsilon\ c_{{i/\epsilon}\,{ j/\epsilon}}: |i-j|=\epsilon, i\in \mathcal{I}, j\in \epsilon\mathbb{Z}^2\setminus \mathcal{I}\Bigr\};
\end{equation}
i.e., $\P^{\alpha,\beta}_\epsilon(\mathcal{I})=\epsilon\, \P^{\alpha,\beta}({1\over\epsilon}\mathcal{I})$.
To study the continuous limit as $\epsilon\to0$ of these energies it will be convenient to identify each subset of $\epsilon\mathbb{Z}^2$ with a measurable subset of $\mathbb{R}^2$, in such a way that equi-boundedness of the energies implies pre-compactness of such sets in the sense of sets of finite perimeter. This identification is as follows: we denote by $Q$ the closed coordinate unit square of center $0$, $Q=[-1/2,1/2]^2$; if $\epsilon>0$ and $ {i}\in\epsilon\mathbb{Z}^2$, we denote by $Q_\epsilon( {i})= {i}+\epsilon Q$ the closed coordinate square with side-length $\epsilon$ and centered in $ {i}$. To a set of indices $\mathcal{I}\subset\epsilon\mathbb{Z}^2$ we associate the set
\begin{equation*}
E_{\mathcal{I}}=\bigcup_{ {i}\in \mathcal{I}}Q_\epsilon( {i}).
\end{equation*}
\\
The space of \emph{admissible sets} related to indices in the two\hbox{-}dimensional square lattice is then defined by
\begin{equation*}
\mathcal{D}_\epsilon:=\left\{E\subseteq\mathbb{R}^2:\quad E=E_\mathcal{I}\text{ for some $\mathcal{I}\subseteq\epsilon\mathbb{Z}^2$}\right\}.
\end{equation*}
For each $E=E_\mathcal{I}\in \mathcal{D}_\epsilon$ we denote
\begin{equation}\label{pabei}
\P^{\alpha,\beta}_\epsilon(E)=\P^{\alpha,\beta}_\epsilon(\mathcal{I}).
\end{equation}
As an easy remark, we note that
\begin{equation}\label{pabes}
\P^{\alpha,\beta}_\epsilon(E)\ge\epsilon\alpha\#\Bigl\{(i,j): |i-j|=\epsilon, i\in \mathcal{I}, j\in \epsilon\mathbb{Z}^2\setminus \mathcal{I}\Bigr\}= \alpha \H^1(\partial E),
\end{equation}
which shows that sequences of sets $E_\epsilon$ with $\sup_\epsilon\P^{\alpha,\beta}_\epsilon(E_\epsilon)<+\infty$ are pre-compact with respect
to the local $L^1$-convergence in $\mathbb{R}^2$ of their characteristic function and their limits are sets of finite perimeter in $\mathbb{R}^2$.
Hence, this defines a meaningful convergence with respect to which compute the $\Gamma$-limit of $\P^{\alpha,\beta}_\epsilon$ as $\epsilon\to 0$.
A general theory for the homogenization of energies (\ref{pabe}), in a more general context, has been developed in \cite{BraPia2} (see also \cite{BCS,BraPia3,BraSo1}), where it is shown that the $\Gamma$-limit's domain is precisely the family of sets of finite perimeter and its general form is
$$
F(E)=\int_{\partial^*E}\varphi(\nu)d\H^1,
$$
with $\varphi$ a convex function positively homogeneous of degree one.
The computation in the case $\alpha=\beta$ (homogeneous spin systems) can be found in \cite{ABC} and gives
$\varphi(\nu)=\alpha\|\nu\|_1$. In our case the presence of the $\beta$-inclusions does not influence the form of the $\Gamma$-limit, as in the following remark.
\begin{oss} [$\Gamma$\hbox{-}convergence of inhomogeneous perimeter energies]
The energies $\text{P}_\epsilon^{\alpha,\beta}$ defined by (\ref{pabe}) $\Gamma$\hbox{-}converge, as $\epsilon\to0$, to the anisotropic crystalline perimeter functional
\begin{equation*}
\P^\alpha(E)=\alpha\int_{\partial^* E}\|\nu\|_1\,d\H^1.
\end{equation*}
This limit is independent of $N_\alpha,N_\beta$, and equals the one obtained when $\beta=\alpha$.
The lower bound for the $\Gamma$-limit is immediately obtained from the case $\alpha=\beta$ in \cite{ABC} after remarking that $\P^{\alpha,\beta}_\epsilon\geq \P_\epsilon^{\alpha,\alpha}$. In order to verify the upper bound, it suffices to note that recovery sequences for the $\Gamma$-limit of $\P_\epsilon^{\alpha,\alpha}$ can be constructed at a scale $N_{\alpha\beta}\epsilon$, thus `avoiding' the $\beta$-connections. To this end, define
$$
{Q}^{N_{\alpha\beta}}_\epsilon=\displaystyle\bigcup\Bigl\{Q_\epsilon( {i}):\
{ {i}\in\epsilon\mathbb{Z}^2},\ 0\leq\| {i}\|_{\infty}<\epsilon N_{\alpha\beta}\Bigr\}.
$$
This is a square of side-length $N_{\alpha\beta}\,\epsilon$ whose boundary intersects only $\alpha$-bonds.
We consider $\P^{N_{\alpha\beta}}_\epsilon$the restriction of $\P_\epsilon^{\alpha,\beta}$ to the class
\begin{equation*}
{\mathcal{D}}^{N_{\alpha\beta}}_\epsilon=\left\{E\subseteq\mathbb{R}^2:\quad E\text{ is a finite union of $\epsilon\mathbb{Z}^2$-translations of }{Q}^{N_{\alpha\beta}}_\epsilon\right\}.
\end{equation*}
Note that we have $\P_\epsilon^{\alpha,\beta}(E)=\P_\epsilon^{\alpha,\alpha}(E)$ for $E\in {\mathcal{D}}^{N_{\alpha\beta}}_\epsilon$, and that
sets in ${\mathcal{D}}^{N_{\alpha\beta}}_\epsilon$ differ from sets in ${\mathcal{D}}_{\epsilonN_{\alpha\beta}}$ by a fixed translation of order $\epsilon$. Hence, we have
(see \cite{GCB} for details on the properties of $\Gamma$-upper limits)
$$
\Gamma\hbox{-}\limsup_{\epsilon\to 0}\P_\epsilon^{\alpha,\beta}(E)\le
\Gamma\hbox{-}\limsup_{\epsilon\to 0}\P^{N_{\alpha\beta}}_\epsilon(E)=\Gamma\hbox{-}\lim_{\epsilon\to 0}\P_{N_{\alpha\beta}\epsilon}^{\alpha,\beta}(E),
$$
and the latter is again equal to $\P^\alpha(E)$. This inequality just states that we can take sets in ${\mathcal{D}}^{N_{\alpha\beta}}_\epsilon$ which are (small translations of) a recovery sequence for $\P_{N_{\alpha\beta}\epsilon}^{\alpha,\beta}(E)$ as a recovery sequence for $\P_\epsilon^{\alpha,\beta}(E)$.
\end{oss}
\subsection{A discrete-in-time minimization scheme}\label{timemin}
For $\mathcal{I}\subset\epsilon\mathbb{Z}^2$ we define the \emph{discrete $\ell^{\infty}$\hbox{-}distance} from $\partial\mathcal{I}$ as
\begin{equation*}
d_\infty^\epsilon( {i},\partial\mathcal{I})=
\begin{cases}
\inf\{\| {i}- {j}\|_\infty: {j}\in\mathcal{I}\}&\text{if $ {i}\not\in\mathcal{I}$}\\
\inf\{\| {i}- {j}\|_\infty: {j}\in\epsilon\mathbb{Z}^2\backslash\mathcal{I}\}&\text{if $ {i}\in\mathcal{I}$}.
\end{cases}
\end{equation*}
Note that we have
$\displaystyle
d_\infty^\epsilon( {i},\partial\mathcal{I})=d_\infty( {i},\partial E_\mathcal{I})+\frac{\epsilon}{2}$,
where $d_\infty$ denotes the usual $\ell^\infty$-distance. The distance can be extended to all $\mathbb{R}^2\backslash\partial E_\mathcal{I}$ by setting
\begin{equation*}
d_\infty^\epsilon(x,\partial\mathcal{I})=d_\infty^\epsilon( {i},\partial\mathcal{I})\quad \text{if }x\in Q_\epsilon( {i}).
\end{equation*}
\\
In the following we will directly work with $E\in\mathcal{D}_\epsilon$, so that the distance can be equivalently defined by
\begin{equation*}
d_\infty^\epsilon(x,\partial E)=d_\infty( {i},\partial E)+\frac{\epsilon}{2},\quad \text{if }x\in Q_\epsilon( {i}).
\end{equation*}
Note that this is well defined as a measurable function, since its definition is unique outside the union of the boundaries of the squares $Q_\epsilon$ (that are a negligible set).\\
We now fix a time step $\tau>0$ and introduce a discrete motion with underlying time step $\tau$ obtained by successive minimization. At each time step we will minimize an energy $\mathcal{F}_{\epsilon,\tau}^{\alpha,\beta}:\mathcal{D}_\epsilon\times\mathcal{D}_\epsilon\to\mathbb{R}$ defined as
\begin{equation}
\mathcal{F}_{\epsilon,\tau}^{\alpha,\beta}(E,F)= \P_\epsilon^{\alpha,\beta}(E)+\frac{1}{\tau}\int_{E\triangle F}d_\infty^\epsilon(x,\partial F)\,dx.
\label{newenergy}
\end{equation}
Note that the integral can be indeed rewritten as a sum on the set of indices $\epsilon\mathbb{Z}^2\cap(E\triangle F)$ (see \cite{BGN}).
Given an initial set $E^0_{\epsilon}$, we define recursively a sequence $E_{\epsilon,\tau}^k$ in $\mathcal{D}_\epsilon$ by requiring the following:
\begin{description}
\item[(i)] $E^0_{\epsilon,\tau}=E^0_{\epsilon}$;
\item[(ii)] $E_{\epsilon,\tau}^{k+1}$ is a minimizer of the functional $\mathcal{F}_{\epsilon,\tau}^{\alpha,\beta}(\cdot,E_{\epsilon,\tau}^k)$.
\end{description}
The \emph{discrete flat flow} associated to functionals $\mathcal{F}_{\epsilon,\tau}^{\alpha,\beta}$ is thus defined by
\begin{equation}\label{disefo}
E_{\epsilon,\tau}(t)=E_{\epsilon,\tau}^{\lfloor t\slash\tau\rfloor}.
\end{equation}
Assuming that the initial data $E^0_\epsilon$ tend, for instance in the Hausdorff sense, to a sufficiently regular set $E_0$, we are interested in identifying the motion described by any converging subsequence of $E_{\epsilon,\tau}(t)$ as $\epsilon,\tau\to0$.
As remarked in the Introduction, the interaction between the two discretization parameters, in time and space, plays a relevant role in such a limiting process. More precisely, the limit motion depends strongly on their relative decrease rate to 0. If $\epsilon\!<\!<\tau$, then we may first let $\epsilon\to0$, so that $\P_\epsilon^{\alpha,\beta}(E)$ can be directly replaced by the limit anisotropic perimeter $\P^\alpha(E)$ and $\frac{1}{\tau}\int_{E\bigtriangleup F}d_\infty^\epsilon(x,\partial F)\,dx$ by $\frac{1}{\tau}\int_{E\bigtriangleup F}d_\infty(x,\partial F)\,dx$. As a consequence the approximated flat motions tend to the solution of the continuous ones studied by {Almgren} and {Taylor} \cite{AT95}. On the other hand, if $\epsilon\!>\!>\tau$ then there is no motion and $E_{\epsilon,\tau}^k\equiv E^0_\epsilon$. Indeed, for any $F\neq E^0_\epsilon$ and for $\tau$ small enough we have
\begin{equation*}
\frac{1}{\tau}\int_{E^0_\epsilon\bigtriangleup F}d_\infty^\epsilon(x,\partial F)\,dx\geq c\frac{\epsilon}{\tau}>\P_\epsilon^{\alpha,\beta}(E^0_\epsilon).
\end{equation*}
In this case the limit motion is the constant state $E_0$. The meaningful regime is the intermediate case $\tau\sim\epsilon$.
\bigskip
\section{Motion of a rectangle}\label{rectangular}
As shown in \cite{BGN} the relevant case is when $\epsilon$ and $\tau$ are of the same order and the initial data are coordinate rectangles $E^0_\epsilon$, which will be the content of this section.
We assume that
\begin{equation*}
\tau=\gamma\epsilon\quad\text{for some }\gamma\in(0,+\infty),
\end{equation*}
and, correspondingly, we omit the dependence on $\tau$ in the notation of
\begin{equation*}
E^k_\epsilon=E^k_{\epsilon,\tau}(=E^k_{\epsilon,\gamma\epsilon}).
\end{equation*}
\smallskip
Due to the lack of uniqueness of minimizers in the discrete minimization scheme, a standard comparison principle cannot hold. We recall a \emph{weak comparison principle} for our motion in the discrete case (see \cite{BGN} for the proof).
\begin{prop}[Discrete weak comparison principle]\label{wcp}
Let $\epsilon>0$ and let $R_\epsilon,K_\epsilon\in\mathcal{D}_\epsilon$ be such that $R_\epsilon\subseteq K_\epsilon$ and $R_\epsilon$ is a coordinate rectangle. Let $K_\epsilon^k$ be a motion from $K_\epsilon$ constructed by successive minimizations. Then $R_\epsilon^k\subseteq K_\epsilon^k$ for all $k\geq1$, where $R_\epsilon^k$ is a motion from $R_\epsilon$ constructed by successively choosing a minimizer of $\mathcal{F}_{\epsilon,\tau}^{\alpha,\alpha}(\cdot,R_\epsilon^{k-1})$ having smallest measure.
\end{prop}
\begin{oss}\label{wpcoss}
The set $\mathbb{R}^2\backslash K_\epsilon^k$ is the $k$-step evolution of the complementary $\mathbb{R}^2\backslash K_\epsilon^k$ of $K_\epsilon$. As a consequence, if we have $R_\epsilon\subseteq\mathbb{R}^2\backslash K_\epsilon^k$, from Proposition \ref{wcp} it follows that
\begin{equation*}
R_\epsilon^k\subseteq\mathbb{R}^2\backslash K_\epsilon^k,\quad \text{for all $k\geq1$.}
\end{equation*}
\end{oss}
\begin{defn}[$\alpha$-type rectangle]
A coordinate rectangle whose sides intersect only $\alpha$\hbox{-}bonds will be called an \emph{$\alpha$-type rectangle}.
\end{defn}
The first result is that coordinate rectangles evolve into $\alpha$-type
rectangles.
\begin{prop}\label{rectangleprop}
If $E^0_{\epsilon}\in\mathcal{D}_\epsilon$ is a coordinate rectangle and $F$ is a minimizer for the minimum problem for $\mathcal{F}^{\alpha,\beta}_{\epsilon,\tau}(\cdot,E^k_\epsilon), k\geq0$, then for all $\delta>0$ $F$ is a coordinate $\alpha$-type rectangle as long as the sides of $E^k_\epsilon$ are larger than $\delta$ and $\epsilon$ is small enough.
\end{prop}
\proof
{\bf Step 1: connectedness of $F$.} We want to prove that each $E_\epsilon^k$ is connected. It will suffice to show this for $F=E_\epsilon^1$. We first need an estimate on the area of the ``small components'' of $E_\epsilon^1$; this estimate will be obtained by using the comparison principle in Proposition \ref{wcp}.
Let $\ell>0$ be the maximum number such that for each point $x\in E_\epsilon^0$ there exists $y\in \mathbb{R}^2$ such that $x\in (y+ Q_\ell)\subseteq E_\epsilon^0$, where $Q_\ell=[-\ell/2,\ell/2]\times[-\ell/2,\ell/2]$, and the same property holds for $x\not\in E_\epsilon^0$. If $E_\epsilon^0=[-L_1/2,L_1/2]\times[-L_2/2,L_2/2]$, we can choose $\ell=\min\{L_1,L_2\}$. By applying Proposition \ref{wcp} and Remark \ref{wpcoss} to the union of squares contained in $E_\epsilon^0$, and to those outside $E_\epsilon^0$, respectively, and taking into account that a side of length $\ell$ shrinks by $\left\lfloor\frac{2\alpha\gamma}{\ell}\right\rfloor\epsilon$ in absence of defects (see \cite{BGN}), it follows that
\begin{equation*}
\text{d}_\H(\partial E_\epsilon^1,\partial E_\epsilon^0)\leq\left(\frac{2\alpha\gamma}{\ell}+1\right)\epsilon.
\end{equation*}
\\
In this way, it is not possible to have a configuration as in Fig.~\ref{fig:1}, with two large components for $E^1_\epsilon$.
\begin{figure}[ht]
\centering
\def190pt{190pt}
\input{fig1.pdf_tex}
\caption{Test set with $E^1_\epsilon$ with two large components.}\label{fig:1}
\end{figure}
Assume by contradiction that $E_\epsilon^1$ is not connected. In this case we should have
only one large component as in Fig.~\ref{fig:2}.
\begin{figure}[ht]
\centering
\def190pt{190pt}
\input{fig2.pdf_tex}
\caption{Small components of $E_\epsilon^1$.}\label{fig:2}
\end{figure}
We consider the decomposition
\begin{equation*}
E_\epsilon^1=E_{0,\epsilon}^1\cup\bigcup_{i=1}^NE_{i,\epsilon}^1,
\end{equation*}
with $E_{0,\epsilon}^1$ the component containing all the points of $E_\epsilon^0$ having distance more than $C'\epsilon$ from $\partial E_\epsilon^0$ for a suitable constant $C'<2\alpha\gamma/\ell+1$.
Therefore for a suitable constant $C''$ we have
\begin{equation*}
d_\infty^\epsilon(x,\partial E_\epsilon^0)\leq C''\epsilon \quad\text{ for all $x\in E_{i,\epsilon}^1$ and $i\geq1$}.
\end{equation*}
By using the isoperimetric inequality, for $\epsilon$ small enough we infer
\begin{equation*}
\frac{1}{\tau}\int_{E_{i,\epsilon}^1}d_\infty^\epsilon(x,\partial E_\epsilon^0)\,dx\leq(C''/\gamma)|E_{i,\epsilon}^1|<C_{\text{iso}}\sqrt{|E_{i,\epsilon}^1|}\leq \P^{\alpha,\alpha}_\epsilon(E_{i,\epsilon}^1)\leq\P_\epsilon^{\alpha,\beta}(E^1_{i,\epsilon}),
\end{equation*}
with $C_{\text{iso}}$ being the constant of the isoperimetric inequality. Thus, we get a contradiction since we can decrease strictly the energy by eliminating the small components of $E_\epsilon^1$ and considering the set $E'=E_{0,\epsilon}^1$ as a competitor.
\smallskip
{\bf Step 2: $\alpha$\hbox{-}rectangularization.} Consider the maximal $\alpha$-type rectangle $R^\alpha$ with each side intersecting $F$. We call the set $F'=F\cup R^\alpha$ the \emph{$\alpha$\hbox{-}rectangularization} of $F$. This set is either an $\alpha$-type rectangle (and in this case we conclude) or it has some protrusions intersecting $\beta$-bonds (Fig.~\ref{fig:4}).
\begin{figure}[htbp]
\centering
\def190pt{190pt}
\input{fig4.pdf_tex}
\caption{$\alpha$-rectangularization.}\label{fig:4}
\end{figure}
In both cases $\P^{\alpha,\beta}_\epsilon(F')\leq\P^{\alpha,\beta}_\epsilon(F)$, and the symmetric difference with $E_\epsilon^0$ decreases. To justify this, note that the $\alpha$\hbox{-}rectangularization reduces (or leaves unchanged) $\P^{\alpha,\alpha}_\epsilon$ and it reduces the symmetric difference.
As a consequence of this observation, we also deduce an \emph{a priori} estimate on the maximal distance between $\partial E^0_\epsilon$ and $\partial E^1_\epsilon$. By the argument above, $F$ contains an $\alpha$-type rectangle $R^\alpha$
and is strictly contained in an $\alpha$-type rectangle $\widetilde R^\alpha$ whose sides have a distance from the corresponding sides of $R^\alpha$ of not more than $(N_\beta+1)\epsilon$.
We only check the \emph{a priori} estimate in the simplifying hypothesis that $E^0_\epsilon$ is of $\alpha$-type and that $E^0_\epsilon$ and $R^\alpha$ are both concentric squares, so that we can express this estimate in terms of the length $L$ of the sides of $E^0_\epsilon$ and the distance between $\partial E^0_\epsilon$ and $\partial R^\alpha$, which can be expressed as $\epsilon N$. Note that we have
$$
\alpha \H^1(\partial E^0_\epsilon)\ge \mathcal{F}_{\epsilon,\tau}^{\alpha,\beta}(E^1_\epsilon,E^0_\epsilon)\ge
\alpha \H^1(\partial R^\alpha)+\frac{1}{\tau}\int_{E^0_\epsilon\setminus \widetilde{R}^\alpha}d_\infty^\epsilon(x,\partial E^0_\epsilon)\,dx,
$$
which translates into
$$
4\alpha L\ge 4\alpha(L-2\epsilon N)+{2L\over\gamma}\epsilon(N-N_\beta)^2+ O(\epsilon^2),
$$
and gives (for $\epsilon$ sufficiently small)
\begin{equation}\label{cielle}
N\le {c_1\over L}+c_2N_\beta=:c(L).
\end{equation}
The same type of estimate holds in the general case taking $L$ the minimal length of sides of $E^0_\epsilon$.
\smallskip
{\bf Step 3: profile of protrusions on $\beta$\hbox{-}squares.}
Now we want to describe the form of the optimal profiles of the boundary of $F$ intersecting $\beta$-squares.
As noted above, $F$ contains an $\alpha$-type rectangle $R^\alpha=[\epsilon m_1,\epsilon M_1]\times [\epsilon m_2,\epsilon M_2]$ and is contained in the $\alpha$-type rectangle $$\widetilde R^\alpha=[\epsilon (m_1-N_\beta),\epsilon (M_1+N_\beta)]\times[\epsilon (m_2-N_\beta),\epsilon (M_2+N_\beta)]$$ whose side exceed the ones of $R^\alpha$ by at most $2\epsilon N_\beta$. We will describe separately the possible profile of $F$ close to each side of $R^\alpha$; e.g., in the rectangle $[\epsilon (m_1-N_\beta),\epsilon (M_1+N_\beta)]\times[\epsilon M_2,\epsilon (M_2+N_\beta)]$ (i.e., close to the upper horizontal side of $R^\alpha$).
\bigskip
We first consider the possible behavior of the boundary of $F$ at a single $\beta$-square~$Q$.
We suppose that such $Q$ is not one of the two extremal squares, for which a slightly different analysis holds.
First, if a portion $\Gamma$ of $\partial F$ intersects $Q$ in exactly two points on opposite vertical sides, then we may consider
in place of $F$ the union of $F$ and all the $\epsilon$-squares with centers $(x,y)$ in $Q\cap\epsilon\mathbb{Z}^2$ and
$$y\le \max \{z_2: z\in \Gamma\}.$$
The new set, pictured in Fig.~\ref{upper}, has both lower perimeter and less symmetric difference with $E^0_\epsilon$.
\begin{figure}[ht]
\centering
\def190pt{280pt}
\input{upper.pdf_tex}
\caption{Envelope of $\partial F$ when intersecting opposite sides.}\label{upper}
\end{figure}
If a portion $\Gamma$ intersects $Q$ in exactly two points on the same side (horizontal or vertical) or adjacent sides,
then we may remove all the $\epsilon$-squares with centers in the portion of $Q\cap F$ with boundary $\Gamma$.
The two cases are pictured in Fig.~\ref{same} and Fig.~\ref{adjacent}, respectively.
\begin{figure}[ht]
\centering
\def190pt{280pt}
\input{same.pdf_tex}
\caption{Removal of $\partial F$ when intersecting one side.}\label{same}
\end{figure}
\begin{figure}[ht]
\centering
\def190pt{280pt}
\input{adjacent.pdf_tex}
\caption{Removal of $\partial F$ when intersecting two adjacent sides.}\label{adjacent}
\end{figure}
This operation decreases the perimeter of at least $\epsilon(\beta-\alpha)$, while at most increases the
bulk term by ${1\over \tau}\epsilon^3 N_\beta^2 c(L)$ ($c(L)$ given by (\ref{cielle})). The total change in the energy is thus
\begin{equation}\label{stima}
-\epsilon(\beta-\alpha) +{1\over \gamma}\epsilon^2N_\beta^2 c(L)\,,
\end{equation}
which is negative if $\epsilon$ is small enough.
As a consequence, either $F\cap Q=\emptyset$ or $\partial F\cap Q$ is a horizontal segment.
The same type of analysis applies to the extremal squares, for which we deduce instead that $F\cap Q$ is a rectangle with one vertex coinciding with a vertex of $\widetilde R^\alpha$.
\bigskip We now consider the interaction of consecutive $\beta$-squares.
Let $Q_1,\ldots, Q_K$ be a maximal array of consecutive $\beta$-squares with $F\cap Q_k\neq\emptyset$ for $k=1,\ldots, K$
and such that $Q_1$ is not a corner square.
If we substitute $F$ with $F\cup R$, where $R$ is the maximal rectangle of $\epsilon$-squares containing all $F\cap Q_k$ and not intersecting other $\beta$-squares, then the corresponding energy has a not larger perimeter part, and a bulk part which is strictly lower if $F\cup R\neq F$. This substitution is pictured in Fig.~\ref{more}.
\begin{figure}[ht]
\centering
\def190pt{420pt}
\input{more.pdf_tex}
\caption{Envelope of $\partial F$ in consecutive squares.}\label{more}
\end{figure}
If the subsequent $\beta$-squares $Q_{K+1}\ldots, Q_{K+K'}$ are a maximal array which do not intersect $F$, then we may further substitute $F\cup R$ with $(F\setminus R)\cup (R+\epsilon N_{\alpha\beta} K'(1,0))$, where we translate $R$ until it meets another portion of $F$ (if any). This translation is pictured in Fig.~\ref{right}.
\begin{figure}[ht]
\centering
\def190pt{420pt}
\input{right.pdf_tex}
\caption{Translation argument to join protrusions.}\label{right}
\end{figure}
Note that if it does meet another portion of $F$, then the change in energy is at most
\begin{equation}\label{stima2}
-2\epsilon\alpha +{1\over \gamma}\epsilon^2N_\beta N_\alpha c(L)\,,
\end{equation}
which is negative if $\epsilon$ is small enough.
In this case at this point we may iterate this analysis since we now have a larger array of consecutive $\beta$-squares intersecting $F$. Note, moreover, that the same argument can be repeated shifting the rectangle $R$ to the left instead than to the right if energetically convenient. As a conclusion, we obtain that $F$ may only either intersect one array of consecutive $\beta$-squares, or two such arrays if they contain the two corner $\beta$-squares; i.e., we have one of the two situations pictured in Fig.~\ref{final}.
\begin{figure}[ht]
\centering
\def190pt{420pt}
\input{final.pdf_tex}
\caption{Profiles of candidate minimal $F$.}\label{final}
\end{figure}
\smallskip
{\bf Step 4: all $\beta$-connections can be removed except those at the four corners.} At this point we are in the situations pictured in Fig.~\ref{final}.
\noindent
If we are as in the upper figure, then by removing all $\epsilon$-squares external to $R^\alpha$ the variation of the energy is less or equal than
\begin{equation*}
-(\beta-\alpha)(N_\beta+1)N\epsilon+c(L)\frac{(N+1)N_{\alpha\beta}N_\beta}{\gamma}\epsilon^2,
\end{equation*}
where $N$ is the number of modified $\beta$-squares. For $\epsilon$ small this variation is negative, showing that $F$ does not contain any protrusion.
If we are as in the lower figure, then we may remove all $\beta$\hbox{-}connections inside the border $\beta$-squares, except those in the two periodicity squares at the corners as in Fig.~\ref{fig:12};
\begin{figure}[ht]
\centering
\def190pt{420pt}
\input{figura11.pdf_tex}
\caption{Removing $\beta$\hbox{-}connections except in the two $\beta$-squares at the corners.}\label{fig:12}
\end{figure}
\noindent
the variation of the energy functional is less or equal than
\begin{equation*}
-(\beta-\alpha)(N_\beta+1)N\epsilon+c(L)\frac{NN_{\alpha\beta}N_\beta}{\gamma}\epsilon^2,
\end{equation*}
\\
where $N$ is the number of modified cells. For $\epsilon$ small this variation is negative, showing that the profile in Fig.~\ref{fig:12} is energetically convenient.
We can repeat this procedure for each side, and finally we obtain that $F$ is
the union of a coordinate $\alpha$-type rectangle $R$ and possibly one to four rectangles $\widetilde{R}_i, i=1,\dots,4$ of side lengths at most $N_{\alpha\beta}\epsilon$
such that the intersection of $\widetilde{R}_i$ with each corner $\beta$-square is a rectangle (see Fig.~\ref{fig:quasirectangle}).
\begin{figure}[ht]
\centering
\def190pt{400pt}
\input{figura12.pdf_tex}
\caption{The set obtained in Step 4.}\label{fig:quasirectangle}
\end{figure}
\smallskip
{\bf Step 5: conclusion.} It remains to prove that the rectangles $\widetilde{R}_i$ in the previous step are actually not there.
This is immediately checked by comparing such an $F$ with $R^\alpha$: if $\widetilde{R}_i\neq \emptyset$ then by removing it the energy changes by at most by
$$
-2\beta\epsilon+{1\over\gamma}c(L) \epsilon^2N_{\alpha\beta}^2,
$$
which is negative for small $\epsilon$.
We finally note that all the estimates above can be iterated and hold uniformly as long as the sides of $E^k_\epsilon$ are larger than $\delta$, since they depend only on $c(\delta)$.
\endproof
The proposition above shows that we may restrict our analysis to $\alpha$-type rectangles; indeed, for fixed $\epsilon$ this assumption is not restrictive until the sides of the rectangles are larger than a constant, which vanishes as $\epsilon\to 0$.
As a consequence, once we suppose the convergence of the initial data, up to subsequences, the discrete
motions $E_{\epsilon,\tau}(t)$ converge as $\epsilon\to 0$ to a limit $E(t)$ such that $E$ is a rectangle for all $t$, up to its extinction time. Note, moreover, that it is not restrictive to suppose that also the initial data are $\alpha$-type rectangles, up to substituting $E^0_\epsilon$ with $E^1_\epsilon$.
\begin{figure}[ht]
\centering
\def190pt{300pt}
\input{negligible.pdf_tex}
\caption{Picture of $E^{k+1}_\epsilon$ inside $E^k_\epsilon$.}\label{BNG-figure4}
\end{figure}
As shown in \cite{BGN}, the motion of each side of $E_\epsilon^k$ can be studied separately, since the constraint of being an $\alpha$-type rectangle does not influence the argument therein, which consists in remarking that the bulk term due to the small corner rectangles in Fig.~\ref{BNG-figure4} is negligible.
As a consequence, we can describe the motion in terms of the length of the sides of $E_\epsilon^k$. This will be done in the following sections.
\subsection{A new {pinning} threshold}\label{newpinning}
We first examine the case when the limit motion is trivial; i.e., all $E_k=E_\epsilon^k$ are the same after a finite number of steps.
This will be done by computing the {\em pinning threshold}; i.e., the critical value of the side length $L$ above which it is energetically not favorable for a side to move. We recall that, in the case $\alpha=\beta$, this threshold is $$\widetilde{L}=2\alpha\gamma.$$
This value is obtained by computing the values for which a side of length $L$ may not move inwards of $\epsilon$ by decreasing the energy. In our case, by the condition that $E_k$ be an $\alpha$-type rectangle, we have to impose instead that it is not energetically favorable to move inwards a side by $(N_\beta+1)\varepsilon$ (see Fig.~\ref{fig:pinning}).
\begin{figure}[htbp]
\centering
\def190pt{190pt}
\input{pinning.pdf_tex}
\caption{Motion is possible if the side can move at least by $(N_\beta+1)\epsilon$.}
\label{fig:pinning}
\end{figure}
We then write the variation of the energy functional $\mathcal{F}^{\alpha,\beta}_{\epsilon,\tau}$ from configuration $A$ to configuration $B$ in Fig.~\ref{fig:pinning}, regarding a side of length $L$.
If we impose it to be positive, we have
\begin{equation*}
-2(N_\beta+1)\alpha\epsilon+\frac{1}{\tau}\displaystyle\sum_{k=1}^{N_\beta+1}(k\epsilon)L\epsilon =
(N_\beta+1)\epsilon\left[-2\alpha+\frac{L}{2\gamma}(N_\beta+2)\right]\geq0
\end{equation*}
and we obtain the pinning threshold
\begin{equation}
\overline{L}:=\frac{4\gamma\alpha}{N_\beta+2}.
\label{threshold}
\end{equation}
\\
Note that this threshold depends on $N_\beta$ and not on the value $\beta>\alpha$ and that, if $N_\beta=0$ (or, otherwise, $\alpha=\beta$), we recover the previous threshold $\widetilde L$.
\subsection{Definition of the effective velocity}\label{newvelocity}
As remarked above, up to an error vanishing as $\epsilon\to0$, the motion of each side is independent of the other ones.
As a consequence, its description can be reduced to a one-dimensional problem, where the unknown represents, e.g., the location of the left-hand vertical side of $E_k$.
Let $x_k$ represents the projection of this side of $E_k$ on the horizontal axis.
The location of $x_{k+1}$ depends on a minimization argument involving
$x_{k}$ and the length $L_k$ of the corresponding side of $E_{k}$.
However, we will see that this latter dependence is locally constant, except for a discrete set of values of $L_k$.
Indeed, for all $Y>0$ (which in our case will be of the form $Y=\gamma/L_k$), consider the minimum problems
\begin{equation}\label{minel}
\min\left\{-2\alpha N+\frac{N(N+1)}{2Y}:N\in\mathbb{N}, \quad [x+N]_{N_{\alpha\beta} }\in\mathbb{Z}_{N_\alpha}\right\},
\end{equation}
for $x\in\{0,\ldots, N_{\alpha\beta}\}$,
where $[z]_{N_{\alpha\beta}}$ denotes the congruence class of $z$ modulo $N_{\alpha\beta}$ and
\begin{equation*}
\mathbb{Z}_{N_\alpha}=\left\{[0]_{N_{\alpha\beta}},\dots,[N_\alpha-1]_{N_{\alpha\beta}}\right\}.
\end{equation*}
Then the set of $Y>0$ for which (\ref{minel}) does not have a unique solution is discrete.
To check this it suffices to remark that the function to minimize
$$-4\alpha XY+{X(X+1)}$$
is a parabola with vertex in
$$
X={2\alpha Y}-{1\over 2}.
$$
The minimizers $N$ are points with $[x+N]_{N_{\alpha\beta} }\in\mathbb{Z}_{N_\alpha}$ of minimal distance from the vertex $X$.
These are not unique in some cases: first if the vertex $X$ is equidistant from two consecutive points in ${\mathbb{Z}}_{N_\alpha}$; i.e.,
if $${2\alpha Y}-{1\over 2}\in {1\over 2}+\mathbb{Z}, $$ or, equivalently,
\begin{equation}\label{evena}
Y\in{1\over 2\alpha}\mathbb{Z}.
\end{equation}
The second case is when we have two points in ${\mathbb{Z}}_{N_\alpha}$ of minimal distance from $X$ which are not consecutive. In this case the distance between these points is $N_\beta+1$, so that we have
$$
{2\alpha Y}-{1\over 2}\in {N_\beta+1\over 2}+\mathbb{Z},
$$
or, equivalenly,
$$
Y\in{1\over 2\alpha}\Bigl({N_\beta\over 2}+\mathbb{Z}\Bigr).
$$
If $N_\beta$ is even then this condition is equivalent to (\ref{evena}), while if $N_\beta$ is odd then we have
\begin{equation}\label{odda}
Y\in {1\over 4\alpha}+{1\over 2\alpha}\mathbb{Z}.
\end{equation}
\defS_\beta{S_\beta}
\begin{defn} We define the (possibly) {\em singular set} $S_{\beta}$ for problems (\ref{minel}) as
$$
S_\beta= {1\over 2\alpha}\Bigl(\mathbb{Z}\cup\Bigl({1\over 2}+\mathbb{Z}\Bigr)\Bigr).
$$
\end{defn}
We will examine the iterated minimizing scheme for $\gamma /L_k=\gamma /L\in (0,+\infty)\setminus S_\beta$ fixed, which reads
\begin{equation}
\begin{cases}
x_{k+1}^L=x_k^L+\overline{N}_k,&k\geq0\\
x_0^L=x^0
\end{cases}
\label{system}
\end{equation}
with $x^0\in\{0,1,\dots,N_{\alpha\beta}-1\}$ and $\overline{N}_k\in\mathbb{N}$ the minimizer of
\begin{equation}
\min\left\{-2\alpha N+\frac{1}{\gamma}\frac{N(N+1)}{2}L:N\in\mathbb{N}, \quad [x^L_k+N]_{N_{\alpha\beta} }\in\mathbb{Z}_{N_\alpha}\right\},
\label{scheme}
\end{equation}
which is unique by the requirement that $\gamma/L\not\in S_\beta$.
After at most $N_\alpha$ steps, $\{x_{k}^L\}_{k\geq0}$ is \emph{periodic modulo} $N_{\alpha\beta}$, as expressed by the following proposition.
\begin{prop}\label{perino}
There exist integers $\overline{k}\le N_\alpha, M\leq N_\alpha$ and $n\geq1$ such that
\begin{equation}\label{periodic}
x_{k+M}^L=x_{k}^L+n\,N_{\alpha\beta}
\qquad\hbox{ for all } k\geq\overline{k}.
\end{equation}
Moreover, the quotient $M/n$ depends only on $\gamma/L$.
\end{prop}
\proof
First remark that, if $x_k^L$ is defined recursively by (\ref{system}), we have
\begin{equation*}
[x_k^L]_{N_{\alpha\beta}}\in{\mathbb{Z}_{N_\alpha}}\qquad\hbox{ for all }k\ge 1.
\end{equation*}
Since $\#\mathbb{Z}_{N_\alpha}=N_\alpha$, there exist integers $0\leq j\leq N_\alpha$ and $l>j$, with $l-j\leq N_\alpha$, such that
\begin{equation}
[x^L_j]_{N_{\alpha\beta}}=[x^L_{l}]_{N_{\alpha\beta}}.
\label{equality}
\end{equation}
Let $l$ be the minimal such $l$.
Define $\overline{k}=j$, $M=l-j$ and $n=\displaystyle\frac{x^L_l-x^L_j}{N_{\alpha\beta}}$ to obtain (\ref{periodic}).
\smallskip
It remains to show the last statement of the theorem. It suffices to show that the quotient is independent of $x_0$.
We start by proving a monotonicity property of the orbits defined in (\ref{system})
with respect to the initial datum: if $\{x_k\}$ and $\{x'_k\}$ are orbits obtained as above, we have
\begin{equation}
\text{if } x_0\le x'_0, \text{then }x_k\leq x'_k,\quad \hbox{for all }k\geq1.
\label{monotone}
\end{equation}
This can be seen iteratively from (\ref{minel}) since the problems with $x=x_{k-1}$ and $x=x'_{k-1}$
consist in a constrained minimization of a parabola and its translation by $x'_{k-1}-x_{k-1}$, and, as
previously remarks, the minimizer
in (\ref{minel}) is the closest point to the vertex of the parabola with $[x+N]_{N_{\alpha\beta} }\in\mathbb{Z}_{N_\alpha}$.
Consider the orbits with initial data $x_0$, $x'_0$ and $x_0+N_{\alpha\beta}$,
and let $n(x)$ and $M(x)$ denote the indices above with initial datum $x\in\{x_0,x'_0,x_0+N_{\alpha\beta}\}$.
Since the orbit with initial datum $x_0+N_{\alpha\beta}$ is the translation by $N_{\alpha\beta}$ of the one with initial datum $x_0$,
we have $n(x_0+N_{\alpha\beta})=n(x_0)$ and $M(x_0+N_{\alpha\beta})=M(x_0)$. Taking into account the ordering of the initial conditions
$$
x_0\le x'_0\le x_0+N_{\alpha\beta},
$$
by (\ref{periodic}) for $k_0$ sufficiently large and taking $k=k_0+T M(x_0) M(x'_0)$ with $T\in\mathbb{N}$,
from $x_k\leq x'_k\leq x_k+N_{\alpha\beta}$
we get
\begin{eqnarray*}
x_{k_0}+Tn(x_0)M(x'_0)N_{\alpha\beta}\Nab&\le& x'_{k_0}+Tn(x'_0)M(x_0)N_{\alpha\beta}\\
&\le& x_{k_0}+Tn(x_0)M(x'_0)N_{\alpha\beta}+N_{\alpha\beta}.
\end{eqnarray*}
In order that this inequality hold for all $T\ge 1$ we must have
$$
n(x_0)M(x'_0)=n(x'_0)M(x_0),
$$
which is the desired equality.
\endproof
\begin{defn}[Effective velocity]\label{velofo} \rm
We define the
\emph{effective velocity function}
$f:(0,+\infty)\setminus S_\beta\longrightarrow[0,+\infty)$ by setting
\begin{equation}
f(Y)=\frac{nN_{\alpha\beta}}{M},
\label{velocityfunction}
\end{equation}
with $M$ and $n$ in (\ref{periodic}) defined by $L$ and $\gamma$ such that $Y=\gamma/L$. By Proposition \ref{perino} this is a good definition.
\label{effvel}
\end{defn}
\begin{oss}\label{meva}
The terminology for formula
(\ref{velocityfunction}) is motivated by the fact that we can define the velocity of a side as a \emph{mean velocity}
averaging on a period; that is,
\begin{equation}
v=\frac{nN_{\alpha\beta}\epsilon}{M\tau}.
\label{meanvel}
\end{equation}
In (\ref{meanvel}) the velocity is the ratio between the minimal (periodic) displacement of the side and the product of the time\hbox{-}scale $\tau$ and the number of steps necessary to describe the minimal period, each of which considered as a 1\hbox{-}time step.
\end{oss}
\begin{oss}[Properties of the velocity function $f$] The velocity function $f$ has the following properties:
\begin{description}
\item[(a)] $f$ is constant on each interval contained in its domain;
\item[(b)] $f(Y)=0$ if $$
Y<\overline Y:=\frac{N_\beta+2}{4\alpha};
$$
in particular
$$
\lim_{\gamma\to 0^+} {1\over \gamma} f\Bigl({\gamma\over L}\Bigr)=0\,.
$$
Note that $(0,\overline Y)\cap S_\beta\neq\emptyset$;
\item[{(}c)] $f(Y)$ is a rational value;
\item[(d)] $f$ is non decreasing;
\item[(e)] we have
$$
\lim_{\gamma\to+\infty} {1\over \gamma} f\Bigl({\gamma\over L}\Bigr)={2\alpha\over L}\,.
$$
\item[(f)] $f(Y)$ is independent of $\beta$ but depends on $N_\beta$.
\end{description}
{\bf (a)} holds since on each component of $(0,+\infty)\setminus S_\beta$ the minimum problems (\ref{minel}) have a unique solution independent of $Y$, so that the values $n$ and $M$ in Proposition \ref{perino} are independent of $Y$. Note, however, that $f(Y)$ may be equal on neighboring components since the corresponding $n$ and $M$ may be equal even without uniqueness in (\ref{perino});
{\bf (b)} holds since we have $\overline{Y}=\gamma/\overline L$, where $\overline{L}$ is the pinning threshold (\ref{threshold}),
and the computation of the pinning threshold is equivalent to the requirement that the orbit be constant after a finite number of steps;
{\bf (c)} is immediate from the formula for $f(Y)$;
{\bf (d)} is again a consequence of the fact that (\ref{minel}) are minimum problems related to a parabola with vertex in $2{\alpha Y}-{1\over 2}$ and the latter is an increasing function of $Y$;
{\bf (e)} using the same argument as in {\bf(d)} above, we deduce in particular that
$$
\Bigl|\overline {N}_k-{2\alpha Y}+{1\over 2}\Bigr|\le N_\beta,
$$
which for $Y=\gamma/L$ implies that
$$
{2\alpha\over L}-{2N_\beta+1\over 2\gamma}\le {1\over \gamma} f\Bigl({\gamma\over L}\Bigr)\le {2\alpha\over L}+{2N_\beta+1\over 2\gamma},
$$
and the desired equality letting $\gamma\to+\infty$;
{\bf (f)} is an immediate consequence of the definition of $f(Y)$.
\end{oss}
\begin{oss}\label{compar}
Let $\gamma/L\in S_\beta$, and let $\{x^L_k\}$ be defined by (\ref{system}) with $\overline {N}_k$ chosen to be a minimizer of (\ref{scheme}), which may be not unique. Then arguing by monotonicity as in {\bf(d)} above, we have $ x^{L^+}_k\le x^L_k\le x^{L^-}_k$, where $L^\pm$ are any two values with $L^-<L<L^+$ and $\gamma/L^\pm$ belonging to the two intervals of $(0,+\infty)\setminus S_\beta$ with one endpoint equal to $L$, and $\{x^{L^\pm}_k\}$ have the same initial data.
\end{oss}
\subsection{Description of the homogenized limit motion.}\label{limitmotion}
The following characterization of any limit motion holds.
\begin{thm}\label{limitmotion1}For all $\epsilon>0$, let $E^0_\epsilon\in\mathcal{D}_\epsilon$ be a coordinate rectangle with sides $S^0_{1,\epsilon},\dots,S^0_{4,\epsilon}$. Assume also that
\begin{equation*}
\lim_{\epsilon\to0^+}\emph{d}_\H(E^0_\epsilon,E_0)=0
\end{equation*}
for some fixed coordinate rectangle $E_0$. Let $\gamma>0$ be fixed and let $E_\epsilon(t)= E_{\epsilon,\gamma\epsilon}(t)$ be the piecewise-constant motion with initial datum $E^0_\epsilon$ defined in {\rm(\ref{disefo})}.
Then, up to a subsequence, $E_\epsilon(t)$ converges as $\epsilon\to0$ to $E(t)$, where $E(t)$ is a coordinate rectangle with sides $S_i(t)$ and such that $E(0)=E_0$. Each $S_i$ moves inward with velocity $v_i(t)$ satisfying
\begin{equation}
v_i(t)
\in\left[\displaystyle\frac{1}{\gamma}f\biggl({\gamma\over L_i(t)}\biggr)^-,\displaystyle\frac{1}{\gamma}f\biggl({\gamma\over L_i(t)}\biggr)^+\right],
\label{vl}
\end{equation}
where $f$ is given by Definition {\rm\ref{velofo}}, $L_i(t):=\H^1(S_i(t))$ denotes the length of the side $S_i(t)$, until the extinction time when $L_i(t)=0$, and $f(Y)^-,f(Y)^+$ are the upper and lower limits of the effective-velocity function at $Y\in (0,+\infty)$.
\end{thm}
\proof We will apply the results of the previous sections with $\tau=\gamma\epsilon$.
Let $S_{\epsilon,i}(t)$ be the sides of $E_\epsilon(t)$, and let $L_{i,\epsilon}^k= \H^1(S_{\epsilon,i}(k\tau))$; i.e., $L_{i,\epsilon}^k$ is the length of the $i$-th side of $E^k_\epsilon$ in the notation of the previous sections. If $\Delta S^k_{\epsilon,i}= \text{d}_\H(S_{\epsilon,i}(\gamma\epsilon k),S_{\epsilon,i}(\gamma\epsilon (k+1))$ denotes the distance from corresponding sides of $E^k_\epsilon$ then note that
$$
L_{i,\epsilon}^{k+1}-L_{i,\epsilon}^k= -\bigl(\Delta S^k_{\epsilon,i-1}+ \Delta S^k_{\epsilon,i+1}\bigr)
$$
(where the indices $i$ rotate cyclically).
By (\ref{cielle}) we have
$$
{\Delta S^k_{\epsilon,i}\over \tau}\le {c_1\over L_{i,\epsilon}^{k}}+ c_2.
$$
This implies that if we define $L_{i,\epsilon}(t)$ as the affine interpolation
in $[k\tau,(k+1)\tau]$ of the values $L_{i,\epsilon}^k$, then $L_{i,\epsilon}(t)$ is a decreasing continuous function of $t$ and the sequence is uniformly Lipschitz continuous on all intervals $[0,T]$ such that $L_{i,\epsilon}(T)\geq c>0$. Hence it converges (up to a subsequence), as $\epsilon\to0$, to a function $L_i(t)$, which is also decreasing. It follows that $E_\epsilon(t)$ converges as $\epsilon\to0$, up to a subsequence and in the Hausdorff sense, to a limit rectangle $E(t)$, for all $t\geq0$.
It remains to justify formula (\ref{vl}) for the velocity $v_i$ of the side $S_i(t)$. Let $[t^-, t^+]$ and $L^\pm _i$ be such that
$\gamma/L^\pm_i\in (0,+\infty)\setminus S_\beta$ and
$$
L^-_i < L_i(t)< L^+_i \qquad \hbox{ for } t^-\!\le t\le t^+.
$$
Then the corresponding $L_{i,\epsilon}(t)$ satisfy the same inequalities for $\epsilon$ small enough. By Remarks \ref{compar} and \ref{meva} we then have
$$
{1\over \gamma} f\Bigl({\gamma\over L^+}\Bigr)\le v_i(t)\le {1\over \gamma} f\Bigl({\gamma\over L^-}\Bigr)\quad\hbox{ for } t^-\!\le t\le t^+.
$$
By optimizing in $L^\pm$, and recalling that $f$ is not decreasing, we obtain (\ref{vl}).
\endproof
\begin{thm}[Unique limit motions]
Let $E_\epsilon,E_0$ be as in the statement of Theorem~{\rm \ref{limitmotion}}. Assume in addition that the lengths $L^0_1,L^0_2$ of the sides of the initial set $E_0$ satisfy one of the following conditions (we assume that $L^0_1\leq L^0_2$):
\begin{itemize}
\item[\emph{(a)}] $L^0_1,L^0_2>\displaystyle\frac{4\alpha\gamma}{N_\beta+2}$ \emph{(total pinning)};
\item[\emph{(b)}] $L^0_1<\displaystyle\frac{4\alpha\gamma}{N_\beta+2}$ and $L^0_2\leq\displaystyle\frac{4\alpha\gamma}{N_\beta+2}$ \emph{(vanishing in finite time)};
\end{itemize}
then $E_\epsilon(t)$ converges locally in time to $E(t)$ as $\epsilon\to 0$, where $E(t)$ is the unique rectangle with sides of lengths $L_1(t)$ and $L_2(t)$ which solve the following system of ordinary differential equations
\begin{equation}\label{unita}
\begin{cases}\displaystyle
\dot{L}_1(t)=-{2\over \gamma}\,f\Bigl({\gamma\over L_2(t)}\Bigr)\\
\\ \displaystyle
\dot{L}_2(t)=-{2\over \gamma}\, f\Bigl({\gamma\over L_1(t)}\Bigr)
\end{cases}
\end{equation}
for almost every $t$, with initial conditions $L_1(0)=L^0_1$ and $L_2(0)=L^0_2$, where $f$ is given by Definition {\rm\ref{velofo}}.
\end{thm}
\proof
In case (a) the statement follows by Theorem~\ref{limitmotion1} noticing that we have $v_1(t)=v_2(t)=0$ for all $t\geq0$, which is equivalent to $\dot{L}_1=\dot{L}_2=0$.
In case (b) the lengths of $L_i$ are strictly decreasing until the extinction time. This implies that the set of $t$ such that
$f ({\gamma/ L_i(t)} )^-\neq f ({\gamma/ L_i(t)} )^+$ is negligible, and (\ref{unita}) follows since $\dot L_i=-2 v_{i+1}$.
\endproof
\begin{oss}[general evolutions]
More general initial data can be considered. Since their treatment follows from Theorem \ref{limitmotion1} as in \cite{BGN}, we do not include the details. We only recall that:
$\bullet$ all velocities $v_i$ satisfying (\ref{vl}) can be obtained, with a proper choice of the initial data $E_{0,\epsilon}$;
$\bullet$ if we take initial data $E_0$ coordinate polyrectangles then the motion can be characterized with the same velocities, with the convention that convex sides move inwards, concave sides move outwards, other sides remain pinned;
$\bullet$ more general initial data $E_0$ can be dealt with once we remark that at level $\epsilon$ the assumption that $E_{0,\epsilon}$ is a polyrectangle is always satisfied.
\end{oss}
\section{Computation of the velocity function}
The velocity function in Definition \ref{effvel} may be not easily described
for generic $N_\alpha$ and $N_\beta$. In this section we compute it, by means of algebraic formulas, in the simpler cases $N_\beta=1$ and $N_\beta=2$, with varying $N_\alpha$. These are prototypes for the cases $N_\beta$ odd and $N_\beta$ even, respectively. We also give two easy examples for $N_\alpha$ fixed and equal to $1$, and we compare the new velocity function with the homogeneous case showing that the inhomogeneities in the lattice may accelerate or decelerate the motion. We can assume, without loss of generality, that $\gamma=1$.
\subsection{The case $N_\beta=1$.}\label{N1}
Let $Y>\overline{Y}=\frac{3}{4\alpha}$. We assume also that $Y$ is not in the singular set;
i.e.,
\begin{equation*}
Y\not\in\left\{\frac{k+j(N_\alpha+1)}{2\alpha},k=1,\dots,N_\alpha-1,j\ge 0\right\}\cup\left\{\frac{N_\alpha+(2j+1)(N_\alpha+1)}{4\alpha},j\ge 0\right\}.
\end{equation*}
As shown by Proposition~\ref{perino}, the minimal period is independent of the starting point of the orbits, so there is no restriction to assume that $x^0=0$ in (\ref{system})-(\ref{scheme}).
We divide the analysis in the three cases (a), (b) and {(}c) below.
\bigskip
$(a)$ If $Y\in \left(\displaystyle\frac{k+j(N_\alpha+1)}{2\alpha},\frac{k+1+j(N_\alpha+1)}{2\alpha}\right), k=1,2,\dots,N_\alpha-1,j\ge 0$, then we denote the minimizer of problem (\ref{scheme}) in the homogeneous case $N_\beta=0$ by $N=k+j(N_\alpha+1)$. The velocity function $f(Y)$ will be characterized by algebraic relations between $N$ and $N_\alpha$.
We have two sub-cases:
$(a_1)$ $N$ and $N_\alpha+1$ are coprime. In this case, by iterating the scheme (\ref{scheme}), after at most $N_\alpha$ steps the side encounters a defect, that is
\begin{equation*}
[nN]_{N_\alpha+1}=[N_\alpha]_{N_\alpha+1}
\end{equation*}
for some $1\leq n\leqN_\alpha$.
In this case, we denote by $\bar{n}\ge0$ the minimal solution of the congruence equation
\begin{equation}
nN\equivN_\alpha\quad \text{mod $(N_\alpha+1),$ $n\geq1$},
\label{congr0}
\end{equation}
and $\bar{k}\geq0$ is given by
\begin{equation*}
\bar{k}=\frac{\bar{n}N-N_\alpha}{N_\alpha+1}.
\end{equation*}
If $Y\in\left(\displaystyle\frac{k+j(N_\alpha+1)}{2\alpha},\frac{2k+2j(N_\alpha+1)+1}{4\alpha}\right)$, then the location of the side at step $n$ is at $N_\alpha-1+\bar{k}(N_\alpha+1)$ (which is equal to $-2$ modulo $N_\alpha+1$).
This computation shows that we can limit our analysis to periodic orbits modulo $N_\alpha+1$ with initial datum equal to $-2$ (or, equivalently, $N_\alpha-1$). The period of such orbits is obtained as follows. We solve the congruence equation
\begin{equation}
nN\equiv1\quad \text{mod }N_\alpha+1,
\label{congr1}
\end{equation}
for $n\geq1$ and denote by $n_{\text{min}}$ the minimal positive solution of equation (\ref{congr1}); that is, the minimal positive integer in the class
\begin{equation*}
\left[N^{\varphi(N_\alpha+1)-1}\right]_{\text{mod $N_\alpha+1$}}.
\end{equation*}
The function $\varphi(n)$ is \emph{Euler's totient function} and it counts the integers $m$ such that $1\leq m<n$ and $m$ has no common divisors with $n$.
If we define
\begin{equation*}
k_{\text{min}}=\frac{n_{\text{min}}N-1}{N_\alpha+1},
\end{equation*}
then we have that
\begin{equation}
\begin{split}
f(Y)&=\frac{k_{\text{min}}(N_\alpha+1)}{\displaystyle\frac{k_{\text{min}}(N_\alpha+1)+1}{\lfloor2\alpha Y\rfloor}}=\left(\frac{k_{\text{min}}(N_\alpha+1)}{k_{\text{min}}(N_\alpha+1)+1}\right){\lfloor2\alpha Y\rfloor}\\
&=\left(\frac{1}{1+\frac{1}{k_{\text{min}}(N_\alpha+1)}}\right){\lfloor2\alpha Y\rfloor}.
\end{split}
\label{dec}
\end{equation}
Note that $f(Y)<\lfloor2\alpha Y\rfloor$, so that the velocity of the side reduces (deceleration) with respect to the homogeneous case. \\
Suppose now that $Y\in\left(\displaystyle\frac{2k+2j(N_\alpha+1)+1}{4\alpha},\frac{k+1+j(N_\alpha+1)}{2\alpha}\right)$, then the location of the side at step $n$ is $N_\alpha+1+\bar{k}(N_\alpha+1)$, which is equal to $0$ modulo $N_\alpha+1$.
We have that
\begin{equation}
f(Y)=\left(\frac{(\bar{k}+1)(N_\alpha+1)}{(\bar{k}+1)(N_\alpha+1)-1}\right){\lfloor2\alpha Y\rfloor}=\left(\frac{1}{1-\frac{1}{(\bar{k}+1)(N_\alpha+1)}}\right){\lfloor2\alpha Y\rfloor}.
\end{equation}
Note that $f(Y)>{\lfloor2\alpha Y\rfloor}$, so the velocity of the side increases (acceleration) with respect to the homogeneous case.
$(a_2)$ $N$ and $N_\alpha+1$ are not coprime. In this case the side does not meet any $\beta$\hbox{-}bond and the velocity function has the same value as in the homogeneous case, i.e.
\begin{equation*}
f(Y)=\lfloor2\alpha Y\rfloor.
\end{equation*}
\medskip
$(b)$ If $Y\in\left(\displaystyle\frac{N_\alpha+j(N_\alpha+1)}{2\alpha},\frac{N_\alpha+(2j+1)(N_\alpha+1)}{4\alpha}\right)$ then we argue as in $(a_1)$.\\
\smallskip
$(c)$ If $Y\in\left(\displaystyle\frac{N_\alpha+(2j+1)(N_\alpha+1)}{4\alpha},\frac{1+(j+1)(N_\alpha+1)}{2\alpha}\right)$, then
\begin{equation*}
f(Y)=N_\alpha+1+j(N_\alpha+1).
\end{equation*}
\\
Note that $f(Y)>\lfloor 2\alpha Y\rfloor$ if $Y\in\left(\displaystyle\frac{N_\alpha+(2j+1)(N_\alpha+1)}{4\alpha},\frac{(j+1)(N_\alpha+1)}{2\alpha}\right)$, while $f(Y)=\lfloor 2\alpha Y\rfloor$ if $Y\in\left(\displaystyle\frac{(j+1)(N_\alpha+1)}{2\alpha},\frac{1+(j+1)(N_\alpha+1)}{2\alpha}\right)$.
\begin{example}[The case $N_\alpha=N_\beta=1$]
In this case the velocity function is given by
\begin{equation*}
f(Y)=
\begin{cases}
0&\mbox{if}\quad Y<\displaystyle\frac{3}{4\alpha},\\
\\
\displaystyle{2k}&\mbox{if}\quad Y\in\left(\displaystyle\frac{4k-1}{4\alpha},\displaystyle\frac{4k+3}{4\alpha}\right),\quad k\geq1;
\end{cases}
\end{equation*}
i.e.,
$$
f(Y)= 2\Bigl\lfloor \alpha Y+{1\over 4}\Bigr\rfloor.
$$
\end{example}
\subsection{The case $N_\beta=2$}\label{N2}
We now study the case $N_\beta=2$. Let $Y>\overline{Y}=\displaystyle\frac{1}{\alpha}$ and we assume also that $Y$ is not in the singular set, i.e.,
\begin{equation*}
Y\not\in\left\{\frac{k+j(N_\alpha+2)}{2\alpha},k=1,\dots,N_\alpha-1,j\ge 0\right\}\cup\left\{\frac{N_\alpha+1+j(N_\alpha+2)}{2\alpha},j\ge 0\right\}.
\end{equation*}
\smallskip
$(a)$ If $Y\in \left(\displaystyle\frac{k+j(N_\alpha+2)}{2\alpha},\frac{k+1+j(N_\alpha+2)}{2\alpha}\right), k=1,2,\dots,N_\alpha-2,j\ge 0$, then $N=k+j(N_\alpha+2)$ and we have two sub-cases:
$(a_1)$ $N$ and $N_\alpha+2$ are coprime. We compute $\bar{k}=\min(k_1,k_2)\geq0$, where $k_1$ is the minimal positive solution of the congruence equation
\begin{equation*}
kN\equivN_\alpha\quad \text{mod }N_\alpha+2,
\end{equation*}
and $k_2$ is the minimal positive solution of the congruence equation
\begin{equation*}
kN\equiv N_\alpha+1\quad \text{mod }N_\alpha+2;
\end{equation*}
that is $k_1$ is the minimal positive integer in the class $\left[N_\alpha N^{\varphi(N_\alpha+2)-1}\right]_{\text{mod $N_\alpha+2$}}$ and $k_2$ is the minimal positive integer in the class $\left[(N_\alpha+1)N^{\varphi(N_\alpha+2)-1}\right]_{\text{mod $N_\alpha+2$}}$.
If $\bar{k}=k_1$, then
\begin{equation}
f(Y)=\left(\frac{k_1(N_\alpha+2)}{k_1(N_\alpha+2)+1}\right){\lfloor 2\alpha Y\rfloor}=\left(\frac{1}{1+\frac{1}{k_1(N_\alpha+2)}}\right){\lfloor 2\alpha Y\rfloor},
\end{equation}
and $f(Y)<\lfloor 2\alpha Y\rfloor$.
If $\bar{k}=k_2$, then
\begin{equation}
f(Y)=\left(\frac{k_2(N_\alpha+2)}{k_2(N_\alpha+2)-1}\right){\lfloor 2\alpha Y\rfloor}=\left(\frac{1}{1-\frac{1}{k_2(N_\alpha+2)}}\right){\lfloor 2\alpha Y\rfloor},
\end{equation}
and $f(Y)>\lfloor 2\alpha Y\rfloor$.
$(a_2)$ $N$ and $N_\alpha+2$ are not coprime. In this case
\begin{equation*}
f(Y)=\lfloor 2\alpha Y\rfloor
\end{equation*}
as in the homogeneous case.
\smallskip
$(b)$ If $Y\in \left(\displaystyle\frac{N_\alpha+1+j(N_\alpha+2)}{2\alpha},\frac{(j+1)(N_\alpha+2)+1}{2\alpha}\right),j\ge 0$, then
\begin{equation*}
f(Y)=(j+1)(N_\alpha+2).
\end{equation*}
Note that, in this case, if $Y\in \left(\displaystyle\frac{N_\alpha+1+j(N_\alpha+2)}{2\alpha},\frac{(j+1)(N_\alpha+2)}{2\alpha}\right)$ then $f(Y)>\lfloor 2\alpha Y\rfloor$, while if
$Y\in \left(\displaystyle\frac{(j+1)(N_\alpha+2)}{2\alpha},\frac{(j+1)(N_\alpha+2)+1}{2\alpha}\right)$ then $f(Y)=\lfloor 2\alpha Y\rfloor$.
\smallskip
$(c)$ If $Y\in \left(\displaystyle\frac{N_\alpha-1+j(N_\alpha+2)}{2\alpha},\frac{N_\alpha+1+j(N_\alpha+2)}{2\alpha}\right),j\ge 0$, then we may argue as in case $(a)$.
\begin{example}[The case $N_\alpha=1, N_\beta=2$]
The velocity function is given by
\begin{equation*}
f(Y)=
\begin{cases}
0&\mbox{if}\quad Y<\displaystyle\frac{1}{\alpha},\\
\\
\displaystyle{3k}&\mbox{if}\quad Y\in\left(\displaystyle\frac{3k-1}{2\alpha},\displaystyle\frac{3k+2}{2\alpha}\right),\quad k\geq1;
\end{cases}
\end{equation*}
i.e.,
$$
f(Y)=
3\Bigl\lfloor {2\over 3}\alpha Y+{1\over 3}\Bigr\rfloor.
$$
\end{example}
|
\part{\sc Introducing the subject and motivations}
The present one is a research paper and it contains some new original results. These latter are mostly of mathematical-geometrical character and in our opinion they might be of some interest both for the mathematical scientific community, as well as for that of the theoretical physicists working in the field of supergravity/superstrings. The motivations for the present study is that of analyzing within a general geometric framework some of the recent results \cite{thesearch} relative to the inclusion of candidate inflaton potentials into extended supergravity, the aim being that of singling out general mathematical patterns that eventually can be exported to other examples that include more fields and more multiplets.
\par
From the physicist's viewpoint the obvious ultimate goal is that of singling out possible chains of inclusions of the inflationary models into hierarchically larger and more (super)-symmetric theories finally sheading light on the appropriate place of the inflaton-dynamics, which is revealed by observational cosmology, within a duely unified theory of all interactions.
\par
Notwithstanding the above mentioned specific motivations, the geometrical results presented in this paper, have an intrinsic mathematical interest and moreover admit different applications than those in cosmology within the very framework of supergravities and their gaugings.
For this reason we have chosen to make this paper self-consistent and readable by a wider audience of both mathematicians and physicists working in fields different from that of supergravity cosmological models, by presenting in a systematic way all the mathematical definitions and structures that we utilize in the original part of our work. In the same spirit, in order to provide our reader with orientation, the paper is subdivided into Parts and sections.
\par
The present first part provides a conceptual introduction and the physical motivations in relation to current research.
\par The second part provides the definitions of K\"ahler-Hodge, Special K\"ahler and Quaternionic K\"ahler manifolds. Then introduces the $c$-map \cite{sabarwhal} and illustrates how in Quaternionic K\"ahler manifolds lying in the image of $c$-map, all the quaternionic structures, in particular the HyperK\"ahler two-forms, the $\mathfrak{su}(2)$-connection and the tri-holomorphic moment maps of isometries can all be constructed purely in terms of Special geometry data.
\par
The third part discusses abelian gaugings of hypermultiplet isometries in $\mathcal{N}=2$ supergravity. Using the general mathematical formulae derived in Part Two we discuss generic properties and features of the ensuing scalar potentials.
\par
The fourth part presents concrete examples. In the particular case of the $c$-map of the $S^3$ model we retrieve the results obtained in \cite{thesearch}. Another relevant Special K\"ahler Geometry that we consider corresponds to that of the symmetric space $\mathrm{Sp(6,\mathbb{R})}/\mathrm{SU(3)\times U(1)}$. For this model we provide an in depth, full fledged construction. Analyzing its Quaternionic K\"ahler extension by means of the $c$-map we are able to generalize the results of \cite{thesearch} showing that they fall into a general pattern. Our detailed construction may have applications both in the cosmological perspective of the present paper and in the classification of Black-Hole solutions, possibly also in other contexts.
\par
Part five summarizes the results obtained in the two considered examples and shows that they unveil a deep and universal structure. We briefly summarize the organization of $\mathcal{N}=2$ scalar manifolds that are symmetric spaces into Tits--Satake universality classes
\cite{titsusataku}. Such a concept already proved to be of high value in relation to the construction and classification of supergravity black hole solutions \cite{noinilpotenti}. It is equally effective and precious in relation to the $c$-map and cosmological models. Indeed relying on these structures we are able to show that the inclusion of Starobinsky-like models into extended supergravity theories has a universal character being associated with the gauging of the universal sub-Tits-Satake subalgebra $\mathbb{G}_{\mathrm{subTS}} \, = \, \mathfrak{sl}(2) \times \mathfrak{sl}(2) \times \mathfrak{sl}(2)$. This also leads to the prediction that available values of $\alpha$ for the so named $\alpha$-attractors \cite{alfatrattori} are just $\alpha \, = \, 1,\, \frac{2}{3},\, \frac{1}{3}$. After this conceptual elaboration, part five contains our conclusions and remarks on further perspectives.
\section{\sc Introduction}
The recent observational results on the power-spectrum of the Cosmological Microwave Background radiation \cite{Ade:2013uln},\cite{Ade:2013zuv},\cite{Hinshaw:2012aka},\cite{biceppo} have stirred renewed interest in one-field inflationary cosmological models \cite{Starobinsky:1980te},\cite{lindefund},\cite{guthfund},\cite{steinhardfund}. Indeed the type of cosmology \cite{cosmology},\cite{pietrocosmobook} that seems to be consistent with observations is that based on the simplest scenario of just one scalar field $\phi$ (\textit{the inflaton}) minimally coupled to Einstein Gravity and endowed with a suitable scalar potential $V(\phi)$.
\par
In view of this, several studies have been devoted to the problem of including into supergravity potentials $V(\phi)$ that produce an early inflationary phase and have good cosmological properties, \cite{johndimitri},\cite{Ketov:2010qz},\cite{Ketov:2012jt},\cite{Kallosh:2013hoa},\cite{Kallosh:2013lkr},\cite{Farakos:2013cqa},
\cite{Kallosh:2013maa},\cite{Kallosh:2013daa},\cite{Kallosh:2013tua},\cite{alfatrattori},\cite{Ferrara:2014cca}. A separate investigation was also devoted to determine a list of one field integrable potentials \cite{augustopietrosasha} and to discuss their possible inclusion into $\mathcal{N}=1$ supergravity by means of suitable superpotentials \cite{noiGaugings}. It was shown that such type of inclusion is quite difficult and can be realized only in very few cases \cite{noiGaugings}.
\par
Notwithstanding the difficulties with the inclusion of inflaton potentials by means of a superpotential (F-terms), approximately one year ago, in the seminal paper \cite{minimalsergioKLP} it was instead pointed out that every positive-definite potential $V(\phi)$ can be minimally included into $\mathcal{N}=1$ supergravity as a D-term (see \cite{cfgv}, \cite{castdauriafre} and \cite{primosashapietro} for the complete structure of matter coupled $\mathcal{N}=1$ supergravity). It suffices to introduce a K\"ahler one-fold $\Sigma$ with an abelian group of isometries $\mathrm{G}_{\mathrm{iso}}$ and a K\"ahler potential $\mathcal{K}$ related to the potential $V$ by an appropriate differential relation which allows to interpret this latter as the square of the moment-map of the holomorphic Killing vector generating $\mathrm{G}_{\mathrm{iso}}$. From the mathematical point of view this new vision stimulated the formulation of the concept of $D$-map and its extensive study in \cite{primosashapietro} and \cite{piesashatwo}. The essential mechanism behind the $D$-map is the just the Brout-Englert-Higgs mechanism as it is realized in supergravity. Gauging an isometry of the K\"ahlerian scalar manifold $\mathcal{M}_{K}$ one generates both mass terms for the fermions and a scalar potential that is the square of the moment map $\mathcal{P}_k$ of the corresponding Killing vector. The gauge vector field $A_\mu$ utilized to gauge the considered isometry becomes massive by eating up one of the two scalar fields composing the Wess-Zumino multiplet, while its partner remains in the Lagrangian as a degree of freedom of spin zero, self interacting by means of the $D$-term potential \cite{VanProeyen:1979ks},\cite{Freedman:1976uk}.
\par
Among the positive definite scalar potentials that can be included in supergravity in this way there are the Starobinsky-like potentials \cite{Starobinsky:1980te}:
\begin{equation}\label{starobombolo}
V_{Starobinsky-like} \, = \, \mbox{const} \, \times \, \left( 1\, - \, \exp\left [ - \, \sqrt{\frac{2}{3 \, \alpha}} \, \phi\right]\right)^2\,.
\end{equation}
When $\alpha \, = \, 1$, the potential (\ref{starobombolo}) emerges in the second derivative supergravity dual of a higher derivative $R+R^2$ supergravity model \cite{Whitt:1984pd},
\cite{Cecotti:1987qe},\cite{Cecotti:1987sa},\cite{Ferrara:2013wka},\cite{Ferrara:2013kca}. Due to the high relevance of these potentials in phenomenology, several studies were devoted to the mechanisms for their inclusion in supergravity\footnote{The inclusion of the original Starobinsky $R+R^2$ model in $\mathcal{N}=2$ supergravity was discussed in \cite{thesearch} (see \cite{Ketov:2014qoa} for an earlier discussion) where it is shown to be dual to an $\mathcal{N}=2$ model with two long massive vector multiplets
on a $\mathcal{N}=2$ Minkowski vacuum. The scalaron is subject to a scalar potential of the form (\ref{starobombolo}) with $\alpha=1$. The models considered here are not dual to the $R+R^2$ Starobinsky model and share with it only the scalaron potential, which we shall refer to as the \emph{Starobinsky potential}. If $\alpha\neq 1$ the potential (\ref{starobombolo}) will be called \emph{Starobinsky-like}.}. In their minimal $D$-term realization, the Starobinsky-like potentials were shown to be generated by the gauging of the parabolic subgroups of $\mathrm{SL(2,\mathbb{R})}$ in a theory where the K\"ahler one-fold $\Sigma$ is the homogeneous space $\frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}$, with a value of the curvature directly related to $\alpha$ by:
\begin{equation}\label{kreptosio}
R_{\alpha} \, = \, - \, \frac{2}{3\,\alpha}
\end{equation}
The case $\alpha \, = \, 1$ which is the proper Starobinsky potential corresponds to a realization of the $\frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}$ geometry which is not only K\"ahlerian but actually Special K\"ahlerian. Indeed it is the case of the $S^3$ model.
\par
In more general terms the minimal supergravity approach and the $D$-map from potentials to K\"ahler geometry posed the question of relating the type of generated potential with the global topology of the isometry whose gauging generates it. This issue was thoroughly studied in \cite{pietrosergiosasha1},\cite{pietrosergiosasha2}. Relying on notions developed by Gromov et al \cite{Gromov1985},\cite{Gromov1987}, in \cite{pietrosergiosasha1},\cite{pietrosergiosasha2} it was shown that the global topology of the isometry group can be characterized as, \textit{elliptic, parabolic or hyperbolic} on general K\"ahler manifolds $\Sigma$ of non-positive curvature and that the three cases lead to different distinctive properties of the moment-maps, that either have a fixed-point at finite distance (elliptic case) or a fixed point on the boundary (parabolic) or two fixed point on the boundary (hyperbolic). In the case of constant negative curvature K\"ahler one-folds, \textit{i.e.} of $\Sigma_\alpha \, = \, \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}$ manifolds, it was shown in \cite{pietrosergiosasha1} and \cite{pietrosergiosasha2} that one generates the following three potentials:
\begin{equation}\label{ginolullobrigo}
V(\phi) \, = \, \left \{ \begin{array}{ccccc}
V_{elliptic}& = & \left( \cosh \left[\sqrt{\frac{2}{3 \, \alpha}} \, \phi\right] \, - \, \kappa\right)^2 & \Leftrightarrow & \mbox{from gauging of an elliptic subgroup} \\
V_{hyperbolic}& = & \left( \sinh \left[\sqrt{\frac{2}{3 \, \alpha}} \, \phi\right] \, - \, \kappa\right)^2 & \Leftrightarrow & \mbox{from gauging of a hyperbolic subgroup}\\
V_{parabolic}& = &\left( \exp\left [ - \, \sqrt{\frac{2}{3 \, \alpha}} \, \phi\right]\, - \, \kappa\right)^2 & \Leftrightarrow & \mbox{from gauging of a parabolic subgroup}\\
\end{array}
\right.
\end{equation}
where $\alpha$ parameterizes the curvature of the one-fold $\Sigma_\alpha$ according to eq. (\ref{kreptosio}) and the parameter $\kappa\, = \, \pm1,0$ is interpreted as the coupling constant of a Fayet Iliopoulos term \cite{Fayet:1974jb}. Utilizing instead a flat K\"ahler one-fold $\Sigma_{flat}\, = \, \mathbb{C}$, it was shown in \cite{pietrosergiosasha1} and \cite{pietrosergiosasha2} that the gauging of an elliptic isometry yields a mexican hat Brout-Englert-Higgs potential, while the quadratic potential of chaotic inflation is generated by a translation parabolic gauging.
\par
In view of the above conclusions an obvious and very much relevant question concerns the possible inclusion of the potentials (\ref{ginolullobrigo}) into extended supergravity, in particular $\mathcal{N}=2$. Such a question amounts to asking whether there are consistent one-field truncations of appropriate gauged extended supergravities that produce the considered potentials. Such a question includes two issues:
\begin{enumerate}
\item the choice of a gauging,
\item the existence of an appropriate consistent truncation.
\end{enumerate}
The first of these two issues is relevant for another important problem, that of constructing and classifying de Sitter or Minkowskian, stable or metastable, vacua in supergravity and it was extensively addressed in that context, \cite{mapietoine},\cite{gkp},\cite{uplift},\cite{kklt},
\cite{scrucca},\cite{noscale}. Let us spend some words on the second issue.
\par
The general form of the scalar field equations in supergravity is that of a $\sigma$-model with a potential, namely the following one:
\begin{equation}\label{guberator}
0 \, = \, \square \, \phi^I \, + \, \Gamma^I_{JK} (\phi)\, \partial_\mu \phi^J \, \partial^\mu \, \phi^K \, + \, \frac{\partial}{\partial \phi^I} \, V(\phi)
\end{equation}
where the scalars $\phi^I$ are coordinates of the target Riemaniann manifold $\mathcal{M}_{scalar}$, by $\Gamma^I_{JK} (\phi)$ we have denoted the Levi-Civita connection on the former and $V$ is the potential. The problem of consistent truncation boils down to the following. We consider the embedding of a $\Sigma_\alpha$ K\"ahlerian one-fold into the scalar manifold:
\begin{equation}\label{pullabacca}
\pi \, : \, \Sigma_\alpha \, \mapsto \, \mathcal{M}_{scalar}
\end{equation}
and we require that the pull-back of the field equations (\ref{guberator}) onto the surface $\Sigma_\alpha $:
\begin{equation}\label{guberator2}
0 \, = \, \phi^\star\left[\square \, \phi^I \, + \, \Gamma^I_{JK} (\phi)\, \partial_\mu \phi^J \, \partial^\mu \, \phi^K \, + \, \frac{\partial}{\partial \phi^I} \, V(\phi)\right]
\end{equation}
be consistent, namely that it reproduces always the same equations for all values of $I$. What are the restrictions on the embedding $\pi$ that guarantee such a consistency? This question can be answered in a general form when $\mathcal{M}_{scalar}$ is a symmetric homogeneous space $\mathrm{G/H}$, as it happens most frequently in supergravity models. Considering as usual the symmetric decomposition of the Lie algebra $\mathbb{G}$:
\begin{eqnarray}
\mathbb{G} & =& \mathbb{H} \, \oplus \, \mathbb{K} \nonumber\\
\left [\mathbb{H} \, , \, \mathbb{H}\right] & \subset & \mathbb{H} \nonumber\\
\left [\mathbb{H} \, , \, \mathbb{K}\right] & \subset & \mathbb{K} \nonumber\\
\left [\mathbb{K} \, , \, \mathbb{K}\right] & \subset & \mathbb{H} \label{simmetricoGH}
\end{eqnarray}
the embedding (\ref{pullabacca}) induces a homomorphism:
\begin{equation}\label{girolamo}
\pi \, : \, \mathfrak{sl}(2,\mathbb{R}) \, \mapsto \, \mathbb{G} \quad ; \quad \pi \, : \, \mathbb{O}(2) \, \mapsto \, \mathbb{H}\quad ; \quad \pi \, : \, \mathbb{K}_2 \, \mapsto \, \mathbb{K}
\end{equation}
where:
\begin{equation}\label{birillo2}
\mathfrak{sl}(2,\mathbb{R}) \, =\, \mathbb{O}(2)\, \oplus \, \mathbb{K}_2
\end{equation}
Let us introduce the centralizer algebra of the image of $\mathbb{O}(2)$ in $\mathbb{H}$, namely
\begin{equation}\label{gunther}
\forall \,\mathfrak{g} \, \in \, \mathbb{H} \quad : \quad \mathfrak{g}\, \in \, \mathbb{N}_\pi \quad \mbox{iff} \quad \left[\mathfrak{g}\, , \, \pi\left(\mathbb{O}(2)\right)\right] \, = \,0
\end{equation}
The subspace $\mathbb{K}$ decomposes into irreducible representations of $\pi\left(\mathbb{O}(2)\right)\, \oplus \, \mathbb{N}_\pi$. The embedding $\pi$ leads to a consistent truncation if such a decomposition has the following structure:
\begin{equation}\label{contrallus}
\mathbb{K} \, = \, \underbrace{\left(\mathbf{2}_1\,|\, \mathbf{1}\right)}_{\pi(\mathbb{K}_2 )} \, \oplus_{i=1}^m \, \left(\mathbf{2}_{q_i}\,|\, \mathbf{D}_i\right) \, \oplus \, \left(\mathbf{1}\,|\, \mathbf{D}_0\right)
\end{equation}
where $\mathbf{2}_{q}$ denotes a doublet representation of $\pi\left(\mathbb{O}(2)\right)$ with charge $q$ and all $\mathbf{D}_i$ and $\mathbf{D}_0$ are transitive representations, namely carry a \textit{color} of $\mathbb{N}_\pi$, the only
$\mathbb{N}_\pi$-singlet being the subspace ${\pi(\mathbb{K}_2 )}$ tangent to the embedded one-fold $\Sigma_\alpha$.
In this case setting to zero all the colored fields (those not in ${\pi(\mathbb{K}_2 )}$ ) is consistent because the equation of a colored field cannot receive contribution from two colorless singlets.
\par
In the case of $\mathcal{N}=2$ supergravity the scalar manifold is actually the direct product of two manifolds: a special K\"ahler manifold $\mathcal{M}_{SK}$ that contains the vector multiplet scalars and a Quaternionic K\"ahler $\mathcal{M}_{Q}$ manifold describing the hypermultiplet scalars. Hence one comes to the question whether the one-fold $\Sigma_\alpha$ associated with the inflaton is to be embedded in $\mathcal{M}_{SK}$ or in $\mathcal{M}_{Q}$. The significant advance introduced by \cite{thesearch} is the scenario in which $\Sigma_\alpha$ goes into $\mathcal{M}_{Q}$ and its abelian isometry is gauged by means of vector muliplets assigned to the so named Minimal Coupling Special Geometry. This choice allows for a generic stabilization of the vector multiplet scalars and allows to focus only on the properties of the Quaternionic K\"ahler manifold $\mathcal{M}_{Q}$. In \cite{thesearch} the authors considered the cases where
\begin{equation}\label{g22su2su2}
\mathcal{M}_{Q} \, = \, \frac{\mathrm{G_{(2,2)}}}{\mathrm{SU(2)} \times \mathrm{SU(2)}}\,\,\,\mbox{and}\,\,\,\,\,\, \mathcal{M}_{Q} \, =\,\frac{{\rm SO}(1,4)}{{\rm SO}(4)}\,,
\end{equation}
and it was shown how all the potentials (\ref{ginolullobrigo}) can be included with possible values of $\alpha\, = \, 1,\,\frac{2}{3},\, \frac{1}{3}$ (the value $\alpha=2/3$ could be obtained only for the quaternionic projective space). In the present paper we address the same question for generic homogeneous symmetric Quaternionic K\"ahler manifolds and we eventually arrive at the same conclusion. The three potentials (\ref{ginolullobrigo}) can always be embedded with the same possible values of $\alpha$. Our main weapons in reaching such a conclusion and proving its generality are two:
\begin{description}
\item[a)] The $c$-map from Special K\"ahler manifolds $\mathcal{M}_{\mathcal{SK}}$ to Quaternionic K\"ahler manifolds $\mathcal{M}_{\mathcal{Q}}$
\item[b)] The Tits-Satake projection and the Tits-Satake universality classes.
\end{description}
Combining these two mathematical instruments we are able to look inside the hypermultiplet manifold and to single out its inner core which is the Special K\"ahler STU-model. Working in this framework we were able to derive simple general formulae for the triholomorphic moment maps which besides their present use in cosmology might admit several other interesting and useful applications. Similarly the Tits Satake structure allows for a deep understanding of the universal character of the Starobinsky like potentials and opens the way to their further uplift to higher $\mathcal{N}$ theories in the perspective of finding \textit{microscopic interpretations} of the gaugings that generates them.
\newpage
\part{\sc Mathematical Theory of the $c$-map}
As announced in the introductory part, the aim of this section of the paper is two-fold. On the one hand we want to
review the general geometric structure of homogeneous symmetric Quaternionic K\"ahler manifolds which are in the image of the c-map and to present
in a unified fashion general analytic formulae for the complex structures, $\mathfrak{su}(2)$-connection and tri-holomorphic moment maps.
The ultimate goal is the use of such mathematical instruments in the quest of retrieving \textit{inflaton potentials} inside properly gauged $\mathcal{N}=2$ supergravity theories. On the other hand we aim at a concise, yet comprehensive presentation of Special K\"ahler geometry, Quaternionic K\"ahler geometry and the $c$-map which might be readable by mathematicians, in particular differential geometers and Lie algebrists. Indeed this mathematical subject was mostly developed by theoretical physicists and it is not widely known in the mathematical community. This is rather unfortunate, especially in the light of its relevance to the issue of the classification of nilpotent orbits, which, instead, is largely explored by mathematicians, and is quite relevant both to the classification of black hole solutions and emerges now as quite important in cosmological issues. Unfortunately the mathematical results on nilpotent orbits are obtained within frameworks that make no reference to the very particular structures of special geometry, quaternionic geometry and the magic relations of the $c$-map. Spreading awareness of this sophisticated and beautiful mathematics among mathematicians might be beneficial to both communities and it is the second aim of the following sections. A further, humbler, yet quite important aim, pursued here, is that of establishing some unified and systematic notations that might be utilized in subsequent publications devoted to a systematic exploration of the reach field of \textit{gaugings}, \textit{consistent truncations}, \textit{de Sitter vacua} and \textit{inflaton potentials}.
\section{\sc Special K\"ahler Geometry}
Special K\"ahler geometry in special coordinates was introduced
in 1984--85 by B. de Wit et al.
and E. Cremmer et al. (see pioneering papers in \cite{SKG}), where the coupling of
$\mathcal{N}=2$ vector multiplets to $\mathcal{N}=2$ supergravity was fully determined. The
more intrinsic definition of special K\"ahler geometry in terms of
symplectic bundles is due to Strominger (1990), who
obtained it in connection with the moduli spaces of
Calabi--Yau compactifications, (see \cite{defiskg}). The coordinate-independent description
and derivation of special K\"ahler geometry in the context of $\mathcal{N}=2$
supergravity is due to Castellani, D'Auria, Ferrara \cite{skgintrinsic}
and to D'Auria, Ferrara, Fre' (1991)\cite{D'Auria:1991fj}. Homogenous symmetric special K\"ahler manifolds were classified before by Cremmer and Van Proyen in \cite{ToineCremmerOld}. An early review in modern mathematical language is provided by \cite{mylecture}.
The structure of isometry group for both Special K\"ahler and Quaternionic K\"ahler manifolds was extensively studied in \cite{specHomgeo},\cite{vanderseypen}
\par
Let us summarize the relevant concepts and definitions
\subsection{\sc Hodge--K\"ahler manifolds}
\def{M(k, \IC)}{{M(k, \relax\,\hbox{$\vrule height1.5ex width.4pt depth0pt\kern-.3em{\rm C}$})}}
Consider a {\sl line bundle} ${\cal L}
{\stackrel{\pi}{\longrightarrow}} {\cal M}$ over a K\"ahler
manifold ${\cal M}$. By definition this is a holomorphic vector
bundle of rank $r=1$. For such bundles the only available Chern
class is the first:
\begin{equation}
c_1 ( {\cal L} ) \, =\, \o{i}{2}
\, {\bar \partial} \,
\left ( \, h^{-1} \, \partial \, h \, \right )\, =
\, \o{i}{2} \,
{\bar \partial} \,\partial \, \mbox{log} \, h
\label{chernclass23}
\end{equation}
where the 1-component real function $h(z,{\bar z})$ is some
hermitian fibre metric on ${\cal L}$. Let $\xi (z)$ be a
holomorphic section of the line bundle ${\cal L}$: noting that
under the action of the operator ${\bar
\partial} \,\partial \, $ the term $\mbox{log} \left ({\bar \xi}({\bar z})
\, \xi (z) \right )$ yields a vanishing contribution, we conclude that
the formula in eq.(\ref{chernclass23}) for the first Chern class can be
re-expressed as follows:
\begin{equation}
c_1 ( {\cal L} ) ~=~\o{i}{2} \,
{\bar \partial} \,\partial \, \mbox{log} \,\parallel \, \xi(z) \, \parallel^2
\label{chernclass24}
\end{equation}
where $\parallel \, \xi(z) \, \parallel^2 ~=~h(z,{\bar z}) \,
{\bar \xi}({\bar z}) \,
\xi (z) $ denotes
the norm of the holomorphic section $\xi (z) $.
\par
Eq.(\ref{chernclass24}) is the starting point for the definition
of Hodge--K\"ahler manifolds. A K\"ahler manifold ${\cal M}$ is a
Hodge manifold if and only if there exists a line bundle ${\cal L}
{\stackrel{\pi}{\longrightarrow}} {\cal M}$ such that its first
Chern class equals the cohomology class of the K\"ahler two-form
$\mathrm{K}$:
\begin{equation}
c_1({\cal L} )~=~\left [ \, \mathrm{K} \, \right ]
\label{chernclass25}
\end{equation}
\par
In local terms this means that there is a holomorphic section $\xi
(z)$ such that we can write
\begin{equation}
\mathrm{K}\, =\, \o{i}{2} \, g_{ij^{\star}} \, dz^{i} \, \wedge
\, d{\bar z}^{j^{\star}} \, = \, \o{i}{2} \, {\bar \partial}
\,\partial \, \mbox{log} \,\parallel \, \xi (z) \,
\parallel^2
\label{chernclass26}
\end{equation}
Recalling the local expression of the K\"ahler metric
in terms of the K\"ahler potential
$ g_{ij^{\star}}\, =\, {\partial}_i \, {\partial}_{j^{\star}}
{\mathcal{K}} (z,{\bar z})$,
it follows from eq.(\ref{chernclass26}) that if the
manifold ${\cal M}$ is a Hodge manifold,
then the exponential of the K\"ahler potential
can be interpreted as the metric
$h(z,{\bar z}) \, = \, \exp \left ( {\cal K} (z,{\bar z})\right )$
on an appropriate line bundle ${\cal L}$.
\par
\subsection{\sc Connection on the line bundle}
On any complex line bundle ${\cal L}$ there is a canonical hermitian connection defined as :
\begin{equation}
\begin{array}{ccccccc}
{\theta}& \equiv & h^{-1} \, \partial \, h = {\o{1}{h}}\, \partial_i h \,
dz^{i} &; &
{\bar \theta}& \equiv & h^{-1} \, {\bar \partial} \, h = {\o{1}{h}} \,
\partial_{i^\star} h \,
d{\bar z}^{i^\star} \cr
\end{array}
\label{canconline}
\end{equation}
For the line-bundle advocated by the Hodge-K\"ahler structure we have
\begin{equation}
\left [ \, {\bar \partial}\,\theta \, \right ] \, = \,
c_1({\cal L}) \, = \, [\mathrm{K}]
\label{curvc1}
\end{equation}
and since the fibre metric $h$ can be identified with the
exponential of the K\"ahler potential we obtain:
\begin{equation}
\begin{array}{ccccccc}
{\theta}& = & \partial \,{\cal K} = \partial_i {\cal K}
dz^{i} & ; &
{\bar \theta}& = & {\bar \partial} \, {\cal K} =
\partial_{i^\star} {\cal K}
d{\bar z}^{i^\star}\cr
\end{array}
\label{curvconline}
\end{equation}
To define special K\"ahler geometry, in addition to the afore-mentioned line--bundle
${\cal L}$ we need a flat holomorphic vector bundle ${\cal SV}
\, \longrightarrow \, {\cal M}$ whose sections play an important role in the construction of the supergravity Lagrangians. For reasons
intrinsic to such constructions the rank of the vector bundle ${\cal SV}$ must be $2\, n_V$ where $n_V$ is the total number of vector fields in the theory. If we have $n$-vector multiplets the total number of vectors is $n_V = n+1$ since, in addition to the vectors of the vector multiplets, we always have the graviphoton sitting in the graviton multiplet. On the other hand the total number of scalars is $2 n$. Suitably paired into $n$-complex fields $z^i$, these scalars span the $n$ complex dimensions of the base manifold ${\cal M}$ to the rank $2n+2$ bundle ${\cal SV}
\, \longrightarrow \, {\cal M}$.
\par
In the sequel we make extensive use of covariant derivatives with
respect to the canonical connection of the line--bundle ${\cal
L}$. Let us review its normalization. As it is well known there
exists a correspondence between line--bundles and
$\mathrm{U(1)}$--bundles. If $\mbox{exp}[f_{\alpha\beta}(z)]$ is
the transition function between two local trivializations of the
line--bundle ${\cal L} {\stackrel{\pi}{\longrightarrow}} {\cal
M}$, the transition function in the corresponding principal
$\mathrm{U(1)}$--bundle ${\cal U} \, \longrightarrow {\cal M}$ is
just $\mbox{exp}[{\rm i}{\rm Im}f_{\alpha\beta}(z)]$ and the
K\"ahler potentials in two different charts are related by: ${\cal
K}_\beta = {\cal K}_\alpha + f_{\alpha\beta} + {\bar
{f}}_{\alpha\beta}$. At the level of connections this
correspondence is formulated by setting: $\mbox{
$\mathrm{U(1)}$--connection} \equiv {\cal Q} \, = \,
\mbox{Im} \theta = -{\o{\rm i}{2}} \left ( \theta - {\bar
\theta} \right)$. If we apply this formula to the case of the
$\mathrm{U(1)}$--bundle ${\cal U} \, \longrightarrow \, {\cal M}$
associated with the line--bundle ${\cal L}$ whose first Chern
class equals the K\"ahler class, we get:
\begin{equation}
{\cal Q} = {\o{\rm i}{2}} \left ( \partial_i {\cal K}
dz^{i} -
\partial_{i^\star} {\cal K}
d{\bar z}^{i^\star} \right )
\label{u1conect}
\end{equation}
Let now
$\Phi (z, \bar z)$ be a section of ${\cal U}^p$. By definition its
covariant derivative is $ \nabla \Phi = (d - i p {\cal Q}) \Phi $
or, in components,
\begin{equation}
\begin{array}{ccccccc}
\nabla_i \Phi &=&
(\partial_i + {1\over 2} p \partial_i {\cal K}) \Phi &; &
\nabla_{i^*}\Phi &=&(\partial_{i^*}-{1\over 2} p \partial_{i^*} {\cal K})
\Phi \cr
\end{array}
\label{scrivo2}
\end{equation}
A covariantly holomorphic section of ${\cal U}$ is defined by the equation:
$ \nabla_{i^*} \Phi = 0 $.
We can easily map each section $\Phi (z, \bar z)$
of ${\cal U}^p$
into a section of the line--bundle ${\cal L}$ by setting:
\begin{equation}
\widetilde{\Phi} = e^{-p {\cal K}/2} \Phi \, .
\label{mappuccia}
\end{equation}
With this position we obtain:
\begin{equation}
\begin{array}{ccccccc}
\nabla_i \widetilde{\Phi}& =&
(\partial_i + p \partial_i {\cal K})
\widetilde{\Phi}& ; &
\nabla_{i^*}\widetilde{\Phi}&=& \partial_{i^*} \widetilde{\Phi}\cr
\end{array}
\end{equation}
Under the map of eq.(\ref{mappuccia}) covariantly holomorphic sections
of ${\cal U}$ flow into holomorphic sections of ${\cal L}$
and viceversa.
\subsection{\sc Special K\"ahler Manifolds}
We are now ready to give the first of two equivalent definitions of special K\"ahler
manifolds:
\begin{definizione}
A Hodge K\"ahler manifold is {\bf Special K\"ahler (of the local type)}
if there exists a completely symmetric holomorphic 3-index section $W_{i
j k}$ of $(T^\star{\cal M})^3 \otimes {\cal L}^2$ (and its
antiholomorphic conjugate $W_{i^* j^* k^*}$) such that the following
identity is satisfied by the Riemann tensor of the Levi--Civita
connection:
\begin{eqnarray}
\partial_{m^*} W_{ijk}& =& 0 \quad \partial_m W_{i^* j^* k^*}
=0 \nonumber \\
\nabla_{[m} W_{i]jk}& =& 0
\quad \nabla_{[m}W_{i^*]j^*k^*}= 0 \nonumber \\
{\cal R}_{i^*j\ell^*k}& =& g_{\ell^*j}g_{ki^*}
+g_{\ell^*k}g_{j i^*} - e^{2 {\cal K}}
W_{i^* \ell^* s^*} W_{t k j} g^{s^*t}
\label{specialone}
\end{eqnarray}
\label{defspecial}
\end{definizione}
In the above equations $\nabla$ denotes the covariant derivative with
respect to both the Levi--Civita and the $\mathrm{U(1)}$ holomorphic connection
of eq.(\ref{u1conect}).
In the case of $W_{ijk}$, the $\mathrm{U(1)}$ weight is $p = 2$.
\par
Out of the $W_{ijk}$ we can construct covariantly holomorphic
sections of weight 2 and - 2 by setting:
\begin{equation}
C_{ijk}\,=\,W_{ijk}\,e^{ {\cal K}} \quad ; \quad
C_{i^\star j^\star k^\star}\,=\,W_{i^\star j^\star k^\star}\,e^{ {\cal K}}
\label{specialissimo}
\end{equation}
The flat bundle mentioned in the previous subsection apparently does not appear in this definition of special geometry.
Yet it is there. It is indeed the essential ingredient in the second definition whose equivalence to the first we shall
shortly provide.
\par
Let ${\cal L} {\stackrel{\pi}{\longrightarrow}} {\cal M}$ denote
the complex line bundle whose first Chern class equals the
cohomology class of the K\"ahler form $\mathrm{K}$ of an
$n$-dimensional Hodge--K\"ahler manifold ${\cal M}$. Let ${\cal
SV} \, \longrightarrow \,{\cal M}$ denote a holomorphic flat
vector bundle of rank $2n+2$ with structural group
$\mathrm{Sp(2n+2,\mathbb{R})}$. Consider tensor bundles of the
type ${\cal H}\,=\,{\cal SV} \otimes {\cal L}$. A typical
holomorphic section of such a bundle will be denoted by ${\Omega}$
and will have the following structure:
\begin{equation}
{\Omega} \, = \, {\twovec{{X}^\Lambda}{{F}_ \Sigma} } \quad
\Lambda,\Sigma =0,1,\dots,n
\label{ololo}
\end{equation}
By definition
the transition functions between two local trivializations
$U_i \subset {\cal M}$ and $U_j \subset {\cal M}$
of the bundle ${\cal H}$ have the following form:
\begin{equation}
{\twovec{X}{ F}}_i=e^{f_{ij}} M_{ij}{\twovec{X}{F}}_j
\end{equation}
where $f_{ij}$ are holomorphic maps $U_i \cap U_j \, \rightarrow
\,\relax\,\hbox{$\vrule height1.5ex width.4pt depth0pt\kern-.3em{\rm C}$} $
while $M_{ij}$ is a constant $\mathrm{Sp(2n+2,\mathbb{R})}$ matrix. For a consistent
definition of the bundle the transition functions are obviously
subject to the cocycle condition on a triple overlap:
$e^{f_{ij}+f_{jk}+f_{ki}} = 1 $ and $ M_{ij} M_{jk} M_{ki} = 1 $.
\par
Let ${\rm i}\langle\ \vert\ \rangle$ be the compatible
hermitian metric on $\cal H$
\begin{equation}
{\rm i}\langle \Omega \, \vert \, \bar \Omega \rangle \, \equiv \,-
{\rm i} \Omega^T \twomat {0} {\relax{\rm 1\kern-.35em 1}} {-\relax{\rm 1\kern-.35em 1}}{0} {\bar \Omega}
\label{compati}
\end{equation}
\begin{definizione}
We say that a Hodge--K\"ahler manifold ${\cal M}$
is {\bf special K\"ahler} if there exists
a bundle ${\cal H}$ of the type described above such that
for some section $\Omega \, \in \, \Gamma({\cal H},{\cal M})$
the K\"ahler two form is given by:
\begin{equation}
\mathrm{K}= \o{\rm i}{2}
\partial \bar \partial \, \mbox{\rm log} \, \left ({\rm i}\langle \Omega \,
\vert \, \bar \Omega
\rangle \right )=\frac{i}{2}\,g_{i, j^*}\,dz^i\wedge d\bar{z}^{j^*}\,.
\label{compati1} .
\end{equation}
\end{definizione}
From the point of view of local properties, eq.(\ref{compati1})
implies that we have an expression for the K\"ahler potential
in terms of the holomorphic section $\Omega$:
\begin{equation}
{\cal K}\, = \, -\mbox{log}\left ({\rm i}\langle \Omega \,
\vert \, \bar \Omega
\rangle \right )\,
=\, -\mbox{log}\left [ {\rm i} \left ({\bar X}^\Lambda F_\Lambda -
{\bar F}_\Sigma X^\Sigma \right ) \right ]
\label{specpot}
\end{equation}
The relation between the two definitions of special manifolds is
obtained by introducing a non--holomorphic section of the bundle
${\cal H}$ according to:
\begin{equation}
V \, = \, \twovec{L^{\Lambda}}{M_\Sigma} \, \equiv \, e^{{\cal K}/2}\Omega
\,= \, e^{{\cal K}/2} \twovec{X^{\Lambda}}{F_\Sigma}
\label{covholsec}
\end{equation}
so that eq.(\ref{specpot}) becomes:
\begin{equation}
1 \, = \, {\rm i}\langle V \,
\vert \, \bar V
\rangle \,
= \, {\rm i} \left ({\bar L}^\Lambda M_\Lambda -
{\bar M}_\Sigma L^\Sigma \right )
\label{specpotuno}
\end{equation}
Since $V$ is related to a holomorphic section by eq.(\ref{covholsec})
it immediately follows that:
\begin{equation}
\nabla_{i^\star} V \, = \, \left ( \partial_{i^\star} - {\o{1}{2}}
\partial_{i^\star}{\cal K} \right ) \, V \, = \, 0
\label{nonsabeo}
\end{equation}
On the other hand, from eq.(\ref{specpotuno}), defining:
\begin{eqnarray}
U_i & = & \nabla_i V = \left ( \partial_{i} + {\o{1}{2}}
\partial_{i}{\cal K} \right ) \, V \equiv
\twovec{f^{\Lambda}_{i} }{h_{\Sigma\vert i}}\nonumber\\
{\bar U}_{i^\star} & = & \nabla_{i^\star}{\bar V} = \left ( \partial_{i^\star} + {\o{1}{2}}
\partial_{i^\star}{\cal K} \right ) \, {\bar V} \equiv
\twovec{{\bar f}^{\Lambda}_{i^\star} }{{\bar h}_{\Sigma\vert i^\star}}
\label{uvector}
\end{eqnarray}
it follows that:
\begin{equation}
\label{ctensor}
\nabla_i U_j = {\rm i} C_{ijk} \, g^{k\ell^\star} \, {\bar U}_{\ell^\star}
\end{equation}
where $\nabla_i$ denotes the covariant derivative containing both
the Levi--Civita connection on the bundle ${\cal TM}$ and the
canonical connection $\theta$ on the line bundle ${\cal L}$.
In eq.(\ref{ctensor}) the symbol $C_{ijk}$ denotes a covariantly
holomorphic (
$\nabla_{\ell^\star}C_{ijk}=0$) section of the bundle
${\cal TM}^3\otimes{\cal L}^2$ that is totally symmetric in its indices.
This tensor can be identified with the tensor of eq.(\ref{specialissimo})
appearing in eq.(\ref{specialone}).
Alternatively, the set of differential equations:
\begin {eqnarray}
&&\nabla _i V = U_i\\
&& \nabla _i U_j = {\rm i} C_{ijk} g^{k \ell^\star} U_{\ell^\star}\\
&& \nabla _{i^\star} U_j = g_{{i^\star}j} V\\
&&\nabla _{i^\star} V =0 \label{defaltern}
\end{eqnarray}
with V satisfying eq.s (\ref{covholsec}, \ref {specpotuno}) give yet
another definition of special geometry.
In particular it is easy to find eq.(\ref{specialone})
as integrability conditions of(\ref{defaltern})
\subsection{\sc The vector kinetic matrix $\mathcal{N}_{\Lambda\Sigma}$ in special geometry}
\label{scrittaN}
In the construction of supergravity actions another essential item is the complex symmetric matrix $\mathcal{N}_{\Lambda\Sigma}$ whose real and imaginary parts are necessary in order to write the kinetic terms
of the vector fields. The matrix $\mathcal{N}_{\Lambda\Sigma}$ constitutes an integral part of the Special Geometry set up and we provide its general definition in the following lines.
Explicitly $\mathcal{N}_{\Lambda\Sigma}$ which, in relation to its interpretation in the case of Calabi-Yau threefolds, is named
the {\it period matrix}, is defined by means of the following relations:
\begin{equation}
{\bar M}_\Lambda = {{\cal N}}_{\Lambda\Sigma}{\bar L}^\Sigma \quad ;
\quad
h_{\Sigma\vert i} = { {\cal N}}_{\Lambda\Sigma} f^\Sigma_i
\label{etamedia}
\end{equation}
which can be solved introducing the two $(n+1)\times (n+1)$ vectors
\begin{equation}
f^\Lambda_I = \twovec{f^\Lambda_i}{{\bar L}^\Lambda} \quad ; \quad
h_{\Lambda \vert I} = \twovec{h_{\Lambda \vert i}}{{\bar M}_\Lambda}
\label{nuovivec}
\end{equation}
and setting:
\begin{equation}
{{\cal N}}_{\Lambda\Sigma}= h_{\Lambda \vert I} \circ \left (
f^{-1} \right )^I_{\phantom{I} \Sigma}
\label{intriscripen}
\end{equation}
\par
Let us now consider the case where the Special K\"ahler manifold $\mathcal{SK}_n$ of complex dimension $n$ has some isometry group
$\mathrm{U}_{\mathcal{SK}}$. Compatibility with the Special Geometry structure requires the existence of a $2n+2$-dimensional symplectic representation of such a group that we name the $\mathbf{W}$ representation.
In other words that there necessarily exists a symplectic embedding of the isometry group $\mathcal{SK}_n$
\begin{equation}
\mathrm{U}_{\mathcal{SK}} \mapsto \mathrm{Sp(2n+2, \mathbb{R})}
\label{sympembed}
\end{equation}
such that for each element $\xi \in \mathrm{U}_{\mathcal{SK}}$ we have its
representation by means of a suitable real symplectic matrix:
\begin{equation}
\xi \mapsto \Lambda_\xi \equiv \left( \begin{array}{cc}
A_\xi & B_\xi \\
C_\xi & D_\xi \
\end{array} \right)
\label{embeddusmatra}
\end{equation}
satisfying the defining relation (in terms of the symplectic antisymmetric metric $\mathbb{C}$):
\begin{equation}
\Lambda_\xi ^T \, \underbrace{\left( \begin{array}{cc}
\mathbf{0}_{n \times n} & { \mathbf{1}}_{n \times n} \\
-{ \mathbf{1}}_{n \times n} & \mathbf{0}_{n \times n} \
\end{array} \right)}_{ \equiv \, \mathbb{C}} \, \Lambda_\xi = \underbrace{\left( \begin{array}{cc}
\mathbf{0}_{n \times n} & { \mathbf{1}}_{n \times n} \\
-{ \mathbf{1}}_{n \times n} & \mathbf{0}_{n \times n} \
\end{array} \right)}_{\mathbb{C}}
\label{definingsympe}
\end{equation}
which implies the following relations on the $n \times n$ blocks:
\begin{eqnarray}
A^T_\xi \, C_\xi - C^T_\xi \, A_\xi & = & 0 \nonumber\\
A^T_\xi \, D_\xi - C^T_\xi \, B_\xi& = & \mathbf{1}\nonumber\\
B^T_\xi \, C_\xi - D^T_\xi \, A_\xi& = & - \mathbf{1}\nonumber\\
B^T_\xi \, D_\xi - D^T_\xi \, B_\xi & = & 0 \label{symplerele}
\end{eqnarray}
Under an element of the isometry group the symplectic section $\Omega$ of Special Geometry transforms
as follows:
\begin{equation}
\Omega\left( \xi \, \cdot \, z\right) \, = \, \Lambda_\xi \, \Omega\left ( z \right )
\end{equation}
As a consequence of its definition, under the same isometry the matrix ${\cal N}$ transforms by means of a generalized linear fractional
transformation:
\begin{equation}
\mathcal{N}\left(\xi \cdot z,\xi \cdot \bar{z}\right) = \left( C_\xi + D_\xi \, \mathcal{N}(z,\bar{z})\right) \, \left( A_\xi + B_\xi \,\mathcal{N}(z,\bar{z})\right) ^{-1}
\label{Ntransfa}
\end{equation}
\subsection{\sc The holomorphic moment map on K\"ahler manifolds}
The concept of holomorphic moment map applies to all K\"ahler manifolds, not necessarily special. Indeed it can be constructed just in terms of the K\"ahler potential without advocating any further structure. In this subsection we review its properties and definition, as usual in order to fix conventions, normalizations and notations.
\par
Let $g_{i {j^\star}}$ be the K\"ahler metric of a K\"ahler
manifold ${\cal M}$ and let us assume that $g_{i {j^\star}}$ admits
a non trivial group of continuous isometries ${\cal G}$
generated by Killing vectors $k_\mathbf{I}^i$ ($\mathbf{I}=1, \ldots, {\rm dim}
\,{\cal G} )$ that define the infinitesimal variation of the complex
coordinates $z^i$ under the group action:
\begin{equation}
\label{urca1}
z^i \to z^i + \epsilon^\mathbf{I} k_\mathbf{I}^i (z)
\end{equation}
Let $k^i_{\mathbf{I}} (z)$ be a basis of holomorphic Killing vectors for
the metric $g_{i{j^\star}}$. Holomorphicity means the following
differential constraint:
\begin{equation}
\partial_{j^*} k^i_{\mathbf{I}} (z)=0
\leftrightarrow \partial_j k^{i^*}_{\mathbf{I}} (\bar z)=0 \label{holly}
\end{equation}
while the generic Killing equation (suppressing the
gauge index $\mathbf{I}$):
\begin{equation}
\nabla_\mu k_\nu +\nabla_\mu k_\nu=0
\end{equation}
in holomorphic indices reads as follows:
\begin{equation}
\begin{array}{ccccccc}
\nabla_i k_{j} + \nabla_j k_{i} &=&0 & ; &
\nabla_{i^*} k_{j} + \nabla_j k_{i^*} &=& 0
\label{killo}
\end{array}
\end{equation}
where the covariant components are defined as
$k_{j }=g_{j i^*} k^{i^*}$ (and similarly for
$k_{i^*}$).
\par
The vectors $k_{\mathbf{I}}^i$ are generators of infinitesimal
holomorphic coordinate transformations $\delta z^i = \epsilon^\mathbf{I} k^i_{\mathbf{I}} (z)$
which leave the metric invariant. In the same way as the metric is
the derivative of a more fundamental
object, the Killing vectors in a K\"ahler manifold are the
derivatives of suitable prepotentials. Indeed the first of
eq.s (\ref{killo}) is automatically satisfied by holomorphic vectors
and the second equation reduces to the following one:
\begin{equation}
k^i_{\mathbf{I}}=i g^{i j^*} \partial_{j^*} {\cal P}_{\mathbf{I}},
\quad {\cal P}^*_{\mathbf{I}} = {\cal P}_{\mathbf{I}}\label{killo1}
\end{equation}
In other words if we can find a real function ${\cal P}^\mathbf{I}$ such
that the expression $i g^{i j^*} \partial_{j^*}
{\cal P}_{(\mathbf{I})}$ is holomorphic, then eq.(\ref{killo1}) defines a
Killing vector.
\par
The construction of the Killing prepotential can be stated in a more
precise geometrical fashion through the notion of {\it moment map}.
Let us review this construction.
\par
Consider a K\"ahlerian manifold ${\cal M}$ of real dimension $2n$.
Consider an isometry group ${\cal G}$ acting on
${\cal M}$ by means of Killing vector
fields $\overrightarrow{X}$ which are holomorphic
with respect to the complex structure
${ J}$ of ${\cal M}$; then these vector
fields preserve also the K\"ahler 2-form
\begin{equation}
\begin{array}{ccc}
\matrix{
{\cal L}_{\scriptscriptstyle\overrightarrow{X}}g = 0 & \leftrightarrow &
\nabla_{(\mu}X_{\nu)}=0 \cr
{\cal L}_{\scriptscriptstyle\overrightarrow{X}}{ J}= 0 &\null &\null \cr }
\Biggr \} & \Rightarrow & 0={\cal L}_{\scriptscriptstyle\overrightarrow{X}}
K = i_{\scriptscriptstyle\overrightarrow{X}}
dK+d(i_{\scriptscriptstyle\overrightarrow{X}}
K) = d(i_{\scriptscriptstyle\overrightarrow{X}}K) \cr
\end{array}
\label{holkillingvectors}
\end{equation}
Here ${\cal L}_{\scriptscriptstyle\overrightarrow{X}}$ and
$i_{\scriptscriptstyle\overrightarrow{X}}$
denote respectively the Lie derivative along
the vector field $\overrightarrow{X}$ and the contraction
(of forms) with it.
\par
If ${\cal M}$ is simply connected,
$d(i_{\overrightarrow{X}}K)=0$ implies the existence
of a function ${\cal P}_{\overrightarrow{X}}$ such
that
\begin{equation}
-\frac{1}{2}d{\cal P}_{\overrightarrow{X}}=
i_{\scriptscriptstyle\overrightarrow{X}}K
\label{mmap}
\end{equation}
The function ${\cal P}_{\overrightarrow{X}}$ is defined up to a constant,
which can be arranged so as to make it equivariant:
\begin{equation}
\overrightarrow{X} {\cal P}_{\overrightarrow{Y}} =
{\cal P}_{[\overrightarrow{X},\overrightarrow{Y}]}
\label{equivarianza}
\end{equation}
${\cal P}_{\overrightarrow{X}}$ constitutes then a {\it moment map}.
This can be regarded as a map
\begin{equation}
{\cal P}: {\cal M} \, \longrightarrow \,
\mathbb{R} \otimes
{\mathbb{G} }^*
\end{equation}
where ${\mathbb{G}}^*$ denotes the dual of the Lie algebra
${\mathbb{G} }$ of the group ${\cal G}$.
Indeed let $x\in {\mathbb{G} }$ be the Lie algebra element
corresponding to the Killing vector $\overrightarrow{X}$; then, for a given
$m\in {\cal M}$
\begin{equation}
\mu (m)\, : \, x \, \longrightarrow \, {\cal P}_{\overrightarrow{X}}(m) \,
\in \, \mathbb{R}
\end{equation}
is a linear functional on ${\mathbb{G}}$.
If we expand
$\overrightarrow{X} = a^\mathbf{I} k_\mathbf{I}$ in a basis of Killing vectors
$k_\mathbf{I}$ such that
\begin{equation}
[k_\mathbf{I}, k_\mathbf{L}]= f_{\mathbf{I} \mathbf{L}}^{\ \ \mathbf{K}} k_\mathbf{K}
\label{blio}
\end{equation}
we have also
\begin{equation}
{\cal P}_{\overrightarrow{X}}\, = \, a^\mathbf{I} {\cal P}_\mathbf{I}
\end{equation}
In the following we use the
shorthand notation ${\cal L}_\mathbf{I}, i_\mathbf{I}$ for the Lie derivative
and the contraction along the chosen basis of Killing vectors $ k_\mathbf{I}$.
\par
From a geometrical point of view the prepotential,
or moment map, ${\cal P}_\mathbf{I}$ is the Hamiltonian function providing the Poissonian
realization of the Lie algebra on the K\"ahler manifold. This
is just another way of stating the already mentioned
{\it equivariance}.
Indeed the very existence of the closed 2-form $K$ guarantees that
every K\"ahler space is a symplectic manifold and that we can define a
Poisson bracket.
\par
Consider eqs.(\ref{killo1}). To every generator of the abstract Lie algebra
${\mathbb{G}}$ we have associated a function ${\cal P}_\mathbf{I}$ on
${\cal M}$; the Poisson bracket of
${\cal P}_\mathbf{I}$ with ${\cal P}_\mathbf{J}$ is defined as follows:
\begin{equation}
\{{\cal P}_\mathbf{I} , {\cal P}_\mathbf{J}\} \equiv 4\pi K
(\mathbf{I}, \mathbf{J})
\end{equation}
where $K(\mathbf{I}, \mathbf{J})
\equiv K (\vec k_\mathbf{I}, \vec k_\mathbf{J})$ is
the value of $K$ along the pair of Killing vectors.
\par
In reference \cite{D'Auria:1991fj} the following
lemma was proved:
\begin{lemma}
{\it{The following identity is true}}:
\begin{equation}
\{{\cal P}_\mathbf{I}, {\cal
P}_\mathbf{J}\}=f_{\mathbf{I}\mathbf{J}}^{\ \ \mathbf{L}}{\cal
P}_\mathbf{L} + C_{\mathbf{I} \mathbf{J}} \label{brack}
\end{equation}
{\it{where $C_{\mathbf{I} \mathbf{J}}$ is a constant fulfilling the
cocycle condition}}
\begin{equation}
f^{\ \ \mathbf{L}}_{\mathbf{I}\mathbf{M}} C_{\mathbf{L} \mathbf{J}} +
f^{\ \ \mathbf{L}}_{\mathbf{M}\mathbf{J}} C_{\mathbf{L} \mathbf{I}}+
f_{\mathbf{J}\mathbf{I}}^{\ \ \mathbf{L}} C_{\mathbf{L} \mathbf{M}}=0
\label{cocy}
\end{equation}
\end{lemma}
If the Lie algebra ${\mathbb{G}}$ has a trivial second cohomology group
$H^2({\mathbb{G}})=0$, then the cocycle $C_{\mathbf{I} \mathbf{J}}$ is a
coboundary; namely we have
\begin{equation}
C_{\mathbf{I} \mathbf{J}} = f^{\ \ \mathbf{L}}_{\mathbf{I} \mathbf{J}} C_\mathbf{L}
\end{equation}
where $C_\mathbf{L}$ are suitable constants. Hence, assuming
$H^2 (\mathbb{G})= 0$
we can reabsorb $C_\mathbf{L}$ in the definition of ${\cal
P}_\mathbf{I}$:
\begin{equation}
{\cal P}_\mathbf{I} \rightarrow {\cal P}_\mathbf{I}+ C_\mathbf{I}
\end{equation}
and we obtain the stronger equation
\begin{equation}
\{{\cal P}_\mathbf{I}, {\cal P}_\mathbf{J}\} =
f_{\mathbf{I}\mathbf{J}}^{\ \ \mathbf{L}} {\cal P}_\mathbf{L}
\label{2.39}
\end{equation}
Note that $H^2({\mathbb{G}}) = 0$ is true for all semi-simple Lie
algebras.
Using eq.(\ref{brack}), eq.(\ref{2.39})
can be rewritten in components as follows:
\begin{equation}
{i\over 2} g_{ij^*}(k^i_\mathbf{I} k^{j^*}_\mathbf{J} -
k^i_\mathbf{J} k^{j^*}_\mathbf{I})=
{1\over 2} f_{\mathbf{I} \mathbf{J}}^{\ \ \mathbf{L}} {\cal
P}_\mathbf{L}
\label{2.40}
\end{equation}
Equation (\ref{2.40}) is identical with the equivariance condition
in eq.(\ref{equivarianza}).
\par
Finally let us recall the explicit general way of solving eq.(\ref{mmap}) obtaining the real valued function ${\cal P}_\mathbf{I}$ which satisfies eq.(\ref{killo1}). In terms of the K\"ahler potential $\mathcal{K}$ we have:
\begin{equation}\label{sisalvichipuo}
\mathcal{P}_{{\bf I}}{}^x=-\frac{i}{2}\left(k_{{\bf I}}^i\partial_i \mathcal{K}-k_{{\bf I}}^{\bar{\imath}}\partial_{\bar{\imath}} \mathcal{K}\right)+{\rm Im}(f_{{\bf I}})\,,
\end{equation}
where $f_{{\bf I}}=f_{{\bf I}}(z)$ is a holomorphic transformation on the line-bundle, defining a compensating K\"ahler transformation:
\begin{equation}
k_{{\bf I}}^i\partial_i \mathcal{K}+k_{{\bf I}}^{\bar{\imath}}\partial_{\bar{\imath}} \mathcal{K}=-f_{{\bf I}}(z)-\bar{f}_{{\bf I}}(\bar{z})\,.\label{sisalvichipuo2}
\end{equation}
We also have:
\begin{eqnarray}
\mathfrak{T}_{{\bf I}}\cdot \Omega &=&\mathfrak{T}_{{\bf I}}\cdot \Omega+f_{{\bf I}}\,\Omega\,,\label{sisalvichipuo30}\\
\mathfrak{T}_{{\bf I}}\cdot V+i\,{\rm Im}(f_{{\bf I}})\,V
&=& k_{{\bf I}}^i\partial_i V+k_{{\bf I}}^{\bar{\imath}}\partial_{\bar{\imath}} V\,,\label{sisalvichipuo3}
\end{eqnarray}
where $\mathfrak{T}_{{\bf I}}\cdot \Omega$ denotes the symplectic action of the isometry on the section $V$. If $\mathfrak{T}_{{\bf I}}$ is represented by the symplectic matrix $(\mathfrak{T}_{{\bf I}})_\alpha{}^\beta=-(\mathfrak{T}_{{\bf I}})^\beta{}_\alpha$, $\alpha,\,\beta=1,\dots,\,2n+2$:
\begin{equation}\label{cosasimplettica}
\mathfrak{T}_{\mathbf{I}}^T \, \mathbb{C} \, + \, \mathbb{C} \, \mathfrak{T}_{\mathbf{I}} \, = \, 0
\end{equation}
we have $(\mathfrak{T}_{{\bf I}}\cdot V)^\alpha=-\mathfrak{T}_{{\bf I}\,\beta}{}^\alpha\, V^\beta=\mathfrak{T}_{{\bf I}}^\alpha{}_\beta\, V^\beta$. From (\ref{sisalvichipuo3}) and (\ref{sisalvichipuo}) we derive the following useful symplectic-invariant expression for the moment maps:
\begin{equation}
\mathcal{P}_{{\bf I}}{}^x=-\bar{V}^\alpha\,\mathfrak{T}_{{\bf I}\,\alpha}{}^\beta\mathbb{C}_{\beta\gamma}\,V^\gamma\,.
\end{equation}
Eq.s (\ref{sisalvichipuo}), (\ref{sisalvichipuo2}), (\ref{sisalvichipuo3}) generalize the corresponding formulae given in sections 7.1 and 7.2 of \cite{Andrianopoli:1996cm}, where the condition $f_{{\bf I}}=0$ was imposed, to gaugings of non-compact isometries which are associated with non-trivial compensating K\"ahler transformations and/or to gauged (non-compact) isometries whose symplectic action is not diagonal.
\section{\sc Quaternionic geometry}
\label{hypgeosec}
Next we turn our attention to the geometry that pertains to the hypermultiplet sector of an
$\mathcal{N}=2$ supersymmetric theory. Each hypermultiplet contains $4$ real scalar fields
and, at least locally, they can be regarded as the
four components of a quaternion. The locality caveat is, in this
case, very substantial because global quaternionic coordinates can be
constructed only occasionally even on those manifolds that are
denominated quaternionic in the mathematical literature
\cite{hklr}, \cite{gal}. Anyhow, what is important is that, in
the hypermultiplet sector, the scalar manifold $\mathcal{QM}$ has
dimension multiple of four:
\begin{equation}
\mbox{dim}_{\bf R} \, \mathcal{QM} \, = \, 4 \, m \,\equiv \, 4 \, \# \,
\mbox{of hypermultiplets}
\label{quatdim}
\end{equation}
and, in some appropriate sense, it has a quaternionic structure.
\par
We name {\it Hypergeometry} that pertaining to the
hypermultiplet sector, irrespectively whether we deal with global or
local $\mathcal{N}$=2 theories. Yet there are two kinds of hypergeometries.
Supersymmetry requires the existence
of a principal $\mathrm{SU}(2)$--bundle
\begin{equation}
{\cal SU} \, \longrightarrow \, \mathcal{QM} \label{su2bundle}
\end{equation}
The bundle ${\cal SU}$ is
{\bf flat} in the {\it rigid supersymmetry case} while its curvature is
proportional to the K\"ahler forms in the {\it local case}.
\par
These two versions of hypergeometry were already known in mathematics prior to
their use \cite{D'Auria:1991fj},\cite{specHomgeo}, \cite{vanderseypen}, \cite{Andrianopoli:1996cm}, \cite{hklr}, \cite{gal} in the context of $\mathcal{N}=2$
supersymmetry and are identified as:
\begin{eqnarray}
\mbox{rigid hypergeometry} & \equiv & \mbox{HyperK\"ahler geometry.}
\nonumber\\ \mbox{local hypergeometry} & \equiv & \mbox{Quaternionic K\"ahler
geometry} \label{picchio}
\end{eqnarray}
\subsection{\sc Quaternionic K\"ahler, versus HyperK\"ahler manifolds}
Both a Quaternionic K\"ahler or a HyperK\"ahler manifold $\mathcal{QM}$
is a $4 m$-dimensional real manifold endowed with a metric $h$:
\begin{equation}
d s^2 = h_{u v} (q) d q^u \otimes d q^v \quad ; \quad u,v=1,\dots,
4 m \label{qmetrica}
\end{equation}
and three complex structures
\begin{equation}
(J^x) \,:~~ T(\mathcal{QM}) \, \longrightarrow \, T(\mathcal{QM}) \qquad
\quad (x=1,2,3)
\end{equation}
that satisfy the quaternionic algebra
\begin{equation}
J^x J^y = - \delta^{xy} \, \relax{\rm 1\kern-.35em 1} \, + \, \epsilon^{xyz} J^z
\label{quaternionetta}
\label{quatalgebra}
\end{equation}
and respect to which the metric is hermitian:
\begin{equation}
\forall \mbox{\bf X} ,\mbox{\bf Y} \in T\mathcal{QM} \,: \quad h
\left( J^x \mbox{\bf X}, J^x \mbox{\bf Y} \right ) = h \left(
\mbox{\bf X}, \mbox{\bf Y} \right ) \quad \quad
(x=1,2,3)
\label{hermit}
\end{equation}
From eq. (\ref{hermit}) it follows that one can introduce a triplet
of 2-forms
\begin{equation}
\begin{array}{ccccccc}
K^x& = &K^x_{u v} d q^u \wedge d q^v & ; & K^x_{uv} &=& h_{uw}
(J^x)^w_v \cr
\end{array}
\label{iperforme}
\end{equation}
that provides the generalization of the concept of K\"ahler form
occurring in the complex case. The triplet $K^x$ is named the {\it
HyperK\"ahler} form. It is an $\mathrm{SU}(2)$ Lie--algebra valued
2--form in the same way as the K\"ahler form is a $\mathrm{U(1)}$
Lie--algebra valued 2--form. In the complex case the definition of
K\"ahler manifold involves the statement that the K\"ahler 2--form is
closed. At the same time in Hodge--K\"ahler manifolds the K\"ahler 2--form can be
identified with the curvature of a line--bundle which in the case of
rigid supersymmetry is flat. Similar steps can be taken also here and
lead to two possibilities: either HyperK\"ahler or Quaternionic K\"ahler manifolds.
\par
Let us introduce a principal $\mathrm{SU}(2)$--bundle ${\cal SU}$ as
defined in eq. (\ref{su2bundle}). Let $\omega^x$ denote a connection
on such a bundle. To obtain either a HyperK\"ahler or a Quaternionic K\"ahler manifold we must impose the condition that the HyperK\"ahler 2--form
is covariantly closed with respect to the connection $\omega^x$:
\begin{equation}
\nabla K^x \equiv d K^x + \epsilon^{x y z} \omega^y \wedge K^z \,
= \, 0 \label{closkform}
\end{equation}
The only difference between the two kinds of geometries resides in
the structure of the ${\cal SU}$--bundle.
\begin{definizione} A
HyperK\"ahler manifold is a $4 m$--dimensional manifold with the
structure described above and such that the ${\cal SU}$--bundle is
{\bf flat}
\end{definizione}
Defining the ${\cal SU}$--curvature by:
\begin{equation}
\Omega^x \, \equiv \, d \omega^x + {1\over 2} \epsilon^{x y z}
\omega^y \wedge \omega^z \label{su2curv}
\end{equation}
in the HyperK\"ahler case we have:
\begin{equation}
\Omega^x \, = \, 0 \label{piattello}
\end{equation}
Viceversa \begin{definizione} A Quaternionic K\"ahler manifold is a $4
m$--dimensional manifold with the structure described above and such
that the curvature of the ${\cal SU}$--bundle is proportional to the
HyperK\"ahler 2--form \end{definizione} Hence, in the quaternionic
case we can write:
\begin{equation}
\Omega^x \, = \, { {\lambda}}\, K^x \label{piegatello}
\end{equation}
where $\lambda$ is a non vanishing real number.
\par
As a consequence of the above structure the manifold $\mathcal{QM}$ has
a holonomy group of the following type:
\begin{eqnarray}
{\rm Hol}(\mathcal{QM})&=& \mathrm{SU}(2)\otimes \mathrm{H} \quad
(\mbox{Quaternionic K\"ahler}) \nonumber \\ {\rm Hol}(\mathcal{QM})&=& \relax{\rm 1\kern-.35em 1}
\otimes \mathrm{H} \quad (\mbox{HyperK\"ahler}) \nonumber \\ \mathrm{H} &
\subset & \mathrm{Sp (2m,\mathbb{R}) }\label{olonomia}
\end{eqnarray}
In both cases, introducing flat indices $\{A,B,C= 1,2\}
\{\alpha,\beta,\gamma = 1,.., 2m\}$ that run, respectively, in the
fundamental representation of $\mathrm{SU}(2)$ and of
$\mathrm{Sp}(2m,\mathbb{R})$, we can find a vielbein 1-form
\begin{equation}
{\cal U}^{A\alpha} = {\cal U}^{A\alpha}_u (q) d q^u
\label{quatvielbein}
\end{equation}
such that
\begin{equation}
h_{uv} = {\cal U}^{A\alpha}_u {\cal U}^{B\beta}_v
\mathbb{C}_{\alpha\beta}\epsilon_{AB} \label{quatmet}
\end{equation}
where $\mathbb{C}_{\alpha \beta} = - \mathbb{C}_{\beta \alpha}$ and $\epsilon_{AB}
= - \epsilon_{BA}$ are, respectively, the flat $\mathrm{Sp}(2m)$ and
$\mathrm{Sp}(2) \sim \mathrm{SU}(2)$ invariant metrics. The vielbein
${\cal U}^{A\alpha}$ is covariantly closed with respect to the
$\mathrm{SU}(2)$-connection $\omega^z$ and to some
$\mathrm{Sp}(2m,\mathbb{R})$-Lie Algebra valued connection
$\Delta^{\alpha\beta} = \Delta^{\beta \alpha}$:
\begin{eqnarray}
\nabla {\cal U}^{A\alpha}& \equiv & d{\cal U}^{A\alpha} +{i\over 2}
\omega^x (\epsilon \sigma_x\epsilon^{-1})^A_{\phantom{A}B}
\wedge{\cal U}^{B\alpha} \nonumber\\ &+& \Delta^{\alpha\beta} \wedge
{\cal U}^{A\gamma} \mathbb{C}_{\beta\gamma} =0 \label{quattorsion}
\end{eqnarray}
\noindent where $(\sigma^x)_A^{\phantom{A}B}$ are the standard Pauli
matrices. Furthermore ${ \cal U}^{A\alpha}$ satisfies the reality
condition:
\begin{equation}
{\cal U}_{A\alpha} \equiv ({\cal U}^{A\alpha})^* = \epsilon_{AB}
\mathbb{C}_{\alpha\beta} {\cal U}^{B\beta} \label{quatreality}
\end{equation}
Eq.(\ref{quatreality}) defines the rule to lower the symplectic
indices by means of the flat symplectic metrics
$\epsilon_{AB}$ and $\mathbb{C}_{\alpha \beta}$. More specifically we can
write a stronger version of eq. (\ref{quatmet}) \cite{sugkgeom_3}:
\begin{eqnarray}
({\cal U}^{A\alpha}_u {\cal U}^{B\beta}_v + {\cal U}^{A\alpha}_v
{\cal
U}^{B\beta}_u)\mathbb{C}_{\alpha\beta}&=& h_{uv} \epsilon^{AB}\nonumber\\
\label{piuforte}
\end{eqnarray}
\noindent We have also the inverse vielbein ${\cal U}^u_{A\alpha}$
defined by the equation
\begin{equation}
{\cal U}^u_{A\alpha} {\cal U}^{A\alpha}_v = \delta^u_v \label{2.64}
\end{equation}
Flattening a pair of indices of the Riemann tensor ${\cal
R}^{uv}_{\phantom{uv}{ts}}$ we obtain
\begin{equation}
{\cal R}^{uv}_{\phantom{uv}{ts}} {\cal U}^{\alpha A}_u {\cal
U}^{\beta B}_v = -\,{{\rm i}\over 2} \Omega^x_{ts} \epsilon^{AC}
(\sigma_x)_C^{\phantom {C}B} \mathbb{C}^{\alpha \beta}+
\mathbb{R}^{\alpha\beta}_{ts}\epsilon^{AB}
\label{2.65}
\end{equation}
\noindent where $\mathbb{R}^{\alpha\beta}_{ts}$ is the field strength
of the $\mathrm{Sp}(2m) $ connection:
\begin{equation}
d \Delta^{\alpha\beta} + \Delta^{\alpha \gamma} \wedge \Delta^{\delta
\beta} \mathbb{C}_{\gamma \delta} \equiv \mathbb{R}^{\alpha\beta} =
\mathbb{R}^{\alpha \beta}_{ts} dq^t \wedge dq^s \label{2.66}
\end{equation}
Eq. (\ref{2.65}) is the explicit statement that the Levi Civita
connection associated with the metric $h$ has a holonomy group
contained in $\mathrm{SU}(2) \otimes \mathrm{Sp}(2m)$. Consider now
eq.s (\ref{quatalgebra}), (\ref{iperforme}) and (\ref{piegatello}).
We easily deduce the following relation:
\begin{equation}
h^{st} K^x_{us} K^y_{tw} = - \delta^{xy} h_{uw} +
\epsilon^{xyz} K^z_{uw}
\label{universala}
\end{equation}
that holds true both in the HyperK\"ahler and in the quaternionic
case. In the latter case, using eq. (\ref{piegatello}), eq.
(\ref{universala}) can be rewritten as follows:
\begin{equation}
h^{st} \Omega^x_{us} \Omega^y_{tw} = - \lambda^2 \delta^{xy} h_{uw} +
\lambda \epsilon^{xyz} \Omega^z_{uw} \label{2.67}
\end{equation}
Eq.(\ref{2.67}) implies that the intrinsic components of the curvature
2-form $\Omega^x$ yield a representation of the quaternion algebra.
In the HyperK\"ahler case such a representation is provided only by
the HyperK\"ahler form. In the quaternionic case we can write:
\begin{equation}
\Omega^x_{A\alpha, B \beta} \equiv \Omega^x_{uv} {\cal U}^u_{A\alpha}
{\cal U}^v_{B\beta} = - i \lambda \mathbb{C}_{\alpha\beta}
(\sigma_x)_A^{\phantom {A}C}\epsilon _{CB} \label{2.68}
\end{equation}
\noindent Alternatively eq.(\ref{2.68}) can be rewritten in an
intrinsic form as
\begin{equation}
\Omega^x =\,-{\rm i}\, \lambda \mathbb{C}_{\alpha\beta} (\sigma
_x)_A^{\phantom {A}C}\epsilon _{CB} {\cal U}^{\alpha A} \wedge {\cal
U}^{\beta B} \label{2.69}
\end{equation}
\noindent whence we also get:
\begin{equation}
{i\over 2} \Omega^x (\sigma_x)_A^{\phantom{A}B} = \lambda{\cal
U}_{A\alpha} \wedge {\cal U}^{B\alpha} \label{2.70}
\end{equation}
\subsection{\sc The triholomorphic moment map on quaternionic manifolds}
Next, following closely the original derivation of \cite{D'Auria:1991fj,mylecture}
let us turn to a discussion of the triholomorphic isometries of the manifold $\mathcal{QM}$
associated with hypermultiplets.
In $D=4$ supergravity the manifold of hypermultiplet scalars $\mathcal{QM}$ is a Quaternionic
K\"ahler manifold and we can gauge only those of its isometries
that are triholomorphic and that either generate an abelian group $\mathcal{G}$
or are \emph{suitably realized} as isometries also on the special manifold $\widehat{\mathcal{SK}}_n$.
This means that on $\mathcal{QM}$ we have Killing vectors:
\begin{equation}
\vec k_\mathbf{I} = k^u_\mathbf{I} {\vec \partial\over \partial q^u}
\label{2.71}
\end{equation}
\noindent
satisfying the same Lie algebra as the corresponding Killing
vectors on $\widehat{\mathcal{SK}}_n$. In other words
\begin{equation}
{\vec{\mathfrak{K}}}_\mathbf{I} =
\hat{k}^i_\mathbf{I} \vec \partial_i + \hat{k}^{i^*}_\mathbf{I}
\vec\partial_{i^*} + k_\mathbf{I}^u \vec\partial_u
\label{2.72}
\end{equation}
\noindent
is a Killing vector of the block diagonal metric:
\begin{equation}
\mathfrak{g} = \left (
\matrix { \widehat{g}_{ij^\star} & \quad 0 \quad \cr \quad 0
\quad & h_{uv} \cr } \right )
\label{2.73}
\end{equation}
defined on the product manifold\footnote{Following the notations described in the introduction, the Special K\"ahler manifold which describes the interaction of vector multiplets is denoted $\widehat{\mathcal{SK} }$ and all the Special Geometry Structures are endowed with a hat in order to distinguish this Special K\"ahler manifold from the other one which is incapsulated into the Quaternionic K\"ahler manifold $\mathcal{QM}$ describing the hypermultiplets when this latter happens to be in the image of the $c$-map.} $\widehat{\mathcal{SK} }\otimes\mathcal{QM}$.
\par
Let us first focus on the manifold $\mathcal{QM}$.
Triholomorphicity means that the Killing vector fields leave
the HyperK\"ahler structure invariant up to $\mathrm{SU(2)}$
rotations in the $\mathrm{SU(2)}$--bundle defined by eq.(\ref{su2bundle}).
Namely:
\begin{equation}
\begin{array}{ccccccc}
{\cal L}_\mathbf{I} K^x & = &\epsilon^{xyz}K^y
W^z_\mathbf{I} & ; &
{\cal L}_\mathbf{I}\omega^x&=& \nabla W^x_\mathbf{I}
\end{array}
\label{cambicchio}
\end{equation}
where $W^x_\mathbf{I}$ is an $\mathrm{SU(2)}$ compensator associated with the
Killing vector $k^u_\mathbf{I}$. The compensator $W^x_\mathbf{I}$ necessarily
fulfills the cocycle condition:
\begin{equation}
{\cal L}_\mathbf{I} W^{x}_\mathbf{J} - {\cal L}_\mathbf{J} W^x_\mathbf{I} + \epsilon^{xyz}
W^y_\mathbf{I} W^z_\mathbf{J} = f_{\mathbf{I} \mathbf{J}}^{\cdot \cdot \mathbf{L}}
W^x_\mathbf{L}
\label{2.75}
\end{equation}
In the HyperK\"ahler case the $\mathrm{SU(2)}$--bundle is flat and the
compensator can be reabsorbed into the definition of the
HyperK\"ahler forms. In other words we can always find a
map
\begin{equation}
\mathcal{QM} \, \longrightarrow \, L^x_{\phantom{x}y} (q)
\, \in \, \mathrm{SO(3)}
\end{equation}
that trivializes the ${\cal SU}$--bundle globally. Redefining:
\begin{equation}
K^{x\prime} \, = \, L^x_{\phantom{x}y} (q) \, K^y
\label{enfantduparadis}
\end{equation}
the new HyperK\"ahler form obeys the stronger equation:
\begin{equation}
{\cal L}_\mathbf{I} K^{x\prime} \, = \, 0
\label{noncambio}
\end{equation}
On the other hand, in the quaternionic case, the non--triviality of the
${\cal SU}$--bundle forbids to eliminate the $W$--compensator
completely. Due to the identification between HyperK\"ahler
forms and $\mathrm{SU(2)}$ curvatures eq.(\ref{cambicchio}) is rewritten
as:
\begin{equation}
\begin{array}{ccccccc}
{\cal L}_\mathbf{I} \Omega^x& = &\epsilon^{xyz}\Omega^y
W^z_\mathbf{I} & ; &
{\cal L}_\mathbf{I}\omega^x&=& \nabla W^x_\mathbf{I}
\end{array}
\label{cambiacchio}
\end{equation}
In both cases, anyhow, and in full analogy with the case of
K\"ahler manifolds, to each Killing vector
we can associate a triplet ${\cal
P}^x_\mathbf{I} (q)$ of 0-form prepotentials.
Indeed we can set:
\begin{equation}
{\bf i}_\mathbf{I} K^x =
- \nabla {\cal P}^x_\mathbf{I} \equiv -(d {\cal
P}^x_\mathbf{I} + \epsilon^{xyz} \omega^y {\cal P}^z_\mathbf{I})
\label{2.76}
\end{equation}
where $\nabla$ denotes the $\mathrm{SU(2)}$ covariant exterior derivative.
\par
As in the K\"ahler case eq.(\ref{2.76}) defines a moment map:
\begin{equation}
{\cal P}: {\cal M} \, \longrightarrow \,
\mathbb{R}^3 \otimes
{\mathcal{G} }^*
\end{equation}
where ${\mathcal{G}}^*$ denotes the dual of the Lie algebra
${\mathcal{G} }$ of the group ${\cal G}$.
Indeed let $x\in {\mathcal{G} }$ be the Lie algebra element
corresponding to the Killing
vector $\overrightarrow{X}$; then, for a given
$m\in {\cal M}$
\begin{equation}
\mu (m)\, : \, x \, \longrightarrow \, {\cal P}_{\overrightarrow{X}}(m) \,
\in \, \mathbb{R}^3
\end{equation}
is a linear functional on ${\mathcal{G}}$. If we expand
$\overrightarrow{X} = a^\mathbf{I} k_\mathbf{I}$ on a basis of Killing vectors
$k_\mathbf{I}$ such that
\begin{equation}
[k_\mathbf{I}, k_\mathbf{L}]= f_{\mathbf{I} \mathbf{L}}^{\ \ \mathbf{K}} k_\mathbf{K}
\label{blioprime}
\end{equation}
and we also choose a basis ${\bf i}_x \, (x=1,2,3)$ for $\mathbb{R}^3$
we get:
\begin{equation}
{\cal P}_{\overrightarrow{X}}\, = \, a^\mathbf{I} {\cal P}_\mathbf{I}^x \, {\bf i}_x
\end{equation}
Furthermore we need a generalization of the equivariance defined
by eq.(\ref{equivarianza})
\begin{equation}
\overrightarrow{X} \circ {\cal P}_{\overrightarrow{Y}} \,= \,
{\cal P}_{[\overrightarrow{X},\overrightarrow{Y}]}
\label{equivarianzina}
\end{equation}
In the HyperK\"ahler case, the left--hand side of
eq.(\ref{equivarianzina})
is defined as the usual action of a vector field on a $0$--form:
\begin{equation}
\overrightarrow{X} \circ {\cal P}_{\overrightarrow{Y}}\, = \, {\bf i}_{\overrightarrow{X}} \, d
{\cal P}_{\overrightarrow{Y}}\, = \,
X^u \, {\frac{\partial}{\partial q^u}} \, {\cal P}_{\overrightarrow{Y}}\,
\end{equation}
The equivariance condition implies
that we can introduce a triholomorphic Poisson bracket defined
as follows:
\begin{equation}
\{{\cal P}_\mathbf{I}, {\cal P}_\mathbf{J}\}^x \equiv 2 K^x (\mathbf{I},
\mathbf{J})
\label{hykapesce}
\end{equation}
leading to the triholomorphic Poissonian realization of the Lie
algebra:
\begin{equation}
\left \{ {\cal P}_\mathbf{I}, {\cal P}_\mathbf{J} \right \}^x \, = \,
f^{\mathbf{K}}_{\phantom{\mathbf{K}}\mathbf{I}\mathbf{J}} \, {\cal P}_\mathbf{K}^{x}
\label{hykapescespada}
\end{equation}
which in components reads:
\begin{equation}
K^x_{uv} \, k^u_\mathbf{I} \, k^v_\mathbf{J} \, = \, {\frac{1}{2}} \,
f^{\mathbf{K}}_{\phantom{\mathbf{K}}\mathbf{I}\mathbf{J}}\, {\cal P}_\mathbf{K}^{x}
\label{hykaide}
\end{equation}
In the quaternionic case, instead, the left--hand side of
eq.(\ref{equivarianzina})
is interpreted as follows:
\begin{equation}
\overrightarrow{X} \circ {\cal P}_{\overrightarrow{Y}}\, = \, {\bf i}_{\overrightarrow{X}}\, \nabla
{\cal P}_{\overrightarrow{Y}}\, = \,
X^u \, {\nabla_u} \, {\cal P}_{\overrightarrow{Y}}\,
\end{equation}
where $\nabla$ is the $\mathrm{SU(2)}$--covariant differential.
Correspondingly, the triholomorphic Poisson bracket is defined
as follows:
\begin{equation}
\{{\cal P}_\mathbf{I}, {\cal P}_\mathbf{J}\}^x \equiv 2 K^x (\mathbf{I},
\mathbf{J}) - { {\lambda}} \, \varepsilon^{xyz} \,
{\cal P}_\mathbf{I}^y \, {\cal P}_\mathbf{J}^z
\label{quatpesce}
\end{equation}
and leads to the Poissonian realization of the Lie algebra
\begin{equation}
\left \{ {\cal P}_\mathbf{I}, {\cal P}_\mathbf{J} \right \}^x \, = \,
f^{\mathbf{K}}_{\phantom{\mathbf{K}}\mathbf{I}\mathbf{J}} \, {\cal P}_\mathbf{K}^{x}
\label{quatpescespada}
\end{equation}
which in components reads:
\begin{equation}
K^x_{uv} \, k^u_\mathbf{I} \, k^v_\mathbf{J} \, - \,
{ \frac{\lambda}{2}} \, \varepsilon^{xyz} \,
{\cal P}_\mathbf{I}^y \, {\cal P}_\mathbf{J}^z\,= \, {\frac{1}{2}} \,
f^{\mathbf{K}}_{\phantom{\mathbf{K}}\mathbf{I}\mathbf{J}}\, {\cal P}_\mathbf{K}^{x}
\label{quatide}
\end{equation}
Eq.(\ref{quatide}), which is the most convenient way of
expressing equivariance in a coordinate basis was originally written
in \cite{D'Auria:1991fj} and has played a fundamental
role in the construction of supersymmetric actions for gauged $\mathcal{N}=2$ supergravity
both in $D=4$ \cite{D'Auria:1991fj,Andrianopoli:1996cm} and in $D=5$ \cite{ceregatta}.
\section{\sc The Quaternionic K\"ahler Geometry in the image of the $c$-map}
The main object of study in the present paper are those Quaternionic K\"ahler manifolds that are in the image of the $c$-map.\footnote{Not all non-compact, homogeneous Quaternionic K\"ahler manifolds which are relevant to supergravity (which are \emph{normal}, i.e. exhibiting a solvable group of isometries having a free and transitive action on it) are in the image of the c-map, the only exception being the quaternionic projective spaces \cite{alek,Cecotti:1988ad,vanderseypen}.} This latter
\begin{equation}\label{cimappo}
\mbox{c-map} \, \, : \,\, \mathcal{SK}_n \, \Longrightarrow \, \mathcal{QM}_{4n+4}
\end{equation}
is a universal construction that starting from an arbitrary Special K\"ahler manifold $\mathcal{SK}_n$ of complex dimension $n$, irrespectively whether it is homogenoeus or not, leads to a unique Quaternionic K\"ahler manifold $\mathcal{QM}_{4n+4}$ of real dimension $4n+4$ which contains $\mathcal{SK}_n$ as a submanifold. The precise modern definition of the $c$-map, originally introduced in \cite{sabarwhal}, is provided below.
\begin{definizione}
Let $\mathcal{SK}_n$ be a special K\"ahler manifold whose complex coordinates we denote by $z^i$ and whose K\"ahler metric we denote by $g_{ij^\star}$. Let moreover $\mathcal{N}_{\Lambda\Sigma}(z,{\bar z})$ be the symmetric period matrix defined by eq.(\ref{intriscripen}), introduce the following set of $4n+4$ coordinates:
\begin{equation}\label{finnico}
\left\{q^u \right\} \, \equiv \, \underbrace{\{U,a\}}_{\mbox{2 real}}\, \bigcup \,\underbrace{\underbrace{\{ z^i\}}_{\mbox{n complex}}}_{\mbox{2n real}} \, \bigcup\, \underbrace{\mathbf{Z} \, = \, \{ Z^\Lambda \, , \, Z_\Sigma \}}_{\mbox{(2n+2) real}}
\end{equation}
Let us further introduce the following $(\mathrm{2n+2})\times(\mathrm{2n+2}) $ matrix ${\cal M}_4^{-1}$:
\begin{eqnarray}
\mathcal{M}_4^{-1} & = &
\left(\begin{array}{c|c}
{\mathrm{Im}}\mathcal{N}\,
+\, {\mathrm{Re}}\mathcal{N} \, { \mathrm{Im}}\mathcal{N}^{-1}\, {\mathrm{Re}}\mathcal{N} & \, -{\mathrm{Re}}\mathcal{N}\,{ \mathrm{Im}}\,\mathcal{N}^{-1}\\
\hline
-\, { \mathrm{Im}}\mathcal{N}^{-1}\,{\mathrm{Re}}\mathcal{N} & { \mathrm{Im}}\mathcal{N}^{-1} \
\end{array}\right) \label{inversem4}
\end{eqnarray}
which depends only on the coordinate of the Special K\"ahler manifold.
The $c$-map image of $\mathcal{SK}_n$ is the unique Quaternionic K\"ahler manifold $\mathcal{QM}_{4n+4}$ whose coordinates are the $q^u$ defined in (\ref{finnico}) and whose metric is given by the following universal formula
\begin{eqnarray}
ds^2_{\mathcal{QM}} &=&\frac{1}{4} \left( d{U}^2+\, 4 g_{ij^\star} \,d{z}^j\, d{{\bar z}}^{j^\star}
+ e^{-2\,U}\,(d{a}+{\bf Z}^T\mathbb{C}d{{\bf
Z}})^2\,-\,2 \, e^{-U}\,d{{\bf
Z}}^T\,\mathcal{M}_4^{-1}\,d{{\bf Z}}\right)
\label{geodaction}
\end{eqnarray}
\end{definizione}
The metric (\ref{geodaction}) has the following positive definite signature
\begin{equation}
\mbox{sign}\left[ds^2_{\mathcal{QM}}\right] \, = \, \left(\underbrace{+,\dots,+}_{4+4\mathrm{n}}\right)
\end{equation}
since the matrix $\mathcal{M}_4^{-1} $ is negative definite.
\par
It is wort mentioning that if we utilize the same $4n+4$ coordinates (\ref{finnico}) and instead of (\ref{geodaction}) we introduce the alternative Lorentzian metric:
\begin{eqnarray}
ds^2_{\mathcal{QM^\star}} &=&\frac{1}{4}\left( d{U}^2+\, 4 g_{ij^\star} \,d{z}^j\, d{{\bar z}}^{j^\star}
+ e^{-2\,U}\,(d{a}+{\bf Z}^T\mathbb{C}d{{\bf
Z}})^2\,+\,2 \, e^{-U}\,d{{\bf
Z}}^T\,\mathcal{M}_4\,d{{\bf Z}}\right)
\label{BHaction}
\end{eqnarray}
that has signature:
\begin{equation}
\mbox{sign}\left[ds^2_{\mathcal{QM}^\star}\right] \, = \, \left(\underbrace{+,\dots,+}_{2\mathrm{n}+2},\underbrace{-,\dots ,-}_{2\mathrm{n}+2}\right)
\end{equation}
we obtain the pseudo-quaternionic manifold $\mathcal{QM}$ which constitutes the target manifold in the $3$-dimensional $\sigma$-model description of $D=4$ supergravity Black-Hole solutions \cite{PietroSashaMarioBH1},\cite{noinilpotenti},\cite{miosasha}. In the case the Special K\"ahler pre-image is a symmetric space $\mathrm{U}_{\mathcal{SK}}/\mathrm{H}_{\mathcal{SK}}$, both $\mathcal{QM}$ and $\mathcal{QM}^\star$ turn out to be symmetric spaces as well, $\mathrm{U}_{Q}/\mathrm{H}_{Q}$ and $\mathrm{U}_{Q}/\mathrm{H}_{Q}^\star$, the numerator group being the same. We will come back to the issue of symmetric homogeneous Quaternionic K\"ahler manifolds in section \ref{omosymmetro}
\subsection{\sc The HyperK\"ahler two-forms and the $\mathfrak{su}(2)$-connection}
The reason why we state that $\mathcal{QM}_{4n+4}$ is Quaternionic K\"ahler is that, by utilizing only the identities of Special K\"ahler Geometry we can construct the three complex structures $J_u^{x|v}$ satisfying the quaternionic algebra (\ref{quaternionetta}) the corresponding HyperK\"ahler two-forms $K^x$ and the $\mathfrak{su}(2)$ connection $\omega^x$ with respect to which they are covariantly constant.
\par
The construction is extremely beautiful and it is the following one.
\par
Consider the K\"ahler connection $\mathcal{Q}$ defined by eq. (\ref{u1conect}) and furthermore introduce the following differential form:
\begin{equation}
\label{Phidiffe}
\Phi \, = \, da + \mathbf{Z}^T \, \mathbb{C}\, \mathrm{d}\mathbf{Z}
\end{equation}
Next define the two dimensional representation of both the $\mathfrak{su}(2)$ connection and of the HyperK\"ahler $2$-forms as it follows:
\begin{eqnarray}
\omega &=& \frac{\rm i}{\sqrt{2}}\,\sum_{x=1}^3 \, \omega^x \, \gamma_x \label{cunnettasu2}\\
\mathbf{K} &=& \frac{\rm i}{\sqrt{2}}\,\sum_{x=1}^3 \, K^x \, \sigma_x \label{HypKalmatra}
\end{eqnarray}
where $\gamma_x$ denotes a basis of $2\times 2$ euclidian $\gamma$-matrices for which we utilize the following basis
which is convenient in the explicit calculations we perform in Part Four\footnote{The chosen $\gamma$-matrices are a permutation of the standard pauli matrices divided by $\sqrt{2}$ and multiplied by $\frac{\rm i}{2}$ can be used as a basis of anti-hermitian generators for the $\mathfrak{su}(2)$ algebra in the fundamental defining representation.}:
\begin{eqnarray}
\gamma_1&=& \left(
\begin{array}{ll}
\frac{1}{\sqrt{2}} & 0 \\
0 & -\frac{1}{\sqrt{2}}
\end{array}
\right) \nonumber\\
\gamma_2&=& \left(
\begin{array}{ll}
0 & -\frac{i}{\sqrt{2}} \\
\frac{i}{\sqrt{2}} & 0
\end{array}
\right)\nonumber\\
\gamma_3 &=& \left(
\begin{array}{ll}
0 & \frac{1}{\sqrt{2}} \\
\frac{1}{\sqrt{2}} & 0
\end{array}
\right) \label{gammini}
\end{eqnarray}
These $\gamma$-matrices satisfy the following Clifford algebra:
\begin{equation}\label{cliffordus}
\left\{ \gamma_x \, , \, \gamma_y \right \} \, = \, \delta^{xy} \, \mathbf{1}_{2 \times 2}
\end{equation}
and $\frac{\rm i}{2} \, \gamma_x$ provide a basis of generators of the $\mathfrak{su}(2)$ algebra.
\par
Having fixed these conventions the expression of the quaternionic $\mathfrak{su}(2)$-connection in terms of Special Geometry structures is encoded in the following expression for the $2\times 2$-matrix valued $1$-form $\omega$. Explicitly we have:
\begin{equation}\label{omegaSu2}
\omega \, = \, \left( \begin{array}{cc}
-\frac{\rm i}{2} \, \mathcal{Q} \, - \, \frac{\rm i}{4} \,e^{-U} \, \Phi & e^{-\frac{U}{2}} \, V^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\\
- \, e^{-\frac{U}{2}} \, \overline{V}^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z} & \frac{\rm i}{2} \, \mathcal{Q} \, + \, \frac{\rm i}{4} \,e^{-U} \, \Phi
\end{array}
\right)
\end{equation}
where $V$ and $\overline{V}$ denote the covariantly holomorphic sections of Special geometry defined in eq.s (\ref{covholsec}).
The curvature of this connection is obtained from a straight-forward calculation:
\begin{eqnarray}
\label{K2per2}
\mathbf{K} &\equiv& d\omega \, + \, \omega \, \wedge \, \omega \nonumber \\
\null &=& \left(\begin{array}{cc}
\mathfrak{u} & \mathfrak{v} \\
- \,\overline{\mathfrak{v}}& -\,\mathfrak{u}
\end{array}
\right)
\end{eqnarray}
the independent $2$-form matrix elements being given by the following explicit formulae:
\begin{eqnarray}
\label{uvvb}
\mathfrak{u} &=& -{\rm i} \frac{1}{2} \, K \, -\frac{1}{8} dS \,\wedge \, d\bar{S}\, - \, e^{-U} \, V^T \, \mathbb{C} \,\mathrm{d}\mathbf{Z}\, \wedge\, \bar{V}^T \, \mathbb{C} \,\mathrm{d}\mathbf{Z} \, - \, \frac{1}{4} \, e^{-U} \,\mathrm{d}\mathbf{Z}^T \, \wedge \, \mathbb{C} \, \mathrm{d}\mathbf{Z} \nonumber\\
\mathfrak{v }&=& e^{-\frac{U}{2}} \left( \, DV^T \, \wedge \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\, - \, \frac{1}{2} dS \, \wedge \,
V^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\right)\nonumber\\
\overline{\mathfrak{v }} &=& e^{-\frac{U}{2}} \left( \, D\overline{V}^T \, \wedge \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\, - \, \frac{1}{2} d\overline{S} \, \wedge \,
\overline{V}^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\right)
\end{eqnarray}
where
\begin{equation}\label{kalleforma}
K \, = \, \frac{ {\rm i}}{2} \, g_{ij^\star} \, dz^i \, \wedge \, d\bar{z}^{j^\star}
\end{equation}
is the K\"ahler $2$-form of the Special K\"ahler submanifold and where we have used the following short hand notations:
\begin{eqnarray}
dS &=& dU \, + \, {\rm i} \, e^{-U}\, \left(da \, + \, \mathbf{Z}^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\right) \label{firbone1} \\
d\overline{S} &=& dU \, - \, {\rm i} \, e^{-U}\, \left(da \, + \, \mathbf{Z}^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\right) \label{firbone2} \\
DV &=& dz^i \, \nabla_i V \label{firbone3} \\
D\overline{V} &=& d\bar{z}^{i^\star} \, \nabla_{i^\star} V \label{firbone4}
\end{eqnarray}
The three HyperK\"ahler forms $K^x$ are easily extracted from eq.s (\ref{K2per2}-\ref{uvvb}) by collecting the coefficients of the $\gamma$-matrix expansion and we need not to write their form which is immediately deduced. The relevant thing is that the components of $K^x$ with an index raised through multiplication with the inverse of the quaternionic metric $h^{uv}$ exactly satisfy the algebra of quaternionic complex structures (\ref{quatalgebra}). Explicitly we have:
\begin{eqnarray}
K^x &=& - \, {\rm i} \, 4 \sqrt{2} \, \mbox{Tr} \, \left( \gamma^x \, \mathbf{K}\right) \, \equiv \, K^x_{uv} \, dq^u \,\wedge \, dq^v \nonumber\\
J^{x|s}_u &=& K^x_{uv}\, h^{vs} \nonumber \\
J^{x|s}_u \, J^{y|v}_s &=& - \delta^{ xy} \, \delta^v_u \, + \, \epsilon^{xyz} \, J^{z|v}_u \label{quatKvera}
\end{eqnarray}
The above formulae are not only the general proof that the Riemaniann manifold $\mathcal{QM}$ defined by the metric (\ref{geodaction}) is indeed a Quaternionic K\"ahler manifold, but, what is most relevant, they also provide an algorithm to write in terms of Special Geometry structures the tri-holomorphic moment map of the principal isometries possessed by $\mathcal{QM}$.
\subsection{\sc Isometries of $\mathcal{QM}$ in the image of the $c$-map and their tri-holomorphic moment maps}
\label{triholoformul}
Let us now consider the isometries of the metric (\ref{geodaction}). There are three type of isometries:
\begin{description}
\item[a)] The isometries of the $(2n+3)$--dimensional Heisenberg algebra $\mathrm{\mathbb{H}eis}$ which is always present and is universal for any $(4n+4)$--dimensional Quaternionic K\"ahler manifold in the image of the $c$-map. We describe it below.
\item[b)] All the isometries of the pre-image Special K\"ahler manifold $\mathcal{SK}_n$ that are promoted to isometries of the image manifold in a way described below.
\item[c)] The additional $2n+4$ isometries that occur only when $\mathcal{SK}_n$ is a symmetric space and such, as a consequence, is also the $c$-map image $\mathcal{QM}_{4n+4}$. We will discuss these isometries
in section (\ref{omosymmetro}).
\end{description}
For the first two types of isometries a) and b) we are able to write general expressions for the tri-holomorphic moment maps that utilize only the structures of Special Geometry. In the case that the additional isometries c) do exist we have another universal formula which can be used for all generators of the isometry algebra $\mathbb{U}_\mathcal{Q}$ and which relies on the identification of the generators of the $\mathfrak{su}(2) \subset \mathbb{H}$ subalgebra with the three complex structures. We will illustrate the details of such an identification while discussing the example of the $S^3$-model.
\par
First of all let us fix the notation writing the general form of a Killing vector. This a tangent vector:
\begin{eqnarray}\label{killingus}
\vec{\mathbf{k}}& = & k^u (q) \, {\partial}_u \nonumber\\
&=& k^\diamond \, \frac{\partial}{\partial U} \, + \, k^i \, \frac{\partial}{\partial z^i}\, + \, k^{i^\star} \, \frac{\partial}{\partial \bar{z}^{i^\star}}\, + \, k^\bullet \, \frac{\partial}{\partial a}\, + \, k^\alpha \, \frac{\partial}{\partial \mathbf{Z}^\alpha}\nonumber\\
& \equiv & k^\diamond \, \partial_\diamond + \, k^i \, \partial_i \, + \, k^{i^\star} \,\partial_{i^\star} + \, k^\bullet \, \partial_\bullet \, + \, k^\alpha \, \partial_\alpha
\end{eqnarray}
with respect to which the Lie derivative of the metric element (\ref{geodaction}) vanishes:
\begin{equation}\label{isoverissima}
\ell_{\vec{\mathbf{k}}}\, ds^2_{\mathcal{QM}} \, = \,0
\end{equation}
\subsubsection{\sc Tri-holomorphic moment maps for the Heisenberg algebra translations}
First let us consider the isometries associated with the Heisenberg algebra. The transformation:
\begin{equation}\label{infinoLam}
Z^\alpha \, \mapsto \, Z^\alpha \, + \, \Lambda^\alpha \quad ; \quad a \, \mapsto \, a \, - \, \Lambda^T \, \mathbb{C} \, \mathbf{Z}
\end{equation}
where $\Lambda^\alpha$ is an arbitrary set of $2n+2$ real infinitesimal parameters is an infinitesimal isometry for the metric $ ds^2_{\mathcal{QM}}$ in (\ref{geodaction}). It corresponds to the following Killing vector:
\begin{eqnarray}\label{KillusW}
\overrightarrow{\mathbf{k}}_{[\Lambda]} & = & \Lambda^\alpha \, \overrightarrow{\mathbf{k}}_\alpha \nonumber\\
& = & \Lambda^\alpha \, \partial_\alpha \, - \, \Lambda^T \, \mathbb{C} \, \mathbf{Z} \, \partial_\bullet
\end{eqnarray}
whose components are immediately deduced by comparison of eq.(\ref{KillusW}) with eq.(\ref{killingus}).
\par
We are interested in determining the expression of the tri-holomorphic moment map $\mathfrak{P}_{[\Lambda]}$ which satisfies the defining equation:
\begin{eqnarray}
\mathbf{i}_{[\Lambda]}\, \mathbf{K} \, \equiv \,
\left(\begin{array}{cc}
\mathbf{i}_{[\Lambda]}\,\mathfrak{u} & \mathbf{i}_{[\Lambda]}\,\mathfrak{v} \\
- \,\mathbf{i}_{[\Lambda]}\,\overline{\mathfrak{v}}& -\,\mathbf{i}_{[\Lambda]}\,\mathfrak{u}
\end{array}
\right) &=& \mathrm{d} \mathfrak{P}_{[\Lambda]} \, + \, \left[ \omega \, , \, \mathfrak{P}_{[\Lambda]}\right ]
\label{pullusLam}
\end{eqnarray}
The general solution to this problem is
\begin{eqnarray}
\label{triholoHeis}
\mathfrak{P}_{[\Lambda]}&=&\left(\begin{array}{cc}
- \, \frac{\rm i}{4}\, e^{-U}\, \Lambda^T \, \mathbb{C} \, \mathbf{Z} &\frac{1}{2} \, e^{-\frac{U}{2}} \, \Lambda^T \, C \, V \\
- \,\frac{1}{2} \, e^{-\frac{U}{2}} \, \Lambda^T \, C \, \overline{V} & \, \frac{\rm i}{4}\, e^{-U}\, \Lambda^T \, \mathbb{C} \, \mathbf{Z}
\end{array}
\right) \nonumber\\
\end{eqnarray}
\subsubsection{\sc Tri-holomorphic moment map for the Heisenberg algebra central charge}
Consider next the isometry associated with the Heisenberg algebra central charge. The transformation:
\begin{equation}\label{infinoZeta}
a \, \mapsto \, a \, + \, \varepsilon
\end{equation}
where $\varepsilon$ is an arbitrary real small parameter is an infinitesimal isometry for the metric $ ds^2_{\mathcal{QM}}$ in (\ref{geodaction}). It corresponds to the following Killing vector:
\begin{eqnarray}\label{KillusZ}
\varepsilon \, \overrightarrow{\mathbf{k}}_{[\bullet]} & = & \varepsilon \, \partial_\bullet
\end{eqnarray}
whose components are immediately deduced by comparison of eq.(\ref{KillusZ}) with eq.(\ref{killingus}).
\par
We are interested in determining the expression of the tri-holomorphic moment map $\mathfrak{P}_{[\bullet]}$ which satisfies the defining equation analogous to eq.(\ref{pullusLam}):
\begin{eqnarray}
\mathbf{i}_{[\bullet]}\, \mathbf{K} &=& \mathrm{d} \mathfrak{P}_{[\bullet]} \, + \, \left[ \omega \, , \, \mathfrak{P}_{[\bullet]}\right ]
\label{pullusBull}
\end{eqnarray}
The solution of this problem is even simpler than in the previous case. Explicitly we obtain:
\begin{eqnarray}
\label{triholoHeisZ}
\mathfrak{P}_{[\bullet]}&=&\left(\begin{array}{cc}
- \, \frac{\rm i}{8}\, e^{-U} &0 \\
0 & \, \frac{\rm i}{8}\, e^{-U}
\end{array}
\right) \nonumber\\
\end{eqnarray}
The explicit expression of the moment maps and Killing vectors associated with the Heisenberg isometries was used in the gauging of abelian subalgebras of the Heisenberg algebra, which is relevant to the description of compactifications of
Type II superstring on a generalized Calabi-Yau manifold \cite{heisenberg}.
\subsubsection{\sc Tri-holomorphic moment map for the extension of $\mathcal{SK}_n$ holomorphic isometries}
Next we consider the question how to write the moment map associated with those isometries that where already present in the original
Special K\"ahler manifold $\mathcal{SK}_n$ which we $c$-mapped to a Quaternionic K\"ahler manifold.
\par
Suppose that ${\mathcal{SK}}_n$ has a certain number of holomorphic Killing vectors $k_{\mathbf{I}}^i(z)$ satisfying eq.s (\ref{killo},\ref{killo1},\ref{holkillingvectors}) necessarily closing some Lie algebra $\mathfrak{g}_{\mathcal{SK}}$ among themselves. Their holomorphic momentum-map is provided by eq.(\ref{sisalvichipuo}). Necessarily every isometry of a special K\"ahler manifold has a linear symplectic $(2n+2)$-dimensional realization on the holomorphic section $\Omega(z)$ up to an overall holomorphic factor. This means that for each holomorphic Killing vector we have (see Eq. (\ref{sisalvichipuo30})):
\begin{equation}\label{bellacosa}
k_{\mathbf{I}}^i(z)\, \partial_i \, \Omega(z) \, = \, \exp\left[f_{\mathbf{I}}(z)\right] \, \mathfrak{T}_{\mathbf{I}} \, \Omega(z) \,.
\end{equation}
where $f_{\mathbf{I}}(z)$ the holomorphic K\"ahler compensator.
Then it can be easily checked that the transformation:
\begin{equation}\label{kQupKil}
z^i \, \mapsto \, z^i \, + \, k^i_{\mathbf{I}}(z) \quad ; \quad \mathbf{Z} \, \mapsto \, \mathbf{Z} \, + \, \mathfrak{T}_{\mathbf{I}} \, \mathbf{Z}
\end{equation}
is an infinitesimal isometry of the metric (\ref{geodaction}) corresponding to the Killing vector:
\begin{equation}\label{promossoKil}
\vec{\mathbf{k}}_{\mathbf{I}} \, = \, k^i_{\mathbf{I}}(z) \, \partial_i \, + \, k^{i^\star}_{\mathbf{I}}(\bar{z}) \,\partial_{i^\star} \, + \, \left(\mathfrak{T}_{\mathbf{I}}\right)^\alpha_{\phantom{\alpha}\beta} \, \mathbf{Z}^\beta \, \partial_\alpha
\end{equation}
Also in this case we are interested in determining the expression of the tri-holomorphic moment map $\mathfrak{P}_{[\mathbf{I}]}$ satisfying the defining equation:
\begin{eqnarray}
\mathbf{i}_{\vec{\mathbf{k}}_{\mathbf{I}}}\, \mathbf{K} &=& \mathrm{d} \mathfrak{P}_{[\mathbf{I}]} \, + \, \left[ \omega \, , \,
\mathfrak{P}_{[\mathbf{I}]}\right ]
\label{pullusBull2}
\end{eqnarray}
The solution is given by the expression below:
\begin{eqnarray}
\label{triholoSK}
\mathfrak{P}_{[\mathbf{I}]}&=&\left(\begin{array}{cc}
\frac{\rm i}{4} \left(\mathcal{P}_{\mathbf{I}} \, + \, \frac{1}{2} \, e^{-U} \, \mathbf{Z}^T \, \mathbb{C} \, \mathfrak{T}_{\mathbf{I}} \, \mathbf{Z} \right)& - \, \frac{1}{2} \, e^{-U/2} \, V^T \, \mathbb{C} \, \mathfrak{T}_{\mathbf{I}} \, \mathbf{Z} \\
\, \frac{1}{2} \, \, e^{-U/2} \,\overline{V}^T \, \mathbb{C} \, \mathfrak{T}_{\mathbf{I}} \, \mathbf{Z} & \, - \frac{\rm i}{4} \left(\mathcal{P}_{\mathbf{I}} \, + \, \frac{1}{2} \, e^{-U} \, \mathbf{Z}^T \, \mathbb{C} \, \mathfrak{T}_{\mathbf{I}} \, \mathbf{Z} \right)
\end{array}
\right) \nonumber\\
\end{eqnarray}
where $\mathcal{P}_{\mathbf{I}}$ is the moment map of the same Killing vector in pure Special Geometry.
\subsection{\sc Homogeneous Symmetric Special Quaternionic K\"ahler manifolds}
\label{omosymmetro}
When the Special K\"ahler manifold $\mathcal{SK}_n$ is a symmetric coset space, it turns out that the metric (\ref{geodaction}) is actually the symmetric metric on an enlarged symmetric coset manifold
\begin{equation}\label{qcosetto}
\mathcal{QM}_{4n+4} \, = \, \frac{\mathrm{U}_Q}{\mathrm{H}_Q} \, \supset \, \frac{\mathrm{U}_{\mathcal{SK}}}{\mathrm{H}_{\mathcal{SK}}}
\end{equation}
\par
Naming $\Lambda[\mathfrak{g}]$ the $\mathbf{W}$-representation of any finite element of the $\mathfrak{g}\in\mathrm{U}_{\mathcal{SK}}$ group, we have that the matrix $\mathcal{M}_4(z,\bar{z})$ transforms as follows:
\begin{equation}\label{traduco}
\mathcal{M}_4\left( \mathfrak{g}\cdot z,\mathfrak{g}\cdot \bar{z} \right)\, = \, \Lambda[\mathfrak{g}] \, \mathcal{M}_4\left(z,\bar{z}\right ) ]\, \Lambda^T[\mathfrak{g}]
\end{equation}
where $\mathfrak{g}\cdot z$ denotes the non linear action of $\mathrm{U}_{\mathcal{SK}}$ on the scalar fields. Since the space $\frac{\mathrm{U}_{\mathcal{SK}}}{\mathrm{H}_{\mathcal{SK}}}$ is homogeneous, choosing any reference point $z_0$ all the others can be reached by a suitable group element $\mathfrak{g}_z$ such that $\mathfrak{g}_z\cdot z_0 \, = \, z$ and we can write:
\begin{equation}\label{turnaconto}
\mathcal{M}_4^{-1}(z,\bar{z}) \, = \, \Lambda^T[\mathfrak{g}_z^{-1}] \, \mathcal{M}_4^{-1}(z_0,\bar{z}_0) ]\, \Lambda[\mathfrak{g}^{-1}_z]
\end{equation}
This allows to introduce a set of $4n+4$ vielbein defined in the following way:
\begin{equation}\label{filibaine}
E^I_{\mathcal{QM}} \, = \, \frac{1}{2} \, \left\{ dU \, , \, \underbrace{e^i(z)}_{2\,n} \, , \, e^{-U} \,\left(d{a}+{\bf Z}^T\mathbb{C}d{{\bf
Z}}\right) \, , \, \underbrace{e^{-\frac {U} {2}}\, \Lambda[\mathfrak{g}_z^{-1}] \, \mathrm{d}\mathbf{Z}}_{2n+2} \right\}
\end{equation}
and rewrite the metric (\ref{geodaction}) as it follows:
\begin{equation}\label{cornish}
ds^2_{\mathcal{QM}} \, = \, E^I_{\mathcal{QM}} \, \mathfrak{q}_{IJ} \, E^J_{\mathcal{QM}}
\end{equation}
where the quadratic symmetric constant tensor $\mathfrak{q}_{IJ}$ has the following form:
\begin{equation}\label{quadrotta}
\mathfrak{q}_{IJ} \, = \, \left( \begin{array}{c|c|c|c}
1 & 0 & 0 & 0 \\
\hline
0 & \delta_{ij} & 0 & 0 \\
\hline
0 & 0 & 1 & 0 \\
\hline
0 & 0 & 0 & -\, 2\, \mathcal{M}_4^{-1}(z_0,\bar{z}_0)
\end{array}
\right)
\end{equation}
The above defined vielbein are endowed with a very special property namely they identically satisfy a set of Maurer Cartan equations:
\begin{equation}\label{MCSolv}
dE^I_{\mathcal{QM}} \, - \, \frac{1}{2} f^I_{\phantom{I}JK} \, E^J_{\mathcal{QM}} \, \wedge \, E^K_{\mathcal{QM}} \, = \, 0
\end{equation}
where $f^I_{\phantom{I}JK}$ are the structure constants of a solvable Lie algebra $\mathfrak{A}$ which can be identified as follows:
\begin{equation}\label{solvableGHalg}
\mathfrak{A} \, = \, Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}} \right)
\end{equation}
In the above equation $Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{\mathcal{Q}}} \right)$ denotes the Lie algebra of the solvable group manifold metrically equivalent to the non-comapact coset manifold $\frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}}$ according to a well developed mathematical theory extensively used in supergravity theories\cite{solvableparam}. In the case ${\mathrm{U}_{\mathcal{SK}}}$ is a \textit{maximally split} real form of a complex Lie algebra, then also ${\mathrm{U}_{\mathcal{Q}}}$ is maximally split and we have:
\begin{equation}\label{solvableGHalg2}
Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}} \right) \, =\, \mbox{Bor}\left ( \mathbb{U}_\mathcal{Q} \right)
\end{equation}
where $\mbox{Bor}\left ( \mathbb{U}_\mathcal{Q} \right)$ denotes the \textit{Borel subalgebra} of the semi-simple Lie algebra $\mathbb{G}$, generated by its Cartan generators and by the step operators associated with all positive roots.
\par
According to the general mathematical theory mentioned above, the very fact that the vielbein (\ref{filibaine}) satisfies the Maurer-Cartan equations of $Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}} \right)$ implies that the metric (\ref{cornish}) is the symmetric metric on the coset manifold $\frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}}$ which therefore admits continuous isometries associated with all the generators of the Lie algebra $\mathbb{U}_\mathcal{Q}$. This latter admits the following general decomposition:
\begin{equation}
\mbox{adj}(\mathbb{U}_{\mathcal{Q}}) =
\mbox{adj}(\mathbb{U}_{\mathcal{SK}})\oplus\mbox{adj}(\mathrm{SL(2,\mathbb{R})_E})\oplus
\mathbf{W}_{(2,\mathbf{W})}
\label{gendecompo}
\end{equation}
where $\mathbf{W}$ is the {\bf symplectic} representation of
$\mathbb{U}_{\mathcal{SK}}$ in which the symplectic section of Special Geometry transforms and which was used to construct the vielbein (\ref{cornish}). Denoting the generators of
$\mathbb{U}_{\mathcal{SK}}$ by $T^a$, the generators of
$\mathrm{SL(2,\mathbb{R})_E}$ by $\mathrm{L^x}$ and denoting by
$\mathbf{W}^{i\alpha}$ the generators in $\mathbf{W}_{(2,\mathbf{W})}$, the commutation
relations that correspond to the decomposition (\ref{gendecompo})
have the following general form \cite{MarioPietroKsenyaKM}:
\begin{eqnarray}
\nonumber && [T^a,T^b] = f^{ab}_{\phantom{ab}c} \, T^c \\
\nonumber && [L^x_E,L^y_E] = f^{xy}_{\phantom{xy}z} \, L^z , \\
&&\nonumber [T^a,\mathbf{W}^{i\alpha}] = (\Lambda^a)^\alpha_{\,\,\,\beta} \, \mathbf{W}^{i\beta},
\\ \nonumber && [L^x_E, \mathbf{W}^{i\alpha}] = (\lambda^x)^i_{\,\, j}\, \mathbf{W}^{j\alpha}, \\
&&[\mathbf{W}^{i\alpha},\mathbf{W}^{j\beta}] = \epsilon^{ij}\, (K_a)^{\alpha\beta}\, T^a + \,
\mathbb{C}^{\alpha\beta}\, k_x^{ij}\, L^x_E \label{genGD3pre}
\end{eqnarray}
where the $2 \times 2$ matrices $(\lambda^x)^i_j$, are the canonical generators of $\mathrm{SL(2,\mathbb{R})}$
in the fundamental, defining representation:
\begin{equation}
\lambda^3 = \left(\begin{array}{cc}
\ft 12 & 0 \\
0 & -\ft 12 \
\end{array} \right) \quad ; \quad \lambda^1 = \left(\begin{array}{cc}
0 & \ft 12 \\
\ft 12 & 0\
\end{array} \right) \quad ; \quad \lambda^2 = \left(\begin{array}{cc}
0 & \ft 12 \\
-\ft 12 & 0\
\end{array} \right)
\label{lambdax}
\end{equation}
while $\Lambda^a$ are the generators
of $\mathbb{U}_{\mathcal{SK}}$ in the symplectic representation $\mathbf{W}$. By
\begin{equation}
\mathbb{C}^{\alpha\beta} \equiv \left( \begin{array}{c|c}
\mathbf{0}_{(n+1)\times (n+1)} & \mathbf{1}_{(n+1)\times (n+1)} \\
\hline
-\mathbf{1}_{(n+1)\times (n+1)} & \mathbf{0}_{(n+1)\times (n+1)} \
\end{array}\right)
\label{omegamatra}
\end{equation}
we denote the antisymmetric symplectic metric in $2n+2$ dimensions, $n$
being the complex dimension of the Special K\"ahler manifold $\frac{\mathbb{U}_{\mathcal{SK}}}{\mathbb{H}_{\mathcal{SK}}}$. The symplectic character
of the representation $\mathbf{W}$ is asserted by the identity:
\begin{equation}
\Lambda^a\, \mathbb{C} + \mathbb{C}\, \left( \Lambda^a \right )^T = 0
\label{Lamsymp}
\end{equation}
The fundamental doublet representation of $\mathrm{SL(2,\mathbb{R})_E}$
is also symplectic and we have denoted by $\epsilon^{ij}= \left( \begin{array}{cc}
0 & 1 \\
-1 & 0
\end{array}\right) $ the
$2$-dimensional symplectic metric, so that:
\begin{equation}
\lambda^x\, \epsilon + \epsilon\, \left( \lambda^x \right )^T = 0,
\label{lamsymp}
\end{equation}
The matrices
$\left(K_a\right)^{\alpha\beta}=\left(K_a\right)^{\beta\alpha}$ and
$\left(k_x\right)^{ij}=\left(k_y\right)^{ji}$ are just symmetric matrices
in one-to-one correspondence with the generators of $\mathbb{U}_{\mathcal{Q}}$ and
$\mathrm{SL(2,\mathbb{R})}$, respectively. Implementing Jacobi
identities we find the following relations:
\begin{eqnarray}
&& \nonumber K_a\Lambda^c +
\Lambda^c K_a = f^{bc}_{\phantom{bc}a}K_b, \quad k_x\lambda^y + \lambda^y k_x
= f^{yz}_{\phantom{yz}x}k_z,
\label{jacobrele}
\end{eqnarray}
which admit the unique solution:
\begin{equation}
K_a = c_1 \, \mathbf{g}_{ab} \,\Lambda^b\mathbb{C}, \quad ; \quad k_x
= c_2 \, \mathbf{g}_{xy} \, \lambda^y \epsilon
\label{uniquesolutK&k}
\end{equation}
where $\mathbf{g}_{ab}$, $\mathbf{g}_{xy}$ are the Cartan-Killing metrics
on the algebras $\mathbb{U}_{\mathcal{SK}}$ and $\mathrm{SL(2,\mathbb{R})}$, respectively
and $c_1$ and $c_2$ are two arbitrary constants. These latter
can always be reabsorbed into the normalization of the generators
$\mathbf{W}^{i\alpha}$ and correspondingly set to one. Hence the algebra
(\ref{genGD3pre}) can always be put into the following elegant form:
\begin{eqnarray}
&& [T^a,T^b] = f^{ab}_{\phantom{ab}c} \, T^c \nonumber\\
&& [L^x,L^y] = f^{xy}_{\phantom{xy}z} \, L^z , \nonumber\\
&&
[T^a,\mathbf{W}^{i\alpha}] = (\Lambda^a)^\alpha_{\,\,\,\beta} \, \mathbf{W}^{i\beta},
\nonumber\\
&& [L^x, \mathbf{W}^{i\alpha}] = (\lambda^x)^i_{\,\, j}\, \mathbf{W}^{j\alpha}, \nonumber \\
&&[\mathbf{W}^{i\alpha},\mathbf{W}^{j\beta}] =
\epsilon^{ij}\, (\Lambda_a)^{\alpha\beta}\, T^a + \, \mathbb{C}^{\alpha\beta}\, \lambda_x^{ij}\, L^x
\label{genGD3}
\end{eqnarray}
where we have used the convention that symplectic indices are raised
and lowered with the symplectic metric, while adjoint representation
indices are raised and lowered with the Cartan-Killing metric.
\par
For the reader's convenience the list of Symmetric Special manifolds and of their Quaternionic K\"ahler counterparts in the image of the c-map is recalled in table \ref{homomodels} which reproduces the results of \cite{ToineCremmerOld}, according to which there is a short list of Symmetric Homogeneous Special manifolds comprising five discrete cases and two infinite series.
\begin{table}
\begin{center}
{\small
\begin{tabular}{||c|c||c||}
\hline
$\mathcal{SK}_n$ & $\mathcal{QM}_{4n+4}$ & $\mbox{dim} \, \mathcal{SK}_n \, = \, $ \\
Special K\"ahler manifold & Quaternionic K\"ahler manifold & $n$ \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}$ & $ \frac{\mathrm{G_{2(2)}}}{\mathrm{SU(2)\times SU(2)}}$ & $n=1$\\
\null & \null & \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{Sp(6,R)}}{\mathrm{SU(3)\times U(1)}}$ & $ \frac{\mathrm{F_{4(4)}}}{\mathrm{USp(6)\times SU(2)}}$ &$n=6$\\
\null & \null & \\
\null & \null &\null \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{SU(3,3)}}{\mathrm{SU(3)\times SU(3) \times U(1)}}$ & $ \frac{\mathrm{E_{6(2)}}}{\mathrm{SU(6)\times SU(2)}}$ &$n=9$\\
\null & \null & \\
\null & \null &\null \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{SO^\star(12)}}{\mathrm{SU(6)\times U(1)}}$ & $ \frac{\mathrm{E_{7(-5)}}}{\mathrm{SO(12)\times SU(2)}}$ & $n=15$ \\
\null & \null & \\
\null & \null &\null \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{E_{7(-25)}}}{\mathrm{E_{6(-78)} \times U(1)}}$ & $ \frac{\mathrm{E_{8(-24)}}}{\mathrm{E_{7(-133)}\times SU(2)}}$ & $n=27$ \\
\null & \null & \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,2+p)}}{\mathrm{SO(2)\times SO(2+p)}}$ & $ \frac{\mathrm{SO(4,4+p)}}{\mathrm{SO(4)\times SO(4+p)}}$ & $n=3+p$ \\
\null & \null & \\
\hline
\null & \null &\null \\
$ \frac{\mathrm{SU(p+1,1)}}{\mathrm{SU(p+1)\times U(1)}}$ & $ \frac{\mathrm{SU(p+2,2)}}{\mathrm{SU(p+2)\times SU(2)}}$ & $n=p+1$ \\
\null & \null &\null \\
\hline
\end{tabular}
}
\caption{List of special K\"ahler symmetric spaces with their Quaternionic K\"ahler c-map images. The number $n$ denotes the complex dimension of the Special K\"ahler preimage. On the other hand $4n+4$ is the real dimension of the Quaternionic K\"ahler c-map image. \label{homomodels}}
\end{center}
\end{table}
\par
Inspecting eq.s (\ref{genGD3}) we immediately realize that the Lie Algebra $\mathbb{U}_{Q}$ contains two universal Heisenberg subalgebras of dimension $(2n+3)$, namely:
\begin{eqnarray}
\mathbb{U}_{\mathcal{Q}} \, \supset \, \mathbb{H}\mathrm{eis}_1 \, &=& \mbox{span}_\mathbb{R}\, \left\{\mathbf{W}^{1\alpha} \, , \, \mathbb{Z}_1 \right\} \quad ; \quad \mathbb{Z}_1 \, =\, L_+ \, \equiv \, L^1\, + \, L^2 \nonumber\\
\null &\null & \left[\mathbf{W}^{1\alpha}\, , \, \mathbf{\mathbf{W}}^{1\beta} \right]\, = \, -\, \frac{1}{2} \, \mathbb{C}^{\alpha\beta} \, \mathbb{Z}_1 \quad ; \quad \left[\mathbb{Z}_1\, , \, \mathbf{W}^{1\beta} \right]\, = \,0 \label{Heinber1}\\
\mathbb{U}_{\mathcal{Q}} \, \supset \, \mathbb{H}\mathrm{eis}_2 \, &=& \mbox{span}_\mathbb{R}\, \left\{\mathbf{W}^{2\alpha} \, , \, \mathbb{Z}_2 \right\} \quad ; \quad \mathbb{Z}_2 \, =\, L_- \, \equiv \, L^1\, - \, L^2 \nonumber\\
\null &\null & \left[\mathbf{W}^{2\alpha}\, , \, \mathbf{W}^{2\beta} \right]\, = \, -\, \frac{1}{2} \, \mathbb{C}^{\alpha\beta} \, \mathbb{Z}_2 \quad ; \quad \left[\mathbb{Z}_2\, , \, \mathbf{W}^{2\beta} \right]\, = \,0 \label{Heinber2}
\end{eqnarray}
The first of these Heisenberg subalgebras of isometries is the universal one that exists for all Quaternionic K\"ahler manifolds $\mathcal{QM}_{4n+4}$ lying in the image of the $c$-map, irrespectively whether the pre-image Special K\"ahler manifold $\mathcal{SK}_n$ is a symmetric space or not. The tri-holomorphic moment map of these isometries was presented in eq.s(\ref{triholoHeis}) and (\ref{triholoHeisZ}). The second Heisenberg algebra exists only in the case when the Quaternionic K\"ahler manifold $\mathcal{QM}_{4n+4}$ is a symmetric space.
\par
From this discussion we also realize that the central charge $\mathbb{Z}_1$ is just the $L_+$ generator of a universal $\mathfrak{sl}(2,R)_E$ Lie algebra that exists only in the symmetric space case and which was named the Ehlers algebra in the context of dimensional reduction analysis from $D=4$ to $D=3$ \cite{PietroSashaMarioBH1}. When $\mathfrak{sl}(2,R)_E$ does exist we can introduce the universal compact generator:
\begin{equation}\label{ruotogrande}
\mathfrak{S}\, \equiv\, L_+ \, - \, L_- \,= \, 2 \, \lambda^2
\end{equation}
which rotates the two sets of Heisenberg translations one into the other:
\begin{equation}\label{tabarro}
\left[ \mathfrak{S}\, , \, \mathbf{W}^{i\alpha}\right ] \, = \, \epsilon^{ij} \, \mathbf{W}^{j\alpha}
\end{equation}
As we shall see, the gauging of this generator is a rather essential ingredient in the inclusion of one-field cosmological models into gauged $\mathcal{N}=2$ supergravity.
\paragraph{The embedding tensor formulation of the gauging.} It is useful to encode the choice of the gauge algebra in an \emph{embedding tensor} $\theta_{\Lambda}{}^{\mathcal{A}}$ \cite{Cordaro:1998tx,Nicolai:2001sv,deWit:2002vt,deWit:2005ub} though which the gauge generators are expressed in terms of the global symmetry ones. If we denote by $\{t_{\mathcal{A}}\}\equiv \{T^a,\,L^x,\,{\bf W}^{i\,\alpha}\}$ the generators of $\mathbb{U}_{\mathcal{Q}}$ and by $X_{\Lambda}$
the gauge generators, since we are gauging only (abelian) isometries of the quaternionic manifold, we can write:
\begin{equation}
X_{\Lambda}=\theta_{\Lambda}{}^{\mathcal{A}}\,t_{\mathcal{A}}\,,
\end{equation}
where the index $\Lambda$ runs over all the vector fields. Of these only a subset will actually gauge the chosen isometries. This subset will be labelled by boldface latin indices ${\bf I},\,{\bf J},\dots$.
The only condition on this tensor originates from the structure of the algebra:
\begin{equation}
[X_{\Lambda},\,X_{\Sigma}]=0\,\,\Rightarrow\,\,\,\,\,\theta_{\Lambda}{}^{\mathcal{A}}\theta_{\Lambda}{}^{\mathcal{B}}\,f_{\mathcal{A}\mathcal{B}}{}^\mathcal{C}=0\,,
\end{equation}
where $f_{\mathcal{A}\mathcal{B}}{}^\mathcal{C}$ are the structure constants of $\mathbb{U}_{\mathcal{Q}}$. The triholomorphic moment maps $\mathcal{P}^x_{\Lambda}$ and the Killing vectors $k_\Lambda^u$ can then be expressed in the following way:
\begin{equation}
\mathcal{P}^x_{\Lambda}=\theta_{\Lambda}{}^{\mathcal{A}}\,\mathcal{P}^x_{\mathcal{A}}\,\,;\,\,\,\,k_\Lambda^u=\theta_{\Lambda}{}^{\mathcal{A}}\,k_{\mathcal{A}}^u\,,
\end{equation}
where $\mathcal{P}^x_{\mathcal{A}},\,k_{\mathcal{A}}^u$ are the \emph{intrinsic} moment maps and Killing vectors associated with the quaternionic isometries.
In order to make the analysis independent of the initial symplectic frame of the vector multiplet sector, it is useful to describe the gauge algebra generators by the (redundant) symplectic notation $X_M=(X_\Lambda,\,X^\Lambda)=\theta_{M}{}^{\mathcal{A}}\,t_{\mathcal{A}}$, see \cite{deWit:2005ub}. The tensor $\theta_{M}{}^{\mathcal{A}}$ should then satisfy the locality constraint:
\begin{equation}
\theta_{\Lambda}{}^{\mathcal{A}}\theta^{\Lambda\,\mathcal{B}}-\theta_{\Lambda}{}^{\mathcal{B}}\theta^{\Lambda\,\mathcal{A}}=0\,,
\end{equation}
which guarantees that the tensor can be rotated, by means of a symplectic transformation, to an \emph{electric frame} in which $\theta^{\Lambda\,\mathcal{A}}=0$. For the restricted kind of gauging that we shall be dealing with, by extending the arguments given in \cite{thesearch}, we can work in the electric frame to start with, with no loss of generality.
\subsubsection{\sc The tri-holomorphic moment map in homogeneous symmetric Quaternionic K\"ahler manifolds}
\label{MezhduAlgGeom}
In the case the Quaternionic K\"ahler manifold $\mathcal{QM}_{4n+4}$ is a homogeneous symmetric space $\frac{\mathrm{U}_\mathcal{Q}}{\mathrm{H}_\mathcal{Q}}$, the tri-holomorphic moment map associated with any generator of $\mathfrak{t} \, \in \, \mathbb{U}_\mathcal{Q}$ of the isometry Lie algebra can be easily constructed by means of the formula:
\begin{equation}\label{generMapformula}
\mathcal{P}_\mathfrak{t}^x \, = \, \mbox{Tr}_{[{\mathbf{fun}}]}\, \left( J^x \, \mathbb{L}_{Solv}^{-1} \, \mathfrak{t} \, \mathbb{L}_{Solv} \right)
\end{equation}
where:
\begin{description}
\item[a)] $J^x$ are the three generators of the $\mathfrak{su}(2)$ factor in the isotropy subalgebra $\mathbb{H}\, = \, \mathfrak{su}(2) \, \oplus \, \mathbb{H}^\prime$, satisfying the quaternionic algebra (\ref{quatKvera}). They should be normalized in such a way as to realize the following condition. Naming:
\begin{equation}\label{MaurAmiCartan}
\Xi \, = \, \mathbb{L}_{Solv}^{-1}(q) \, \mathrm{d} \mathbb{L}_{Solv}(q)
\end{equation}
the Maurer Cartan differential one-form its projection on $J^x$ should precisely yield the $\mathfrak{su}(2)$ one-form defined in eq. (\ref{omegaSu2}):
\begin{equation}\label{omegacorrusco}
\omega \, = \, - \, \frac{{\rm i}}{ \sqrt{2} N_f } \, \sum_{x=1}^3 \,\mbox{Tr}_{[\mathbf{fun}]}\, \left( J^x \, \Xi\right) \, \gamma_x\, = \, \left( \begin{array}{cc}
-\frac{\rm i}{2} \, \mathcal{Q} \, - \, \frac{\rm i}{4} \,e^{-U} \, \Phi & e^{-\frac{U}{2}} \, V^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z}\\
- \, e^{-\frac{U}{2}} \, \overline{V}^T \, \mathbb{C} \, \mathrm{d}\mathbf{Z} & \frac{\rm i}{2} \, \mathcal{Q} \, + \, \frac{\rm i}{4} \,e^{-U} \, \Phi
\end{array}
\right)
\end{equation}
In the above equation, which provides the precise link between the $c$-map description and the coset manifold description of the same geometry, $N_f \, = \, \mbox{dim}\,\mathbf{fun}$ denotes the dimension of the fundamental representation of $\mathbb{U}_\mathcal{Q}$.
\item[b)] The solvable coset representative $\mathbb{L}_{Solv}(q)$ is obtained by exponentiation of the Solvable Lie algebra:
\begin{equation}\label{solvoexpo}
\mathbb{L}_{Solv}(q) \, \simeq \, \exp \left[ q \, \cdot \, Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathcal{H}_\mathcal{Q}}\right) \right]
\end{equation}
but the detailed exponentiation rule has to be determined in such a way that projecting the same Maurer Cartan form (\ref{MaurAmiCartan}) along an appropriate basis of generators $T_{I|Solv}$ of the solvable Lie algebra $Solv\left( \frac{\mathrm{U}_\mathcal{Q}}{\mathcal{H}_\mathcal{Q}}\right)$ we precisely obtain the vielbein $E^I_{QM}$ defined in eq.(\ref{filibaine}). This summarized in the following general equations:
\begin{eqnarray}\label{sopore}
E^I_{\mathcal{QM}} & = & \mbox{Tr}_{\mathbf{[fun]}} \left( T^I_{Solv} \, \Xi \right) \nonumber\\
\delta^{I}_{J} & = & \mbox{Tr}_{\mathbf{[fun]}} \left( T^I_{Solv} \, T_{I|Solv} \right) \nonumber\\
\Xi & = & E^I_{\mathcal{QM}} \, T_{I|Solv}
\end{eqnarray}
\end{description}
In eq.(\ref{sopore}) by $T^I_{Solv}$ we have denoted the conjugate (with respect to the trace) of the solvable Lie algebra generators.
\par
A general comment is in order. The precise calibration of the basis of the solvable generators $T^I_{Solv}$ and of their exponentiation outlined in eq.(\ref{solvoexpo}) which allows the identification (\ref{sopore}) is a necessary and quite laborious task in order to establish the bridge between the general $c$-map description of the quaternionic geometry and its actual realization in each symmetric coset model. This is also an unavoidable step in order to give a precise meaning to the very handy formula (\ref{generMapformula}) for the tri-holomorphic map. It should also be noted that although (\ref{generMapformula}) covers all the cases, the result of such a purely algebraic calculation is difficult to be guessed a priori. Hence educated guesses on the choice of generators whose gauging produces a priori determined features are difficult to be inferred from (\ref{generMapformula}). The analytic structure of the tri-holomorphic moment map instead is much clearer in the $c$-map framework of formulae (\ref{triholoHeis},\ref{triholoHeisZ},\ref{triholoSK}). The use of both languages and the construction of the precise bridge between them in each model is therefore an essential ingredient to understand the nature and the properties of candidate gaugings in whatever physical application.
\newpage
\part{\sc Abelian Gaugings and General Properties of their Potentials in the $c$-map Framework}
\label{grandiscussia}
As we stressed in the introduction the inclusion into $\mathcal{N}=2$ supergravity obtained in \cite{thesearch} of inflaton potentials such as the Starobinsky potential\footnote{Just as in \cite{thesearch} we mention scalar fields that typically have non canonical kinetic terms}
\begin{equation}\label{starotto}
V_{Starobinsky}(\phi) \, \equiv \, \left(1 \, - \, \exp\left[-\phi\right]\right)^2
\end{equation}
is not occasional and limited to the case of hypermultiplets lying in $\frac{\mathrm{G_{(2,2)}}}{\mathrm{SU(2)} \times \mathrm{SU(2)}}$, rather it follows a general pattern that can be uncovered and relies on the properties of the $c$-map. In this way the mechanisms of the \cite{thesearch} can be generalized to
larger Quaternionic K\"ahler manifolds opening a quite interesting new playground for the search of inflaton potentials that can be classified and understood in their geometrical origin.
\par
Let us schematically summarize the main ingredients of the approach pioneered in \cite{thesearch} whose generalization we pursue in this paper:
\begin{description}
\item[A)] The inflaton field $\phi$ is assumed to belong to the hypermultiplet Quaternionic K\"ahler manifold $\mathcal{QM}$.
\item[B)] In analogy with the construction in \cite{thesearch}, we require the graviphoton not to be minimally coupled to any other field. This condition originally followed from the general argument that in the dual to the $R+R^2$ supergravity the central charge is gauged. This will amount to a constraint on the form of the embedding tensor $\theta$ defining the gauge algebra.
\item[C)] The inflaton potential is generated by the gauging of an abelian subalgebra $\mathcal{A} \subset \mbox{iso} \left[\mathcal{QM}\right]$ of the isometry algebra of the hypermultipet manifold.
\item[D)] Since $\mathcal{A}$ is abelian it is not required to have any action on the vector multiplet scalars $\omega^i$ which are inert. Actually it is quite desirable that the potential $V_{gauging}$ generated by the gauging allows to fix all the $\omega^i$ to their values at some reference point, say $\omega^i \, = \, 0$:
\begin{equation}\label{fixing}
\left. \frac{\partial}{\partial \omega^i} \, V_{gauging} \right|_{\omega^i \, = \, 0} \, = \, 0
\end{equation}
As shown in \cite{thesearch}, one can generically guarantee the fixing conditions (\ref{fixing}) if the Special K\"ahler Geometry of the vector multiplets is chosen to be that of the so named Minimal Coupling, defined below in eq.s (\ref{minicup1}-\ref{minicup3}).
\item[E)] With the above choice of the vector multiplet geometry, after fixing the scalars $\omega^i$ the effective potential reduces to a sum of squares of the tri-holomorphic moment maps $P_{\mathcal{A}}^x$ which still depend on the variables $\left\{Z,U,a,z^i,{\bar z}^{i^\star}\right\}$. In order to approach effective potentials recognizable also as $\mathcal{N}=1$ supergravity potentials one would like to be able to fix all the Heisenberg fields $\mathbf{Z}$ (and possibly also the other fields $U$ and $a$) to zero, remaining only with the complex fields $\left ( z^i,{\bar z}^{i^\star}\right)$ of the inner Special K\"ahler manifold. Looking at the general form (\ref{triholoHeis}) of the tri-holomorphic moment map for the Heisenberg algebra generators and (\ref{triholoSK}) for the tri-holomorphic moment map of the inner Special K\"ahler isometries we immediately realize that, gauging these isometries \emph{separately}, the condition:
\begin{equation}\label{secondfixing}
\left. \frac{\partial}{\partial \mathbf{Z}^\alpha} \sum_{\mathfrak{t} \,\in \, \mathcal{A}} \, \left(\mathcal{P}_{\mathfrak{t}}\right)^2 \right|_{\mathbf{Z}\, = \, 0} \, = \, 0
\end{equation}
is always satisfied. A gauge generator which is a combination of a translation in the Heisenberg algebra and a Special K\"ahler isometry, yields in general a scalar potential exhibiting linear terms in ${\bf Z}$, so that (\ref{secondfixing}) provides a non-trivial constraint.\par
The definition of the locus $\mathcal{L}$ involves setting to zero a certain number of fields $\phi^r$ belonging to $\mathcal{SK}_n$ so that we should also realize the consistency condition:
\begin{equation}\label{thirdfixing}
\left. \frac{\partial}{\partial \phi^r} \sum_{\mathfrak{t} \,\in \, \mathcal{A}} \, \left(\mathcal{P}_{\mathfrak{t}}\right)^2 \right|_{\begin{array}{ccc}\mathbf{Z}& = & 0\\
\phi^r & = & 0 \\
\end{array}} \, = \, 0\,.
\end{equation}
As mentioned earlier, the gauging yielding Starobinsky-like potentials need also involve the compact generator $\mathfrak{S}$. As we shall show in the following, if the gauged isometry is a combination of $\mathfrak{S}$ and an $\mathcal{SK}_n$ isometry, (\ref{secondfixing}) poses no constraint on the gauging.
\item[F)] A favorite, though not mandatory, choice corresponds to looking for abelian generators of $\mbox{iso}\left[\mathcal{SK}_n\right]$ such that the locus which satisfies conditions (\ref{thirdfixing}) is defined by setting to zero all the axions $p_r$, namely all the fields associated with nilpotent generators of the solvable Lie algebra of $\mathcal{SK}_n$. The inclusion of the Starobinsky potential in supergravity was obtained in \cite{thesearch} precisely in this way. In section \ref{generalonuovo} we show a generalization of the same mechanism in the case of a bigger manifold $\mathcal{QM}_{4n+4}$, obtaining what can be denominated a multi Starobinsky model.
\item[G)] \textbf{The $U$-problem}. If we use only the type of isometries yielding the tri-holomorphic moment maps (\ref{triholoHeis}), (\ref{triholoHeisZ}) and (\ref{triholoSK}) we face a serious problem with the fields $U$. It appears only through exponentials all of the same sign ($\exp[ - 2 \, U]$ or $\exp[ - \, U]$ in front of perfect squares. Hence the field $U$ cannot be stabilized unless all such squares are zero which means no residual potential. To overcome such a problem one should have moment maps with the opposite sign of $U$ in the exponential and this can happen only by introducing in the gauging either $L^E_-$ or generators $\mathbf{W}^{2,\alpha}$ this means that such generators should exist, namely the manifold $\mathcal{QM}_{4n+4}$ should be a symmetric space. In \cite{thesearch} the $U$-problem was solved by adding to a parabolic generator of a $\mathcal{SK}_n$-isometry the universal compact generator (\ref{ruotogrande}). As we have emphasized the Ehlers subalgebra exists in all symmetric spaces and so does the compact generator (\ref{ruotogrande}). This implies that the mechanism leading to the inclusion of the Starobinsky model found in \cite{thesearch} is actually rather universal and can be generalized in several ways.
\end{description}
The above discussion provides a framework for the search of other inflaton potentials.
\section{\sc Minimal Coupling Special Geometry}
\label{minicoup}
In this section we shortly describe the structure of the Minimal Coupling Special K\"ahler manifold $\mathcal{MSK}_{p+1}$, mostly in order to fix our conventions and to establish our notations. As announced in the introduction, this kind of Special Geometry is our favorite choice for the vector multiplet sector of the $\mathcal{N}=2$ lagrangian which allows us to construct an entire class of theories where the vector multiplet scalars can be stabilized and the effective potential of an abelian gauging is reduced only to the hypermultiplet sector. In view of such a use of $\mathcal{MSK}_{p+1}$, all items of its Special Geometry will be denoted with a hat, and its complex coordinates will be named $\omega_i$ rather than $z^i$. However it is clear that $\mathcal{MSK}_{p+1}$ might also be used as $c$-map preimage of a Quaternionic K\"ahler manifold describing hypermultiplets.
\par
As a manifold $\mathcal{MSK}_{p+1}$ is the following coset:
\begin{equation}\label{minicup1}
\mathcal{MSK}_{p+1} \, = \, \frac{\mathrm{SU(1,p+1)}}{\mathrm{U(1)} \times \mathrm{SU(p+1)}}
\end{equation}
In terms of the complex coordinates $\omega^i$ a convenient choice of the $(2 \,p \,+\, 4)$-dimensional holomorphic symplect section is the following one:
\begin{equation}\label{minicup2}
\widehat{\Omega} \, = \, \left( \begin{array}{c}
\widehat{X}^\Lambda \\
\hline
\widehat{F}_\Sigma
\end{array}
\right) \, = \,\left( \begin{array}{c}
1 \\
\omega^i\\
\hline
-{\rm i}\\
{\rm i}\, \omega^i
\end{array}
\right) \quad ; \quad (i\,= \, 1, \dots \, p+1 )
\end{equation}
which leads to the following K\"ahler potential:
\begin{equation}\label{minicup3}
\widehat{\mathcal{K}} \, = \, - \, \log \,\left[ - \, {\rm i} \widehat{\Omega} \, \widehat{\mathbb{C}} \, \widehat{{\overline{\Omega}}}\right ] \, = \, - \, \log \, \left[\,2 \,\left( 1 \, - \, \omega\, \cdot \,\bar{\omega}\right)\right]
\end{equation}
and to the following K\"ahler metric:
\begin{equation}\label{minicup4}
\widehat{g}_{ij^\star} \, = \, \partial_i\,\partial_{j^\star} \,\widehat{\mathcal{K}}\, = \, \frac{1}{\left(1-\omega\cdot\ \bar{\omega} \right)^2} \, \left( \delta^{ij}\, \left(1-\omega\cdot\ \bar{\omega} \right)\, + \, \bar{\omega}^i \, \omega^j\right)
\end{equation}
Defining the K\"ahler covariant derivatives of the covariantly holomorphic sections as in eq.s (\ref{uvector}) we obtain three results that are very important for the discussion of reduced scalar potentials in the present paper.
Firstly we get:
\begin{equation}\label{minicup4barra}
\nabla_i \, \widehat{U}_j \, \equiv \, \nabla_i \, \nabla_j \widehat{V} \, = \,0
\end{equation}
which compared with eq.(\ref{defaltern}) implies the vanishing of the three-index symmetric tensor $\widehat{C}_{ijk}$. This unique property of the special K\"ahler manifold $\mathcal{MSK}_{p+1}$ defined by eq.(\ref{minicup1}) is the reason why it has been named the Minimal Coupling Special Geometry, the interpretation of the tensor $C_{ijk}$ in phenomenological applications being that of Yukawa couplings of the gauginos. In ref. \cite{thesearch} it was shown that the vanishing of $\widehat{C}_{ijk}$ guarantees the consistency (see eq. (3.10) of the quoted reference) of the truncation of the classical supergravity theory to the hypermultiplet quaternionic scalars by fixing the vector multiplet scalars to the origin of their manifold:
\begin{equation}\label{golubchika}
\omega^i \, = \, 0
\end{equation}
Secondly we evaluate the the covariantly symplectic holomorphic section in the origin of the manifold and we obtain:
\begin{equation}\label{miniculpan1}
\left. \widehat{V} \right|_{\omega\, =\,0} \, = \,\frac{1}{\sqrt{2}}\,\left\{\begin{array}{c|c||c|c}
1 & 0 & -\,{\rm i} & 0
\end{array}
\right\}
\end{equation}
In the same point we have:
\begin{equation}\label{fantamini}
\left.\left(\widehat{g}^{ij^\star}\,\nabla_i \widehat{V}^\alpha \,\nabla_{j^\star} \widehat{\overline{V}}\right)\right|_{\omega\, =\,0} \, = \, \frac{1}{2} \, \left(\begin{array}{c|c||c|c}
0 & 0 & 0 & 0 \\
\hline
0 & \mathbf{1}_{(p+1)\times(p+1)} & 0 & {\rm i} \,\mathbf{1}_{(p+1)\times(p+1)} \\
\hline
\hline
0 & 0 & 0 & 0 \\
\hline
0 & - \,{\rm i} \,\mathbf{1}_{(p+1)\times(p+1)} & 0 & \mathbf{1}_{(p+1)\times(p+1)}
\end{array}
\right)
\end{equation}
\subsection{\sc Gauging abelian isometries of the hypermultiplets}
Relying on these results we see that if the hypermultiplet Quaternionic manifold $\mathcal{QM}_{4m}$ possesses a $p+1$-dimensional abelian Lie algebra of isometries, we can always gauge them by using, for the vector multiplets, the Special K\"ahler manifold
$\mathcal{MSK}_{p+1}$ introducing also the following embedding tensor:
\begin{equation}\label{gioisco}
\theta_{M}^{I} \, \equiv \, \left\{\theta_{\Lambda}^I\, , \, \theta^{\Sigma|I }\right\} \, = \, \left \{\theta_0{}^{I} \, = \,0 \, , \, \theta_J{}^{I} \, = \, \delta^I_J \, , \, \theta^{\Sigma|I }\, = \, 0\right\}\,.
\end{equation}
Notice that the choice of setting $\theta_0{}^{I} \, = \,0$ follows from the requirement $B)$ that the graviphoton should not be gauged. This indeed amounts to
requiring:
\begin{equation}
\left. \widehat{V} \right|_{\omega\, =\,0}^M\,\theta_M{}^{\mathcal{A}}=0\,\,\,\Rightarrow\,\,\,\,\,\,\theta_0{}^{I} \, = \,0\,.
\end{equation}
In such a theory the scalar potential has the following general form:
\begin{equation}\label{cornettoalgida}
\mathcal{V}_{scalar}(\omega, \bar{\omega},q) \, = \, 4 \, k_I^u k_J^v \, h_{uv} \, \widehat{V}^I \, \widehat{\overline{V}}{}^J
\, + \,
\left(\widehat{g}^{ij^\star}\,\nabla_i \widehat{V}^I \,\nabla_{j^\star} \widehat{\overline{V}}{}^J \, - \, 3 \,
\widehat{V}^I \,\widehat{\overline{V}}{}^J \right) \, \mathcal{P}^x_I \, \mathcal{P}^x_J
\end{equation}
setting $\omega^i \, = \, 0$ is a consistent truncation and the reduced potential takes the following universal general form
which is positive definite by construction:
\begin{equation}\label{gartolini}
\mathcal{V}_{scalar}(0, 0,q) \, = \, \sum_{I=1}^{p+1} \,\mathcal{P}^x_I(q) \, \mathcal{P}^x_I(q)
\end{equation}
In the next section \ref{starobin1} we reconsider the derivation of the Starobinsky potential obtained in \cite{thesearch} from a parabolic gauging as a master example that can be generalized to bigger manifolds.
\subsection{\sc The Starobinsky potential}
\label{starobin1}
Last year a great deal of activity was devoted to the inclusion of phenomenologically interesting inflaton potentials into $\mathcal{N}=1$ supergravity as we recalled in the introduction. A first wave of investigations considered the possible generation of potentials by means of suitably chosen superpotentials, subsequently, after an important new viewpoint was introduced in \cite{minimalsergioKLP} and was subsequently developed in
\cite{primosashapietro},\cite{piesashatwo},\cite{Ferrara:2013wka},\cite{Ferrara:2013kca},
\cite{pietrosergiosasha1},\cite{pietrosergiosasha2}, it became clear that positive definite inflaton potentials can be generated by the gauging of some isometry of the K\"ahler manifold of scalar multiplets. Such potentials have the form of squares of K\"ahler moment maps. In \cite{pietrosergiosasha1} this mechanism was applied to the case of constant curvature one-dimensional K\"ahler manifolds and it was shown that Starobinsky-like potentials \cite{Starobinsky:1980te} emerge from the moment map of a parabolic isometry in $\mathrm{SL(2,\mathbb{R})} \simeq \mathrm{SU(1,1)}$ with the addition of a Fayet Iliopoulos term. In particular the standard Starobinsky model that is dual to an $R+R^2$ supergravity emerges from gauging the parabolic shift isometry of an $\frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}$ manifold with K\"ahler potential $\mathcal{K}\, = \, - 3\, \log (z-{\bar z})$ which is precisely the Special K\"ahler manifold $S^3$. Let us now consider eq.(\ref{triholoSK}) and we can learn an important lesson. If in the $c$-map image of some $\mathcal{SK}$ Special K\"ahler manifold, for instance the $S^3$ model, we gauge, according to the scheme discussed in section \ref{grandiscussia}, some nilpotent Lie algebra element $\mathfrak{N}_+\, \in \, \mathbb{U}_{\mathcal{SK}} \, \subset \mathbb{U}_\mathcal{Q}$ identical with the parabolic shift generator that we would have gauged in $\mathcal{N}=1$ supergravity, (for instance the generator $L_+\, \in \, \mathfrak{sl}(2,\mathbb{R})$ in the case of the $S^3$ model), we obtain a moment map that contains precisely the $\mathcal{P}_\mathbf{I}$ of the $\mathcal{N}=1$ case, modified by $\mathbf{Z}$ dependent terms. In case the $\mathbf{Z}$ can be stabilized to zero the remaining effective potential is that of the corresponding $\mathcal{N}=1$ theory, apart from the Fayet Iliopoulos term. There are two remaining problems. The generation of a Fayet Iliopoulos term and the stabilization of the $U$ field. They are solved in one stroke by modifying the parabolic generator of the inner Special K\"ahler isometry with the addition of the universal Ehlers rotation (\ref{ruotogrande}).
\par
Let us see how this works.
\par
With reference to eq.s (\ref{generillini}) let us consider the following generator:
\begin{equation}\label{costianovo}
\mathfrak{p}\, = \, \mathfrak{N}_+ \, + \, \kappa \, \mathfrak{S}
\end{equation}
where $\mathfrak{N}_+$ is the previously mentioned nilpotent element of the Special K\"ahler subalgebra ($\mathfrak{N}_+^r \, = \, 0$, for some positive integer $r$) and $\kappa$ is a parameter. Let us then calculate the tri-holomorphic moment map $\mathcal{P}^x_\mathfrak{p}$ according to formula (\ref{fisterone}).
\par
Because of the linearity of the momentum map in Lie algebra elements we have:
\begin{eqnarray}
\mathfrak{P}_\mathfrak{p} &=& \mathfrak{P}_{\mathfrak{N}_+} \, + \, \mathfrak{P}_{\mathfrak{S}}\nonumber \\
\mathfrak{P}_{\mathfrak{N}_+} &=& \left( \begin{array}{c|c}
\frac{\rm i}{4} \, \mathcal{P}_{\mathfrak{N}_+} \, + \, \mathcal{O} \, \left( \mathbf{Z}^2\right)& \mathcal{O} \, \left( \mathbf{Z}\right) \\
\hline
\mathcal{O} \, \left( \mathbf{Z}\right) & \, - \, \frac{\rm i}{4} \, \mathcal{P}_{\mathfrak{N}_+}\, - \, \mathcal{O} \, \left( \mathbf{Z}^2\right)
\end{array}
\right) \nonumber\\
\mathfrak{P}_{\mathfrak{S}} &=& \left( \begin{array}{c|c}
\frac{\rm i}{8} \, e^{-U} \,\left(1 + a^2 \, + \, e^{2U}\right)\, + \, \mathcal{O} \, \left( \mathbf{Z}^2\right)& \mathcal{O} \, \left( \mathbf{Z}\right) \\
\hline
\mathcal{O} \, \left( \mathbf{Z}\right) & \, - \, \frac{\rm i}{8} \, e^{-U} \,\left(1 + a^2 \, + \, e^{2U}\right)\, - \, \mathcal{O} \, \left( \mathbf{Z}^2\right)
\end{array}
\right) \nonumber\\
\label{gomorratano}
\end{eqnarray}
where $\mathcal{P}_{\mathfrak{N}_+}$ is the K\"ahlerian moment map of the Killing vector associated with the generator $\mathfrak{N}_+$ as defined in eq.(\ref{sisalvichipuo}).
It is evident by the above completely universal formulae that the potential:
\begin{equation}\label{potentus}
V_{gauging} \, = \, \mbox{const} \, \mbox{Tr} \left[ \, \mathfrak{P}_\mathfrak{p}\, \cdot \, \mathfrak{P}_\mathfrak{p} \right]
\end{equation}
possesses the following universal property:
\begin{equation}\label{Zetaseneva}
\left. \frac{\partial}{\partial \mathbf{Z}^\alpha} \, V_{gauging}\right|_{\mathbf{Z}=0} \, = \,0
\end{equation}
allowing for a consistent truncation of the Heisenberg fields. After such truncation we find:
\begin{equation}\label{Vzero}
V_{eff}(U,a, z,\bar z) \, =\, \left.V_{gauging}\right|_{\mathbf{Z}=0} \, = \, \mbox{const} \, \times \, \left[ \mathcal{P}_{\mathfrak{N}_+} \, + \, \frac{\kappa}{2} \, e^{-U} \left(1+a^2+e^{2U}\right)\right]^2
\end{equation}
From equation (\ref{Vzero}) we further learn that we can consistently truncate the fields $a$ and $U$ setting them to zero since
\begin{equation}\label{ciurlonelmanicozzo}
\left. \frac{\partial}{\partial U} \, V_{eff}\right|_{U=a=0} \, = \,0 \quad ; \quad \left. \frac{\partial}{\partial a} \, V_{eff}\right|_{U=a=0} \, = \,0
\end{equation}
We find:
\begin{equation}\label{curiosone}
V_{infl}(z,\bar z) \, \equiv \, V_{eff}(0,0, z,\bar z) \, = \, \left( \, \mathcal{P}_{\mathfrak{N}_+} \, + \, \kappa \right)^2
\end{equation}
which clearly shows how the universal generator $\mathfrak{S}$ provides, after stabilization of the $U$ field, the mechanism that generates the Fayet Iliopoulos term \cite{Fayet:1974jb} essential for inflation.
\newpage
\part{\sc Examples}
As an illustration of the general patterns and mechanisms described in the previous pages we consider two examples of Quaternionic K\"ahler manifolds $\mathcal{QM}_{4n+4}$ obtained from the $c$-map of two homogeneous symmetric Special K\"ahler manifolds $\mathcal{SK}_n$.
\begin{enumerate}
\item The manifold $\frac{\mathrm{G_{(2,2)}}}{\mathrm{SU(2)} \times \mathrm{SU(2)}}$ which is the $c$-map image of the Special K\"ahler manifold $\frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}$ with cubic embedding of $\mathrm{SU(1,1)}$ in $\mathrm{Sp(4,\mathbb{R})}$. In this case $n=1$ and the corresponding coupling of one vector multiplet to supergravity is usually named the $S^3$ model in the literature.
\item The manifold $\frac{\mathrm{F_{(4,4)}}}{\mathrm{SU(2)} \times \mathrm{USp(6)}}$ which is the $c$-map image of the Special K\"ahler manifold $\frac{\mathrm{Sp(6,\mathbb{R})}}{\mathrm{SU(3) \times U(1)}}$ . In this case $n=6$.
\end{enumerate}
For these two models we provide a full fledged construction of all the geometrical items and in particular we realize the bridge between the algebraic description and the analytic one advocated at the end of section \ref{MezhduAlgGeom}. This allows us to discuss a couple of examples of gaugings. In particular in the case of the of the first model which we utilized as a calibration device for our general formulae we retrieve the inclusion of the Starobinsky model first demonstrated in \cite{thesearch}.
\par
The detailed construction of the second model, which we plan to utilize in future publications for an extensive and possibly exhaustive analysis of gaugings in larger hypermultiplet spaces, is utilized in the present paper to provide an example of generalization of the results of \cite{thesearch} by means of the inclusion of a multi Starobinsky model.
\section{\sc The $S^3$ model and its quaternionic image $\frac{\mathrm{G_{(2,2)}}}{\mathrm{SU(2)} \times \mathrm{SU(2)}}$}
In this which is the simplest example $n=1$, namely the Special K\"ahler manifold has complex dimension $1$ and it can be identified with the time honored Poincar\'e Lobachevsky plane:
\begin{equation}\label{tripini}
\mathcal{SK}_{1} \, = \, \frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}
\end{equation}
\subsection{\sc The special K\"ahler structure of $S^3$}
The corresponding K\"ahler potential is:
\begin{equation}\label{kelero1}
\mathcal{K} \, = \, - \, \log \, \left[ (z \, - \, \bar{z})^3\right ]
\end{equation}
which leads to the K\"ahler metric:
\begin{equation}\label{kelero2}
g_{z\bar{z}} \, = \, \frac{3}{4} \, \frac{1}{(z \, - \, \bar{z})^2}
\end{equation}
Setting:
\begin{equation}\label{realcordo}
z \, = \, {\rm i} \, \exp[h] \, + \, y
\end{equation}
we get:
\begin{equation}\label{metrullone}
g_{z\bar{z}} \, = \, \frac{3}{2} \, \left(dh^2 \, + \, \exp[-2h] \, dy^2\right)
\end{equation}
In the notations of \cite{PietroSashaMarioBH1} the holomorphic symplectic section governing this special geometry is given
by the following four component vector:
\begin{equation}\label{seziona}
\Omega \, = \,\left\{-\sqrt{3}z^2,z^3,\sqrt{3} z,1\right\}
\end{equation}
In this case the $\mathbf{W}$-representation is the spin $j\, = \, \frac{3}{2}$ of the $\mathrm{SL(2,\mathbb{R})} \sim \mathrm{SU(1,1)}$ group that happens to be
four dimensional symplectic:
\begin{equation}
\label{frilli}
\mathrm{SL(2,\mathbb{R})} \, \ni \,\left(\begin{array}{ll}
a & b \\
c & d
\end{array} \right) \, \Longrightarrow \, \left(
\begin{array}{llll}
d a^2+2 b c a & -\sqrt{3} a^2
c & -c b^2-2 a d b &
-\sqrt{3} b^2 d \\
-\sqrt{3} a^2 b & a^3 &
\sqrt{3} a b^2 & b^3 \\
-b c^2-2 a d c & \sqrt{3} a
c^2 & a d^2+2 b c d &
\sqrt{3} b d^2 \\
-\sqrt{3} c^2 d & c^3 &
\sqrt{3} c d^2 & d^3
\end{array}
\right) \, \in \, \mathrm{Sp(4,\mathbb{R})}
\end{equation}
the preserved symplectic metric being the following one:
\begin{equation}\label{goriaci}
\mathbb{C} \, = \, \left(
\begin{array}{llll}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
-1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}
\right)
\end{equation}
According to the general rule the K\"ahler potential (\ref{kelero1}) is retrieved by setting:
\begin{equation}\label{fagano}
\mathcal{K}(z,{\bar z}) \, = \, - \log \left[ - {\rm i} \Omega \, \mathbb{C} \, \
\overline{\Omega} \right ]
\end{equation}
\subsection{\sc The matrix $\mathcal{M}_4^{-1}$ and the $c$-map}
For the $S^3$ model the matrix $\mathcal{M}_4$ and its inverse have the following explicit appearance:
\begin{equation}\label{m4diretto}
\mathcal{M}_4 \, = \, \left(
\begin{array}{llll}
\frac{4 i z {\bar z} \left(z^2+4 {\bar z}
z+{\bar z}^2\right)}{(z-{\bar z})^3} & -\frac{4 i \sqrt{3} z^2
{\bar z}^2 (z+{\bar z})}{(z-{\bar z})^3} & -\frac{i
(z+{\bar z}) \left(z^2+10 {\bar z}
z+{\bar z}^2\right)}{(z-{\bar z})^3} & -\frac{2 i \sqrt{3}
(z+{\bar z})^2}{(z-{\bar z})^3} \\
-\frac{4 i \sqrt{3} z^2 {\bar z}^2 (z+{\bar z})}{(z-{\bar z})^3} &
\frac{8 i z^3 {\bar z}^3}{(z-{\bar z})^3} & \frac{2 i \sqrt{3} z
{\bar z} (z+{\bar z})^2}{(z-{\bar z})^3} & \frac{i
(z+{\bar z})^3}{(z-{\bar z})^3} \\
-\frac{i (z+{\bar z}) \left(z^2+10 {\bar z}
z+{\bar z}^2\right)}{(z-{\bar z})^3} & \frac{2 i \sqrt{3} z
{\bar z} (z+{\bar z})^2}{(z-{\bar z})^3} & \frac{4 i \left(z^2+4
{\bar z} z+{\bar z}^2\right)}{(z-{\bar z})^3} & \frac{4 i
\sqrt{3} (z+{\bar z})}{(z-{\bar z})^3} \\
-\frac{2 i \sqrt{3} (z+{\bar z})^2}{(z-{\bar z})^3} & \frac{i
(z+{\bar z})^3}{(z-{\bar z})^3} & \frac{4 i \sqrt{3}
(z+{\bar z})}{(z-{\bar z})^3} & \frac{8 i}{(z-{\bar z})^3}
\end{array}
\right)
\end{equation}
its inverse being:
\begin{equation}\label{m4inverso}
\mathcal{M}_4^{-1} \, = \,\left(
\begin{array}{llll}
\frac{4 i \left(z^2+4 {\bar z}
z+{\bar z}^2\right)}{(z-{\bar z})^3} & \frac{4 i \sqrt{3}
(z+{\bar z})}{(z-{\bar z})^3} & \frac{i \left(z^3+11 {\bar z}
z^2+11 {\bar z}^2 z+{\bar z}^3\right)}{(z-{\bar z})^3} &
-\frac{2 i \sqrt{3} z {\bar z} (z+{\bar z})^2}{(z-{\bar z})^3}
\\
\frac{4 i \sqrt{3} (z+{\bar z})}{(z-{\bar z})^3} & \frac{8
i}{(z-{\bar z})^3} & \frac{2 i \sqrt{3}
(z+{\bar z})^2}{(z-{\bar z})^3} & -\frac{i
(z+{\bar z})^3}{(z-{\bar z})^3} \\
\frac{i \left(z^3+11 {\bar z} z^2+11 {\bar z}^2
z+{\bar z}^3\right)}{(z-{\bar z})^3} & \frac{2 i \sqrt{3}
(z+{\bar z})^2}{(z-{\bar z})^3} & \frac{4 i z {\bar z}
\left(z^2+4 {\bar z} z+{\bar z}^2\right)}{(z-{\bar z})^3} &
-\frac{4 i \sqrt{3} z^2 {\bar z}^2 (z+{\bar z})}{(z-{\bar z})^3}
\\
-\frac{2 i \sqrt{3} z {\bar z} (z+{\bar z})^2}{(z-{\bar z})^3} &
-\frac{i (z+{\bar z})^3}{(z-{\bar z})^3} & -\frac{4 i \sqrt{3}
z^2 {\bar z}^2 (z+{\bar z})}{(z-{\bar z})^3} & \frac{8 i z^3
{\bar z}^3}{(z-{\bar z})^3}
\end{array}
\right)
\end{equation}
Furthermore, in this case a convenient reference point is given by $z_0 \, = \, {\rm i}$ that can be mapped into any point of the upper complex plane by means of the element:
\begin{equation}\label{trasluco}
\mathfrak{g}_z \, = \, \ \left(
\begin{array}{ll}
e^{h/2} & e^{-h/2} y \\
0 & e^{-h/2}
\end{array}
\right) \, \in \, \mathrm{SL(2,\mathbb{R})}
\end{equation}
acting by means of fractional linear transformations. The explicit form of the $\Lambda(\mathfrak{g})$ matrix in the $\mathbf{W}$-representation was given in eq.(\ref{frilli}). This provides us with all the necessary information in order to write down the explicit form of the $E^I_{\mathcal{QM}} $ vielbein for the $S^3$ case.
\subsection{\sc The vielbein and the borellian Maurer Cartan equations}
They are the following ones:
\begin{equation}\label{EBHfilo}
E^I_{\mathcal{QM}} \, = \, \frac{1}{2} \left(
\begin{array}{l}
\mathrm{dU} \\
\sqrt{3} \mathrm{dh} \\
\sqrt{3} \mathrm{dy} e^{-h} \\
e^{-U} \left(\mathrm{da}+\mathrm{dZ}_3 Z_1+\mathrm{dZ}_4 Z_2-\mathrm{dZ}_1 Z_3-\mathrm{dZ}_2
Z_4\right) \\
\sqrt{2} e^{-\frac{h}{2}-\frac{U}{2}}
\left(\mathrm{dZ}_1+y \left(2
\mathrm{dZ}_3-\sqrt{3} y
\mathrm{dZ}_4\right)\right) \\
\sqrt{2} e^{-\frac{3 h}{2}-\frac{U}{2}}
\left(\left(\sqrt{3} \mathrm{dZ}_3-y
\mathrm{dZ}_4\right) y^2+\sqrt{3} \mathrm{dZ}_1
y+\mathrm{dZ}_2\right)\\
\sqrt{2} e^{\frac{h-U}{2}}
\left(\mathrm{dZ}_3-\sqrt{3} y
\mathrm{dZ}_4\right) \\
\sqrt{2} e^{\frac{3 h}{2}-\frac{U}{2}}
\mathrm{dZ}_4
\end{array}
\right)
\end{equation}
Furthermore we find $\mathcal{M}_4^{-1}({\rm i},-{\rm i}) \, = \, - \, \mathbf{1}_{4\times4}$ so that the quadratic form (\ref{quadrotta}) is just:
\begin{equation}\label{doremito}
\mathfrak{q}_{AB} \, = \, \mbox{diag} \,\left( 1,1,1,1,1,1,1,1\right)
\end{equation}
The next step consists of calculating the geometry of the space described by the above vielbein and flat metric (\ref{doremito}). To this effect we have first to calculate the contorsion, namely the exterior derivatives of the vielbein and then using such a result the spin connection $\omega^{IJ}$, finally the curvature two-form from which we extract the Riemann and the Ricci tensor.
\par
Addressing the first step, namely the contorsion, we have the first important surprise. The exterior derivatives of the vielbein are expressed in terms of wedge-quadratic products of the same vielbein with constant numerical coefficients. This means that the above constructed vielbein satisfy a set of Maurer Cartan equations describing a Lie algebra, namely\footnote{Note that here, for simplicity we have dropped the suffix $\mathcal{SK}$. This is done for simplicity since there is no risk of confusion.}:
\begin{equation}\label{glorioso}
dE^I \, - \, \frac{1}{2} \, f_{JK}^{\phantom{JK}I} \, E^I \, \wedge \, E^J \, = \, 0
\end{equation}
the tensor $f_{BC}^{\phantom{BC}A}$ being the structure constants of such a Lie algebra. Explicitly for the $S^3$ model we get:
\begin{equation}\label{primocartano}
\begin{array}{l}
0 \, = \, dE^{1} \\
0 \, = \, dE^{2} \\
0 \, = \, dE^{3}\, + \, 2\,\frac{E^{2}\wedge
E^{3}}{\sqrt{3}} \\
0 \, = \, dE^{4}\, +\, 2 \, E^{1}\wedge
E^{4}\, -\, 2 \, E^{5}\wedge
E^{7}\, -\, 2 \, E^{6}\wedge E^{8}
\\
0 \, = \, dE^{5}+ E^{1}\wedge
E^{5}\, +\, \frac{E^{2}\wedge
E^{5}}{ \sqrt{3}}-\frac{4 \,
E^{3}\wedge E^{7}}{\sqrt{3}} \\
0 \, = \, dE^{6}+ E^{1}\wedge
E^{6}+ \sqrt{3}
E^{2}\wedge
E^{6}\, -\, 2\, E^{3}\wedge E^{5}
\\
0 \, = \, dE^{7}+ E^{1}\wedge
E^{7}-\frac{E^{2}\wedge
E^{7}}{\sqrt{3}}\, +\, 2 \, E^{3}\wedge
E^{8} \\
0 \, = \, dE^{8}+E^{1}\wedge
E^{8}- \sqrt{3}
E^{2}\wedge E^{8}
\end{array}
\end{equation}
Hence it arises the following question: which Lie algebra is described by such Maurer Cartan equations?
Utilizing the standard method of diagonalizing the adjoint action of the two commuting generators $H_{1,2}$ dual to $E^{1,2}$ we find that the eigenvalues are just the positive roots of $\mathfrak{g}_{2,2}$:
\begin{equation}
\label{g2rootsystem}
\begin{array}{rclcrcl}
\alpha_1&=&(1,0)&;&\alpha_2&=&\frac{\sqrt{3}}{2}\,(-\sqrt{3},1)\\
\alpha_3 \, =\, \alpha_1+\alpha_2&=&\frac{1}{2}\,(-1,\sqrt{3}) &;&
\alpha_4 \, = \, 2\,\alpha_1+\alpha_2 &=&
\frac{1}{2}\,(1,\sqrt{3}) \\
\alpha_5 \, = \, 3\,\alpha_1+\alpha_2&=&\frac{\sqrt{3}}{2}\,(\sqrt{3},1)&;&\alpha_6 \, = \, 3\,\alpha_1+2\,\alpha_2 &=&
(0,\sqrt{3})\
\end{array}
\end{equation}
As it is well known the complex Lie algebra $\mathfrak{g}_2(\mathbb{C})$ has rank two and it is defined by the $2\times 2$ Cartan matrix encoded in the following Dynkin diagram:
\begin{center}
\begin{picture}(110,30)
\put (-60,20){$\mathfrak{g}_2$}
\put (10,23){\circle {10}}
\put (15,25){\line (1,0){20}}
\put (15,23){\line (1,0){20}}
\put (20.5,19){{\LARGE$>$}}
\put (15,20.5){\line (1,0){20}}
\put (40,23){\circle {10}}
\put (65,21){$=\quad\quad\left (\begin{array}{cc}
2 & -3\\
-1 & 2
\end{array} \right)$}
\end{picture}
\end{center}
The real form $\mathfrak{g}_{2,2}$ is the maximally split form of the above complex Lie algebra.
With a little bit of more work we can put eq.s(\ref{primocartano}) into the standard Cartan Weyl form for the Borel subalgebra of $\mathfrak{g}_{2,2}$, composed by the Cartan generators and by all the positive root step operators. Naming $T_J$ the generators dual to the vielbein $E^I$ such that $E^I(T_J) \, = \delta^I_J$, we find that the appropriate identifications are the following ones:
\begin{equation}\label{interpretaziag22}
\begin{array}{l}
T_2 \, = \, 2 \, \frac{\mathcal{H}_1}{\sqrt{3}} \\
T_1 \, = \, 2 \, \frac{\mathcal{H}_2}{\sqrt{3}} \\
T_3 \, = \, 2 \, E^{\alpha _1} \\
T_4 \, = \, 2 \, E^{\alpha _6} \\
T_8 \, = \, 2 \, E^{\alpha _2} \\
T_7 \, = \, 2 \, E^{\alpha _3 } \\
T_5 \, = \, 2 \, E^{\alpha _4} \\
T_6 \, = \, 2 \, E^{\alpha _5}
\end{array}
\end{equation}
We conclude that the manifold on which the metric (\ref{geodaction}) is constructed is homeomorphic to the solvable group-manifold $\mbox{Bor}(\mathfrak{g}_{2,2})$.
\subsection{\sc The spin connection}
Next, calculating the Levi-Civita spin connection from its definition, namely the vanishing torsion condition:
\begin{equation}\label{surcallo}
0 \, = \, dE^I \, + \, \omega^{IJ} \, \wedge \, E^J
\end{equation}
we find the following result:
\begin{equation}\label{ristolone}
\omega^{IJ} \, = \, \left(
\begin{array}{llllllll}
0 & 0 & 0 & E^{4} &
\frac{E^{5}}{2} &
\frac{E^{6}}{2} &
\frac{E^{7}}{2} &
\frac{E^{8}}{2} \\
0 & 0 & \frac{E^{3}}{\sqrt{3}} & 0 &
\frac{E^{5}}{2 \sqrt{3}} & \frac{1}{2}
\sqrt{3} E^{6} & -\frac{E^{7}}{2
\sqrt{3}} & -\frac{1}{2} \sqrt{3} E^{8}
\\
0 & -\frac{E^{3}}{\sqrt{3}} & 0 & 0 &
-\frac{E^{6}}{2}-\frac{E^{7}}{\sqrt{3}} & -\frac{E^{5}}{2} &
\frac{E^{8}}{2}-\frac{E^{5}}{\sqrt{3}} & \frac{E^{7}}{2} \\
-E^{4} & 0 & 0 & 0 &
\frac{E^{7}}{2} &
\frac{E^{8}}{2} &
-\frac{E^{5}}{2} &
-\frac{E^{6}}{2} \\
-\frac{E^{5}}{2} & -\frac{E^{5}}{2
\sqrt{3}} &
\frac{E^{6}}{2}+\frac{E^{7}}{\sqrt{3}} & -\frac{E^{7}}{2} & 0 &
\frac{E^{3}}{2} &
-\frac{E^{3}}{\sqrt{3}}-\frac{E^4}{2} & 0 \\
-\frac{E^{6}}{2} & -\frac{1}{2} \sqrt{3} \,
E^{6} & \frac{E^{5}}{2} &
-\frac{E^{8}}{2} &
-\frac{E^{3}}{2} & 0 & 0 &
-\frac{E^{4}}{2} \\
-\frac{E^{7}}{2} & \frac{E^{7}}{2
\sqrt{3}} &
\frac{E^{5}}{\sqrt{3}}-\frac{E^{
8}}{2} & \frac{E^{5}}{2} &
\frac{E^{3}}{\sqrt{3}}+\frac{E^{
4}}{2} & 0 & 0 & \frac{E^{3}}{2} \\
-\frac{E^{8}}{2} & \frac{1}{2} \sqrt{3}
E^{8} & -\frac{E^{7}}{2} &
\frac{E^{6}}{2} & 0 &
\frac{E^{4}}{2} &
-\frac{E^{3}}{2} & 0
\end{array}
\right)
\end{equation}
which can be decomposed in the way we now describe.
\subsection{\sc Holonomy algebra and decompostion of the spin connection}
Let us introduce two triplets $J^x_{[I]}$ and $J^x_{[II]}$ of $8 \times 8$ matrices that can be read off explicitly as the coefficients of $\alpha_x$ and $\beta_x$ in the following linear combinations:
\begin{eqnarray}\label{gringo1}
&\sum_{x=1}^3 \, \alpha_x \, J^x_{[I]} \, = \, & \nonumber\\
& \left(
\begin{array}{llllllll}
0 & 0 & 0 & -\frac{\alpha _1}{2} & -\frac{1}{4}
\sqrt{3} \alpha _3 & -\frac{\alpha _2}{4} &
\frac{\sqrt{3} \alpha _2}{4} & -\frac{\alpha
_3}{4} \\
0 & 0 & \frac{\alpha _1}{2} & 0 & -\frac{\alpha
_3}{4} & -\frac{1}{4} \sqrt{3} \alpha _2 &
-\frac{\alpha _2}{4} & \frac{\sqrt{3} \alpha
_3}{4} \\
0 & -\frac{\alpha _1}{2} & 0 & 0 & -\frac{\alpha
_2}{4} & \frac{\sqrt{3} \alpha _3}{4} &
\frac{\alpha _3}{4} & \frac{\sqrt{3} \alpha
_2}{4} \\
\frac{\alpha _1}{2} & 0 & 0 & 0 & \frac{\sqrt{3}
\alpha _2}{4} & -\frac{\alpha _3}{4} &
\frac{\sqrt{3} \alpha _3}{4} & \frac{\alpha
_2}{4} \\
\frac{\sqrt{3} \alpha _3}{4} & \frac{\alpha
_3}{4} & \frac{\alpha _2}{4} & -\frac{1}{4}
\sqrt{3} \alpha _2 & 0 & \frac{\sqrt{3} \alpha
_1}{4} & -\frac{\alpha _1}{4} & 0 \\
\frac{\alpha _2}{4} & \frac{\sqrt{3} \alpha
_2}{4} & -\frac{1}{4} \sqrt{3} \alpha _3 &
\frac{\alpha _3}{4} & -\frac{1}{4} \sqrt{3}
\alpha _1 & 0 & 0 & \frac{\alpha _1}{4} \\
-\frac{1}{4} \sqrt{3} \alpha _2 & \frac{\alpha
_2}{4} & -\frac{\alpha _3}{4} & -\frac{1}{4}
\sqrt{3} \alpha _3 & \frac{\alpha _1}{4} & 0 &
0 & \frac{\sqrt{3} \alpha _1}{4} \\
\frac{\alpha _3}{4} & -\frac{1}{4} \sqrt{3}
\alpha _3 & -\frac{1}{4} \sqrt{3} \alpha _2 &
-\frac{\alpha _2}{4} & 0 & -\frac{\alpha
_1}{4} & -\frac{1}{4} \sqrt{3} \alpha _1 & 0
\end{array}
\right) & \nonumber\\
\end{eqnarray}
\begin{eqnarray}\label{gringo2}
&\sum_{x=1}^3 \, \beta_x \, J^x_{[II]} \, = \, & \nonumber\\
& \left(
\begin{array}{llllllll}
0 & 0 & 0 & \frac{3 \beta _1}{2} & -\frac{1}{4}
\sqrt{3} \beta _3 & -\frac{3 \beta _2}{4} &
-\frac{1}{4} \sqrt{3} \beta _2 & \frac{3 \beta
_3}{4} \\
0 & 0 & \frac{\beta _1}{2} & 0 & -\frac{\beta
_3}{4} & -\frac{3}{4} \sqrt{3} \beta _2 &
\frac{\beta _2}{4} & -\frac{3}{4} \sqrt{3}
\beta _3 \\
0 & -\frac{\beta _1}{2} & 0 & 0 & \frac{5 \beta
_2}{4} & \frac{\sqrt{3} \beta _3}{4} & \frac{5
\beta _3}{4} & -\frac{1}{4} \sqrt{3} \beta _2
\\
-\frac{3 \beta _1}{2} & 0 & 0 & 0 & -\frac{1}{4}
\sqrt{3} \beta _2 & \frac{3 \beta _3}{4} &
\frac{\sqrt{3} \beta _3}{4} & \frac{3 \beta
_2}{4} \\
\frac{\sqrt{3} \beta _3}{4} & \frac{\beta _3}{4}
& -\frac{5 \beta _2}{4} & \frac{\sqrt{3} \beta
_2}{4} & 0 & \frac{\sqrt{3} \beta _1}{4} &
-\frac{5 \beta _1}{4} & 0 \\
\frac{3 \beta _2}{4} & \frac{3 \sqrt{3} \beta
_2}{4} & -\frac{1}{4} \sqrt{3} \beta _3 &
-\frac{3 \beta _3}{4} & -\frac{1}{4} \sqrt{3}
\beta _1 & 0 & 0 & -\frac{3 \beta _1}{4} \\
\frac{\sqrt{3} \beta _2}{4} & -\frac{\beta
_2}{4} & -\frac{5 \beta _3}{4} & -\frac{1}{4}
\sqrt{3} \beta _3 & \frac{5 \beta _1}{4} & 0 &
0 & \frac{\sqrt{3} \beta _1}{4} \\
-\frac{3 \beta _3}{4} & \frac{3 \sqrt{3} \beta
_3}{4} & \frac{\sqrt{3} \beta _2}{4} &
-\frac{3 \beta _2}{4} & 0 & \frac{3 \beta
_1}{4} & -\frac{1}{4} \sqrt{3} \beta _1 & 0
\end{array}
\right) & \nonumber\\
\end{eqnarray}
Both triplets form an 8-dimensional representation of the $\mathfrak{su}(2)$ Lie algebra and the two triplets commute with each other:
\begin{eqnarray}
\left[ \,J^x_{[I]} \, , \, J^y_{[I]} \right ] &=& \epsilon^{xyz} \, J^y_{[I]} \nonumber\\
\left[ \,J^x_{[II]} \, , \, J^y_{[II]} \right ] &=& \epsilon^{xyz} \, J^y_{[II]} \nonumber\\
\left[ \,J^x_{[I]} \, , \, J^y_{[II]} \right ] &=&0 \label{guglielmotell}
\end{eqnarray}
Furthermore all matrices are antisymmetric so that the two Lie algebras $\mathfrak{su}_{\mathrm{I}}(2)$ and $\mathfrak{su}_{\mathrm{II}}(2)$ are both subalgebras of $\mathfrak{so}(8)$.
The distinction between these two representations becomes clear when we calculate the Casimir operator for both of them. We obtain:
\begin{equation}\label{gaglioffo}
\sum_{x=1}^3 \, J^x_{[I]}\cdot J^x_{[I]} \, = \, - \, \frac{3}{4} \, \mathbf{1} \quad ; \quad \sum_{x=1}^3 \, J^x_{[II]}\cdot J^x_{[II]} \, = \, - \, \frac{15}{4} \, \mathbf{1}
\end{equation}
Hence the first $\mathfrak{su}_{\mathrm{I}}(2)$ Lie algebra in realized on the considered eight--dimensional space in the $j=\frac{1}{2}$ representation, while the second
$\mathfrak{su}_{\mathrm{II}}(2)$ Lie algebra in realized on the same space in the $j=\frac{3}{2}$. In other words, with respect to both subalgebras of $\mathfrak{so}(8)$, the fundamental representation decomposes as follows:
\begin{equation}\label{cirimello}
\mathbf{8} \, \stackrel{\mathfrak{su}_{\mathrm{I}}(2) \oplus \mathfrak{su}_{\mathrm{II}}(2) \subset \mathfrak{so}(8)}{\Longrightarrow} \, (\mathbf{2},\mathbf{4})
\end{equation}
By direct calculation we verify that the spin connection displayed in equation (\ref{ristolone}) has the following structure:
\begin{equation}\label{decompospincon}
\omega \, = \, \omega^{[I]}_x \, J^x_{[I]} \, \oplus \, \omega^{[II]}_x \, J^x_{[II]}
\end{equation}
where:
\begin{equation}\label{cucurucu}
\omega^{[I]}_x \, = \, \left(
\begin{array}{l}
\sqrt{3} \mathrm{E}^3- \mathrm{E}^4 \\
\sqrt{3} \mathrm{E}^7- \mathrm{E}^6 \\
- \mathrm{E}^8- \sqrt{3} \mathrm{E}^5
\end{array}
\right) \quad ; \quad \omega^{[II]}_x \, = \, \left(
\begin{array}{l}
\frac{\mathrm{E}^4}{2}+\frac{\mathrm{E}^3}{2
\sqrt{3}} \\
-\frac{\mathrm{E}^7}{2
\sqrt{3}}-\frac{\mathrm{E}^6}{2} \\
\frac{\mathrm{E}^8}{2}-\frac{\mathrm{E}^5}{2
\sqrt{3}}
\end{array}
\right)
\end{equation}
This structure clearly demonstrates the reduced holonomy of the Quaternionic K\"ahler manifold. Indeed, according to eq.(\ref{cirimello}) the vielbein transforms in the doublet of $\mathfrak{su}_{\mathrm{I}}(2)$ tensored with the fundamental representation of $\mathfrak{sp}(4,\mathbb{R})$. In the present case the symplectic $4 \times 4$ matrices are actually reduced to the subalgebra $\mathfrak{su}_{\mathrm{II}}(2) \subset \mathfrak{sp}(4,\mathbb{R})$ with respect to which the fundamental of $\mathfrak{sp}(4,\mathbb{R})$ remains irreducible and coincides with the $j=\frac{3}{2}$ representation of $\mathfrak{su}_{\mathrm{II}}(2)$. The above discussion can be summarized by the statement:
\begin{equation}\label{ollonomio}
\mathfrak{so}(8) \, \subset \, \mathfrak{su}(2) \, \oplus \, \mathfrak{usp}(4) \, \subset \, \mathrm{Hol} \, = \, \mathfrak{su}_{\mathrm{I}}(2) \, \oplus \, \mathfrak{su}_{\mathrm{II}}(2)
\end{equation}
by definition the holonomy algebra being the Lie algebra in which the Levi-Civita spin connection takes values.
\subsection{\sc Structure of the isotropy subalgebra $\mathbb{H}$}
It remains to single out the structure of the denominator subalgebra ${\mathbb{H}} \, \subset \, \mathbb{U} \, \equiv \, {\mathfrak{g}}_{2,2}$ in the orthogonal decomposition:
\begin{equation}\label{ortogonallodecompo}
\mathbb{U} \, = \, \mathbb{H} \, \oplus \, \mathbb{K} \quad ; \quad \left \{ \begin{array}{ccc}
\left[ \mathbb{H} \, , \, \mathbb{H} \right] & \subset & \mathbb{H} \\
\left[ \mathbb{H} \, , \, \mathbb{K} \right] & \subset & \mathbb{K} \\
\left[ \mathbb{K} \, , \, \mathbb{K} \right] & \subset & \mathbb{H} \\
\end{array}\right.
\end{equation}
Since our quaternionic K\"ahler manifold is a symmetric space it follows that Lie algebra $\mathbb{H}$ must be isomorphic with the holonomy algebra $\mathrm{Hol} \, = \,\mathfrak{su}_{\mathrm{I}}(2) \oplus \mathfrak{su}_{\mathrm{II}}(2)$ that we have calculated in the previous subsection.
By definition the Lie algebra $\mathbb{H}$ is the maximal compact subalgebra which for maximal split algebras has a universal definition in terms of the step operators associated with the positive roots $E^\alpha$ and their conjugates $E^{-\alpha}$. In the case of $\mathfrak{g}_{2,2}$ which has six positive roots we can write:
\begin{equation}\label{Hfavola}
\mathbb{H} \, \equiv \,\mbox{span}_{\mathbb{R}} \left \{ E^{\alpha _1} - E^{\alpha _1}\, , \, E^{\alpha _2} - E^{\alpha _2} \, , \, E^{\alpha _3} - E^{\alpha _3} \, , \, E^{\alpha _4} - E^{\alpha _4}\, , \, E^{\alpha _5} - E^{\alpha _5} \, , \, E^{\alpha _6} - E^{\alpha _6} \right \}
\end{equation}
The structure of (\ref{Hfavola}) is the following:
\begin{equation}\label{fattucco}
\mathbb{H} \, = \, \mathfrak{su}_{\mathrm{I}}(2) \, \oplus \, \mathfrak{su}_{\mathrm{II}}(2)
\end{equation}
where the generators of the two subalgebras are:
\begin{equation}\label{friccouno}
j^x_{[I]} \, = \,\left(
\begin{array}{l}
\frac{-3 \mathrm{E}^{-\alpha_1}+3
\mathrm{E}^{\alpha_1}+\sqrt{3}
\left(\mathrm{E}^{\alpha_6}-\mathrm{E}^{-\alpha_6}\right)}{6 \sqrt{2}} \\
\frac{3 \mathrm{E}^{-\alpha_3}-3
\mathrm{E}^{\alpha_3}+\sqrt{3}
\left(\mathrm{E}^{-\alpha_5}-\mathrm{E}^{\alpha_5}\right)}{6 \sqrt{2}} \\
\frac{\sqrt{3} \left(\mathrm{E}^{\alpha_2}-\mathrm{E}^{-\alpha_2}\right)+3
\left(\mathrm{E}^{-\alpha_4}-\mathrm{E}^{\alpha_4}\right)}{6 \sqrt{2}}
\end{array}
\right)
\end{equation}
and
\begin{equation}\label{friccodue}
j^x_{[II]} \, = \,\left(
\begin{array}{l}
\frac{-\mathrm{E}^{-\alpha_1}+\mathrm{E}^{\alpha_1}+\sqrt{3} \left(\mathrm{E}^{-\alpha_6}-\mathrm{E}^{\alpha_6}\right)}{2 \sqrt{2}}
\\
\frac{-\mathrm{E}^{-\alpha_3}+\mathrm{E}^{\alpha_3}+\sqrt{3} \left(\mathrm{E}^{-\alpha_5}-\mathrm{E}^{\alpha_5}\right)}{2 \sqrt{2}}
\\
\frac{\sqrt{3} \left(\mathrm{E}^{-\alpha_2}-\mathrm{E}^{\alpha_2}\right)+\mathrm{E}^{-\alpha_4}-\mathrm{E}^{\alpha_4}}{2 \sqrt{2}}
\end{array}
\right)
\end{equation}
and satisfy among themselves the same relations (\ref{guglielmotell}) as their homologous generators $ J^x_{[I]}$ and $ J^x_{[II]}$.
In eq.(\ref{fattucco}) we have used the same notation as in eq.(\ref{ollonomio}) using the obligatory homomorphism between the the holonomy algebra $\mathrm{Hol}$ and the isotropy subalgebra $\mathbb{H}$. The precise correspondence between generators of one algebra and generators of the other will be establishe in the next subsection by means of the use of the coset representative.
\subsection{\sc The coset representative}
The next step in the development of the coset approach is the construction of the solvable coset representative $\mathbb{L}_{Solv}(\phi)$, advocated in eq.s(\ref{solvoexpo}-\ref{sopore} ), namely a coordinate dependent element of the Borel group of $\mathfrak{g}_{(2,2)}$ such that the Maurer Cartan form
\begin{equation}\label{rupiaindu}
\Xi \, = \, \mathbb{L}_{Solv}(\phi)^{-1} \, \mathrm{d}\mathbb{L}_{Solv}(\phi)
\end{equation}
projected along the Borel algebra generators, as given in eq.(\ref{interpretaziag22}), reproduces the vielbein of eq.(\ref{EBHfilo}).
The appropriate coset representative is obtained by exponentiating the Borel Lie algebra and the precise recipe is provided below.
First define:
\begin{eqnarray}\label{generillini}
\mathbf{L}^E_0 & = & \frac{1}{\sqrt{3}} \,\mathcal{H}_2 \quad ; \quad \mathbf{L}^E_+ \, = \, -\sqrt{\frac{2}{3}}\, E^{\alpha_6} \nonumber\\
\mathbf{L}_0 & = &\,\mathcal{H}_1 \quad ; \quad \mathbf{L}_+ \, = \, \sqrt{2} E^{\alpha_1} \nonumber\\
\mathbf{W}^I & = & \sqrt{\frac{2}{3}} \, \left \{ E^{\alpha_4} \, , \, E^{\alpha_5} \, , \, - \, E^{\alpha_3}\, , \, - \, E^{\alpha_2} \, , \, \right \}
\end{eqnarray}
and then set:
\begin{eqnarray}\label{fornarina}
\mathbb{L}& = & \exp\left [a \mathbf{L}^E_+\right]\,\cdot \, \exp\left [\sqrt{2} \left(Z_1 \, \mathbf{W}^1 \, + \, Z_3 \, \mathbf{W}^3\right )\right]\,\cdot \, \nonumber\\
&&\,\cdot \,\exp\left [\sqrt{2} \left(Z_1 \, \mathbf{W}^1 \, + \, Z_3 \, \mathbf{W}^3\right )\right]\,\cdot \,
\exp\left [y \mathbf{L}_+\right]\,\cdot \,\exp\left [h \mathbf{L}_0\right]\,\cdot \,\exp\left [U \mathbf{L}^E_0\right]
\end{eqnarray}
By explicit evaluation we obtain the result displayed in the appendix in formulae (\ref{fornoalegna}) and (\ref{cosettuspresento})
and we verify that, if we set:
\begin{equation}\label{sequenzus}
\mathfrak{T}_I \, = \, \left\{ \mathbf{L}^E_0 \, , \, \mathbf{L}_0 \, , \, \mathbf{L}^E_+ \, , \, \mathbf{L}_+ \, , \,\mathbf{W}_1\, , \,\mathbf{W}_2\, , \,\mathbf{W}_3\, , \,\mathbf{W}_4\right\}
\end{equation}
upon substitution of (\ref{cosettuspresento}) into the Maurer Cartan form (\ref{rupiaindu})
we obtain:
\begin{equation}\label{rupiamoresca}
\mathbb{L}_{Solv}(\phi)^{-1} \, \mathrm{d}\mathbb{L}_{Solv}(\phi) \, = \, \sum_{I=1}^8 \, \mathfrak{T}_I \, E^I_{\mathcal{QM}}
\end{equation}
the forms $E^I_{\mathcal{QM}}$ being given in equation (\ref{EBHfilo}). Alternatively we can also write:
\begin{equation}\label{curlandico}
\mathbb{L}_{Solv}(\phi)^{-1} \, \mathrm{d}\mathbb{L}_{Solv}(\phi) \, = \, \sum_{x=1}^3 \, \left (\omega^{[I]}_x \, j^x_{[I]}\, \oplus \, \omega^{[II]}_x \, j^x_{[II]}\right) \, \oplus \, \sum_{I}^8 \, \mathbf{T}_I \, E^I_{\mathcal{SQ}}
\end{equation}
In the above equation $\omega^{[I]}_x$ and $\omega^{[II]}_x$ are the components of the spin connections given in eq. (\ref{cucurucu}),
$j^x_{[I]}$ and $j^x_{[II]}$ are the generators of $\mathbb{H}$ defined in eq. (\ref{friccouno},\ref{friccodue}) and $\mathbf{T}_I $ denotes a suitable base of generators in the $\mathbb{K}$ subspace of $\mathfrak{g}_{(2,2)}$ defined as:
\begin{eqnarray}\label{cofimus}
\mathbb{K} & \equiv & ,\mbox{span}_{\mathbb{R}} \left \{ \mathcal{H}_1 \, , \, \mathcal{H}_2 \, , \, E^{\alpha _1} + E^{\alpha _1}\, , \, E^{\alpha _2} + E^{\alpha _2} \, , \, E^{\alpha _3} + E^{\alpha _3} \right. \nonumber\\
& &\left. E^{\alpha _4} + E^{\alpha _4}\, , \, E^{\alpha _5} + E^{\alpha _5} \, , \, E^{\alpha _6} + E^{\alpha _6} \right \}
\end{eqnarray}
The precise form of the generators $\mathbf{T}_I $ is not relevant to our purposes and we omit it. The key point is instead the identification of the generators $j^x_{[I]}$ of $\mathbb{H}$ with generators $J^x_{[I]}$ of the holonomy algebra. This provides us with the knowledge of the quaternionic complex structures within the algebra $\mathbb{U}_\mathcal{Q}$ and allows to calculate the tri-holomorphic moment map of any generator $\mathbf{t}\, \in \, \mathbb{U}_\mathcal{Q}$ by means of the formula (\ref{generMapformula}) which in our case reads:
\begin{equation}\label{fisterone}
\mathcal{P}_{\mathbf{t}}^x \, = \,\frac{1}{2} \mbox{Tr}_{\mathbf{7}} \left ( j^x_{[I]} \, \mathbb{L}_{Solv}^{-1} \, \mathbf{t} \, \mathbb{L}_{Solv} \right )
\end{equation}
having denoted by $\mathbf{7}$ the $7$-dimensional fundamental representation of $\mathfrak{g}_{(2,2)}$.
\paragraph{The Starobinsky potential}
As an immediate application of eq.(\ref{fisterone}) one can retrieve the results of \cite{thesearch} on the inclusion of the Starobinsky potential into supergravity. In section \ref{starobin1} we presented a general discussion of the gaugings of nilpotent generators in the Special K\"ahler subalgebra $\mathbb{U}_{\mathcal{SK}} \, \subset \, \mathbb{U}_\mathcal{Q}$. In the present case where $\mathbb{U}_{\mathcal{SK}} \, = \, \mathfrak{sl}(2,\mathbb{R})$ the only available nilpotent operator is $L_+$ and from the general formula (\ref{sisalvichipuo}) applied to the case where the metric is given by (\ref{kelero2}) and the complex coordinate is parameterized as in eq.(\ref{realcordo}) we find:
\begin{equation}\label{cubitagorio}
\mathcal{P}_{L_+} \, = \, \mbox{const} \, \times \, \exp[-h] \, = \, \mbox{const} \, \times \, \left(\mbox{Im} \,z\right)^{-1}
\end{equation}
This result inserted into the general formula (\ref{curiosone}) yields
\begin{equation}\label{starobbo}
V(h) \, = \, \mbox{const} \, \times \, \left(\exp[-h] \, + \, \kappa \right)^2
\end{equation}
which is indeed the Starobinsky potential, since, once expressed in terms of $h$, the K\"ahler potential is exactly $\mathcal{K}\, = \, 3 \, h$.
The same result is directly obtained with precise coefficients by inserting in eq.(\ref{fisterone}) the $7$-dimensional image of $L_+$ in the fundamental representation of $\mathfrak{g}_{(2,2)}$.
\newpage
\section{\sc The $\mathrm{Sp(6,\mathbb{R})}/\mathrm{SU(3)\times U(1)}$ - model and its c-map image.}
Next we consider the Special K\"ahler manifold
\begin{equation}\label{gurto}
\mathcal{M}_{\mathop{\rm {}Sp} 6} \, = \, \frac{\mathop{\rm {}Sp}(6,\mathbb{R})}{\mathrm{SU(3)\times U(1)}}
\end{equation}
and its c-map image which is the following quaternionic manifold:
\begin{equation}\label{quatergurto}
\mbox{$c$-map} \quad : \quad \mathcal{M}_{\mathop{\rm {}Sp} 6} \, \mapsto \, \mathcal{QM}_{F4} \, \equiv \, \frac{\mathrm{F_{(4,4)}}}{\mathrm{SU(2)} \times \mathrm{USp(6)}}
\end{equation}
$\mathcal{M}_{\mathop{\rm {}Sp} 6}$ belongs to the magic square of exceptional special K\"ahler manifolds whose quaternionic $c$-map is a homogeneous symmetric space having, as it is evident from (\ref{quatergurto}), an exceptional Lie group as isometry group.
\par
We begin by illustrating some general properties of this remarkable manifold. First of all, in order to discuss them adequately we need to choose a basis for the $\mathfrak{sp}(6,\mathbb{R})$ Lie algebra. Since we are not interested in solving Lax equations we do not choose the basis where the matrices of the Borel subalgebra are upper triangular. We rather use the basis where the symplectic preserved metric is the standard one, namely:
\begin{equation}\label{Cmatra}
\mathbb{C} \, = \, \left(
\begin{array}{llllll}
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
-1 & 0 & 0 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 & 0 \\
0 & 0 & -1 & 0 & 0 & 0
\end{array}
\right)
\end{equation}
This traditional choice allows to describe in a simple way other aspects of the manifold geometry that are more relevant to our present purposes.
\par
According to the above choice, an element of the $\mathop{\rm {}Sp}(6,\mathbb{R})$ group and an element of the $\mathfrak{sp}(6,\mathbb{R})$ Lie-algebra are matrices respectively fulfilling the following two constraints:
\begin{equation}\label{giluro}
\left( \begin{array}{c|c}
A & B \\
\hline
C & D
\end{array}
\right)^T \, \mathbb{C} \, \left( \begin{array}{c|c}
A & B \\
\hline
C & D
\end{array}
\right) \, = \, \mathbb{C} \quad ; \quad \left( \begin{array}{c|c}
\mathbf{A} & \mathbf{B} \\
\hline
\mathbf{C} & \mathbf{D}
\end{array}
\right)^T \, \mathbb{C} + \mathbb{C} \left( \begin{array}{c|c}
\mathbf{A} & \mathbf{B} \\
\hline
\mathbf{C} & \mathbf{D}
\end{array}
\right) \, = \, 0
\end{equation}
where $A,B,C,D$, $\mathbf{A,B,C,D}$ are $3\times 3$ blocks. By means of the so called Cayley transformation
\begin{equation}\label{caylus}
\mathcal{C} \, = \, \left(
\begin{array}{llllll}
\frac{1}{\sqrt{2}} & 0 & 0 & \frac{i}{\sqrt{2}} & 0 & 0 \\
0 & \frac{1}{\sqrt{2}} & 0 & 0 & \frac{i}{\sqrt{2}} & 0 \\
0 & 0 & \frac{1}{\sqrt{2}} & 0 & 0 & \frac{i}{\sqrt{2}} \\
\frac{1}{\sqrt{2}} & 0 & 0 & -\frac{i}{\sqrt{2}} & 0 & 0 \\
0 & \frac{1}{\sqrt{2}} & 0 & 0 & -\frac{i}{\sqrt{2}} & 0 \\
0 & 0 & \frac{1}{\sqrt{2}} & 0 & 0 & -\frac{i}{\sqrt{2}}
\end{array}
\right)
\end{equation}
a real element of the symplectic group (or algebra) can be mapped into a matrix that is simultaneously symplectic and pseudounitary:
\begin{equation}\label{grillo}
\mathcal{S} \, = \, \mathcal{C}^\dagger \, \, \left( \begin{array}{c|c}
{A} & {B} \\
\hline
{C} & {D}
\end{array}
\right) \,\mathcal{C} \, = \, \left(\begin{array}{cc}
U_0 & U_1^\star \\
U_1 & U_0 ^\star
\end{array}
\right) \,\, \in \, \mathop{\rm {}Sp}(6,\mathbb{C}) \bigcap \mathrm{SU(3,3)}
\end{equation}
The diagonal blocks $U_0 \in \mathrm{U(3)}$ span the $\mathrm{H}$-subgroup of the coset (\ref{gurto}). This allows to introduce a set projective coordinates that parameterize the points of the manifold (\ref{gurto}) and have a nice fractional linear transformation under the action of the group $\mathop{\rm {}Sp}(6,\mathbb{R})$. Given any coset parameterization
\begin{equation}\label{cosetpar}
\left( \begin{array}{c|c}
{A}(\phi) & {B}(\phi) \\
\hline
{C}(\phi) & {D}(\phi)
\end{array}
\right) \, \in \, \mathop{\rm {}Sp}(6,\mathbb{R})
\end{equation}
namely a family of symplectic group elements depending on 12 parameters $\phi^i$ such that each different choice of the $\phi^i$ provides a representative of a different equivalence class in (\ref{gurto}), we can construct the following, \textit{symmetric complex matrix}:
\begin{equation}\label{Zmatra}
Z(\phi) \, \equiv \, \left(A(\phi) \, - \, {\rm i} \,B(\phi) \right) \, \left( C(\phi) \, - \, {\rm i} D(\phi) \right)^{-1}
\end{equation}
which has a very simple transformation under the action of the symplectic group. Let us consider the action of any element of $\mathop{\rm {}Sp}(6,\mathbb{R})$ on the coset representative. We have:
\begin{equation}\label{canalito}
\underbrace{\left( \begin{array}{c|c}
{\hat{A}} & {\hat{B}} \\
\hline
{\hat{C}} & {\hat{D}}
\end{array}
\right)}_{= \mathfrak{g}\in \,\mathop{\rm {}Sp}(6,\mathbb{R})} \left( \begin{array}{c|c}
{A}(\phi) & {B}(\phi) \\
\hline
{C}(\phi) & {D}(\phi)
\end{array}
\right) \, = \, \left( \begin{array}{c|c}
{A}(\phi^\prime) & {B}(\phi^\prime) \\
\hline
{C}(\phi^\prime) & {D}(\phi^\prime)
\end{array}
\right) \, \mathrm{H}(\phi, \mathfrak{g})
\end{equation}
where $\phi^\prime$ is the label of a new equivalence class and $\mathrm{H}(\phi, \mathfrak{g}) \, \in \, \mathrm{U(3)}$ is a suitable $\mathrm{H}$-compensator. Calculating the matrix $Z(\phi^\prime)$ according to the definition (\ref{Zmatra}) we find that it is related to $Z(\phi)$ by a simple linear fractional transformation (generalized to matrices):
\begin{equation}\label{gourmet}
Z(\phi^\prime) \, = \, \left( A Z(\phi) \, + \, B\right) \, \left( C Z(\phi) \, + \, D\right)^{-1}
\end{equation}
Formula (\ref{gourmet}) is of crucial relevance and requires several comments. From a mathematical point of view, (\ref{gourmet}) is the well known generalization of the action of the $\mathrm{SL(2,\mathbb{R})}\, \simeq \, \mathop{\rm {}Sp}(2,\mathbb{R})$ group on the upper complex plane of Poincar\'e-Lobachevsky. The complex numbers $z$ with positive imaginary parts (${\rm Im} z > 0$) are replaced by the complex symmetric matrices $Z_{ij}$ whose imaginary part is positive definite. Such matrices constitute the so named \textbf{upper Siegel plane}, which indeed is homeomorphic to the coset $\mathop{\rm {}Sp}(2n,\mathbb{R})/\mathrm{U(n)}$. From the physical point of view (\ref{gourmet}) is just identical to the Gaillard-Zumino formula for the construction of the kinetic matrix $\mathcal{N}_{\Lambda\Sigma}$ which appears in the lagrangian of the vector fields in $\mathcal{N}=2$ supergravity and is rooted in the structure of special K\"ahler geometry. Indeed for any special K\"ahler manifold $\mathcal{M}_{n}$ of complex dimension $n$ that is also a symmetric space $\mathrm{G/H}$, there exists a so named $\mathbf{W}$-representation of $\mathrm{G}$, which is symplectic, has dimension $2n+2$ and hosts the electric and magnetic field strengths of the model. Such a representation defines a symplectic embedding:
\begin{equation}\label{gongolini}
\mathrm{G} \, \rightarrow \, \mathop{\rm {}Sp}\left( 2n+2,\mathbb{R}\right )
\end{equation}
which associates to any coset representative $\mathfrak{g}(\phi) \in \mathrm{G/H}$ its corresponding symplectic $(2n+2) \times (2n +2)$
representation $\left( \begin{array}{c|c}
{A}(\phi) & {B}(\phi) \\
\hline
{C}(\phi) & {D}(\phi)
\end{array}
\right)$. From this latter, utilizing the recipe provided by formula (\ref{gourmet}) we obtain an $(n+1) \times (n+1)$ complex symmetric matrix to be identified with the appropriate $\mathcal{N}$ kinetic matrix largely discussed and utilized in section \ref{scrittaN}.
\par
The peculiarity of the $\mathcal{N}=2$ model under investigation is that the original isometry group $\mathrm{G}$ is already symplectic so that we can utilize the Gaillard-Zumino formula (\ref{gourmet}) in the fundamental $6$ dimensional representation in order to construct a Siegel parametrization of the coset in terms of a symmetric complex $ 3 \times 3$ matrix $Z$. The $\mathbf{W}$-representation is the $\mathbf{14}^\prime$ and this defines the embedding:
\begin{equation}\label{gallettoalladiavola}
\mathop{\rm {}Sp}(6,\mathbb{R}) \mapsto \mathop{\rm {}Sp}(14,\mathbb{R})
\end{equation}
from which we can construct the $ 7\times 7 $ kinetic matrix $\mathcal{N}(Z)$.
\paragraph{\sc The transitive action of $\mathop{\rm {}Sp}(6,\mathbb{R})$ on the upper Siegel plane.} Before proceeding with the actual construction of the Lie algebra let us comment on the transitive action of the symplectic group on the Siegel plane. Focusing on the the formula (\ref{gourmet}), consider the $\mathop{\rm {}Sp}(6,\mathbb{R})$ parabolic subgroup composed by the following matrices:
\begin{equation}\label{caripollo}
\mathfrak{g}(B) \, = \, \left(\begin{array}{c|c}
\mathbf{1}_{3\times 3} & B \\
\hline
\mathbf{0}_{3\times 3} & \mathbf{1}_{3\times 3}
\end{array}
\right)
\end{equation}
where $B$ is symmetric and real. By means of such a subgroup we can always map a generic $Z$ matric into one that has vanishing real part $\mbox{Re} Z \, = \,0$.
Next consider the action on the residual imaginary part of $Z$ of the $\mathrm{GL(3,\mathbb{R})} \subset \mathop{\rm {}Sp}(6,\mathbb{R})$ subgroup
composed by the matrices:
\begin{equation}\label{caripollo2}
\mathfrak{g}(B) \, = \, \left(\begin{array}{c|c}
\mathcal{A} & \mathbf{0}_{3\times 3} \\
\hline
\mathbf{0}_{3\times 3} & \left(\mathcal{A}^T\right) ^{-1}
\end{array}
\right) \quad ; \quad \mathcal{A} \, \in \, \mathrm{GL(3,\mathbb{R})}
\end{equation}
We obtain:
\begin{equation}\label{struzzo}
\mbox{Im} Z \, \mapsto \, \mathcal{A} \, \mbox{Im} Z \, \mathcal{A}^T
\end{equation}
Choosing $\mathcal{A} \, = \, (\mbox{Im} Z)^{\frac 12}$, which is always possible since $\mbox{Im} Z$ is positive definite we can reduce the imaginary part to the identity matrix. This shows the transitive action of the symplectic group on the Siegel plane and also provides a nice coset parameterization of the coset manifold. Indeed we can introduce the following matrix:
\begin{equation}\label{godereccio}
\mathfrak{g}(Z) \, \equiv \, \left (\begin{array}{c|c}
\left(\mbox{Im} Z\right)^{\frac 12} &\mbox{Re} Z \, \left(\mbox{Im} Z\right)^{-\frac 12} \\
\hline
\mathbf{0} & \left(\mbox{Im} Z\right)^{-\frac 12}
\end{array}
\right)
\end{equation}
which maps the origin of the manifold ${\rm i} \mathbf{1}_{3\times 3}$ in the complex symmetric matrix $Z$.
\subsection{\sc The $\mathfrak{sp}(6,\mathbb{R})$ Lie algebra}
From the point of view of the Dynkin classification the Lie algebra $\mathfrak{sp}(6,\mathbb{R})$ is the maximally split real section of the complex Lie algebra $C_3$ whose Dynkin diagram is displayed in fig.\ref{C3dynk}.
\begin{figure}
\centering
\begin{picture}(90,50)
\put (-70,35){$C_3$} \put
(10,35){\circle {10}} \put (7,20){$\alpha_1$} \put (15,35){\line
(1,0){20}} \put (40,35){\circle {10}} \put (37,20){$\alpha_2$}\put (45,38){\line (1,0){20}}
\put (55,35){\line (1,1){10}} \put (55,35){\line (1,-1){10}}\put (45,33){\line (1,0){20}} \put
(70,35){\circle {10}} \put (67,20){$\alpha_{3}$}
\end{picture}
\caption{The Dynkin diagram of $C_{3}$. \label{C3dynk}}
\end{figure}
The root system is composed of $18$-roots whose subset of $9$ positive ones is displayed here below:
\begin{equation}\label{cornette}
\left[\begin{array}{c}
\alpha_1 \\
\alpha_2 \\
\alpha_3\\
\alpha_4 \\
\alpha_5 \\
\alpha_6 \\
\alpha_7 \\
\alpha_8 \\
\alpha_9
\end{array}
\right] \, = \, \left[
\begin{array}{ll}
\alpha _1 & \{1,-1,0\} \\
\alpha _2 & \{0,1,-1\} \\
\alpha _3 & \{0,0,2\} \\
\alpha _1+\alpha _2 &
\{1,0,-1\} \\
\alpha _2+\alpha _3 &
\{0,1,1\} \\
\alpha _1+\alpha _2+\alpha _3
& \{1,0,1\} \\
2 \alpha _2+\alpha _3 &
\{0,2,0\} \\
\alpha _1+2 \alpha _2+\alpha
_3 & \{1,1,0\} \\
2 \alpha _1+2 \alpha
_2+\alpha _3 & \{2,0,0\}
\end{array}
\right]
\end{equation}
The simple roots are the first three. Of the remaining $6$ we have provided both their expression in terms of the simple roots and their realization as three-vectors in $\mathbb{R}^3$. Such a realization is spelled out also for the simple roots.
Next we present the basis of $6\times6$ matrices that fulfill the standard commutation relations of the Lie Algebra in the Cartan Weyl basis.
\paragraph{Cartan Generators.} The Cartan generators are named $\mathcal{H}^i$ and can be easily read-off from the following formula:
\begin{equation}\label{cartolini}
\sum_{i=1}^{3} \, h_i \, \mathcal{H}^i \, = \,\left(
\begin{array}{llllll}
h_1 & 0 & 0 & 0 & 0 & 0 \\
0 & h_2 & 0 & 0 & 0 & 0 \\
0 & 0 & h_3 & 0 & 0 & 0 \\
0 & 0 & 0 & -h_1 & 0 & 0 \\
0 & 0 & 0 & 0 & -h_2 & 0 \\
0 & 0 & 0 & 0 & 0 & -h_3
\end{array}
\right)
\end{equation}
by collecting the coefficient of the parameter $h_i$.
\paragraph{\sc Positive Root Step Operators.} The step operator associated with the positive root $\alpha_i$ is named $\mathcal{E}^{\alpha_i}$ and can be easily read-off from the following formula:
\begin{equation}\label{steppinisu}
\sum_{i=1}^{9} \, a_i \, \mathcal{E}^{\alpha_i} \, = \,\left(
\begin{array}{llllll}
0 & a_1 & a_4 & \sqrt{2} a_9
& a_8 & a_6 \\
0 & 0 & a_2 & a_8 & \sqrt{2}
a_7 & a_5 \\
0 & 0 & 0 & a_6 & a_5 &
\sqrt{2} a_3 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & -a_1 & 0 & 0 \\
0 & 0 & 0 & -a_4 & -a_2 & 0
\end{array}
\right)
\end{equation}
by collecting the coefficient of the parameter $a_i$.
\paragraph{\sc Negative Root Step Operators.} The step operator associated with the negative root $-\alpha_i$ is named $\mathcal{E}^{-\alpha_i}$ and can be easily read-off from the following formula:
\begin{equation}\label{steppinigiu}
\sum_{i=1}^{9} \, b_i \, \mathcal{E}^{-\alpha_i} \, = \,\left(
\begin{array}{llllll}
0 & 0 & 0 & 0 & 0 & 0 \\
b_1 & 0 & 0 & 0 & 0 & 0 \\
b_4 & b_2 & 0 & 0 & 0 & 0 \\
\sqrt{2} b_9 & b_8 & b_6 & 0
& -b_1 & -b_4 \\
b_8 & \sqrt{2} b_7 & b_5 & 0
& 0 & -b_2 \\
b_6 & b_5 & \sqrt{2} b_3 & 0
& 0 & 0
\end{array}
\right)
\end{equation}
by collecting the coefficient of the parameter $b_i$.
\subsection{\sc The representation $14^\prime$}
\label{seziona14}
The $\mathbf{14}^\prime$ representation of $\mathfrak{sp}(6,\mathbb{R})$ which plays the role $\mathbf{W}$-representation for the special manifold under consideration is defined as the representation obeyed by the three-times antisymmetric tensors with vanishing $\mathbb{C}$-traces, namely:
\begin{equation}\label{goliutico}
\underbrace{t_{ABC}}_{\mbox{antisymmetric in} A,B,C} \times \quad \mathbb{C}^{BC} \, = \,0
\end{equation}
The generators are constructed in the appendix, subsection \ref{seziona14}, and displayed in eq.s (\ref{cartolini14}), (\ref{steppinisu14}) and (\ref{steppinigiu14}).
\subsection{\sc The holomorphic symplectic section and its transformation in the $14^\prime$}
\label{holoseziona}
In order to construct the special geometry of the manifold (\ref{gurto}) we need to introduce the holomorphic symplectic section that, by definition, should transform in the $14^\prime$ representation of $\mathop{\rm {}Sp}(6,\mathbb{R})$. To this effect, we choose as special coordinates the components of the symmetric complex matrix defined by eq.(\ref{gourmet}) and we choose a lexicographic order to enumerate its independent components, namely we set:
\begin{equation}\label{specoord}
Z \, = \, \left(
\begin{array}{lll}
z_1 & z_2 & z_3 \\
z_2 & z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right)
\end{equation}
Next we introduce the holomorphic prepotential defined by:
\begin{eqnarray}
\mathcal{F} & \equiv & Z_{a,i} \, Z_{b,j} \, Z_{c,k} \, \epsilon^{abc} \,\epsilon^{ijk}\nonumber \\
\null &=& -6 \left(z_6 z_2^2-2 z_3 z_5
z_2+z_3^2 z_4+z_1
\left(z_5^2-z_4
z_6\right)\right)
\end{eqnarray}
and we can introduce a first ansatz for the symplectic section by writing:
\begin{eqnarray}
\widetilde{\Omega} &=& \left \{1, \, z^I , \, \mathcal{F} , \, \frac{\partial \mathcal{F} }{\partial z^J} \right\} \nonumber \\
\null &=& \left\{1,z_1,z_2,z_3,z_4,z_5,z
_6,-6 \left(z_6 z_2^2-2 z_3
z_5 z_2+z_3^2 z_4+z_1
\left(z_5^2-z_4
z_6\right)\right), \right.\nonumber\\
&&\left.6 z_4
z_6-6 z_5^2,12 \left(z_3
z_5-z_2 z_6\right),12
\left(z_2 z_5-z_3
z_4\right),6 z_1 z_6-6
z_3^2,12 \left(z_2 z_3-z_1
z_5\right),6 z_1 z_4-6
z_2^2\right\}\nonumber\\
\label{preomega}
\end{eqnarray}
In order to match the transformation of this holomorphic section with the transformations of the $\mathbf{14}^\prime$ representation as we defined it in subsection \ref{holoseziona} we still need a change of basis. Consider the following matrix
{\scriptsize
\begin{equation}\label{paola}
\mathfrak{S}\, = \, \left(
\begin{array}{llllllllllllll}
0 & 0 & 0 & 0 & 0 & 0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \frac{1}{3 \sqrt{2}} & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\frac{1}{3 \sqrt{2}} & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & \frac{1}{3 \sqrt{2}} & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \frac{1}{6} & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\frac{1}{6} & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \frac{1}{6} & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \frac{1}{3 \sqrt{2}} \\
0 & 0 & 0 & 0 & -\sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & -2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & -2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0
\end{array}
\right)
\end{equation}
}
and define:
\begin{equation}\label{gargamelle}
\Omega\left( Z\right) \, = \, \mathfrak{S}\, \widetilde{\Omega}(Z) \, = \, \left(
\begin{array}{l}
\sqrt{2} z_6 \\
\sqrt{2} \left(z_1 z_6-z_3^2\right) \\
\sqrt{2} \left(z_5^2-z_4 z_6\right) \\
-\sqrt{2} \left(z_6 z_2^2-2 z_3 z_5 z_2+z_3^2 z_4+z_1 \left(z_5^2-z_4 z_6\right)\right) \\
2 z_2 z_3-2 z_1 z_5 \\
2 z_3 z_4-2 z_2 z_5 \\
2 z_3 z_5-2 z_2 z_6 \\
\sqrt{2} \left(z_1 z_4-z_2^2\right) \\
-\sqrt{2} z_4 \\
\sqrt{2} z_1 \\
\sqrt{2} \\
-2 z_5 \\
2 z_3 \\
-2 z_2
\end{array}
\right)
\end{equation}
Naming $\mathcal{D}_{14}\left[g\right]$ the 14-dimensional representation of a finite element $g , \in \, \mathop{\rm {}Sp}(6,\mathbb{R})$ of the symplectic group that corresponds to the representation of the algebra as we constructed it above, the holomorphic symplectic section (\ref{gargamelle}) transforms in the following way:
\begin{equation}\label{buonotrasformo}
\Omega \left[ (A\,Z \, + \, B) \, (C \, Z \, + \, D )^{-1} \right] \, = \, \frac{1}{\mbox{Det}\left(C \, Z \, + \, D\right)} \, \mathcal{D}_{14}\left[\left(\begin{array}{c|c}
A & B \\
\hline
C & D
\end{array}\right)
\right] \, \Omega[Z]
\end{equation}
The formula (\ref{buonotrasformo}) can be in particular applied to the case where the original $Z$ is the origin of the coset manifold:
$Z_0 \, = \, {\rm i} \, \mathbf{1}_{3\times 3}$. In that case, recalling eq. (\ref{godereccio}) we find:
\begin{equation}\label{frittellaprima}
\Omega[Z_0] \, = \, \left\{i \sqrt{2},-\sqrt{2},\sqrt{2},-i
\sqrt{2},0,0,0,-\sqrt{2},-i \sqrt{2},i
\sqrt{2},\sqrt{2},0,0,0\right\}
\end{equation}
and
\begin{equation}\label{caviccio}
\Omega[Z_0] \, = \, \sqrt{\mbox{Det}\left[\mbox{Im} \, Z\right] } \, \times \, \mathcal{D}_{14}\left[ \left (\begin{array}{c|c}
\left(\mbox{Im} Z\right)^{\frac 12} &\mbox{Re} Z \, \left(\mbox{Im} Z\right)^{-\frac 12} \\
\hline
\mathbf{0} & \left(\mbox{Im} Z\right)^{-\frac 12}
\end{array}
\right)\right] \, \cdot \, \Omega[Z_0]
\end{equation}
\subsection{\sc The K\"ahler potential and the metric}
Provided with this information we can now write the explicit form of the K\"ahler potential and of the K\"ahler metric for the manifold (\ref{gurto}) according to the rules of special K\"ahler geometry. We have:
\begin{eqnarray}
\mathcal{K} & \equiv & - \log \left( {\rm i} \Omega[Z] \, \mathbb{C}_{14} \, \overline{\Omega}[\bar{Z}]\right) \nonumber\\
\null &=& -\log \left(2 i \left(-z_6
z_2^2+{\bar z}_6 z_2^2+2 z_6
{\bar z}_2 z_2-2 z_5 {\bar z}_3
z_2+2 {\bar z}_3 {\bar z}_5 z_2-2
{\bar z}_2 {\bar z}_6 z_2-z_6
{\bar z}_2^2-z_4
{\bar z}_3^2+{\bar z}_1
{\bar z}_5^2+\right.\right. \nonumber\\
&&\left.\left. z_5^2 {\bar z}_1-z_4
z_6 {\bar z}_1+2 z_5 {\bar z}_2
{\bar z}_3+{\bar z}_3^2
{\bar z}_4+z_6 {\bar z}_1
{\bar z}_4 +z_3^2
\left({\bar z}_4-z_4\right)-2 z_5
{\bar z}_1 {\bar z}_5-2
{\bar z}_2 {\bar z}_3
{\bar z}_5 \right.\right.\nonumber\\
&&\left.\left.+2 z_3 \left(-z_5
{\bar z}_2+{\bar z}_5
{\bar z}_2+z_4
{\bar z}_3-{\bar z}_3
{\bar z}_4+z_2
\left(z_5-{\bar z}_5\right)\right)+
{\bar z}_2^2 {\bar z}_6+z_4
{\bar z}_1 {\bar z}_6-{\bar z}_1
{\bar z}_4 {\bar z}_6 \right.\right.\nonumber\\
&&\left.\left. -z_1
\left(z_5^2-2 {\bar z}_5
z_5+{\bar z}_5^2+z_6
{\bar z}_4-{\bar z}_4
{\bar z}_6+z_4
\left({\bar z}_6-z_6\right)\right)\right)\right) \nonumber\\
\end{eqnarray}
and the line element on the manifold, in terms of the special coordinates $z_i$ takes the standard form:
\begin{equation}\label{gumilevo}
ds_K^2 \, = \, \frac{\partial}{\partial z^i} \, \frac{\partial}{\partial \bar{z}^{j}} \, \mathcal{K} \, dz^i \otimes d\bar{z}^j
\end{equation}
The explicit form of $ds_K^2$ in terms of the special coordinate $z^i$ can be worked out by simple derivatives, yet its explicit form is quite lengthy and so much involved that we think it better not to display it. For the purposes that we pursue we rather prefer to write the form of the metric in terms of solvable real coordinates.
\subsubsection{\sc The solvable parametrization}
The transition to a solvable parametrization of the coset is rather simple. Let us define the solvable coset representative as the product of the exponentials of all the generators of the Borel subalgebra of $\mathfrak{sp}(6,\mathbb{R})$:
\begin{eqnarray}
& \mathbb{L}(h,p)\, = \, \prod_{i=1}^9 \, \exp\left [ p_{10-i} \, \mathcal{E}^{\alpha_{10-i}}\right ] \, \prod_{j=3}^3 \exp\left [ h_{j} \, \mathcal{H}^{j}\right ] \, = \, & \nonumber\\
& \mbox{\scriptsize $\left(
\begin{array}{l|l|l|l|l|l}
e^{h_1} & e^{h_2} p_1 & e^{h_3} p_4 &
\begin{array}{l}
e^{-h_1} \left(\sqrt{2} p_1 p_2 p_3
p_4 \right.\\
\left. +\left(p_1 p_2-p_4\right) p_6\right.\\
\left.-p_1
p_8+\sqrt{2} p_9 \right)
\end{array}
& \begin{array}{c}
e^{-h_2}
\left(-\sqrt{2} p_2 p_3 p_4\right. \\
\left. -p_2
p_6+p_8\right)
\end{array}
& e^{-h_3}
\left(\sqrt{2} p_3 p_4+p_6\right) \\
\hline
0 & e^{h_2} & e^{h_3} p_2 &
\begin{array}{c}
e^{-h_1}
\left(\left(p_1 p_2-p_4\right)
p_5\right. \\
\left.-\sqrt{2} p_1 p_7+p_8\right)
\end{array}
&
\begin{array}{c}
e^{-h_2} \left(\sqrt{2} p_7 \right. \\
\left. -p_2
p_5\right)
\end{array}
& e^{-h_3} p_5
\\
\hline
0 & 0 & e^{h_3} & e^{-h_1}
\left(\sqrt{2} p_1 p_2 p_3-p_1
p_5+p_6\right) & e^{-h_2}
\left(p_5-\sqrt{2} p_2 p_3\right) &
\sqrt{2} e^{-h_3} p_3 \\
\hline
0 & 0 & 0 & e^{-h_1} & 0 & 0 \\
\hline
0 & 0 & 0 & -e^{-h_1} p_1 & e^{-h_2} &
0 \\
\hline
0 & 0 & 0 & e^{-h_1} \left(p_1
p_2-p_4\right) & -e^{-h_2} p_2 &
e^{-h_3}
\end{array}
\right) $}& \nonumber\\
\label{baltico}
\end{eqnarray}
The real coordinates of the manifold are now the $12$ parameters:
\begin{equation}\label{coordine}
\mbox{coordinates} \, \equiv \, \left\{ h_1, \dots , h_3, \, p_1 , \, \dots \, p_9 \right \}
\end{equation}
Extracting the complex matrix $Z$ from the symplectic matrix $\mathbb{L}(h,p)$ we find:
\begin{eqnarray}\label{zph}
& Z(h,p) \, = \, & \nonumber\\
& \mbox{\scriptsize $\left(
\begin{array}{lll}
i e^{2 h_2} p_1^2+i e^{2
h_1}+\left(\sqrt{2} p_3+i e^{2
h_3}\right) p_4^2+\sqrt{2} p_9 & i
e^{2 h_2} p_1+i e^{2 h_3} p_2
p_4+p_8 & \left(\sqrt{2} p_3+i e^{2
h_3}\right) p_4+p_6 \\
i e^{2 h_2} p_1+i e^{2 h_3} p_2
p_4+p_8 & i e^{2 h_3} p_2^2+i e^{2
h_2}+\sqrt{2} p_7 & i e^{2 h_3}
p_2+p_5 \\
\left(\sqrt{2} p_3+i e^{2 h_3}\right)
p_4+p_6 & i e^{2 h_3} p_2+p_5 &
\sqrt{2} p_3+i e^{2 h_3}
\end{array}
\right)$} &
\end{eqnarray}
which defines the coordinate transformation from the special to the solvable coordinates:
\begin{equation}\label{carlinus}
\left(
\begin{array}{l}
z_1 \\
z_2 \\
z_3 \\
z_4 \\
z_5 \\
z_6
\end{array}
\right)\, = \, \left(
\begin{array}{l}
\sqrt{2} p_3 p_4^2+i \left(e^{2 h_2}
p_1^2+e^{2 h_1}+e^{2 h_3}
p_4^2\right)+\sqrt{2} p_9 \\
i \left(e^{2 h_2} p_1+e^{2 h_3} p_2
p_4\right)+p_8 \\
i e^{2 h_3} p_4+\sqrt{2} p_3 p_4+p_6
\\
i \left(e^{2 h_3} p_2^2+e^{2
h_2}\right)+\sqrt{2} p_7 \\
i e^{2 h_3} p_2+p_5 \\
\sqrt{2} p_3+i e^{2 h_3}
\end{array}
\right)
\end{equation}
Inserting such a coordinate transformation into the K\"ahler metric (\ref{gumilevo}) we obtain its form in terms of the real coordinates
(\ref{coordine}). For the explicit form of the metric, we refer the reader to the appendix, eq. (\ref{formidabile}).
The complete metric is quite formidable (\ref{formidabile}) since it contains a total of 100 terms. It has however quite simple properties when we sit in the neighborhood of the coset origin, in particular at $p_i \sim 0$. In this case it drastically simplifies and becomes diagonal:
\begin{eqnarray}\label{pippa}
&&ds_K^2 \, \stackrel{p_i \rightarrow 0}{\Longrightarrow} \, \mathrm{dh}_1^2+\mathrm{dh}_2^2+\mathrm{dh}_3
^2+\frac{1}{2} e^{2 h_2-2 h_1}
\mathrm{dp}_1^2+\frac{1}{2} e^{2 h_3-2
h_2} \mathrm{dp}_2^2+\frac{1}{2} e^{-4
h_3} \mathrm{dp}_3^2+\frac{1}{2} e^{2
h_3-2 h_1} \mathrm{dp}_4^2\nonumber\\
&&+\frac{1}{2}
e^{-2 h_2-2 h_3}
\mathrm{dp}_5^2+\frac{1}{2} e^{-2
h_1-2 h_3} \mathrm{dp}_6^2+\frac{1}{2}
e^{-4 h_2} \mathrm{dp}_7^2+\frac{1}{2}
e^{-2 h_1-2 h_2}
\mathrm{dp}_8^2+\frac{1}{2} e^{-4 h_1}
\mathrm{dp}_9^2
\end{eqnarray}
which shows that it is positive definite as it should be. It is also interesting to note that if the truncation to the Cartan is permitted by the potential, then we just have three dilatons with canonical kinetic terms.
\subsection{\sc The quartic invariant in the $14^\prime$}
Of crucial relevance for the analysis of Black Hole charges and in general for the classification of orbits in the $\mathbf{W}$-representation is the quartic symplectic invariant. Given a $14$-vector
\begin{equation}\label{gongolando}
\mathcal{Q} \, = \, \left\{ q_1,\, q_2 \, \dots ,\,q_{14}\right\}
\end{equation}
the standard form of this invariant can be expressed in the following manifestly ${\rm Sp}(6,\mathbb{R})$-invariant form (see for instance \cite{Ferrara:2013zga})
\begin{eqnarray}
\mathfrak{J}_4(\mathcal{Q}) &=& -\frac{n_V(2n_V+1)}{6d}\,(\Lambda_a)_{\alpha\beta}\,(\Lambda^a)_{\gamma\delta}\,\mathcal{Q}^\alpha\,\mathcal{Q}^\beta\,\mathcal{Q}^\gamma\,\mathcal{Q}^\delta\,,
\end{eqnarray}
where in our case $n_V=7$ and $d={\rm dim}{\rm Sp}(6,\mathbb{R})=21$, the symplectic indices are raised and lowered by $\mathbb{C}_14^{\alpha\beta}$ and $\mathbb{C}_{14\,\alpha\beta}$ and the index $a$ is raised by the inverse of $\eta_{ab}\equiv
{\rm Tr}(\Lambda_a\,\Lambda_b)$.
The explicit form of $\mathfrak{J}_4(\mathcal{Q})$ reads:
\begin{eqnarray}
\mathfrak{J}_4(\mathcal{Q}) &=& -2 q_1 q_9 q_5^2+2 q_3 q_{11} q_5^2-2 \sqrt{2} q_6 q_7 q_{11} q_5-2 q_1 q_8 q_{12}
q_5+2 q_2 q_9 q_{12} q_5-2 q_3 q_{10} q_{12} q_5\nonumber\\
&&+2 q_4 q_{11} q_{12} q_5-2
\sqrt{2} q_7 q_9 q_{13} q_5 +2 \sqrt{2} q_1 q_6 q_{14} q_5+2 \sqrt{2} q_3 q_{13}
q_{14} q_5+q_1^2 q_8^2\nonumber\\
&&+q_2^2 q_9^2+q_3^2 q_{10}^2+q_4^2 q_{11}^2+2 q_2 q_8
q_{12}^2-2 q_4 q_{10} q_{12}^2-2 q_3 q_8 q_{13}^2+2 q_4 q_9 q_{13}^2-2 q_2 q_3
q_{14}^2\nonumber\\
&&-2 q_1 q_4 q_{14}^2+2 q_1 q_2 q_8 q_9+2 q_1 q_6^2 q_{10}+2 q_1 q_3 q_8
q_{10}-2 q_7^2 q_9 q_{10}-2 q_2 q_3 q_9 q_{10}\nonumber\\
&&-4 q_1 q_4 q_9 q_{10}-2 q_2 q_6^2
q_{11}-2 q_7^2 q_8 q_{11}-4 q_2 q_3 q_8 q_{11}\nonumber\\
&&-2 q_1 q_4 q_8 q_{11}+2 q_2 q_4
q_9 q_{11}+2 q_3 q_4 q_{10} q_{11}\nonumber\\
&&+2 \sqrt{2} q_6 q_7 q_{10} q_{12}-2 q_1 q_6
q_8 q_{13}-2 q_2 q_6 q_9 q_{13}+2 q_3 q_6 q_{10} q_{13}\nonumber\\
&&+2 q_4 q_6 q_{11}
q_{13}-2 \sqrt{2} q_7 q_8 q_{12} q_{13}+2 q_1 q_7 q_8 q_{14}\nonumber\\
&&+2 q_2 q_7 q_9
q_{14}+2 q_3 q_7 q_{10} q_{14}+2 q_4 q_7 q_{11} q_{14}-2 \sqrt{2} q_2 q_6 q_{12}
q_{14}+2 \sqrt{2} q_4 q_{12} q_{13} q_{14}\nonumber\\
\label{I4inv}
\end{eqnarray}
\subsection{\sc Truncation to the $STU$-model}
\label{truncazia}
Next we analyze how the $STU$-model is embedded into the $\mathop{\rm {}Sp}(6,\mathbb{R})$-model.
At the level of the special coordinates the truncation to the $STU$-model is very simply done. It suffices to
set to zero the complex coordinates $z_2,z_3,z_5$ keeping only $z_1,z_4,z_6$ that can be identified with the fields $S,T,U$. When we do so the symplectic section reduces as follows:
\begin{equation}\label{stuSymsec}
\Omega\left[Z \left(
\begin{array}{ccc}
z_1 & 0 & 0 \\
0 & z_4 & 0 \\
0 & 0 & z_6 \\
\end{array}
\right)
\right] \, = \, \left(
\begin{array}{l}
\sqrt{2} z_6 \\
\sqrt{2} z_1 z_6 \\
-\sqrt{2} z_4 z_6 \\
\sqrt{2} z_1 z_4 z_6 \\
0 \\
0 \\
0 \\
\sqrt{2} z_1 z_4 \\
-\sqrt{2} z_4 \\
\sqrt{2} z_1 \\
\sqrt{2} \\
0 \\
0 \\
0
\end{array}
\right)
\end{equation}
and the K\"ahler potential reduces to:
\begin{equation}\label{stuKalpot}
\mathcal{K} \, \rightarrow\, -\log \left[2 i \left(z_1-{\bar z}_1\right) \left(z_4-{\bar z}_4\right)
\left(z_6-{\bar z}_6\right)\right]
\end{equation}
which yields three copies of the Poincar\'e metric, one for each of the three $\frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}$
submanifolds.
\par
The result (\ref{stuSymsec}) is in agreement with the decomposition of the $\mathbf{14}^\prime$ of $\mathfrak{sp}(6,\mathbb{R})$ with respect to the three subalgebras $\mathfrak{sl}(2)$:
\begin{equation}\label{decumpo}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2) \times \mathfrak{sl}(2) \times \mathfrak{sl}(2)}{\Longrightarrow} \, \left(\mathbf{2,2,2}\right)
\oplus \left(\mathbf{2,1,1}\right) \oplus \left(\mathbf{1,2,1}\right) \oplus \left(\mathbf{1,1,2}\right)
\end{equation}
From (\ref{stuSymsec}) we also learn that the directions $\{1,2,3,4,8,9,10,11\}$ of the $\mathbf{14}^\prime$ vector space span the representation $\left(\mathbf{2,2,2}\right)$, while the directions $\{5,6,7,12,13,14\}$ of the same space span the representations $\left(\mathbf{2,1,1}\right) \oplus \left(\mathbf{1,2,1}\right) \oplus \left(\mathbf{1,1,2}\right)$.
The adjoint representation of $\mathfrak{sp}(6,\mathbb{R})$ decomposes instead in the following way:
\begin{eqnarray}
\mbox{adj} \,\left[\mathfrak{sp}(6,\mathbb{R})\right] & \stackrel{\mathfrak{sl}(2) \times \mathfrak{sl}(2) \times \mathfrak{sl}(2)}{\Longrightarrow} &
\left(\mathbf{3,1,1}\right) \oplus \left(\mathbf{1,3,1}\right) \oplus \left(\mathbf{1,1,3}\right) \nonumber\\
&&
\oplus \left(\mathbf{2,2,1}\right) \oplus \left(\mathbf{2,1,2}\right) \oplus \left(\mathbf{1,2,2}\right) \label{adjointo}
\end{eqnarray}
as it is evident by a quick inspection of the roots (\ref{cornette}). In terms of the Cartan-Weyl basis the three
$\mathrm{\mathfrak{sl}(2,\mathbb{R})}$ subalgebra contains the three Cartan generators $\mathcal{H}_i$ and the step operators
$\mathcal{E}^{\pm \alpha_3}$, $\mathcal{E}^{\pm \alpha_7}$ , $\mathcal{E}^{\pm \alpha_9}$. The remaining 12 step operators
span the representation $\left(\mathbf{2,2,1}\right) \oplus \left(\mathbf{2,1,2}\right) \oplus \left(\mathbf{1,2,2}\right)$, namely:
\begin{equation}\label{spannone}
\left(\mathbf{2,2,1}\right) \oplus \left(\mathbf{2,1,2}\right) \oplus \left(\mathbf{1,2,2}\right) \, = \, \mbox{span} \left[
\mathcal{E}^{\pm \alpha_1}, \mathcal{E}^{\pm \alpha_2} , \mathcal{E}^{\pm \alpha_4} ,\mathcal{E}^{\pm \alpha_5}, \mathcal{E}^{\pm \alpha_6} , \mathcal{E}^{\pm \alpha_8}
\right]
\end{equation}
The explicit form of an $\mathfrak{sp}(6,\mathbb{R})$ Lie algebra element reduced to the $\mathfrak{sl}(2)^3$ subalgebra is the following one:
\begin{equation}\label{pagnocco}
\left(
\begin{array}{llllll}
h_1 & 0 & 0 & b_1 & 0 & 0 \\
0 & h_2 & 0 & 0 & b_2 & 0 \\
0 & 0 & h_3 & 0 & 0 & b_3 \\
c_1 & 0 & 0 & -h_1 & 0 & 0 \\
0 & c_2 & 0 & 0 & -h_2 & 0 \\
0 & 0 & c_3 & 0 & 0 & -h_3
\end{array}
\right) \, \in \, \mathfrak{sl}(2) \otimes \mathfrak{sl}(2) \otimes \mathfrak{sl}(2) \, \subset \, \mathfrak{sp}(6,\mathbb{R})
\end{equation}
\subsection{\sc Reduction of the charge vector to the $(2,2,2)$}
In order to study the orbits of the charge vectors in the $\mathbf{14}^\prime$ our first step consists of reducing it to normal form, namely to the $(2,2,2)$ representation. We claim that for generic charge vectors this is always possible by means of $\mathop{\rm {}Sp}(6,\mathbb{R})$ rotations generated by elements of the $\left(\mathbf{2,2,1}\right) \oplus \left(\mathbf{2,1,2}\right) \oplus \left(\mathbf{1,2,2}\right)$ subspace. To show this let us consider the six dimensional compact Lie algebra element:
\begin{eqnarray}\label{cappotto}
\mathbb{K}_\psi & = & \psi_1 \,\left(\mathcal{E}^{\alpha_1} -\mathcal{E}^{-\alpha_1}\right) +\psi_2 \, \left(\mathcal{E}^{\alpha_2} -\mathcal{E}^{-\alpha_2}\right) +\psi_3 \, \left(\mathcal{E}^{\alpha_4} -\mathcal{E}^{-\alpha_4}\right)\nonumber\\
&&\psi_4 \,\left(\mathcal{E}^{\alpha_5} -\mathcal{E}^{-\alpha_5}\right) +\psi_5 \, \left(\mathcal{E}^{\alpha_6} -\mathcal{E}^{-\alpha_6}\right) +\psi_6 \, \left(\mathcal{E}^{\alpha_8} -\mathcal{E}^{-\alpha_8}\right)
\end{eqnarray}
and a generic charge vector that has components only in the $\mathbf{(2,2,2)}$ subspace.
\begin{equation}\label{toroidallo}
\mathcal{Q}_{2,2,2} \, = \, \left\{\Theta _1,\Theta _2,\Theta _3,\Theta _4,0,0,0,\Theta _5,\Theta _6,\Theta
_7,\Theta _8,0,0,0\right\}
\end{equation}
If we apply the $\mathbf{14}^\prime$ representation of $\mathbb{K}_\psi$ to the charge vector $\mathcal{Q}_N$ we obtain:
\begin{equation}\label{gorgonzola}
\mathcal{D}_{14}\left( \mathbb{K}_\psi\right) \,\mathcal{Q}_{2,2,2} \, = \, \left(
\begin{array}{l}
0 \\
0 \\
0 \\
0 \\
-\sqrt{2} \Theta _2 \psi _2+\sqrt{2} \Theta _5 \psi _2-\sqrt{2} \Theta _4 \psi
_4-\sqrt{2} \Theta _7 \psi _4 \\
-\sqrt{2} \Theta _3 \psi _3-\sqrt{2} \Theta _5 \psi _3+\sqrt{2} \Theta _4 \psi
_5-\sqrt{2} \Theta _6 \psi _5 \\
\sqrt{2} \Theta _2 \psi _1+\sqrt{2} \Theta _3 \psi _1-\sqrt{2} \Theta _1 \psi
_6-\sqrt{2} \Theta _4 \psi _6 \\
0 \\
0 \\
0 \\
0 \\
-\sqrt{2} \Theta _1 \psi _2-\sqrt{2} \Theta _6 \psi _2+\sqrt{2} \Theta _3 \psi
_4-\sqrt{2} \Theta _8 \psi _4 \\
\sqrt{2} \Theta _1 \psi _3-\sqrt{2} \Theta _7 \psi _3+\sqrt{2} \Theta _2 \psi
_5+\sqrt{2} \Theta _8 \psi _5 \\
\sqrt{2} \Theta _6 \psi _1+\sqrt{2} \Theta _7 \psi _1-\sqrt{2} \Theta _5 \psi
_6-\sqrt{2} \Theta _8 \psi _6
\end{array}
\right)
\end{equation}
which clearly shows that the six parameters $\psi_{1,\dots,6}$ are sufficient to generate arbitrary components $\{5,6,7,12,13,14\}$ of the charge vector starting from vanishing ones. Reverting the path this means that by means of the same rotations, apart from singular orbits that deserve a separate study we can always fix the gauge where the six components $\{5,6,7,12,13,14\}$ vanish.
\subsubsection{\sc Further reduction to normal form of the charge vector}
Once the charge vector is reduced to $(2,2,2)$ representation, we can further act on it with the $\mathrm{SL(2,\mathbb{R})}^3$ group in order to further reduce its components. By using the three parameters of the abelian translation group $\mathbb{R}^3$ contained in $\mathrm{SL(2,\mathbb{R})}^3$ we can put to zero three of the eight charges and a possible normal form of the charge vector is the following one:
\begin{equation}\label{normaldata}
\mathcal{Q}_{N} \, = \, \left\{0,P_1,P_2,P_3,0,0,0,P_4,0,0,P_5,0,0,0\right\}
\end{equation}
The corresponding quartic invariant is:
\begin{equation}\label{normaldata2}
\mathfrak{J}_4 \left(\mathcal{Q}_{N}\right)\, = \, P_3^2 P_5^2-4 P_1 P_2 P_4 P_5
\end{equation}
\section{\sc The $\frac{\mathrm{F_{(4,4)}}}{\mathrm{SU(2)} \times \mathrm{USp(6)}}$ quaternionic K\"ahler manifold}
Let us now come to the $c$-map image of the Special K\"ahler manifold (\ref{gurto}), namely to the quaternionic K\"ahler manifold (\ref{quatergurto}). The $F_{(4,4)}$ Lie algebra has rank four and its structure is codified in the Dynkin diagram presented in fig.\ref{F4dynk}.
\begin{figure}
\centering
\begin{picture}(90,50)
\put (-70,35){$F_4$} \put
(10,35){\circle {10}} \put (7,20){$\beta_4$} \put (15,35){\line
(1,0){20}} \put (40,35){\circle {10}} \put (37,20){$\beta_3$}\put (45,38){\line (1,0){20}}
\put (55,35){\line (1,1){10}} \put (55,35){\line (1,-1){10}}\put (45,33){\line (1,0){20}} \put
(70,35){\circle {10}} \put (67,20){$\beta_{2}$} \put
(75,35){\line (1,0){20}} \put (100,35){\circle {10}} \put (100,35){\circle {9}}\put (100,35){\circle {8}}\put (100,35){\circle {10}}\put (100,35){\circle {7}}
\put (100,35){\circle {6}}\put (100,35){\circle {5}}\put (100,35){\circle {4}}\put (100,35){\circle {3}}\put (100,35){\circle {2}}\put (100,35){\circle {1}}\put
(97,20){$\beta_{1}$}
\end{picture}
$$ \begin{array}{l}\psi \, = \, \beta_{24} \, = \, 2\beta_1 +3\beta_2 +4\beta_3+2\beta_4 \\(\psi \, , \,\beta_1) = 2
\quad; \quad (\psi \, , \, \beta_i ) = 0 \quad i \ne 1 \
\end{array} $$
\vskip 1cm \caption{The Dynkin diagram of $F_{4(4)}$. The only root which is not orthogonal
to the highest root is $\beta_V = \beta_1$. The root $\beta_V = \beta_1 $
is the highest weight of the $\mathbf{W}$-representation of $\mathfrak{sp}(6,\mathbb{R})$ \label{F4dynk}}
\end{figure}
The complete set of positive roots contains $24$ elements that are listed below:
\begin{equation}\label{radicioneF4}
\begin{array}{lllll}
\beta _1 & = & \beta _1 & = & \{-1,-1,-1,1\} \\
\beta _2 & = & \beta _2 & = & \{0,0,2,0\} \\
\beta _3 & = & \beta _3 & = & \{0,1,-1,0\} \\
\beta _4 & = & \beta _4 & = & \{1,-1,0,0\} \\
\beta _5 & = & \beta _1+\beta _2 & = & \{-1,-1,1,1\} \\
\beta _6 & = & \beta _2+\beta _3 & = & \{0,1,1,0\} \\
\beta _7 & = & \beta _3+\beta _4 & = & \{1,0,-1,0\} \\
\beta _8 & = & \beta _1+\beta _2+\beta _3 & = &
\{-1,0,0,1\} \\
\beta _9 & = & \beta _2+2 \beta _3 & = & \{0,2,0,0\} \\
\beta _{10} & = & \beta _2+\beta _3+\beta _4 & = &
\{1,0,1,0\} \\
\beta _{11} & = & \beta _1+\beta _2+2 \beta _3 & = &
\{-1,1,-1,1\} \\
\beta _{12} & = & \beta _1+\beta _2+\beta _3+\beta _4 & = &
\{0,-1,0,1\} \\
\beta _{13} & = & \beta _2+2 \beta _3+\beta _4 & = &
\{1,1,0,0\} \\
\beta _{14} & = & \beta _1+2 \beta _2+2 \beta _3 & = &
\{-1,1,1,1\} \\
\beta _{15} & = & \beta _1+\beta _2+2 \beta _3+\beta _4 & =
& \{0,0,-1,1\} \\
\beta _{16} & = & \beta _2+2 \beta _3+2 \beta _4 & = &
\{2,0,0,0\} \\
\beta _{17} & = & \beta _1+2 \beta _2+2 \beta _3+\beta _4 &
= & \{0,0,1,1\} \\
\beta _{18} & = & \beta _1+\beta _2+2 \beta _3+2 \beta _4 &
= & \{1,-1,-1,1\} \\
\beta _{19} & = & \beta _1+2 \beta _2+3 \beta _3+\beta _4 &
= & \{0,1,0,1\} \\
\beta _{20} & = & \beta _1+2 \beta _2+2 \beta _3+2 \beta _4
& = & \{1,-1,1,1\} \\
\beta _{21} & = & \beta _1+2 \beta _2+3 \beta _3+2 \beta _4
& = & \{1,0,0,1\} \\
\beta _{22} & = & \beta _1+2 \beta _2+4 \beta _3+2 \beta _4
& = & \{1,1,-1,1\} \\
\beta _{23} & = & \beta _1+3 \beta _2+4 \beta _3+2 \beta _4
& = & \{1,1,1,1\} \\
\beta _{24} & = & 2 \beta _1+3 \beta _2+4 \beta _3+2 \beta
_4 & = & \{0,0,0,2\}
\end{array}
\end{equation}
In eq.(\ref{radicioneF4}) the first column is the name of the root, the second column gives its decomposition in terms of simple roots, while the last column provides the component of the root vector in $\mathbb{R}^4$.
\par
The standard Cartan-Weyl form of the Lie algebra is as follows:
\begin{eqnarray}
\left[ \mathcal{H}_i \, , \, E^{\pm\beta}\right] &=& \pm \, \beta^i \, E^{\pm\beta_I} \, \label{weylus1} \\
\left[ E^{\beta} \, , \, E^{-\,\beta}\right] &=& \, \beta \, \cdot \, \mathcal{H} \, \label{weylus2} \\
\left[ E^{\beta}\, , \, E^{\gamma}\right] &=& \left\{ \begin{array}{lc}N_{\beta \gamma} \, E^{\beta \, + \, \gamma}& \mbox{if $\beta+\gamma$ is a root}\\
0 &\mbox{if $\beta+\gamma$ is not a root}\end{array} \right. \label{weylus3}
\end{eqnarray}
where $N_{\beta \gamma}$ are numbers that can be predicted from Lie algebra theory. They are irrelevant, since they follows from commutators, when one has an explicit matrix realization of all the Cartan generators $\mathcal{H}_i$ and of the step operators $E^{\pm\beta}$. The fundamental representation of $F_{(4,4)}$ is $26$-dimensional and real. We have constructed in a MATHEMATICA code all the $26 \times 26$ matrix representations of the $52$ generators of the Cartan Weyl basis. Obviously we can not present them here because they are too big. However all the rest of the construction can be easily presented in terms of these Weyl generators sand this is what we presently do.
\subsection{\sc The maximal compact subalgebra $\mathbb{H} \, = \, \mathfrak{su}(2) \oplus \mathfrak{usp}(6)$}
The maximal compact subalgebra $\mathbb{H}$ of a maximally split simple Lie algebra such as $F_{(4,4)}$, is just the real span of all the independent compact generators $E^{\beta_i} \, - \, E^{- \beta_i}$. In our case we have $24$ positive roots and we can write:
\begin{equation}\label{algebraH}
\mathbb{H} \, = \, \mbox{span}_\mathbb{R} \, \left\{ H_1\, , \, H_2 \, , \, \dots \, , \, H_{24} \right\}
\end{equation}
where we have defined:
\begin{equation}\label{Higene}
H_i \, = \, E^{\beta_i} \, - \, E^{- \beta_i}
\end{equation}
the positive roots being numbered as in eq.(\ref{radicioneF4}). We know from theory that this maximal compact subalgebra has the structure:
\begin{equation}\label{furtivo}
\mathbb{H} \, = \, \mathfrak{su}(2) \oplus \mathfrak{usp}(6)
\end{equation}
It is important to derive an explicit basis of generators satisfying the standard commutation relations of the two simple factors in eq.(\ref{furtivo}) for holonomy calculations of the coset manifold. Particularly important are the three generators $J^x$ of the $\mathfrak{su}(2)$ subalgebra since they will act as quaternionic complex structures in the calculation of the tri-holomorphic moment map. By means of standard techniques of diagonalization of the adjoint action of generators we have retrieved the required basis rearrangement.
\subsubsection{\sc The $\mathfrak{su}(2)$ Lie algebra} The three generators $J^x$ have tho following explicit form:
\begin{eqnarray}
J^1 &=& \frac{H_1-H_{14}+H_{20}-H_{22}}{4 \sqrt{2}} \nonumber\\
J^2 &=& \frac{H_5+H_{11}-H_{18}+H_{23}}{4 \sqrt{2}}\nonumber \\
J^3 &=& -\frac{H_2-H_9+H_{16}+H_{24}}{4 \sqrt{2}}\label{su2generati}
\end{eqnarray}
and close the standard commutation relations:
\begin{equation}\label{stundasu2}
\left[ J^x \, , \, J^y \right] \, = \, \epsilon^{xyz} \, J^y
\end{equation}
\subsubsection{\sc The $\mathfrak{usp}(6)$ Lie algebra} The $21$ generators of the $\mathfrak{usp}(6)$ Lie algebra are given by the following combinations.
First we have three mutually commuting generators (the compact Cartan generators):
\begin{equation}\label{cartacantabene}
\left[ \mathcal{L}^i \, , \, \mathcal{L}^j \right]\, = \, 0
\end{equation}
that are given by the following combinations:
\begin{equation}\label{compacarta}
\begin{array}{lll}
\mathcal{L}^1 & = &
-\frac{H_2}{2}-\frac{H_9}{2}+\frac{H_{16
}}{2}-\frac{H_{24}}{2} \\
\mathcal{L}^2 & = &
-\frac{H_2}{2}+\frac{H_9}{2}+\frac{H_{16
}}{2}+\frac{H_{24}}{2} \\
\mathcal{L}^3 & = &
\frac{H_2}{2}+\frac{H_9}{2}+\frac{H_{16}
}{2}-\frac{H_{24}}{2}
\end{array}
\end{equation}
Secondly we have $9$ pairs of generators $\left\{X_{i}\, , \, Y_{i}\right\}$ which are in correspondence with the $9$ positive roots of the $\mathfrak{sp}(6,C)$ Lie algebra (see eq.(\ref{cornette}). Explicitly we have:
\begin{equation}\label{coppietteXY}
\begin{array}{lllllll}
X_1 & = & H_{10} & ; &
Y_1 & = & H_7 \\
X_2 & = & H_4 & ; & Y_2 &
= & -H_{13} \\
X_3 & = & H_6 & ; & Y_3 &
= & -H_3 \\
X_4 & = &
-H_1+H_{14}+H_{20}-H_{22} & ; &
Y_4 & = &
-H_5-H_{11}-H_{18}+H_{23} \\
X_5 & = & H_{21} & ; &
Y_5 & = & -H_8 \\
X_6 & = & H_1+H_{14}+H_{20}+H_{22}
& ; & Y_6 & = &
H_5-H_{11}-H_{18}-H_{23} \\
X_7 & = &
-H_1-H_{14}+H_{20}+H_{22} & ; &
Y_7 & = &
H_5-H_{11}+H_{18}+H_{23} \\
X_8 & = & H_{17} & ; &
Y_8 & = & H_{15} \\
X_9 & = & H_{12} & ; &
Y_9 & = & H_{19}
\end{array}
\end{equation}
The commutation relations with the compact Cartan generators are as follows:
\begin{equation}\label{cirocondo}
\left[ \mathcal{L}^i \, , \, X_I \right] \, = \, \alpha_I^i \, Y_I \quad ; \quad \left[ \mathcal{L}^i \, , \, Y_I \right] \, = \, - \, \alpha_I^i \, X_I
\end{equation}
where $\alpha_I$ are the roots of eq.(\ref{cornette}). The remaining commutation relations mix the $Y$ and the $X$ among themselves and reproduce the Cartan generators.
\subsection{\sc The subalgebra $\mathfrak{sl}(2,\mathbb{R})_E \, \oplus \, \mathfrak{sp}(6,\mathbb{R})$ and the $\mathbf{W}$-generators}
Of great relevance in all applications of the (pseudo)-quaternionic geometry either in the construction of Black-Hole solutions or in the quest of inflaton potentials by means of the gauging of hypermultiplet isometries is the identification of the subalgebra:
\begin{equation}\label{cannula}
\mathfrak{sl}(2,\mathbb{R})_E \, \oplus \, \mathfrak{sp}(6,\mathbb{R}) \, \subset \, \mathfrak{f}_{(4,4)}
\end{equation}
and the recasting of $\mathfrak{f}_{(4,4)}$ in the general form \ref{genGD3} by means of the identification of the $\mathbf{W}$-generators.
\par
To this effect a very powerful tool is provided by the comparison of the $\mathfrak{f}_{(4,4)}$ root system displayed in eq.(\ref{radicioneF4}) with the $\mathfrak{sp}(6,\mathbb{R})$ root system displayed in eq.(\ref{cornette}). The step operators associated with the highest (lowest) root $\pm\beta_{24}$ are the only ones that have a grading $\pm 2$ with respect to the fourth Cartan generator $\mathcal{H}_4$. These three operators close among themselves the Lie algebra $\mathfrak{sl}(2,\mathbb{R})_E $. There are $9$ roots that have grading zero with respect to $\mathcal{H}_4$. Projected onto the plane $\mathcal{H}_4\, = \,0$ these $9$ roots form, together with their negatives, a $\mathfrak{sp}(6,\mathbb{R})$ root system. Correspondingly the $\mathfrak{sp}(6,\mathbb{R})$ subalgebra is generated by the step operators associated with these $9$ roots (and with their negaives) plus the first $3$ Cartan generators. Finally there are $14$ positive roots $\beta$ that have have grading $1$ with respect to $\mathcal{H}_4$. The step operators associated with these $14$ roots form the $\mathbf{W}$-generators with index $1$ of $\mathrm{SL(2,\mathbb{R})}_E$. Their partners with index $2$ are provided by the corresponding negative root step operators.
\par
It is quite important to arrange the generators $\mathbf{W}$ in such a way that under any element $\mathfrak{g }\, \in \, \mathfrak{sp}(6,\mathbb{R}) \, \subset \, \mathfrak{f}_{(4,4)} $ they transform exactly with the $\mathcal{D}_{14}(\mathfrak{g})$ matrices defined in eq.s(\ref{cartolini14}), (\ref{steppinisu14}) and (\ref{steppinigiu14}).
\par
The precise definition of all the generators that satisfy the specified requirements is given below.
\subsubsection{\sc The Ehlers subalgebra $\mathfrak{sl}(2,\mathbb{R})_E$.} The standard commutation relations:
\begin{eqnarray}
\left[ L_0^E \, , \, L^E_\pm\right] &=& \pm \, L^E_\pm \\
\left[ L^E_+\, , \, L^E_- \right] &=& 2 \, L^E_0 \label{EhlersF4}
\end{eqnarray}
are satisfied by the following generators:
\begin{eqnarray}
L_0^E&=& \frac{1}{2} \, \mathcal{H}_4 \nonumber \\
L_+^E &=& \frac{1}{\sqrt{2}} \, \mathcal{E}^{\beta_{24}}\nonumber \\
L_- ^E&=& \frac{1}{\sqrt{2}} \, \mathcal{E}^{- \,\beta_{24}} \label{EhlersF4BIS}
\end{eqnarray}
\subsubsection{\sc The subalgebra $\mathfrak{sp}(6,\mathbb{R})$.} The Cartan generators are the following ones:
\begin{eqnarray}
\mathcal{H}_1 &=& \mathcal{H}_1 \nonumber\\
\mathcal{H}_2 &=& \mathcal{H}_2 \nonumber\\
\mathcal{H}_3 &=& \mathcal{H}_3 \label{samicarti}
\end{eqnarray}
while the step operators are identified as follows
\begin{equation}
\begin{array}{lll}
\mathcal{E}^{\pm \alpha _1} & = & \mathcal{E}^{\pm\beta_{4}} \\
\mathcal{E}^{\pm \alpha _2} & = & \mathcal{E}^{\pm\beta_{3}} \\
\mathcal{E}^{\pm \alpha _3} & = & \mathcal{E}^{\pm\beta_{2}} \\
\mathcal{E}^{\pm \alpha _4} & = & \mathcal{E}^{\pm\beta_{7}} \\
\mathcal{E}^{\pm \alpha _5} & = & -\mathcal{E}^{\pm\beta_{6}}\\
\mathcal{E}^{\pm \alpha _6} & = & \mathcal{E}^{\pm\beta_{10}}\\
\mathcal{E}^{\pm \alpha _7} & = & -\mathcal{E}^{\pm\beta_{9}}\\
\mathcal{E}^{\pm \alpha _8} & = & \mathcal{E}^{\pm\beta_{13}}\\
\mathcal{E}^{\pm \alpha _9} & = & \mathcal{E}^{\pm\beta_{16}}
\end{array}\label{fognato}
\end{equation}
We would like to attract the attention of the reader on the two minus signs introduced in the identifications (\ref{fognato}). Together with the other minus signs that appear below in the identification of the $W$-generators these signs are essential in order for the transformations of the $W$.s to be identical with those given by the previously defined $\mathcal{D}_{14}(\mathfrak{g})$ matrices.
\subsubsection{\sc The $\mathbf{W}$-generators} Casting the $\mathfrak{f}_{(4,4)}$ Lie algebra in the general form (\ref{genGD3}) is completed by the identification of the $\mathbf{W}$-generators. We find:
\begin{equation}\label{widentifio}
\begin{array}{lll}
\mathbf{W}^{1,1} & = & \mathcal{E}^{ \beta_{5}} \\
\mathbf{W}^{1,2} & = & \mathcal{E}^{ \beta_{20}} \\
\mathbf{W}^{1,3} & = & \mathcal{E}^{ \beta_{14}} \\
\mathbf{W}^{1,4} & = & -\mathcal{E}^{ \beta_{23}} \\
\mathbf{W}^{1,5} & = & \mathcal{E}^{ \beta_{21}} \\
\mathbf{W}^{1,6} & = & \mathcal{E}^{ \beta_{19}} \\
\mathbf{W}^{1,7} & = & -\mathcal{E}^{ \beta_{17}} \\
\mathbf{W}^{1,8} & = & -\mathcal{E}^{ \beta_{22}} \\
\mathbf{W}^{1,9} & = & -\mathcal{E}^{ \beta_{11}} \\
\mathbf{W}^{1,10} & = & -\mathcal{E}^{ \beta_{18}} \\
\mathbf{W}^{1,11} & = & -\mathcal{E}^{ \beta_{1}} \\
\mathbf{W}^{1,12} & = & -\mathcal{E}^{ \beta_{8}} \\
\mathbf{W}^{1,13} & = & -\mathcal{E}^{ \beta_{12}} \\
\mathbf{W}^{1,14} & = & -\mathcal{E}^{ \beta_{15}}
\end{array}
\end{equation}
and for all $\mathfrak{g }\, \in \, \mathfrak{sp}(6,\mathbb{R}) \, \subset \, \mathfrak{f}_{(4,4)} $ we have:
\begin{equation}\label{ciucciatiquesta}
\left[ \mathfrak{g } \, , \, \mathbf{W}^{1,\alpha} \right] \, = \, \mathcal{D}_{14}(\mathfrak{g})^\alpha_{\phantom{\alpha}\gamma} \, \mathbf{W}^{1,\gamma}
\end{equation}
The generators $\mathbf{W}^{2,\alpha}$ are then easily obtained from by means of a rotation with the unique compact generator of the Ehlers subalgebra introduced in eq.(\ref{ruotogrande}):
\begin{equation}\label{fungatorepazzo}
\left[ \mathfrak{S} \, , \, \mathbf{W}^{1,\alpha} \right] \, = \,\mathbf{W}^{2,\alpha}
\end{equation}
\subsection{\sc The solvable coset representative}
The precise constructions of the previous sections enable us to introduce the solvable coset representative $\mathbb{L}_{Solv}\left(a,U,h,p,Z\right)$ of the manifold (\ref{quatergurto}) such that the Maurer Cartan form:
\begin{equation}\label{maurocartaneggia}
\Xi \, \equiv \, \mathbb{L}_{Solv}^{-1} \, \mathrm{d} \mathbb{L}_{Solv}
\end{equation}
decomposed along the generators of the Borel Lie algebra:
\begin{eqnarray}\label{solubilino}
\Xi & = & E^I_{\mathcal{QM}} \, T_I\nonumber\\
T_I & = & \left\{ \underbrace{L_0^E \, , \, L^E_+ }_{2 \,\hookrightarrow\, Solv\left[\mathfrak{sl}(2)\right]} \, , \, \underbrace{\mathcal{H}^i \, , \, \mathcal{E}^{\alpha_i}}_{12 \, \hookrightarrow \, Solv\left[\mathfrak{sp}(6)\right]} \, ,\, \underbrace{\mathbf{W}^{1\alpha}}_{14 \, \hookrightarrow \, \mathbb{H}\mathrm{eis}}\right\}
\end{eqnarray}
provides the vielbein $E^I_{\mathcal{QM}}$ mentioned in eq.(\ref{filibaine}) and by squaring the metric (\ref{cornish}).
\par
In full analogy with eq.s (\ref{fornarina}) and (\ref{baltico}) we write:
\begin{eqnarray}\label{SolvCosF4}
\mathbb{L}_{Solv} & = & \exp \left[ a \, L^E_+\right] \, \cdot \, \exp\left[\sum_{j=1}^7 \, \mathbf{Z}_{2j-1} \, \mathbf{W}^{1,2j-1}\right] \, \cdot \, \exp\left[\sum_{j=1}^7 \, \mathbf{Z}_{2j} \, \mathbf{W}^{1,2j}\right] \, \times \nonumber \\
&& \times \prod_{i=1}^9 \, \exp\left [ p_{10-i} \, \mathcal{E}^{\alpha_{10-i}}\right ] \, \cdot \, \prod_{j=3}^3 \exp\left [ h_{j} \, \mathcal{H}^{j}\right ] \, \cdot \, \exp\left[U \, L^E_0 \right]
\end{eqnarray}
The explicit expression of $\mathbb{L}_{Solv}$ in the fundamental $26$-dimensional representation is obviously very large but it can be dealt with by means of an appropriate MATHEMATICA code.
\par
We are finally in the position of calculating the tri-holomorphic moment map of any element $ \mathfrak{t}\, \in \, \mathfrak{f}_{(4,4)}$
of the isometry Lie algebra of $\mathcal{QM}$ through the formula:
\begin{equation}\label{cordiglierAndina}
\mathcal{P}^x_\mathfrak{t} \, = \, \mbox{Tr}_{\mathbf{26}} \, \left( J^x \, \mathbb{L}_{Solv}^{-1} \, \mathfrak{t} \, \mathbb{L}_{Solv} \right)
\end{equation}
\subsection{\sc The example of the inclusion of multi Starobinsky models}
\label{generalonuovo}
In section \ref{truncazia} we studied the truncation of the $\mathfrak{sp}(6,\mathbb{R})$ model to the STU model. There we showed that setting to zero the three complex coordinates $z_2,z_3,z_5$, the remaining ones $z_1,z_4,z_6$ span the STU model, namely they parameterize three copies of the Lobachevsky-Poincar\'e hyperbolic plane. Inspecting eq. (\ref{carlinus}) we also see that the three coordinates $z_1,z_4,z_6$ are the only ones that survive when all the axions $p_i$ are set to zero. We also recall from sect. \ref{truncazia} that the three parabolic generators of the three $\mathrm{SL(2,\mathbb{\mathbb{R}})}$ groups spanning the STU model are $\mathcal{E}^{\alpha_3},\, \mathcal{E}^{\alpha_7},\,\mathcal{E}^{\alpha_9}$ whose identification with $\mathfrak{f}_{(4,4)}$ generators is provided by eq.(\ref{fognato}). Correspondingly we introduce the following generator:
\begin{equation}\label{carnevaleSTU}
\mathfrak{t}_{STU}\, = \, \beta_3 \, \mathcal{E}^{\alpha_3}\, + \, \beta_2 \, \mathcal{E}^{\alpha_7}\, + \, \beta_1 \, \mathcal{E}^{\alpha_9}\, - \, \kappa \, \mathfrak{S}
\end{equation}
and we calculate its tri-holomorphic moment map, by means of eq.(\ref{cordiglierAndina}). Defining the potential:
\begin{equation}\label{gogamigoga}
V_{STU} \, = \, \sum_{x=1}^3 \, \left(\mathcal{P}^x_{\mathfrak{t}_{STU}}\right)^2
\end{equation}
We can verify that:
\begin{eqnarray}
\left. \frac{\partial}{\partial \mathbf{Z}^\alpha} \, V_{STU}\right|_{\mathbf{Z}=U=a =0} & = &0 \nonumber\\
\left. \frac{\partial}{\partial U} \, V_{STU}\right|_{\mathbf{Z}=U=a =0} & = &0 \nonumber\\
\left. \frac{\partial}{\partial a} \, V_{STU}\right|_{\mathbf{Z}=U=a =0} & = &0 \label{UaZetaseneva}
\end{eqnarray}
Hence we can consistently truncate $U$, $a$ and the Heisenberg fields $\mathbf{Z}$. We find:
\begin{equation}\label{multistarobbo}
\left. V_{STU}\right|_{\mathbf{Z}=U=a =0} \, = \, \frac{9}{4} \left( \, 2\kappa \, - \, \sqrt{2} \sum_{i=1}^3
\beta_i \,e^{-2 h_i} \right)^2
\end{equation}
The above potential can be named a multi-Starobinsky model with three independent dilatons.
\par
First of all let us note that in the above model the absolute value of $\beta_i$ is irrelevant since we can always reabsorb it by a constant shift $h_i \to h_i \, - \, \log|\beta_i|$. The only relevant thing are the signs of $\beta_i$ including in this notion also zero, namely $\beta_i $ can be $\pm 1$ or $0$. Secondly we observe that when all the non vanishing $\beta_i$ have the same sign we can make a consistent one field truncation to
\begin{equation}\label{correlo}
h_i \, = \, h \quad ; \quad \mbox{for all $i$ such that $\beta_i \ne 0$}
\end{equation}
After this truncation the potential (\ref{multistarobbo}) becomes the following:
\begin{equation}\label{riducione}
V_{eff} \, = \, \frac{9}{4} \left(\, 2\kappa \, - \, \sqrt{2} \, q\, e^{-2 h}\right)^2
\end{equation}
where $q$ is the number of equal sign non zero $\beta_i$, which obviously can take only three values $q=1,2,3$. In order to compare this result with the definition of $\alpha$-attractors introduced in \cite{alfatrattori}, we just have to compare the potential (\ref{riducione}) with the normalization of the scalar kinetic terms in the lagrangian:
\begin{equation}\label{kinescalo}
\mathcal{L} \, = \, \dots \, + \, \frac{1}{4} \, (\partial U)^2 \, + \, (\partial h_1)^2\, + \, (\partial h_2)^2 \, + \, (\partial h_3)^2 \, + \, \dots
\end{equation}
which follows from eq.s(\ref{geodaction}, \ref{formidabile}). Renaming $h\, = \, \frac{1}{\sqrt{2 \, q}} \, \phi$, so that the new field $\phi$ has canonical kinetic term $\frac{1}{2} \, (\partial \phi)^2$, we obtain a potential:
\begin{equation}\label{bambolone}
V_{eff} \, = \, \mbox{const} \, \times \, \left( 2\kappa \, - \, \sqrt{2} \, q\, \exp\left [ - \, \sqrt{\frac{2}{q}} \, \phi\right]\right)^2
\end{equation}
which, in the notation of \cite{alfatrattori}, corresponds to $\alpha = \frac{q}{3} $, namely to:
\begin{equation}\label{valoretti}
\alpha \, = \, 1 \, , \, \frac{2}{3} \, , \, \frac{1}{3}
\end{equation}
The above result has been obtained by gauging only one generator, namely (\ref{carnevaleSTU}). Correspondingly we have generated Starobinsky-like models with only one massive vector that is the gauge vector associated with the gauged generator. There is another way of obtaining the same potential but with $q$-massive vectors (one for each constituent Starobinsky model with $q=\frac{1}{3}$). This is very simply understood remarking that the $\mathfrak{f}_{(4,4)}$ algebra contains an $\mathfrak{sl}(2,\mathbb{R})^4$ subalgebra singled out as follows:
\begin{equation}\label{governolato}
\mathfrak{f}_{(4,4)} \, \supset \, \mathfrak{sl}(2,\mathbb{R})_E \, \oplus \, \underbrace{\mathfrak{sl}(2,\mathbb{R})_S \, \oplus \, \mathfrak{sl}(2,\mathbb{R})_T\, \oplus \, \mathfrak{sl}(2,\mathbb{R})_U}_{\subset \, \mathfrak{sp}(6,\mathbb{R})}
\end{equation}
where $\mathfrak{sl}(2,\mathbb{R})_S \, \oplus \, \mathfrak{sl}(2,\mathbb{R})_T\, \oplus \, \mathfrak{sl}(2,\mathbb{R})_U$ describes the STU model embedded in the K\"ahler manifold (\ref{gurto}). These four $\mathfrak{sl}(2,\mathbb{R})$ algebras are completely symmetric among themselves and the gauging of their generators produce identical results. So we can introduce the abelian gauge algebra spanned by the following three commuting generators:
\begin{eqnarray}
\mathfrak{t}_{S} &=&\beta_3 \, \mathcal{E}^{\alpha_3}\, - \, \kappa_3 \, \mathfrak{S} \nonumber \\
\mathfrak{t}_{T} &=& \beta_2 \, \mathcal{E}^{\alpha_7}\, - \, \kappa_2 \, \mathfrak{S}\nonumber\\
\mathfrak{t}_{U} &=& \, \beta_1 \, \mathcal{E}^{\alpha_9}\, - \, \kappa_1\, \mathfrak{S}
\end{eqnarray}
Gauging with three separate vectors each of the above generators we obtain a new potential:
\begin{equation}\label{STUnew}
\widehat{V}_{STU} \, = \, \sum_{x=1}^3 \, \left(\mathcal{P}^x_{\mathfrak{t}_{S}}\right)^2 \, + \, \sum_{x=1}^3 \, \left(\mathcal{P}^x_{\mathfrak{t}_{T}}\right)^2 \, + \, \sum_{x=1}^3 \, \left(\mathcal{P}^x_{\mathfrak{t}_{U}}\right)^2 \,
\end{equation}
that has the same property as the potential (\ref{gogamigoga}), namely it allows us to truncate consistently to zero all the axions $p_i$, all the Heisenberg fields $\mathbf{Z}^\alpha$ and the Taub NUT field $a$. The reduced potential after such a truncation has the form:
\begin{equation}\label{ciurlone}
\widehat{V}_{red}\, = \, \frac{9}{4} \, \sum_{i=1}^3 \, \left( \, 2\kappa_i \, - \, \sqrt{2}
e^{-2 h_i} \beta_i \right)^2
\end{equation}
As we already remarked before, the absolute value of the $\beta_i$ parameters is irrelevant: what matters is only the relative signs of the $\beta_i$ with respect to the sign of their corresponding $\kappa_i$. If for all non vanishing $\beta_i$ we have $\frac{\beta_i}{\kappa_i} \, = 1$, then we can consistently perform the same truncation (\ref{correlo}) as before and we reobtain the potentials (\ref{bambolone}) with the same spectrum of $\alpha$-values (\ref{valoretti}). The difference with the previous case is, as we emphasized at the beginning o this discussion, that now the number of massive fields is $q$, namely as many as the elementary non trivial constituent Starobinsky-like models.
\subsection{\sc Nilpotent gaugings and truncations}
\label{orbitando}
Let us now put the above obtained results in the general framework discussed in sect.\ref{starobin1}. The issue is the classification of orbits of nilpotent operators and the question whether for each of these orbits we can find a consistent one-field reduction that produces a Starobinsky-like model with an appropriate value of $\alpha$.
\par
To answer this question we have followed the algorithm described in the second paper of \cite{noinilpotenti}. According to a general mathematical set, up to conjugation, every nilpotent orbit is associated with a standard triple $\left \{ x,y,h \right \}$ satisfying the standard commutation relations of the $\mathfrak{sl}(2)$ Lie algebra, namely:
\begin{equation}\label{basictriple}
\left [ h \, , \, x\right ] \, = \, x \quad ; \quad \left [ h \, , \, y\right ] \, = -\, y
\quad ; \quad \left [ x \, , \, y \right ] \, = \, 2 \, h
\end{equation}
Interesting for us is the classification of nilpotent orbits in the K\"ahler subalgebra $\mathfrak{sp}(6,\mathbb{R})$ and, according to the above mathematical theory, this is just the classification of embeddings of an $\mathfrak{sl}(2)$ Lie algebra in the ambient one, modulo conjugation by the full group $\mathrm{Sp(6,\mathbb{R})}$.
The second relevant point emphasized in \cite{noinilpotenti} is that embeddings of subalgebras $\mathfrak{h} \subset \mathfrak{g}$ are characterized by the branching law of any representation of $\mathfrak{g}$ into irreducible representations of $\mathfrak{h}$. Clearly two embeddings might be conjugate only if their branching laws are identical. Embeddings with different branching laws necessarily belong to different orbits. In the case of the $\mathfrak{sl}(2) \sim \mathfrak{so}(1,2)$ Lie algebra, irreducible representations are uniquely identified by their spin $j$, so that the branching law is expressed by listing the angular momenta $\left\{ j_1 , j_2 , \dots j_n\right \}$ of the irreducible blocks into which any representation of the original algebra, for instance the fundamental, decomposes with respect to the embedded subalgebra. The dimensions of each irreducible module is $2j+1$ so that an a priori constraint on the labels $\left\{ j_1 , j_2 , \dots j_n\right \}$ characterizing an irreducible orbit of $\mathfrak{sp}(6,\mathbb{R})$ is the summation rule:
\begin{equation}\label{summarulla}
\sum_{i=1}^{n} (2 j_i +1) \, = \, 6\, = \, \mbox{dimension of the fundamental representation}
\end{equation}
Therefore we have considered all possible partitions of the number $6$ into integers and for each partition we have constructed a candidate $h$ element in the Cartan subalgebra of $\mathfrak{sp}(6,\mathbb{R})$ containing as eigenvalues all the $J_3$ values of the corresponding $\left\{ j_1 , j_2 , \dots j_n\right \}$ representation. To clarify what we mean by this it suffices to consider the example of the first partition $6=6$. In this case the $6$ dimensional representation of $\mathfrak{sl}(2)$ is the $j \, = \, \frac{5}{2}$ and the $6$ eigenvalues are $\pm \frac{5}{2}$, $\pm \frac{3}{2}$, $\pm \frac{1}{2}$. Having so fixed the so named central element $h$ of the candidate standard triplet we have tried to construct the corresponding $x$ and $y$. Imposing the standard commutation relations (\ref{basictriple}) one obtains quadratic equations on the coefficients of the linear combinations expressing the candidate $x$ and $y$ that may have or may not have solutions. If the solutions exist, then the corresponding standard triple is found, the orbit exists and we have constructed one representative $x$.
\par
Next, given the existing orbits and the corresponding standard triples, for each of them we have constructed a Lobachevsky complex plane immersed in the Special K\"ahler manifold $\mathcal{M}_{Sp6}$ defined by eq.(\ref{gurto}). The construction is very simple. One calculates the group element $\mathfrak{g}(\lambda,\psi) \, \in \, \mathfrak{sp}(6,\mathbb{R})$ defined below:
\begin{equation}\label{lambopsi}
\mathfrak{g}(\lambda,\psi) \, = \, \exp\left[ \psi \, x\right] \, \cdot \, \, \exp\left[ \lambda \, h\right] \, = \, \left (\begin{array}{c|c}
\mathbf{A}(\lambda,\psi) & \mathbf{B}(\lambda,\psi) \\
\hline
\mathbf{C}(\lambda,\psi) & \mathbf{D}(\lambda,\psi)
\end{array}
\right)
\end{equation}
and using equation (\ref{Zmatra}), we write:
\begin{eqnarray}\label{embeddone}
Z(\lambda\, , \, \psi) & = & \left( \mathbf{A}(\lambda,\psi) \, - \, {\rm i} \mathbf{B}(\lambda,\psi)\right) \, \cdot \, \left( \mathbf{C}(\lambda,\psi) \, - \, {\rm i} \mathbf{D}(\lambda,\psi)\right)^{-1} \, \nonumber\\
& \equiv & \left( \begin{array}{ccc}
z_1(\lambda,\psi) & z_2(\lambda,\psi) & z_3(\lambda,\psi) \\
z_2(\lambda,\psi) & z_4(\lambda,\psi) & z_5(\lambda,\psi) \\
z_3(\lambda,\psi) & z_5(\lambda,\psi) & z_6(\lambda,\psi)
\end{array}
\right)
\end{eqnarray}
which defines the explicit embedding:
\begin{equation}\label{embeddus}
\phi \, : \, \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}} \, \rightarrow \, \frac{\mathrm{Sp(6,\mathbb{R})}}{\mathrm{SU(3)} \times \mathrm{U(1)}} \, \equiv \, \mathcal{M}_{Sp6}
\end{equation}
of the Lobachevsky plane in $\mathcal{M}_{Sp6}$. Indeed from (\ref{embeddone}) we read off the parameterization of the complex coordinates $z_i$ ($i=1,\dots ,6$) as functions of $\lambda \, = \, \log \mbox{Im}\, w$ and $\psi \, = \, \mbox{Re}\,w$, the complex variable $w$ being the local variable over the embedded Poincar\'e-Lobachevsky plane.
\par
The question is whether the field equations of the scalar fields:
\begin{equation}\label{finocchione}
\partial_i \, \partial_{j^\star} \,\mathcal{K}\, \partial^\mu\partial_\mu \,\bar{z}^{j^\star} \, + \, \partial_i \, \partial_{j^\star}\,\partial_{k^\star} \, \mathcal{K} \, \partial_\mu z^{j^\star} \, \partial_\mu z^{k^\star}\, - \, \frac{1}{4}\,\partial_i \, V_{gauging}\left(z\, ,\, {\bar z}\right) \, = \, 0
\end{equation}
admit first a consistent reduction to the complex scalar field $w$ and then a consistent truncation to a vanishing axion $\psi \, = \, 0$. Consistency of the truncation can be verified or disproved in the following simple way. The pull-back on the immersed surface $\phi^\star \left(\frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\right) \, \subset \, \mathcal{M}_{Sp6}$ of the twelve field equations (\ref{finocchione}) (six complex equations) should be consistent among themselves and be identical with the two field equations obtained from the variation of the pull-back $\phi^\star(\mathcal{L})$ on the immersed surface of the Lagrangian $\mathcal{L}$ from which eq.s (\ref{finocchione}) derive, namely:
\begin{equation}\label{lagrandona}
\mathcal{L} \, = \, 4 \, \partial_i \, \partial_{j^\star} \, \mathcal{K}\, \partial_\mu \,{z}^{i} \,\partial^\mu \,\bar{z}^{j^\star} \, - \, V_{gauging}\left(z\, ,\, {\bar z}\right)
\end{equation}
In other words, defining $ w\, = \, {\rm i} \, e^{\lambda} \, + \, \psi $, the truncation is consistent if the following diagram is commutative:
\begin{eqnarray}
&&\begin{array}{ccc}
\mathcal{L}(z,{\bar z}) & \stackrel{\phi^\star}{\Longrightarrow} & \phi^\star \mathcal{L}(w,{\bar w}) \\
\downarrow & \null & \downarrow \\
\partial^\mu \, \frac{\partial \mathcal{L}}{\partial (\partial_\mu z)} \, - \, \frac{\partial \mathcal{L}}{\partial z} & \stackrel{\phi^\star}{\Longrightarrow} & \partial^\mu \, \frac{\partial \phi^\star \mathcal{L}}{\partial (\partial_\mu w)} \, - \, \frac{\partial \phi^\star \mathcal{L}}{\partial w}
\end{array} \label{coomodiag}
\end{eqnarray}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Partition&J.s&Orbit Name&One field reduction \\
\hline
6=6 & $\left(\frac{5}{2}\right)$ & $\mathfrak{O}_1$ & NO \\
6=5+1 & $\left(2,0\right)$ & Orbit does not exist & NO \\
6=4+2 & $\left(\frac{3}{2},\frac{1}{2}\right)$ & $\mathfrak{O}_2$ & NO \\
6=3+3 & $\left(1,1\right)$ & $\mathfrak{O}_3$ & NO \\
6=3+2+1 & $\left(1,\frac{1}{2},0\right)$ &Orbit does not exist & NO \\
6=3+1+1+1 & $\left(1,0,0,0\right)$ & Orbit does not exist & NO \\
6=2+2+2 & $\left(\frac{1}{2},\frac{1}{2},\frac{1}{2}\right)$ & $\mathfrak{O}_4$ & YES \\
6=2+2+1+1 & $\left(\frac{1}{2},\frac{1}{2},0,0\right)$ & $\mathfrak{O}_5$ & YES \\
6=2+1+1+1+1 & $\left(\frac{1}{2},0,0,0,0\right)$ & $\mathfrak{O}_6$ & YES \\
\hline
\end{tabular}
\end{center}
In the above table we have summarized the results of this simple investigation. There is a total of six orbits (up to possible further splitting in Weyl group orbits which we have not analyzed) and for each of them the corresponding immersion formulae in the $\mathcal{M}_{\mathrm{Sp6}}$ manifolds are those described below.
\paragraph{Orbit $\mathfrak{O}_1$: ($j=\frac{5}{2}$).}
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
-6 \psi ^5+10 i e^{\lambda } \psi ^4+5 i e^{3 \lambda } \psi ^2+i
e^{5 \lambda } & \sqrt{5} \psi \left(3 \psi ^3-4 i e^{\lambda
} \psi ^2-i e^{3 \lambda }\right) & i \sqrt{10} \left(i \psi
+e^{\lambda }\right) \psi ^2 \\
\sqrt{5} \psi \left(3 \psi ^3-4 i e^{\lambda } \psi ^2-i e^{3
\lambda }\right) & i \left(8 i \psi ^3+8 e^{\lambda } \psi
^2+e^{3 \lambda }\right) & \sqrt{2} \psi \left(3 \psi -2 i
e^{\lambda }\right) \\
i \sqrt{10} \left(i \psi +e^{\lambda }\right) \psi ^2 & \sqrt{2}
\psi \left(3 \psi -2 i e^{\lambda }\right) & i e^{\lambda }-3
\psi
\end{array}
\right) \nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, + \, \psi \label{immerO1}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittaprima}
\phi^\star \mathcal{L} \, = \, 35 \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, \frac{1}{4}\, g^2 \, \left(3\,e^{-\lambda} \, - \, \kappa\right)^2
\end{equation}
The pull-backs of the scalar field equations are inconsistent among themselves and differ from the equations derived from the pull-back of the lagrangian (\ref{orbittaprima}), hence the truncation is not consistent. No Starobinsky--like model can be obtained from this orbit.
\par
One might wonder whether the inconsistency is due to the particularly chosen coset representative (\ref{lambopsi}) and to the explicit form of the embedding (\ref{immerO1}) which turns out to be non-holomorphic. To clarify such a doubt and show that the inconsistency of the equations is an intrinsic property of the orbit, we have addressed the problem from a different view point which leads to a perfectly holomorphic embedding of the Lobachevsky plane associated with the considered orbit into the target Special K\"ahler manifold (\ref{gurto}).
\par
The argument is the following one. Having fixed the embedding $\mathfrak{sl}(2,\mathbb{R})\mapsto \mathfrak{sp}(6,\mathbb{R})$ at the level of the fundamental representation $\mathbf{6}$ it is fixed also in all other representations and we can wonder what is the branching rule of the $\mathbf{W}$-representation $\mathbf{14}^\prime$ such an embedding. By direct evaluation of the Casimir we obtain the following branching:
\begin{equation}\label{frikandello1}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, \left( j\, =\, \frac{9}{2}\right) \, \oplus \, \left( j\, =\, \frac{3}{2}\right)
\end{equation}
This means that the symplectic section (\ref{gargamelle}) splits into the sum of two vectors, one lying in the 10-dimensional space of the first representation, the other in the $4$-dimensional space of the second representation. Imposing the vanishing of the lowest spin representation introduces a set of $4$ holomorphic constraints on the six coordinates $z_i$. By construction these constraints are $\mathfrak{sl}(2,\mathbb{R})$ invariant: therefore the sought for Lobachevsky plane certainly lies in the complex two-folds defined by the vanishing of these constraints. With a little bit of work one can further eliminate one of the two remaining complex coordinates in such a way that the ten entries of the $ \left( j\, =\, \frac{9}{2}\right)$ representation correspond to all the powers $w^r$, with $r=0, 1, \dots , 9$ of a complex parameter $w$. Because of this very property $w$ can be interpreted as the local coordinate of the sought for Lobachevsky plane embedded in the K\"ahler manifold (\ref{gurto}) according to the specified orbit. Indeed if $w$ transforms by fractional linear transformation under some algebra $\mathfrak{sl}(2)$, then the $2j+1$ first powers of $w$ provide a basis for the $j$-representation of that $\mathfrak{sl}(2)$. Viceversa, if a vector, which is known to transform in the $j$-represenation of a given $\mathfrak{sl}(2)$ (up to an overall function of $w$), is made by linear combinations of the first $2j+1$ powers of a coordinate $w$, then that $w$ is the local coordinate on a Lobachevsky plane transitive under the action of that very $\mathfrak{sl}(2)$.
\par
In our case the four holomorphic constraints that express the vanishing of the $j=\frac{3}{2}$ representation inside the $14^\prime$ are the following ones:
\begin{eqnarray}
\sqrt{\frac{2}{7}} \left(\sqrt{5} \left(z_4
z_6-z_5^2\right)-2 z_2\right) &=& 0 \nonumber \\
\frac{8 z_3-\sqrt{10} z_4}{\sqrt{21}}&=& 0 \nonumber\\
\sqrt{\frac{2}{7}} \left(\sqrt{5} z_1+2 z_3 z_5-2 z_2
z_6\right) &=& 0 \nonumber \\
\frac{-\sqrt{10} z_3^2+8 z_4 z_3-8 z_2 z_5+\sqrt{10}
z_1 z_6}{\sqrt{21}}&=& 0 \label{bamboccione}
\end{eqnarray}
The explicit form of (\ref{bamboccione}) obviously depends on the standard triple chosen as representative of the orbit, yet for whatever representative the four constraints are holomorphic. The next point consists in solving (\ref{bamboccione}) in terms of a parameter $w$ so that the complementary set of ten polynomials of the $z_i$ spanning the $j=\frac{9}{2}$ representation provide all the powers of $w$ from $0$ to $9$.
The requested solution is given by:
\begin{equation}\label{sicumerolo}
z_1\to \frac{3 w^5}{16}, \, \, z_2\to \frac{3\sqrt{5} w^4}{16},\, \, z_3\to \frac{1}{4}
\sqrt{\frac{5}{2}} w^3,\, \, z_4\to w^3,\, \,z_5\to \frac{3w^2}{2 \sqrt{2}},\, \, z_6\to \frac{3 w}{2}
\end{equation}
Implementing the transformation (\ref{sicumerolo}) in the symplectic section (\ref{gargamelle}) one finds:
\begin{equation}\label{novemezzisezia}
\Omega[Z] \, \stackrel{\phi}{\Longrightarrow} \, \Omega_{\frac{9}{2}}[w] \, = \, \left(
\begin{array}{l}
\frac{3 w}{\sqrt{2}} \\
\frac{w^6}{4 \sqrt{2}} \\
-\frac{3 w^4}{4 \sqrt{2}} \\
\frac{w^9}{256 \sqrt{2}} \\
-\frac{3 w^7}{32 \sqrt{2}} \\
-\frac{1}{16} \sqrt{\frac{5}{2}} w^6 \\
-\frac{3}{16} \sqrt{5} w^5 \\
\frac{3 w^8}{128 \sqrt{2}} \\
-\sqrt{2} w^3 \\
\frac{3 w^5}{8 \sqrt{2}} \\
\sqrt{2} \\
-\frac{3 w^2}{\sqrt{2}} \\
\frac{1}{2} \sqrt{\frac{5}{2}} w^3 \\
-\frac{3}{8} \sqrt{5} w^4
\end{array}\right)
\end{equation}
which as requested contains all the powers of $w$ and has vanishing projection on the $j=\frac{3}{2}$ representation. Calculating the K\"ahler potential from such a section we obtain:
\begin{equation}\label{furutto}
\mathcal{K}_{\frac{9}{2}} \, = \, - \, \log \,\left(\bar{\Omega}_{\frac{9}{2}}[\bar{w}]\,\mathbb{C}_{14} \, \Omega_{\frac{9}{2}}[w] \right) \, = \, \log\left ( -\frac{\rm i}{256} \,(w-\bar{w})^9\right)
\end{equation}
Now the question of consistency can be readdressed in the present context. Implementing the substitution (\ref{sicumerolo}) in the six complex equations (\ref{finocchione}) (with for instance vanishing potential) do we obtain six consistent equations or not? The answer is no. The six equations (\ref{finocchione}) are inconsistent and this confirms in a holomorphic set up the same result we had previously obtained in the direct approach of eq.s (\ref{lambopsi}-\ref{embeddone}). Hence the $\mathfrak{sl}(2)$ embedding of orbit $\mathfrak{O}_1$ leads to inconsistent truncations and has to be excluded.
\paragraph{Orbit $\mathfrak{O}_2$: ($j_1=\frac{3}{2}, \, j_2=\frac{1}{2}$).}
For the second orbit, the direct approach (\ref{lambopsi}-\ref{embeddone}) leads to:
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
\left(e^{\lambda }-i \psi \right)^2 \left(2 \psi -i e^{\lambda
}\right) & 0 & \sqrt{3} \psi \left(\psi +i e^{\lambda }\right)
\\
0 & -\psi -i e^{\lambda } & 0 \\
\sqrt{3} \psi \left(\psi +i e^{\lambda }\right) & 0 & -2 \psi -i
e^{\lambda }
\end{array}
\right) \nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, + \, \psi \label{immerO2}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittaseconda}
\phi^\star \mathcal{L} \, = \, 11 \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, \frac{1}{4}\, g^2 \, \left(3\,e^{-\lambda} \, - \, \kappa\right)^2
\end{equation}
Also in this case the pull-back of the scalar field equations yields an inconsistent set and there is no truncation. No Starobinsky--like model can be obtained from this orbit. In a similar way to the previous case we can discuss the same issue in a holomorphic set up. The branching rule of the $\mathbf{14}^\prime$ representation in the considered embedding is the following one:
\begin{equation}\label{frikandello2}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, \left( j\, =\, \frac{5}{2}\right) \, \oplus \, \left( j\, =\, \frac{3}{2}\right) \, \oplus \, \left( j\, =\, \frac{3}{2}\right)
\end{equation}
and we can impose holomorphic constraints that suppress the two lowest spin representations $\left( j\, =\, \frac{3}{2}\right)$ leaving only the top one $\left( j\, =\, \frac{5}{2}\right)$ spanned by the powers of a parameter $w$ from $0$ to $5$.
Such a holomorphic embedding is given:
\begin{equation}\label{frikandellus2}
z_1 \frac{2 w^3}{3^{3/4}}, \, \, z_2\to 0,\, \, z_3\to w^2,z_4\to
\frac{w}{\sqrt[4]{3}},\, \, z_5\to 0,\, \, z_6\to \frac{2
w}{\sqrt[4]{3}}
\end{equation}
Substitution of eq.s (\ref{frikandellus2}) into the field equations (\ref{finocchione}) confirms that their pull-back on this surface is inconsistent.
\paragraph{Orbit $\mathfrak{O}_3$: ($j_1=1, \, j_2=1$).}
For the third orbit, the direct approach (\ref{lambopsi}-\ref{embeddone}) leads to
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
-i e^{2 \lambda } & -\psi ^2 & -\sqrt{2} \psi \\
-\psi ^2 & -i \left(2 \psi ^2+e^{2 \lambda }\right) & -i \sqrt{2}
\psi \\
-\sqrt{2} \psi & -i \sqrt{2} \psi & -i
\end{array}
\right) \nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, + \, \psi \label{immerO3}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittaterza}
\phi^\star \mathcal{L} \, = \, 8 \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, \frac{1}{4}\, g^2 \, \kappa^2
\end{equation}
while the pull-back of the scalar field equations is an inconsistent set. Hence this truncation is not consistent and no Starobinsky--like model can be obtained from this orbit. As in the previous two cases we can confirm the same result in a holomorphic set up, yet we consider it useless to repeat once more the same type of calculations. What is relevant to mention in view of our subsequent considerations
is the branching rule of the $\mathbf{14}^\prime$ representation under this forbidden embedding leading to inconsistent field equations::
\begin{equation}\label{frikandello3}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, \left( j\, =\, 2\right) \, \oplus \, \left( j\, =\, 2\right) \,
\oplus \, 4 \, \times \, \left( j\, =\, 0\right)
\end{equation}
\paragraph{Orbit $\mathfrak{O}_4$: ($j_1=\frac{1}{2}, \, j_2=\frac{1}{2},j_3=\frac{1}{2}$).} For the fourth orbit, the direct approach (\ref{lambopsi}-\ref{embeddone}) leads to
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
i e^{\lambda }-\psi & 0 & 0 \\
0 & i e^{\lambda }-\psi & 0 \\
0 & 0 & i e^{\lambda }-\psi
\end{array}
\right)
\nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, - \, \psi \label{immerO4}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittaquarta}
\phi^\star \mathcal{L} \, = \, 3 \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, g^2 \, \frac{1}{4} \,\left( 3 \, e^{-\lambda}\, - \, 2\, \kappa\right)^2
\end{equation}
The pull-back of the scalar field equations produces equations consistent among themselves which coincide with the equations derived from the pull-back of the lagrangian (\ref{orbittaquarta}), hence the truncation is consistent. We reobtain the Starobinsky model discussed in the previous section with $q=3$ and hence with $\alpha \, = \, 1$. In this case the consistent truncation is already produced form holomorphic constraints. Indeed equation (\ref{immerO4}) can be summarized as:
\begin{equation}\label{holomorph4}
z_2 \, = \, z_3 \, = \, z_5 \, = \, 0 \quad ; \quad z_1 \, = \, z_4 \, = \, z_6 \, = \, w
\end{equation}
It is interesting and important for our future consideration to mention the branching rule of the $\mathbf{14}^\prime$ representation under this $\mathfrak{sl}(2)$ subalgebra:
\begin{equation}\label{frikandello4}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, \left( j\, =\, \frac{3}{2}\right) \,
\oplus \, 5 \, \times \, \left( j\, =\, \frac{1}{2}\right)
\end{equation}
and the constraints (\ref{holomorph4}) precisely are the conditions under which the five representations $\left( j\, =\, \frac{1}{2}\right)$ vanish and we are left with the representation $\left( j\, =\, \frac{3}{2}\right)$ duely spanned by the powers $1,w,w^2,w^3$.
\paragraph{Orbit $\mathfrak{O}_5$: ($j_1=\frac{1}{2}, \, j_2=\frac{1}{2},j_3=0$).} For the fifth orbit, the direct approach (\ref{lambopsi}-\ref{embeddone}) leads to
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
i e^{\lambda }-\psi & 0 & 0 \\
0 & i e^{\lambda }-\psi & 0 \\
0 & 0 & i
\end{array}
\right) \nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, - \, \psi \label{immerO5}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittaquinta}
\phi^\star \mathcal{L} \, = \, 2 \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, g^2 \, \left(e^{-\lambda} \, - \, \kappa\right)^2
\end{equation}
The pull-back of the scalar field equations yields a consistent system identical with the field equations derived from the pull-back of the lagrangian (\ref{orbittasesta}), hence the truncation is consistent. We reobtain the Starobinsky--like model discussed in the previous section with $q=2$ and hence with $\alpha \, = \, \frac{2}{3}$.
\par
In this, as in the previous case, the consistent truncation is produced from holomorphic constraints. Indeed equation (\ref{immerO4}) can be summarized as:
\begin{equation}
z_2 \, = \, z_3 \, = \, z_5 \, = \, 0 \quad ; \quad z_1 \, = \, z_4 \, = \, w \quad ; \quad z_6 \, = \, {\rm i} \label{belgiuro}
\end{equation}
In this case the branching rule of the $\mathbf{14}^\prime$ representation under the considered $\mathfrak{sl}(2)$ subalgebra is the following one:
\begin{equation}\label{frikandello5}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, \left( j\, =\, 1\right) \,
\oplus \,\left( j\, =\, 1\right)\, \oplus \, 2 \, \times \, \left( j\, =\, \frac{1}{2}\right) \, + \, 4 \, \times \, \left( j\, =\, 0\right)
\end{equation}
and the constraint (\ref{belgiuro}) guarantees that the singlets and the $\left( j\, =\, \frac{1}{2}\right)$ representations are all set to zero.s
\paragraph{Orbit $\mathfrak{O}_6$: ($j_1=\frac{1}{2}, \, j_2=0,\, j_3=0$).} For the sixth orbit, the direct approach (\ref{lambopsi}-\ref{embeddone}) leads to
\begin{eqnarray}
&& \left( \begin{array}{ccc}
z_1& z_2 & z_3\\
z_2& z_4 & z_5 \\
z_3 & z_5 & z_6
\end{array}
\right) = \, \nonumber\\
&& \left(
\begin{array}{lll}
\psi +i e^{\lambda } & 0 & 0 \\
0 & i & 0 \\
0 & 0 & i
\end{array}
\right) \nonumber\\
&& w \,=\, {\rm i} \, e^\lambda \, + \, \psi \label{immerO6}
\end{eqnarray}
The pull-back of the lagrangian is the following one:
\begin{equation}\label{orbittasesta}
\phi^\star \mathcal{L} \, = \, \left( \partial^\mu\psi\, \partial^\mu \psi \, e^{-2\lambda} \, + \, \partial^\mu\lambda\, \partial^\mu \lambda\right) \, - \, \frac{1}{4}\, g^2 \, \left( e^{\lambda}\, + \, \kappa\right)^2
\end{equation}
The pull-back of the scalar field equations yields a consistent system coinciding with the equations derived from the pull-back of the lagrangian (\ref{orbittasesta}). So we have a consistent truncation and we reobtain the Starobinsky--like model discussed in the previous section with $q=1$. It corresponds to $\alpha \, = \, \frac{1}{3}$. Equation (\ref{immerO6}) can be summarized as:
\begin{equation}\label{holomorph6}
z_2 \, = \, z_3 \, = \, z_5 \, = \, 0 \quad ; \quad z_1 \, = \, w \quad ; \quad z_4 \, = \, z_6 \, = \, {\rm i}
\end{equation}
The branching of the $\mathbf{14}^\prime$ dimensional representation under this $\mathfrak{sl}(2)$ subalgebra is the following one:
\begin{equation}\label{frikandello6}
\mathbf{14}^\prime \, \stackrel{\mathfrak{sl}(2,\mathbb{R})}{\longrightarrow} \, 5 \, \times \,\left( j\, =\, \frac{1}{2}\right) \,
\oplus \, 4 \, \times \, \left( j\, =\, 0\right)
\end{equation}
\subsubsection{\sc Conclusion of the above discussion}
This concludes our preliminary study of the orbits and shows that the embedded Starobinsky-like models described in section \ref{generalonuovo} exhaust the list of possible embeddings, the values of $\alpha\, = \, 1, \frac{2}{3} ,\frac{1}{3}$ being, apparently the only admissible ones. Next let us observe that the branching rules of the $\mathbf{14}^\prime$ dimensional representation which lead to consistent truncations, namely, (\ref{frikandello4},\ref{frikandello5},\ref{frikandello6}) are the only possible ones that we can obtain by embedding:
\begin{equation}\label{grullonevero}
\mathfrak{sl}(2) \, \mapsto \, \mathfrak{sl}(2) \, \times \, \mathfrak{sl}(2) \,\times \, \mathfrak{sl}(2)
\end{equation}
if the considered $\mathbf{14}^\prime$ representation of $\mathfrak{sl}(2) \, \times \, \mathfrak{sl}(2) \,\times \, \mathfrak{sl}(2)$ is the following one:
\begin{equation}\label{cannoleccio}
\mathbf{14}^\prime \, = \, \left(\frac{1}{2} \, , \, \frac{1}{2} \, , \, \frac{1}{2}\right)\, \oplus \, \left(\frac{1}{2},0,0\right) \, \oplus \,
\left(0, \frac{1}{2},0\right) \, \oplus \, \left(0, 0, \frac{1}{2}\right)
\end{equation}
This has a profound meaning. It implies that the only consistent truncations occur when the $\mathfrak{sl}(2)$ Lie algebra is embedded in the \textit{sub-Tits-Satake} Lie algebra, which as we discuss in the conclusive part is universal for all $\mathcal{N}=2$ models. This allows us to make the bold statement that the only values of $\alpha$ one can obtain form the gauging of hypermultiplet isometries in any supergravity theory based on symmetric manifolds is just $\alpha \, = \, 1, \frac{2}{3} , \, \frac{1}{3}$.
\newpage
\part{\sc Conclusions and perspectives}
\par
Considering the structure of the $c$-map and the results relative to the inclusion of Starobinsky like potentials that we have concretely obtained in the case of the $\mathrm{Sp(6,\mathbb{R})}$-model, we interpret them within the framework provided by Tits-Satake subalgebras and Tits Satake universality classes. This allows us to advocate that the mechanism underlying the generation of such cosmological potentials is universal and the prediction on the possible values of $\alpha$ equally general. Actually the analysis we are going to present in the next section supports the following general conclusion: the cosmological potentials classified in \cite{pietrosergiosasha1}, that follow from the gauging of a generator respectively elliptic, hyperbolic or parabolic of a constant curvature K\"ahler surface admit a universal uplifting to all $\mathcal{N}=2$ models based on symmetric spaces for the hypermultiplets, provided the curvature of the K\"ahler surface is duely quantized. Indeed there exist consistent one-field truncation only if the gauging occurs inside the universal \textit{sub-Tits-Satake} Lie algebra.
\section{\sc The Tits Satake projection}
In most cases of lower supersymmetry, neither the algebra $\mathbb{U}_{\mathcal{SK}}$ nor the
algebra $\mathbb{U}_{\mathcal{QM}}$ are \textbf{maximally split}. In short
this means that the non-compact rank $r_{nc} < r $ is less than
the rank of $\mathbb{U}$, namely not all the Cartan generators are
non-compact. Rigorously $r_{nc}$ is defined as follows:
\begin{equation}
r_{nc}\, = \, \mbox{rank} \left( \mathrm{U/H}\right) \, \equiv \, \mbox{dim} \,
\mathcal{H}^{n.c.} \quad ; \quad \mathcal{H}^{n.c.} \, \equiv \,
\mbox{CSA}_{\mathbb{U}(\mathbb{C})} \, \bigcap \, \mathbb{K}
\label{rncdefi}
\end{equation}
When this happens it means that the structure of both black-hole-like and cosmological-like solutions of supergravity is effectively determined by a
\textit{maximally split subalgebra} $\mathbb{U}^{TS} \subset
\mathbb{U}$ named the \textit{Tits Satake} subalgebra of
$\mathbb{U}$, whose rank is equal to $r_{nc}$. Effectively determined
does not mean that solutions of the big system
coincide with those
of the smaller system rather it means that the former can be obtained from the
latter by means of rotations of a compact subgroup $\mathrm{G_{paint} \subset U}$ of the big group
which we name the \textit{paint group}, for whose precise definition we refer the reader to \cite{titsusataku}.
Here we just emphasize few important facts, relevant for our goals. To this effect we recall that the Tits Satake algebra is obtained from the
original algebra via a projection of the root system of $\mathbb{U}$ onto the subspace orthogonal to the compact part of the Cartan subalgebra
of $\mathbb{U}^{TS}$:
\begin{equation}
\Pi^{TS} \quad ; \quad \Delta_\mathbb{U} \,\mapsto \,
\overline{\Delta}_{\mathbb{U}^{TS}}
\label{Tsproj}
\end{equation}
In euclidian geometry $\overline{\Delta}_{\mathbb{U}^{TS}}$ is just a collection of vectors in $r_{nc}$ dimensions; a priori there is no reason why it should be the root system of another Lie algebra. Yet in almost all cases, $\overline{\Delta}_{\mathbb{U}^{TS}}$ turns out to be a Lie algebra root system and the maximal split Lie algebra corresponding to it, $\mathbb{U}^{TS}$, is, by definition, the Tits Satake subalgebra of the original non maximally split Lie algebra: $\mathbb{U}^{TS} \subset \mathbb{U}$. Such algebras $\mathbb{U}$ are called \textit{non-exotic}. The \textit{exotic} non compact algebras are those for which the system $\overline{\Delta}_{\mathbb{U}^{TS}}$ is not an admissible root system. In such cases there is no Tits Satake subalgebra $\mathbb{U}^{TS}$. Exotic subalgebra are very few and in supergravity they appear only in three instances that display additional peculiarities relevant for the black-hole and cosmological solutions. As for $\mathcal{N}=2$ theories the only exotic homogeneous symmetric Special K\"ahler manifolds are those of the Minimal Coupling series discussed in section \ref{minicoup}. Exotic are also the Quaternionic K\"ahler manifolds in the $c$-map image of the former.
\par
For the non exotic models we have that the decomposition (\ref{gendecompo}) commutes with the projection,
namely:
\begin{equation}
\begin{array}{rcl}
\mbox{adj}(\mathbb{U}_{\mathcal{QM}}) &=&
\mbox{adj}(\mathbb{U}_{\mathcal{SK}})\oplus\mbox{adj}({\mathfrak{sl}(2,\mathbb{R})_E})\oplus
\mathbf{W}_{(2,W)} \\
\null &\Downarrow & \null\\
\mbox{adj}(\mathbb{U}^{TS}_{\mathcal{QM}}) &=&
\mbox{adj}(\mathbb{U}^{TS}_{\mathcal{SK}})\oplus\mbox{adj}({\mathfrak{sl}(2,\mathbb{R})_E})\oplus
\mathbf{W}_{(2,W^{TS})} \\
\end{array}
\label{gendecompo2}
\end{equation}
In other words the projection leaves the $A_1$ Ehlers subalgebra untouched and has a non trivial effect only
on the algebra $\mathbb{U}_{\mathcal{SK}}$. Furthermore the image under the projection of the highest root
of $\mathbb{U}$ is the highest root of $\mathbb{U}^{TS}$:
\begin{equation}
\Pi^{TS} \quad : \quad \psi \, \rightarrow \, \psi^{TS}
\label{Pionpsi}
\end{equation}
The reason why the Tits Satake projection is relevant to us is that the classification of nilpotent orbits (standard triples) and hence of abelian gaugings depends only on the Tits Satake subalgebra and therefore is universal for all members of the same Tits Satake universality class. By this name we mean all algebras that share the same Tits Satake projection. A similar property was extensively used in \cite{noinilpotenti} in the discussion of extremal black-hole solutions. Indeed the classification of extremal black-holes also boils down to the classification of nilpotent orbits, so that the mathematical problem at stake is just the same.
\par
Having clarified these points we can proceed with the classification of homogeneous symmetric spaces relevant to $\mathcal{N}=2$ supergravity either in the vector multiplet sector (Special K\"ahler) or in the hypermultiplet sector (Quaternionic K\"ahler). These spaces are listed in table \ref{homomodelTS}.
\begin{table}
\begin{center}
{\tiny
\begin{tabular}{|l|c|c||c|c||c|c||c||}
\hline
\null & TS & TS & coset &coset & Paint & subP & susy\\
$\#$ & $\mathcal{SK}_{TS}$ & $\mathcal{QM}_{TS}$ & $\mathcal{SK}$ & $\mathcal{QM}$ & Group & Group & \\
\hline
\hline
1 &\null & \null & \null & \null &\null & \null & \null \\
\null & $ \frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}$ & $ \frac{\mathrm{G_{2(2)}}}{\mathrm{SU(2)\times SU(2)}}$ & $ \frac{\mathrm{SU(1,1)}}{\mathrm{U(1)}}$ & $ \frac{\mathrm{G_{2(2)}}}{\mathrm{SU(2)\times SU(2)}}$ & $1$ & $1$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & n=1 \\
\hline
\hline
2 &\null & \null & \null & \null &\null & \null & \null \\
\null & \null & \null & $ \frac{\mathrm{Sp(6,R)}}{\mathrm{SU(3)\times U(1)}}$ & $ \frac{\mathrm{F_{4(4)}}}{\mathrm{USp(6)\times SU(2)}}$ & $1$ & $1$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & $n=6$ \\
\cline{1-1} \cline{4-8}
3 &\null & \null & \null & \null &\null & \null & \null \\
\null & \null & \null & $ \frac{\mathrm{SU(3,3)}}{\mathrm{SU(3)\times SU(3) \times U(1)}}$ & $ \frac{\mathrm{E_{6(2)}}}{\mathrm{SU(6)\times SU(2)}}$ & $\mathrm{SO(2)\times SO(2)}$ & $1$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & $n=9$ \\
\cline{1-1} \cline{4-8}
4 &\null & \null & \null & \null &\null & \null & $\mathcal{N}=6$ \\
\null & $ \frac{\mathrm{Sp(6,R)}}{\mathrm{SU(3)\times U(1)}}$ & $ \frac{\mathrm{F_{4(4)}}}{\mathrm{Sp(6,R)\times SL(2,R)}}$ & $ \frac{\mathrm{SO^\star(12)}}{\mathrm{SU(6)\times U(1)}}$ & $ \frac{\mathrm{E_{7(-5)}}}{\mathrm{SO(12)\times SU(2)}}$ & $\mathrm{SO(3)\times SO(3)}$ & $\mathrm{SO(3)_{d}}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &$\mathrm{\times SO(3)}$ & \null & n=15 \\
\cline{1-1} \cline{4-8}
5 &\null & \null & \null & \null &\null & \null & \null \\
\null & \null & \null & $ \frac{\mathrm{E_{7(-25)}}}{\mathrm{E_{6(-78)} \times U(1)}}$ & $ \frac{\mathrm{E_{8(-24)}}}{\mathrm{E_{7(-133)}\times SU(2)}}$ & $\mathrm{SO(8)}$ & $G_{2(2)}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & $n=27$ \\
\hline
\hline
\hline
6 & \null \null & \null & \null & \null &\null & \null & \null \\
\null & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,1)}}{\mathrm{SO(2)}}$ & $ \frac{\mathrm{SO(4,3)}}{\mathrm{SO(4)\times SO(3)}}$ & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,1)}}{\mathrm{SO(2)}}$ & $ \frac{\mathrm{SO(4,3)}}{\mathrm{SO(4)\times SO(3)}}$ & $\mathrm{1}$ & $\mathrm{1}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & n=2 \\
\hline
7 & \null \null & \null & \null & \null &\null & \null & \null \\
\null & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,2)}}{\mathrm{SO(2)\times SO(2)}}$ & $ \frac{\mathrm{SO(4,4)}}{\mathrm{SO(4)\times SO(4)}}$ & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,2)}}{\mathrm{SO(2)\times SO(2)}}$ & $ \frac{\mathrm{SO(4,4)}}{\mathrm{SO(4)\times SO(4)}}$ & $\mathrm{1}$ & $\mathrm{1}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & n=3 \\
\hline
8 & \null \null & \null & \null & \null &\null & \null & \null \\
\null & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,3)}}{\mathrm{SO(2)\times SO(3)}}$ & $ \frac{\mathrm{SO(4,5)}}{\mathrm{SO(4)\times SO(5)}}$ & $ \frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\times\frac{\mathrm{SO(2,2+p)}}{\mathrm{SO(2)\times SO(2+p)}}$ & $ \frac{\mathrm{SO(4,4+p)}}{\mathrm{SO(4)\times SO(4+p)}}$ & $\mathrm{SO(p)}$ & $\mathrm{SO(p-1)}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null & \null & n=3+p \\
\hline
\hline
\hline
$\mathrm{exot}$&\null & \null & \null & \null &\null & \null & \null \\
\null \null & $ bc_1$ & $ bc_2$ & $ \frac{\mathrm{SU(p+1,1)}}{\mathrm{SU(p+1)\times U(1)}}$ & $ \frac{\mathrm{SU(p+2,2)}}{\mathrm{SU(p+1,1)\times SL(2,R)_{h^\star}}}$ & $\mathrm{U(1)\times U(1) \times U(p)}$ & $\mathrm{U(p-1)}$ & $\mathcal{N}=2$ \\
\null &\null & \null & \null & \null &\null& \null & n=p+1 \\
\hline
\end{tabular}
}
\caption{The first eight rows of this table list the \textit{non-exotic} homogenous symmetric Special K\"ahler manifolds. They are displayed together with their Quaternionic K\"ahler $c$-map images and are organized in five Tits Satake universality classes. Non exotic means that the Tits Satake projection of the root system is a standard Lie Algebra root system. Of the five universality classes three contain only one maximally split element, one contains four elements filling the Tits magic square, while the last class contains an infinite number of elements. Within each class the models are distinguished by the different structure of the Paint Group and of its subPaint subgroup. The Paint group is a $c$-map invariant. It is the same in the Special K\"ahler and in the Quaternionic K\"ahler case. The last line displays the unique family of exotic Special K\"ahler symmetric spaces for which the Tits Satake projection of the root system is not a root system. They correspond to the Minimal Coupling Models discussed in the main text. Their $c$-map images of the exotic models are also exotic. Notwithstanding this anomaly the concept of Paint Group, according to its definition as group of external automorphisms of the solvable Lie algebra generating the non compact coset manifold still exists. The Paint group is the same for the K\"ahler and the quaternionic K\"ahler case as in the non exotic cases. \label{homomodelTS}}
\end{center}
\end{table}
In table \ref{homomodelTS} we have also listed the Paint groups and the subpaint groups. These latter are always compact and their different structures is what distinguishes the different elements belonging to the same class. As it was shown in \cite{titsusataku}, these groups are dimensional reduction invariant and therefore $c$-map invariant, namely they are the same in $\mathrm{U}_{\mathcal{SK}}$ and in $\mathrm{U}_{\mathcal{QM}}$. Hence the representation $\mathbf{W}$ which, as we have seen, hosts the symplectic section of Special Geometry and regulates, by means of its branching, the existence or non existence of consistent truncations, can be decomposed with respect to the Tits Satake subalgebra and the Paint group revealing a regularity structure inside each Tits Satake universality class which is what allows us to draw general conclusions and make universal predictions. In the case of black-holes the same Tits-Satake decomposition of the $\mathbf{W}$-representation is at the heart of the classification of \textit{charge orbits} of the hole, as we extensively discussed in \cite{noinilpotenti}.
\section{\sc Tits Satake Universality classes and the embedding of Starobinsky-like models}
In the present section we consider the decomposition of the $\mathbf{W}$-representations with respect to Tits-Satake subalgebras and Paint groups for all the non-exotic models.
\par
In \cite{titsusataku} the \textit{paint algebra} was defined as the algebra of external automorphisms of the solvable Lie algebra $\mathop{\rm Solv}\nolimits_\mathcal{M}$ generating the non-compact symmetric space: $\mathcal{M}\, = \, \mathrm{U/H}$, namely
\begin{equation}
\mathbb{G}_{\mathrm{paint}} \, = \, \mathrm{Aut}_{\mathrm{Ext}} \, \left[
\mathop{\rm Solv}\nolimits_\mathcal{M}\right]. \label{pittureFuori}
\end{equation}
where:
\begin{equation}
\mathrm{Aut}_{\mathrm{Ext}} \, \left[ \mathop{\rm Solv}\nolimits_\mathcal{M}\right] \,
\equiv \, \frac{\mathrm{Aut} \, \left[
\mathop{\rm Solv}\nolimits_\mathcal{M}\right]}{\mathop{\rm Solv}\nolimits_\mathcal{M}},
\label{outerauto}
\end{equation}
Given the paint algebra $\mathbb{G}_{\mathrm{paint}} \, \subset \, \mathbb{U}$ and the Tits Satake subalgebra $\mathbb{G}_{\mathrm{TS}}\, \subset \, \mathbb{U}$, whose construction we have briefly recalled above, following \cite{titsusataku}
one introduces the \textit{sub Tits Satake} and \textit{sub paint} algebras as the centralizers of the paint algebra and of the Tits Satake algebra, respectively. In other words we have:
\begin{equation}\label{subTs}
\mathfrak{s} \, \in \, \mathbb{G}_{\mathrm{subTS}} \, \subset \, \mathbb{G}_{\mathrm{TS}}\, \subset \, \mathbb{U} \quad \Leftrightarrow \, \quad \left[ \mathfrak{s} \, , \, \mathbb{G}_{\mathrm{paint}}\right] \, = \, 0
\end{equation}
and
\begin{equation}\label{subpaint}
\mathfrak{t} \, \in \, \mathbb{G}_{\mathrm{subpaint}} \, \subset \, \mathbb{G}_{\mathrm{paint}}\, \subset \, \mathbb{U} \quad \Leftrightarrow \, \quad \left[ \mathfrak{t} \, , \, \mathbb{G}_{\mathrm{TS}}\right] \, = \, 0
\end{equation}
A very important property of the paint and subpaint algebras is that they are conserved int the $c$-map, namely they are the same for $\mathbb{U}_{\mathcal{SK}}$ and $\mathbb{U}_{\mathcal{QM}}$.
\par
In the next lines we analyze the decomposition of the $\mathbf{W}$-representations with respect to these subalgebras for each Tits Satake universality class of non maximally split models. In the case of maximally split models there is no paint algebra and there is nothing with respect to which to decompose.
\subsection{\sc Universality class $\mathfrak{sp}(6,\mathbb{R})\Rightarrow \mathfrak{f}_{4(4)}$}
In this case the sub Tits Satake Lie algebra is
\begin{equation}\label{cuffio}
\mathbb{G}_{\mathrm{subTS}} \, = \, \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{sl}(2,\mathbb{R}) \subset \mathfrak{sp}(6,\mathbb{R}) \, = \, \mathbb{G}_{\mathrm{TS}}
\end{equation}
and the $\mathbf{W}$-representation of the maximally split model decomposes as follows:
\begin{equation}\label{scompo14}
\mathbf{14}^\prime \, \stackrel{\mathbb{G}_{\mathrm{subTS}}}{\Longrightarrow} \, (\mathbf{2},\mathbf{1},\mathbf{1}) \oplus (\mathbf{1},\mathbf{2},\mathbf{1}) \oplus (\mathbf{1},\mathbf{1},\mathbf{2}) \oplus (\mathbf{2},\mathbf{2},\mathbf{2})
\end{equation}
This decomposition combines in the following way with the paint group representations in the various models belonging to the same universality class.
\subsubsection{\sc $\mathfrak{su}(3,3)$ model}
For this case the paint algebra is
\begin{equation}\label{furadino}
\mathbb{G}_{\mathrm{paint}} \, = \, \mathfrak{so}(2) \oplus \mathfrak{so}(2)
\end{equation}
and the $\mathbf{W}$-representation is the $\mathbf{20}$ dimensional of $\mathfrak{su}(3,3)$ corresponding to an antisymmetric tensor with a reality condition of the form:
\begin{equation}\label{realata}
t_{\alpha\beta\gamma}^\star = \frac{1}{3!} \, \epsilon_{\alpha\beta\gamma\delta\eta\theta} \, t_{\delta\eta\theta}
\end{equation}
The decomposition of this representation with respect to the Lie algebra $ \mathbb{G}_{\mathrm{paint}}\oplus {\mathbb{G}_{\mathrm{subTS}}}$ is the following one:
\begin{equation}\label{ruppatoA}
\mathbf{20} \, \stackrel{\mathbb{G}_{\mathrm{paint}} \oplus \mathbb{G}_{\mathrm{subTS}}}{\Longrightarrow} \, (2,q_1|\mathbf{2},\mathbf{1},\mathbf{1}) \oplus (2,q_2|\mathbf{1},\mathbf{2},\mathbf{1}) \oplus (2,q_3|\mathbf{1},\mathbf{1},\mathbf{2}) \oplus (1,0|\mathbf{2},\mathbf{2},\mathbf{2})
\end{equation}
where $(2,q)$ means a doublet of $\mathfrak{so}(2)\oplus\mathfrak{so}(2)$ with a
certain grading $q$ with respect to the generators, while $(1,0)$
means the singlet that has $0$ grading with respect to both
generators. The subpaint algebra in this case is $
\mathbb{G}_{\mathrm{subpaint}}\, = \,0$ and the decomposition of
the same $\mathbf{W}$-representation with respect to
$\mathbb{G}_{\mathrm{subpaint}} \oplus \mathbb{G}_{\mathrm{TS}}$
is:
\begin{equation}\label{ruppatoB}
\mathbf{20} \, \stackrel{\mathbb{G}_{\mathrm{subpaint}} \oplus \mathbb{G}_{\mathrm{TS}}}{\Longrightarrow} \, \mathbf{6} \, \oplus \, \mathbf{14}
\end{equation}
This follows from the decomposition of the $\mathbf{6}$ of $\mathfrak{sp}(6,\mathbf{R})$ with respect to the sub Tits Satake algebra (\ref{cuffio}):
\begin{equation}\label{seifracco}
\mathbf{6} \, \stackrel{\mathbb{G}_{\mathrm{subTS}}}{\Longrightarrow} \, (\mathbf{2},\mathbf{1},\mathbf{1}) \oplus (\mathbf{1},\mathbf{2},\mathbf{1}) \oplus (\mathbf{1},\mathbf{1},\mathbf{2})
\end{equation}
\subsubsection{\sc $\mathfrak{so}^\star(12)$ model}
For this case the paint algebra is
\begin{equation}\label{furadinotwo}
\mathbb{G}_{\mathrm{paint}} \, = \, \mathfrak{so}(3) \oplus \mathfrak{so}(3) \oplus \mathfrak{so}(3)
\end{equation}
and the $\mathbf{W}$-representation is the $\mathbf{32}_s $
dimensional spinorial representation of $\mathfrak{so}^\star(12)$. The
decomposition of this representation with respect to the Lie
algebra $ \mathbb{G}_{\mathrm{paint}}\oplus
{\mathbb{G}_{\mathrm{subTS}}}$ is the following one:
\begin{equation}\label{ruppatoAtwo}
\mathbf{32}_s \, \stackrel{\mathbb{G}_{\mathrm{paint}} \oplus \mathbb{G}_{\mathrm{subTS}}}{\Longrightarrow} \, (\underline{2},\underline{2},\underline{1}|\mathbf{2},\mathbf{1},\mathbf{1}) \oplus (\underline{2},\underline{1},\underline{2}|\mathbf{1},\mathbf{2},\mathbf{1}) \oplus (\underline{1},\underline{1},\underline{2}|\mathbf{1},\mathbf{1},\mathbf{2}) \oplus (\underline{1},\underline{1},\underline{1}|\mathbf{2},\mathbf{2},\mathbf{2})
\end{equation}
where $\underline{2}$ means the doublet spinor representation of
$\mathfrak{so}(3)$. The subpaint algebra in this case is $
\mathbb{G}_{\mathrm{paint}}\, = \,\mathfrak{so}(3)_{\mathrm{diag}}$ and the
decomposition of the same $\mathbf{W}$-representation with respect
to $\mathbb{G}_{\mathrm{subpaint}} \oplus
\mathbb{G}_{\mathrm{TS}}$ is:
\begin{equation}\label{ruppatoBtwo}
\mathbf{32}_s \, \stackrel{\mathbb{G}_{\mathrm{TS}}\oplus \mathbb{G}_{\mathrm{subpaint}}}{\Longrightarrow} \, (\mathbf{6}|\underline{3}) \, \oplus \, (\mathbf{14}^\prime|\underline{1})
\end{equation}
This follows from the decomposition of the product $\underline{2} \times \underline{2}$ of $\mathfrak{so}(3)_{\mathrm{diag}}$ times the Tits Satake algebra (\ref{cuffio}):
\begin{equation}\label{seifraccotwo}
\underline{2} \times \underline{2} \, = \, \underline{3} \, \oplus \, \underline{1}
\end{equation}
\subsubsection{\sc $\mathfrak{e}_{7(-25)}$ model}
For this case the paint algebra is
\begin{equation}\label{furadinothree}
\mathbb{G}_{\mathrm{paint}} \, = \, \mathfrak{so}(8)
\end{equation}
and the $\mathbf{W}$-representation is the fundamental $\mathbf{56} $ dimensional representation of $\mathfrak{e}_{7(-25)}$
The decomposition of this representation with respect to the Lie algebra $ \mathbb{G}_{\mathrm{paint}}\oplus {\mathbb{G}_{\mathrm{subTS}}}$ is the following one:
\begin{equation}\label{ruppatoAthree}
\mathbf{56} \, \stackrel{\mathbb{G}_{\mathrm{paint}} \oplus \mathbb{G}_{\mathrm{subTS}}}{\Longrightarrow} \, (\mathbf{8}_v|\mathbf{2},\mathbf{1},\mathbf{1}) \oplus (\mathbf{8}_s|\mathbf{1},\mathbf{2},\mathbf{1}) \oplus (\mathbf{8}_c|\mathbf{1},\mathbf{1},\mathbf{2}) \oplus (\mathbf{1}|\mathbf{2},\mathbf{2},\mathbf{2})
\end{equation}
where $\mathbf{8}_{v,s,c}$ are the three inequivalent eight-dimensional representations of $\mathfrak{so}(8)$, the vector, the spinor and the conjugate spinor. The subpaint algebra in this case is $ \mathbb{G}_{\mathrm{paint}}\, = \,\mathfrak{g}_{2(-14)}$ with respect to which all three $8$-dimensional representations of $\mathfrak{so}(8)$ branch as follows:
\begin{equation}\label{seifraccothree}
\mathbf{8}_{v,s,c} \, \stackrel{\mathfrak{g}_{2(-14)}}{\Longrightarrow} \, \mathbf{7 }\, \oplus \, \mathbf{1}
\end{equation}
In view of this the decomposition of the same $\mathbf{W}$-representation with respect to $\mathbb{G}_{\mathrm{subpaint}} \oplus \mathbb{G}_{\mathrm{TS}}$ is:
\begin{equation}\label{ruppatoBthree}
\mathbf{56} \, \stackrel{\mathbb{G}_{\mathrm{TS}} \oplus \mathbb{G}_{\mathrm{subpaint}}}{\Longrightarrow} \, (\mathbf{6}|\mathbf{7}) \, \oplus \, (\mathbf{14}^\prime|\mathbf{1})
\end{equation}
\subsection{\sc Universality class $\mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(2,3) \Rightarrow \mathfrak{so}(4,5)$}
This case corresponds to one of the possible infinite families of $\mathcal{N}=2$ theories with a symmetric homogeneous special K\"ahler manifold and a number of vector multiplets larger than three ($n=3+p$). The other infinite family corresponds instead to one of the three exotic models.
\par
The generic element of this infinite class corresponds to the following algebras:
\begin{eqnarray}
\mathbb{U}_{\mathcal{SK}} &=& \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(2,2+p)\nonumber\\
\mathbb{U}_{\mathcal{QM}} &=& \mathfrak{so}(4,4+p) \label{feldane}
\end{eqnarray}
In this case the sub Tits Satake algebra is:
\begin{equation}\label{casilino1}
\mathbb{G}_{\mathrm{subTS}} \, = \, \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{sl}(2,\mathbb{R}) \, \simeq \, \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(2,2) \, \, \subset \, \mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(2,3) \, = \, \mathbb{G}_{\mathrm{TS}}
\end{equation}
an the paint and subpaint algebras are as follows:
\begin{eqnarray}
\mathbb{G}_{\mathrm{paint}} &=& \mathfrak{so}(p) \nonumber\\
\mathbb{G}_{\mathrm{subpaint}} &=& \mathfrak{so}(p-1) \label{felanina1}
\end{eqnarray}
The symplectic $\mathbf{W}$ representation of $\mathbb{U}_{\mathcal{SK}}$ is the tensor product of the fundamental representation of $\mathfrak{sl}(2)$ with the fundamental vector representation of $\mathfrak{so}(2,2+p)$, namely
\begin{equation}\label{erast1}
\mathbf{W} \, = \, \left( \mathbf{2 | 4}+p\right) \quad ; \quad \mbox{dim} \,\mathbf{W} \, = \, 8+2p
\end{equation}
The decomposition of this representation with respect to $\mathbb{G}_{\mathrm{subTS}} \oplus \mathbb{G}_{\mathrm{subpaint}}$ is the following one:
\begin{equation}\label{decompusOne}
\mathbf{W} \, \stackrel{\mathbb{G}_{\mathrm{subTS}} \oplus \mathbb{G}_{\mathbf{subpaint}}}{\Longrightarrow} \, \left(\mathbf{2,2,2|1}\right) \oplus \left(\mathbf{2,1,1|1}\right) \oplus \left(\mathbf{2,1,1}|p-1\right)
\end{equation}
where $\mathbf{2,2,2}$ denotes the tensor product of the three fundamental representations of $\mathfrak{sl}(2,\mathbb{R})^3$. Similarly $\mathbf{2,1,1}$ denotes the doublet of the first $\mathfrak{sl}(2,\mathbb{R})$ tensored with the singlets of the following two $\mathfrak{sl}(2,\mathbb{R})$ algebras. The representations appearing in (\ref{decompusOne}) can be grouped in order to reconstruct full representations either of the complete Tits Satake or of the complete paint algebras. In this way one obtains:
\begin{eqnarray}
\mathbf{W} & \stackrel{\mathbb{G}_{\mathrm{subTS}} \oplus \mathbb{G}_{\mathbf{paint}}}{\Longrightarrow} & \left(\mathbf{2,2,2|1}\right) \oplus \left(\mathbf{2,1,1}|p+1\right)\nonumber\\
\mathbf{W} & \stackrel{\mathbb{G}_{\mathrm{TS}} \oplus \mathbb{G}_{\mathbf{subpaint}}}{\Longrightarrow} & \left(\mathbf{2,5|1}\right) \oplus \left(\mathbf{2,1}|p-1\right)
\end{eqnarray}
\subsection{\sc $\mathbf{W}$-representations of the maximally split non exotic models}
In the previous subsections we have analyzed the Tits-Satake decomposition of the $\mathbf{W}$-representation for all those models that are non maximally split. The remaining models are the maximally split ones for which there is no paint algebra and the Tits Satake projection is the identity map. There are essentially four type of models:
\begin{enumerate}
\item The $\mathrm{SU(1,1)}$ non exotic model where the $\mathbf{W}$-representation is the $j=\ft 32$ of $\mathfrak{so}(1,2)\sim \mathfrak{su}(1,1)$
\item The $\mathrm{Sp(6,\mathbb{R})}$ model where the $\mathbf{W}$-representation is the $\mathbf{14}^\prime$ (antisymmetric symplectic traceless three-tensor).
\item The models $\mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(q,q)$ where the $\mathbf{W}$-representation is the $\left(\mathrm{2,2q}\right)$, namely the tensor product of the two fundamentals.
\item The models $\mathfrak{sl}(2,\mathbb{R}) \oplus \mathfrak{so}(q,q+1)$ where the $\mathbf{W}$-representation is the $\left(\mathrm{2,2q+1}\right)$, namely the tensor product of the two fundamentals.
\end{enumerate}
Therefore, for the above maximally split models, we need the classification of $\mathrm{U_{\mathcal{SK}}}$ orbits in the mentioned $\mathbf{W}$-representations. Actually such orbits are sufficient also for the non maximally split models. Indeed each of the above $4$-models correspond to one Tits Satake universality class
and, within each universality class, the only relevant part of the $\mathbf{W}$-representation is the subpaint group singlet which is universal for all members of the class. This is precisely what we verified in the previous subsections.
\par
For instance for all members of the universality class of $\mathrm{Sp(6,\mathbb{R})}$, the $\mathbf{W}$-representation splits as follows with respect to the subalgebra $\mathfrak{sp}(6,\mathbb{R})\oplus \mathbb{G}_{\mathrm{subpaint}}$:
\begin{equation}\label{gribochky}
\mathbf{W} \, \stackrel{\mathfrak{sp}(6,\mathbb{R})\oplus \mathbb{G}_{\mathrm{subpaint}}}{\Longrightarrow} \, \left( \mathbf{6}\, | \, \mathcal{D}_{\mathrm{subpaint}}\right) \, + \, \left( \mathbf{14}^\prime \, | \, \mathbf{1}_{\mathrm{subpaint}}\right)
\end{equation}
where the representation $\mathcal{D}_{\mathrm{subpaint}}$ is the following one for the three non-maximally split members of the class:
\begin{equation}\label{reppisubpitturi}
\mathcal{D}_{\mathrm{subpaint}} \, = \, \left\{ \begin{array}{ccccc}
\mathbf{1} & \mbox{of} & \mathbf{1} & \mbox{for the} & \mathfrak{su}(3,3)-\mbox{model} \\
\mathbf{3} & \mbox{of} & \mathfrak{so}(3) & \mbox{for the} & \mathfrak{so}^\star(12)-\mbox{model} \\
\mathbf{7} & \mbox{of} & \mathfrak{g}_{2(-14)} & \mbox{for the} & \mathfrak{e}_{7(-25)}-\mbox{model}\\
\end{array}\right.
\end{equation}
Clearly the condition:
\begin{equation}\label{condosputta}
\left( \mathbf{6}\, | \, \mathcal{D}_{\mathrm{subpaint}}\right)\, = \, 0
\end{equation}
imposed on a vector in the $\mathbf{W}$-representation breaks the group $\mathrm{U}_{\mathcal{SK}}$ to its Tits Satake subgroup. The key point is that each $\mathbf{W}$-orbit of the big group $\mathrm{U}_{\mathcal{SK}}$ crosses the locus (\ref{condosputta}) so that the classification of $\mathrm{Sp(6,\mathbb{R})}$ orbits in the $\mathbf{14}^\prime$-representation exhausts the classification of $\mathbf{W}$-orbits for all members of the universality class.
\par
In order to prove that the gauge (\ref{condosputta}) is always reachable it suffices to show that the representation $\left( \mathbf{6}\, | \,\mathcal{D}_{\mathrm{subpaint}}\right )$ always appears at least once in the decomposition of the Lie algebra $\mathbb{U}_{\mathcal{SK}}$ with respect to the subalgebra $\mathfrak{sp}(6,\mathbb{R})\oplus \mathbb{G}_{\mathrm{subpaint}}$. The corresponding parameters of the big group can be used to set to zero the projection of the $\mathbf{W}$-vector onto $\left( \mathbf{6}\, | \,\mathcal{D}_{\mathrm{subpaint}}\right )$.
\par
The required condition is easily verified since we have:
\begin{eqnarray}
\underbrace{\mbox{adj}\, \mathfrak{su}(3,3)}_{\mathbf{35}} &\stackrel{\mathfrak{sp}(6,\mathbb{R})}{\Longrightarrow} \,& \underbrace{\mbox{adj}\, \mathfrak{sp}(6,\mathbb{R})}_{\mathbf{21}} \, \oplus \, \mathbf{6} \, \oplus \, \mathbf{6} \, \oplus \, \mathbf{1} \, \oplus \, \mathbf{1}\nonumber \\
\underbrace{\mbox{adj}\, \mathfrak{so}^\star(12)}_{\mathbf{66}} &\stackrel{\mathfrak{sp}(6,\mathbb{R})\oplus \mathfrak{so}(3)}{\Longrightarrow} \,& \underbrace{\mbox{adj}\, \mathfrak{sp}(6,\mathbb{R})}_{\mathbf{21}} \, \oplus \, \underbrace{\mbox{adj}\, \mathfrak{so}(3)}_{\mathbf{3}}\,
\oplus \, \left(\mathbf{6} , \mathbf{3}\right)\, \oplus \, \left(\mathbf{6} , \mathbf{3}\right)\oplus \, \left(\mathbf{1},\mathbf{3}\right)\,\oplus \, \left(\mathbf{1},\mathbf{3}\right)\,\nonumber \\
\underbrace{\mbox{adj}\, \mathfrak{e}_{7(-25)}}_{\mathbf{133}} &\stackrel{\mathfrak{sp}(6,\mathbb{R})\oplus \mathfrak{g}_{2(-14)}}{\Longrightarrow} \,& \underbrace{\mbox{adj}\, \mathfrak{sp}(6,\mathbb{R})}_{\mathbf{21}} \, \oplus \, \underbrace{\mbox{adj}\, \mathfrak{g}_{2(-14)}}_{\mathbf{14}}\,
\oplus \, \left(\mathbf{6} , \mathbf{7}\right)\, \oplus \, \left(\mathbf{6} , \mathbf{7}\right)\oplus \, \left(\mathbf{1},\mathbf{7}\right)\,\oplus \, \left(\mathbf{1},\mathbf{7}\right)\,\nonumber \\
\end{eqnarray}
The reader cannot avoid being impressed by the striking similarity of the above decompositions which encode the very essence of Tits Satake universality. Indeed the representations of the common Tits Satake subalgebra appearing in the decomposition of the adjoint are the same for all members of the class. They are simply uniformly assigned to the fundamental representation of the subpaint algebra which is different in the three cases. The representation $\left( \mathbf{6}\, | \,\mathcal{D}_{\mathrm{subpaint}}\right )$ appears twice in these decompositions and can be used to reach the gauge (\ref{condosputta}) as we claimed above.
\paragraph{\sc Holomorphic consistent truncations} The next point to remark is that the condition (\ref{condosputta}) has another important interpretation if applied to the holomorphic section of special geometry. The key point is the following numerical identity valid for all members of the universality class:
\begin{equation}\label{curiosona}
\mbox{dim} \, \frac{\mathrm{U}_{\mathcal{SK}}}{\mathrm{H}_{\mathcal{SK}}}\, = \,
\mbox{dim} \, \frac{\mathrm{U}^{TS}_{\mathcal{SK}}}{\mathrm{H}^{TS}_{\mathcal{SK}}}\, \oplus \, 6 \, \times \, \mbox{dim}\, \mathcal{D}_{\mathrm{subpaint}}
\end{equation}
This means that if we decompose the symplectic section of the big group according to the Tits-Satake subalgebra and we impose on it the condition (\ref{condosputta}) we just obtain the right number of holomorphic constraints to project onto the submanifold $\frac{\mathrm{U}_{\mathcal{SK}}}{\mathrm{H}_{\mathcal{SK}}}$. At the level of field equations this is certainly a consistent truncation, since we project onto the singlets of the subpaint group.
\par
On the other hand if we decompose the $\mathbf{W}$-representation with respect to the sub-Tits-Satake subalgebra $\mathfrak{sl}(2)\times \mathfrak{sl}(2) \times \mathfrak{sl}(2)$ we have the branching rule:
\begin{equation}\label{fillopona}
\mathbf{W} \, \rightarrow \, \left(\mathcal{D}_1 |\mathbf{2},\mathbf{1},\mathbf{1}\right)\, \oplus \, \left(\mathcal{D}_2 |\mathbf{1},\mathbf{2},\mathbf{1}\right)\, \oplus \,\left(\mathcal{D}_3 |\mathbf{1},\mathbf{1},\mathbf{2}\right) \, \oplus \, \left(\mathbf{1} |\mathbf{2},\mathbf{2},\mathbf{2}\right)
\end{equation}
where $\mathcal{D}_{1,2,3}$ are three suitable representations of the Paint Group. Imposing on the symplectic section of the big model
the constraints:
\begin{eqnarray}
\left(\mathcal{D}_1 |\mathbf{2},\mathbf{1},\mathbf{1}\right)&=& 0 \nonumber\\
\left(\mathcal{D}_2 |\mathbf{1},\mathbf{2},\mathbf{1}\right) &=& 0 \nonumber\\
\left(\mathcal{D}_3 |\mathbf{1},\mathbf{1},\mathbf{2}\right) &=& 0
\end{eqnarray}
yields precisely the correct number of holomorphic constraints that restrict the considered Special K\"ahler manifold to the Special K\"ahler manifold of the STU-model namely to $\left(\frac{\mathrm{SL(2,\mathbb{R})}}{\mathrm{SO(2)}}\right)^3$. This follows from the numerical identity true for all members of the universality class:
\begin{equation}\label{guttallaxa}
\mbox{dim} \, \frac{\mathrm{U}_{\mathcal{SK}}}{\mathrm{H}_{\mathcal{SK}}}\, = \, \sum_{i=1}^3 \, 2 \, \times \, \mbox{dim} \, \mathcal{D}_i \, + \, 6
\end{equation}
The reason why the truncation to the STU-model is always a consistent truncation at the level of field equations is obvious in this set up. It corresponds to the truncation to the Paint Group singlets.
\paragraph{\sc $\mathbf{W}$-representations for the remaining models}
For the models of type $\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+p)$ having $\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+1)$ as Tits Satake subalgebra and $\mathfrak{so}(p-1)$ as subpaint algebra the decomposition of the $\mathbf{W}$-representation is the following one:
\begin{equation}\label{funghifritti}
\mathbf{W} \, = \, \left(\mathbf{2,2q+p}\right) \, \stackrel{\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+1)\oplus \mathfrak{so}(p-1)}{\Longrightarrow} \, \left(\mathbf{2,2q+1}|\mathbf{1}\right) \, \oplus \, \left(\mathbf{2,1}|\mathbf{p-1}\right)
\end{equation}
and the question is whether each $\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+p)$ orbit in the $\left(\mathbf{2,2q+p}\right)$ representation intersects the $\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+1)\oplus \mathfrak{so}(p-1)$-invariant locus:
\begin{equation}\label{fittone}
\left(\mathbf{2,1}|\mathbf{p-1}\right) \, = \,0
\end{equation}
The answer is yes since we always have enough parameters in the coset
\begin{equation}\label{ciabatta}
\frac{\mathrm{SL(2,\mathbb{R})}\times \mathrm{SO(q,q+p)}}{\mathrm{SL(2,\mathbb{R})\times SO(q,q+1)\times SO(p-1})}
\end{equation}
to reach the desired gauge (\ref{fittone}). Indeed let us observe the decomposition:
\begin{equation}\label{fruttodimare}
\mbox{adj}\, \left[\mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(q,q+p)\right] \, = \, \mbox{adj}\, \left[\mathfrak{sl}(2,\mathbb{R})\right] \oplus \mbox{adj}\, \left[\mathfrak{so}(q,q+1)\right] \, \oplus \mbox{adj}\, \left[\mathfrak{so}(p-1)\right] \, \oplus \, \left(\mathbf{1,2q+1 | p-1}\right)
\end{equation}
The $2q+1$ vectors of $\mathfrak{so}(p-1)$ appearing in (\ref{fruttodimare}) are certainly sufficient to set to zero the $2$ vectors of $\mathfrak{so}(p-1)$ appearing in $\mathbf{W}$.
\par
Relevant for the case of $\mathcal{N}\,=2\,$ supersymmetry is the value $q=2$ and in this case the sub-Tits-Satake Lie algebra is :
\begin{equation}\label{grilloparlante}
\mathbb{G}_{\mathrm{subTS}} \, = \, \mathfrak{sl}(2,\mathbb{R})\oplus \mathfrak{so}(2,2) \, = \, \mathfrak{sl}(2,\mathbb{R})\oplus\mathfrak{sl}(2,\mathbb{R})\oplus\mathfrak{sl}(2,\mathbb{R})
\end{equation}
namely it is once again the Lie algebra of the STU-model. Reduction to the STU-model is consistent for the same reason as in the other universality classes: it corresponds to truncation to Paint Group singlets.
\subsection{\sc Gaugings with consistent one-field truncations}
On the basis of the analysis presented in the previous section we arrive at the following conclusion. By gauging a nilpotent element of the isometry subalgebra of $\mathcal{SK}$ inside $\mathcal{QM}$ we generate a potential. The structure of the theory depends on the nilpotent orbit, namely on the embedding of an $\mathfrak{sl}(2)$ Lie algebra in $\mathbb{U}_{\mathcal{SK}}$ and there are many ways of doing this (the orbits), yet the gauged theory will admit a one-field truncation if and only if the $\mathfrak{sl}(2)$ is embedded into the sub Tits Satake Lie algebra:
\begin{equation}\label{gongolato}
\mathfrak{sl}(2) \, \mapsto \, \mathbb{G}_{\mathrm{subTS}}\, \subset \, \mathbb{U}_{\mathcal{SK}}
\end{equation}
There are only three different embeddings of $\mathfrak{sl}(2)$ into $\left( \mathfrak{sl}(2) \right)^3$ and these correspond to the three admissible values $\alpha\, = \, 1 , \, \frac{2}{3}, \, \frac{1}{3}$ in the Starobinsky-like model.
\section{\sc Conclusions}
In this paper we have analyzed in detail the structure of the $c$-map from Special K\"ahler manifolds to Quaternionic K\"ahler manifolds in connection with the abelian gaugings of hypermultiplet isometries in $\mathcal{N}=2$ supergravity and the generation of effective one field potentials that might describe inflaton dynamics.
\par
The motivations of such a study have been put forward in the introduction. Here we try to summarize the main results we have obtained and the perspectives for future investigations that have emerged.
\paragraph{\sc Results}
\begin{description}
\item[I] The $c$-map description of Quaternionic K\"ahler manifolds $\mathcal{QM}_{4n+4}$ allows to distribute the isometries of
$\mathcal{QM}_{4n+4}$ into three classes:
\begin{description}
\item[Ia)] The isometries associated with the Heisenberg algebra which exists in all cases, even when the internal Special K\"ahler manifold $\mathcal{SK}_{n}$ has no continuous symmetries as it might happen when we deal the moduli space of a Calabi-Yau three-fold and the Heisenberg fields are the Ramond--Ramond scalars of a superstring compactification. The gauging of such perturbative isometries produces one field potentials that are of the pure exponential type.
\item[Ib)] The isometries of the inner Special K\"ahler manifold $\mathcal{SK}_{n}$ that can always be promoted to isometries of the full Quaternionic manifold $\mathcal{QM}_{4n+4}$. Gauging these isometries one obtains a potential that at $\mathbf{Z}^\alpha \, =\, a \, = \, U \, = \, 0$ is identical with the potential obtained from the gauging of the same K\"ahler isometries in $\mathcal{N}=1$ supersymmetry. Hence one would like to be able to stabilize these fields. As for $\mathbf{Z}^\alpha \, =\, a \, = \, 0$ there is no problem. We can always truncate them. The field $U$ instead, that in superstring interpretations of the hypermultiplets can be identified with the coupling constant dilaton field, appears through exponentials all of the same sign in these perturbative gaugings and cannot be stabilized.
\item[Ic)] The non perturbative isometries that mix the Heisenberg symmetries with the K\"ahler symmetries. These exist only when $\mathcal{QM}_{4n+4}$ is a symmetric homogeneous space and the Lie algebra of the full isometry group takes the universal form (\ref{genGD3}). Gauging a compact non perturbative generator appears to be the only way of introducing exponentials with opposite sign of the field $U$ allowing for its stabilization.
\end{description}
\item[II] The Starobinsky-like potential:
\begin{equation}\label{bambolone2}
V_{Starobinsky} \, = \, \mbox{const} \, \times \, \left( 1\, - \, \exp\left [ - \, \sqrt{\frac{2}{3 \, \alpha}} \, \phi\right]\right)^2
\end{equation}
can be obtained universally from all homogeneous symmetric Quaternionic K\"ahler manifolds by means of an admissible embedding of a $\mathfrak{sl}(2,\mathbb{R})$ Lie algebra into the Special K\"ahler subalgebra $\mathbb{U}_\mathcal{SK}\,\subset \, \mathbb{U}_\mathcal{Q}$. The problem of embedding of $\mathfrak{sl}(2,\mathbb{R})$ algebras is the same as the problem of classifying standard triples and nilpotent algebras, yet the embedding must also be admissible, in the sense that it should allow for consistent one-field truncations. In the case of $\mathcal{SK} \, = \, \mathfrak{sp}(6,\mathbb{R})$ we exhausted the analysis of orbits and showed that admissible subalgebras correspond to embeddings of $\mathfrak{sl}(2)$ into the maximal subalgebra $\left(\mathfrak{sl}(2)\right)^3$ yielding $\alpha \, = \, 1, \, \frac{2}{3}, \, \frac{1}{3}$. By means of arguments based on Tits Satake universality classes, we have advocated that this result is general for all supergravity theories where the hypermultiplets are described by a homogeneous symmetric space.
\item[III] The above results rely on the use of the minimal coupling Special Geometry for the description of the vector multiplets. Staying within such a framework one can gauge by different vector fields the generators of a maximal set of mutually commuting $\mathfrak{sl}(2,\mathbb{R})$ Lie algebras and obtain as many massive fields. The number of massive vector fields appears therefore to be equal to the rank of the Tits Satake subalgebra $\mathbb{G}_{\mathrm{subTS}} \, \subset \, \mathbb{U}_\mathcal{SK}\,\subset \, \mathbb{U}_\mathcal{Q}$. The result of \cite{thesearch} where it was found that the maximal number of massive vector fields in this sort of gauging is one is confirmed by this general rule. In \cite{thesearch} we have $\mathbb{G}_{\mathrm{subTS}}\, =\, \mathfrak{sl}(2,\mathbb{R}) \, = \, \mathbb{U}_\mathcal{SK}\,\subset \, \mathbb{U}_\mathcal{Q} \, \equiv \, \mathfrak{g}_{(2,2)}$ and $\mbox{rank}\,\mathfrak{sl}(2,\mathbb{R})\, = \, 1$. In all other cases the rank of the sub Tits Satake Lie algebra is three.
\end{description}
\paragraph{\sc Perspectives and Generalizations}
In order to improve our understanding of the possible embedding of inflaton dynamics within larger unified theories, we consider necessary and feasible to explore the following directions.
\begin{description}
\item[A)] Enlarge the scope of gaugings of hypermultiplet isometries to non abelian algebras and consider the attractive situation where the Special K\"ahler manifold describing the vector multiplets and that, whose $c$-map provides the Quaternionic K\"ahler manifold description of the hypermultiplets are just the same or at least belong to the same class of homogeneous Special K\"ahler spaces.
\item[B] Utilize the above $\mathcal{N}=2$ set up to promote the embedding of inflaton dynamics from $\mathcal{N}=2$ to higher $\mathcal{N}$, in particular to $\mathcal{N}=3$ and $\mathcal{N}=4$.
\item[C] Consider such scenarios as superstring compactification on $T^2 \times K3$ and try to interpret the gaugings that produce inflaton dynamics in terms of fluxes.
\end{description}
We plan to address such questions in forthcoming publications.
\par
\section*{Acknowledgements}
We are grateful to our friends Sergio Ferrara and Antoine Van Proyen for useful discussions during the completion of this work.
The work of A.S. was supported in part by the RFBR Grants
No. 13-02-91330-NNIO-a, No. 13-02-90602-Arm-a and by the
Heisenberg-Landau program.
\newpage
|
\section*{Acknowledgements}
\noindent We express our gratitude to our colleagues in the CERN
accelerator departments for the excellent performance of the LHC. We
thank the technical and administrative staff at the LHCb
institutes. We acknowledge support from CERN and from the national
agencies: CAPES, CNPq, FAPERJ and FINEP (Brazil); NSFC (China);
CNRS/IN2P3 (France); BMBF, DFG, HGF and MPG (Germany); SFI (Ireland); INFN (Italy);
FOM and NWO (The Netherlands); MNiSW and NCN (Poland); MEN/IFA (Romania);
MinES and FANO (Russia); MinECo (Spain); SNSF and SER (Switzerland);
NASU (Ukraine); STFC (United Kingdom); NSF (USA).
The Tier1 computing centres are supported by IN2P3 (France), KIT and BMBF
(Germany), INFN (Italy), NWO and SURF (The Netherlands), PIC (Spain), GridPP
(United Kingdom).
We are indebted to the communities behind the multiple open
source software packages on which we depend. We are also thankful for the
computing resources and the access to software R\&D tools provided by Yandex LLC (Russia).
Individual groups or members have received support from
EPLANET, Marie Sk\l{}odowska-Curie Actions and ERC (European Union),
Conseil g\'{e}n\'{e}ral de Haute-Savoie, Labex ENIGMASS and OCEVU,
R\'{e}gion Auvergne (France), RFBR (Russia), XuntaGal and GENCAT (Spain), Royal Society and Royal
Commission for the Exhibition of 1851 (United Kingdom).
\ifx\mcitethebibliography\mciteundefinedmacro
\PackageError{LHCb.bst}{mciteplus.sty has not been loaded}
{This bibstyle requires the use of the mciteplus package.}\fi
\providecommand{\href}[2]{#2}
\begin{mcitethebibliography}{10}
\mciteSetBstSublistMode{n}
\mciteSetBstMaxWidthForm{subitem}{\alph{mcitesubitemcount})}
\mciteSetBstSublistLabelBeginEnd{\mcitemaxwidthsubitemform\space}
{\relax}{\relax}
\bibitem{Aubert:2003fg}
BaBar collaboration, B.~Aubert {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of a narrow meson
decaying to $D_s^+ \pi^0$ at a mass of $2.32 {\,Ge\kern -0.1em V\!/}c^2$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.90.242001}{Phys.\ Rev.\ Lett.\
\textbf{90} (2003) 242001}, \href{http://arxiv.org/abs/hep-ex/0304021}{{\tt
arXiv:hep-ex/0304021}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Besson:2003cp}
CLEO collaboration, D.~Besson {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of a narrow resonance
of mass $2.46 {\,Ge\kern -0.1em V\!/}c^2$ decaying to $D^{*+}_s \pi^0$ and
confirmation of the $D^*_{sJ}(2317)$ state}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.68.032002}{Phys.\ Rev.\
\textbf{D68} (2003) 032002}, Erratum
\href{http://dx.doi.org/10.1103/PhysRevD.75.119908}{ibid.\ \textbf{D75}
(2007) 119908}, \href{http://arxiv.org/abs/hep-ex/0305100}{{\tt
arXiv:hep-ex/0305100}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aubert:2006mh}
BaBar collaboration, B.~Aubert {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of a new $D_s$ meson
decaying to $DK$ at a mass of $2.86 {\,Ge\kern -0.1em V\!/}c^2$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.97.222001}{Phys.\ Rev.\ Lett.\
\textbf{97} (2006) 222001}, \href{http://arxiv.org/abs/hep-ex/0607082}{{\tt
arXiv:hep-ex/0607082}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Brodzicka:2007aa}
Belle collaboration, J.~Brodzicka {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of a new $D_{sJ}$
meson in ${\ensuremath{\Bu}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\Dbar{}^0}}\xspace{\ensuremath{\D^0}}\xspace{\ensuremath{\kaon^+}}\xspace$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.100.092001}{Phys.\ Rev.\
Lett.\ \textbf{100} (2008) 092001},
\href{http://arxiv.org/abs/0707.3491}{{\tt arXiv:0707.3491}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aubert:2009ah}
BaBar collaboration, B.~Aubert {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of $D_{sJ}$ decays to
$D^{*}K$ in inclusive $e^{+}e^{-}$ interactions}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.80.092003}{Phys.\ Rev.\
\textbf{D80} (2009) 092003}, \href{http://arxiv.org/abs/0908.0806}{{\tt
arXiv:0908.0806}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-016}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of $D_{sJ}$ decays to
$D^+K^0_S$ and $D^0K^+$ final states in $pp$ collisions}},
}{}\href{http://dx.doi.org/10.1007/JHEP10(2012)151}{JHEP \textbf{10} (2012)
151}, \href{http://arxiv.org/abs/1207.6016}{{\tt arXiv:1207.6016}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Swanson:2006st}
E.~S. Swanson, \ifthenelse{\boolean{articletitles}}{\emph{{The new heavy
mesons: a status report}},
}{}\href{http://dx.doi.org/10.1016/j.physrep.2006.04.003}{Phys.\ Rept.\
\textbf{429} (2006) 243}, \href{http://arxiv.org/abs/hep-ph/0601110}{{\tt
arXiv:hep-ph/0601110}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Rosner:2006jz}
J.~L. Rosner, \ifthenelse{\boolean{articletitles}}{\emph{{Hadron spectroscopy:
theory and experiment}},
}{}\href{http://dx.doi.org/10.1088/0954-3899/34/7/S07}{J.\ Phys.\
\textbf{G34} (2007) S127}, \href{http://arxiv.org/abs/hep-ph/0609195}{{\tt
arXiv:hep-ph/0609195}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Klempt:2007cp}
E.~Klempt and A.~Zaitsev,
\ifthenelse{\boolean{articletitles}}{\emph{{Glueballs, hybrids, multiquarks:
experimental facts versus QCD inspired concepts}},
}{}\href{http://dx.doi.org/10.1016/j.physrep.2007.07.006}{Phys.\ Rept.\
\textbf{454} (2007) 1}, \href{http://arxiv.org/abs/0708.4016}{{\tt
arXiv:0708.4016}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Colangelo:2012xi}
P.~Colangelo, F.~De~Fazio, F.~Giannuzzi, and S.~Nicotri,
\ifthenelse{\boolean{articletitles}}{\emph{{New meson spectroscopy with open
charm and beauty}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.86.054024}{Phys.\ Rev.\
\textbf{D86} (2012) 054024}, \href{http://arxiv.org/abs/1207.6940}{{\tt
arXiv:1207.6940}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Wagner:1974gw}
F.~Wagner, M.~Tabak, and D.~M. Chew,
\ifthenelse{\boolean{articletitles}}{\emph{{An amplitude analysis for the
reaction $\pi^+ {\ensuremath{\Pp}}\xspace \ensuremath{\rightarrow}\xspace \pi^+ \pi^- \pi^0 \Delta^{++}$ at $7 {\,Ge\kern
-0.1em V}$}}, }{}\href{http://dx.doi.org/10.1016/0370-2693(75)90637-1}{Phys.\
Lett.\ \textbf{B58} (1975) 201}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aston:1988rf}
D.~Aston {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Spin parity
determination of the $\phi_{J}(1850)$ from $K^- p$ interactions at $11
{\,Ge\kern -0.1em V\!/}c$}},
}{}\href{http://dx.doi.org/10.1016/0370-2693(88)90439-X}{Phys.\ Lett.\
\textbf{B208} (1988) 324}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Brandenburg:1975ft}
G.~W. Brandenburg {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Determination of the $K^*(1800)$
spin parity}},
}{}\href{http://dx.doi.org/10.1016/0370-2693(76)90711-5}{Phys.\ Lett.\
\textbf{B60} (1976) 478}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Baldi:1976ua}
R.~Baldi {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Observation
of the $K^*(1780)$ in the reaction $K^+ {\ensuremath{\Pp}}\xspace \ensuremath{\rightarrow}\xspace K^0_S \pi^+ {\ensuremath{\Pp}}\xspace$ at
$10 {\,Ge\kern -0.1em V\!/}c$}},
}{}\href{http://dx.doi.org/10.1016/0370-2693(76)90279-3}{Phys.\ Lett.\
\textbf{B63} (1976) 344}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{blatt-weisskopf}
J.~Blatt and V.~E. Weisskopf, {\em Theoretical nuclear physics}, J. Wiley (New
York), 1952\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Dalitz:1953cp}
R.~H. Dalitz, \ifthenelse{\boolean{articletitles}}{\emph{{On the analysis of
tau-meson data and the nature of the tau-meson}},
}{}\href{http://dx.doi.org/10.1080/14786441008520365}{Phil.\ Mag.\
\textbf{44} (1953) 1068}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2014-036}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz plot analysis of
$B^0_s\ensuremath{\rightarrow}\xspace\bar{D}^0K^-\pi^+$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.90.072003}{Phys.\ Rev.\
\textbf{D90} (2014) 072003}, \href{http://arxiv.org/abs/1407.7712}{{\tt
arXiv:1407.7712}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Alves:2008zz}
LHCb collaboration, A.~A. Alves~Jr.\ {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{The \mbox{LHCb}\xspace detector at the LHC}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/3/08/S08005}{JINST \textbf{3}
(2008) S08005}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2012-004}
R.~Aaij {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{The \mbox{LHCb}\xspace
trigger and its performance in 2011}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/8/04/P04022}{JINST \textbf{8}
(2013) P04022}, \href{http://arxiv.org/abs/1211.3055}{{\tt
arXiv:1211.3055}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-056}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Search for the decay $B^0_s \ensuremath{\rightarrow}\xspace
D^{*\mp}\pi^\pm$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.87.071101}{Phys.\ Rev.\
\textbf{D87} (2013) 071101(R)}, \href{http://arxiv.org/abs/1302.6446}{{\tt
arXiv:1302.6446}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2013-022}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurements of the branching
fractions of the decays $B^0_s \ensuremath{\rightarrow}\xspace \bar{D}^0 K^- \pi^+$ and $B^0 \ensuremath{\rightarrow}\xspace
\bar{D}^0 K^+ \pi^-$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.87.112009}{Phys.\ Rev.\
\textbf{D87} (2013) 112009}, \href{http://arxiv.org/abs/1304.6317}{{\tt
arXiv:1304.6317}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Feindt:2006pm}
M.~Feindt and U.~Kerzel, \ifthenelse{\boolean{articletitles}}{\emph{{The
NeuroBayes neural network package}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.11.166}{Nucl.\ Instrum.\
Meth.\ \textbf{A559} (2006) 190}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Pivk:2004ty}
M.~Pivk and F.~R. Le~Diberder,
\ifthenelse{\boolean{articletitles}}{\emph{{sPlot: a statistical tool to
unfold data distributions}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.08.106}{Nucl.\ Instrum.\
Meth.\ \textbf{A555} (2005) 356},
\href{http://arxiv.org/abs/physics/0402083}{{\tt
arXiv:physics/0402083}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-025}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{First evidence for the
annihilation decay mode $B^+ \ensuremath{\rightarrow}\xspace D_s^+\phi$}},
}{}\href{http://dx.doi.org/10.1007/JHEP02(2013)043}{JHEP \textbf{02} (2013)
043}, \href{http://arxiv.org/abs/1210.1089}{{\tt arXiv:1210.1089}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-050}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{First observations of
$\bar{B}^0_s \ensuremath{\rightarrow}\xspace D^+D^-$, $D_s^+D^-$ and $D^0\bar{D}^0$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.87.092007}{Phys.\ Rev.\
\textbf{D87} (2013) 092007}, \href{http://arxiv.org/abs/1302.5854}{{\tt
arXiv:1302.5854}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-048}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurements of the
$\Lambda_b^0$, $\Xi_b^-$, and $\Omega_b^-$ baryon masses}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.110.182001}{Phys.\ Rev.\
Lett.\ \textbf{110} (2013) 182001},
\href{http://arxiv.org/abs/1302.1072}{{\tt arXiv:1302.1072}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCB-PAPER-2013-011}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Precision measurement of $D$
meson mass differences}},
}{}\href{http://dx.doi.org/10.1007/JHEP06(2013)065}{JHEP \textbf{06} (2013)
065}, \href{http://arxiv.org/abs/1304.6865}{{\tt arXiv:1304.6865}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{PDG2012}
Particle Data Group, J.~Beringer {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{\href{http://pdg.lbl.gov/}{Review
of particle physics}}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.86.010001}{Phys.\ Rev.\
\textbf{D86} (2012) 010001}, {and 2013 partial update for the 2014
edition}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Hulsbergen:2005pu}
W.~D. Hulsbergen, \ifthenelse{\boolean{articletitles}}{\emph{{Decay chain
fitting with a Kalman filter}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.06.078}{Nucl.\ Instrum.\
Meth.\ \textbf{A552} (2005) 566},
\href{http://arxiv.org/abs/physics/0503191}{{\tt
arXiv:physics/0503191}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2013-056}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of beauty baryon decays to
$D^0 p h^-$ and $\Lambda_c^+ h^-$ final states}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.89.032001}{Phys.\ Rev.\
\textbf{D89} (2014) 032001}, \href{http://arxiv.org/abs/1311.4823}{{\tt
arXiv:1311.4823}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Skwarnicki:1986xj}
T.~Skwarnicki, {\em {A study of the radiative cascade transitions between the
Upsilon-prime and Upsilon resonances}}, PhD thesis, Institute of Nuclear
Physics, Krakow, 1986,
{\href{http://inspirehep.net/record/230779/files/230779.pdf}{DESY-F31-86-02}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Fleming:1964zz}
G.~N. Fleming, \ifthenelse{\boolean{articletitles}}{\emph{{Recoupling effects
in the isobar model. 1. General formalism for three-pion scattering}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.135.B551}{Phys.\ Rev.\
\textbf{135} (1964) B551}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Morgan:1968zza}
D.~Morgan, \ifthenelse{\boolean{articletitles}}{\emph{{Phenomenological
analysis of $I=\frac{1}{2}$ single-pion production processes in the energy
range 500 to 700 MeV}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.166.1731}{Phys.\ Rev.\
\textbf{166} (1968) 1731}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Herndon:1973yn}
D.~Herndon, P.~Soding, and R.~J. Cashmore,
\ifthenelse{\boolean{articletitles}}{\emph{{Generalised isobar model
formalism}}, }{}\href{http://dx.doi.org/10.1103/PhysRevD.11.3165}{Phys.\
Rev.\ \textbf{D11} (1975) 3165}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{lass}
LASS collaboration, D.~Aston {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{A study of $K^- \pi^+$ scattering
in the reaction $K^- p \ensuremath{\rightarrow}\xspace K^- \pi^+ n$ at $11 {\,Ge\kern -0.1em V\!/}c$}},
}{}\href{http://dx.doi.org/10.1016/0550-3213(88)90028-4}{Nucl.\ Phys.\
\textbf{B296} (1988) 493}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Zemach:1963bc}
C.~Zemach, \ifthenelse{\boolean{articletitles}}{\emph{{Three-pion decays of
unstable particles}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.133.B1201}{Phys.\ Rev.\
\textbf{133} (1964) B1201}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Zemach:1968zz}
C.~Zemach, \ifthenelse{\boolean{articletitles}}{\emph{{Use of angular-momentum
tensors}}, }{}\href{http://dx.doi.org/10.1103/PhysRev.140.B97}{Phys.\ Rev.\
\textbf{140} (1965) B97}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\end{mcitethebibliography}
\newpage
\centerline{\large\bf LHCb collaboration}
\begin{flushleft}
\small
R.~Aaij$^{41}$,
B.~Adeva$^{37}$,
M.~Adinolfi$^{46}$,
A.~Affolder$^{52}$,
Z.~Ajaltouni$^{5}$,
S.~Akar$^{6}$,
J.~Albrecht$^{9}$,
F.~Alessio$^{38}$,
M.~Alexander$^{51}$,
S.~Ali$^{41}$,
G.~Alkhazov$^{30}$,
P.~Alvarez~Cartelle$^{37}$,
A.A.~Alves~Jr$^{25,38}$,
S.~Amato$^{2}$,
S.~Amerio$^{22}$,
Y.~Amhis$^{7}$,
L.~An$^{3}$,
L.~Anderlini$^{17,g}$,
J.~Anderson$^{40}$,
R.~Andreassen$^{57}$,
M.~Andreotti$^{16,f}$,
J.E.~Andrews$^{58}$,
R.B.~Appleby$^{54}$,
O.~Aquines~Gutierrez$^{10}$,
F.~Archilli$^{38}$,
A.~Artamonov$^{35}$,
M.~Artuso$^{59}$,
E.~Aslanides$^{6}$,
G.~Auriemma$^{25,n}$,
M.~Baalouch$^{5}$,
S.~Bachmann$^{11}$,
J.J.~Back$^{48}$,
A.~Badalov$^{36}$,
C.~Baesso$^{60}$,
W.~Baldini$^{16}$,
R.J.~Barlow$^{54}$,
C.~Barschel$^{38}$,
S.~Barsuk$^{7}$,
W.~Barter$^{47}$,
V.~Batozskaya$^{28}$,
V.~Battista$^{39}$,
A.~Bay$^{39}$,
L.~Beaucourt$^{4}$,
J.~Beddow$^{51}$,
F.~Bedeschi$^{23}$,
I.~Bediaga$^{1}$,
S.~Belogurov$^{31}$,
K.~Belous$^{35}$,
I.~Belyaev$^{31}$,
E.~Ben-Haim$^{8}$,
G.~Bencivenni$^{18}$,
S.~Benson$^{38}$,
J.~Benton$^{46}$,
A.~Berezhnoy$^{32}$,
R.~Bernet$^{40}$,
M.-O.~Bettler$^{47}$,
M.~van~Beuzekom$^{41}$,
A.~Bien$^{11}$,
S.~Bifani$^{45}$,
T.~Bird$^{54}$,
A.~Bizzeti$^{17,i}$,
P.M.~Bj\o rnstad$^{54}$,
T.~Blake$^{48}$,
F.~Blanc$^{39}$,
J.~Blouw$^{10}$,
S.~Blusk$^{59}$,
V.~Bocci$^{25}$,
A.~Bondar$^{34}$,
N.~Bondar$^{30,38}$,
W.~Bonivento$^{15,38}$,
S.~Borghi$^{54}$,
A.~Borgia$^{59}$,
M.~Borsato$^{7}$,
T.J.V.~Bowcock$^{52}$,
E.~Bowen$^{40}$,
C.~Bozzi$^{16}$,
T.~Brambach$^{9}$,
J.~van~den~Brand$^{42}$,
J.~Bressieux$^{39}$,
D.~Brett$^{54}$,
M.~Britsch$^{10}$,
T.~Britton$^{59}$,
J.~Brodzicka$^{54}$,
N.H.~Brook$^{46}$,
H.~Brown$^{52}$,
A.~Bursche$^{40}$,
G.~Busetto$^{22,r}$,
J.~Buytaert$^{38}$,
S.~Cadeddu$^{15}$,
R.~Calabrese$^{16,f}$,
M.~Calvi$^{20,k}$,
M.~Calvo~Gomez$^{36,p}$,
P.~Campana$^{18,38}$,
D.~Campora~Perez$^{38}$,
A.~Carbone$^{14,d}$,
G.~Carboni$^{24,l}$,
R.~Cardinale$^{19,38,j}$,
A.~Cardini$^{15}$,
L.~Carson$^{50}$,
K.~Carvalho~Akiba$^{2}$,
G.~Casse$^{52}$,
L.~Cassina$^{20}$,
L.~Castillo~Garcia$^{38}$,
M.~Cattaneo$^{38}$,
Ch.~Cauet$^{9}$,
R.~Cenci$^{58}$,
M.~Charles$^{8}$,
Ph.~Charpentier$^{38}$,
M. ~Chefdeville$^{4}$,
S.~Chen$^{54}$,
S.-F.~Cheung$^{55}$,
N.~Chiapolini$^{40}$,
M.~Chrzaszcz$^{40,26}$,
K.~Ciba$^{38}$,
X.~Cid~Vidal$^{38}$,
G.~Ciezarek$^{53}$,
P.E.L.~Clarke$^{50}$,
M.~Clemencic$^{38}$,
H.V.~Cliff$^{47}$,
J.~Closier$^{38}$,
V.~Coco$^{38}$,
J.~Cogan$^{6}$,
E.~Cogneras$^{5}$,
P.~Collins$^{38}$,
A.~Comerma-Montells$^{11}$,
A.~Contu$^{15}$,
A.~Cook$^{46}$,
M.~Coombes$^{46}$,
S.~Coquereau$^{8}$,
G.~Corti$^{38}$,
M.~Corvo$^{16,f}$,
I.~Counts$^{56}$,
B.~Couturier$^{38}$,
G.A.~Cowan$^{50}$,
D.C.~Craik$^{48}$,
M.~Cruz~Torres$^{60}$,
S.~Cunliffe$^{53}$,
R.~Currie$^{50}$,
C.~D'Ambrosio$^{38}$,
J.~Dalseno$^{46}$,
P.~David$^{8}$,
P.N.Y.~David$^{41}$,
A.~Davis$^{57}$,
K.~De~Bruyn$^{41}$,
S.~De~Capua$^{54}$,
M.~De~Cian$^{11}$,
J.M.~De~Miranda$^{1}$,
L.~De~Paula$^{2}$,
W.~De~Silva$^{57}$,
P.~De~Simone$^{18}$,
D.~Decamp$^{4}$,
M.~Deckenhoff$^{9}$,
L.~Del~Buono$^{8}$,
N.~D\'{e}l\'{e}age$^{4}$,
D.~Derkach$^{55}$,
O.~Deschamps$^{5}$,
F.~Dettori$^{38}$,
A.~Di~Canto$^{38}$,
H.~Dijkstra$^{38}$,
S.~Donleavy$^{52}$,
F.~Dordei$^{11}$,
M.~Dorigo$^{39}$,
A.~Dosil~Su\'{a}rez$^{37}$,
D.~Dossett$^{48}$,
A.~Dovbnya$^{43}$,
K.~Dreimanis$^{52}$,
G.~Dujany$^{54}$,
F.~Dupertuis$^{39}$,
P.~Durante$^{38}$,
R.~Dzhelyadin$^{35}$,
A.~Dziurda$^{26}$,
A.~Dzyuba$^{30}$,
S.~Easo$^{49,38}$,
U.~Egede$^{53}$,
V.~Egorychev$^{31}$,
S.~Eidelman$^{34}$,
S.~Eisenhardt$^{50}$,
U.~Eitschberger$^{9}$,
R.~Ekelhof$^{9}$,
L.~Eklund$^{51}$,
I.~El~Rifai$^{5}$,
Ch.~Elsasser$^{40}$,
S.~Ely$^{59}$,
S.~Esen$^{11}$,
H.-M.~Evans$^{47}$,
T.~Evans$^{55}$,
A.~Falabella$^{14}$,
C.~F\"{a}rber$^{11}$,
C.~Farinelli$^{41}$,
N.~Farley$^{45}$,
S.~Farry$^{52}$,
RF~Fay$^{52}$,
D.~Ferguson$^{50}$,
V.~Fernandez~Albor$^{37}$,
F.~Ferreira~Rodrigues$^{1}$,
M.~Ferro-Luzzi$^{38}$,
S.~Filippov$^{33}$,
M.~Fiore$^{16,f}$,
M.~Fiorini$^{16,f}$,
M.~Firlej$^{27}$,
C.~Fitzpatrick$^{39}$,
T.~Fiutowski$^{27}$,
M.~Fontana$^{10}$,
F.~Fontanelli$^{19,j}$,
R.~Forty$^{38}$,
O.~Francisco$^{2}$,
M.~Frank$^{38}$,
C.~Frei$^{38}$,
M.~Frosini$^{17,38,g}$,
J.~Fu$^{21,38}$,
E.~Furfaro$^{24,l}$,
A.~Gallas~Torreira$^{37}$,
D.~Galli$^{14,d}$,
S.~Gallorini$^{22}$,
S.~Gambetta$^{19,j}$,
M.~Gandelman$^{2}$,
P.~Gandini$^{59}$,
Y.~Gao$^{3}$,
J.~Garc\'{i}a~Pardi\~{n}as$^{37}$,
J.~Garofoli$^{59}$,
J.~Garra~Tico$^{47}$,
L.~Garrido$^{36}$,
C.~Gaspar$^{38}$,
R.~Gauld$^{55}$,
L.~Gavardi$^{9}$,
G.~Gavrilov$^{30}$,
A.~Geraci$^{21,v}$,
E.~Gersabeck$^{11}$,
M.~Gersabeck$^{54}$,
T.~Gershon$^{48}$,
Ph.~Ghez$^{4}$,
A.~Gianelle$^{22}$,
S.~Gian\`{i}$^{39}$,
V.~Gibson$^{47}$,
L.~Giubega$^{29}$,
V.V.~Gligorov$^{38}$,
C.~G\"{o}bel$^{60}$,
D.~Golubkov$^{31}$,
A.~Golutvin$^{53,31,38}$,
A.~Gomes$^{1,a}$,
C.~Gotti$^{20}$,
M.~Grabalosa~G\'{a}ndara$^{5}$,
R.~Graciani~Diaz$^{36}$,
L.A.~Granado~Cardoso$^{38}$,
E.~Graug\'{e}s$^{36}$,
G.~Graziani$^{17}$,
A.~Grecu$^{29}$,
E.~Greening$^{55}$,
S.~Gregson$^{47}$,
P.~Griffith$^{45}$,
L.~Grillo$^{11}$,
O.~Gr\"{u}nberg$^{62}$,
B.~Gui$^{59}$,
E.~Gushchin$^{33}$,
Yu.~Guz$^{35,38}$,
T.~Gys$^{38}$,
C.~Hadjivasiliou$^{59}$,
G.~Haefeli$^{39}$,
C.~Haen$^{38}$,
S.C.~Haines$^{47}$,
S.~Hall$^{53}$,
B.~Hamilton$^{58}$,
T.~Hampson$^{46}$,
X.~Han$^{11}$,
S.~Hansmann-Menzemer$^{11}$,
N.~Harnew$^{55}$,
S.T.~Harnew$^{46}$,
J.~Harrison$^{54}$,
J.~He$^{38}$,
T.~Head$^{38}$,
V.~Heijne$^{41}$,
K.~Hennessy$^{52}$,
P.~Henrard$^{5}$,
L.~Henry$^{8}$,
J.A.~Hernando~Morata$^{37}$,
E.~van~Herwijnen$^{38}$,
M.~He\ss$^{62}$,
A.~Hicheur$^{1}$,
D.~Hill$^{55}$,
M.~Hoballah$^{5}$,
C.~Hombach$^{54}$,
W.~Hulsbergen$^{41}$,
P.~Hunt$^{55}$,
N.~Hussain$^{55}$,
D.~Hutchcroft$^{52}$,
D.~Hynds$^{51}$,
M.~Idzik$^{27}$,
P.~Ilten$^{56}$,
R.~Jacobsson$^{38}$,
A.~Jaeger$^{11}$,
J.~Jalocha$^{55}$,
E.~Jans$^{41}$,
P.~Jaton$^{39}$,
A.~Jawahery$^{58}$,
F.~Jing$^{3}$,
M.~John$^{55}$,
D.~Johnson$^{38}$,
C.R.~Jones$^{47}$,
C.~Joram$^{38}$,
B.~Jost$^{38}$,
N.~Jurik$^{59}$,
M.~Kaballo$^{9}$,
S.~Kandybei$^{43}$,
W.~Kanso$^{6}$,
M.~Karacson$^{38}$,
T.M.~Karbach$^{38}$,
S.~Karodia$^{51}$,
M.~Kelsey$^{59}$,
I.R.~Kenyon$^{45}$,
T.~Ketel$^{42}$,
B.~Khanji$^{20}$,
C.~Khurewathanakul$^{39}$,
S.~Klaver$^{54}$,
K.~Klimaszewski$^{28}$,
O.~Kochebina$^{7}$,
M.~Kolpin$^{11}$,
I.~Komarov$^{39}$,
R.F.~Koopman$^{42}$,
P.~Koppenburg$^{41,38}$,
M.~Korolev$^{32}$,
A.~Kozlinskiy$^{41}$,
L.~Kravchuk$^{33}$,
K.~Kreplin$^{11}$,
M.~Kreps$^{48}$,
G.~Krocker$^{11}$,
P.~Krokovny$^{34}$,
F.~Kruse$^{9}$,
W.~Kucewicz$^{26,o}$,
M.~Kucharczyk$^{20,26,38,k}$,
V.~Kudryavtsev$^{34}$,
K.~Kurek$^{28}$,
T.~Kvaratskheliya$^{31}$,
V.N.~La~Thi$^{39}$,
D.~Lacarrere$^{38}$,
G.~Lafferty$^{54}$,
A.~Lai$^{15}$,
D.~Lambert$^{50}$,
R.W.~Lambert$^{42}$,
G.~Lanfranchi$^{18}$,
C.~Langenbruch$^{48}$,
B.~Langhans$^{38}$,
T.~Latham$^{48}$,
C.~Lazzeroni$^{45}$,
R.~Le~Gac$^{6}$,
J.~van~Leerdam$^{41}$,
J.-P.~Lees$^{4}$,
R.~Lef\`{e}vre$^{5}$,
A.~Leflat$^{32}$,
J.~Lefran\c{c}ois$^{7}$,
S.~Leo$^{23}$,
O.~Leroy$^{6}$,
T.~Lesiak$^{26}$,
B.~Leverington$^{11}$,
Y.~Li$^{3}$,
T.~Likhomanenko$^{63}$,
M.~Liles$^{52}$,
R.~Lindner$^{38}$,
C.~Linn$^{38}$,
F.~Lionetto$^{40}$,
B.~Liu$^{15}$,
S.~Lohn$^{38}$,
I.~Longstaff$^{51}$,
J.H.~Lopes$^{2}$,
N.~Lopez-March$^{39}$,
P.~Lowdon$^{40}$,
H.~Lu$^{3}$,
D.~Lucchesi$^{22,r}$,
H.~Luo$^{50}$,
A.~Lupato$^{22}$,
E.~Luppi$^{16,f}$,
O.~Lupton$^{55}$,
F.~Machefert$^{7}$,
I.V.~Machikhiliyan$^{31}$,
F.~Maciuc$^{29}$,
O.~Maev$^{30}$,
S.~Malde$^{55}$,
A.~Malinin$^{63}$,
G.~Manca$^{15,e}$,
G.~Mancinelli$^{6}$,
A.~Mapelli$^{38}$,
J.~Maratas$^{5}$,
J.F.~Marchand$^{4}$,
U.~Marconi$^{14}$,
C.~Marin~Benito$^{36}$,
P.~Marino$^{23,t}$,
R.~M\"{a}rki$^{39}$,
J.~Marks$^{11}$,
G.~Martellotti$^{25}$,
A.~Martens$^{8}$,
A.~Mart\'{i}n~S\'{a}nchez$^{7}$,
M.~Martinelli$^{39}$,
D.~Martinez~Santos$^{42}$,
F.~Martinez~Vidal$^{64}$,
D.~Martins~Tostes$^{2}$,
A.~Massafferri$^{1}$,
R.~Matev$^{38}$,
Z.~Mathe$^{38}$,
C.~Matteuzzi$^{20}$,
A.~Mazurov$^{16,f}$,
M.~McCann$^{53}$,
J.~McCarthy$^{45}$,
A.~McNab$^{54}$,
R.~McNulty$^{12}$,
B.~McSkelly$^{52}$,
B.~Meadows$^{57}$,
F.~Meier$^{9}$,
M.~Meissner$^{11}$,
M.~Merk$^{41}$,
D.A.~Milanes$^{8}$,
M.-N.~Minard$^{4}$,
N.~Moggi$^{14}$,
J.~Molina~Rodriguez$^{60}$,
S.~Monteil$^{5}$,
M.~Morandin$^{22}$,
P.~Morawski$^{27}$,
A.~Mord\`{a}$^{6}$,
M.J.~Morello$^{23,t}$,
J.~Moron$^{27}$,
A.-B.~Morris$^{50}$,
R.~Mountain$^{59}$,
F.~Muheim$^{50}$,
K.~M\"{u}ller$^{40}$,
M.~Mussini$^{14}$,
B.~Muster$^{39}$,
P.~Naik$^{46}$,
T.~Nakada$^{39}$,
R.~Nandakumar$^{49}$,
I.~Nasteva$^{2}$,
M.~Needham$^{50}$,
N.~Neri$^{21}$,
S.~Neubert$^{38}$,
N.~Neufeld$^{38}$,
M.~Neuner$^{11}$,
A.D.~Nguyen$^{39}$,
T.D.~Nguyen$^{39}$,
C.~Nguyen-Mau$^{39,q}$,
M.~Nicol$^{7}$,
V.~Niess$^{5}$,
R.~Niet$^{9}$,
N.~Nikitin$^{32}$,
T.~Nikodem$^{11}$,
A.~Novoselov$^{35}$,
D.P.~O'Hanlon$^{48}$,
A.~Oblakowska-Mucha$^{27}$,
V.~Obraztsov$^{35}$,
S.~Oggero$^{41}$,
S.~Ogilvy$^{51}$,
O.~Okhrimenko$^{44}$,
R.~Oldeman$^{15,e}$,
G.~Onderwater$^{65}$,
M.~Orlandea$^{29}$,
J.M.~Otalora~Goicochea$^{2}$,
P.~Owen$^{53}$,
A.~Oyanguren$^{64}$,
B.K.~Pal$^{59}$,
A.~Palano$^{13,c}$,
F.~Palombo$^{21,u}$,
M.~Palutan$^{18}$,
J.~Panman$^{38}$,
A.~Papanestis$^{49,38}$,
M.~Pappagallo$^{51}$,
L.L.~Pappalardo$^{16,f}$,
C.~Parkes$^{54}$,
C.J.~Parkinson$^{9,45}$,
G.~Passaleva$^{17}$,
G.D.~Patel$^{52}$,
M.~Patel$^{53}$,
C.~Patrignani$^{19,j}$,
A.~Pazos~Alvarez$^{37}$,
A.~Pearce$^{54}$,
A.~Pellegrino$^{41}$,
M.~Pepe~Altarelli$^{38}$,
S.~Perazzini$^{14,d}$,
E.~Perez~Trigo$^{37}$,
P.~Perret$^{5}$,
M.~Perrin-Terrin$^{6}$,
L.~Pescatore$^{45}$,
E.~Pesen$^{66}$,
K.~Petridis$^{53}$,
A.~Petrolini$^{19,j}$,
E.~Picatoste~Olloqui$^{36}$,
B.~Pietrzyk$^{4}$,
T.~Pila\v{r}$^{48}$,
D.~Pinci$^{25}$,
A.~Pistone$^{19}$,
S.~Playfer$^{50}$,
M.~Plo~Casasus$^{37}$,
F.~Polci$^{8}$,
A.~Poluektov$^{48,34}$,
E.~Polycarpo$^{2}$,
A.~Popov$^{35}$,
D.~Popov$^{10}$,
B.~Popovici$^{29}$,
C.~Potterat$^{2}$,
E.~Price$^{46}$,
J.~Prisciandaro$^{39}$,
A.~Pritchard$^{52}$,
C.~Prouve$^{46}$,
V.~Pugatch$^{44}$,
A.~Puig~Navarro$^{39}$,
G.~Punzi$^{23,s}$,
W.~Qian$^{4}$,
B.~Rachwal$^{26}$,
J.H.~Rademacker$^{46}$,
B.~Rakotomiaramanana$^{39}$,
M.~Rama$^{18}$,
M.S.~Rangel$^{2}$,
I.~Raniuk$^{43}$,
N.~Rauschmayr$^{38}$,
G.~Raven$^{42}$,
S.~Reichert$^{54}$,
M.M.~Reid$^{48}$,
A.C.~dos~Reis$^{1}$,
S.~Ricciardi$^{49}$,
S.~Richards$^{46}$,
M.~Rihl$^{38}$,
K.~Rinnert$^{52}$,
V.~Rives~Molina$^{36}$,
D.A.~Roa~Romero$^{5}$,
P.~Robbe$^{7}$,
A.B.~Rodrigues$^{1}$,
E.~Rodrigues$^{54}$,
P.~Rodriguez~Perez$^{54}$,
S.~Roiser$^{38}$,
V.~Romanovsky$^{35}$,
A.~Romero~Vidal$^{37}$,
M.~Rotondo$^{22}$,
J.~Rouvinet$^{39}$,
T.~Ruf$^{38}$,
F.~Ruffini$^{23}$,
H.~Ruiz$^{36}$,
P.~Ruiz~Valls$^{64}$,
J.J.~Saborido~Silva$^{37}$,
N.~Sagidova$^{30}$,
P.~Sail$^{51}$,
B.~Saitta$^{15,e}$,
V.~Salustino~Guimaraes$^{2}$,
C.~Sanchez~Mayordomo$^{64}$,
B.~Sanmartin~Sedes$^{37}$,
R.~Santacesaria$^{25}$,
C.~Santamarina~Rios$^{37}$,
E.~Santovetti$^{24,l}$,
A.~Sarti$^{18,m}$,
C.~Satriano$^{25,n}$,
A.~Satta$^{24}$,
D.M.~Saunders$^{46}$,
M.~Savrie$^{16,f}$,
D.~Savrina$^{31,32}$,
M.~Schiller$^{42}$,
H.~Schindler$^{38}$,
M.~Schlupp$^{9}$,
M.~Schmelling$^{10}$,
B.~Schmidt$^{38}$,
O.~Schneider$^{39}$,
A.~Schopper$^{38}$,
M.-H.~Schune$^{7}$,
R.~Schwemmer$^{38}$,
B.~Sciascia$^{18}$,
A.~Sciubba$^{25}$,
M.~Seco$^{37}$,
A.~Semennikov$^{31}$,
I.~Sepp$^{53}$,
N.~Serra$^{40}$,
J.~Serrano$^{6}$,
L.~Sestini$^{22}$,
P.~Seyfert$^{11}$,
M.~Shapkin$^{35}$,
I.~Shapoval$^{16,43,f}$,
Y.~Shcheglov$^{30}$,
T.~Shears$^{52}$,
L.~Shekhtman$^{34}$,
V.~Shevchenko$^{63}$,
A.~Shires$^{9}$,
R.~Silva~Coutinho$^{48}$,
G.~Simi$^{22}$,
M.~Sirendi$^{47}$,
N.~Skidmore$^{46}$,
T.~Skwarnicki$^{59}$,
N.A.~Smith$^{52}$,
E.~Smith$^{55,49}$,
E.~Smith$^{53}$,
J.~Smith$^{47}$,
M.~Smith$^{54}$,
H.~Snoek$^{41}$,
M.D.~Sokoloff$^{57}$,
F.J.P.~Soler$^{51}$,
F.~Soomro$^{39}$,
D.~Souza$^{46}$,
B.~Souza~De~Paula$^{2}$,
B.~Spaan$^{9}$,
A.~Sparkes$^{50}$,
P.~Spradlin$^{51}$,
S.~Sridharan$^{38}$,
F.~Stagni$^{38}$,
M.~Stahl$^{11}$,
S.~Stahl$^{11}$,
O.~Steinkamp$^{40}$,
O.~Stenyakin$^{35}$,
S.~Stevenson$^{55}$,
S.~Stoica$^{29}$,
S.~Stone$^{59}$,
B.~Storaci$^{40}$,
S.~Stracka$^{23,38}$,
M.~Straticiuc$^{29}$,
U.~Straumann$^{40}$,
R.~Stroili$^{22}$,
V.K.~Subbiah$^{38}$,
L.~Sun$^{57}$,
W.~Sutcliffe$^{53}$,
K.~Swientek$^{27}$,
S.~Swientek$^{9}$,
V.~Syropoulos$^{42}$,
M.~Szczekowski$^{28}$,
P.~Szczypka$^{39,38}$,
D.~Szilard$^{2}$,
T.~Szumlak$^{27}$,
S.~T'Jampens$^{4}$,
M.~Teklishyn$^{7}$,
G.~Tellarini$^{16,f}$,
F.~Teubert$^{38}$,
C.~Thomas$^{55}$,
E.~Thomas$^{38}$,
J.~van~Tilburg$^{41}$,
V.~Tisserand$^{4}$,
M.~Tobin$^{39}$,
S.~Tolk$^{42}$,
L.~Tomassetti$^{16,f}$,
D.~Tonelli$^{38}$,
S.~Topp-Joergensen$^{55}$,
N.~Torr$^{55}$,
E.~Tournefier$^{4}$,
S.~Tourneur$^{39}$,
M.T.~Tran$^{39}$,
M.~Tresch$^{40}$,
A.~Tsaregorodtsev$^{6}$,
P.~Tsopelas$^{41}$,
N.~Tuning$^{41}$,
M.~Ubeda~Garcia$^{38}$,
A.~Ukleja$^{28}$,
A.~Ustyuzhanin$^{63}$,
U.~Uwer$^{11}$,
V.~Vagnoni$^{14}$,
G.~Valenti$^{14}$,
A.~Vallier$^{7}$,
R.~Vazquez~Gomez$^{18}$,
P.~Vazquez~Regueiro$^{37}$,
C.~V\'{a}zquez~Sierra$^{37}$,
S.~Vecchi$^{16}$,
J.J.~Velthuis$^{46}$,
M.~Veltri$^{17,h}$,
G.~Veneziano$^{39}$,
M.~Vesterinen$^{11}$,
B.~Viaud$^{7}$,
D.~Vieira$^{2}$,
M.~Vieites~Diaz$^{37}$,
X.~Vilasis-Cardona$^{36,p}$,
A.~Vollhardt$^{40}$,
D.~Volyanskyy$^{10}$,
D.~Voong$^{46}$,
A.~Vorobyev$^{30}$,
V.~Vorobyev$^{34}$,
C.~Vo\ss$^{62}$,
H.~Voss$^{10}$,
J.A.~de~Vries$^{41}$,
R.~Waldi$^{62}$,
C.~Wallace$^{48}$,
R.~Wallace$^{12}$,
J.~Walsh$^{23}$,
S.~Wandernoth$^{11}$,
J.~Wang$^{59}$,
D.R.~Ward$^{47}$,
N.K.~Watson$^{45}$,
D.~Websdale$^{53}$,
M.~Whitehead$^{48}$,
J.~Wicht$^{38}$,
D.~Wiedner$^{11}$,
G.~Wilkinson$^{55}$,
M.P.~Williams$^{45}$,
M.~Williams$^{56}$,
F.F.~Wilson$^{49}$,
J.~Wimberley$^{58}$,
J.~Wishahi$^{9}$,
W.~Wislicki$^{28}$,
M.~Witek$^{26}$,
G.~Wormser$^{7}$,
S.A.~Wotton$^{47}$,
S.~Wright$^{47}$,
S.~Wu$^{3}$,
K.~Wyllie$^{38}$,
Y.~Xie$^{61}$,
Z.~Xing$^{59}$,
Z.~Xu$^{39}$,
Z.~Yang$^{3}$,
X.~Yuan$^{3}$,
O.~Yushchenko$^{35}$,
M.~Zangoli$^{14}$,
M.~Zavertyaev$^{10,b}$,
L.~Zhang$^{59}$,
W.C.~Zhang$^{12}$,
Y.~Zhang$^{3}$,
A.~Zhelezov$^{11}$,
A.~Zhokhov$^{31}$,
L.~Zhong$^{3}$,
A.~Zvyagin$^{38}$.\bigskip
{\footnotesize \it
$ ^{1}$Centro Brasileiro de Pesquisas F\'{i}sicas (CBPF), Rio de Janeiro, Brazil\\
$ ^{2}$Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, Brazil\\
$ ^{3}$Center for High Energy Physics, Tsinghua University, Beijing, China\\
$ ^{4}$LAPP, Universit\'{e} de Savoie, CNRS/IN2P3, Annecy-Le-Vieux, France\\
$ ^{5}$Clermont Universit\'{e}, Universit\'{e} Blaise Pascal, CNRS/IN2P3, LPC, Clermont-Ferrand, France\\
$ ^{6}$CPPM, Aix-Marseille Universit\'{e}, CNRS/IN2P3, Marseille, France\\
$ ^{7}$LAL, Universit\'{e} Paris-Sud, CNRS/IN2P3, Orsay, France\\
$ ^{8}$LPNHE, Universit\'{e} Pierre et Marie Curie, Universit\'{e} Paris Diderot, CNRS/IN2P3, Paris, France\\
$ ^{9}$Fakult\"{a}t Physik, Technische Universit\"{a}t Dortmund, Dortmund, Germany\\
$ ^{10}$Max-Planck-Institut f\"{u}r Kernphysik (MPIK), Heidelberg, Germany\\
$ ^{11}$Physikalisches Institut, Ruprecht-Karls-Universit\"{a}t Heidelberg, Heidelberg, Germany\\
$ ^{12}$School of Physics, University College Dublin, Dublin, Ireland\\
$ ^{13}$Sezione INFN di Bari, Bari, Italy\\
$ ^{14}$Sezione INFN di Bologna, Bologna, Italy\\
$ ^{15}$Sezione INFN di Cagliari, Cagliari, Italy\\
$ ^{16}$Sezione INFN di Ferrara, Ferrara, Italy\\
$ ^{17}$Sezione INFN di Firenze, Firenze, Italy\\
$ ^{18}$Laboratori Nazionali dell'INFN di Frascati, Frascati, Italy\\
$ ^{19}$Sezione INFN di Genova, Genova, Italy\\
$ ^{20}$Sezione INFN di Milano Bicocca, Milano, Italy\\
$ ^{21}$Sezione INFN di Milano, Milano, Italy\\
$ ^{22}$Sezione INFN di Padova, Padova, Italy\\
$ ^{23}$Sezione INFN di Pisa, Pisa, Italy\\
$ ^{24}$Sezione INFN di Roma Tor Vergata, Roma, Italy\\
$ ^{25}$Sezione INFN di Roma La Sapienza, Roma, Italy\\
$ ^{26}$Henryk Niewodniczanski Institute of Nuclear Physics Polish Academy of Sciences, Krak\'{o}w, Poland\\
$ ^{27}$AGH - University of Science and Technology, Faculty of Physics and Applied Computer Science, Krak\'{o}w, Poland\\
$ ^{28}$National Center for Nuclear Research (NCBJ), Warsaw, Poland\\
$ ^{29}$Horia Hulubei National Institute of Physics and Nuclear Engineering, Bucharest-Magurele, Romania\\
$ ^{30}$Petersburg Nuclear Physics Institute (PNPI), Gatchina, Russia\\
$ ^{31}$Institute of Theoretical and Experimental Physics (ITEP), Moscow, Russia\\
$ ^{32}$Institute of Nuclear Physics, Moscow State University (SINP MSU), Moscow, Russia\\
$ ^{33}$Institute for Nuclear Research of the Russian Academy of Sciences (INR RAN), Moscow, Russia\\
$ ^{34}$Budker Institute of Nuclear Physics (SB RAS) and Novosibirsk State University, Novosibirsk, Russia\\
$ ^{35}$Institute for High Energy Physics (IHEP), Protvino, Russia\\
$ ^{36}$Universitat de Barcelona, Barcelona, Spain\\
$ ^{37}$Universidad de Santiago de Compostela, Santiago de Compostela, Spain\\
$ ^{38}$European Organization for Nuclear Research (CERN), Geneva, Switzerland\\
$ ^{39}$Ecole Polytechnique F\'{e}d\'{e}rale de Lausanne (EPFL), Lausanne, Switzerland\\
$ ^{40}$Physik-Institut, Universit\"{a}t Z\"{u}rich, Z\"{u}rich, Switzerland\\
$ ^{41}$Nikhef National Institute for Subatomic Physics, Amsterdam, The Netherlands\\
$ ^{42}$Nikhef National Institute for Subatomic Physics and VU University Amsterdam, Amsterdam, The Netherlands\\
$ ^{43}$NSC Kharkiv Institute of Physics and Technology (NSC KIPT), Kharkiv, Ukraine\\
$ ^{44}$Institute for Nuclear Research of the National Academy of Sciences (KINR), Kyiv, Ukraine\\
$ ^{45}$University of Birmingham, Birmingham, United Kingdom\\
$ ^{46}$H.H. Wills Physics Laboratory, University of Bristol, Bristol, United Kingdom\\
$ ^{47}$Cavendish Laboratory, University of Cambridge, Cambridge, United Kingdom\\
$ ^{48}$Department of Physics, University of Warwick, Coventry, United Kingdom\\
$ ^{49}$STFC Rutherford Appleton Laboratory, Didcot, United Kingdom\\
$ ^{50}$School of Physics and Astronomy, University of Edinburgh, Edinburgh, United Kingdom\\
$ ^{51}$School of Physics and Astronomy, University of Glasgow, Glasgow, United Kingdom\\
$ ^{52}$Oliver Lodge Laboratory, University of Liverpool, Liverpool, United Kingdom\\
$ ^{53}$Imperial College London, London, United Kingdom\\
$ ^{54}$School of Physics and Astronomy, University of Manchester, Manchester, United Kingdom\\
$ ^{55}$Department of Physics, University of Oxford, Oxford, United Kingdom\\
$ ^{56}$Massachusetts Institute of Technology, Cambridge, MA, United States\\
$ ^{57}$University of Cincinnati, Cincinnati, OH, United States\\
$ ^{58}$University of Maryland, College Park, MD, United States\\
$ ^{59}$Syracuse University, Syracuse, NY, United States\\
$ ^{60}$Pontif\'{i}cia Universidade Cat\'{o}lica do Rio de Janeiro (PUC-Rio), Rio de Janeiro, Brazil, associated to $^{2}$\\
$ ^{61}$Institute of Particle Physics, Central China Normal University, Wuhan, Hubei, China, associated to $^{3}$\\
$ ^{62}$Institut f\"{u}r Physik, Universit\"{a}t Rostock, Rostock, Germany, associated to $^{11}$\\
$ ^{63}$National Research Centre Kurchatov Institute, Moscow, Russia, associated to $^{31}$\\
$ ^{64}$Instituto de Fisica Corpuscular (IFIC), Universitat de Valencia-CSIC, Valencia, Spain, associated to $^{36}$\\
$ ^{65}$KVI - University of Groningen, Groningen, The Netherlands, associated to $^{41}$\\
$ ^{66}$Celal Bayar University, Manisa, Turkey, associated to $^{38}$\\
\bigskip
$ ^{a}$Universidade Federal do Tri\^{a}ngulo Mineiro (UFTM), Uberaba-MG, Brazil\\
$ ^{b}$P.N. Lebedev Physical Institute, Russian Academy of Science (LPI RAS), Moscow, Russia\\
$ ^{c}$Universit\`{a} di Bari, Bari, Italy\\
$ ^{d}$Universit\`{a} di Bologna, Bologna, Italy\\
$ ^{e}$Universit\`{a} di Cagliari, Cagliari, Italy\\
$ ^{f}$Universit\`{a} di Ferrara, Ferrara, Italy\\
$ ^{g}$Universit\`{a} di Firenze, Firenze, Italy\\
$ ^{h}$Universit\`{a} di Urbino, Urbino, Italy\\
$ ^{i}$Universit\`{a} di Modena e Reggio Emilia, Modena, Italy\\
$ ^{j}$Universit\`{a} di Genova, Genova, Italy\\
$ ^{k}$Universit\`{a} di Milano Bicocca, Milano, Italy\\
$ ^{l}$Universit\`{a} di Roma Tor Vergata, Roma, Italy\\
$ ^{m}$Universit\`{a} di Roma La Sapienza, Roma, Italy\\
$ ^{n}$Universit\`{a} della Basilicata, Potenza, Italy\\
$ ^{o}$AGH - University of Science and Technology, Faculty of Computer Science, Electronics and Telecommunications, Krak\'{o}w, Poland\\
$ ^{p}$LIFAELS, La Salle, Universitat Ramon Llull, Barcelona, Spain\\
$ ^{q}$Hanoi University of Science, Hanoi, Viet Nam\\
$ ^{r}$Universit\`{a} di Padova, Padova, Italy\\
$ ^{s}$Universit\`{a} di Pisa, Pisa, Italy\\
$ ^{t}$Scuola Normale Superiore, Pisa, Italy\\
$ ^{u}$Universit\`{a} degli Studi di Milano, Milano, Italy\\
$ ^{v}$Politecnico di Milano, Milano, Italy\\
}
\end{flushleft}
\end{document}
|
\section{Introduction}\label{intro}
Leibniz algebras were defined by Loday in 1993 \cite{loday, loday2}. In recent years it has been a common theme to extend various results from Lie algebras to Leibniz algebras \cite{ao, ayupov, omirov}. Several authors have proven results on nilpotency and related concepts which can be used to help extend properties of Lie algebras to Leibniz algebras.
Specifically, variations of Engel's theorem for Leibniz algebras have been proven by different authors \cite{barnesengel, jacobsonleib} and Barnes has proven Levi's theorem for Leibniz algebras \cite{barneslevi}. Additionally, Barnes has shown that left-multiplication by any minimal ideal of a Leibniz algebra is either zero or anticommutative \cite{barnesleib}.
In an effort to classify Lie algebras, many authors place various restrictions on the nilradical \cite{cs, nw, rw, wld}. In \cite{tw}, Tremblay and Winternitz study solvable Lie algebras with triangular nilradical. It is the goal of this paper to extend these results to the Leibniz setting.
Recent work has been done on classification of certain classes of Leibniz algebras \cite{aor, chelsie-allison, heisen, clok, clok2}. In \cite{heisen}, a subset of the authors of this work found a complete classification of all Leibniz algebras whose nilradical is Heisenberg. In particular, this includes a classification of all Leibniz algebras whose nilradical is the triangular Lie algebra $T(3)$, since $T(3)$ is the three-dimensional Heisenberg algebra. For this reason our primary example will be Leibniz algebras whose nilradical is the triangular algebra $T(4)$.
\section{Preliminaries}
A Leibniz algebra, $L$, is a vector space over a field (which we will take to be $\mathbb{C}$ or $\mathbb{R}$) with a bilinear operation (which we will call multiplication) defined by $[x,y]$ which satisfies the Leibniz identity
\begin{equation}\label{Jacobi}
[x,[y,z]] = [[x,y],z] + [y,[x,z]]
\end{equation}
for all $x,y,z \in L$. In other words $L_x$, left-multiplication by $x$, is a derivation. Some authors choose to impose this property on $R_x$, right-multiplication by $x$, instead. Such an algebra is called a ``right'' Leibniz algebra, but we will consider only ``left'' Leibniz algebras (which satisfy \eqref{Jacobi}). $L$ is a Lie algebra if additionally $[x,y]=-[y,x]$.
The derived series of a Leibniz (Lie) algebra $L$ is defined by $L^{(1)}=[L,L]$, $L^{(n+1)}=[L^{(n)},L^{(n)}]$ for $n\ge 1$. $L$ is called solvable if $L^{(n)}=0$ for some $n$. The lower-central series of $L$ is defined by $L^2 = [L,L]$, $L^{n+1}=[L,L^n]$ for $n>1$. $L$ is called nilpotent if $L^n=0$ for some $n$. It should be noted that if $L$ is nilpotent, then $L$ must be solvable.
The nilradical of $L$ is defined to be the (unique) maximal nilpotent ideal of $L$, denoted by $\nr(L)$. It is a classical result that if $L$ is solvable, then $L^2 = [L,L] \subseteq \nr(L)$. From \cite{mubar}, we have that
\begin{equation}\label{dimension}
\dim (\nr(L)) \geq \frac{1}{2} \dim (L).
\end{equation}
The triangular algebra $T(n)$ is the $\frac{1}{2}n(n-1)$-dimensional Lie algebra whose basis is the set of strictly upper-triangular matrices, $\{N_{ik} \vert 1 \leq i < k \leq n \}$ defined by multiplications
\begin{equation}\label{tri}
[N_{ik},N_{ab}]=\delta_{ka}N_{ib} - \delta_{bi}N_{ak}.
\end{equation}
The left-annihilator of a Leibniz algebra $L$ is the ideal $\Ann_\ell(L) = \left\{x\in L\mid [x,y]=0\ \forall y\in L\right\}$. Note that the elements $[x,x]$ and $[x,y] + [y,x]$ are in $\Ann_\ell(L)$, for all $x,y\in L$, because of \eqref{Jacobi}.
An element $x$ in a Leibniz algebra $L$ is nilpotent if both $(L_x)^n = (R_x)^n = 0$ for some $n$. In other words, for all $y$ in $L$
\begin{equation*}
[x,\cdots[x,[x,y]]] = 0 = [[[y,x],x]\cdots,x].
\end{equation*}
A set of matrices $\{X_\alpha\}$ is called linearly nilindependent if no non-zero linear combination of them is nilpotent. In other words, if
\begin{equation*}
X = \displaystyle\sum_{\alpha=1}^f c_\alpha X_\alpha,
\end{equation*}
then $X^n=0$ implies that $c_\alpha=0$ for all $\alpha$. A set of elements of a Leibniz algebra $L$ is called linearly nilindependent if no non-zero linear combination of them is a nilpotent element of $L$.
\section{Classification}
Let $T(n)$ be the $\frac{1}{2}n(n-1)$-dimensional triangular (Lie) algebra over the field $F$ ($\mathbb{C}$ or $\mathbb{R}$) with basis $\{N_{ik} \vert 1 \leq i < k \leq n \}$ and products given by \eqref{tri}. We will extend $T(n)$ to a solvable Leibniz algebra $L$ of dimension $\frac{1}{2}n(n-1) + f$ by appending linearly nilindependent elements $\{X^1, \ldots, X^f\}$. In doing so, we will construct an indecomposable Leibniz algebra whose nilradical is $T(n)$.
We construct the vector $N = (N_{12} N_{23} \cdots N_{(n-1)n} N_{13} \cdots N_{(n-2)n} \cdots N_{1n})^T$ whose components are the basis elements of the nilradical ordered along consecutive off-diagonals ($N_{i(i+1)}$ in order, then $N_{i(i+2)}$ in order, \ldots). Then since $[L,L] \subseteq \nr(L)$, the brackets of $L$ are given by \eqref{tri} and
\begin{align*}
[X_\alpha,N_{ik}] &= A^\alpha_{ik,pq} N_{pq}\\
[N_{ik},X_\alpha] &= B^\alpha_{ik,pq} N_{pq}\\
[X_\alpha,X_\beta] &= \sigma^{\alpha\beta}_{pq} N_{pq}.
\end{align*}
using Einstein summation notation on repeated indices (from here onward), where $1\leq \alpha, \beta \leq f$, $A^\alpha_{ik,pq}, \sigma^{\alpha \beta}_{pq} \in F$. Note that $A^\alpha \in F^{r \times r}$, $N \in T(n)^{r \times 1}$ where $r = \frac{1}{2}n(n-1)$.
To classify Leibniz algebras $L(n,f)$ we must classify the matrices $A^\alpha$ and $B^\alpha$ and the constants $\sigma^{\alpha \beta}_{pq}$. The Jacobi identities for the triples $\{X_\alpha, N_{ik}, N_{ab}\}$, $\{N_{ik}, N_{ab}, X_\alpha\}$, $\{N_{ik}, X_\alpha, N_{ab}\}$ with $1 \leq \alpha \leq f$, $1 \leq i < k \leq n$, $1 \leq a < b \leq n$ give us respectively
\begin{align}
\tag{4a}\label{2.14}\delta_{ka} A^\alpha_{ib,pq} N_{pq} - \delta_{bi} A^\alpha_{ak,pq} N_{pq} + A^\alpha_{ik,bq} N_{aq} - A^\alpha_{ik,pa} N_{pb} - A^\alpha_{ab,kq} N_{iq} + A^\alpha_{ab,pi} N_{pk} &= 0\\
\tag{4b}\label{un2.14}\delta_{ka} B^\alpha_{ib,pq} N_{pq} - \delta_{bi} B^\alpha_{ak,pq} N_{pq} + B^\alpha_{ik,bq} N_{aq} - B^\alpha_{ik,pa} N_{pb} - B^\alpha_{ab,kq} N_{iq} + B^\alpha_{ab,pi} N_{pk} &= 0\\
\tag{4c}\label{2.14twist}\delta_{ka} A^\alpha_{ib,pq} N_{pq} - \delta_{bi} A^\alpha_{ak,pq} N_{pq} + A^\alpha_{ik,bq} N_{aq} - A^\alpha_{ik,pa} N_{pb} + B^\alpha_{ab,kq} N_{iq} - B^\alpha_{ab,pi} N_{pk} &= 0.
\end{align}
\addtocounter{equation}{1}
As a consequence of \eqref{2.14} and \eqref{2.14twist}, we also have that
$$A^\alpha_{ab,pi} N_{pk} - A^\alpha_{ab,kq} N_{iq} = - (B^\alpha_{ab,pi} N_{pk} - B^\alpha_{ab,kq} N_{iq}).$$
Thus $A^\alpha_{ab,pi} = - B^\alpha_{ab,pi}$ if $p<i$ and $A^\alpha_{ab,kq} = - B^\alpha_{ab,kq}$ if $k<q$. Therefore
\begin{equation}\label{2.14cor}
A^\alpha_{ab,ik} = - B^\alpha_{ab,ik} \quad\forall ab,ik \text{ except } ik=1n.
\end{equation}
Similarly the Jacobi identities for the triples $\{X_\alpha, X_\beta, N_{ab}\}$, $\{X_\alpha, N_{ik}, X_\beta\}$, $\{N_{ik}, X_\alpha, X_\beta\}$ with $1 \leq \alpha, \beta \leq f$ and $1 \leq i < k \leq n$ give us respectively
\begin{align}
\tag{6a}\label{2.15}[A^\alpha, A^\beta]_{ik,pq} N_{pq} =& \phantom{-(}\sigma^{\alpha \beta}_{kq} N_{iq} - \sigma^{\alpha \beta}_{pi} N_{pk}\\
\tag{6b}\label{2.15twist}[A^\alpha, B^\beta]_{ik,pq} N_{pq} =& - (\sigma^{\alpha \beta}_{kq} N_{iq} - \sigma^{\alpha \beta}_{pi} N_{pk})\\
\tag{6c}\label{un2.15}(\BbA^\alpha+B^\alphaB^\beta)_{ik,pq} N_{pq} =& \phantom{-(}\sigma^{\alpha \beta}_{kq} N_{iq} - \sigma^{\alpha \beta}_{pi} N_{pk}.
\end{align}
\addtocounter{equation}{1}
Unlike the Lie case, these give nontrivial relations for $f=1$ or $\alpha=\beta$ when $f>1$.
The Jacobi identity for the triple $\{X_\alpha, X_\beta, X^\gamma\}$ with $1 \leq \alpha, \beta, \gamma \leq f$ gives us
\begin{equation}\label{2.16}
\sigma^{\beta \gamma}_{pq} A^\alpha_{pq,ik} - \sigma^{\alpha \beta}_{pq} B^\gamma_{pq,ik} - \sigma^{\alpha \gamma}_{pq} A^\beta_{pq,ik} = 0.
\end{equation}
Again, we do not require $\alpha, \beta, \gamma$ to be distinct, which in particular gives nontrivial relations for $f \geq 1$.
In order to simplify the matrices $A^\alpha$ and $B^\alpha$ and the constants $\sigma^{\alpha \beta}_{pq}$ we will make use of several transformations which leave the commutation relations \eqref{tri} invariant. Namely
\begin{itemize}
\item Redefining the elements of the extension:
\begin{equation}\label{2.17}
\begin{split}
&\hspace{23pt}X_\alpha \longrightarrow X_\alpha + \mu^\alpha_{pq} N_{pq}, \quad \mu^\alpha_{pq} \in F \\
\Rightarrow
&\begin{cases}
A^\alpha_{ik,ab} \longrightarrow A^\alpha_{ik,ab} + \delta_{kb}\mu^\alpha_{ai} - \delta_{ia}\mu^\alpha_{kb}\\
B^\alpha_{ik,ab} \longrightarrow B^\alpha_{ik,ab} - \delta_{kb}\mu^\alpha_{ai} + \delta_{ia}\mu^\alpha_{kb}.
\end{cases}
\end{split}
\end{equation}
\item Changing the basis of $\nr(L)$:
\begin{equation}\label{2.18}
\begin{split}
&\hspace{17pt}N \longrightarrow GN, \quad G \in GL(r,F) \\
\Rightarrow
&\begin{cases}
A^\alpha \longrightarrow GA^\alpha G^{-1}\\
B^\alpha \longrightarrow GB^\alpha G^{-1}.
\end{cases}
\end{split}
\end{equation}
\item Taking a linear combination of the elements $X_\alpha$.
\end{itemize}
The matrix $G$ must satisfy certain restrictions discussed later in order to preserve the commutation relations \eqref{tri} of $\nr(L)$.
Note that $N_{1n}$ is not used in \eqref{2.17} since it commutes with all the elements in $\nr(L)$. Since \eqref{2.16} gives relations between the matrices $A^\alpha$, $B^\alpha$ and the constants $\sigma^{\alpha \beta}_{pq}$, the unused constant $\mu^\alpha_{1n}$ can be used to scale the constants $\sigma^{\alpha \beta}_{pq}$ when $f \geq 2$:
\begin{equation}\label{2.19}
\begin{split}
X_\alpha &\longrightarrow X_\alpha + \mu^\alpha_{1n} N_{1n}, \quad \mu^\alpha_{1n} \in F \\
\Rightarrow \sigma^{\alpha \beta}_{pq} &\longrightarrow \sigma^{\alpha \beta}_{pq} + \mu^\beta_{1n}A^\alpha_{1n,pq} + \mu^\alpha_{1n}B^\beta_{1n,pq}.
\end{split}
\end{equation}
In this transformation $A^\alpha$ is invariant, so we will be able to simplify some constants $\sigma^{\alpha \beta}_{pq}$.
\section{Extensions of $T(4)$}
In this paper we will focus on triangular algebras $T(n)$ with $n\geq 4$ because:
\begin{itemize}
\item $T(2)$ is a one-dimensional algebra (hence by \eqref{dimension} $L$ has dimension at most 2) and the Jacobi identity gives that the only family of solvable non-Lie Leibniz algebras with one-dimensional nilradical are given by $L(c)= \langle a,b \rangle$ with $[a,a]=[a,b]=0$, $[b,a]=ca$, $[b,b]=a$ where $c \neq 0 \in F$.
\item $T(3)$ is a Heisenberg Lie algebra, and Leibniz algebras with Heisenberg nilradical were classified in \cite{heisen}.
\end{itemize}
Now we will consider the case when $n = 4$. In particular, $N = (N_{12} N_{23} N_{34} N_{13} N_{24} N_{14})^T$ and $r = 6$.
We can proceed by considering the relations in \eqref{2.14} and \eqref{un2.14} for $1 \leq i < k \leq 4$, $1 \leq a < b \leq 4$, $k \neq a$, $b \neq i$
\begin{align}
\tag{11a}\label{3.3}A^\alpha_{ik,bq} N_{aq} - A^\alpha_{ik,pa} N_{pb} - A^\alpha_{ab,kq} N_{iq} + A^\alpha_{ab,pi} N_{pk} &= 0\\
\tag{11b}\label{un3.3}B^\alpha_{ik,bq} N_{aq} - B^\alpha_{ik,pa} N_{pb} - B^\alpha_{ab,kq} N_{iq} + B^\alpha_{ab,pi} N_{pk} &= 0.
\end{align}
\addtocounter{equation}{1}
Similarly for $1 \leq i < k = a < b \leq 4$, we obtain
\begin{align}
\tag{12a}\label{3.2}A^\alpha_{ib,pq} N_{pq} + A^\alpha_{ik,bq} N_{kq} - A^\alpha_{ik,pk} N_{pb} - A^\alpha_{kb,kq} N_{iq} + A^\alpha_{kb,pi} N_{pk} &= 0\\
\tag{12b}\label{un3.2}B^\alpha_{ib,pq} N_{pq} + B^\alpha_{ik,bq} N_{kq} - B^\alpha_{ik,pk} N_{pb} - B^\alpha_{kb,kq} N_{iq} + B^\alpha_{kb,pi} N_{pk} &= 0.
\end{align}
\addtocounter{equation}{1}
Using the linear independence of the $N_{ik}$ with equation \eqref{3.3}, we can obtain relationships among the entries of the matrices $A^\alpha$, summarized in the matrix below. For example, letting $ik = 12$ and $ab = 34$, the coefficient of $N_{14}$ gives $A^\alpha_{12,13} + A^\alpha_{34,24} = 0$, the coefficient of $N_{13}$ gives $A^\alpha_{34,23} = 0$, and the coefficient of $N_{24}$ gives $A^\alpha_{12,23} = 0$.
Using equation \eqref{un3.3}, we obtain the same relationships among the entries of $B^\alpha$.
\begin{align*}
A^\alpha =&
\begin{pmatrix}
*&0&A^\alpha_{12,34}&A^\alpha_{12,13}&*&*\\
0&*&0&*&*&*\\
A^\alpha_{34,12}&0&*&*&-(A^\alpha_{12,13})&*\\
0&0&0&*&(A^\alpha_{12,34})&*\\
0&0&0&(A^\alpha_{34,12})&*&*\\
0&0&0&0&0&*
\end{pmatrix}
\end{align*}
Applying \eqref{3.2} and \eqref{un3.2} in the same way, $A^\alpha$ becomes:
\begin{align*}
A^\alpha =&
\begin{pmatrix}
A^\alpha_{12,12}&0&0&A^\alpha_{12,13}&*&*\\
&A^\alpha_{23,23}&0&A^\alpha_{23,13}&A^\alpha_{23,24}&*\\
&&A^\alpha_{34,34}&*&-(A^\alpha_{12,13})&*\\
&&&A^\alpha_{12,12} + A^\alpha_{23,23}&0&(A^\alpha_{23,24})\\
&&&&A^\alpha_{23,23} + A^\alpha_{34,34}&(A^\alpha_{23,13})\\
&&&&&A^\alpha_{12,12} + A^\alpha_{23, 23} + A^\alpha_{34,34}
\end{pmatrix}
\end{align*}
As before, $B^\alpha$ has the same form above. This with \eqref{2.14cor} implies that $B^\alpha_{13,14} = B^\alpha_{23,24} = -A^\alpha_{23,24} = -A^\alpha_{13,14}$. Similarly, $B^\alpha_{24,14} = -A^\alpha_{24,14}$ and $B^\alpha_{14,14} = -A^\alpha_{14,14}$.
We can further simplify matrix $A^\alpha$ by performing the transformation specified in \eqref{2.17}. Choosing $\mu^\alpha_{12} = -A^\alpha_{23,13}$, $\mu^\alpha_{23} = A^\alpha_{12,13}$, $\mu^\alpha_{34} = A^\alpha_{23,24}$, $\mu^\alpha_{13} = -A^\alpha_{34,14}$, and $\mu^\alpha_{24} = -A^\alpha_{12,14}$ leads to
$$A^\alpha_{23,13} = A^\alpha_{12,13} = A^\alpha_{23,24} = A^\alpha_{34,14} = A^\alpha_{12,14} = 0.$$
This gives us the matrices
\begin{align*}
A^\alpha =&
\begin{pmatrix}
A^\alpha_{12,12}&0&0&0&A^\alpha_{12,24}&0\\
&A^\alpha_{23,23}&0&0&0&A^\alpha_{23,14}\\
&&A^\alpha_{34,34}&A^\alpha_{34,13}&0&0\\
&&&A^\alpha_{13,13}&0&0\\
&&&&A^\alpha_{24,24}&0\\
\phantom{-A^\alpha_{12,12}}&\phantom{-A^\alpha_{23,23}}&\phantom{-A^\alpha_{34,34}}&\phantom{-A^\alpha_{13,13}}&\phantom{-A^\alpha_{24,24}}&\,\,A^\alpha_{14,14}\,\,
\end{pmatrix}\\
B^\alpha =&
\begin{pmatrix}
-A^\alpha_{12,12}&0&0&0&-A^\alpha_{12,24}&B^\alpha_{12,14}\\
&-A^\alpha_{23,23}&0&0&0&B^\alpha_{23,14}\\
&&-A^\alpha_{34,34}&-A^\alpha_{34,13}&0&B^\alpha_{34,14}\\
&&&-A^\alpha_{13,13}&0&0\\
&&&&-A^\alpha_{24,24}&0\\
&&&&&-A^\alpha_{14,14}
\end{pmatrix}
\end{align*}
$$A^\alpha_{ik,ik} = \sum^{k-1}_{p = i}A^\alpha_{p(p+1),p(p+1)}.$$
Note that $A^\alpha_{12,12}$, $A^\alpha_{23,23}$, and $A^\alpha_{34,34}$ cannot simultaneously equal 0, otherwise the nilradical would no longer be $T(4)$. The nilindependence among the $A^\alpha$ implies that $T(4)$ can have at most a three-dimensional extension, since there are three parameters on the diagonal.
The form of the matrices $A^\alpha$ implies that the only nonzero elements of $[A^\alpha, A^\beta]$ are
$$[A^\alpha, A^\beta]_{12,24},\quad [A^\alpha, A^\beta]_{23,14},\quad [A^\alpha, A^\beta]_{34,13}.$$
The linear independence of the $N_{ik}$ with equation \eqref{2.15}, yields
\begin{align}
\label{Acommute}[A^\alpha,A^\beta] =& \ 0\\
\label{3.12}[X_\alpha,X_\beta] =& \ \sigma^{\alpha \beta}_{14} N_{14}.
\end{align}
For example, letting $ik = 12$ in equation \eqref{2.15}, the coefficient of $N_{24}$ implies that $[A^\alpha,A^\beta]_{12,24}=0$ and the coefficient of $N_{14}$ implies that $\sigma^{\alpha \beta}_{24} = [A^\alpha,A^\beta]_{12,14} = 0$. Henceforth we will abbreviate $\sigma^{\alpha \beta}_{14} = \sigma^{\alpha \beta}$ as all other $\sigma^{\alpha \beta}_{pq} = 0$. Since the $A^\alpha$ commute by \eqref{Acommute}, \eqref{2.15} and \eqref{2.15twist} imply
\begin{equation}\label{ABcommute}
[A^\alpha,B^\beta]=0.
\end{equation}
Considering \eqref{ABcommute} componentwise, we find that $B^\beta_{12,14}(A^\alpha_{14,14}-A^\alpha_{12,12}) = B^\beta_{34,14}(A^\alpha_{14,14}-A^\alpha_{34,34}) = (A^\beta_{23,14} + B^\beta_{23,14})(A^\alpha_{14,14}-A^\alpha_{23,23}) = 0$.
Furthermore, by \eqref{ABcommute} and \eqref{un2.15}, $0=B^\alphaA^\beta+\BbB^\alpha=(A^\beta+B^\beta)B^\alpha$. Componentwise, this tells us that
\begin{equation}\label{LindseyLemma}
0 = B^\beta_{12,14}A^\alpha_{14,14} = B^\beta_{34,14}A^\alpha_{14,14} = (A^\beta_{23,14} + B^\beta_{23,14})A^\alpha_{14,14}.
\end{equation}
In particular, if $B^\beta$ has a nontrivial off-diagonal entry, then
\begin{equation}\label{offdiag}
\begin{cases}
B^\beta_{12,14}\ne 0 &\Rightarrow A^\alpha_{12,12}=A^\alpha_{14,14}=0,\ \forall\alpha \\
B^\beta_{34,14}\ne 0 &\Rightarrow A^\alpha_{34,34}=A^\alpha_{14,14}=0,\ \forall\alpha \\
B^\beta_{23,14}\ne -A^\beta_{23,14} &\Rightarrow A^\alpha_{23,23}=A^\alpha_{14,14}=0,\ \forall\alpha.
\end{cases}
\end{equation}
The form of the matrices $A^\alpha$ and $B^\alpha$ imply that \eqref{2.16} becomes
\begin{equation}\label{3.13}
\sigma^{\alpha \beta}A^\gamma_{14,14}-\sigma^{\alpha \gamma}A^\beta_{14,14}+\sigma^{\beta \gamma}A^\alpha_{14,14}=0.
\end{equation}
By adding equations of form \eqref{3.13}, we get
\begin{equation}\label{Lieish}
(\sigma^{\beta \gamma} + \sigma^{\gamma \beta})A^\alpha_{14,14} = 0.
\end{equation}
For example $(\sigma^{12}+\sigma^{21})A^\alpha_{14,14}=0$ is obtained by adding \eqref{3.13} with $\beta=1, \gamma=2$ to \eqref{3.13} with $\beta=2, \gamma=1$.
On a related note, since $[X_\beta,X_\beta] \in \Ann_\ell(L)$, we have $0=[[X_\beta,X_\beta],X_\alpha]=[\sigma^{\beta \beta}N_{14},X_\alpha]= - \sigma^{\beta \beta} A^\alpha_{14,14} N_{14}$. Thus,
\begin{equation}\label{Lieish2}
\sigma^{\beta \beta} A^\alpha_{14,14} = 0.
\end{equation}
As a consequence of \eqref{2.14cor}, \eqref{3.12}, \eqref{LindseyLemma}, \eqref{Lieish}, and \eqref{Lieish2}, we have the following result.
\begin{lemma}\label{megalem}
If $A^\alpha_{14,14}\ne0$ for any $\alpha=1,\ldots,f$, then the Leibniz algebra is a Lie algebra.
\end{lemma}
The form of the matrices $A^\alpha$ and $B^\alpha$ imply that the transformation \eqref{2.19} becomes
\begin{equation}\label{3.14}
\sigma^{\alpha \beta} \longrightarrow \sigma^{\alpha \beta} + \mu^\beta_{14}A^\alpha_{14,14} - \mu^\alpha_{14}A^\beta_{14,14}.
\end{equation}
For $f=2$, suppose $A^\alpha_{14,14} \neq 0$ for some $\alpha$, and without loss of generality assume that $\alpha=1$. Then choosing $\mu^1_{14}=0$ and $\mu^2_{14}=-\frac{\sigma^{12}}{A^1_{14,14}}$; \eqref{3.14} makes $\sigma^{12}=0$. For $f=3$, suppose $A^\alpha_{14,14} \neq 0$ for some $\alpha$, and without loss of generality assume that $\alpha=1$. Then choosing $\mu^1_{14}=0$ and $\mu^\beta_{14}=-\frac{\sigma^{1\beta}}{A^1_{14,14}}$ for $\beta=2,3$; \eqref{3.14} makes $\sigma^{1\beta}=0$. By \eqref{3.13}, we also have $\sigma^{23}=0$. Combining these results for $f=2$, 3 and employing \eqref{Lieish2} for $f=1$, we have:
\begin{equation}\label{3.15}
[X_\alpha,X_\beta] =
\begin{cases}
\sigma^{\alpha \beta}N_{14} & \text{if }A^1_{14,14}=\cdots=A^f_{14,14}=0 \\
0 & \text{otherwise.}
\end{cases}
\end{equation}
Now we will utilize matrices $G$ to simplify the structure of matrices $A^\alpha$ and $B^\alpha$. Perform transformation \eqref{2.18}, given by $N\longrightarrow G_1 N$, with
$$G_1 = \begin{pmatrix}
1&0&0&0&g_1&g_0\\
&1&0&0&0&g_2 \\
& &1&g_3&0&g_4\\
& & &1&0&0\\
& & & &1&0\\
& & & & &1
\end{pmatrix}.$$
Observe that $G_1$ acts invariantly on the commutation relations \eqref{tri}. It does, however, transform matrices $A^\alpha$ and $B^\alpha$ by $A^\alpha\longrightarrow G_1 A^\alpha G_1^{-1}$ and $B^\alpha\longrightarrow G_1 B^\alpha G_1^{-1}$, respectively. In particular, $G_1$ transforms the following components
\begin{eqnarray*}
\begin{cases}
A^\alpha_{12,24} \longrightarrow A^\alpha_{12,24} + g_1(A^\alpha_{24,24}-A^\alpha_{12,12}) \\
A^\alpha_{23,14} \longrightarrow A^\alpha_{23,14} + g_2(A^\alpha_{14,14}-A^\alpha_{23,23}) \\
A^\alpha_{34,13} \longrightarrow A^\alpha_{34,13} + g_3(A^\alpha_{13,13}-A^\alpha_{34,34})
\end{cases}\\
\begin{cases}
B^\alpha_{12,14} \longrightarrow B^\alpha_{12,14} - g_0(A^\alpha_{14,14}-A^\alpha_{12,12}) \\
B^\alpha_{23,14} \longrightarrow B^\alpha_{23,14} - g_2(A^\alpha_{14,14}-A^\alpha_{23,23}) \\
B^\alpha_{34,14} \longrightarrow B^\alpha_{34,14} - g_4(A^\alpha_{14,14}-A^\alpha_{34,34}).
\end{cases}
\end{eqnarray*}
We use the matrix $G_1$ to eliminate some entries in $A^\alpha$ and $B^\alpha$. However, if $B^\alpha_{12,14}$ or $B^\alpha_{34,14}$ is not zero, then by \eqref{offdiag}, $G_1$ leaves that entry of $B^\alpha$ invariant. Hence, we can use $g_1$, $g_2$, and $g_3$ to eliminate at most 3 off-diagonal elements.
Note: In this step, we consider only transformations of the form $G_1=\begin{pmatrix} I&*\\0&I\end{pmatrix}$. Any other transformation which leaves \eqref{tri} and the form of $A^\alpha$ invariant, but eliminates $B^\alpha_{12,14}$ or $B^\alpha_{34,14}$,
would provide an isomorphism from %
a non-Lie Leibniz algebra to a Lie algebra.
It is, however, possible to scale such entries, which we will consider in the next case.
Let the diagonal matrix $G_2$ be
$$G_2=\begin{pmatrix}
g_{12}&&&&&\\
&g_{23}&&&&\\
&&g_{34}&&&\\
&&&g_{12}g_{23}&&\\
&&&&g_{23}g_{34}&\\
&&&&&g_{12}g_{23}g_{34}
\end{pmatrix},\quad g_{ik}\in F\backslash\{0\}.$$
Note that $G_2$ preserves commutation relations \eqref{tri}. Our transformation of $\nr(L)$ will be defined by $G=G_2 G_1$. Observe that $G_2$ transforms $A^\alpha$ and $B^\alpha$ by $A^\alpha_{ik,ab}\longrightarrow \dfrac{g_{ik}}{g_{ab}}A^\alpha_{ik,ab}$ and $B^\alpha_{ik,ab}\longrightarrow \dfrac{g_{ik}}{g_{ab}}B^\alpha_{ik,ab}$, respectively, where $g_{ik}=(G_2)_{ik}=\prod\limits^{k-1}_{j=i}g_{j(j+1)}$. Hence, we can scale up to three nonzero off-diagonal elements to 1. In the case of Lie algebras, it may be necessary to scale to $\pm 1$ over $F=\mathbb{R}$. This issue does not arise in Leibniz algebras of non-Lie type, because we have greater restrictions on the number of nonzero entries.
\subsection{Leibniz algebras $L(4,1)$}
The Lie cases for $\nr(L)=T(4)$, with $f=1$, have been previously classified in \cite{tw}, so we will focus on the Leibniz algebras of non-Lie type. We know that all such algebras will have $A^1_{14,14}=0$ by Lemma \ref{megalem}, where $A=A^1$ will be of the form found in \cite{tw}. Since $A$ is not nilpotent and $A_{14,14}=0$, we know that there is at most 1 nonzero off-diagonal entry in $A$. Altogether, there are 10 classes of Leibniz algebras of non-Lie type. Of these, there are 2 two-dimensional families, and 8 one-dimensional families. The matrices $A$ and $B$ for these can be found in Table \ref{L41} in the appendix.
\subsection{Leibniz algebras $L(4,2)$}
Again, all Lie cases for $\nr(L)=T(4)$, with $f=2$, were classified in \cite{tw}. So, focusing on Leibniz algebras of non-Lie type, we again require that $A^1_{14,14}=A^2_{14,14}=0$ by Lemma \ref{megalem}. As a result, if $B^\alpha_{ik,14}\ne-A^\alpha_{ik,14}$, for any $\alpha$ or pair $ik$, then $A^\beta_{ik,ik}=A^\beta_{14,14}=0$ $\forall \beta$, which makes it impossible to have two linearly nilindependent matrices $A^1$ and $A^2$. Therefore, there is 1 four-dimensional family of Leibniz algebras of non-Lie type, and their matrices $A^1=-B^1$ and $A^2=-B^2$ can be found in Table \ref{L42} in the appendix.
\subsection{Leibniz algebras $L(4,3)$}
There is only one Lie algebra that is a three-dimensional extension of $T(4)$, again given in \cite{tw}. Since we cannot have three linearly nilindependent matrices of form $A^\alpha$ with $A^1_{14,14} = A^2_{14,14} = A^3_{14,14} = 0$, it is impossible to have a three-dimensional extention of $T(4)$ that is of non-Lie type, by Lemma \ref{megalem}.
\begin{theorem}\label{L4f}
Every Leibniz algebra $L(4,f)$ is either of Lie type, or is isomorphic to precisely one algebra represented in Table \ref{L41} or Table \ref{L42}.
\end{theorem}
\section{Solvable Lie algebras $L(n,f)$ for $n\ge4$}
We are now going to consider Leibniz algebras $L$ with $\nr(L)=T(n)$. Recall from \eqref{2.14cor}, $A^\alpha_{ik,ab} = -B^\alpha_{ik,ab}$ for all $ab\ne 1n$. We have the following result:
\begin{lemma}\label{Astructure}
Matrices $A^\alpha=(A^\alpha_{ik,ab})$ and $B^\alpha=(B^\alpha_{ik,ab})$, $1\le i<k\le n$, $1\le a < b\le n$ have the following properties.
\begin{enumerate}
\item[i.] $A^\alpha$ and $B^\alpha$ are upper-triangular.
\item[ii.] The only off-diagonal elements of $A^\alpha$ and $B^\alpha$ which may not be eliminated by an appropriate transformation on $X_\alpha$ are:
\begin{align*}\label{lem1.2A}
A^\alpha_{12,2n},\quad A^\alpha_{j(j+1),1n} \ (2\le j\le n-2),\quad A^\alpha_{(n-1)n,1(n-1)}, \\
B^\alpha_{12,2n},\quad B^\alpha_{j(j+1),1n} \ (1\le j\le n-1),\quad B^\alpha_{(n-1)n,1(n-1)}.
\end{align*}
\item[iii.] The diagonal elements $A^\alpha_{i(i+1),i(i+1)}$ and $B^\alpha_{i(i+1),i(i+1)}$, $1\le i\le n-1$, are free. The remaining diagonal elements of $A^\alpha$ and $B^\alpha$ satisfy
\begin{equation*}\label{lem1.3}
A^\alpha_{ik,ik} = \sum\limits^{k-1}_{j=i}A^\alpha_{j(j+1),j(j+1)},\quad B^\alpha_{ik,ik} = \sum\limits^{k-1}_{j=i}B^\alpha_{j(j+1),j(j+1)},\quad k>i+1.
\end{equation*}
\end{enumerate}
\end{lemma}
\begin{proof}
The form of the matrices $A^\alpha$ given in Lemma \ref{Astructure} follows from \eqref{2.14} by induction on $n$, as shown in \cite{tw}. Similarly, properties $i.$ and $iii.$ follow for $B^\alpha$ from \eqref{un2.14}. Property $ii.$ for matrices $B^\alpha$ follows from \eqref{2.14cor}.
\end{proof}
As a consequence of property $iii.$ and \eqref{2.14cor}, we have that $A^\alpha_{1n,1n}=-B^\alpha_{1n,1n}$.
Lemma \ref{Astructure} asserts that $A^\alpha$ has $n-1$ free entries on the diagonal and nonzero off-diagonal entries possible in only $n-1$ locations, represented by $*$ in the matrix below. The form of $B^\alpha$ is the same, save for nonzero off-diagonal entries possible in two additional locations, represented by $b_1$ and $b_2$ in the matrix below.
$\hfill\left(
\begin{array}{ccccc|ccccc}
* &&&& & && &*&b_1 \\
& * &&& & && &&* \\
&& \ddots && & && &&\vdots \\
&&& * & & && && * \\
&&&& * & && *&&b_2 \\
\hline
&&&& & *&&&&\\
&&&& & & \ddots &&& \\
&&&& & &&*&&\\
&&&& & &&&*&\phantom{\ddots}\\
\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&\phantom{\ddots}&*
\end{array}
\right)
\hfill$
\begin{lemma}
The maximum degree of an extension of $T(n)$ is $f=n-1$.
\end{lemma}
\begin{proof}
The proof follows from the fact that the $A^\alpha$ are nilindependent and that we have, at most, $n-1$ parameters along the diagonal.
\end{proof}
The form of the matrices $A^\alpha$ implies that the only nonzero elements of $[A^\alpha, A^\beta]$ are
$$[A^\alpha,A^\beta]_{12,2n},\quad [A^\alpha,A^\beta]_{j(j+1),1n} \ (2\le j\le n-2),\quad [A^\alpha,A^\beta]_{(n-1)n,1(n-1)}.$$
As before, the linear independence of the $N_{ik}$ with equations \eqref{2.15} and \eqref{2.15twist}, yields
\begin{align}
\label{Acommute2}[A^\alpha,A^\beta] =& \ 0\\
\label{ABcommute2}[A^\alpha,B^\beta]=& \ 0\\
\nonumber [X_\alpha,X_\beta] =& \ \sigma^{\alpha \beta}_{1n} N_{1n}.
\end{align}
From Lemma \ref{Astructure}, \eqref{2.16} becomes
\begin{equation*}\label{3.13b}
\sigma^{\alpha \beta}A^\gamma_{1n,1n}-\sigma^{\alpha \gamma}A^\beta_{1n,1n}+\sigma^{\beta \gamma}A^\alpha_{1n,1n}=0.
\end{equation*}
\begin{lemma}\label{commutelem}
Matrices $A^\alpha$ and $B^\alpha$ can be transformed to a canonical form satisfying
\begin{align*
[A^\alpha, A^\beta] &= 0 \\
[A^\alpha, B^\beta] &= 0 \\
[X_\alpha,X_\beta] &=
\begin{cases}
\sigma^{\alpha \beta}N_{1n} & \text{if }A^1_{1n,1n}=\cdots=A^f_{1n,1n}=0 \\
0 & \text{otherwise.}
\end{cases}
\end{align*}
\end{lemma}
\begin{proof}
The first two identities of this lemma have already been shown. The argument for the third identity is the same as \eqref{3.15}, using $(\sigma^{\beta \gamma} + \sigma^{\gamma \beta})A^\alpha_{1n,1n} = 0$ and $\sigma^{\beta \beta} A^\alpha_{1n,1n} = 0$.
\end{proof}
\subsection{Change of basis in $\nr(L(n,f))$}
As before, we perform the transformation \eqref{2.18} on $N$ by use of the matrix $G_1$, with $G_1$ defined to be all zeroes except $(G_1)_{ik,ik} = 1$, and $(G_1)_{12,1n}, (G_1)_{(n-1)n,1n}, (G_1)_{ab,ik} \in F$ where $\{ab,ik\}=\{12,2n\}, \{(n-1)n,1(n-1)\}, \{j(j+1),1n\}$ for $2 \leq j \leq n-2$. Therefore, the zero entries of $G_1$ are the off-diagonal entries of $B^\alpha$ that are guaranteed to be zero. Note that the transformation given by $G_1$ preserves commutation relation \eqref{tri}. Matrices $A^\alpha$ and $B^\alpha$ are transformed by conjugation with $G_1$, leaving the diagonal elements invariant and giving
\begin{eqnarray*}
\begin{split}
A^\alpha_{ik,ab} &\longrightarrow A^\alpha_{ik,ab} + (G_1)_{ik,ab}(A^\alpha_{ab,ab}-A^\alpha_{ik,ik}) \\
B^\alpha_{ik,ab} &\longrightarrow B^\alpha_{ik,ab} - (G_1)_{ik,ab}(A^\alpha_{ab,ab}-A^\alpha_{ik,ik}).
\end{split}
\end{eqnarray*}
By \eqref{ABcommute2}, we have that $0=[A^\alpha,B^\beta]_{12,1n} = B^\beta_{12,1n}(A^\alpha_{12,12}-A^\alpha_{1n,1n})$ and $0=[A^\alpha,B^\beta]_{(n-1)n,1n} = B^\beta_{(n-1)n,1n}(A^\alpha_{(n-1)n,(n-1)n}-A^\alpha_{1n,1n})$. Consequently, $G_1$ cannot eliminate the entries $B^\beta_{12,1n}$ and $B^\beta_{(n-1)n,1n}$.
\begin{lemma}\label{lemma4}
Matrices $A^\alpha$ and $B^\alpha$ will have a nonzero off-diagonal entry $A^\alpha_{ik,ab}$ or $B^\alpha_{ik,ab}$, respectively, only if
$$A^\beta_{ik,ik}=A^\beta_{ab,ab},\quad \forall\beta=1,\ldots,f.$$
\end{lemma}
\begin{proof}
For $A^\alpha$, the proof of Lemma \ref{lemma4} follows from \eqref{Acommute2}, as shown in \cite{tw}. Similarly for $B^\alpha$, the proof of Lemma \ref{lemma4} follows from considering \eqref{ABcommute2} componentwise. Namely, we find that $0=[A^\beta,B^\alpha]_{12,1n}=B^\alpha_{12,1n}(A^\beta_{1n,1n}-A^\beta_{12,12})$,
and similar identities for $B^\alpha_{12,2n}$ and $B^\alpha_{(n-1)n,1(n-1)}$.
The commutation relations for $[A^\alpha,B^\beta]_{j(j+1),1n}$ gives that $B^\alpha_{j(j+1),1n}(A^\beta_{1n,1n}-A^\beta_{j(j+1),j(j+1)})=-A^\beta_{j(j+1),1n}(A^\alpha_{1n,1n}-A^\alpha_{j(j+1),j(j+1)})=0$.
\end{proof}
Now consider a second transformation $G_2$ given by $N \longrightarrow G_2 N$, where $G_2$ is the diagonal matrix $(G_2)_{ik,ik} = g_{ik}$ and $g_{ik}=\prod\limits_{j=i}^{k-1} g_{j(j+1)}$. The matrices $A^\alpha$ and $B^\alpha$ are transformed by conjugation by $G_2$. Thus $A^\alpha_{ik,ab} \longrightarrow \dfrac{g_{ik}}{g_{ab}} A^\alpha_{ik,ab}$ and $B^\alpha_{ik,ab} \longrightarrow \dfrac{g_{ik}}{g_{ab}} B^\alpha_{ik,ab}$. This transformation can be used to scale up to $n-1$ nonzero off-diagonal elements to 1. For Lie algebras over the field $F=\mathbb{R}$ it may be that some entries have to be scaled to $-1$.
Since the only non-Lie cases occur when $A^\beta_{1n,1n}=0$ for all $\beta$, then $B^\alpha_{ik,1n} \ne -A^\alpha_{ik,1n}$ implies $A^\beta_{ik,ik}=0$ for all $\beta$ by Lemma \ref{lemma4}. Since the extensions $X_\alpha$ are required to be nilindependent, this imposes restrictions on the degree $f$ of non-Lie extensions of $T(n)$. In particular, this implies that in the maximal case $f=n-1$, $L(n,n-1)$ must be Lie. Such algebras have been classified in \cite{tw}, and in fact there is a unique algebra $L(n,n-1)$ where all $A^\alpha$ are diagonal and the $X_\alpha$ commute.
\begin{theorem}
Every solvable Leibniz algebra $L(n,f)$ with triangular nilradical $T(n)$ has dimension $d=\frac{1}{2}n(n-1) + f$ with $1 \leq f \leq n-1$. It can be written in a basis $\{X_\alpha, N_{ik}\}$ with $\alpha= 1, \ldots, f$, $1 \leq i < k \leq n$ satisfying
$$[N_{ik},N_{ab}]=\delta_{ka}N_{ib} - \delta_{bi}N_{ak}$$
$$[X_\alpha,N_{ik}]=A^\alpha_{ik,pq}N_{pq}$$
$$[N_{ik},X_\alpha]=B^\alpha_{ik,pq}N_{pq}$$
$$[X_\alpha,X_\beta]=\sigma^{\alpha \beta} N_{1n}.$$
Furthermore, the matrices $A^\alpha$ and $B^\alpha$ and the constants $\sigma^{\alpha \beta}$ satisfy:
\begin{enumerate}
\item[i.] The matrices $A^\alpha$ are linearly nilindependent and $A^\alpha$ and $B^\alpha$ have the form specified in Lemma \ref{Astructure}. $A^\alpha$ commutes with all these matrices, i.e. $[A^\alpha,A^\beta]=[A^\alpha,B^\beta]=0$.
\item[ii.] $B^\alpha_{ik,ab}=-A^\alpha_{ik,ab}$ for $ab \neq 1n$, and $B^\alpha_{1n,1n}=-A^\alpha_{1n,1n}$.
\item[iii.] $L$ is Lie and all $\sigma^{\alpha \beta}=0$, unless $A^\gamma_{1n,1n}=0$ for $\gamma=1, \ldots, f$.
\item[iv.] The remaining off-diagonal elements $A^\alpha_{ik,ab}$ and $B^\alpha_{ik,ab}$ are zero, unless $A^\beta_{ik,ik}=A^\beta_{ab,ab}$ for $\beta=1, \ldots, f$.
\item[v.] In the maximal case $f=n-1$, there is only one algebra, which is isomorphic to the Lie algebra with all $A^\alpha$ diagonal where all $X_\alpha$ commute.
\end{enumerate}
\end{theorem}
\vspace{12pt}
\noindent{\bf Acknowledgements.}
The authors gratefully acknowledge the support of the Departments of Mathematics at Spring Hill College, the University of Texas at Tyler, and West Virginia University: Institute of Technology.
\newpage
\begin{table}[b]
\tiny
\caption{The Leibniz algebras $L(4,1)$ of non-Lie type
\begin{tabular}{lllc}
\hline
No. & $A$ & $B$ & $\sigma$, parameters \\
\hline\hline
(1)
&
\left(\begin{array}{cccccc}
\ 1 & & &&& \\
& a & &&& \\
& & -1-a &&&\\
& & & 1+a && \\
& & & &-1& \\
& & & && 0\\
\end{array}\right)$ & $B=-A$ &
$\begin{array}{l}\sigma^{11} \neq 0 \in F\\ a\ne-1 \in F\end{array}$ \\
\hline
(2)
&
\left(\begin{array}{cccccc}
\ 1 & & &&& \\
& 0 & \ &&& \\
& & -1\ &&&\\
& & & 1 && \\
& & &&-1& \\
& & & && 0\\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
\ -1 & & &&& \\
& 0 & \ &&& 1\\
& & 1\ &&&\\
& & & -1 && \\
& & & &1& \\
& & & && 0\\
\end{array}\right)$ &
$\sigma^{11}\in F$ \\
\hline
(3)
&
\left(\begin{array}{cccccc}
\ 1 & & &&& \\
& -1 & &&& \\
& & 0 &&&\\
& & & 0 && \\
& & &&-1& \\
& & & && 0\\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
\ -1 & & &&& \\
& 1& &&& \\
& & 0 &&&1\\
& & & 0&& \\
& & & &1& \\
& & & && 0\\
\end{array}\right)$ &
$\sigma^{11} \in F$ \\
\hlin
(4) &
\left(\begin{array}{cccccc}
0 & & & & & \\
& 1 & & & & \\
& & -1 & & & \\
& & & 1 & & \\
& & & & 0 & \\
& & & & & 0 \\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
0 & & & & &1 \\
& -1 & & & & \\
& & 1 & & & \\
& & & -1 & & \\
& & & & 0 & \\
& & & & & 0 \\
\end{array}\right)$ & $\sigma^{11} \in F$ \\
\hline
(5) &
\left(\begin{array}{cccccc}
0 & & & & & \\
& 1 & & & & \\
& & -1 & & & \\
& & & 1 & & \\
& & & &0 & \\
& & & & & 0 \\
\end{array}\right)$ & $B = -A$ & $\sigma^{11} \neq 0 \in F$ \\
\hlin
(6) &
\left(\begin{array}{cccccc}
0 & & & & 1 & \\
& 1 & & & & \\
& & -1 & & & \\
& & & 1 & & \\
& & & & 0 & \\
& & & & & 0 \\
\end{array}\right)$ & $B = -A$ & $\sigma^{11} \neq 0 \in F$ \\
\hline
(7) &
\left(\begin{array}{cccccc}
0 & & & & 1 & \\
& 1 & & & & \\
& & -1 & & & \\
& & & 1 & & \\
& & & &0 & \\
& & & & & 0 \\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
0 & & & & -1 & 1 \\
& -1 & & & & \\
& & 1 & & & \\
& & & -1 & & \\
& & & & 0 & \\
& & & & & 0 \\
\end{array}\right)$ & $\sigma^{11} \in F$ \\
\hlin
(8) &
\left(\begin{array}{cccccc}
1 & & & & & \\
& 0 & & & & 1 \\
& & -1 & & & \\
& & & 1 & & \\
& & & & -1 & \\
& & & & & 0 \\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
-1 & & & & & \\
& 0 & & & & b \\
& & 1 & & & \\
& & & -1 & & \\
& & & & 1 & \\
& & & & & 0 \\
\end{array}\right)$ &
$\begin{array}{c}
\sigma^{11}, b \in F\\
\\
\sigma^{11}, b+1 \text{ not both zero}
\end{array}
$ \\
\hlin
(9) &
\left(\begin{array}{cccccc}
1 & & & & & \\
& -1 & & & & \\
& & 0 & 1 & & \\
& & & 0 & & \\
& & & & -1 & \\
& & & & & 0 \\
\end{array}\right)$ & $B= -A$ & $\sigma^{11} \neq 0 \in F$ \\
\hline
(10) &
\left(\begin{array}{cccccc}
1 & & & & & \\
& -1 & & & & \\
& & 0 & 1 & & \\
& & & 0 & & \\
& & & & -1 & \\
& & & & & 0 \\
\end{array}\right)$ &
\left(\begin{array}{cccccc}
-1 & & & & & \\
& 1 & & & & \\
& & 0 & -1 & & 1 \\
& & & 0 & & \\
& & & & 1 & \\
& & & & & 0 \\
\end{array}\right)$ & $\sigma^{11} \in F$ \\
\hline\hline
\end{tabular}
\label{L41}
\end{table}
\begin{table}[th]
\tiny
\caption{The Leibniz algebras $L(4,2)$ of non-Lie type}
\begin{tabular}{lllc}
\hline
No. &
$A^1=-B^1$ & $A^2=-B^2$ & $\sigma$ \\
\hline\hline
(11) &
$\left(\begin{array}{cccccc}
1 & & & & & \\
& 0 & & & & \\
& & -1& & & \\
& & & 1 & & \\
& & & & -1& \\
& & & & & 0 \\
\end{array}\right)$
&
$\left(\begin{array}{cccccc}
0 & & & & & \\
& 1 & & & & \\
& & -1& & & \\
& & & 1 & & \\
& & & & 0 & \\
& & & & & 0 \\
\end{array}\right)$
&
$\begin{array}{c}
\sigma^{11}, \sigma^{22}, \sigma^{12}, \sigma^{21} \in F\\
\\
\sigma^{11}, \sigma^{22}, \sigma^{12}+\sigma^{21}\text{ not all zero}
\end{array}
$ \\
\hline\hline
\end{tabular}
\label{L42}
\end{table}
\newpage
|
\section{Introduction}\label{sec:introduction}
Let $X$ be a smooth complex projective curve of genus $g\geq 1$, and let $G$ be a complex reductive group.
The $G$-character variety of $X$ is defined as the moduli space of semisimple representations of
$\pi_{1}(X)$ into $G$, that is,
$$
\cM (G)= \{(A_{1},B_{1},\ldots,A_{g},B_{g}) \in G^{2g} \,| \prod_{i=1}^{g}[A_{i},B_{i}]=\Id\}/ / G.
$$
For the complex linear groups $G=\GL(n,\CC), \SL(n,\CC)$, the representations of $\pi_1(X)$ into $G$ can be understood
as $G$-local systems $E\to X$, hence defining
a flat bundle $E$ which has $\deg E=0$. A natural generalization consists of allowing bundles $E$ of
non-zero degree $d$. The $G$-local systems on $X$ correspond to representations $\rho:\pi_1(X - \{p_0\}) \to G$,
where $p_0\in X$ is a point, and $\rho(\gamma)= \frac{d}{n} \Id$, $\gamma$ a loop around $p_0$, giving
rise to the moduli space of twisted representations
$$
\cM^d (G)= \{(A_{1},B_{1},\ldots,A_{g},B_{g}) \in G^{2g} \,| \prod_{i=1}^{g}[A_{i},B_{i}]=e^{2\pi \imat d/n} \Id \}/ / G.
$$
A related object is the moduli space of parabolic representations with one puncture, given by those
representations whose monodromy around $p_0$ is $\rho(\gamma)=\xi=\text{diag}(\lambda_1,\ldots, \lambda_n)$,
where $\lambda_1\cdots \lambda_n =1$.
The parabolic character variety is
$$
\cM_{\xi}(G) = \{ (A_{1},B_{1},\ldots, A_{g},B_{g})\in G^{2g} \ |
\prod_{i=1}^{g}[A_{i},B_{i}] = \xi \} / /\Stab(\xi) .
$$
These spaces are naturally generalized by the character varieties with arbitrary monodromy $C\in G$,
$$
\cM_{C}(G) = \{ (A_{1},B_{1},\ldots, A_{g},B_{g})\in G^{2g} \ |
\prod_{i=1}^{g}[A_{i},B_{i}] = C \} / /\Stab(C) .
$$
Here $C$ can be central, diagonalizable, or of Jordan type.
\medskip
The space $\cM^d(G)$ is homeomorphic to the moduli space of $G$-Higgs bundles $\mathcal{H}^d(G)$,
which parametrizes pairs $(E,\Phi)$, consisting of a vector bundle
$E\to X$ of degree $d$ and rank $n$ (with fixed determinant in the case $G=\SL(n,\CC)$),
and a homomorphism $\Phi:E\to E\otimes K_X$, known as the Higgs field (in the case $G=\SL(n,\CC)$,
the Higgs field has trace $0$). This homeomorphism has been used to compute the
cohomology of the moduli space $\cM^d(G)$ when $\gcd(n,d)=1$.
Poincar\'e polynomials for $G=\SL(2,\CC)$ were computed in \cite{hitchin:1987}, for $G=\SL(3,\CC)$
in \cite{gothen:1994} and for $G=\GL(4,\CC)$ in \cite{garciaprada-heinloth-schmitt:2011}. A recursive formula for the motive of the moduli space of Higgs bundles of arbitrary rank and degree coprime to the rank has been given in \cite{garciaprada-heinloth:2013}. In particular, this gives the Betti numbers of the character variety for arbitrary coprime rank and degree.
There is a similar correspondence between the parabolic character variety and the moduli
space of parabolic Higgs bundles. The Poincar\'e polynomials of the moduli of parabolic
Higgs bundles are given in \cite{by} for $G=\SL(2,\CC)$ and in \cite{garciaprada-gothen-munoz:2004} for
$G=\SL(3,\CC)$ and $\GL(3,\CC)$.
However, the homeomorphism $\cM^d(G) \cong \mathcal{H}^d(G)$ is far from being an algebraic isomorphism,
and hence the algebro-geometric information carried by these moduli spaces is different. Natural
refinements of the Poincar\'e polynomials which take into account the algebraic structure
are the E-polynomials (or Hodge-Deligne polynomials), which
encode the information of the dimensions of the Hodge decomposition of the cohomology in a nice way.
The fact that the E-polynomials of $\cM^d(G)$ and $\mathcal{H}^d(G)$ have different natures has deep
implications on the Mirror Symmetry phenomena exhibited in the non-abelian Hodge theory of a
curve \cite{hausel-thaddeus:2003}.
Hausel and Rodriguez-Villegas started the computation of the E-polynomials of $G$-character
varieties focusing on $G=\GL(n,\CC), \SL(n,\CC)$ and $\PGL(n,\CC)$, using arithmetic methods inspired by the Weil
conjectures. In \cite{hausel-rvillegas:2007} they obtained the E-polynomials of $\cM^{d}(G)$
for $G=\GL(n,\CC)$ (i.e., for $C=e^{2\pi\imat d/n} \Id$), in terms of a simple generating function.
Following these methods, Mereb \cite{mereb:2010} studied this case for $\SL(n,\CC)$, giving
an explicit formula for the E-polynomial in the case $G=\SL(2,\CC)$, while for $\SL(n,\CC)$
these polynomials are given in terms of a generating function.
Also E-polynomials of the parabolic character varieties for $G=\GL(n,\CC)$
have been obtained by Hausel, Letellier and Rodriguez-Villegas, for semisimple conjugacy classes at the punctures
\cite{hausel-letellier-rvillegas:2011}.
Logares, Newstead and the second author introduced a geometric technique in \cite{lomune}
to compute the E-polynomial of character varieties
by using stratifications and also handling fibrations
which are locally trivial in the analytic topology but not in the Zariski topology. The
main results of \cite{lomune} are explicit formulas for the E-polynomials
of character varieties for $G=\SL(2,\CC)$ and $g=1,2$. Actually, the geometric technique
allows for dealing with character varieties in which the holonomy around the puncture
is not diagonalizable (in this case there is no correspondence with
a Higgs bundle moduli space as mentioned above). In \cite{mamu}, the authors extend the
theory to compute the E-polynomial of the character variety for $G=\SL(2,\CC)$ and $g=3$,
and holonomy $\pm \Id$. Here we complete the general case $g\geq 3$ and any holonomy.
We say that a variety $Z$ is of \textit{balanced type} (also called of Hodge-Tate type in the literature) if its mixed Hodge structure $H^{k,p,q}_c(Z)$ is non-zero only for $p=q$. Our first result is the following
\begin{thm}\label{thm:balanced}
All character varieties $\mathcal{M}_{C}(\SL(2,\mathbb{C}))$ are of balanced type.
\end{thm}
The main result of the paper is the computation of the E-polynomials of the character varietes for $\SL(2,\mathbb{C})$ and any genus.
\begin{thm}\label{thm:main}
Let $X$ be a complex curve of genus $g\geq 1$. Let $\cM_C=\cM_C(\SL(2,\CC))$ be the
character variety corresponding to $C\in \SL(2,\CC)$. The E-polynomials are as follows:
\begin{align*}
e(\mathcal{M}_{\Id})= & \, (q^{3}-q)^{2g-2}+(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2}-2^{2g}q^{2g-2} \\
& +\frac{1}{2}q^{2g-2}(q+2^{2g}-1)((q+1)^{2g-2}+(q-1)^{2g-2}) + \frac{1}{2}q((q+1)^{2g-1}+(q-1)^{2g-1}) \\
\end{align*}
\begin{align*}
e(\mathcal{M}_{-\Id}) = & \, (q^3-q)^{2g-2}+(q^2-1)^{2g-2}-2^{2g-1}(q^{2}+q)^{2g-2}+(2^{2g-1}-1)(q^{2}-q)^{2g-2} \\
e(\mathcal{M}_{J_{+}}) = & \, (q^{3}-q)^{2g-2}(q^{2}-1)+(2^{2g-1}-1)(q-1)(q^{2}-q)^{2g-2}-2^{2g-1}(q+1)(q^{2}+q)^{2g-2}\\
& +\frac{1}{2}q^{2g-2}(q-1)\left((q-1)^{2g-1}-(q+1)^{2g-1} \right) \\
e(\mathcal{M}_{J_{-}}) =& \, (q^{3}-q)^{2g-2}(q^{2}-1)+(2^{2g-1}-1)(q-1)(q^{2}-q)^{2g-2}+2^{2g-1}(q+1)(q^{2}+q)^{2g-2} \\
e(\mathcal{M}_{\xi_{\lambda}}) = & \, (q^{3}-q)^{2g-2}(q^{2}+q)+(q^{2}-1)^{2g-2}(q+1)+(2^{2g}-2)(q^{2}-q)^{2g-2}q,
\end{align*}
for $J_+=\begin{pmatrix} 1& 1\\ 0 &1\end{pmatrix}$,
$J_-=\begin{pmatrix} - 1& 1\\ 0 &- 1\end{pmatrix}$ and
$\xi_\lambda=\begin{pmatrix} \lambda & 0\\ 0 &\lambda^{-1}\end{pmatrix}$, $\lambda\neq 0,\pm 1$,
and with $q=\, uv$.
\end{thm}
This generalizes the formulas of \cite{lomune} for $g=1,2$ (the formula for $e(\cM_{\Id})$ for $g=2$ in Theorem 1.2 of
\cite{lomune} has a misprint; the polynomial should be $e(\cM_{\Id})=q^6+ 17q^4+q^2+ 1$,
which is correctly written in Section 8 of \cite{lomune}). It also generalizes the formulas in \cite{mamu} for
$e(\cM_{\Id})$ and $e(\cM_{-\Id})$ for $g=3$. The formula for $e(\cM_{-\Id})$ and any $g$ coincides with
that of \cite{mereb:2010}. The other E-polynomials are new, including the case of parabolic $\SL(2,\mathbb{C})$-character varieties $\mathcal{M}_{\xi_\lambda}$.
We use the information for the cases $g=1,2$ from \cite{lomune}
as building blocks. The E-polynomial for the character variety for a curve $X_g$ of
genus $g\geq 3$ will be computed inductively. The basic idea, as will be clear throughout the paper,
is to decompose $X_g=X_{g-1}\# X_1$ as a connected sum. From the information for
$X_g$ one gets information for $X_{g-1}$ with a hole, and this is used in turn to
compute the E-polynomial corresponding to $X_{g+1}=X_{g-1}\# X_2$. The
E-polynomials of $X_1,X_2$ with a puncture from \cite{lomune} come into play here.
This induction has as starting point the curve $X_3$ of genus $g=3$ (with no
puncture), which is computed in \cite{mamu}. This special case has its special
features, and has to be treated separately. In particular, in that case the
techniques to compute E-polynomials of analytically locally trivial fibrations use
a base of dimension $2$. In all other cases ($g=1,2$ in \cite{lomune}, and
the induction for $g\geq 4$ treated here) a base of dimension $1$ suffices.
Theorem \ref{thm:main} allows us to prove the following relation
of the E-polynomials of various character varieties, conjectured by T. Hausel.
\begin{cor} \label{cor:Hausel}
For any genus $g\geq 1$, we have
$$
e(\mathcal{M}_{J_{-}})+(q+1)e(\mathcal{M}_{-\Id})=e(\mathcal{M}_{\xi_{\lambda}}).
$$
\end{cor}
As a byproduct of our analysis, we also obtain the behaviour of the E-polynomial of the parabolic character
variety ($G=\SL(2,\CC)$)
$$
\cM_{\xi_\lambda}=\cM_{\xi_\lambda} (G)= \{(A_{1},B_{1},\ldots,A_{g},B_{g}) \in G^{2g} \,| \prod_{i=1}^{g}[A_{i},B_{i}]=
\begin{pmatrix} \lambda & 0 \\ 0 &\lambda^{-1} \end{pmatrix} \}/ / G,
$$
when $\lambda$ varies in $\CC-\{0,\pm 1\}$. This is of relevance for Mirror Symmetry phenomena. It is given by the following formula.
\begin{thm}
Let $X$ be a curve of genus $g\geq 1$. Then
\begin{equation*}\label{eqn:RcM}
R(\mathcal{M}_{\xi_{\lambda}})= \left( (q^{3}-q)^{2g-2}(q^{2}+q)+(q+1)(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2} \right) T
+ \left( (2^{2g}-1)q(q^{2}-q)^{2g-2} \right) N,
\end{equation*}
which means that the E-polynomial of the invariant part of the cohomology is the polynomial accompanying $T$,
and the E-polynomial of the non-invariant part is the polynomial accompanying $N$.
\end{thm}
We end up giving some consequences of Theorem \ref{thm:main} in Section \ref{sec:topol}. Notably,
\begin{cor}
The E-polynomials of $e(\mathcal{M}_{-\Id})$, $e(\mathcal{M}_{\xi_{\lambda}})$ are palindromic.
\end{cor}
\medskip
The present arguments can be used to compute the E-polynomials of the $\PGL(2,\CC)$-character varieties of
surface groups for arbitrary genus, which will appear in \cite{ma}.
\noindent\textbf{Acknowledgements.}
We would like to thank Peter Newstead, Tam\'as Hausel, Martin Mereb, Nigel Hitchin, Marina Logares and Oscar Garc\'{\i}a-Prada for useful conversations.
This work has been partially supported by MICINN (Spain) Project MTM2010-17389. The first author was also supported by a FPU scholarship from the Spanish Ministerio de Educaci\'on.
\section{E-polynomials}\label{sec:method}
We start by giving the definition of E-polynomials.
A pure Hodge structure of weight $k$ consists of a finite dimensional complex vector space
$H$ with a real structure, and a decomposition $H=\bigoplus_{k=p+q} H^{p,q}$
such that $H^{q,p}=\overline{H^{p,q}}$, the bar meaning complex conjugation on $H$.
A Hodge structure of weight $k$ gives rise to the so-called Hodge filtration, which is a descending filtration
$F^{p}=\bigoplus_{s\ge p}H^{s,k-s}$. We define $\Gr^{p}_{F}(H):=F^{p}/ F^{p+1}=H^{p,k-p}$.
A mixed Hodge structure consists of a finite dimensional complex vector space $H$ with a real structure,
an ascending (weight) filtration $\ldots \subset W_{k-1}\subset W_k \subset \ldots \subset H$
(defined over $\RR$) and a descending (Hodge) filtration $F$ such that $F$ induces a pure Hodge structure
of weight $k$ on each $\Gr^{W}_{k}(H)=W_{k}/W_{k-1}$. We define
$$
H^{p,q}:= \Gr^{p}_{F}\Gr^{W}_{p+q}(H)
$$
and write $h^{p,q}$ for the {\em Hodge number} $h^{p,q} :=\dim H^{p,q}$.
Let $Z$ be any quasi-projective algebraic variety (maybe non-smooth or non-compact).
The cohomology groups $H^k(Z)$ and the cohomology groups with compact support
$H^k_c(Z)$ are endowed with mixed Hodge structures \cite{Deligne2,Deligne3}.
We define the {\em Hodge numbers} of $Z$ by
$h^{k,p,q}_{c}(Z) = h^{p,q}(H_{c}^k(Z))$.
The Hodge-Deligne polynomial, or E-polynomial, is defined as
$$
e(Z)=e(Z)(u,v):=\sum _{p,q,k} (-1)^{k}h^{k,p,q}_{c}(Z) u^{p}v^{q}.
$$
When $h_c^{k,p,q}=0$ for $p\neq q$, the polynomial $e(Z)$ depends only on the product $uv$.
This will happen in all the cases that we shall investigate here. In this situation,
we use the variable $q=uv$. If this happens, we say that the variety is {\it of balanced type}.
For instance, $e(\CC^n)=q^n$.
The key property of Hodge-Deligne polynomials that permits their calculation is that they are additive for
stratifications of $Z$. If $Z$ is a complex algebraic variety and
$Z=\bigsqcup_{i=1}^{n}Z_{i}$, where all $Z_i$ are locally closed in $Z$, then $e(Z)=\sum_{i=1}^{n}e(Z_{i})$.
There is another useful property that we shall use often: if there is an action of $\ZZ_2$ on $X$, then we have
polynomials $e(X)^+, e(X)^-$, which are the E-polynomials of the invariant and anti-invariant parts of the cohomology
of $X$, respectively. More concretely, $e(X)^+=e(X/\ZZ_2)$ and $e(X)^-=e(X)-e(X)^+$. Then if $\ZZ_2$ acts
on $X$ and on $Y$, we have the equality (see \cite[Proposition 2.6]{lomune})
\begin{equation}\label{eqn:+-}
e((X\times Y)/\ZZ_2) =e(X)^+e(Y)^+ +e(X)^-e(Y)^-\, .
\end{equation}
Suppose that
\begin{equation}\label{fibration}
F \longrightarrow Z \overset{\pi}{\longrightarrow} B
\end{equation}
is a fibration locally trivial in the analytic topology, and with $F$ of balanced type.
The fibration defines a local system $\mathcal{H}^{k}_{c}$,
whose fibers are the cohomology groups $H^{k}_{c}(F_{b})$, where $b\in B$, $F_{b}=\pi^{-1}(b)$.
Associated to the fibration, there is a monodromy representation
\begin{equation}\label{eqn:exxtra}
\rho : \pi_{1}(B) \longrightarrow \GL(H^{k,p,p}_c(F)).
\end{equation}
Suppose that the monodromy group $\Gamma=\mathrm{im}\,(\rho)$ is an abelian and finite group.
Then $H_c^{k,p,p}(F)$ are modules over the representation ring $R(\Gamma)$. So there is a well
defined element, the {\em Hodge monodromy representation},
\begin{equation}\label{eqn:Hodge-mon-rep}
R(Z) := \sum (-1)^k H_c^{k,p,p}(F)\, q^p \in R(\Gamma)[q] \, .
\end{equation}
As the monodromy representation (\ref{eqn:exxtra}) has finite image, there is a finite covering
$B_\rho \to B$ such that the pull-back fibration has trivial monodromy. We have the
following result.
\begin{thm}[{\cite[Theorem 2]{mamu}}] \label{thm:general-fibr}
Suppose that $B_\rho$ is of balanced type. Then $Z$ is of balanced type.
There is a $\ZZ[q]$-linear map
$$
e: R(\Gamma)[q] \to \ZZ[q]
$$
satisfying the property that $e(R(Z))=e(Z)$
\end{thm}
This can be rephrased as follows.
Let $S_1,\ldots, S_N$ be the
irreducible representations of $\Gamma$ (there are $N=\# \Gamma$ of them, and all
of them are one-dimensional). These are generators of $R(\Gamma)$ as a free abelian group.
Write $s_i(q)=e(S_i)$, $1\leq i\leq N$. Then if we
write the Hodge monodromy representation of (\ref{eqn:Hodge-mon-rep}) as
$$
R(Z)= a_1(q) S_1 + \ldots a_N(q) S_N,
$$
then we have
$$
e(Z)= a_1(q) s_1(q) + \ldots +a_N(q) s_N(q) .
$$
We shall only use
Theorem \ref{thm:general-fibr} in two situations. First, when
$\pi:Z\to B$ is a fibre bundle with fibre $F$ such that the action of $\pi_1(B)$ on $H_c^*(F)$ is trivial.
Then $R(Z)=e(F) T$, where $T$ is the trivial local system and
$e(Z)=e(F) e(B)$ (this result appears in \cite[Proposition 2.4]{lomune}).
In particular, this happens when
$Z$ is a $G$-space with isotropy $H<G$ such that $G/H\to Z\to B$ is a fiber bundle,
and $G$, $H$ are connected algebraic groups. Then $e(Z)=e(B)e(G)/e(H)$
Second, when $B$ is one-dimensional. So we have a fibration
$$
F\longrightarrow Z \longrightarrow B=\CC-\{q_1,\ldots, q_\ell\},
$$
with monodromy $\rho$ and Hodge monodromy representation $e(R(Z))\in R(\Gamma)[q]$.
Assuming that $B_\rho$ is a rational curve, we write
$R(Z)=a_1(q) T+ a_2(q) S_2+ \ldots + a_N(q)S_N$, where $T$ is the trivial representation
and $S_2, \ldots, S_N$ are the non-trivial representations. Then $e(T)=q-\ell$ and $e(S_i)=-(\ell-1)$,
$2 \leq i\leq N$. Hence
\begin{equation}\label{eqn:dos}
e(Z)=(q- \ell) a_1(q) -(\ell -1) \sum_{i=2}^N a_i(q)=(q-1)\, e(F)^{inv} - (\ell-1) e(F),
\end{equation}
where $e(F)^{inv}=a_1(q)$ is the E-polynomial of
the invariant part of the cohomology of $F$ and $e(F)=\sum_{i=1}^N a_i(q)$. See \cite[Corollary 3]{mamu}.
\section{Stratifying the space of representations} \label{sec:strata}
Let $g\geq 1$ be any natural number. We define the following sets:
\begin{itemize}
\item $\overline{X}{}_{0}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \in \SL(2,\CC)^{2g} \mid \prod_{i=1}^{g}[A_{i},B_{i}]= \Id\}$.
\item $\overline{X}{}_{1}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g})\in \SL(2,\CC)^{2g} \mid \prod_{i=1}^{g}[A_{i},B_{i}]= -\Id\}$.
\item $\overline{X}{}_{2}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \in \SL(2,\CC)^{2g}\mid \prod_{i=1}^{g}[A_{i},B_{i}]= J_+=
\begin{pmatrix} 1& 1\\ 0 & 1 \end{pmatrix}\}$.
\item $\overline{X}{}_{3}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \in \SL(2,\CC)^{2g}\mid \prod_{i=1}^{g}[A_{i},B_{i}]= J_-=
\begin{pmatrix} -1& 1\\ 0 & -1 \end{pmatrix}\}$.
\item $\overline{X}{}_{4,\lambda}^{g}=
\lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \in \SL(2,\CC)^{2g}\mid \prod_{i=1}^{g}[A_{i},B_{i}]= \xi_\lambda=
\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix} \rbrace$, where $\lambda\in \CC-\{ 0, \pm 1\}$.
\item $\overline{X}{}_{4}^{g}=
\lbrace (A_{1},B_{1},\ldots, A_{g},B_{g},\lambda) \in \SL(2,\CC)^{2g}\times (\CC-\{ 0, \pm 1\}) \mid
\prod_{i=1}^{g}[A_{i},B_{i}]= \begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix} \rbrace$.
\end{itemize}
There is a natural fibration
$$
\overline{X}{}_4^g \longrightarrow \CC-\{ 0, \pm 1\}
$$
whose fiber are $\overline{X}{}_{4,\lambda}^{g}$.
There is an action of $\ZZ_2$ on $\overline{X}{}_4$ given by
$ (A_{1},\ldots,B_{g},\lambda) \mapsto (P^{-1}_0A_1P_0,\ldots, P^{-1}_0B_gP_0,\lambda^{-1})$, with $P_0=
\left(\begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right)$. There is an induced fibration
$$
\overline{X}{}_4^g/\ZZ_2 \longrightarrow ( \CC-\{ 0, \pm 1\})/\ZZ_2 \cong \CC-\{\pm 2\},
$$
where the basis is parametrized by $t=\lambda+\lambda^{-1}$.
For $g=1$, we shall denote $\overline{X}{}_0=\overline{X}{}_0^1, \overline{X}{}_1=\overline{X}{}_1^1$, etc., following the notations of \cite{lomune} and \cite{mamu}.
For $g=2$, we shall denote $\overline{Y}_0=\overline{X}{}_0^2, \overline{Y}_1=\overline{X}{}_1^2$, etc., following the notations of \cite[Section 4]{mamu}.
We collect the information for $g=1,2$ from \cite{lomune} and \cite{mamu}.
\begin{itemize}
\item $e(\overline{X}{}_{0}) = q^{4}+4q^{3}-q^2-4q$
\item $e(\overline{X}{}_1) =q^{3}-q $
\item $e(\overline{X}{}_2) = q^3-2q^2-3 q $
\item $e(\overline{X}{}_3) =q^3 +3 q^2 $
\item $R(\overline{X}{}_{4}/\mathbb{Z}_{2})=q^{3}T-3q S_{2}+3q^{2}S_{-2}-S_{0}$
\item $e(\overline{Y}_0)= q^9+q^8+12q^7+2q^6-3q^4-12q^3-q$.
\item $e(\overline{Y}_1)= q^9-3q^7-30q^6+30q^4+3q^3-q$.
\item $e(\overline{Y}_2)= q^9-3q^7-4q^6-39q^5-4q^4-15q^3$.
\item $e(\overline{Y}_3)= q^9-3q^7+15q^6+6q^5+45q^4$.
\item $R(\overline{Y}_{4}/\ZZ_2)=(q^9-3q^7+6q^5)T-(45q^5+15q^3)S_{2}+(15q^6+45q^4)S_{-2}+(-6q^4+3q^2-1)S_{0}$.
\end{itemize}
Here the monodromy group is $\Gamma=\ZZ_2\times\ZZ_2$, generated by the loops $\gamma_{\pm 2}$ around the punctures $\pm 2$ of
$\CC-\{\pm 2\}$. The ring $R(\Gamma)$ is generated by the trivial representation $T$, the representations $S_{\pm 2}$ which
are non-trivial around $\pm2$ and trivial around $\mp 2$, and the representation $S_0=S_2\otimes S_{-2}$.
\medskip
Now we shall set up an induction. Assume that for all $k< g$, the Hodge monodromy representation of $\overline{X}{}_4^k/\ZZ_2$
is in $R(\Gamma)[q]$. We write
$$
R(\overline{X}{}_4^k/\ZZ_2)=a_k T+ b_k S_2 + c_k S_{-2} + d_k S_0,
$$
for some polynomials $a_k,b_k,c_k,d_k \in \ZZ[q]$.
Take $k,h< g$. Fix some $C=\Id, -\Id, J_+,J_-$ or $\xi_\lambda$. Then
\begin{equation} \label{eqn:uno}
\prod_{i=1}^{k+h}[A_{i},B_{i}]= C \iff \prod_{i=1}^{k}[A_{i},B_{i}]= C \prod_{i=1}^{h}[B_{k+i},A_{k+i}].
\end{equation}
\section{Computation of $e(\overline{X}{}_0^{k+h})$}\label{subsec:0}
Using (\ref{eqn:uno}), we stratify $\overline{X}{}_{0}^{k+h}= \bigsqcup W_{i}$, where
\begin{itemize}
\item $W_{0}=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}] =\Id \rbrace \cong \overline{X}{}_{0}^{k}\times \overline{X}{}_{0}^h$.
\item $W_{1}=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}] =-\Id \rbrace \cong \overline{X}{}_{1}^{k}\times \overline{X}{}_{1}^h$.
\item $W_2=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim J_{+} \rbrace \cong
\PGL(2,\CC)/U \times \overline{X}{}_{2}^{k}\times \overline{X}{}_{2}^h$, where $U=\{ \begin{pmatrix} 1 & x \\ 0 & 1 \end{pmatrix}\} \cong \CC$.
\item $W_3=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim J_{-} \rbrace\cong
\PGL(2,\CC)/U \times \overline{X}{}_{3}^{k}\times \overline{X}{}_{3}^h$.
\item $W_{4}= \lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim
\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}, \lambda \neq 0, \pm 1 \rbrace$.
\end{itemize}
To compute $e(W_4)$, we define
$$
\overline{W}{}_4= \lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h},\lambda) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= \prod_{i=1}^{h}[B_{k+i},A_{k+i}]=
\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}, \lambda \neq 0, \pm 1 \rbrace,
$$
which produces the fibration
$$
\overline{W}{}_{4}/\ZZ_2 \rightarrow \CC -\{ \pm 2\},
$$
whose Hodge monodromy representation is
\begin{align} \label{eqn:tres}
R(\overline{W}{}_{4}/\ZZ_2) =& \, R(\overline{X}{}_{4}^{k}/\ZZ_2) \otimes R(\overline{X}{}_{4}^h/\ZZ_2) \nonumber \\
= &\, (a_{k}a_h+b_{k}b_h+c_{k}c_h+d_{k}d_h)T + (a_{k}b_h+b_{k}a_h+c_{k}d_h+d_{k}c_h)S_{2} \\
& + (a_{k}c_h+b_{k}d_h+c_{k}a_h+d_{k}b_h)S_{-2} + (a_{k}d_h+b_{k}c_h+c_{k}b_h+d_{k}a_h)S_{0}. \nonumber
\end{align}
From $R(\overline{W}{}_{4}/\ZZ_2)$ there is a standard procedure to obtain $e(W_4)$.
For brevity, write $R(\overline{W}{}_{4}/\ZZ_2)=A T+ BS_2+ CS_{-2}+DS_0$, where
\begin{align} \label{eqn:tres.5}
A&= a_{k}a_h+b_{k}b_h+c_{k}c_h+d_{k}d_h\nonumber \\
B&= a_{k}b_h+b_{k}a_h+c_{k}d_h+d_{k}c_h \\
C &= a_{k}c_h+b_{k}d_h+c_{k}a_h+d_{k}b_h \nonumber\\
D&= a_{k}d_h+b_{k}c_h+c_{k}b_h+d_{k}a_h \nonumber
\end{align}
First, using (\ref{eqn:dos}), we have that $e(\overline{W}{}_{4}/\ZZ_2)=(q-2)A-(B+C+D)$.
On the other hand, the $2:1$-cover $\CC-\{0,\pm1\} \to \CC -\{ \pm 2\}$ allows us to
deduce that $R(\overline{W}{}_{4})=(A+D) T+ (B+C)N$, where $T$ is the trivial representation,
and $N$ is the representation which is non-trivial and of order two around the origin. So
using (\ref{eqn:dos}) again, we have that $e(\overline{W}{}_{4})=(q-3)(A+D)-2(B+C)$.
Now note that
$$
W_4 \cong (\PGL(2,\CC) /D \times \overline{W}{}_{4}) /\ZZ_2,
$$
where $D\cong \CC^*$ are the diagonal matrices.
By \cite[Proposition 3.2]{lomune}, we have that $ e(\PGL(2, \CC)/D)^+=q^2$ and $ e(\PGL(2, \CC)/D)^-=q$. Using (\ref{eqn:+-}),
\begin{align}\label{eqn:RX4-RX4Z2->eX4}
e(W_4) &= q^2 e(\overline{W}{}_ 4)^+ + q \, e(\overline{W}{}_4)^- \nonumber \\
&= q^2 e(\overline{W}{}_4/\ZZ_2) + q( e(\overline{W}{}_4)-e(\overline{W}{}_4/\ZZ_2)) \nonumber \\
&= (q^2-q)e(\overline{W}{}_4/\ZZ_2) +q \, e(\overline{W}{}_{4}) \\
&= (q^2-q)((q-2)A-(B+C+D)) + q ((q-3)(A+D)-2(B+C)) \nonumber \\
&= (q^3-2q^2-q)A-(q^2+q)(B+C)-2qD. \nonumber
\end{align}
All together, recalling also that $e(\PGL(2,\CC))=q^3-q$ and so $e(\PGL(2,\CC)/U)=q^2-1$, we have
\begin{align} \label{eqn:eX0}
e(\overline{X}{}_{0}^{k+h}) &= e(\overline{X}{}_0^k)e(\overline{X}{}_0^h)+ e(\overline{X}{}_1^k)e(\overline{X}{}_1^h)+
(q^2-1) e(\overline{X}{}_2^k)e(\overline{X}{}_2^h)+ (q^2-1) e(\overline{X}{}_3^k)e(\overline{X}{}_3^h)+ e(W_4).
\end{align}
Setting $k=g-1$, $h=1$, and substituting the values $A,B,C,D$ from (\ref{eqn:tres.5}) into (\ref{eqn:RX4-RX4Z2->eX4}),
and then the values of $e(\overline{X}{}_j^1)$ and $a_1,b_1,c_1,d_1$ from Section \ref{sec:strata}, we have
\begin{align}\label{eqn:eX0.} \tag{$\alpha$}
e_0^g=e(\overline{X}{}_{0}^{g}) = & \, (q^4 + 4 q^3 - q^2 - 4 q) e_0^{g-1}+ (q^3-q) e_1^{g-1} \nonumber \\
&+ ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_2^{g-1}+ (q^5+3q^4-q^3-3q^2) e_3^{g-1} \nonumber \\
&+ (q^6-2q^5-4q^4+3q^2+2q) a_{g-1} + (-q^5-4q^4+4q^2+q)b_{g-1} \nonumber\\
&+ (2q^5-7q^4-3q^3+7q^2-q) c_{g-1} + (-5q^4-q^3+5q^2-q) d_{g-1} \nonumber
\end{align}
\section{Computation of $e(\overline{X}{}_1^{k+h})$}\label{subsec:1}
We do something similar to the previous case. We stratify $\overline{X}{}_{1}^{k+h}= \bigsqcup W'_{i}$, where
\begin{itemize}
\item $W'_{0}=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= - \prod_{i=1}^{h}[B_{k+i},A_{k+i}] =\Id \rbrace \cong \overline{X}{}_{0}^{k}\times \overline{X}{}_{1}^h$.
\item $W'_{1}=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= -\prod_{i=1}^{h}[B_{k+i},A_{k+i}] =-\Id \rbrace \cong \overline{X}{}_{1}^{k}\times \overline{X}{}_{0}^h$.
\item $W'_2=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= -\prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim J_{+} \rbrace
\cong\PGL(2,\CC)/U \times \overline{X}{}_{2}^{k}\times \overline{X}{}_{3}^h$, where $U=\{ \begin{pmatrix} 1 & x \\ 0 & 1 \end{pmatrix}\} \cong \CC$.
\item $W'_3=\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= - \prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim J_{-} \rbrace
\cong\PGL(2,\CC)/U \times \overline{X}{}_{3}^{k}\times \overline{X}{}_{2}^h$.
\item $W'_{4}= \lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= - \prod_{i=1}^{h}[B_{k+i},A_{k+i}] \sim
\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}, \lambda \neq 0, \pm 1 \rbrace$.
\end{itemize}
To compute $e(W'_4)$, we define $\overline{W}{}'_4=
\lbrace (A_{1},B_1,\ldots,A_{k+h},B_{k+h},\lambda) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= - \prod_{i=1}^{h}[B_{k+i},A_{k+i}]=
\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}, \lambda \neq 0, \pm 1 \rbrace$, which
produces the fibration
$$
\overline{W}{}'_{4}/\ZZ_2 \rightarrow \CC -\{ \pm 2\},
$$
whose Hodge monodromy representation is given as (where $\tau(\lambda)=-\lambda$),
\begin{align} \label{eqn:tres'}
R(\overline{W}{}'_{4}/\ZZ_2) =& \, R(\overline{X}{}_{4}^{k}/\ZZ_2) \otimes \tau^* R(\overline{X}{}_{4}^h/\ZZ_2) \nonumber \\
= & \, (a_{k}T+b_{k}S_{2}+c_{k}S_{-2}+d_{k}S_{0}) \otimes (a_hT+c_hS_{2}+b_hS_{-2}+d_hS_{0}) \nonumber\\
= & \, (a_{k}a_h+b_{k}c_h+c_{k}b_h+d_{k}d_h)T + (a_{k}c_h+b_{k}a_h+c_{k}d_h+d_{k}b_h)S_{2}\\
& + (a_{k}b_h+b_{k}d_h+c_{k}a_h+d_{k}c_h)S_{-2} + (a_{k}d_h+b_{k}b_h+c_{k}c_h+d_{k}a_h)S_{0} \nonumber
\end{align}
We write $R(\overline{W}{}'_{4}/\ZZ_2)=A' T+ B'S_2+ C'S_{-2}+D'S_0$, with
\begin{align} \label{eqn:tres'.5}
A'&= a_{k}a_h+b_{k}c_h+c_{k}b_h+d_{k}d_h\nonumber \\
B'&= a_{k}c_h+b_{k}a_h+c_{k}d_h+d_{k}b_h \\
C' &=a_{k}b_h+b_{k}d_h+c_{k}a_h+d_{k}c_h \nonumber\\
D'&=a_{k}d_h+b_{k}b_h+c_{k}c_h+d_{k}a_h \nonumber
\end{align}
We use (\ref{eqn:RX4-RX4Z2->eX4}) to get
\begin{align*}
e(W'_4) &= (q^3-2q^2-q)A'-(q^2+q)(B'+C')-2qD'.
\end{align*}
Finally
\begin{align} \label{eqn:eX1}
e(\overline{X}{}_{1}^{k+h}) &= e(\overline{X}{}_0^k)e(\overline{X}{}_1^h)+ e(\overline{X}{}_1^k)e(\overline{X}{}_0^h)+
(q^2-1) e(\overline{X}{}_2^k)e(\overline{X}{}_3^h)+ (q^2-1) e(\overline{X}{}_3^k)e(\overline{X}{}_2^h)+ e(W'_4).
\end{align}
Setting $k=g-1$, $h=1$, and substituting the values $A',B',C',D'$ from (\ref{eqn:tres'.5})
and the values of $e(\overline{X}{}_j^1)$ and $a_1,b_1,c_1,d_1$ from Section \ref{sec:strata}, we have
\begin{align}\label{eqn:eX1.} \tag{$\beta$}
e_1^g=e(\overline{X}{}_{1}^{g}) = & \, (q^3-q)e_0^{g-1}+ (q^4 + 4 q^3 - q^2 - 4 q) e_1^{g-1} \nonumber \\
&+ (q^5+3q^4-q^3-3q^2)e_2^{g-1}+ ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_3^{g-1} \nonumber \\
&+(q^6-2q^5-4q^4+3q^2+2q) a_{g-1} + (2q^5-7q^4-3q^3+7q^2+q) b_{g-1} \nonumber\\
&+ (-q^5-4q^4+4q^2+q) c_{g-1} + (-5q^4-q^3+5q^2+q) d_{g-1} \nonumber
\end{align}
\section{Computation of $e(\overline{X}{}_2^{k+h})$}\label{subsec:2}
Now we consider
$$
Z=\lbrace (A_{1},B_{1},\ldots, A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= J_{+}\prod_{i=1}^{h}[B_{k+i},A_{k+i}] \rbrace.
$$
We write
$$
\nu =\prod_{i=1}^{h}[B_{k+i},A_{k+i}] =\begin{pmatrix} a & b \\ c & d \end{pmatrix}, \qquad
\delta = \prod_{i=1}^{k}[A_{i},B_{i}] = J_{+}\nu = \begin{pmatrix} a+c & b+d \\ c & d \end{pmatrix}.
$$
Let $t_{1}=\tr \nu$, $t_{2}=\tr \delta$. Note that $c=t_{2}-t_{1}$.
We use the stratification defined as in \cite[Section 11]{lomune}, according to the values of the pair $(t_{1},t_{2})$.
\begin{itemize}
\item $Z_{1}$ given by $(t_{1},t_{2})=(2,2)$. In this case $c=0$, $a=d=1, b\in \CC$. If $b\neq 0,-1$, both are of Jordan type,
whereas if $b=0,1$ one of them is of Jordan type and the other is equal to $\Id$. We get
$$
e(Z_{1})=(q-2)e(\overline{X}{}^k_{2})e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{2}^k)e(\overline{X}{}_{0}^{h})+e(\overline{X}{}_{0}^k)e(\overline{X}{}_{2}^{h}).
$$
\item $Z_{2}$ given by $(t_{1},t_{2})=(-2,-2)$. Analogously
$$
e(Z_{2}) = (q-2)e(\overline{X}{}_{3}^k)e(\overline{X}{}_{3}^{h})+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{1}^{h})+e(\overline{X}{}_{1}^k)e(\overline{X}{}_{3}^{h}).
$$
\item $Z_{3}$ given by $(t_{1},t_{2})=(2,-2),(-2,2)$. Now $c\neq 0$. The action of $U\cong \CC$ allows to fix
$d=0$. Both $\nu,\delta$ are of Jordan type. So we obtain
$$
e(Z_{3})=q(e(\overline{X}{}^k_{2})e(\overline{X}{}_{3}^{h})+e(\overline{X}{}^k_{3})e(\overline{X}{}_{2}^{h})).
$$
\item $Z_{4}$ given by the subcases:
\begin{itemize}
\item $Z_{4,1}$ given by $(2,t_{2}), t_{2}\neq \pm 2$ and $(-2,t_{2}), t_{2}\neq \pm 2$. We focus on the first case.
It must be $c\neq 0$. The group $U\cong \CC$ acts freely on the matrix $\nu$, which is of Jordan type. Note also
that $\delta$ is diagonalizable. The second case is similar with $\nu\sim J_-$. We thus get
$$
e(Z_{4,1})= q(e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{3}^{h}))e(\overline{X}{}_{4}^k/\ZZ_{2}).
$$
\item $Z_{4,2}$ given by $(t_{1},2),t_{1}\neq \pm 2$ and $(t_{1},-2), t_{1}\neq \pm 2$. It is completely similar, interchanging the roles of $\nu$ and $\delta$.
$$
e(Z_{4,2}) =q(e(\overline{X}{}_{2}^k)+e(\overline{X}{}_{3}^k))e(\overline{X}{}_{4}^{h}/\ZZ_{2}).
$$
\end{itemize}
\item $Z_{5}$ corresponding to $t_{1}=t_{2}\neq \pm 2$. Now
$$
\eta=\left(\begin{array}{cc}
a & b\\
0 & a^{-1}
\end{array}\right), \ \delta=\left(\begin{array}{cc}
a & b+a^{-1}\\
0 & a^{-1}
\end{array}\right),
$$
Therefore $e(Z_5)=q\, e(\overline{Z}_5)$, where $\overline{Z}_{5}$ is a fibration over $a\in \CC- \{0,\pm 1 \}$ whose fibers are
$\overline{X}{}_{4,a}^k\times \overline{X}{}_{4,a}^{h}$. Thus the
Hodge monodromy representation is given in (\ref{eqn:tres}),
\begin{align*}
R(\overline{Z}_{5}/\ZZ_2) & =A T+ BS_2+CS_{-2}+DS_0 ,\\
R(\overline{Z}_{5}) & =(A+D)T+(B+C)N, \\
e(Z_5) &= q\, e(\overline{Z}_{5})=q((q-3)(A+D)-2(B+C)).
\end{align*}
\item $Z_{6}$ corresponding to the open stratum $t_{1},t_{2}\neq \pm 2, t_{1}\neq t_2$.
As $c\neq 0$, we can arrange $d=0$ by using the action of $U\cong \CC$. Both
$\delta$ and $\nu$ are diagonalizable matrices. If we ignore for a while the condition
$t_{1}\neq t_{2}$, the total space is isomorphic to $\CC \times \overline{X}{}_{4}^k/\ZZ_2 \times \overline{X}{}_{4}^{h}/\ZZ_2$.
The fibration over the diagonal $(t_{1},t_{1})$ has total space isomorphic $\CC \times (\overline{Z}_{5}/\ZZ_{2})$. Thus
\begin{align*}
e(\overline{Z}_{5}/\ZZ_2) &=(q-2)A-(B+C+D), \\
e(Z_{6}) &=q(e(\overline{X}{}_{4}^k/\ZZ_2)e(\overline{X}{}_{4}^{h}/\ZZ_2)-e(\overline{Z}_{5}/\ZZ_2)).
\end{align*}
\end{itemize}
Adding all up,
\begin{align*}
e(\overline{X}{}_{2}^{k+h}) = & \,e(\overline{X}{}_{2}^k)e(\overline{X}{}_{0}^{h})+e(\overline{X}{}_{0}^k)e(\overline{X}{}_{2}^{h})-2e(\overline{X}{}^k_{2})e(\overline{X}{}_{2}^{h})
+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{1}^{h})+e(\overline{X}{}_{1}^k)e(\overline{X}{}_{3}^{h})-2e(\overline{X}{}_{3}^k)e(\overline{X}{}_{3}^{h}) \\
& +q\, (e(\overline{X}{}^k_{2})+e(\overline{X}{}^k_{3})+e(\overline{X}{}_{4}^k/\ZZ_{2}))( e(\overline{X}{}_{2}^{h})+ e(\overline{X}{}_{3}^{h})+e(\overline{X}{}_{4}^{h}/\ZZ_{2}))
+q\,( e(\overline{Z}_{5})- e(\overline{Z}_{5}/\ZZ_2)).
\end{align*}
Setting $k=g-1$, $h=1$, and substituting the values $A,B,C,D$ from (\ref{eqn:tres.5})
and the values of $e(\overline{X}{}_j^1)$ and $a_1,b_1,c_1,d_1$ from Section \ref{sec:strata}, we have
\begin{align}\label{eqn:eX2.} \tag{$\gamma$}
e_2^g=e(\overline{X}{}_2^{g}) = & \,
(q^3 - 2 q^2 - 3 q)e_0^{g-1}+ (q^3 + 3 q^2) e_1^{g-1} \nonumber \\
&+ (q^5+q^4+ 3 q^2 + 3 q)e_2^{g-1}+ (q^5-3 q^3- 6 q^2 ) e_3^{g-1} \nonumber \\
& +(q^6-2q^5-3q^4+q^3+3q^2)a_{g-1}+(-q^5+2q^4-4q^3+3q^2) b_{g-1}\nonumber \\
& +(-q^5-q^4-4q^3+6q^2) c_{g-1}+(-2q^4-q^3+3q^2)d_{g-1} \nonumber
\end{align}
\section{Computation of $e(\overline{X}{}_3^{k+h})$}\label{subsec:3}
This is similar to the previous case. Consider
$$
Z'=\lbrace (A_{1},B_{1},\ldots, A_{k+h},B_{k+h}) \mid \prod_{i=1}^{k}[A_{i},B_{i}]= J_{-}\prod_{i=1}^{h}[B_{k+i},A_{k+i}] \rbrace.
$$
We write
$$
\nu =\prod_{i=1}^{h}[B_{k+i},A_{k+i}] =\begin{pmatrix} a & b \\ c & d \end{pmatrix}, \qquad
\delta = \prod_{i=1}^{k}[A_{i},B_{i}] = J_{-}\nu = \begin{pmatrix} -a+c & -b+d \\ -c & -d \end{pmatrix}.
$$
Let $t_{1}=\tr \nu$, $t_{2}=\tr \delta$. In this case, $c=t_{2}+t_{1}$. Stratifying as in Section \ref{subsec:2}, we get
\begin{itemize}
\item $Z'_{1}$ given by $(t_{1},t_{2})=(2,-2)$. We obtain
$$
e(Z'_{1})=(q-2)e(\overline{X}{}^h_{2})e(\overline{X}{}_{3}^{k})+e(\overline{X}{}_{2}^h)e(\overline{X}{}_{1}^{k})+e(\overline{X}{}_{0}^h)e(\overline{X}{}_{3}^{k}).
$$
\item $Z'_{2}$ given by $(t_{1},t_{2})=(-2,2)$. Analogously
$$
e(Z'_{2})=(q-2)e(\overline{X}{}^h_{3})e(\overline{X}{}_{2}^{k}) +e(\overline{X}{}^h_{3})e(\overline{X}{}_{0}^{k})+e(\overline{X}{}^h_{1})e(\overline{X}{}_{2}^{k}).
$$
\item $Z'_{3}$ given by the two values $(t_{1},t_{2})=(2,2),(-2,-2)$. We get
$$
e(Z'_{3})=q(e(\overline{X}{}^k_{2})e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{3}^{h})).
$$
\item $Z'_{4}$, divided into two possible cases:
\begin{itemize}
\item $Z'_{4,1}$ given by the lines $t_{1}=2$, $t_{2}\neq \pm 2$ and $t_{1}=-2$, $t_{2}\neq \pm 2$. We get
$$
e(Z'_{4,1})=q(e(\overline{X}{}^h_{2})+e(\overline{X}{}^h_{3}))e(\overline{X}{}_{4}^{k}/\ZZ_2).
$$
\item $Z'_{4,2}$ given by the lines $t_{2}=2$, $t_{1}\neq \pm 2$ and $t_{2}=-2$, $t_{1}\neq \pm 2$. We obtain
$$
e(Z'_{4,2})=q(e(\overline{X}{}_{2}^{k})+e(\overline{X}{}_{3}^{k}))e(\overline{X}{}_{4}^h/\ZZ_2).
$$
\end{itemize}
\item $Z'_{5}$ corresponding to $t_{1}=-t_{2}\neq \pm 2$. So $c=0$ and
$$
\eta=\left(\begin{array}{cc}
a & b\\
0 & a^{-1}
\end{array}\right), \ \delta=\left(\begin{array}{cc}
-a & -b+a^{-1}\\
0 & -a^{-1}
\end{array}\right).
$$
Therefore $e(Z'_5)=q\, e(\overline{Z}{}'_5)$, where $\overline{Z}{}'_{5}$ is a fibration over
$a\in \CC- \{0,\pm 1 \}$ whose fibers are
$\overline{X}{}_{4,a}^k\times \overline{X}{}_{4,-a}^{h}$. Thus the Hodge monodromy representation is given in (\ref{eqn:tres'}),
\begin{align*}
R(\overline{Z}{}'_{5}/\ZZ_2) & =A' T+ B'S_2+C'S_{-2}+D'S_0 ,\\
R(\overline{Z}{}'_{5}) & =(A'+D')T+(B'+C')N, \\
e(Z'_5) &= q\, e(\overline{Z}{}'_{5})=q((q-3)(A'+D')-2(B'+C')).
\end{align*}
\item $Z'_{6}$ corresponding to the open stratum $t_{1},t_{2}\neq \pm 2$, $t_{1}\neq -t_{2}$. The action of $U\cong \CC$ can be used to set $d=0$.
The total space, ignoring the condition $t_{1}\neq -t_{2}$, gives a contribution of $e(\overline{X}{}_{4}^k)e(\overline{X}{}_{4}^{h})$.
The fibration over the diagonal $(t_{1},-t_{1})$ has total space isomorphic
$\CC \times (\overline{Z}{}'_{5}/\ZZ_{2})$. Thus
\begin{align*}
e(\overline{Z}{}'_{5}/\ZZ_2) &=(q-2)A'-(B'+C'+D') ,\\
e(Z'_{6}) &=q(e(\overline{X}{}_{4}^k/\ZZ_2)e(\overline{X}{}_{4}^{h}/\ZZ_2)-e(\overline{Z}'_{5}/\ZZ_2)).
\end{align*}
\end{itemize}
Adding all up,
\begin{align*}
e(\overline{X}{}_{2}^{k+h}) = & \,e(\overline{X}{}_{2}^k)e(\overline{X}{}_{1}^{h})+e(\overline{X}{}_{0}^k)e(\overline{X}{}_{3}^{h})-2e(\overline{X}{}^k_{2})e(\overline{X}{}_{2}^{h})
+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{0}^{h})+e(\overline{X}{}_{1}^k)e(\overline{X}{}_{2}^{h})-2e(\overline{X}{}_{2}^k)e(\overline{X}{}_{3}^{h}) \\
& +q\, (e(\overline{X}{}^k_{2})+e(\overline{X}{}^k_{3})+e(\overline{X}{}_{4}^k/\ZZ_{2}))( e(\overline{X}{}_{2}^{h})+ e(\overline{X}{}_{3}^{h})+e(\overline{X}{}_{4}^{h}/\ZZ_{2}))
+q\,( e(\overline{Z}{}'_{5})- e(\overline{Z}{}'_{5}/\ZZ_2)).
\end{align*}
Setting $k=g-1$, $h=1$, and substituting the values $A',B',C',D'$ from (\ref{eqn:tres'.5})
and the values of $e(\overline{X}{}_j^1)$ and $a_1,b_1,c_1,d_1$ from Section \ref{sec:strata}, we have:
\begin{align}\label{eqn:eX2.} \tag{$\eta$}
e_3^g=e(\overline{X}{}_3^{g}) = & \,
(q^3 +3 q^2 )e_0^{g-1}+ (q^3 -2 q^2-3q) e_1^{g-1} \nonumber \\
&+ (q^5-3 q^3- 6 q^2 ) e_2^{g-1}+ (q^5+q^4+ 3 q^2 + 3 q)e_3^{g-1} \nonumber \\
& + (q^{6}-2q^5-3q^4+q^3+3q^2)a_{g-1} + (-q^5-q^4-4q^3+6q^2) b_{g-1} \nonumber \\
& +(-q^5+2q^4-4q^3+3q^2) c_{g-1}+(-2q^4-q^3+3q^2) d_{g-1}\nonumber
\end{align}
\section{Computation of $R(\overline{X}{}_4^{k+h})$}\label{subsec:4}
Now we move to the stratum $\overline{X}{}_4^g$. This one is controlled by a Hodge monodromy representation
$R(\overline{X}{}_4^g/\ZZ_2)$. We start by computing $R(\overline{X}{}_4^g)$.
As before, we write
$$
\overline{X}{}_{4}^{k+h}= \lbrace (A_{1},B_{1},\ldots, A_{k+h},B_{k+h},\lambda) \mid
\prod_{i=1}^{k}[A_{i},B_{i}]= \xi_\lambda \prod_{i=1}^{h}[B_{k+i},A_{k+i}] , \lambda\neq 0, \pm 1 \},
$$
where $\xi_\lambda=\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}$.
We are going to study the fibration $\overline{X}{}_{4}^{k+h}\to \CC-\{0,\pm 1\}$, with fiber
$\overline{X}{}_{4,\lambda}^{k+h}$.
Let
$$
\nu =\prod_{i=1}^{h}[B_{k+i},A_{k+i}] =\begin{pmatrix} a & b \\ c & d \end{pmatrix}, \qquad
\delta = \prod_{i=1}^{k}[A_{i},B_{i}] = \xi_\lambda \nu =\begin{pmatrix} \lambda a & \lambda b \\ \lambda^{-1} c & \lambda^{-1}d \end{pmatrix}.
$$
Note that $t_{1}=\tr \nu$, $t_{2}= \tr \delta$ and $\lambda$ determine $a,d$, and $bc=ad-1$.
We follow the stratification in terms of the traces $(t_{1},t_{2})$ given in \cite[Section 10]{lomune} for the genus $2$ case.
We decompose $\overline{X}{}_{4,\lambda}=\bigsqcup_{j=1}^7 Z_{j,\lambda}$, where
\begin{itemize}
\item $Z_{1,\lambda}$ corresponding to $t_{1}=\pm 2, t_{2}=\pm 2$. In this case both $\nu,\delta$ are of Jordan type.
Focus on the case $(t_1,t_2)=(2,2)$, the other cases being similar. Taking an adequate basis,
$$
\nu=\begin{pmatrix} 1 & x \\ 0 & 1\end{pmatrix}, \quad \delta=\begin{pmatrix} 1 & 0 \\ y & 1\end{pmatrix},
$$
for certain $x,y\in \CC^*$. We can fix $x=1$ by rescaling the basis vectors.
Since $\delta\nu^{-1}=\begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}$, we obtain that
$\lambda+\lambda^{-1}=2-xy$ , so $y$ is also fixed. When varying $\lambda$, we see that there is no monodromy around the punctures.
Therefore, taking also care of all four possibilities for $(t_1,t_2)$, we have
\begin{align*}
R(Z_{1}) & =e(Z_{1,\lambda})T \\ &=
(q-1)(e(\overline{X}{}_{2}^k)e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{2}^k)e(\overline{X}{}_{3}^{h})
+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{3}^k)e(\overline{X}{}_{3}^{h}))T \\
&=(q-1)(e(\overline{X}{}_{2}^k)+e(\overline{X}{}_{3}^k))(e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{3}^{h}))T ,
\end{align*}
where $T$ is the trivial representation.
\item $Z_{2,\lambda}$ corresponding to $(t_{1},t_2)=(2, \lambda+\lambda^{-1})$ and $(t_{1},t_{2})=(-2,-\lambda-\lambda^{-1})$. We focus on the first case.
In this situation $bc=0$, so there are three possibilities: either $b=c=0$ (in which case $\nu=\Id$) or $b=0,c\neq 0$ or $b\neq 0,c=0$ (in either case
there is a parameter in $\CC^*$ and $\nu\sim J_+$). In all cases, there is no monodromy for $\nu$ as $\lambda$ moves in
$\CC-\{0,\pm 1\}$. On the other hand,
$\delta\sim \begin{pmatrix} \lambda & 0 \\ 0 & \lambda^{-1} \end{pmatrix}$ gives a contribution $R(\overline{X}{}^{h}_{4})$.
The second case is analogous, changing $\Id$ by $-\Id$ and $J_{+}$ by $J_{-}$. Therefore
\begin{align*}
R(Z_{2}) & = (e(\overline{X}{}^k_{0})+2(q-1)e(\overline{X}{}^k_{2})+e( \overline{X}{}^k_{1})+2(q-1)e(\overline{X}{}^k_{3}))R(\overline{X}{}_{4}^{h}).
\end{align*}
\item $Z_{3,\lambda}$ corresponding to $(t_{1},t_2)=(\lambda+\lambda^{-1},2)$ and $(t_{1},t_{2})=(-\lambda-\lambda^{-1},-2)$.
This is completely analogous to the previous case, so
\begin{align*}
R(Z_{3}) & = (e(\overline{X}{}^h_{0})+2(q-1)e(\overline{X}{}^h_{2})+e( \overline{X}{}^h_{1})+2(q-1)e(\overline{X}{}^h_{3}))R(\overline{X}{}_{4}^{k}).
\end{align*}
\item $Z_{4,\lambda}$ defined by $t_{1}=2, t_{2}\neq \pm 2, \lambda+\lambda^{-1}$ and $t_{1}=-2, t_{2}\neq \pm 2, -\lambda-\lambda^{-1}$.
For each $\lambda$, $(t_{1},t_{2})$ move in (two) punctured lines $\{ (t_{1},t_{2}) \mid t_{1}=\pm 2, t_{2}\neq \pm 2, \pm (\lambda+\lambda^{-1}) \}$,
where $\nu$ is of Jordan
type and $\delta$ is of diagonal type. Both families can be trivialized, giving a contribution of
$e(\overline{X}{}_{2}^k)$ times $e(\overline{X}{}_{4}^{h}/\mathbb{Z}_{2})$ for one line, and
$e(\overline{X}{}_{3}^k)$ times $e(\overline{X}{}^{h}_{4}/\ZZ_2)$ for the other line. The missing fiber $\overline{X}{}_{4,\lambda}^{h}$ over $\lambda+\lambda^{-1}$,
which needs to be removed, has monodromy given by $R(\overline{X}{}_{4}^{h})$ as $\lambda$ varies. Finally, there is a $(q-1)$ factor due to the fact that $bc\neq 0$. Therefore
\begin{align*}
R(Z_{4}) = & (q-1)(e(\overline{X}{}_{2}^k)+e(\overline{X}{}_{3}^k))(e(\overline{X}{}_{4}^{h}/\mathbb{Z}_{2})T-R(\overline{X}{}^{h}_{4})) .
\end{align*}
\item $Z_{5,\lambda}$ defined by $t_{2}=2, t_{1}\neq \pm 2, \lambda +\lambda^{-1}$ and $t_{2}=-2, t_{1}\neq \pm 2,
-\lambda-\lambda^{-1}$. Similarly to $Z_{4,\lambda}$, we obtain
\begin{align*}
R(Z_{5}) & = (q-1)(e(\overline{X}{}_{2}^h)+e(\overline{X}{}_{3}^h))(e(\overline{X}{}_{4}^{k}/\mathbb{Z}_{2})T-R(\overline{X}{}^{k}_{4})) .
\end{align*}
\item $Z_{6,\lambda}$. This stratum corresponds to the set $\{ (t_{1},t_{2})\mid t_{1},t_{2}\neq \pm 2, ad=1 \}$,
which is a hyperbola $H_\lambda$ for every $\lambda$ (see \cite[Figure 1, Section 10]{lomune}).
There is a contribution of $2q-1$ which accounts for $bc=0$. Parametrizing
$H_\lambda$ by $\mu \in \CC^* -\{ \pm 1, \pm \lambda^{-1} \}$ as in \cite[Section 10]{lomune},
we obtain a fibration over $\CC^* -\{ \pm 1, \pm \lambda^{-1} \}$
whose fiber over $\mu$ is $\overline{X}{}^k_{4,\mu}\times \overline{X}{}^{h}_{4,\lambda\mu}$, for each $\lambda$.
When $\lambda$ varies over $\CC -\{0, \pm 1 \}$,
we can extend the local system trivially to the cases $\lambda,\mu = \pm 1$. This extension
can be regarded as a local system over the set of $(\lambda,\mu) \in \CC^*\times \CC^*$,
$$
\overline{Z}_{6}=\overline{X}{}^k_{4}\times m^{*}\overline{X}{}^{h}_{4} \longrightarrow \CC^*\times\CC^*\, ,
$$
where $m: \CC^*\times \CC^* \rightarrow \CC^*$ maps $(\lambda,\mu)\mapsto \lambda\mu$.
The Hodge monodromy representation of $\overline{Z}_{6}$ belongs to $R( \mathbb{Z}_{2}\times \mathbb{Z}_{2})[q]$ (with generators
$N_{1},N_{2}$ denoting the representation which is not trivial over the generator of the
fundamental group of the first and second copies of $\CC^*$, respectively, and
$N_{12}=N_1\otimes N_2$). So we get
\begin{align*}
R_{ \CC^*\times\CC^*} (\overline{Z}_{6}) & =((a_k+d_k)T+(b_k+c_k)N_{2})\otimes ((a_h+d_h)T+(b_h+c_h)N_{12}) \\
& =(a_h+d_h)(a_k+d_k) T +(b_h+c_h)(b_k+c_k) N_{1} +(a_h+d_h) (b_k+c_k) N_{2}+ (b_h+c_h)(a_k+d_k)N_{12}.
\end{align*}
To obtain the Hodge monodromy representation over $\lambda \in \CC^*$, we use the projection
$\pi_{1}:\CC^*\times\CC^* \rightarrow \CC^*$, $(\lambda,\mu)\mapsto \lambda$, which maps
$T\mapsto e(T) T=(q-1)T$, $N_2 \mapsto e(N_2)T=0$, $N_1\mapsto e(T) N=(q-1)N$,
$N_{12} \mapsto e(N_2)N=0$ for the representations. Therefore $R_{\CC^*}(\overline{Z}_6)=
(q-1)((a_h+d_h)(a_k+d_k) T +(b_h+c_h)(b_k+c_k) N)$. Now we have to substract the contribution
from the sets $\mu=\pm1,\pm\lambda^{-1}$. The first two yield $-2e(\overline{X}{}_4^k)R(\overline{X}{}_4^h)$ and
the second two yield $-2e(\overline{X}{}_4^h)R(\overline{X}{}_4^k)$. Therefore
\begin{align*}
R(\overline{Z}_{6}) = &(q-1)((a_h+d_h)(a_k+d_k) T +(b_h+c_h)(b_k+c_k) N)-2e(\overline{X}{}_{4,\lambda}^k)R(\overline{X}{}_4^h)-2e(\overline{X}{}_{4,\lambda}^h)R(\overline{X}{}_4^k),\\
R(Z_6) = &(2q-1) R(\overline{Z}_{6}) .
\end{align*}
\item $Z_{7,\lambda}$ corresponding to the open stratum given by the set of $(t_{1},t_{2})$ such that $t_{i} \neq \pm 2$,
$i=1,2$ and $(t_{1},t_{2})\not\in H_\lambda$. If we forget about the condition $(t_{1},t_{2})\in H_\lambda$, $Z_{7,\lambda}$ is a fibration over
$(t_{1},t_{2})$ with fiber isomorphic to $\overline{X}{}^h_{4,\mu_{1}}\times \overline{X}{}^{k}_{4,\mu_{2}}$, $t_{i}=\mu_i+\mu^{-1}_i$, $i=1,2$.
Its monodromy is trivial, as the local system is trivial when $\lambda$ varies. The contribution over $H_\lambda$, already computed in the previous
stratum, is $R(\overline{Z}_6)$. So we get
\begin{align*}
R(Z_{7}) =& (q-1)(e(\overline{X}{}^k_{4}/\mathbb{Z}_{2})e(\overline{X}{}^h_{4}/\ZZ_2)T-R(\overline{Z}_{6})) .
\end{align*}
\end{itemize}
Adding all the pieces, we get
\begin{align*}
R(\overline{X}{}^{k+h}_{4}) =&
(q-1)(e(\overline{X}{}_{2}^k)+e(\overline{X}{}_{3}^k)+e(\overline{X}{}^k_{4}/\mathbb{Z}_{2}))(e(\overline{X}{}_{2}^{h})+e(\overline{X}{}_{3}^{h})+e(\overline{X}{}^h_{4}/\mathbb{Z}_{2}))T\nonumber \\
&+ (e(\overline{X}{}^k_{0})+e( \overline{X}{}^k_{1})+(q-1)e(\overline{X}{}^k_{2})+(q-1)e(\overline{X}{}^k_{3}))R(\overline{X}{}_{4}^{h}) \\
& +(e(\overline{X}{}^h_{0})+e( \overline{X}{}^h_{1})+(q-1)e(\overline{X}{}^h_{2})+(q-1)e(\overline{X}{}^h_{3}))R(\overline{X}{}_{4}^{k}) + q R(\overline{Z}_{6}). \nonumber
\end{align*}
Setting $k=g-1$, $h=1$, we have:
\begin{align} \label{eqn:R4}
R(\overline{X}{}^{g}_{4}) = & \, \Big(
(q^3-1)e_0^{g-1}+(q^3-1)e_1^{g-1}+(q^5-3q^3+2q^2)e_2^{g-1}+(q^5-3q^3+2q^2)e_3^{g-1} \nonumber \\
&+(q^6- 2q^5 -2q^4 + 4q^3- 3 q^2 +2)a_{g-1}
+(- q^5 -q^4 + 2q^3- 2 q^2 +q+1)(b_{g-1}+c_{g-1}) \nonumber\\
&+(-q^4-2q^2+2q+1)d_{g-1} \Big)T \\
&+ \Big( (3q^2-3q)e_0^{g-1}+(3q^2-3q)e_1^{g-1}+(3 q^3 - 6 q^2 + 3 q)e_2^{g-1}+(3 q^3 - 6 q^2 + 3 q)e_3^{g-1} \nonumber\\
&+(-6q^3+6q^2)(a_{g-1}+d_{g-1}) +( 4q^4-14q^3+ 10 q^2) (b_{g-1}+c_{g-1}) \Big)N \nonumber
\end{align}
\section{Computation of $R(\overline{X}{}_4^{g}/\ZZ_2)$}\label{subsec:5}
\begin{lem} \label{lem:RR}
Suppose that $R(\overline{X}{}_4^k/\ZZ_2)=a_k T+b_k S_2+ c_k S_{-2}+ d_k S_0$, for all $k<g$.
Then the Hodge monodromy representation $R(\overline{X}{}^g_{4}/\mathbb{Z}_{2})$ is of the form
$R(\overline{X}{}^g_{4}/\mathbb{Z}_{2})= a_gT+b_gS_{2}+c_gS_{-2}+d_gS_{0}$, for some polynomials $a_g,b_g,c_g,d_g\in \ZZ[q]$.
\end{lem}
\begin{proof}
The Hodge monodromy representation $R(\overline{X}{}^g_{4}/\mathbb{Z}_{2})$ lies in the representation ring of the
fundamental group of $\CC-\{\pm 2\}$. Under the double cover $\CC-\{0,\pm 1\} \to \CC-\{ \pm 2\}$, it reduces
to $R(\overline{X}{}^g_{4})$. By Section \ref{subsec:4}, $R(\overline{X}{}^g_{4})$ is of order $2$. Hence
$R(\overline{X}{}^g_{4}/\mathbb{Z}_{2})$ has only monodromy of order $2$ over the loops $\gamma_{\pm 2}$
around the points $\pm 2$. This is the statement of the lemma.
\end{proof}
\begin{prop} \label{prop:balanced}
All $\overline{X}{}_0^g$, $\overline{X}{}_1^g$, $\overline{X}{}_2^g$, $\overline{X}{}_3^g$, $\overline{X}{}_4^g$ and $\overline{X}{}_{4,\lambda}^g$ are of balanced type.
\end{prop}
\begin{proof}
By \cite[Proposition 2.8]{lomune}, if $Z= \bigsqcup Z_i$ and all $Z_i$ are of balanced type, then $Z$ is of balanced type.
Also, if $\ZZ_2$ acts on $Z$ and $Z$ is of balanced type, so is $Z/\ZZ_2$. Also Theorem \ref{thm:general-fibr} says
that if $F\to Z \to B$ is a fibration with $F$ of balanced type, with
either $B=\CC-\{0,\pm 1\}$ and Hodge monodromy $R(Z)=a T+ b N$ or
$B=\CC-\{\pm 2\}$ and Hodge monodromy $R(Z)=a T+ b S_2+cS_{-2}+dS_0$, then $Z$ is of balanced type.
In \cite{lomune} it is proved that the result holds for $g=1,2$. Also $\SL(2,\CC)$, $\PGL(2,\CC)$, $\PGL(2,\CC)/D$,
$\PGL(2,\CC)/U$ are of balanced type. Now assume that all
$\overline{X}{}_0^k$, $\overline{X}{}_1^k$, $\overline{X}{}_2^k$, $\overline{X}{}_3^k$, $\overline{X}{}_4^k$ and $\overline{X}{}_{4,\lambda}^k$ are of balanced type
for $k<g$. A look at the description of all strata for which we compute the E-polynomials
in Sections \ref{subsec:0}--\ref{subsec:3} convinces us that $\overline{X}{}_0^g$, $\overline{X}{}_1^g$, $\overline{X}{}_2^g$, $\overline{X}{}_3^g$
are of balanced type. The same is true for $\overline{X}{}_{4,\lambda}^g$ by the stratification in Section \ref{subsec:4}.
Finally formula (\ref{eqn:R4}) gives us that $\overline{X}{}_4^g$ is also of balanced type.
\end{proof}
\medskip
Now to find the four polynomials $a_g,b_g,c_g,d_g\in \ZZ[q]$, we need four equations.
Two come from the fact that $R(\overline{X}{}^g_{4})=(a_g+d_g)T+(b_g+c_g)N$. From (\ref{eqn:R4}), we have
\begin{align}\label{eqn:aaa}
a_g+d_g =&
(q^3-1)e_0^{g-1}+(q^3-1)e_1^{g-1}+(q^5-3q^3+2q^2)e_2^{g-1}+(q^5-3q^3+2q^2)e_3^{g-1} \\
&+(q^6- 2 q^5 -2 q^4 + 4 q^3- 3 q^2 +2)a_{g-1}
+(-q^5-q^4+2q^3-2q^2+q+1) (b_{g-1}+c_{g-1}) \nonumber\\ &+(-q^4-2q^2+2q+1)d_{g-1} \nonumber\\
b_g+d_g =& (3q^2-3q)e_0^{g-1}+(3q^2-3q)e_1^{g-1}+(3 q^3 - 6 q^2 + 3 q)e_2^{g-1}+(3 q^3 - 6 q^2 + 3 q)e_3^{g-1}
\label{eqn:bbb}\\
&+(-6q^3+6q^2)(a_{g-1}+d_{g-1}) +( 4q^4-14q^3+ 10 q^2) (b_{g-1}+c_{g-1}).\nonumber
\end{align}
One more equation is obtained by computing $e(\overline{X}{}_0^{g+1})$ with $k=g$, $h=1$ and with $k=g-1$, $h=2$, and
equating. From ($\alpha$) with $k=g$, we get
\begin{align}\label{eqn:cc}
e_0^{g+1}= & \, (q^4 + 4 q^3 - q^2 - 4 q) e_0^{g}+ (q^3-q) e_1^{g} \nonumber \\
&+ ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_2^{g}+ (q^5+3q^4-q^3-3q^2) e_3^{g} \\
&+ (q^6-2q^5-4q^4+3q^2+2q) a_{g} + (-q^5-4q^4+4q^2+q)b_{g} \nonumber\\
&+ (2q^5-7q^4-3q^3+7q^2+q) c_{g} + (-5q^4-q^3+5q^2+q) d_{g} \nonumber
\end{align}
Using (\ref{eqn:eX0}) with $k=g-1$, $h=2$, and the
values of $e(\overline{X}{}_j^2)$ and $a_2,b_2,c_2,d_2$ from Section \ref{sec:strata}, we have
\begin{align}\label{eqn:cc'}
e_0^{g+1}= & \, (q^9+ q^8+12q^7+2q^6-3q^4-12q^3-q) e_0^{g-1}
+ (q^9-3q^7-30q^6+30q^4+3q^3-q) e_1^{g-1}\nonumber \\
&+ ( q^{11}-4q^9-4q^8-36q^7+24q^5+4q^4+15q^3)e_2^{g-1} \nonumber \\
&+ (q^{11}-4q^9+15q^8+9q^7+30q^6-6q^5-45q^4) e_3^{g-1} \nonumber \\
&+ (q^{12}- 2 q^{11}- 4 q^{10}+ 6 q^9 - 6 q^8+ 18 q^7 - 6 q^6 - 18 q^5+ 15 q^4- 6 q^3+2 q ) a_{g-1} \\
&+ (- q^{11}- q^{10}+ 3 q^9 - 42 q^8+ 54 q^7 + 30 q^6- 54 q^5 + 12 q^4 - 3 q^3 + q^2 +q )b_{g-1} \nonumber \\
&+ (- q^{11}- q^{10}+ 18 q^9 - 27 q^8 + 24 q^7- 39 q^5+ 27 q^4- 3 q^3+ q^2 +q ) c_{g-1} \nonumber\\
&+ (- 2 q^{10}- 9 q^8 + 24 q^7 - 21 q^5+ 9 q^4 - 4 q^3+ 2 q^2 +q ) d_{g-1} \nonumber
\end{align}
A fourth equation are obtained by computing $e(\overline{X}{}_1^{g+1})$ with $k=g$, $h=1$ and with $k=g-1$, $h=2$, and
equating. From ($\beta$) with $k=g$, we get
\begin{align}\label{eqn:dd}
e_1^{g+1}=& \, (q^3-q)e_0^{g}+ (q^4 + 4 q^3 - q^2 - 4 q) e_1^{g} \nonumber\\
&+ (q^5+3q^4-q^3-3q^2)e_2^{g}+ ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_3^{g} \\
& +(q^6-2q^5-4q^4+3q^2+2q) a_{g} + (2q^5-7q^4-3q^3+7q^2+q) b_{g} \nonumber\\
&+ (-q^5-4q^4+4q^2+q) c_{g} + (-5q^4-q^3+5q^2+q) d_{g} \nonumber
\end{align}
Using (\ref{eqn:eX1}) with $k=g-1$, $h=2$, and the
values of $e(\overline{X}{}_j^2)$ and $a_2,b_2,c_2,d_2$ from Section \ref{sec:strata}, we have
\begin{align}\label{eqn:dd'}
e_1^{g+1}= & \, (q^9-3q^7-30q^6+30q^4+3q^3-q) e_0^{g-1} +(q^9+ q^8+12q^7+2q^6-3q^4-12q^3-q) e_1^{g-1} \nonumber\\
&+ (q^{11}-4q^9+15q^8+9q^7+30q^6-6q^5-45q^4) e_2^{g-1} \nonumber \\
&+ (q^{11}-4q^9-4q^8-36q^7+24q^5+4q^4+15q^3)e_3^{g-1} \\
&+ (q^{12}- 2 q^{11}- 4 q^{10}+ 6 q^9 - 6 q^8+ 18 q^7 - 6 q^6 - 18 q^5+ 15 q^4- 6 q^3+2 q ) a_{g-1} \nonumber\\
&+ (- q^{11}- q^{10}+ 18 q^9 - 27 q^8 + 24 q^7- 39 q^5+ 27 q^4- 3 q^3+ q^2 +q ) c_{g-1} \nonumber\\
&+ (- q^{11}- q^{10}+ 3 q^9 - 42 q^8+ 54 q^7 + 30 q^6- 54 q^5 + 12 q^4 - 3 q^3 + q^2 +q )b_{g-1} \nonumber\\
&+ (- 2 q^{10}- 9 q^8 + 24 q^7 - 21 q^5+ 9 q^4 - 4 q^3+ 2 q^2 +q ) d_{g-1} \nonumber
\end{align}
The solutions to (\ref{eqn:aaa}), (\ref{eqn:bbb}), (\ref{eqn:cc})=(\ref{eqn:cc'}) and (\ref{eqn:dd})=(\ref{eqn:dd'}),
and using the values of ($\alpha$), ($\beta$), ($\gamma$) and ($\delta$), are given by
\begin{align*}
a_g =& q^3 e_{0}^{g-1} + q^3 e_{1}^{g-1} +(q^5-3q^3) e_{2}^{g-1} + (q^5-3q^3)e_{3}^{g-1} \\
& + (q^6-2q^5-2q^4+4q^3+q^2)a_{g-1} + (-q^5-q^4+2q^3)b_{g-1} +(-q^5-q^4+2q^3)c_{g-1} -2q^4 d_{g-1}\\
b_g =&-3q e_{0}^{g-1} +3q^2 e_{1}^{g-1} +(3q^3+3q) e_{2}^{g-1} -6q^2 e_{3}^{g-1} \\
& +(-3q^3+3q^2)a_{g-1} + (4q^4-6q^3+4q^2)b_{g-1} + (-8q^3+6q^2)c_{g-1} + (q^2-3q^3+3q^2)d_{g-1} \\
c_g =& 3q^2 e_{0}^{g-1} -3q e_{1}^{g-1} -6q^2 e_{2}^{g-1} +(3q^3+3q) e_{3}^{g-1}\\
& (-3q^3+3q^2)a_{g-1} + (-8q^3+6q^2)b_{g-1} + (4q^4-6q^3+4q^2) c_{g-1} + (-3q^3+3q^2) d_{g-1} \\
d_g =& -e_{0}^{g-1} -e_{1}^{g-1} +2q^{2}e_{2}^{g-1} +2q^{2}e_{3}^{g-1}\\
& +(-4q^2+2)a_{g-1} + (-2q^2+q+1)b_{g-1} + (-2q^{2}+q+1)c_{g-1} + (q^{4}-2q^{2}+2q+1)d_{g-1}
\end{align*}
We put this together with equations ($\alpha$), ($\beta$), ($\gamma$) and ($\delta$)
\begin{align*}
e^g_0 =& (q^4 + 4 q^3 - q^2 - 4 q) e_0^{g-1}+ (q^3-q) e_1^{g-1} + ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_2^{g-1}+ (q^5+3q^4-q^3-3q^2) e_3^{g-1} \\
&+ (q^6-2q^5-4q^4+3q^2+2q) a_{g-1} + (-q^5-4q^4+4q^2+q)b_{g-1} \\
&+ (2q^5-7q^4-3q^3+7q^2-q) c_{g-1} + (-5q^4-q^3+5q^2-q) d_{g-1}\\
e^g_1 =& (q^3-q)e_0^{g-1}+ (q^4 + 4 q^3 - q^2 - 4 q) e_1^{g-1} + (q^5+3q^4-q^3-3q^2)e_2^{g-1}+ ( q^5- 2 q^4 - 4 q^3+ 2 q^2 +3 q ) e_3^{g-1} \\
&+(q^6-2q^5-4q^4+3q^2+2q) a_{g-1} + (2q^5-7q^4-3q^3+7q^2+q) b_{g-1} \\
&+ (-q^5-4q^4+4q^2+q) c_{g-1} + (-5q^4-q^3+5q^2+q) d_{g-1} \\
e^g_2 =& (q^3 - 2 q^2 - 3 q)e_0^{g-1}+ (q^3 + 3 q^2) e_1^{g-1} + (q^5+q^4+ 3 q^2 + 3 q)e_2^{g-1}+ (q^5-3 q^3- 6 q^2 ) e_3^{g-1} \\
& +(q^6-2q^5-3q^4+q^3+3q^2)a_{g-1}+(-q^5+2q^4-4q^3+3q^2) b_{g-1} \\
& +(-q^5-q^4-4q^3+6q^2) c_{g-1}+(-2q^4-q^3+3q^2)d_{g-1} \\
e^g_3=& (q^3 +3 q^2 )e_0^{g-1}+ (q^3 -2 q^2-3q) e_1^{g-1} + (q^5-3 q^3- 6 q^2 ) e_2^{g-1}+ (q^5+q^4+ 3 q^2 + 3 q)e_3^{g-1} \\
& + (q^{6}-2q^5-3q^4+q^3+3q^2)a_{g-1} + (-q^5-q^4-4q^3+6q^2) b_{g-1} \\
& +(-q^5+2q^4-4q^3+3q^2) c_{g-1}+(-2q^4-q^3+3q^2) d_{g-1}
\end{align*}
Hence there is a $8\times 8$-matrix $M$ such that if we write
$v_g=( e^g_0, e^g_1, e^g_2, e^g_3,a_g,b_g,c_g,d_g)^t$,
\begin{equation}\label{eqn:matrix}
v_g =M v_{g-1},
\end{equation}
for all $g\geq 3$. $M$ is the following matrix
\begin{equation} \label{matrixg-1-->g}
\scalemath{0.68}{\left( \begin{array}{c@{\hspace{2em}}c@{\hspace{2em}}c@{\hspace{2em}}c@{\hspace{2em}}c@{\hspace{2em}}c@{\hspace{2em}}c@{\hspace{2em}}c}
q^4+4q^3 & q^3-q & q^5 -2q^4-4q^3 & q^5+3q^4 & q^6 -2q^5 -4q^4 & -q^5-4q^4 & 2q^5 -7q^4 -3q^3 & -5q^4-q^3 \\
-q^2-4q & & +2q^2+3q & -q^3 -3q^2 & +3q^2 +2q & +4q^2+q & +7q^2+q & +5q^2 +q \\
& & & & & & & \\
q^3 -q & q^4 +4q^3 & q^5+3q^4 & q^5-2q^4-4q^3 & q^6-2q^5-4q^4 & 2q^5-7q^4-3q^3 & -q^5-4q^4 & -5q^4-q^3 \\
& -q^2-4q & -q^3-3q^2 & +2q^2+3q & +3q^2+2q & +7q^2 +q & +4q^2+q & +5q^2+q \\
& & & & & & & \\
q^3-2q^2 & q^3+3q^2 & q^5+q^4 & q^5 -3q^3 & q^6-2q^5-3q^4 & -q^5+2q^4 & -q^5 -q^4 & -2q^4 -q^3 \\
-3q & & +3q^2+3q & -6q^2 & +q^3+3q^2 & -4q^3+3q^2 & -4q^3 +6q^2 & +3q^2 \\
& & & & & & & \\
q^3+3q^2 & q^3-2q^2 & q^5-3q^3 & q^5+q^4 & q^6-2q^5-3q^4 & -q^5-q^4 & -q^5 +2q^4 & -2q^4 -q^3 \\
& -3q & -6q^2 & +3q^2+3q & +q^3+3q^2 & -4q^3+6q^2 & -4q^3 +3q^2 & +3q^2 \\
& & & & & & & \\
q^3 & q^3 & q^5-3q^3 & q^5-3q^3 & q^6 -2q^5-2q^4 & -q^5-q^4 & -q^5-q^4 & -2q^4 \\
& & & & +4q^3+q^2 & +2q^3 & +2q^3 & \\
& & & & & & & \\
-3q & 3q^2 & 3q^2+3 & -6q^2 & -3q^3+3q^2 & 4q^4-6q^3 +4q^2 & -8q^3+6q^2 & -3q^3+3q^2 \\
& & & & & & & \\
3q^2 & -3q & -6q^2 & 3q^3+3q & -3q^3+3q^2 & -8q^3+6q^2 & 4q^4-6q^3+4q^2 & -3q^3+3q^2 \\
& & & & & & & \\
-1 & -1 & 2q^2 & 2q^2 & -4q^2+2 & -2q^2+q+1 & -2q^2+q+1 & q^4 -2q^2 \\
& & & & & & & +2q+1
\end{array} \right)}.
\end{equation}
The starting vector is given in Section \ref{sec:strata},
$ v_2= (e_0^2,e_1^2,e_2^2,e_3^2, a_2,b_2,c_2,d_2)^t=
( q^9+q^8+12q^7+2q^6-3q^4-12q^3-q,
q^9-3q^7-30q^6+30q^4+3q^3-q,
q^9-3q^7-4q^6-39q^5-4q^4-15q^3,
q^9-3q^7+15q^6+6q^5+45q^4,
q^9-3q^7+6q^5,-(45q^5+15q^3),15q^6+45q^4,-6q^4+3q^2-1)^t$.
If we write $v_1=(e_0^1,e_1^1,e_2^1,e_3^1, a_1,b_1,c_1,d_1)^t=
( q^{4}+4q^{3}-q^2-4q,q^{3}-q , q^3-2q^2-3 q,q^3 +3 q^2 ,q^{3},-3q,3q^{2},-1)^t$
and
$$
v_0=(1,0,0,0,0,0,0,0)^t,
$$
then equation (\ref{eqn:matrix}) holds for all $g\geq 1$. So
$$
v_g=M^g v_0 .
$$
\begin{rem}
As in \cite{lomune}, we can stratify $\SL(2,\mathbb{C})^{2g} = \bigsqcup_{i=0}^{4} X_{i}^{g}$, with
\begin{itemize}
\item $X_0^g=\overline{X}{}_{0}^{g}$, $e(X_0^g)=e_0^g$.
\item $X_1^g=\overline{X}{}_{1}^{g}$, $e(X_1^g)=e_1^g$.
\item $X_{2}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \mid \prod_{i=1}^{g}[A_{i},B_{i}]\sim J_+\rbrace \cong
(\PGL(2,\CC)/U) \times \overline{X}{}_2^g$. So $e(X_2^g)=(q^2-1)e_2^g$.
\item $X_{3}^{g}= \lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \mid \prod_{i=1}^{g}[A_{i},B_{i}]\sim J_- \}
\cong (\PGL(2,\CC)/U) \times \overline{X}{}_3^g$. So $e(X_3^g)=(q^2-1)e_3^g$.
\item $X_{4}^{g}=
\lbrace (A_{1},B_{1},\ldots, A_{g},B_{g}) \mid
\prod_{i=1}^{g}[A_{i},B_{i}] \sim \xi_\lambda,$ for some $\lambda\in \CC-\{0,\pm 1\}\rbrace$. Here
$X_4^g \cong ( \PGL(2,\CC)/D \times \overline{X}{}_4^g )/\ZZ_2$. Using (\ref{eqn:RX4-RX4Z2->eX4}),
we have $e(X_4^g)=(q^3-2q^2-q)a_g-(q^2+q)(b_g+c_g)-2q d_g$.
\end{itemize}
Therefore it must be
\begin{equation}\label{EpolyStratificationSL2Cg}
(q^3-q)^{2g} = e_{0}^{g}+e_{1}^{g}+(q^2-1)(e_{2}^{g}+e_{3}^{g})+(q^3-2q^2-q)a_{g}-(q^2+q)(b_{g}+c_{g})-2qd_{g}.
\end{equation}
We can prove (\ref{EpolyStratificationSL2Cg}) numerically by induction on $g\geq 0$, using
(\ref{eqn:matrix}). The equation (\ref{EpolyStratificationSL2Cg}) is certainly true for $g=0$. Suppose it holds
for $g-1$ and let $w=(w_{0},\ldots,w_{7})^t=Mv_{g-1}$. Then
an easy computation gives
\begin{align*}
w_{0}&+w_{1}+(q^2-1)(w_{2}+w_{3})+(q^3-2q^2-q)w_{4}-(q^2+q)(w_{5}+w_{6})-2qw_{7} \\ & = (q^3-q)^2(e_{0}^{g-1}+e_{1}^{g-1}+(q^2-1)(e_{2}^{g-1}+e_{3}^{g-1})+(q^3-2q^2-q)a_{g-1}-(q^2+q)(b_{g-1}+c_{g-1})-2qd_{g-1}) \\ & = (q^3-q)^{2} (q^3-q)^{2g-2} =(q^3-q)^{2g},
\end{align*}
so equation (\ref{EpolyStratificationSL2Cg}) holds for $v_{g}=w=Mv_{g-1}$.
\end{rem}
We start by proving Corollary \ref{cor:Hausel} using (\ref{matrixg-1-->g}).
\begin{thm}
For every $g\geq 1$, we have
$e(\mathcal{M}_{J_{-}})+(q+1)e(\mathcal{M}_{-\Id})=e(\mathcal{M}_{\xi_{\lambda}})$.
\end{thm}
\begin{proof}
First, $\cM_{-\Id}=\overline{X}{}_1^g/\PGL(2,\CC)$, so $e(\cM_{-\Id})=e_1^g/(q^3-q)$. Second,
$\cM_{J_+}=\overline{X}{}_2^g/U$, so $e(\cM_{J_+})=e_3^g/q$. And third,
$\cM_{\xi_\lambda}= \overline{X}{}_{4,\lambda}^g /D$, so $e(\cM_{\xi_\lambda})=e(\overline{X}{}_{4,\lambda}^g) /(q-1)=
(a_g+b_g+c_g+d_g)/(q-1)$.
The assertion is thus equivalent to
\begin{equation} \label{HauselIdentity}
(q^2-1)e_{3}^{g}+(q+1)e_{1}^{g}=(q^2+q)(a_{g}+b_{g}+c_{g}+d_{g}),
\end{equation}
for all $g\geq 1$. We proceed by induction starting with $g=0$, where it obviously holds.
If we assume that (\ref{HauselIdentity}) holds for $g-1$, then using (\ref{matrixg-1-->g}),
\begin{align*}
(q^2+q) & (a_{g}+ b_{g}+c_{g}+d_{g})-(q^2-1)e_{3}^{g}-(q+1)e_{1}^{g} \\
& = -q^2(q+1)(q-1)^2e_{1}^{g-1} -q^2(q+1)(q-1)^3e_{3}^{g-1}
+q^3(q+1)(q-1)^2(a_{g-1}+b_{g-1}+c_{g-1}+d_{g-1}) \\
& = q^2(q-1)^2((q^2+q)(a_{g-1}+b_{g-1}+c_{g-1}+d_{g-1}) -(q^2-1)e_{3}^{g-1}-(q+1)e_{1}^{g-1}) = 0,
\end{align*}
by induction hypothesis.
\end{proof}
Since $v_{g}=M^{g}v_{0}$, we can obtain closed formulas for $e_{0}^{g},e_{1}^{g},e_{2}^{g},e_{3}^{g},a_{g},b_{g},c_{g},d_{g}$. We summarize them in the following
\begin{prop} \label{Polynomialseig}
For all $g\geq 1$,
\begin{align*}
e_{0}^{g} & = (q^3-q)\left( (q^3-q)^{2g-2}+(q^2-1)^{2g-2}-(q^2-q)^{2g-2}+\frac{1}{2}q^{2g-2}(q+2^{2g}-1)\left( (q+1)^{2g-2}+(q-1)^{2g-2} \right) \right) \\
e_{1}^{g} & = (q^{3}-q)\left( (q^3-q)^{2g-2}+(q^2-1)^{2g-2}-2^{2g-1}(q^{2}+q)^{2g-2}+(2^{2g-1}-1)(q^{2}-q)^{2g-2}) \right)\\
e_{2}^{g} & = (q^{3}-q)^{2g-1}+(2^{2g-1}-1)(q^{2}-q)^{2g-1}-2^{2g-1}(q^{2}+q)^{2g-1}+\frac{1}{2}q^{2g-1}(q-1) \left( (q-1)^{2g-1}-(q+1)^{2g-1} \right) \\
e_{3}^{g} & = (q^{3}-q)^{2g-1}+(2^{2g-1}-1)(q^{2}-q)^{2g-1}+2^{2g-1}(q^{2}+q)^{2g-1} \\
a_{g} & = (q^{3}-q)^{2g-1}+\frac{1}{2}q^{2g-1} \left( (q+1)^{2g-1}-(q-1)^{2g-1} \right) \\
b_{g} & = 2^{2g-1}(q^{2}-q)^{2g-1}-2^{2g-1}(q^{2}+q)^{2g-1} +\frac{1}{2}q^{2g-1} \left( (q+1)^{2g-1}-(q-1)^{2g-1} \right) \\
c_{g} & = 2^{2g-1}(q^{2}-q)^{2g-1}+2^{2g-1}(q^{2}+q)^{2g-1}-\frac{1}{2}q^{2g-1} \left( (q+1)^{2g-1}+(q-1)^{2g-1} \right) \\
d_{g} & = (q^{2}-1)^{2g-1}-\frac{1}{2}q^{2g-1} \left( (q+1)^{2g-1} +(q-1)^{2g-1} \right),
\end{align*}
and also
\begin{align*}
e_{4,\xi_{\lambda}}^{g}& = a_{g}+b_{g}+c_{g}+d_{g} =(q^{3}-q)^{2g-1}+(q^{2}-1)^{2g-1}+(2^{2g}-2)(q^{2}-q)^{2g-1}.
\end{align*}
\end{prop}
\begin{proof}
We know that $v_{g}=M^{g}v_{0}$, where $M$ is given in \ref{matrixg-1-->g}.
There exists a matrix $Q$ with entries in the fraction field of $\mathbb{Z}[q]$ such that $M=QDQ^{-1}$,
where $D$ is the diagonal matrix
$$
D=\begin{pmatrix}
(q^2-q)^2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & (q^2+q)^2 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 4(q^2-q)^2 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 4(q^2+q)^2 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & (q^2-1)^2 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & (q^3-q)^2 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & (q^2-q)^2 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & (q^2+q)^2
\end{pmatrix}
$$
As $M^{g}=QD^{g}Q^{-1}$, a straightforward computation gives the desired formulas.
\end{proof}
Proposition \ref{Polynomialseig} gives us the E-polynomials of all the moduli spaces where the quotient is geometric dividing by the E-polynomials of the respective stabilizers. We obtain
\begin{thm}
For all $g\geq 1$,
\begin{align*}
e(\mathcal{M}_{-\Id}) = & \, e_{1}^{g}/(q^3-q) = (q^3-q)^{2g-2}+(q^2-1)^{2g-2}-2^{2g-1}(q^{2}+q)^{2g-2}+(2^{2g-1}-1)(q^{2}-q)^{2g-2} \\
e(\mathcal{M}_{J_{+}}) = & \, e_{2}^{g}/q = (q^{3}-q)^{2g-2}(q^{2}-1)+(2^{2g-1}-1)(q-1)(q^{2}-q)^{2g-2}\\
& \qquad \quad -2^{2g-1}(q+1)(q^{2}+q)^{2g-2}+\frac{1}{2}q^{2g-2}(q-1)\left((q-1)^{2g-1}-(q+1)^{2g-1} \right) \\
e(\mathcal{M}_{J_{-}}) =& \, e_{3}^{g}/q = (q^{3}-q)^{2g-2}(q^{2}-1)+(2^{2g-1}-1)(q-1)(q^{2}-q)^{2g-2}+2^{2g-1}(q+1)(q^{2}+q)^{2g-2} \\
e(\mathcal{M}_{\xi_{\lambda}}) = & \, e_{4,\xi_{\lambda}}^{g}/q-1 = (q^{3}-q)^{2g-2}(q^{2}+q)+(q^{2}-1)^{2g-2}(q+1)+(2^{2g}-2)(q^{2}-q)^{2g-2}q.
\end{align*}
\end{thm}
Note that $e(\mathcal{M}_{-\Id}^{g})$ agrees with the result obtained by arithmetic methods in \cite{mereb:2010}.
\begin{cor} \label{cor:hodgemonoparabolic}
For $g\geq 1$, the behaviour of the E-polynomial of the parabolic character variety $\mathcal{M}_{\xi_{\lambda}}^{g}$ is given by
$$
R(\mathcal{M}_{\xi_{\lambda}})=
\left( (q^{3}-q)^{2g-2}(q^{2}+q)+(q+1)(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2} \right) T
+ \left( (2^{2g}-1)q(q^{2}-q)^{2g-2} \right) N.
$$
\end{cor}
\begin{proof}
From Proposition \ref{Polynomialseig} we get that
\begin{align*}
R(\overline{X}{}_{4}^{g})& =(a_{g}+d_{g})T+(b_{g}+c_{g})N \\
& = ((q^{3}-q)^{2g-1}+(q^{2}-1)^{2g-1}-(q^{2}-q)^{2g-1})T + ((2^{2g}-1)(q^{2}-q)^{2g-1})N.
\end{align*}
The result is obtained dividing by $e(\Stab(\xi_{\lambda}))=q-1$.
\end{proof}
To complete the proof of Theorem \ref{thm:main}, it remains the following
\begin{thm} \label{thm:id}
For all $g\geq 1$, we have
\begin{align*}
e(\mathcal{M}_{\Id})= & \, (q^{3}-q)^{2g-2}+(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2}-2^{2g}q^{2g-2} \\
& +\frac{1}{2}q^{2g-2}(q+2^{2g}-1)((q+1)^{2g-2}+(q-1)^{2g-2}) + \frac{1}{2}q((q+1)^{2g-1}+(q-1)^{2g-1}).
\end{align*}
\end{thm}
\begin{proof}
We need to distinguish between reducible and irreducible orbit since we have to take a GIT quotient to compute $e(\mathcal{M}_{\Id})$ and identify those orbits whose closures intersect.
We follow the method described in \cite{mamu} and compute the reducible locus. The E-polynomial of the irreducible locus is obtained
by sustracting the contribution of the reducible part from the E-polynomial of the total space $e_{0}^{g}$.
A reducible representation given by $(A_{1},B_{1},A_{2},B_{2},\ldots,A_{g},B_{g}) \in \SL(2,\mathbb{C})^{2g}$ is S-equivalent to
\begin{equation} \label{eqn:redrepres}
\left( \begin{pmatrix} \lambda_{1} & 0 \\ 0 & \lambda_{1}^{-1} \end{pmatrix}, \begin{pmatrix} \lambda_{2} & 0 \\ 0 & \lambda_{2}^{-1} \end{pmatrix}, \ldots, \begin{pmatrix} \lambda_{2g} & 0 \\ 0 & \lambda_{2g}^{-1} \end{pmatrix} \right)
\end{equation}
under the $\ZZ_2$-action $(\lambda_{1},\lambda_{2},\ldots, \lambda_{2g}) \sim (\lambda_{1}^{-1}, \lambda_{2}^{-1}, \ldots, \lambda_{2g}^{-1})$.
We have that $e(\CC^*)^{+}=q, e(\CC^*)^{-}=-1$, so
\begin{align*}
e(\mathcal{M}_{\Id}^{red}) & = e((\CC^*)^{2g}/\ZZ_2) \\
& = (e(\CC^*)^{+})^{2g}+ \binom{2g}{2}(e(\CC^*)^{+})^{2g-2}(e(\CC^*)^{-})^{2}+\ldots + \binom{2g}{2g-2}(e(\CC^*)^{+})^{2}(e(\CC^*)^{-})^{2g-2}+ (e(\CC^*)^{-})^{2g} \\
& = \frac{1}{2} \left( (q-1)^{2g} +(q+1)^{2g} \right).
\end{align*}
A reducible representation occurs if there is a common eigenvector. With respect to a suitable basis, the representation takes the form
$$
\left( \begin{pmatrix} \lambda_{1} & a_{1} \\ 0 & \lambda_{1}^{-1} \end{pmatrix}, \begin{pmatrix} \lambda_{2} & a_{2} \\ 0 & \lambda_{2}^{-1} \end{pmatrix},
\ldots, \begin{pmatrix} \lambda_{2g} & a_{2g} \\ 0 & \lambda_{2g}^{-1} \end{pmatrix} \right),
$$
which is a set parametrized by $(\CC^* \times \CC)^{2g}$. The condition $\prod_{i=1}^{g}[A_{i},B_{i}]=\Id$ is rewritten as
\begin{equation} \label{eqn:redlocusrelation}
\sum_{i=1}^{g} \lambda_{2i}(\lambda_{2i-1}^{2}-1)a_{2i} - \lambda_{2i-1}(\lambda_{2i}^{2}-1)a_{2i-1} = 0.
\end{equation}
There are four cases:
\begin{itemize}
\item $R_{1}$, given by $(a_{1},a_{2},\ldots,a_{2g}) \in \langle (\lambda_{1}-\lambda_{1}^{-1},\lambda_{2}-\lambda_{2}^{-1}, \ldots, \lambda_{2g}-\lambda_{2g}^{-1}) \rangle$,
and $(\lambda_{1},\lambda_{2}, \ldots, \lambda_{2g}) \neq (\pm 1, \pm 1, \ldots, \pm 1)$. In this case we can conjugate the representation to the diagonal form
(\ref{eqn:redrepres}) and assume that $a_{i}=0$.
The stabilizer of this stratum is the set of diagonal matrices $D \subset \PGL(2,\mathbb{C})$. Writing $A:=(\CC^*)^{2g}- \lbrace (\pm 1, \pm 1, \ldots, \pm 1) \rbrace$,
the stratum is isomorphic to $(A\times \PGL(2,\mathbb{C})/D) / \ZZ_2$, where the $\ZZ_2$-action is given by the permutation of the two basis vectors.
Since $e(\PGL(2,\mathbb{C})/D)^{+}=q^{2}, e(\PGL(2,\mathbb{C})/D)^{-}=q$ and
\begin{align*}
e(A)^{+} & = \frac{1}{2}\left((q-1)^{2g}+(q+1)^{2g} \right)-2^{2g} \\
e(A)^{-} & = e(A)-e(A)^{+}= \frac{1}{2}\left( (q-1)^{2g}-(q+1)^{2g} \right),
\end{align*}
we obtain
\begin{align*}
e(R_{1}) & = e(\PGL(2,\mathbb{C})/D)^{+}e(A)^{+}+e(\PGL(2,\mathbb{C})/D)^{-}e(A)^{-} \\
& = q^2 \left( \frac{1}{2}((q-1)^{2g}+(q+1)^{2g})-2^{2g}\right) +q\left( \frac{1}{2} ((q-1)^{2g}-(q+1)^{2g}) \right) \\
& = (q^3-q)\frac{1}{2}\left( (q-1)^{2g-1}+(q+1)^{2g-1} \right)-2^{2g}q^2.
\end{align*}
\item $R_{2}$, given by $(a_{1},a_{2},\ldots,a_{2g}) \not \in \langle (\lambda_{1}-\lambda_{1}^{-1},\lambda_{2}-\lambda_{2}^{-1}, \ldots, \lambda_{2g}-\lambda_{2g}^{-1}) \rangle$,
and $(\lambda_{1},\lambda_{2}, \ldots, \lambda_{2g}) \neq (\pm 1, \pm 1, \ldots, \pm 1)$. Equation (\ref{eqn:redlocusrelation}) defines a hyperplane $H\subset \CC^{2g}$
and the condition for $(a_{1},a_{2},\ldots, a_{2g})$ defines a line $l \subset H$. Writing $U'\cong D\times U$ for the upper triangular matrices,
we have a surjective map $A \times (H - l) \times \PGL(2,\mathbb{C}) \longrightarrow R_{2}$ with fiber isomorphic to $U'$. Hence
\begin{align*}
e(R_{2}) & = ((q-1)^{2g}-2^{2g})(q^{2g-1}-q)(q^{3}-q)/(q^{2}-q) \\
& = (q+1)(q^{2g-1}-q)((q-1)^{2g}-2^{2g}).
\end{align*}
\item $R_{3}$, given by $(a_{1},a_{2},\ldots,a_{2g})=(0,0,\ldots, 0)$, and $(\lambda_{1},\lambda_{2}, \ldots, \lambda_{2g}) = (\pm 1, \pm 1, \ldots, \pm 1)$, corresponding to the case where $A_{i}=B_{i}=\pm \Id$. The stratum consists of $2^{2g}$ points, so
$$
e(R_{3})=2^{2g}.
$$
\item $R_{4}$, given by $(a_{1},a_{2},\ldots,a_{2g}) \neq (0,0, \ldots , 0)$, and $(\lambda_{1},\lambda_{2}, \ldots, \lambda_{2g}) = (\pm 1, \pm 1, \ldots, \pm 1)$. In this case, there is at least a matrix of Jordan type, so the diagonal matrices $D$ act projectivizing the set $(a_{1},a_{2},\ldots,a_{2g})\in \CC^{2g}-\lbrace (0,0,\ldots,0) \rbrace$. The stabilizer is isomorphic to $U$. Therefore
\begin{align*}
e(R_{4}) & = 2^{2g}e(\mathbb{P}^{2g-1})e(\PGL(2,\mathbb{C})/U) \\
& = 2^{2g}(q^{2g-1}+q^{2g-2}+\ldots + 1)(q^{2}-1) \\
& = 2^{2g}(q^{2g}-1)(q+1).
\end{align*}
\end{itemize}
The total E-polynomial of the reducible locus $R$ is thus
\begin{align*}
e(R) & = e(R_{1})+e(R_{2})+e(R_{3})+e(R_{4}) \\
& =(q^3-q)\left( \frac{1}{2}((q+1)^{2g-1}-(q-1)^{2g-1})+2^{2g}q^{2g-2}+(q-1)(q^{2}-q)^{2g-2} \right).
\end{align*}
We obtain the E-polynomial of the irreducible part as
\begin{align*}
e(I) = & \, e_{0}^{g}-e(R) \\
= &\, (q^3-q) \left( (q^3-q)^{2g-2}+(q^{2}-1)^{2g-2}-(q^{2}-q)^{2g-2}+\frac{1}{2}q^{2g-2}(q+2^{2g}-1)((q+1)^{2g-2}+(q-1)^{2g-2}) \right. \\
& \left. -2^{2g}q^{2g-2}-(q-1)(q^{2}-q)^{2g-2}-\frac{1}{2}((q+1)^{2g-1}-(q-1)^{2g-1}) \right),
\end{align*}
and
\begin{align*}
e(\mathcal{M}_{\Id}^{irr}) =&\, e(I)/(q^{3}-q) \\
= &\, (q^{3}-q)^{2g-2}+(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2}-2^{2g}q^{2g-2} \\
& +\frac{1}{2}q^{2g-2}(q+2^{2g}-1)((q+1)^{2g-2}+(q-1)^{2g-2}) - \frac{1}{2}((q+1)^{2g-1}-(q-1)^{2g-1}).
\end{align*}
Finally,
\begin{align*}
e(\mathcal{M}_{\Id}) =&\, e(\mathcal{M}_{\Id}^{irr})+e(\mathcal{M}_{\Id}^{red}) \\
=&\, (q^{3}-q)^{2g-2}+(q^{2}-1)^{2g-2}-q(q^{2}-q)^{2g-2}-2^{2g}q^{2g-2} \\
& +\frac{1}{2}q^{2g-2}(q+2^{2g}-1)((q+1)^{2g-2}+(q-1)^{2g-2}) + \frac{1}{2}q((q+1)^{2g-1}+(q-1)^{2g-1}).
\end{align*}
\end{proof}
\section{Topological consequences} \label{sec:topol}
In this section, we extract some information from the formulas in Theorem \ref{thm:main}.
We start by a proof of Theorem \ref{thm:balanced}.
\noindent{\textit{Proof of Theorem \ref{thm:balanced}.\/}
In the case of $\cM_{-\Id}, \cM_{J_\pm}$ and $\cM_{\xi_\lambda}$,
the result follows readily from Proposition \ref{prop:balanced}. For instance, for $\cM_{-\Id}$ we
have that $\cM_{-\Id}=\overline{X}{}_1/ \PGL(2,\CC)$, where $\overline{X}{}_1$ and $G=\PGL(2, \CC)$ are of balanced type.
Hence the classifying space $BG$ is also of balanced type and there is a homotopy fibration
$\overline{X}{}_1 \to \overline{X}{}_1/G \to BG$. The Leray spectral sequence gives that $\overline{X}{}_1/G$ must be of balanced
type (a similar argument appears in the proof of Proposition 7.2 of \cite{Mu-e}).
In the case of $\cM_{\Id}$, the description in Theorem \ref{thm:id} yields that
$R$ is of balanced type. Hence $I$ is also of balanced type. The same argument as above
proves that $\cM_{\Id}^{irr}=I/\PGL(2,\CC)$ is of balanced type. As $\cM_{\Id}^{red}$ is
clearly of balanced type, so is $\cM_{\Id}$.
\hfill $\Box$
\medskip
\begin{cor}
Let $X$ be a complex curve of genus $g\geq 2$. The Euler characteristic of $\cM_C=\cM_C(\SL(2,\CC))$ is given by
\begin{align*}
&\chi(\mathcal{M}_{\Id})= 2^{4g-3}-3\cdot 2^{2g-2} \\
&\chi(\mathcal{M}_{-\Id}) = -2^{4g-3} \\
&\chi(\mathcal{M}_{J_{+}}) = -2^{4g-2} \\
&\chi(\mathcal{M}_{J_{-}}) =2^ {4g-2}\\
&\chi(\mathcal{M}_{\xi_{\lambda}}) = 0.
\end{align*}
\end{cor}
\begin{proof}
The Euler characteristic is obtained by setting $q=1$ in $e(\cM_C)$ given in Theorem \ref{thm:main}.
\end{proof}
\begin{cor}
Let $X$ be a complex curve of genus $g\geq 2$. Then $\cM_{\Id}$ and $\cM_{-\Id}$ are of dimension $6g-6$ and
$\cM_{J_+}$, $\cM_{J_-}$ and $\cM_{\xi_\lambda}$ are of dimension $6g-4$. All of them have a unique component of
maximal dimension.
\end{cor}
\begin{proof}
From Theorem \ref{thm:main}, we get
\begin{align*}
&e(\mathcal{M}_{\Id})= q^{6g-6}+ \ldots + 1 \\
&e(\mathcal{M}_{-\Id}) = q^{6g-6}+ \ldots + 1\\
&e(\mathcal{M}_{J_{+}}) = q^{6g-4}+ \ldots +(1-2^{2g-1})q^{2g-2} \\
&e(\mathcal{M}_{J_{-}}) = q^{6g-4}+ \ldots +(2^{2g}-1) q^{2g-1}\\
&e(\mathcal{M}_{\xi_{\lambda}}) = q^{6g-4}+ \ldots + 1
\end{align*}
where we have written the monomials of maximum and minimum degrees in each case.
The degree of the polynomial gives the dimension of the character variety, and the coefficient (which is always $1$)
gives the number of irreducible components.
\end{proof}
\begin{cor}
Let $X$ be a complex curve of genus $g\geq 1$. Then $e(\mathcal{M}_{-\Id})$, $e(\mathcal{M}_{\xi_{\lambda}})$,
and its invariant and non-invariant part given in Cororally \ref{cor:hodgemonoparabolic}, are palindromic polynomials.
\end{cor}
\begin{proof}
Let $d=6g-4$. If we write $R(\mathcal{M}_{\xi_{\lambda}})=AT+BN$, with
\begin{align*}
A & = (q^3-q)^{2g-2}(q^2+q)+(q+1)(q^2-1)^{2g-2}-q(q^{2}-q)^{2g-2}, \\
B & = (2^{2g}-1)q(q^{2}-q)^{2g-2},
\end{align*}
given in Corollary \ref{cor:hodgemonoparabolic}, then one only has to check that $q^dA(q^{-1})= A(q)$ and $q^d B(q^{-1})=B(q)$, which is straightforward.
The computation for $e(\mathcal{M}_{-\Id})$ is analogous and it is also given in \cite[Section 4.4]{mereb:2010}.
\end{proof}
|
\section{Introduction}
The use of the ROS middleware is a growing trend in
robotics in general, in particular in experimental branches of robotics
such as modular robotics, fields robotics, and the vast area of cyber-physical systems (for example applied to welfare technology).
Our main area of interest is in experimental robotics and cyber-physical systems. When building ``robot controllers'' for the
aforementioned systems there are numerous suitable technological platforms.
Given specific requirements we can choose an appropriate standardized
approach, for example emphasizing flexibility and ease of development
by using a generic middleware --- such as ROS --- or emphasizing real-time
performance and direct hardware access by using approaches based on
dedicated, embedded hardware. So far ROS and hard real-time
embedded systems have however not been easily uniteable while retaining the same overall communication and processing methodology at all levels.
In this paper we present an approach aimed at tackling the schism
between high-level, flexible software and low-level, real-time
software. The key idea of our approach is to enable software
components written for a high-level publish-subscribe software
architecture to be automatically migrated to a dedicated hardware
architecture implemented using programmable logic. Our approach is
based on the Unity framework, a unified software/hardware framework
based on FPGAs for quickly interfacing high-level software to low-level
robotics hardware. The vision of Unity is to enable non-expert
users to build high-quality interface and control systems using FPGAs
and to interface them to high-level software frameworks, thereby
providing a framework for speeding up and increasing innovation in
experimental robotics. This paper presents the overall vision and the
initial work on the implementation of an architecture supporting a
generative approach, based on a declarative specification of how
software components are mapped to a hardware architecture; the actual
language design is left as future work.
\section{Context: Unity and FPGAs}
The traditional approach to building a control system in experimental
robotics is mainly based on microcontrollers (MCU's) and PC's. This
approach has numerous advantages, mainly: (1)~developers are familiar
with the programming methodology; (2)~good tools, libraries and
frameworks from commercial vendors and the open-source community; and
(3)~the availability of cheap and simple MCU-based systems like the
Arduino, as well as more powerfull ARM based systems. Despite the advantages of this approach, there are also
inherent limitations to the sequential-style processing and fixed
hardware (HW) architecture, which can significantly limit reuse of HW
as well as real-time capabilities, design freedom and flexibility.
We prefer FPGAs and hybrid FPGA-MCU SoC systems over pure MCUs: we
find FPGAs superior to MCUs in many performance areas relevant to
experimental robotics, except for price and library support. FPGAs can
provide deterministic hard real-time performance no matter the
complexity or scale of the implemented algorithms~\cite{fernandes,
pordel, toscher}. On an FPGA the architecture is designed by the
developer, providing increased flexibility that can reduce the need
for costly software abstractions on higher levels \cite{falsig1,
falsig4, falsig3, falsig2} and reduce or eliminate the need for
external support logic. FPGAs are however not commonly used; we
believe the reason to be partly historical: people stick to
technologies they know. Moreover, FPGAs suffer from a lack of good,
open-source, vendor-independent HDL-component libraries suited for
robotics, and a high degree of complexity associated with FPGA
programming, caused partly by complex tools and a different
programming methodology compared to the traditional Von-Neumann style.
\begin{figure*}[t]
\begin{center}
\center\includegraphics[width=15.3cm]{traditional_vs_unity.pdf}
\caption{Unity Link compared to a traditional MCU-based architecture (UML 2.0 component diagram notation)}
\label{fig:unity-vs-mcu}
\end{center}
\end{figure*}
We have proposed the Unity framework as a means to facilitating
FPGA-based development for experimental
robotics~\cite{sdir8,iros13}. Unity is an open-source framework
consisting of reference HW designs, gateware (GW, VHDL) and SW
libraries, all targeted at providing a complete framework for easy
development, with standard cases covered by model-based code
generation of all the necessary FPGA GW and PC SW needed to interface
electronics with a high-level software framework. The Unity framework
is a work-in-progress: The modular HW designs include single nodes,
distributed nodes, sensor interfaces and generic motor controllers. On
the GW side we have a growing library of VHDL modules, including
servo- and brushless DC motor controllers, a real-time network based
on a shared memory model, a complete FPGA-based real-time operating
system~\cite{hartos}, as well as a modular and reconfigurable FPGA-PC
interface called Unity-Link~\cite{iros13}. The use of Unity compared
to a traditional MCU-based approach, exemplified with a PC connected
to low-level hardware using ROS-serial, is illustrated in
Figure~\ref{fig:unity-vs-mcu}.
We believe that a generic FPGA or FPGA-MCU SoC based module will be
more flexible and therefore more easily reused for various tasks,
compared to a standard off-the-shelf MCU system, since the various
hardware interfaces needed are decoupled from (i.e., not locked to)
specific pin locations, and therefore virtually only the pin count
limits the number and types of interfaces that are possible when using
programmable logic. Unity is an evolution of the
TosNet framework, which is the basis for the real-time network and
other specific components~\cite{falsig1,falsig4,falsig3,falsig2}.
\section{Automatic Migration of ROS Components to FPGAs}
\begin{figure*}[t]
\begin{center}
\center\includegraphics[width=14.8cm]{migration-before.png}
\caption{Example: Robot software before migration of selected
components, part of the control loop is implemented in ROS. (UML 2.0)}
\label{fig:migration-before}
\end{center}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\center\includegraphics[width=14.8cm]{migration-after.png}
\caption{Example: Robot software after migration of selected
components, the ``maintain position'' and ``filter'' components have
been migrated to the FPGA. (UML 2.0)}
\label{fig:migration-after}
\end{center}
\end{figure*}
We are currently investigating the idea of automatically migrating
networks of ROS components\footnote{Throughout this paper we
consistently use the term ``ROS component'' to refer to ROS nodes:
we believe our approach is applicable to other component-based
middlewares as well, and hence prefer the technology-independent
term ``component.''} to our FPGA-based architecture. The Unity
framework already provides a standardized platform on which gateware
components can be interconnected, and Unity-Link provides automatic
integration on ROS components with gateware components using a
publish-subscribe infrastructure~\cite{iros13}. There is however no
support for migrating a ROS component, or a set of ROS components,
from the PC to the FPGA, without completely reimplementing the
functionality of each of the components, and furthermore using
the Unity framework to connect them internally on the FPGA. Note that
we are not concerned with dynamic migration: we simply want to make it
easy for the developer to statically change the deployment of functionality
between the flexible PC platform and the real-time FPGA platform.
We propose that migration of a given ROS component from the PC-based
platform to the FPGA can be done by recompiling the component to run
on either a softcore or hard-IP CPU embedded in the FPGA. The HartOS real-time operating system~\cite{hartos}
will be used to execute the threads of the component and to handle
external events. A substrate that provides the ROS API and a few
selected parts of the standard POSIX API\footnote{Only a
small subset of the POSIX API will be relevant, as well as feasible for a processing system utilizing the HartOS kernel. We assume our
approach is primarily relevant for ROS components having a fairly
small amount of interaction with the operating system.} will be used
on the embedded CPU, enabling a ROS component e.g.\ implemented in
C$++$ to execute on the CPU after a simple recompilation.
Publish-subscribe messages can be routed between the CPU and a PC running ROS using Unity-Link. A high performance hard-IP CPU, like e.g.\ the dual-core ARM-A9 in a Xilinx Zynq device, could as a second option also run a full linux system with ROS, and thereby support native (non-recompiled) ROS components. By providing the same memory-mapped publish/subscribe and service-call IP interfaces on both the small softcore and Hard-IP CPU's, no matter the software environment executed on them, Unity will allow both high and low performance processors, and PC's using Unity-link, to communicate with GW components directly utilizing ROS' own communication paradigm, thereby enabling easy migration between
execution paradigms.
A set of ROS components that communicate using publish-subscribe can
similarly be migrated to the FPGA. Each component is placed on a
softcore CPU, depending on the performance requirements they can be
placed on the same or different CPUs. If they are placed on the same
CPU, HartOS is used for scheduling CPU-time between the components,
and communication can be performed directly between the components
(taking care to preserve communication semantics). If components are
placed on different CPUs, a shared memory component is used
to propagate publish/subscribe messages between the nodes: each topic
uses a specific address in the shared memory, enabling a complete
decoupling of the execution of publishers and subscribers. Service
calls can be handled similarly, however rather than using a shared
memory, a generic address-data bus can be used to provide a
point-to-point connection between components that need to communicate.
As an example, consider a first revision of the robot software
architecture for a two-wheeled balancing robot shown in
Fig.~\ref{fig:migration-before}. Low-level control and hardware
interfacing is done in the FPGA using the Unity framework, and
consists of low-level hardware interface components and a generic PID
controller. Unity-Link connects low-level control and sensor
interfaces to ROS using publish/subscribe and service calls.
High-level control is implemented in ROS, and concerns navigation,
movement, and balancing of the robot. Real-time operation of the
``maintain position'' component is ensured by using a suitably fast
PC. Now assume that --- although initial experiments showed that this
worked fine --- after experimenting with the robot in a realistic
scenario it is found that control is unstable because real-time
deadlines are sometimes missed. To solve the issue using our
approach, the ``maintain position'' component is moved to a softcore
CPU on the FPGA, as illustrated in Fig~\ref{fig:migration-after}.
Moreover, due to the use of standard interfaces that are interoperable
between ROS and the FPGA, the software filter component is
transparently replaced by a functionally equivalent gateware component from the Unity library. All components on the FPGA execute in hard real-time, making control of the robot predictable.
\section{Discussion and Status}
The migration is intended to be automatic, in the sense that given a
declarative specification of how a set of ROS components should be
mapped to a real-time architecture, our system will generate substrate
code, configuration files, and VHDL components such that the ROS
components can be directly recompiled to run on the FPGA. This
declarative specification will thus need to model which components are
to be deployed to which softcore CPU, how much time is to be assigned
to each thread of each component, and how the components are to
communicate with each other and with the rest of the ROS system.
We are currently extending the Unity framework to support multiple ROS
components executing and communicating in real-time on one or more
softcore CPUs on one or more FPGAs connected by a real-time network. Once
this framework is complete, we will augment it with a model-based code
generator than can automatically generate the complete set of code
artifacts needed to support the execution of the ROS components on the
FPGA. We expect that the task of implementing the framework and
corresponding generator is significantly reduced by building on top of
the standardized Unity architecture and using Unity Link to interface
the FPGA-based ROS components to the rest of the ROS system.
|
\section{Introduction}
Early-type galaxies (ETGs) are often considered as the final stage of galaxy evolution, with virtually little or no cold interstellar medium left. Recently, more and more evidence is appearing that a substantial fraction of the ETGs do contain substantial reservoirs of cold interstellar matter.
Systematic searches for atomic and molecular gas in ETGs have yielded widely varying detection rates, corresponding to different selection criteria and sensitivity limits \citep[e.g.,][]{2007MNRAS.377.1795C, 2009A&A...498..407G, 2010MNRAS.409..500O, 2010ApJ...725..100W}. The most complete effort to make a census of cold interstellar gas in ETGs in the Local Universe is the ATLAS$^{\text{3D}}$ project \citep{2011MNRAS.413..813C}, which targets a volume-limited sample of 260 ETGs. Molecular gas was detected in 22\% of the galaxies in the sample \citep{2011MNRAS.414..940Y}, whereas the atomic gas detection rate depended strongly on the environment, and increased from 10\% in the Virgo Cluster to 40\% for field ETGs \citep{2012MNRAS.422.1835S}. The frequently observed kinematic misalignments between stars and gas suggests an external origin for the gas in many ETGs \citep[e.g.][]{1993ApJ...419..544S, 2001AJ....121..140K, 2006MNRAS.366.1151S,2008ApJ...676..317Y, 2011MNRAS.417..882D}. Similarly, there is now plenty of evidence for the presence of cold interstellar dust in a significant fraction of the ETG population. Deep optical imaging surveys have shown that many ETGs possess dust features in a variety of morphological forms \citep{1994A&AS..105..341G, 1995AJ....110.2027V, 2001AJ....121.2928T, 2001AJ....121..808C, 2006ApJS..164..334F, 2014NewA...30...51K}.
In a small fraction of the ETG population, the dust is organized in prominent and large-scale dust lanes \citep{2007A&A...461..103P, 2008MNRAS.390..969F, 2010MNRAS.409..727F}. These so-called dust-lane ETGs are considered to be the remnants of recent minor mergers between ETGs and gas-rich satellites \citep{1981MNRAS.196..747H, 2012MNRAS.423...49K, 2013MNRAS.435.1463K, 2012MNRAS.423...59S}. Combining new ATCA observations and archival data, \citet{2002AJ....123..729O} studied the H{\sc{i}} properties of a sample of 9 dust-lane ETGs. They found huge H{\sc{i}} reservoirs in 6 galaxies, with masses up to several times 10$^9~M_\odot$ and H{\sc{i}}-to-dust mass ratios of more than 1000. On the other side of the spectrum, they did not detect H{\sc{i}} in the remaining 3 dust-lane ETGs, which implied unusually low H{\sc{i}}-to-dust mass ratios, down to $M_{\text{H{\sc{i}}}}/M_{\text{d}} < 20$. They suggested that the cold interstellar medium should mainly be in molecular rather than atomic form in these galaxies.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{NGC5485-images-1.pdf}\\[0.5ex]
\includegraphics[width=\columnwidth]{NGC5485-images-2.pdf}
\caption{SDSS $g$ band, PACS 100 and 160 $\mu$m, and SPIRE 250 $\mu$m images of NGC\,5485. The field of view of each image is $2\times2$ arcmin$^2$, and the beam FWHM is indicated in the three Herschel images. In the optical image, a shell-like feature is clearly visible some 20 arcsec towards the SE of the nucleus, and its counterpart is detectable in the PACS images as well.}
\label{NGC5485-images.pdf}
\end{figure}
In this Letter we focus on NGC\,5485, a dust-lane ETG at a distance of 25.2 Mpc\footnote{Throughout this Letter we use this distance for NGC\,5485; all luminosities and masses are converted to be consistent with this value.} \citep{2001ApJ...546..681T, 2011MNRAS.413..813C}. It is home to a prominent dust lane, perpendicular to the photometric major axis (Fig.~{\ref{NGC5485-images.pdf}}, top left). Given the prominent dust lane, one would expect NGC\,5485 to contain a large cold interstellar matter reservoir. The galaxy is detected by IRAS, with a corresponding dust mass of about $10^6~M_\odot$ \citep{2007A&A...461..103P, 2010MNRAS.407.2475F}. Similarly to a third of the dust-lane ETGs from the \citet{2002AJ....123..729O} sample, NGC\,5485 has a low H{\sc{i}} content: in the frame of the ATLAS$^{\text{3D}}$ campaign, \citet{2012MNRAS.422.1835S} observed the galaxy at 21~cm with the WSRT, but did not detect it. The corresponding upper limit to the atomic gas mass is $M_{\text{H{\sc{i}}}} < 1.5\times10^7~M_\odot$. Surprisingly, also no molecular gas is detected: \citet{2011MNRAS.414..940Y} observed NGC\,5485 in the CO(1-0) and CO(2-1) lines with the IRAM 30m telescope, but did not detect it in either line. The resulting upper limit on the molecular gas mass is $M_{\text{H}_2} < 4.0\times10^7~M_\odot$.
The combination of these gas and dust mass estimates results in a cold gas-to-dust ratio lower than about 50. This value is probably still strongly overestimated, as IRAS was insensitive to cool dust. In this Letter, we present new far-infrared and submm imaging data for NGC\,5485, taken with the PACS and SPIRE instruments onboard the Herschel Space Observatory. In \S2 we present our observations and the resulting flux densities, and we combine these new data with ancillary data to make a solid measurement of the dust mass, and hence an accurate upper limit on the gas-to-dust ratio. In \S3 we discuss different scenarios to explain this gas-to-dust ratio upper limit, including a potential overestimate of the dust mass, and underestimate of the cold gas mass, and a number of possible alternative options. Finally, \S4 summarizes our conclusions.
\section{Observations and dust mass determination}
NGC\,5485 was observed with the Herschel Space Observatory \citep{2010A&A...518L...1P} in the frame of the {\it{Far-infraRed Investigation of Early-type galaxies with Dust Lanes}} (FRIEDL) program. We used the PACS \citep{2010A&A...518L...2P} and SPIRE \citep{2010A&A...518L...3G} photometers in scan mode, both with their nominal scan speed, i.e. 20 arcsec/s for PACS and 30 arcsec/s for SPIRE. The size of the map was chosen to be $8\times8$~arcmin$^2$. For the PACS map, four cross-scans (i.e. four nominal and four orthogonal scans) were performed, while the SPIRE map was observed with a single cross-scan. The data reduction was done with HIPE version 12.0.0 and includes the same steps as in \citet{2013A&A...556A..54V}. For the PACS data, it includes the use of the Scanamorphos version 23 \citep{2013PASP..125.1126R} in order to make optimal use of the redundancy in the observational data. Figure~{\ref{NGC5485-images.pdf}} shows the resulting Herschel maps at 100, 160 and 250 $\mu$m. At these wavelengths, the FWHM is 7, 11 and 18 arcsec respectively, and the dust emission from the galaxy is resolved along the direction of the dust lane. At 350 and 500 $\mu$m, where the FWHM increases to 24 and 36 arcsec, the emission is not resolved anymore.
Global flux densities were determined through aperture photometry using the DS9/Funtools program FUNCTS, following the same strategy as described in detail by \citet{2013A&A...556A..54V}. We followed the approach of \citet{2012ApJ...745...95D} to determine the background level and uncertainty estimates. The Herschel flux densities are $533\pm 44$, $797\pm52$, $518\pm46$, $288\pm33$ and $108\pm23$ mJy at 100, 160, 250, 350 and 500 $\mu$m.
To determine the dust mass, we first used a simple modified blackbody model, with dust emissivity index $\beta=2$ and $\kappa_\nu=0.192$~m$^2$\,kg$^{-1}$ at 350~$\mu$m. These values correspond to what is probably the most widely adopted physical dust model \citep{2003ARA&A..41..241D, 2007ApJ...657..810D}, and they have been widely used to interpret Herschel SEDs \citep[e.g.,][]{2012ApJ...745...95D, 2013MNRAS.428.1880A, 2013A&A...556A..54V, 2014A&A...565A...4H}. Applying such a simple modified blackbody fit to the PACS and SPIRE flux densities, we find $T_{\text{d}} = 18.8\pm0.4$~K and $M_{\text{d}} = (3.79\pm0.31)\times10^6~M_\odot$.
Apart from a simple modified blackbody fit, we also used the MAGPHYS code \citep{2008MNRAS.388.1595D} to determine the dust mass. MAGPHYS has been used extensively to model the panchromatic SEDs of galaxies \citep[e.g.,][]{2012MNRAS.427..703S, 2013ApJ...768...90L, 2013MNRAS.433..695C, 2014A&A...567A..71V}. We combined the Herschel PACS and SPIRE flux densities with optical {\it{ugriz}} flux densities for NGC\,5485 from SDSS, near-infrared {\it{JHK}}$_{\text{s}}$ flux densities from 2MASS, mid-infrared flux densities from WISE, and far-infrared flux densities at 60 and 100 $\mu$m from IRAS. The result of the MAGPHYS fit can be found in Figure~{\ref{NGC5485-MagPhys.pdf}}. The most important property for our goals is the total dust mass, for which we found $M_{\text{d}} = (3.81^{+0.80}_{-0.64})\times10^6~M_\odot$, completely consistent with the value obtained from our modified blackbody fit (we adopted the same dust model).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{NGC5485-MagPhys.pdf}
\caption{A MAGPHYS fit to the UV-submm spectral energy distribution of NGC\,5485. The top panel shows the SED, where the red stars represent the SDSS, 2MASS, WISE, IRAS and Herschel data points, the blue solid line is the intrinsic stellar SED, and the solid black line is the model SED including dust extinction and emission. The bottom panel shows the residuals between the data and the model. The $\chi^2$ value is dominated by the IRAS 100 $\mu$m flux density, which is inconsistent with the PACS 100 $\mu$m flux density.}
\label{NGC5485-MagPhys.pdf}
\end{figure}
\section{Discussion}
Combining the dust mass resulting from the Herschel flux densities with the molecular and atomic gas measurements discussed in \S1, we find a cold gas-to-dust upper limit $M_{\text{gas}}/M_{\text{d}} < 14.5$. Such an gas-to-dust ratio is extremely low -- typical values range from several thousands for low-metallicity dwarf galaxies to about 100 for solar metallicity galaxies \citep[e.g.,][]{2009ApJ...701.1965M, 2011A&A...532A..56G, 2011A&A...535A..13M, 2014A&A...563A..31R}.
One possible explanation for the extreme gas-to-dust ratio upper limit for NGC\,5485 is that the dust mass as determined from the Herschel flux densities is overestimated. There are several possible systematic effects that we should investigate in some detail.
A first observation is that the new dust mass we have derived is a factor 3 to 4 larger than the IRAS-based dust masses of \citet{2007A&A...461..103P} and \citet{2010MNRAS.407.2475F}. And these dust masses are themselves more than an order higher than the dust masses the same authors derived using extinction in optical images, for which they found several times $10^4~M_\odot$. However, we are convinced that our new dust mass based on Herschel FIR/submm data is more reliable than those based on IRAS and optical extinction. IRAS-based dust masses are most certainly an underestimate of the true dust mass, since IRAS is insensitive to cool dust at temperature below about 25~K \citep[e.g.,][]{1995A&A...298..784G, 1996ApJ...468..571T}. Herschel observations have decisively shown that the typical cold dust temperature in ETGs is around 20~K \citep{2012ApJ...748..123S, 2012MNRAS.419.2545R, 2013A&A...552A...8D}. For NGC\,5485, we find an effective dust temperature of 19~K, which indeed indicates that IRAS missed the bulk of the dust. On the other hand, inferring the amount of dust from optical extinction maps is complicated due to the complexity of the star-dust geometry and the effects of scattering \citep[e.g.,][]{1992ApJ...393..611W, 2001MNRAS.326..733B}. In dust-lane ETGs, dust mass estimates based on simple extinction recipes applied to optical images can underestimate the true dust mass by an order of magnitude or more \citep{1995A&A...298..784G, 2007A&A...461..103P, 2008MNRAS.390..969F, 2010MNRAS.409..727F}.
We should also critically consider the possibility that the observed Herschel flux densities are overestimated. NGC\,5485 has a galactic latitude of 59$^\circ$, so there are no foreground subtraction issues due to strong cirrus that have plagued some other Herschel extragalactic observations \citep[e.g.,][]{2010MNRAS.409..102D, 2012A&A...546A..34F}. In optical and NIR images, NGC\,5485 seems to have a shell-like structure some 20 arcsec towards to SE of the nucleus (see left panel of Figure~{\ref{NGC5485-images.pdf}}). This feature has a clear counterpart in the PACS maps, and could be an edge-on background galaxy rather than a feature linked to NGC\,5485 itself. In any case, the contribution of this source to the Herschel flux densities of NGC\,5485 is marginal. In the PACS 100 $\mu$m map, where the resolution is sufficient to separate this feature from the dust emission in the dust lane, we find that it contributes no more than 10\% to the total flux density.
Finally, a last aspect to take into consideration is that we modeled the FIR/submm emission as a single modified blackbody at a single temperature, whereas in reality, the FIR/submm emission results from a complicated weighted sum of modified blackbodies at a range of temperatures. This is a systematic effect that we cannot avoid due to our lack of resolution. However, the total dust mass is typically {\em{underestimated}} when it is based on the integrated SED \citep{2011A&A...536A..88G, 2012MNRAS.425..763G}, so it does not help to explain the low gas-to-dust ratio.
On the other hand, the low gas-to-dust ratio might be the result of a systematic underestimate of the gas mass. In any case, there does not seem to be any reason to doubt the reliability of the H{\sc{i}} and CO upper limits of \citet{2012MNRAS.422.1835S} and \citet{2011MNRAS.414..940Y}, respectively. The noise levels obtained in both observation campaigns, with the WSRT and IRAM 30m telescope respectively, are almost exactly the median value of the entire ATLAS$^{\text{3D}}$ study. One source of uncertainty on the derivation of the molecular gas mass upper limit is the conversion factor needed to convert CO(1-0) intensities to the H$_2$ masses, the so-called $X_{\text{CO}}$ factor. This factor has been the subject of a substantial debate in the recent literature \citep[e.g.,][]{1996A&A...308L..21S, 1997A&A...328..471I, 2011ApJ...737...12L, 2012ApJ...756...40S, 2013ApJ...777....5S}. The authoritative review by \citet{2013ARA&A..51..207B} suggests the value $X_{\text{CO}} = 2\times10^{20}$ cm$^{-2}$ (K~km~s$^{-1}$)$^{-1}$ for the Milky Way and other "normal galaxies". In their determination of the upper limit for NGC\,5485, \citet{2011MNRAS.414..940Y} used $X_{\text{CO}} = 3\times10^{20}$ cm$^{-2}$ (K~km~s$^{-1}$)$^{-1}$, i.e.\ 50\% larger than the standard value suggested by \citet{2013ARA&A..51..207B}. Using the standard value would decrease the molecular gas mass upper limit by factor 1.5, and it would decrease the gas-to-dust ratio upper limit by a factor 1.32 to $M_{\text{gas}}/M_{\text{d}} < 11$.
In principle, most of the gas in NGC\,5485 could be ionized rather than neutral. An interesting comparison case here is the study of the elliptical galaxy NGC\,4125 by \citet{2013ApJ...776L..30W}. Combining Herschel-based dust estimates with H{\sc{i}} and CO non-detections, they also found a similar cold gas-to-dust ratio upper limit as we obtain here. However, based on a [N{\sc{ii}}]/[C{\sc{ii}}] ratio consistent with a low-density ionized medium and the (crude) assumption that the ionized gas is distributed uniformly over a 4.2 kpc diameter sphere, \citet{2013ApJ...776L..30W} argued that a significant fraction of the gas in NGC\,4125 is warm ionized rather than cold neutral gas. For NGC\,5485, this option seems excluded, though. The galaxy was imaged in H$\alpha$ by \citet{2010MNRAS.407.2475F}. They detected an inclined disc of ionized gas that nicely follows the morphology of the dust lane, which seems to support that at least part of the dust might be associated with warm ionized gas. The total amount of ionized hydrogen gas traced in the H$\alpha$ map is $M_{\text{H{\sc{ii}}}} = (2.0\pm0.6)\times10^4~M_\odot$, far too little to be a significant contributor to the total gas budget. \citet{2010MNRAS.407.2475F} even warn that their H{\sc{ii}} masses are likely overestimated by a factor 2-3 due to confusion of the H$\alpha$ and [N{\sc{ii}}] lines. On the other hand, H$\alpha$ photons are subject to dust attenuation in the dust lane, which could cause the H{\sc{ii}} mass in NGC\,5485 to be underestimated. However, in order to bring the H{\sc{ii}} mass to the level required for a typical gas-to-dust ratio of about 100, unrealistically high attenuation values $A_{\text{V}}\gg10$ are needed \citep{1994ApJ...429..582C}.
Alternatively, a major part of the gas budget in NGC\,5484 could be in the form of hot X-ray emitting ionized gas. While dust grains are expected to be destroyed through sputtering by thermal collisions with energetic ions, the coexistence of dust and a hot X-ray gas is not completely impossible \citep[e.g.,][]{2003ApJ...585L.121T, 2007PASJ...59..107K, 2008ApJ...677..249L}. NGC\,5485, however, was not detected in the ROSAT all-sky survey: \citet{1999MNRAS.302..209B} quote an upper limit of $L_{\text{X}} < 10^{39.8}~{\text{erg}}\,{\text{s}}^{-1}$. If we insert the K-band luminosity $L_{\text{K}} = 5.65\times10^{10}~L_\odot$ into the latest calibration between the K-band luminosity of ETGs and the expected unresolved X-ray emission due to low-mass X-ray binaries (LMXBs) by \citet{2011ApJ...729...12B}, we find exactly the same number, $L_{\text{X,LXMB}} = 10^{39.8}~{\text{erg}}\,{\text{s}}^{-1}$. This implies that any X-ray emission from NGC\,5485 most probably originates from LMXBs rather than from hot ionized gas. Moreover, the fact that the dust is distributed in a well-defined dust lane makes a physical association with a potential hot diffuse halo rather unlikely, although it is not impossible that a fraction of the dust is distributed diffusely over the galaxy \citep{1995A&A...298..784G}.
A final, more exotic, option could be that NGC\,5485 contains a reservoir of molecular gas, but that this gas is not emitting CO line emission. Both diffuse $\gamma$-ray emission \citep{2005Sci...307.1292G, 2010ApJ...710..133A, 2012ApJ...750....3A} and combined gas and dust observations \citep{2011A&A...536A..19P, 2012A&A...543A.103P} have revealed the presence of a substantial amount of so-called dark gas in our own Milky Way. The presence of CO-dark gas has been found many years ago in low-metallicity dwarf galaxies. At low metallicities, most of the carbon is locked up in neutral or singly ionized carbon rather than CO, which makes [C{\sc{ii}}] a more reliable tracer the cold H$_2$ reservoir \citep{1991ApJ...373..423S, 2010ApJ...724..957S, 1997ApJ...483..200M, 2013PASP..125..600M}. As a relatively massive lenticular galaxy, NGC\,5485 is not expected to contain copious amounts of CO-dark gas. Being a dust-lane ETG, however, it has most probably acquired most of its dust and gas during a recent minor merger \citep{2012MNRAS.423...49K}. The recent merger scenario is also supported by its rather exceptional kinematical structure, which shows strong minor-axis rotation \citep{1988A&A...195L...5W, 2011MNRAS.414.2923K, 2011MNRAS.414..888E}.
In principle, it is possible that it has accreted a dwarf galaxy with a substantial CO-dark molecular gas reservoir. An appropriate example of such a dwarf galaxy would be the SMC, which contains a dust mass of $3\times10^5~M_\odot$, but hardly any CO-emitting molecular gas \citep{2001PASJ...53L..45M, 2007ApJ...658.1027L, 2011ApJ...737...12L}. It does, however, contain about $4\times10^8~M_\odot$ of H{\sc{i}} gas \citep{1999MNRAS.302..417S, 2011ApJ...741...12B}, so if a minor merger with an SMC-type galaxy would be the origin of the dust lane in NGC\,5485, somehow a large fraction of this atomic gas should have been lost during the merger event. Atomic gas is generally more loosely bound than molecular gas, so it could be more easily be stripped during interactions (but it would probably still be visible as a tidal tail). Another argument against such a scenario is that the gas-to-dust ratios of low-metallicity dwarf galaxies are typically up to an order of magnitude higher than those of giant galaxies \citep{2014A&A...563A..31R}, which makes the apparently low gas-to-dust ratio in NGC\,5485 even more puzzling. Whether such a merging scenario is plausible or even possible, is a challenge that could be investigated using detailed merger hydrodynamical simulations. Also deeper H{\sc{i}} and CO observations, sensitive enough to trace possible diffuse gas, would be useful to unveil the nature of the ISM in this peculiar system.
\section{Conclusions}
We have discussed the interstellar dust and gas properties of the dust-lane ETG NGC\,5485. We present new Herschel PACS and SPIRE imaging of NGC\,5485, taken in the frame of the {\it{Far-infraRed Investigation of Early-type galaxies with Dust Lanes}} (FRIEDL) program. Using both standard modified blackbody model fits and the MAGPHYS spectral energy distribution modeling code, we obtain a dust mass $M_{\text{d}} = 3.8\times10^6~M_\odot$. The combination of the dust mass and H{\sc{i}} and CO stringent upper limits, obtained in the frame of the ATLAS$^{\text{3D}}$ survey, leads to an exceptionally low gas-to-dust ratio, $M_{\text{gas}}/M_{\text{d}} < 14.5$, almost an order of magnitude lower than the canonical value of the Milky Way.
We have investigated different possible explanations for this extreme gas-to-dust ratio. We have critically checked the reliability of the dust mass estimate, but neither the lack of spatial resolution in the FIR/submm imaging, nor a possible contamination of background or companion sources affect the dust mass. Similarly, the reliability of the cold gas mass is scrutinized. The main source of uncertainty here is the notorious CO-to-H$_2$ conversion factor, and if we would assume the "standard" value rather than the slightly higher value adopted by the ATLAS$^{\text{3D}}$ consortium, the gas-to-dust ratio upper limit would even by decreased to $M_{\text{gas}}/M_{\text{d}} < 11.0$.
Finally, we investigate less obvious scenarios to explain the lack of cold gas in NGC\,5485. Based on H$\alpha$ and X-ray observations, we can discard the possibility that the bulk of the gas is in a warm of hot ionized medium. One possible option is the presence of a component of CO-dark molecular gas. In principle, the extreme gas-to-dust ratio could be the result of a merger with an SMC-type metal-poor dwarf galaxy, if a substantial fraction of the H{\sc{i}} could have been lost during the interaction, but it remains to be investigated whether such a scenario is possible.
\section*{Acknowledgements}
The authors are indebted to the anonymous referee for providing insightful and constructive comments. MB, JF and TH acknowledge the financial support of the Belgian Science Policy Office (BELSPO) through the PRODEX project "Herschel-PACS Guaranteed Time and Open Time Programs: Science Exploitation" (C90370). MB, FA, SV and IDL acknowledge the support from the Flemish Fund for Scientific Research (FWO-Vlaanderen). IP thanks the Mexican foundation CONACyT for financial support.
PACS has been developed by a consortium of institutes led by MPE (Germany) and including UVIE (Austria); KU Leuven, CSL, IMEC (Belgium); CEA, LAM (France); MPIA (Germany); INAF-IFSI/OAA/OAP/OAT, LENS, SISSA (Italy); IAC (Spain). This development has been supported by the funding agencies BMVIT (Austria), ESA-PRODEX (Belgium), CEA/CNES (France), DLR (Germany), ASI/INAF (Italy), and CICYT/MCYT (Spain). SPIRE has been developed by a consortium of institutes led by Cardiff University (UK) and including Univ. Lethbridge (Canada); NAOC (China); CEA, LAM (France); IFSI, Univ.\ Padua (Italy); IAC (Spain); Stockholm Observatory (Sweden); Imperial College London, RAL, UCL-MSSL, UKATC, Univ.\ Sussex (UK); and Caltech, JPL, NHSC, Univ.\ Colorado (USA). This development has been supported by national funding agencies: CSA (Canada); NAOC (China); CEA, CNES, CNRS (France); ASI (Italy); MCINN (Spain); SNSB (Sweden); STFC (UK); and NASA (USA).
|
\section{Introduction}
In recent 16 years,
cosmic acceleration \cite{Riess98,spergel03,Tegmark04,Komatsu09,Percival10,Drinkwater10}
has become one of the most important issues in modern cosmology.
To explain this puzzle,
one can either introduce an unknown energy component
(i.e. dark energy (DE) \cite{quint,phantom,k,Chaplygin,ngcg,tachyonic,HDE,hessence,YMC,hscalar,cq,others1,others2,others3,others4}),
or modify Einstein's general relativity (i.e. modified gravity (MG) \cite{SH,PR,DGP,GB,Galileon,FR,FT1,FT2,FRT}).
For recent reviews, see \cite{CST,FTH,Linder,CK,Uzan,Tsujikawa,NO,LLWW,CFPS,YWBook}.
One of the most powerful probes to illuminate the mystery of cosmic acceleration
is Type Ia supernovae (SNe Ia).
Several high-quality supernova (SN) datasets had been released in recent years \cite{Union,Constitution,Union2,Union2.1}.
The Supernova Legacy Survey three-year (SNLS3) data \cite{Guy10} were released in 2010.
Soon after, using SNLS3 dataset,
Conley et al. \cite{Conley11} and Sullivan et al. \cite{Sullivan11}
presented the SN-only cosmological results and the joint cosmological constraints, respectively.
Unlike other SN group, the SNLS team treated two important quantities,
stretch-luminosity parameter $\alpha$ and color-luminosity parameter $\beta$ of SNe Ia,
as free model parameters.
A critical challenge is the control of the systematic uncertainties of SNe Ia.
One of the most important factors that yield systematic uncertainties is the potential SN evolution,
i.e. the possibility for the redshift evolution of $\alpha$ and $\beta$.
So far, it is found that $\alpha$ is still consistent with a constant,
but the hints for the evolution of $\beta$ have been found in \cite{Astier06,Kessler09,Marriner11,Scolnic1,Scolnic2}.
In \cite{Mohlabeng}, using a linear $\beta(z) = \beta_0 + \beta_1 z$,
Mohlabeng and Ralston studied the case of Union2.1 dataset
and found that $\beta$ deviates from a constant at 7$\sigma$ confidence levels (CL).
In \cite{WangWang}, Wang \& Wang found that, for the SNLS3 data,
$\beta$ increases significantly with $z$ at the 6$\sigma$ CL;
moreover, they proved that this conclusion is insensitive to the lightcurve fitter models,
or the functional form of $\beta(z)$ assumed.
These studies show that the evolution of $\beta$ is a common phenomenon for various SN datasets,
and should be taken into account seriously.
It is very important to study the effects of a time-varying $\beta$ on parameter estimation of cosmological models.
In \cite{WangNew}, Wang, Li \& Zhang explored this issue
by considering the $\Lambda$-cold-dark-matter ($\Lambda$CDM) model, the $w$CDM model, and the Chevallier-Polarski-Linder (CPL) model.
Then, in \cite{WangNew2}, Wang, Geng, Hu \& Zhang studied the case of holographic dark energy (HDE) model,
which is a physically plausible DE candidate based on the holographic principle.
Next, in \cite{WangNew3}, Wang, Wang, Geng \& Zhang extended the discussion
to the case of considering the interaction between dark sectors.
It is found that, for all these models, $\beta$ deviates from a constant at $\sim$ 6$\sigma$ CL;
in addition, a time-varying $\beta$ will significantly change the confidence ranges of various cosmological parameters.
It must be stressed that, in previous studies, only DE models are adopted to explore the issue of varying $\beta$.
To do a comprehensive analysis on the cosmological consequences of a time-varying $\beta$,
it is necessary to extend the discussions to the case of MG,
which is another important approach to explaining cosmic acceleration.
So in this paper, we explore the effects of a time-varying $\beta$ on the cosmological constraints of three popular MG models,
including Dvali-Gabadadze-Porrati (DGP) model \cite{DGP} and two $f(T)$ models \cite{FT1,FT2}.
In addition to the SNLS3 data,
we also use the Planck distance prior data \cite{WangWangCMB} of the cosmic microwave background (CMB),
the galaxy clustering (GC) data
from Sloan Digital Sky Survey (SDSS) data release 7 (DR7) \cite{ChuangWang12} and Baryon Oscillation Spectroscopic Survey (BOSS) \cite{Chuang13},
as well as the direct measurement of Hubble constant $H_0$ from the Hubble Space Telescope (HST) observation \cite{Riess11}.
We describe our method in Sec.~II, present our results in Sec.~III, and conclude in Sec.~IV.
In this paper, we assume today's scale factor $a_{0}=1$, thus the redshift $z=a^{-1}-1$.
The subscript ``0'' always indicates the present value of the corresponding quantity,
and the natural units are used.
\section{Methodology}
\label{sec:method}
In this section, we introduce the theoretical models we considered
and the observational data we used in this paper.
\subsection{Theoretical models}
In this paper, we consider a Friedmann-Lematre-Robertson-Walker Universe with a non-zero spatial curvature.
We investigate three popular MG models, including the DGP model, the power-law type $f(T)$ model, and the exponential type $f(T)$ model.
\begin{itemize}
\item DGP model
\end{itemize}
The DGP model is a braneworld model \cite{DGP},
where gravity is altered at immense distances by slow leakage of gravity off from our three-dimensional universe.
For this model, the dimensionless Hubble parameter $E(z)\equiv H(z)/H_{0}$ is given by
\begin{eqnarray}
E(z) = \left[ \sqrt{\Omega_{m0}(1+z)^3+\Omega_{r0}(1+z)^4+\Omega_{k0}(1+z)^2+\Omega_{rc}}+\sqrt{\Omega_{rc}}\right],
\end{eqnarray}
where $\Omega_{rc} = (1-\Omega_{m0}-\Omega_{r0}-\Omega_{k0})^2/4$.
Here $\Omega_{m0}$, $\Omega_{r0}$ and $\Omega_{k0}$
are the present fractional densities of matter, radiation and curvature, respectively.
In addition, we have $\Omega_{r0}=\Omega_{m0} / (1+z_{\rm eq})$,
and $z_{\rm eq}=2.5\times 10^4 \Omega_{m0} h^2 (T_{\rm cmb}/2.7\,{\rm K})^{-4}$ with
$T_{\rm cmb}=2.7255\,{\rm K}$.
\begin{itemize}
\item $f(T)$ models
\end{itemize}
In the $f(T)$ gravity theory \cite{FT1,FT2},
the torsion scalar $T$ in the Lagrangian density is replaced by a generalized function $T + f(T)$,
then the corresponding action $S$ can be written as
\begin{equation}\label{fT:action}
S=\frac{1}{16\pi G}\int d^{4}x\sqrt{-g}[T+f(T)]+S_{m}+S_{r}+S_{k},
\end{equation}
where $S_{m}$, $S_{r}$ and $S_{k}$ are the actions of matter, radiation and curvature, respectively.
Since the torsion scalar $T$ and the Hubble expansion rate $H$ satisfy the relation $T=-6H^2$,
the modified Friedmann equation can be written as \cite{FT2}
\begin{equation}\label{fT:F.e.}
H^{2}=\frac{8\pi G}{3}(\rho_{m}+\rho_{r}+\rho_{k})-\frac{f}{6}-2H^{2}f_{T}.
\end{equation}
Here $\rho_{m}$, $\rho_{r}$ and $\rho_{k}$ denote the energy densities of matter, radiation and curvature, respectively;
beside, $f_{T}\equiv df/dT$ is the derivative of $f(T)$ with respect to $T$.
Making use of the Hubble constant $H_0$, Eq. (\ref{fT:F.e.}) can be rewritten as
\begin{equation}\label{fT:E}
E(z)^{2}=\frac{1}{1+2f_{T}}\big(\Omega_{m0}(1+z)^{3}+\Omega_{r0}(1+z)^{4}+\Omega_{k0}(1+z)^{2}-\frac{f}{6H_{0}^{2}}\big).
\end{equation}
Here we consider two $f(T)$ models:
one is a power law form model proposed in \cite{FT1}, the other is an exponential form model proposed by Linder \cite{FT2}.
For simplicity, hereafter we will call them $f(T)_{PL}$ model and $f(T)_{EXP}$ model, respectively.
The $f(T)_{PL}$ model assumes the following ansatz of $f(T)$,
\begin{equation} \label{eq:fT1}
f(T) = \alpha (-T)^{n}.
\end{equation}
Here $n$ is a free model parameter.
Using Eq. (\ref{fT:F.e.}), the value of $\alpha$ can be fixed as
\begin{equation}
\alpha = (6 H_{0}^{2})^{1-n} \frac{1-\Omega_{m0}-\Omega_{r0}-\Omega_{k0}}{2n-1}.
\end{equation}
For the $f(T)_{PL}$ model, Eq. (\ref{fT:E}) becomes
\begin{equation}\label{fTpl:E}
E(z)^{2}=\Omega_{m0}(1+z)^{3}+\Omega_{r0}(1+z)^{4}+\Omega_{k0}(1+z)^{2}+(1-\Omega_{m0}-\Omega_{r0}-\Omega_{k0})E(z)^{2n}.
\end{equation}
Moreover, this equation can be rewritten as
\begin{equation}\label{fTpl:dEdz}
\frac{dE}{dz}=\frac{3\Omega_{m0}(1+z)^{2}+4\Omega_{r0}(1+z)^{3}+2\Omega_{k0}(1+z)}{2E-2n(1-\Omega_{m0}-\Omega_{r0}-\Omega_{k0})E^{2n-1}}.
\end{equation}
Making use of the initial condition $E(0) = 1$ and numerically solving Eq. (\ref{fTpl:dEdz}),
the evolution of $E(z)$ for the $f(T)_{PL}$ model can be easily obtained.
The $f(T)_{EXP}$ model adopts the following ansatz of $f(T)$ \cite{FT2},
\begin{equation} \label{fT2}
f(T) = m T_0 \big(1-e^{-p\sqrt{T/T_{0}}}\big).
\end{equation}
Here $p$ is a free model parameter, $T_0 = -6H^{2}_{0}$, and
\begin{equation} \label{c}
m=\frac{1-\Omega_{m0}-\Omega_{r0}-\Omega_{k0}}{1-(1+p)e^{-p}}.
\end{equation}
For the $f(T)_{EXP}$ model, Eq. (\ref{fT:E}) becomes
\begin{equation}\label{fTexp:E}
E(z)^{2}=\Omega_{m0}(1+z)^{3}+\Omega_{r0}(1+z)^{4}+\Omega_{k0}(1+z)^{2}+m\Big(1-\big(1+pE(z)\big)e^{-pE(z)}\Big).
\end{equation}
Moreover, this equation can be rewritten as
\begin{equation}\label{fTexp:dEdz}
\frac{dE}{dz}=\frac{3\Omega_{m0}(1+z)^{2}+4\Omega_{r0}(1+z)^{3}+2\Omega_{k0}(1+z)}{2E-mp^{2}E(z)e^{-pE(z)}}.
\end{equation}
Making use of the initial condition $E(0) = 1$ and numerically solving Eq. (\ref{fTexp:dEdz}),
the evolution of $E(z)$ for the $f(T)_{EXP}$ model can be obtained, too.
\subsection{Observational data}
In this subsection, we introduce how to calculate the $\chi^2$ function for SNLS3 data in detail.
For the SNLS3 sample, the observable is $m_B$, which is the rest-frame peak B-band magnitude of the SN.
By considering three functional forms (linear case, quadratic case, and step function case),
Wang \& Wang \cite{WangWang} showed that the evolutions of $\alpha$ and $\beta$
are insensitive to functional form of $\alpha$ and $\beta$ assumed.
So in this paper,
we just adopt a constant $\alpha$ and a linear $\beta(z) = \beta_{0} + \beta_{1} z$.
Then, the predicted magnitude of a SN becomes
\begin{equation}
m_{\rm mod}=5 \log_{10}{\cal D}_L(z)
- \alpha (s-1) +\beta(z) {\cal C} + {\cal M},
\end{equation}
where $s$ and ${\cal C}$ are the stretch measure and the color measure for the SN light curve.
Here ${\cal M}$ is a parameter representing some combination of SN absolute magnitude $M$ and Hubble constant $H_0$.
It must be emphasized that,
to include host-galaxy information in the cosmological fits,
Conley et al. \cite{Conley11} split the SNLS3 sample based on host-galaxy stellar mass at $10^{10} M_{\odot}$,
and made ${\cal M}$ to be different for the two samples.
Therefore, unlike other SN samples, there are two values of ${\cal M}$, ${\cal M}_1$ and ${\cal M}_2$, for the SNLS3 data
(for the details, see the subsections $3.2$ and $5.8$ of \cite{Conley11}).
Moreover, Conley et al. removed ${\cal M}_1$ and ${\cal M}_2$ from cosmological fits by analytically marginalizing over them
(for more details, see the appendix C of \cite{Conley11},
as well as the the public code which is available at https://tspace.library.utoronto.ca/handle/1807/24512).
In this paper, we just follow the recipe of Ref. \cite{Conley11};
following Ref. \cite{Conley11}, we do not report the values of ${\cal M}_1$ and ${\cal M}_2$.
The luminosity distance ${\cal D}_L(z)$ is defined as
\begin{equation}
{\cal D}_L(z)\equiv H_0 (1+z_{\rm hel}) r(z),
\end{equation}
where $z$ and $z_{\rm hel}$ are the CMB restframe and heliocentric redshifts of SN.
In addition, the comoving distance $r(z)$ is given by
\begin{equation}
\label{eq:r(z)}
r(z)=H_0^{-1}\, |\Omega_{k0}|^{-1/2} {\rm sinn}\big (|\Omega_{k0}|^{1/2}\, \Gamma(z)\big ),
\end{equation}
where $\Gamma(z)=\int_0^z\frac{dz'}{E(z')}$,
and ${\rm sinn}(x)=\sin(x)$, $x$, $\sinh(x)$ for $\Omega_{k0}<0$, $\Omega_{k0}=0$, and $\Omega_{k0}>0$ respectively.
For a set of $N$ SNe with correlated errors, the $\chi^2$ function is
\begin{equation}
\label{eq:chi2_SN}
\chi^2_{SN}=\Delta \mbox{\bf m}^T \cdot \mbox{\bf C}^{-1} \cdot \Delta\mbox{\bf m},
\end{equation}
where $\Delta m \equiv m_B-m_{\rm mod}$ is a vector with $N$ components,
and $\mbox{\bf C}$ is the $N\times N$ covariance matrix of the SN, given by
\begin{equation}
\mbox{\bf C}=\mbox{\bf D}_{\rm stat}+\mbox{\bf C}_{\rm stat}+\mbox{\bf C}_{\rm sys}.
\end{equation}
$\mbox{\bf D}_{\rm stat}$ is the diagonal part of the statistical uncertainty, given by \cite{Conley11}
\begin{eqnarray}
\mbox{\bf D}_{{\rm stat},ii}&=&\sigma^2_{m_B,i}+\sigma^2_{\rm int}
+ \sigma^2_{\rm lensing}+ \sigma^2_{{\rm host}\,{\rm correction}} + \left[\frac{5(1+z_i)}{z_i(1+z_i/2)\ln 10}\right]^2 \sigma^2_{z,i} \nonumber\\
&& +\alpha^2 \sigma^2_{s,i}+\beta(z_i)^2 \sigma^2_{{\cal C},i} + 2 \alpha C_{m_B s,i} - 2 \beta(z_i) C_{m_B {\cal C},i} -2\alpha \beta(z_i) C_{s {\cal C},i},
\end{eqnarray}
where $C_{m_B s,i}$, $C_{m_B {\cal C},i}$, and $C_{s {\cal C},i}$
are the covariances between $m_B$, $s$, and ${\cal C}$ for the $i$-th SN,
$\beta_i=\beta(z_i)$ are the values of $\beta$ for the $i$-th SN.
Notice that $\sigma^2_{z,i}$ includes a peculiar velocity residual of 0.0005
(i.e., 150$\,$km/s) added in quadrature.
Following Ref. \cite{Conley11}, we fix the intrinsic scatter $\sigma_{int}$ to ensure that $\chi^2/dof=1$.
Varying $\sigma_{int}$ could have a significant impact on parameter estimation, see \cite{Kim2011} for details.
We define $\mbox{\bf V} \equiv \mbox{\bf C}_{\rm stat} + \mbox{\bf C}_{\rm sys}$,
where $\mbox{\bf C}_{\rm stat}$ and $\mbox{\bf C}_{\rm sys}$
are the statistical and systematic covariance matrices, respectively.
After treating $\beta$ as a function of $z$,
$\mbox{\bf V}$ is given in the form,
\begin{eqnarray}
\mbox{\bf V}_{ij}&=&V_{0,ij}+\alpha^2 V_{a,ij} + \beta_i\beta_j V_{b,ij} +\alpha V_{0a,ij} +\alpha V_{0a,ji}
-\beta_j V_{0b,ij} -\beta_i V_{0b,ji} -\alpha \beta_j V_{ab,ij} - \alpha \beta_i V_{ab,ji}.
\end{eqnarray}
It must be stressed that, while $V_0$, $V_{a}$, $V_{b}$, and $V_{0a}$
are the same as the ``normal'' covariance matrices
given by the SNLS data archive, $V_{0b}$, and $V_{ab}$ are {\it not} the same as the ones given there.
This is because the original matrices of SNLS3 are produced by assuming $\beta$ is constant.
We have used the $V_{0b}$ and $V_{ab}$ matrices for the ``Combined'' set
that are applicable when varying $\beta(z)$ (A.~Conley, private communication, 2013).
To improve the cosmological constraints,
we also use some other cosmological observations,
including the Planck distance prior data \cite{WangWangCMB},
the galaxy clustering (GC) data extracted from SDSS DR7 \cite{ChuangWang12} and BOSS \cite{Chuang13},
as well as the direct measurement of Hubble constant $H_0=73.8\pm 2.4~{\rm km/s/Mpc}$ from the HST observations \cite{Riess11}.
For the details of including these data into the $\chi^2$ analysis, see Refs. \cite{WangNew,WangNew2,WangNew3}.
Now the total $\chi^2$ function is
\begin{equation}
\chi^2=\chi^2_{SN}+\chi^2_{CMB}+\chi^2_{GC}+\chi^2_{H0}.
\end{equation}
Finally, we perform an MCMC likelihood analysis \cite{COSMOMC}
to obtain ${\cal O}$($10^6$) samples for each model considered in this paper.
\section{Results}
\subsection{Evolution of $\beta$}
\label{sec:varbeta}
In this subsection,
we explore the evolution of $\beta$ by considering the DGP model, the $f(T)_{PL}$ model and the $f(T)_{EXP}$ model.
As mentioned above, to explore the evolution of $\beta$,
we study the case of constant $\alpha$ and linear $\beta(z) = \beta_{0} + \beta_{1} z$;
for comparison, the case of constant $\alpha$ and constant $\beta$ is also taken into account.
\begin{itemize}
\item SN-only case
\end{itemize}
Firstly, we discuss the results given by the SN data alone.
Notice that the reduced Hubble constant $h$ has been marginalized during the $\chi^2$ fitting process of SNe Ia.
In Table \ref{table1}, we list the fitting results for various constant $\beta$ and linear $\beta(z)$ cases,
where only the SNLS3 SN data are used.
The most obvious feature of this table is that varying $\beta$ can significantly improve the fitting results of various MG models:
for all the models, adding a parameter of $\beta$ can reduce the best-fit values of $\chi^2$ by $\sim$ 36.
Based on the Wilk's theorem, 36 units of $\chi^2$ is equivalent to a Gaussian fluctuation of 6$\sigma$.
Therefore, for the case of using SNLS3 data alone,
a constant $\beta$ is ruled out at 6$\sigma$ CL for all the MG models.
This result is consistent with the results of dark energy cases \cite{WangNew,WangNew2,WangNew3},
showing that the evolution of $\beta$ is completely independent of the cosmological model in the background.
Therefore, by taking into account the MG models,
we further confirm the redshift-evolution of $\beta$ for the SNLS3 data.
\begin{table*}\tiny
\caption{Fitting results for various constant $\beta$ and linear $\beta(z)$ cases,
where only the SN data are used.}
\label{table1}
\begin{tabular}{ccccccccc}
\hline\hline &\multicolumn{2}{c}{DGP}&&\multicolumn{2}{c}{$f(T)_{PL}$}&&\multicolumn{2}{c}{$f(T)_{EXP}$} \\
\cline{2-3}\cline{5-6}\cline{8-9}
Parameter & Const $\beta$ & Linear $\beta(z)$ && Const $\beta$ & Linear $\beta(z)$ && Const $\beta$ & Linear $\beta(z)$ \\ \hline
$\alpha$ & $1.425^{+0.112}_{-0.100}$
& $1.409^{+0.108}_{-0.092}$&
& $1.424^{+0.107}_{-0.095}$
& $1.416^{+0.100}_{-0.098}$&
& $1.463^{+0.055}_{-0.150}$
& $1.392^{+0.098}_{-0.049}$\\
$\beta_0$ & $3.264^{+0.107}_{-0.110}$
& $1.450^{+0.345}_{-0.406}$&
& $3.259^{+0.142}_{-0.103}$
& $1.437^{+0.354}_{-0.387}$&
& $3.289^{+0.082}_{-0.122}$
& $1.466^{+0.235}_{-0.319}$\\
$\beta_1$ &
& $5.075^{+1.151}_{-0.923}$&
&
& $5.095^{+1.081}_{-0.938}$&
&
& $5.056^{+0.905}_{-0.636}$\\
$\Omega_{m0}$ & $0.121^{+0.075}_{-0.082}$
& $0.021^{+0.157}_{-0.011}$&
& $0.259^{+0.239}_{-0.249}$
& $0.052^{+0.324}_{-0.042}$&
& $0.468^{+0.050}_{-0.458}$
& $0.523^{+0.052}_{-0.513}$\\
$\Omega_{k0}$ & $0.0621^{+0.2375}_{-0.1947}$
& $0.3754^{+0.1361}_{-0.3647}$&
& $-0.7040^{+1.4111}_{-0.3698}$
& $0.5037^{+0.3332}_{-1.4052}$&
& $-1.0433^{+1.4011}_{-0.0751}$
& $-1.1034^{+1.7594}_{-0.1069}$\\
$n$ &
& &
& $1.2279^{+0.5949}_{-4.2276}$
& $-0.0334^{+1.3375}_{-2.9645}$&
&
& \\
$p$ &
& &
&
& &
& $ -0.5068^{+10.5054}_{-1.5496}$
& $0.0374^{+9.9595}_{-0.6623}$\\
\hline $\chi^{2}_{min}$ & 419.758 & 383.621 && 419.567 & 383.622 && 419.340 & 383.161 \\
\hline
\end{tabular}
\end{table*}
In Fig. \ref{fig1}, using the SN data alone,
we plot 1D marginalized probability distributions of $\beta_{1}$ (panel a),
as well as 1$\sigma$ confidence constraints of $\beta(z)$
for the DGP model (panel b), the $f(T)_{PL}$ model (panel c), and the $f(T)_{EXP}$ model (panel d).
To make a comparison, the best-fit results of various constant $\beta$ cases are also shown.
The panel A shows that, for all the MG models, $\beta_{1}>0$ at 6$\sigma$ CL;
while the panels B, C, and D show that, $\beta(z)$ rapidly increases with $z$.
These results further confirm that the evolution of $\beta$ is independent of the MG models,
showing that the importance of considering evolution of $\beta$ in the cosmology-fits.
\begin{figure}
\includegraphics[scale=0.25, angle=0]{Fig1d1.pdf}
\includegraphics[scale=0.25, angle=0]{Fig1d2.pdf}
\includegraphics[scale=0.25, angle=0]{Fig1d3.pdf}
\includegraphics[scale=0.25, angle=0]{Fig1d4.pdf}
\caption{\label{fig1}\footnotesize%
1D marginalized probability distributions of $\beta_{1}$ (panel a),
as well as 1$\sigma$ confidence constraints of $\beta(z)$ for the DGP (panel b), the $f(T)_{PL}$ (panel c), and the $f(T)_{EXP}$ model (panel d).
Only the SN data are used in the analysis. To make a comparison, the best-fit results of various constant $\beta$ cases are also plotted.}
\end{figure}
It should be pointed out that the evolutionary behavior of $\beta(z)$ depends on the SN samples used.
In \cite{Mohlabeng}, Mohlabeng and Ralston found that, for the Union2.1 SN data, $\beta(z)$ decreases with $z$.
This is similar to the case of Pan-STARRS1 SN data \cite{Scolnic2}.
It is of great interest to study why different SN data give different evolutionary behaviors of $\beta(z)$,
and some numerical simulation studies may be required to solve this problem.
We will study this issue in future works.
\begin{itemize}
\item SN+CMB+GC+$H_{0}$ case
\end{itemize}
Next, let us discuss the results given by the SN+CMB+GC+$H_{0}$ data.
It should be mentioned that, in order to use the Planck distance priors data,
two new model parameters, reduced Hubble parameter $h$ and radiation parameter $\Omega_{b}h^{2}$ must be added.
In Table \ref{table2}, we make a comparison for the fitting results of constant $\beta$ and linear $\beta(z)$ cases,
where the SN+CMB+GC+$H_{0}$ data are used.
Again, we see that varying $\beta$ can significantly improve the fitting results of various MG models:
for the $f(T)_{PL}$ and the $f(T)_{EXP}$ models,
adding a parameter of $\beta$ will reduce the best-fit values of $\chi^2$ by $\sim$ 36;
for the DGP model, adding a parameter of $\beta$ will reduce the best-fit value of $\chi^2$ by $\sim$ 47.
Therefore, the conclusion of $\beta_1\neq0$ still holds true for the SN+CMB+GC+$H_{0}$ case.
In addition, only using the SN data,
the $\chi^{2}_{min}$ value of DGP model is almost the same to the results of $f(T)_{PL}$ and $f(T)_{EXP}$ models;
once taking into account other observational data,
the $\chi^{2}_{min}$ value of DGP model becomes significantly larger than the results of $f(T)_{PL}$ and $f(T)_{EXP}$ models.
This implies that, for the DGP model,
the cosmological constraints given by the SN data is significantly inconsistent with the cosmological constraints given by other cosmological observations.
Therefore, we can conclude that the DGP model is strongly disfavored by the current cosmological observations;
this is consistent with the conclusions of many previous works \cite{Fang08,Rubin09,Li:2009jx,ZWS}.
\begin{table*}\tiny
\caption{Fitting results for various constant $\beta$ and linear $\beta(z)$ cases,
where the SN+CMB+GC+$H_0$ data are used.}
\label{table2}
\begin{tabular}{ccccccccc}
\hline\hline &\multicolumn{2}{c}{DGP}&&\multicolumn{2}{c}{$f(T)_{PL}$}&&\multicolumn{2}{c}{$f(T)_{EXP}$} \\
\cline{2-3}\cline{5-6}\cline{8-9}
Parameter & Const $\beta$ & Linear $\beta(z)$ && Const $\beta$ & Linear $\beta(z)$ && Const $\beta$ & Linear $\beta(z)$ \\ \hline
$\alpha$ & $1.406^{+0.112}_{-0.097}$
& $1.398^{+0.098}_{-0.087}$&
& $1.414^{+0.110}_{-0.086}$
& $1.394^{+0.120}_{-0.073}$&
& $1.433^{+0.089}_{-0.111}$
& $1.415^{+0.098}_{-0.096}$\\
$\beta_0$ & $3.250^{+0.123}_{-0.089}$
& $1.101^{+0.416}_{-0.310}$&
& $3.266^{+0.098}_{-0.105}$
& $1.506^{+0.309}_{-0.408}$&
& $3.243^{+0.116}_{-0.098}$
& $1.448^{+0.343}_{-0.341}$\\
$\beta_1$ &
& $6.063^{+0.887}_{-1.148}$&
&
& $4.950^{+1.045}_{-0.885}$&
&
& $5.093^{+1.006}_{-0.919}$\\
$\Omega_{m0}$ & $0.299^{+0.012}_{-0.011}$
& $0.303^{+0.012}_{-0.010}$&
& $0.264^{+0.013}_{-0.009}$
& $0.281^{+0.017}_{-0.011}$&
& $0.276^{+0.009}_{-0.011}$
& $0.279^{+0.010}_{-0.010}$\\
$\Omega_{k0}$ & $0.0310^{+0.0034}_{-0.0040}$
& $0.0303^{+0.0037}_{-0.0038}$&
& $-0.002^{+0.0044}_{-0.0044}$
& $0.0016^{+0.0050}_{-0.0051}$&
& $0.0050^{+0.0031}_{-0.0033}$
& $0.0051^{+0.0028}_{-0.0035}$\\
$n$ &
& &
& $-0.465^{+0.271}_{-0.300}$
& $-0.083^{+0.202}_{-0.236}$&
&
& \\
$p$ &
& &
&
& &
& $9.0888^{+0.9118}_{-3.8316}$
& $7.1324^{+2.8669}_{-2.7097}$\\
$h$ & $0.683^{+0.011}_{-0.012}$
& $0.678^{+0.010}_{-0.012}$&
& $0.731^{+0.011}_{-0.017}$
& $0.707^{+0.014}_{-0.021}$&
& $0.712^{+0.013}_{-0.011}$
& $0.710^{+0.012}_{-0.012}$\\
$\Omega_{b}h^{2}$ & $0.0224^{+0.00031}_{-0.00027}$
& $0.0224^{+0.00026}_{-0.00030}$&
& $0.0222^{+0.00032}_{-0.00025}$
& $0.0223^{+0.00028}_{-0.00027}$&
& $0.0223^{+0.00033}_{-0.00024}$
& $0.0223^{+0.00030}_{-0.00027}$\\
\hline $\chi^{2}_{min}$ & 455.965 & 408.834 && 421.941 & 386.965 && 425.410 & 388.878 \\
\hline
\end{tabular}
\end{table*}
In Fig. \ref{fig2}, using SN+CMB+GC+$H_0$ data,
we plot 1D marginalized probability distributions of $\beta_{1}$ (left panel)
and 1$\sigma$ confidence constraints of $\beta(z)$ (right panel) for various MG models.
The left panel shows that, for all the MG models, $\beta_{1}>0$ at 6$\sigma$ CL;
while the right panel show that $\beta(z)$ rapidly increases with $z$ for all the models.
This result is just the same to the result of SN-only case.
In addition, it is also consistent with the results of dark energy cases \cite{WangNew,WangNew2,WangNew3}.
Notice that the evolution of $\beta$ for the DGP model is slightly different from the results for the $f(T)_{PL}$ and $f(T)_{EXP}$ models;
this maybe due to the possible degeneracy between $\Omega_{m0}$ and $\beta(z)$.
But this slight difference has no influence on the conclusion of time-varying $\beta$.
\begin{figure}
\includegraphics[scale=0.25, angle=0]{Fig2d1.pdf}
\includegraphics[scale=0.25, angle=0]{Fig2d2.pdf}
\caption{\label{fig2}\footnotesize%
1D marginalized probability distributions of $\beta_{1}$ (left panel)
and 1$\sigma$ confidence constraints of $\beta(z)$ (right panel) for various MG models.
The SN+CMB+GC+$H_0$ data are used in the analysis.}
\end{figure}
\subsection{Effects of time-varying $\beta$}
\label{sec:betaeff}
In this subsection, we discuss the effects of varying $\beta$ on parameter estimations of various MG models.
For simplicity, in this subsection we just use the SN+CMB+GC+$H_0$ data.
\begin{itemize}
\item DGP model
\end{itemize}
Firstly, let us discuss the results of DGP model.
An advantage of this model is that it has the same parameter number with the simplest $\Lambda$CDM model.
In Fig. \ref{fig3},
we plot 1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel), for the DGP model.
We find that varying $\beta$ yields a larger $\Omega_{m0}$ and a smaller $h$:
the best-fit results of constant $\beta$ case are $\Omega_{m0}=0.299$ and $h=0.683$,
while best-fit results of the linear $\beta(z)$ case are $\Omega_{m0}=0.303$ and $h=0.678$.
This result is consistent with the conclusions of dark energy cases \cite{WangNew,WangNew2,WangNew3}.
\begin{figure}
\includegraphics[scale=0.25, angle=0]{Fig3d1.pdf}
\includegraphics[scale=0.25, angle=0]{Fig3d2.pdf}
\caption{\label{fig3}\footnotesize%
1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel),
given by the SN+CMB+GC+$H_0$ data, for the DGP model.
Both the results of constant $\beta$ and linear $\beta(z)$ cases are presented.}
\end{figure}
\begin{itemize}
\item $f(T)_{PL}$ model
\end{itemize}
Then, let us turn to the case of $f(T)_{PL}$ model.
This model has an additional model parameter $n$.
In Fig. \ref{fig4},
we plot 1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel), for the $f(T)_{PL}$ model.
It can be seen that varying $\beta$ also yields a larger $\Omega_{m0}$ and a smaller $h$ for this case:
the best-fit results of constant $\beta$ case are $\Omega_{m0}=0.264$ and $h=0.731$,
while best-fit results of the linear $\beta(z)$ case are $\Omega_{m0}=0.281$ and $h=0.707$.
This result is consistent with the result of Fig. \ref{fig3}.
\begin{figure}
\includegraphics[scale=0.25, angle=0]{Fig4d1.pdf}
\includegraphics[scale=0.25, angle=0]{Fig4d2.pdf}
\caption{\label{fig4}\footnotesize%
1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel),
given by the SN+CMB+GC+$H_0$ data, for the $f(T)_{PL}$ model.
Both the results of constant $\beta$ and linear $\beta(z)$ cases are presented.}
\end{figure}
In Fig. \ref{fig5},
we plot the 1$\sigma$ and 2$\sigma$ confidence contours of $\{\Omega_{m0},n\}$, for the $f(T)_{PL}$ model.
From this figure, one can see that varying $\beta$ yields a larger $n$:
the best-fit value of constant $\beta$ case is $n=-0.465$,
while best-fit value of the linear $\beta(z)$ case is $n=-0.083$.
Moreover, it can be seen that a time-varying $\beta$ significantly change the shapes of 1$\sigma$ and 2$\sigma$ confidence regions;
this implies that ignoring the evolution of $\beta$ will cause systematic bias.
\begin{figure}
\psfig{file=Fig5.pdf,width=3.5in}\\
\caption{\label{fig5}\footnotesize%
The 1$\sigma$ and 2$\sigma$ confidence contours of $\{\Omega_{m0},n\}$, for the $f(T)_{PL}$ model.
Both the results of constant $\beta$ and linear $\beta(z)$ cases are presented.
}
\end{figure}
\begin{itemize}
\item $f(T)_{EXP}$ model
\end{itemize}
Finally, we turn to the $f(T)_{EXP}$ model, which has an additional model parameter $p$.
In Fig. \ref{fig6},
we plot 1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel),
for the $f(T)_{EXP}$ model.
Again, we see that varying $\beta$ yields a larger $\Omega_{m0}$ and a smaller $h$:
the best-fit results of constant $\beta$ case are $\Omega_{m0}=0.276$ and $h=0.712$,
while best-fit results of the linear $\beta(z)$ case are $\Omega_{m0}=0.279$ and $h=0.710$.
This result is consistent with the results of Fig. \ref{fig3} and Fig. \ref{fig4}.
\begin{figure}
\includegraphics[scale=0.25, angle=0]{Fig6d1.pdf}
\includegraphics[scale=0.25, angle=0]{Fig6d2.pdf}
\caption{\label{fig6}\footnotesize%
1D marginalized probability distributions of $\Omega_{m0}$ (left panel) and $h$ (right panel),
given by the SN+CMB+GC+$H_0$ data, for the $f(T)_{EXP}$ model.
Both the results of constant $\beta$ and linear $\beta(z)$ cases are presented.}
\end{figure}
In Fig. \ref{fig7},
we plot the 1$\sigma$ and 2$\sigma$ confidence contours of $\{\Omega_{m0},p\}$, for the $f(T)_{EXP}$ model.
It can be seen that varying $\beta$ yields a smaller $p$;
in addition, a time-varying $\beta$ will change the shapes of 1$\sigma$ and 2$\sigma$ confidence regions.
\begin{figure}
\psfig{file=Fig7.pdf,width=3.5in}\\
\caption{\label{fig7}\footnotesize%
The 1$\sigma$ and 2$\sigma$ confidence contours of $\{\Omega_{m0},p\}$, for the $f(T)_{EXP}$ model.
Both the results of constant $\beta$ and linear $\beta(z)$ cases are presented.
}
\end{figure}
According to Figs. \ref{fig3}, \ref{fig4} and \ref{fig6},
we can conclude that a time-varying $\beta$ always yields a larger $\Omega_{m0}$ and a smaller $h$.
In addition, based on Figs. \ref{fig5} and \ref{fig7},
we can conclude that varying $\beta$ significantly changes the shapes of 1$\sigma$ and 2$\sigma$ confidence regions, and thus corrects systematic bias.
These two conclusions are independent of the cosmological models in the background.
\section{Discussion and summary}
In recent years,
the control of the systematic uncertainties of SNe Ia has drawn more and more attention.
One of the most important systematic uncertainties for SNe Ia is the potential SN evolution.
The hints for the evolution of $\beta$ have been found \cite{Astier06,Kessler09,Marriner11,Scolnic1,Scolnic2,Mohlabeng}.
In \cite{WangWang}, using the SNLS3 data,
Wang \& Wang found strong evidence for the redshift-evolution of $\beta$;
moreover, they proved that the evolution of $\beta$ is insensitive to the lightcurve fitter models,
or the functional form of $\beta(z)$ assumed.
It is clear that a time-varying $\beta$ will have significant impact on parameter estimation.
Adopting a constant $\alpha$ and a linear $\beta(z) = \beta_{0} + \beta_{1} z$,
Wang, Li \& Zhang \cite{WangNew} explored this issue by considering $\Lambda$CDM model, $w$CDM model, and CPL model.
Then, Wang, Geng, Hu \& Zhang \cite{WangNew2} studied this issue in the framework of HDE model,
which is a physically plausible DE candidate based on the holographic principle.
Soon after, Wang, Wang, Geng \& Zhang \cite{WangNew3} extended the corresponding discussion
to the case of considering the interaction between dark sectors.
It is found that, for all these models, $\beta$ deviates from a constant at $\sim$ 6$\sigma$ CL;
in addition, a time-varying $\beta$ will significantly change the confidence ranges of various cosmological parameters.
It must be stressed that, in previous studies, only DE models are adopted to explore the issue of varying $\beta$.
To do a comprehensive analysis on the cosmological consequences of a time-varying $\beta$,
it is necessary to extend the discussions to the case of MG.
So in this paper, we explore the effects of a time-varying $\beta$ on the cosmological constraints of three popular MG models,
including DGP model, $f(T)_{PL}$ model and $f(T)_{EXP}$ model.
In addition to the SNLS3 SN data,
we also use the Planck distance priors data,
the GC data extracted from SDSS DR7 and BOSS,
as well as the direct measurement of Hubble constant $H_0$ from the HST observation.
In this paper,
we further confirm the evidence of redshift-evolution of $\beta$ for the SNLS3 data.
We find that, for both the cases of using the SN data alone and using the combination of all data,
adding a parameter of $\beta$ can reduce $\chi^2$ by $\sim$ 36 for all the MG models,
showing that a constant $\beta$ is ruled out at 6$\sigma$ CL.
Moreover, we find that a time-varying $\beta$ always yields a larger $\Omega_{m0}$ and a smaller $h$;
in addition, it significantly changes the shapes of 1$\sigma$ and 2$\sigma$ confidence regions of various MG models,
and thus corrects systematic bias for the parameter estimation.
The conclusions of our paper are consistent with the results of DE cases,
showing that the conclusion of time-varying $\beta$ holds true for both DE and MG models.
In other words, $\beta$'s evolution is completely independent of the cosmological models in the background.
Therefore, our work highlights the importance of considering the evolution of $\beta$ in the cosmology-fits.
In this paper, only the potential SN evolution is taken into account.
Some other factors, such as the evolution of $\sigma_{int}$ \cite{Kim2011},
may also cause systematic uncertainties for SNe Ia.
This issue deserves further study in future.
\begin{acknowledgments}
We are grateful to Dr. Alex Conley for providing us with the SNLS3 covariance matrices that
allow redshift-dependent $\beta$. We acknowledge the use of CosmoMC.
SW is supported by the Fundamental Research Funds for the Central Universities under Grant No. N130305007.
XZ is supported by the National Natural Science Foundation of China under Grant No. 11175042
and the Fundamental Research Funds for the Central Universities under Grant No. N120505003.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
In this paper the problem of asymptotically stabilizing sets using
output feedback is investigated. Many control objectives can be
accomplished by stabilizing an appropriate invariant set, often a
submanifold, in the state space of a control system. This point of
view is relevant in applications such as output
regulation~\cite{Dav76},~\cite{Fra77},~\cite{IsiByr90},
synchronization~\cite{PogSanNij02}, formation control problems for
multi-agent systems~\cite{ElhMag13},~\cite{KriFraBro09} and path
following~\cite{FreRobShiJoh08},~\cite{NieFulMag10}. Topological
obstructions for submanifold stabilization using full-state feedback
were characterized in~\cite{Man10}. Output feedback controllers are
necessary whenever the state of the system is not available for
feedback. This is common in applications where, due to economic or
technological reasons, sensors cannot measure a system's entire state.
The most natural approach to stabilizing sets using output feedback is
to find an observable, i.e., available for feedback, function that
yields a well-defined relative degree whose associated zero dynamics
manifold coincides with the set to be stabilized. If such an
observable function exists, then the set stabilization problem becomes
an output stabilization problem which can be solved using classical
and well-understood output feedback control design
techniques~\cite{AndPra09},~\cite{AtaKha99},~\cite{MarPraIsi07},~\cite{TeePra95}. The
main contribution of this paper are necessary and sufficient
conditions for the existence of such an observable function
(Theorem~\ref{thm:main}).
In the case of full-state feedback, i.e., the full information case,
the above approach to set stabilization was studied
in~\cite{NieMag08}. There we sought a coordinate and feedback
transformation locally bringing the control system to a ``normal
form,'' in which the system's dynamics are decomposed into two
cascade-connected subsystems. In our normal form, the driving system
is linear, time-invariant, and controllable. It models the dynamics
``transversal'' to the target set in the sense that, in transformed
coordinates, the target set corresponds to the origin of this linear
system. We refer to the driving subsystem as the transversal
subsystem. On the other hand, the restriction of the driven system to
the target set represents the ``tangential'' motion of the control
system on the set, and for this reason such restriction is referred to
as the tangential subsystem. The process of bringing the original
control system to the normal form just described is called local
transverse feedback linearization (LTFL). This terminology originated
with the work of Andrzej Banaszuk and John Hauser
in~\cite{BanHau95}
\subsection{Contributions}
The contributions of the paper are the following. 1) The results
in~\cite{NieMag06},~\cite{NieMag08} are extended to the partial
information case in Theorem~\ref{thm:main}. In that work we assumed
that the full state of the control system is available for feedback
and that, in particular, the local transverse output was permitted to
be a function of the entire state. In this paper we assume that the
only information available for feedback is modeled by a fixed
output. We refer to this as the partial information case. This work is
complementary to the
papers~\cite{AtaKha99},~\cite{MarPraIsi07},~\cite{MarPraIsi10}~\cite{DelliPriscoli2009},~\cite{TeePra95}. In
those papers output feedback controllers are designed for systems in a
given normal form. The main result of this work, motivated by set
stabilization problems, provides necessary and sufficient conditions
under which a system can be brought into the aforementioned normal
form in which the control design techniques can be applied. 2)
Sufficient conditions under which a global version of the problem can
be solved are presented in Section~\ref{sec:global}. 3) In
Section~\ref{sec:disturbances} we show how the results of this paper
can be used to facilitate control design for systems affected by
unmeasured disturbances. A preliminary version of this paper appeared
in~\cite{NieMag12}.
\section{Motivating example}
\label{sec:motivation}
Consider a system
\begin{equation}
\left[
\begin{array}{c}
\dot x_1 \\
\dot x_2 \\
\dot x_3 \\
\dot x_4 \\
\dot x_5
\end{array}
\right] = \left[
\begin{array}{c}
x_4 \\
-x_3-x_2^3\\
x_2\\
0\\
x_1
\end{array}
\right] + \left[
\begin{array}{c}
x_1\\0\\0\\1\\x_5
\end{array}
\right]u
\label{eq:example1}
\end{equation}
with output
\begin{equation}
y = h(x) = \left[
\begin{array}{c}
x_4 \\
x_5
\end{array}
\right].
\label{eq:ex_out}
\end{equation}
We are interested in locally stabilizing the invariant set
\[
\Gamma^\star = \set{x \in \Real^5 : x_1 = x_4 = x_5 = 0}.
\]
As discussed in Section~\ref{sec:intro}, the most direct approach to
locally stabilizing this set using output feedback is to seek an
observable function that yields, in a \nbhd of a point $x_0 \in
\Gamma^\star$, a well-defined relative degree whose associated zero
dynamics manifold coincides with $\Gamma^\star$. If such a function
exists, then zeroing the function locally solves the set stabilization
problem (if the trajectories of the closed-loop system are
bounded). Furthermore, if the function is observable,
system~\eqref{eq:example1} is feedback equivalent to a system that
fits the framework of well-known and ``standard'' output feedback
control design approaches~\cite{AtaKha99},~\cite{TeePra95} that can
zero the observable function.
Specifically, in this example we seek a function $\lambda: \Real^5 \to
\Real$ that, in a \nbhd of a point $x_0 \in \Gamma^\star$, has the
following properties.
\begin{enumerate}
\item The function $\lambda$ yields a well-defined relative degree at
$x_0 \in \Gamma^\star$.
\item The zero dynamics manifold of System~\eqref{eq:example1} with
output $\lambda$ coincides with the target set $\Gamma^\star$ in a
\nbhd of $x_0$.
\item The function $\lambda$ is observable. In other words, the
function $\lambda$ can be expressed as a composition $\lambda =
\tilde{\lambda} \circ h$ of a sufficiently smooth function
$\tilde{\lambda} : \Real^2 \to \Real$ with the
output~\eqref{eq:ex_out}.
\end{enumerate}
A natural first attempt to finding a function with the aforementioned
properties is to check if any of the constraints that define
$\Gamma^\star$ satisfy the conditions enumerated above. In this
example neither of the functions $x_1$ and $x_5$ yield a well-defined
relative degree at any point on $\Gamma^\star$. Furthermore, $x_1$ is
not observable. The constraint function $x_4$ does yield a
well-defined relative degree and is observable, however the zero
dynamics manifold associated to the output $x_4$ does not equal
$\Gamma^\star$. Hence making $x_4 \To 0 $ does not ensure that the set
$\Gamma^\star$ is locally attractive. These facts mean that it is not
clear whether or not a function $\lambda$ that satisfies the three
conditions above exists and, therefore, it is not clear whether or not
the above program can be carried out.
The main contribution of this paper is to provide, given a control
system with fixed output and an invariant set, necessary and
sufficient conditions for the existence of a function $\lambda$ that
satisfies the three conditions listed above. In this example the
function
\begin{equation}
\lambda(x) = \tilde{\lambda} \circ h(x) = x_5\e^{-x_4}
\label{eq:lambda_ex}
\end{equation}
meets the above criteria with $\tilde{\lambda}(y) = y_2\e^{-y_1}$. We
now illustrate how this function facilities output feedback
stabilization of $\Gamma^\star$.
Using the observable function~\eqref{eq:lambda_ex} define the
coordinate transformation
\[
\left[
\begin{array}{c}
\eta_1 \\
\eta_2 \\
\xi_1 \\
\xi_2 \\
\xi_3
\end{array}
\right] \coloneqq \left[
\begin{array}{c}
x_2 \\
x_3\\
x_5\e^{-x_4}\\
x_1\e^{-x_4}\\
x_4\e^{-x_4}
\end{array}
\right]
\]
which, by the inverse function theorem, is a diffeomorphism of a \nbhd
of any point $x \in \Real^5$. The system in $(\eta, \xi)$-coordinates
reads
\[
\left[
\begin{array}{c}
\dot \eta_1 \\
\dot \eta_2 \\
\dot \xi_1 \\
\dot \xi_2 \\
\dot \xi_3
\end{array}
\right] = \left[
\begin{array}{c}
-\eta_2-\eta_1^3 \\
\eta_1\\
\xi_2\\
\xi_3\\
0
\end{array}
\right] + \left[
\begin{array}{c}
0\\0\\0\\0\\1
\end{array}
\right]\phi(\eta, \xi)u
\]
where $\phi(\eta, \xi)=\left.\left(1 - x_4\right)\e^{-x_4}\right|_{x =
T^{-1}(\eta, \xi)}$ and where $\xi_1 = \lambda(x)$ is available for
feedback. For this system there are various approaches one can take to
stabilize the $\xi$-subsystem. For example, one can view the
stabilization problem as the study of a system with unknown high
frequency gain for which the techniques in~\cite{Nuss83} along with
the switching strategy in~\cite{ILCOWE91} can be employed to stabilize
$\Gamma^\star$. Alternatively, using the results in~\cite{AtaKha99},
there exists a dynamic feedback that stabilizes $\Gamma^\star$ using
only measurements of $\xi_1$. The ``high-gain'' observer used
in~\cite{AtaKha99} takes the form
\[
\dot{\hat{\xi}} = \left[\begin{array}{ccc}0 & 1 & 0\\ 0 & 0 & 1\\0 &
0 & 0\end{array}\right]\hat{\xi} + \left[
\begin{array}{c}
0\\0\\1
\end{array}
\right]\phi_0(\hat{\xi})u + \left[
\begin{array}{c}
\frac{\alpha_1}{\varepsilon}\\\frac{\alpha_2}{\varepsilon^2}\\\frac{\alpha_3}{\varepsilon^3}
\end{array}
\right]\left(\xi_1
- \hat{\xi}_1\right)
\]
where $\hat{\xi} \in \Real^3$ is an estimate of $\xi$,
$\phi_0(\hat{\xi}) = (1 - \hat{\xi}_3)$ is the nominal, inexact, model
of $\phi(\eta, \xi)$, $\varepsilon > 0$ is a high-gain parameter and
the constants $\alpha_i$ are chosen so that the polynomial $s^3 +
\alpha_1s^2 + \alpha_2s + \alpha_3$ is Hurwitz. The control law is
chosen as
\[
u = \frac{1}{\phi_0(\hat{\xi})}\left(-k_1\hat{\xi}_1 -k_2\hat{\xi}_2 -k_3\hat{\xi}_3 \right)
\]
with $k_i > 0$, $i \in \set{1,2,3}$.
The applicability of the above approach to output feedback control
design depends crucially on the existence of the observable
function~\eqref{eq:lambda_ex}. Therefore, a key challenge in output
feedback stabilization of invariant sets is finding such a function
and, most importantly, in determining whether it exists or not. This
paper completely solves the latter question for single-input systems.
\section{Preliminaries}
This section presents the notation used throughout the
paper. Section~\ref{sec:bundles} contains supporting material needed
to prove the main result. Section~\ref{sec:vectorfields} provides
definitions for the concept of invariance used in this paper and the
Lie derivative and Lie bracket.
\subsection{Notation}
Let $\col{(x_1 \ldots, x_k)} \coloneqq \left[\begin{array}{ccc}x_1 &
\cdots & x_n\end{array}\right]^\top$ where ${}^\top$ denotes
transpose. Let $x$ and $y$ be two column vectors, define $\col{(x,y)}
\coloneqq \left[\begin{array}{cc}x^\top &
y^\top\end{array}\right]^\top$. If $x \in \Real^n$ then $\|x\|$
denotes the Euclidean norm. If $\mathscr{V}$ and $\mathscr{W}$ are
subspaces of the finite-dimensional vector space $\mathscr{X}$, the
notation $\mathscr{V} \oplus \mathscr{W}$ (internal direct sum)
represents the subspace $\mathscr{V} + \mathscr{W}$ when $\mathscr{V}$
and $\mathscr{W}$ are independent.
If $f$ is a scalar-valued function from an open set $U \subseteq
\Real^n$ into $\Real$, and $k$ times continuously differentiable for
at every $x \in U$, then $f$ is of differentiability class $C^k$ on
$U$, denoted $f \in C^k(U)$ or $f \in C^k$ when the domain of $f$ is
clear. If $f$ is $C^k$ for all $k$, then $f$ is $C^\infty$ or
smooth. If $f : U \subseteq \Real^n \to V \subseteq \Real^m$ is a
continuously differentiable map, then for each $x \in U$, the
derivative of $f$ at $x$, denoted $\D f_x$ , is a linear map $\D f_x :
\Real^n \to \Real^m$. Its matrix representation is the Jacobian matrix
of $f$ evaluated at $x$. If $U$ is an open set of $\Real^n$, let
$\Diff{U}$ denote the family of diffeomorphism with domain $U$.
For brevity, the term submanifold is used in place of embedded
submanifold of $\Real^n$. If $M$ is a smooth manifold and $ p \in M$,
we denote by $T_pM$ the tangent space to $M$ at $p$ and by $TM$ the
tangent bundle of $M$. The cotangent space to $M$ at $p$ is denoted by
$T^\star_pM$ and the cotangent bundle is written as $T^\star M$.
\begin{definition}
\label{def:distribution}
A smooth distribution $D$ on a manifold $M$ is an assignment to each
$p \in M$ of a subspace $D(p) \subseteq T_pM$ which varies smoothly as
a function of $p$. A point $p \in M$ is a regular point of the smooth
distribution $D$ if there exists a \nbhd $U$ containing $p$ for which
$\dim{(D(q))}$ is constant for all $q \in U$. In this case, $D$ is
said to be {nonsingular} on $U$. Similarly, a codistribution $\Omega$
on $M$ assigns at each $p \in M$ a subspace $\Omega(p) \subseteq
T^\star_pM$.
\end{definition}
Given a smooth distribution $D$, we let $\inv{(D)}$ be its involutive
closure (the smallest involutive distribution containing $D$). The
codistribution $\ann{(D)}$ is the annihilator of $D$, i.e., an
assignment to each $p \in M$ of a subspace $\ann{(D)}(p) \subseteq
T_p^\star M$ with the property that if $\sigma \in \ann{(D)}(p)$ and
$\tau \in D(p)$, then $\sigma(\tau) = 0$.
\subsection{Vector bundles}
\label{sec:bundles}
If $D$ is a distribution defined on $\Real^n$ and $N$ is a submanifold
we at times consider objects like $TN + D$ and $TN \cap D$. These
objects are examples of real vector bundles, more precisely,
subbundles of $\left.T\Real^n\right|_N$. They are defined, for each $p
\in N$, by $T_pN+ D(p)$ and $T_p N \cap D(p)$, respectively. These
subbundles, and the operations on them, can be defined formally using
the framework of vector bundles~\cite{Hir76, Lee02, Spi05}.
\begin{definition}
A $n$-dimensional (real) vector bundle is a map
\[
\pi : E \rightarrow B
\]
of manifolds $E$ and $B$ such that, for any $b \in B$, the inverse
image $\pi^{-1}(b)$ has the structure of the $n$-dimensional vector
space $\Real^n$ having the following property of local triviality: For
each $b \in B$, there exists a \nbhd $U$ of $b$ in $B$ and a
diffeomorphism
\[
h: \pi^{-1}(U) \rightarrow U \times \Real^n
\]
such that for every $b^\prime \in U$ the assignment of $x \in
\pi^{-1}(b^\prime)$ to $h(x) = (b^\prime, \hat{h}(x))$ is an
isomorphism of $\pi^{-1}(b^\prime)$ to $\{b^\prime\} \times
\Real^n$. The manifold $E$ is called the total space, $B$ is
called the base space and the vector space $E_b \coloneqq \pi^{-1}(b)$
is called the fibre over $b$.
\label{def:vecbundle}
\end{definition}
All of the vector bundles encountered in this paper are
finite-dimensional and real. We typically denote a vector bundle
$(\pi, E, B)$ by $E$ alone. Given a bundle $(\pi, E, B)$, for each
$b\in B$ we can replace the fibre $\pi^{-1}(b)$ with different vector
spaces. In this paper we will only consider the simplest case, we
replace each vector space $\pi^{-1}(b)$ with its dual space.
\begin{definition}
Let $\xi = (\pi, E, B)$ be a vector bundle. The dual bundle to $\xi$,
is $\xi^\star = \left(\pi^\star, E^\star, B\right)$ where
\[
E^\star \coloneqq \bigcup_{b \in B} \left(\pi^{-1}(b)\right)^\star,
\]
and $\pi^\star: E^\star \rightarrow B$ is the natural projection
$\pi^\star : \left(\pi^{-1}(p)\right)^\star \mapsto p$.
\end{definition}
When this construction is applied to the tangent bundle $TM$ of a
manifold $M$, the resulting bundle is the cotangent bundle $T^\star M$
of $M$.
\begin{definition}
Let $\eta=(\pi_F, F, B)$ be a subbundle of the smooth vector bundle
$\xi = (\pi, E, B)$. The annihilator $\ann{(\eta)}$ of $\eta$, is the
subbundle of $\xi^\star$ whose fibres are defined at each $b \in B$ by
\[
\begin{aligned}
\ann{\left(F_b\right)} &\coloneqq \{e^\star \in E^\star_b: e^\star(f) =
0, \; \forall \ f \in F_b\}.
\end{aligned}
\]
\label{def:annvb}
\end{definition}
Recall that, if $\mathscr{X}$ is a finite dimensional vector space,
then $\left(\mathscr{X}^\star\right)^\star =:
\mathscr{X}^{\star\star}$ is canonically isomorphic to
$\mathscr{X}$. Using this fact, and applying
Definition~\ref{def:annvb} twice to the vector bundle $\xi$, we obtain
the following.
\begin{proposition}
Let $\xi = (\pi, E, B)$ be a smooth vector bundle over $B$. Then
\[
\ann{(\ann{(\xi)})} = \xi.
\]
\label{prop:annann}
\end{proposition}
Proposition~\ref{prop:annann} implies that, if $\eta^\star$ is a
subbundle of the cotangent bundle $T^\star M$, then
$\ann{\left(\eta^\star\right)}$ is a subbundle of $TM$, the tangent
bundle to $M$. The following results, needed in this paper, can be
found in~\cite{Hir76, Lee02, Spi05}.
\begin{proposition}
Let $\xi = (\pi, E, B)$, $\xi_1 = (\pi_{E_1}, E_1, B)$ and $\xi_2 =
(\pi_{E_2}, E_2, B)$ be vector bundles such that $\xi_2 \subseteq
\xi_1 \subseteq \xi$, then,
\[
\ann{(\xi)} \subseteq \ann{(\xi_1)} \subseteq \ann{(\xi_2)} \subseteq
\xi^\star.
\]
\label{prop:ann12}
\end{proposition}
\begin{proposition}
Let $\xi_1 = (\pi_{E_1}, E_1, B)$, $\xi_2 = (\pi_{E_2}, E_2, B)$ be
subbundles of the smooth vector bundle $\xi = (\pi, E, B)$. If $\xi_1
+ \xi_2 = \left(\pi_{F}, F, B\right)$ is also a subbundle of $\xi$,
then
\[
\ann{\left(\xi_1 + \xi_2\right)} = \ann{(\xi_1)}\cap \ann{(\xi_2)}.
\]
\label{prop:annint}
\end{proposition}
\subsection{Invariant sets, Lie derivatives, Lie brackets}
\label{sec:vectorfields}
Denote the set of all $C^\infty$-vector fields on a smooth manifold
$M$ by ${\mathsf V}(M)$. Given $v \in {\mathsf V}(M)$ and a point $p
\in M$, we denote the maximal integral curve, or flow, generated by
the vector field $v$ through the point $p$ as $\phi^v_t(x)$.
\begin{definition}
A set $N \subset M$ is said to be invariant under $v \in {\mathsf
V}(M)$ if
\[
\left(p \in N\right) \Rightarrow (\forall t \geq 0)(\phi^v_t(p)
\in N).
\]
\label{def:invariant}
\end{definition}
The property of invariance in Definition~\ref{def:invariant} is
sometimes called positive or forward invariance because $N$ is
invariant for $t \geq 0$. When $N$ is a closed submanifold invariance
for $t \geq 0$ is equivalent to invariance for $t \in \Real$. If $N$
is an $n$-dimensional submanifold of $M$ expressed as
$N = \{p \in M: \phi(p) = 0\}$,
where $\phi(p) = \col(\phi_1(p), \ldots, \phi_{m-n}(p))$ is a smooth
map $M \rightarrow \Real^{m-n}$, and $0$ is a regular value of
$\phi$, then there is a particularly simple criterion for invariance.
\begin{theorem}
Let $\phi : M \rightarrow \Real^{m-n}$ be a smooth map, and $0$
be a regular value of $\phi$. Let $v \in {\mathsf V}(M)$, then, $N =
\phi^{-1}(0)$ is invariant under $v$ if, and only if,
\[
\left(\D\phi_i\right)_p(v(p)) = 0
\]
for all $i \in \{1, \ldots, m-n\}$ and all $p \in \phi^{-1}(0)$.
\label{thm:inv_sub}
\end{theorem}
Geometrically, the theorem asserts that $N$ is invariant under $v$ if
and only if $v$ is tangent to $N$, everywhere on $N$. The same is true
for general closed submanifolds of $M$.
\begin{definition}
If $v \in {\mathsf V}(M)$ and $\lambda\in C^\infty(M)$ then the
derivative of $\lambda$ along $v$ is a function $L_v\lambda: M
\rightarrow \Real$ defined by
\[
L_v\lambda(p) = \lim_{h\rightarrow
0}\frac{1}{h}\left[\lambda(\phi^v_h(p)) - \lambda(p)\right]
\]
and called the Lie or directional derivative of $\lambda$ along $v$ at
$p$. It is an element of $C^\infty(M)$.
\label{def:liederivative}
\end{definition}
\begin{definition}
If $f, g \in {\mathsf V}(M)$, then the Lie bracket of $f$ and $g$ is
a vector field $\liebr{f}{g} \in {\mathsf V}(M)$ defined by the
relation
\[
\left(\forall \lambda \in C^\infty(M)\right) \qquad L_{\left[f,
g\right]}\lambda = L_f (L_g \lambda) - L_g (L_f \lambda).
\]
\label{def:liebracket}
\end{definition}
Definitions~\ref{def:liederivative} and~\ref{def:liebracket} are implicit
in that they do not directly indicate how to compute, respectively,
the Lie derivative and Lie bracket.
If $\lambda \in C^\infty(\Real^n)$ and $v \in \mathsf{V}(\Real^n)$
then $L_v\lambda(x)$ is computed as
\[
L_v\lambda(x) = \left(\D\lambda\right)_x(v(x)).
\]
If $f, g \in {\mathsf V}(\Real^n)$, the Lie bracket of $f$ and $g$ is
computed as
\[
\liebr{f}{g}(x) = \D g_x\left(f(x)\right) - \D f_x\left(g(x)\right),
\]
where $\D f_x$, $\D g_x$ are the derivative maps of the vector
functions $f, g: \Real^n \rightarrow \Real^n$. We use the following
standard notation for iterated Lie derivatives and Lie brackets
\[
\begin{aligned}
&L^0_g\lambda \coloneqq \lambda, \; \;
\; L^k_g\lambda \coloneqq L_g(L^{k-1}_g\lambda),\\
&L_gL_f\lambda \coloneqq L_g(L_f\lambda), \\ &ad^0_fg \coloneqq g, \; \; \;
ad^k_fg \coloneqq \left[f, ad^{k-1}_fg\right], \; \; \; k \geq 1.
\end{aligned}
\]
\section{Problem formulation}
\label{sec:problem}
Consider a control system modeled by equations of the form
\begin{equation}
\dot{x} = f(x) + g(x)u.
\label{eq:system}
\end{equation}
Here $x \in \Real^n$ is the state and $u \in \Real$ is the control
input. The vector fields $f$ and $g : \Real^n \rightarrow T\Real^n$
are smooth ($C^\infty$). Suppose that the state $x$ is not available
for feedback but, rather, the only available information is given by a
smooth vector output
\begin{equation}
\begin{aligned}
y = h(x), \qquad h: \Real^n \rightarrow \Real^p.
\end{aligned}
\label{eq:output}
\end{equation}
We assume that the component functions $\col{(h_1(x), \ldots,
h_p(x))}$ of the output $h(x)$ are linearly independent,
i.e., we assume that $\D h_x$ has rank $p$ for all $x \in \Real^n$.
Define the following distributions associated with control
system~\eqref{eq:system}
\begin{equation}
\begin{aligned}
\Gg_i \coloneqq &\Sp\{ad^j_fg : 0 \leq j \leq i\}.
\end{aligned}
\label{eq:G}
\end{equation}
To the output~\eqref{eq:output} we associate the nonsingular,
involutive, $(n-p)$-dimensional distribution
\begin{equation}
\label{eq:W}
\W \coloneqq \ann{\left(\Sp{\left\{\D h_1, \ldots ,
\D h_p\right\}}\right)}.
\end{equation}
Suppose that we are given a submanifold $\Gamma^\star \subset \Real^n$
of dimension $0 < n^\star < n$ which is either invariant under the
vector field $f(x)$ in~\eqref{eq:system} or controlled invariant,
i.e., it can be made invariant by appropriate choice of smooth
feedback.
\begin{definition}
\label{def:cis}
A closed connected submanifold $N \subset \Real^n$ is called
controlled invariant for~\eqref{eq:system} if there exists a smooth
feedback $\overline{u}: N \rightarrow \Real$ making $N$ an invariant
set for the closed-loop system
\end{definition}
In this paper, as in~\cite{NieMag08}, we treat the controlled
invariant set $\Gamma^\star$ as given data. Often, however, one is
given a set $\Gamma \subset \Real^n$, perhaps defined by virtual
constraints or design goals, and then one must pare away pieces of
$\Gamma$ until all that remains is the maximal controlled invariant
submanifold $\Gamma^\star$ contained in $\Gamma$. We now state the
problem considered in this paper.
\begin{problem*}{Local Transverse Feedback Linearization with Partial
Information (LTFLPI) Problem }
Given a smooth single-input system~\eqref{eq:system} with smooth
output~\eqref{eq:output}, a closed, connected, embedded,
$n^\star$-dimensional controlled invariant submanifold $\Gamma^\star
\subset \Real^n$ and a point $x_0 \in \Gamma^\star$, find, if
possible, a diffeomorphism $\Xi \in \Diff{U}$
\begin{equation}
\begin{aligned}
\Xi : U &\to \Xi(U) \subset (\Gamma^\star \cap U) \times \Real^{n -
n^\star}\\
x &\mapsto (\eta, \xi)
\end{aligned}
\end{equation}
where $U$ is a \nbhd of $x_0$, such that
\begin{itemize}
\item[(i)] The restriction of $\Xi$ to $\Gamma^\star \cap U$ is
\[
\left.\Xi\right|_{\Gamma^\star \cap U} : x \mapsto (\eta, 0).
\]
\item[(ii)] The dynamics of system~\eqref{eq:system} in $(\eta,
\xi)$-coordinates reads
\begin{equation}
\begin{aligned}
\dot{\eta} &= f_0(\eta, \xi)\\
\dot{\xi} &= A\xi + b(a_1(\eta, \xi) + a_2(\eta, \xi)u),
\end{aligned}
\label{eq:partial_info}
\end{equation}
where the pair $(A,b)$ is in Brunovsk\'{y} normal form (one chain of
integrators) and $a_2(\eta, \xi) \neq 0$ in $\Xi(U)$.
\item[(iii)] The first component of $\xi$, denoted $\xi_1$ is
observable, i.e., there exists a function $\tilde{\lambda} : h(U)
\subseteq \Real^p \to \Real$ such that
\[
\xi_1(x) = \tilde{\lambda} \circ h(x).
\]
\end{itemize}
\end{problem*}
As illustrated in Section~\ref{sec:motivation}, solving LTFLPI is
relevant for stabilizing the set $\Gamma^\star$ using output feedback.
To understand this claim, suppose that LFTLPI is solvable at $x_0 \in
\Gamma^\star$. Let $\lambda(x) \coloneqq \tilde{\lambda} \circ
h(x)$. Using this function we partially define the diffeomorphism
$\Xi(x)$ by letting $\xi \coloneqq \col{(\lambda(x), L_f\lambda(x), }$
$\ldots, L_f^{n - n^\star - 1}\lambda(x))$. Choose $n - n^\star$
additional independent functions $\eta_i \coloneqq \phi_i(x)$, $i \in
\{1, \ldots, n - n^\star\}$, to complete the coordinate transformation
$\Xi: U \rightarrow \Real^{n^\star} \times \Real^{n - n^\star}$, $x
\mapsto (\eta, \xi)$. In the single-input case, since
$\ann{\left(\Sp{\set{g}}\right)}$ is spanned by exact differentials,
the functions $\phi_i(x)$ can always be chosen (see~\cite{Isi95}) so
that their time derivative along the control system do not depend on
$u$, i.e., so that for all $x \in U$ and all $i \in \{1, \ldots,
n^\star\}$, $L_g\phi_i(x) = 0$. After applying this coordinate
transformation, in $(\eta, \xi)$-coordinates the system is modeled by
equation~\eqref{eq:partial_info}.
If the entire state $x$ is available for feedback, as in the full
information case, then the regular feedback transformation $u =
-\frac{a_1(\eta, \xi)}{a_2(\eta, \xi)} + \frac{v}{a_2(\eta, \xi)}$
yields a system of the form
\begin{equation}
\begin{aligned}
\dot{\eta} &= f_0(\eta, \xi)\\
\dot{\xi} &= A\xi + bv
\end{aligned}
\label{eq:full_info}
\end{equation}
and we say that system~\eqref{eq:system} has been locally
transversely feedback linearized with respect to the set
$\Gamma^\star$. In this case, stabilizing the subspace $\xi = 0$ in
$(\eta, \xi)$-coordinates corresponds to stabilizing the set
$\Gamma^\star \cap U$ in original coordinates (if the trajectories of
the closed-loop system are bounded). For this reason we call the
$\xi$-subsystem of~\eqref{eq:partial_info} the transverse dynamics
of~\eqref{eq:system} with respect to $\Gamma^\star$. Stabilizing
$\xi=0$ can be achieved easily using the auxiliary control input $v$
since the pair $(A,b)$ is controllable. On the target set, the system
dynamics are governed by the ordinary differential equation
\begin{equation}
\dot{\eta} = f_0(\eta, 0).
\label{eq:tang}
\end{equation}
For this reason the dynamics~\eqref{eq:tang} are called the tangential
dynamics of~\eqref{eq:system} with respect to $\Gamma^\star$.
In the partial information case the state $x$ is not available for
feedback, the only available information is given by the output
function~\eqref{eq:output}. In this case $(\eta, \xi)$ is not
available for feedback, the feedback transformation above cannot be
implemented, and it may be impossible to stabilize the $\xi$
subsystem. In the partial information case the $\xi$-subsystem before
feedback transformation is
\begin{equation}
\label{eq:xi_subsys}
\begin{aligned}
\dot \xi_1 &= \xi_2\\
&\cdots\\
\dot \xi_{n-n^\star -1} &= \xi_{n-n^\star}\\
\dot \xi_{n-n^\star} &= a_1(\eta, \xi) + a_2(\eta, \xi)u.
\end{aligned}
\end{equation}
Since $\xi_1 = \lambda(x) = \tilde{\lambda}(h(x))$, it is available
for feedback. For system~\eqref{eq:partial_info} with $\xi_1$ measured
and $a_2(\eta, \xi)$ sign-definite, the results
in~\cite{AtaKha99},~\cite{MarPraIsi07},~\cite{MarPraIsi10}~\cite{DelliPriscoli2009},~\cite{TeePra95}
assert the existence of a dynamic feedback
\[
\begin{aligned}
&\dot \zeta = \varphi(\zeta, \xi_1)\\
&u = \varrho(\zeta, \xi_1)
\end{aligned}
\]
capable of stabilizing the origin of~\eqref{eq:xi_subsys}. The
selection of an appropriate output feedback design framework depends
crucially on the properties of tangential system~\eqref{eq:tang}. In
the simplest case, the results of~\cite{AtaKha99} can be used whenever
$\phi(\eta, \xi, u) \coloneqq a_1(\eta, \xi) + a_2(\eta, \xi)u$ and
$f_0(\eta, \xi)$ are locally Lipschitz with $\phi(0,0,0)=0$,
$f_0(0,0)=0$ and the tangential dynamics~\eqref{eq:tang} are minimum
phase. Alternatively, if practical stability is sought then the
results of~\cite[Section 6]{TeePra95} can be used, again provided the
tangential subsystem is minimum phase.
In the cases when the tangential dynamics do not necessarily converge
to zero but remain otherwise bounded, the results
in~\cite{MarPraIsi07} are relevant. In~\cite{MarPraIsi07} a
weak-minimum phase assumption is made on the tangential subsystem and
the function $a_1(\eta, \xi)$ is not necessarily known. Furthermore,
they require that the tangential dynamics $f_0(\eta, \xi)$ have the
form $f_0(\eta, \xi_1)$. Sufficient conditions for this additional
property to hold are given in
Corollary~\ref{prop:global_normal_form}. Finally, in cases where the
tangential system has the form $f_0(\eta, \xi_1)$ and $\DER{a_2(\eta,
\xi)}{\xi_i} \equiv 0$, $i \in \set{2, \ldots, n-n^\star}$ and
$\DER{a_2(\eta, \xi)}{\eta} \equiv 0$, the results in~\cite{ByrIsi91},
see also~\cite{AndPra09}, are applicable.
Once a system is expressed in the normal form~\eqref{eq:partial_info}
there are many other output stabilization techniques one can
consider. The survey~\cite{AndPra09} gives an excellent overview of
the available techniques while also classifying them as direct or
indirect approaches. Conceptually, the direct design approach is
preferable because the state feedback stabilizing control law is known
and therefore the estimation scheme can focus on estimating the
control signal directly. On the other hand, the indirect approach is
far more common in the research literature and in particular, the
approach used in Section~\ref{sec:motivation} is an example of
``domination via a dominant model''~\cite{AndPra09}. Motivated by
these observations, we seek conditions guaranteeing the existence of
an observable transverse output function.
\section{Main result}
\label{sec:main}
The next result, an obvious consequence of~\cite[Theorem
4.1]{NieMag06} or~\cite[Theorem 3.1]{NieMag08}, shows that LFTLPI is
solvable if and only if there exists a ``virtual output'' function
yielding a well-defined relative degree.
\begin{theorem} LTFLPI is solvable at $x_0\in\Gamma^\star$ if and only
if there exists a smooth $\Real$-valued function $\tilde{\lambda}$,
defined on a \nbhd of $h(x_0)$ in $\Real^p$ satisfying
\begin{itemize}
\item[(a)] for some \nbhd $U$ of $x_0 \in \Real^n$, $\Gamma^\star \cap
U \subseteq \{x\in U: \tilde{\lambda} \circ h (x)=0\}$, and
\item[(b)] the system
\begin{equation}
\begin{aligned}
&\dot x=f(x) + g(x) u \\
&y^\prime = \lambda(x) = \tilde{\lambda} \circ h (x)
\end{aligned}
\label{eq:virtual_output}
\end{equation}
has relative degree $n - n^\star$ at $x_0$.
\end{itemize}
Moreover, if LTFLPI is solvable, then there exists a \nbhd $V
\subseteq U$ of $x_0$ such that, on $V$, a connected component
$\mathcal{Z}^\star$ of the zero dynamics manifold
of~\eqref{eq:virtual_output} coincides with $\Gamma^\star$ :
$\mathcal{Z}^\star \cap V = \Gamma^\star \cap V$.
\label{thm:rel_iff}
\end{theorem}
The proof of Theorem~\ref{thm:rel_iff} is omitted because it is almost
identical to the proof of~\cite[Theorem 4.1]{NieMag06}.
\begin{definition}
Let $\tilde{\lambda}$ be a smooth $\Real$-valued functions
satisfying the conditions of Theorem~\ref{thm:rel_iff}. The map
$\lambda(x) \coloneqq \tilde{\lambda} \circ h(x)$ is called a local
observable transverse output of~\eqref{eq:system},~\eqref{eq:output}
with respect to $\Gamma^\star$.
\end{definition}
The main result of this paper, presented next, gives necessary and
sufficient conditions for the existence of an observable transverse
output.
\begin{theorem}
\label{thm:main}
Suppose that $\inv{\left(\Gg_{n-n^\star -2}+ \W\right)}$ is regular at
$x_0 \in \Gamma^\star$. Then LTFLPI is solvable at $x_0$ for
system~\eqref{eq:system} if and only if
\begin{itemize}
\item[(a)] $T_{x_0}\gstar \oplus \Gg_{n-n^\star -1}(x_0) = T_{x_0}\Real^n$
\item[(b)] there exists an open \nbhd $U$ of $x_0$ in $\Real^n$ such
that, $\left(\forall x \in \gstar \cap U\right)$,
\[
\dim{\left(T_x\gstar \oplus \Gg_{n-n^\star -2}(x)\right)} =
\dim{\left(T_x\gstar \oplus \inv{\left(\Gg_{n-n^\star -2}+
\W\right)}(x)\right)}.
\]
\end{itemize}
\end{theorem}
\begin{proof}
Suppose that LTFLPI is solvable at $x_0 \in \Gamma^\star$. Condition
(a) is coordinate and feedback invariant so it suffices to show that
it holds for system~\eqref{eq:full_info}. Let $V \coloneqq
\Xi{(\Gamma^\star \cap U)}$. By the properties of the normal
form~\eqref{eq:partial_info}, $\Xi(x_0) = \col{(p_0, 0)}$. Hence
in $(\eta,\xi)$-coordinates
\[
T_{p_0}V + \Gg_{{n} - n^\star - 1}(\col{(p_0, 0)})= \image{\left(\begin{bmatrix}
I_{n^\star} & \star & \star & \ldots & \star \\ 0_{n-n^\star\times
n^\star}& b & Ab & \ldots & A^{n-n^\star-1}b
\end{bmatrix}\right).}
\]
Since $(A,b)$ is a controllable pair it immediately follows that $T_pV
+ \Gg_{{n} - n^\star - 1}(\col{(p, 0)}) = T_p\Real^n$.
Furthermore, since~\eqref{eq:system} is a single-input system and by
the definition~\eqref{eq:G} of $\Gg_i$, $\dim{(\Gg_{{n} - n^\star -
1})} \leq n - n^\star$, and therefore the subspaces
$T_{x_0}\Gamma^\star$ and $\Gg_{{n} - n^\star - 1}(x_0)$ are
independent which proves that condition $(a)$ is necessary.
We are left to show that condition $(b)$ is necessary. Since LTFLPI
is solvable and $\xi_1(x) = \lambda(x) = \tilde{\lambda} \circ h(x)$,
we have
$\Gamma^\star \cap U = \{x \in \Real^{{n}}: \xi(x) = 0\} \subseteq \{x
\in \Real^{{n}}: {\lambda}(x) = 0\}$
so that, for all $x \in \Gamma^\star \cap U$ and for any $v \in
T_{x}\Gamma^\star$,
$L_v{\lambda}(x) = 0$.
This implies that $\D\lambda \in \ann{(T\Gamma^\star)}$. Furthermore,
since ${\lambda}(x)$ yields a well-defined relative degree of
$n-n^\star$ at $x_0$, for any $x$ in an open \nbhd of $x_0$, without
loss of generality $U$,
$L_g{\lambda}(x) = L_{ad_fg}{\lambda}(x) = \cdots =
L_{ad^{n-n^\star-2}_fg}{\lambda}(x) = 0$
and
$L_{ad^{n-n^\star - 1}_fg}{\lambda}(x) \neq 0$.
This means that, in a \nbhd of $x_0$, without loss of generality
$U$,
\[
\D{\lambda} \in \ann{\left(\Gg_{{n} - n^\star - 2}\right)}, \; \; \;
\D{\lambda} \not\in \ann{\left(\Gg_{{n} - n^\star - 1}\right)}.
\]
By the chain rule, $\D\lambda_x = \D\tilde{\lambda}_{h(x)} \circ \D
h_x,$
so that, for any vector field $w \in \W$, and all $x \in U$,
\[
\D\lambda_x(w(x)) = \D\tilde{\lambda}_{h(x)} \circ \D h_x(w(x)) = 0.
\]
In other words, $\D\lambda \in \ann{(\W)} = \Sp\{\D h_1, \ldots, \D
h_p\}$. This shows that, in $U$,
\[
\begin{aligned}
&\D\lambda \in \ann{\left(\Gg_{{n} - n^\star - 2}\right)} \cap \ann{\left(\W\right)}\\
&\Rightarrow \; \Gg_{{n} - n^\star - 2}+ \W \subseteq \ann{(\D\lambda)}.
\end{aligned}
\]
The distribution $\ann{(\D\lambda)}$ is involutive since its
annihilator is spanned by smooth, exact one-forms. Therefore
$\inv{(\Gg_{{n} - n^\star - 2}+ \W)} \subseteq \ann{(\D\lambda)}$
and by Proposition~\ref{prop:ann12},
\[
\begin{aligned}
&\D\lambda \in \ann{\left(\inv{(\Gg_{{n} - n^\star - 2}+ \W)}\right)}.
\end{aligned}
\]
This shows that, on $\Gamma^\star \cap U$, $d\lambda \in
\ann{(T\Gamma^\star)} \cap \ann{\left(\inv{(\Gg_{n - n^\star -
2}+\W)}\right)}$. Thus, by Proposition~\ref{prop:annint},
\[
\begin{aligned}
\D\lambda &\in \ann{\left(T\Gamma^\star +
\inv{(\Gg_{n - n^\star - 2}+ \W)}\right)}
\end{aligned}
\]
which implies that on $\Gamma^\star \cap U$,
\begin{equation}
\dim{\left(\ann{\left(T\Gamma^\star + \inv{\Gg_{n - n^\star - 2}+
W}\right)}\right)} \geq 1.
\label{eq:dim_ann}
\end{equation}
Therefore, by~\eqref{eq:dim_ann} and Proposition~\ref{prop:annann}, at
any point on $x \in \Gamma^\star \cap U$
\[
\dim{\left(T_{x}\Gamma^\star + \inv{(\Gg_{n - n^\star - 2}+
\W)}(x)\right)} < {n}.
\]
We have already shown that condition $(a)$ is necessary and therefore
on $\Gamma^\star \cap U$,
\begin{equation}
\dim{\left( T_x\Gamma^\star \oplus \Gg_{n -n^\star - 2}(x)\right)} = {n} -1.
\label{eq:dim_TVG}
\end{equation}
Therefore, by~\eqref{eq:dim_ann} and~\eqref{eq:dim_TVG}, we have that
for any point in $\Gamma^\star \cap U$
\begin{equation*}
\begin{aligned}
{n} -1 &= \dim{\left( T_{x}\Gamma^\star + \Gg_{n-n^\star -
2}\right)(x)}\\ &\leq \dim{\left(T_{x}\Gamma^\star +
\Gg_{n-n^\star - 2}(x) + \W(x)\right)}\\ &\leq \dim{\left(
T_{x}\Gamma^\star + \inv{\left(\Gg_{n -n^\star - 2}+ \W\right)}(x)
\right)} < n.
\end{aligned}
\end{equation*}
which proves the necessity of condition $(b)$.
We now turn to the proof of sufficiency. Conditions $(a)$ and $(b)$,
and the regularity of $\inv{(\Gg_{n-n^\star-2} + \W)}$ at $x_0$ imply
that $T\Gamma^\star \cap \inv{\left(\Gg_{n-n^\star-2}+\W\right)}$ is a
smooth nonsingular distribution near $x_0$. Using an argument
identical to that in the proof of~\cite[Lemma 4.5]{NieMag08}, it can
be shown that, by taking $U$ sufficiently small, there exists a smooth
nonsingular distribution $\Gg^\tang \subset
\inv{(\Gg_{n-n^\star-2}+\W)}$ on $U$ enjoying the two properties
\[
\begin{aligned}
\left(\forall x \in U\right) \;
&\inv{\left(\Gg_{n-n^\star-2}+W\right)}(x) = \Gg^\tang(x) \oplus
\Gg_{n -
n^\star-2}(x),\\
\left(\forall x \in \Gamma^\star \cap U\right) \;
&\left.\Gg^\tang\right|_{\Gamma^\star \cap U}(x) = T_x\Gamma^\star
\cap \inv{\left(\Gg_{n-n^\star-2}+\W\right)}(x).
\end{aligned}
\]
Let $V \coloneqq \Gamma^\star \cap U$ and let $w_1, \ldots, w_\mu$ be
a set of local generators for $\Gg^\tang$ on $U$. Similarly, there
exist $n^\star - \mu$ vector fields $\{v_1, \ldots, v_{n^\star -
\mu}\}$, $v_i : V \To TV$, such that, for all $x \in V$,
\[
T_xV = \left.\Gg^\tang\right|_V(x) \oplus \Sp\{v_1, \ldots, v_{n^\star - \mu}\}(x).
\]
Now we have a collection of $n$ linearly independent vector fields,
\begin{equation}
\underbrace{\underbrace{\overbrace{v_1, \ldots, v_{n^\star - \mu}}^{TV
/ \left.\Gg^\tang\right|_V}; \overbrace{w_1, \ldots, w_{\mu}}^{\Gg^\tang}}_{TV}; \overbrace{g, \ldots,
ad^{n-n^\star-2}_fg}^{\Gg_{n - n^\star -2}},
ad_f^{n-n^\star-1}g}_{T\Real^n}
\label{eq:array_info}
\end{equation}
which we use to generate a coordinate transformation analogous to that
used in Theorem~\cite[Theorem 3.2]{NieMag08}. We work our way from
left to right in the list~\eqref{eq:array_info} starting with the
group of vector fields spanning $TV/\left.\Gg^\tang\right|_V$. Define
the map $S_{\varnothing} \coloneqq \left(s_1, \ldots, s_{n^\star -
\mu}\right) \mapsto \Phi^\varnothing_{S_\varnothing}(x_0)$ as
\[
\Phi^{\varnothing}_{S_\varnothing}(x_0) = \phi^{v_{n^\star -
\mu}}_{s_{n^\star - \mu}} \circ \cdots \circ \phi^{v_1}_{s_1}(x_0).
\]
Next define $S^\parallel \coloneqq \left(s^\tang_{1}, \ldots,
s^\tang_{\mu}\right) \mapsto \Phi^{\tang}_{S^\tang}(x)$, as
\[
\Phi^\tang_{S^\tang}(x) \coloneqq \phi^{w_{\mu}}_{s^\tang_{\mu}} \circ
\cdots \circ \phi^{w_1}_{s^\tang_{1}}(x),
\]
and the map $S^\tran \coloneqq \left(s^\tran_{0}, \ldots,
s^\tran_{n-n^\star - 2}\right) \mapsto \Phi^\tran_{S^\tran}(x)$, as
\[
\Phi^\tran_{S^\tran}(x) \coloneqq \phi^{g}_{s^\tran_{0}} \circ \cdots
\circ \phi^{ad^{n-n^\star -2}_{f}g}_{s^\tran_{n - n^\star -2}}(x).
\]
Finally, let $s \coloneqq (S_\varnothing, s^\tran_{n - n^\star -1},
S^\tran, S^\tang) \mapsto \Phi_s(x_0)$, with domain a \nbhd $U$ of
$s=0$, be defined as
\begin{equation}
\begin{aligned}
\Phi_s(x_0) \coloneqq \Phi^{\tang}_{S^\tang} \circ \Phi^{\tran}_{S^\tran}
\circ \phi^{ad^{n-n^\star -1}_{f}g}_{s^\tran_{ n - n^\star - 1}}\circ
\Phi^{\varnothing}_{S_{\varnothing}}(x_0).
\end{aligned}
\label{eq:flows}
\end{equation}
Since the vector fields in the list~\eqref{eq:array_info} are linearly
independent near $x_0$, it follows from the inverse function theorem
that there exists a \nbhd $U$ of $s=0$ such that~\eqref{eq:flows} is a
diffeomorphism onto its image. Let
\begin{equation}
\lambda(x) = s^\tran_{n - n^\star - 1}(x).
\label{eq:exoutput}
\end{equation}
We will not show that~\eqref{eq:exoutput} yields a well-defined
relative degree of ${n}-\ns$ at $x_0$ because the arguments are
similar to the proof~\cite[Theorem 3.2]{NieMag08}. Instead, we focus
on showing that there exists a function $\tilde{\lambda}$ such that
$\lambda = \tilde{\lambda}(h(x))$. The function~\eqref{eq:exoutput}
yields relative degree $n -n ^\star$ on near $x_0$, so in particular
\[
L_g{\lambda}(x) = L_{ad_fg}{\lambda}(x) = \cdots =
L_{ad^{n-n^\star-2}_fg}{\lambda}(x) = 0.
\]
Furthermore, since $\Gamma^\star \subset \lambda^{-1}(0)$, these facts imply that,
\[
\begin{aligned}
\D\lambda \in \ann{(T\Gamma^\star)} \cap \ann{({\Gg}_{n - n^\star
-2})} &= \ann{(T\Gamma^\star + {\Gg}_{n - n^\star -2})}\\ &=
\ann{(T\Gamma^\star + \inv{(\Gg_{{n} - n^\star - 2} + W)})}
\\&\subseteq \ann{(T\Gamma^\star + \Gg_{{n} - n^\star - 2} + W)}\\
&= \ann{(T\Gamma^\star)}\cap \ann{(\Gg_{{n} - n^\star - 2} + W)}.
\end{aligned}
\]
Therefore,
\[
\begin{aligned}
\D\lambda &\in \ann{(\Gg_{{n} - n^\star - 2})} \cap \ann{(W)}\\&=
\ann{(\Gg_{{n} - n^\star - 2})} \cap \Sp\{\D h_1, \ldots,
\D h_p\}
\end{aligned}
\]
so
\[
\D\lambda = \sum^{p}_{i=1}\sigma_i(x)\D h_i(x)
\]
which implies that $\lambda = \tilde{\lambda}(h(x))$ and $\sigma_i(x)
= \left.\DER{\tilde{\lambda}}{y_i}\right|_{y = h(x)}$.
\end{proof}
\begin{remark}
When the state $x$ is available for feedback, i.e., $h(x) = x$, $\W
= \set{0}$ and $\Gamma^\star = \set{x_0}$ is an equilibrium point of
the open-loop system, the conditions of Theorem~\ref{thm:main}
coincide with the necessary and sufficient conditions for solving
the state-space exact feedback linearization problem~\cite[Theorem
4.2.3]{Isi95}.
\end{remark}
\begin{remark}
The direct generalization of Theorem~\ref{thm:main} to multi-input,
multi-output systems gives sufficient, not necessary, conditions
under which the MIMO version of LTFLPI is solvable. The MIMO proof
of necessity is an open problem and the subject of future research.
\end{remark}
\subsection*{Example}
We now return to the motivating example of
Section~\ref{sec:motivation} to illustrate the application of
Theorem~\ref{thm:main}. In Section~\ref{sec:motivation} we had a
system of the form~\eqref{eq:system},~\eqref{eq:output} with $f(x) =
\col{\left(x_4, -x_3-x_2^3, x_2, 0, x_1\right)}$, $g(x) =
\col{\left(x_1, 0, 0, 1, x_5\right)}$, $h(x) = \col{\left(x_4,
x_5\right)}$. The target set is given by $\Gamma^\star = \set{x
\in \Real^4 : x_1 = x_4 = x_5 = 0}$. We now use
Theorem~\ref{thm:main} to justify the discussion from
Section~\ref{sec:motivation} and show that LTFLPI is solvable, for the
given system and set, in a \nbhd of the origin. First note that
$n^\star = 1$ and that
$(\forall x \in \Gamma^\star) \; \; T_x\Gamma^\star = \Sp{\set{v_1,
v_2}}(x) = \Sp\{e_2, e_3\}$
where $e_2$ and $e_3$ are the second and third natural basis vectors
for $\Real^5$.
Furthermore
\[
\Gg_{2}(x) = \Sp{\set{g, ad_fg, ad^2_fg}}(x) = \Sp{\set{
\left[
\begin{array}{c}
x_1\\0\\0\\1\\x_5
\end{array}
\right], \;
\left[
\begin{array}{c}
x_4 - 1\\0\\0\\0\\0
\end{array}
\right], \;
\left[
\begin{array}{c}
0\\0\\0\\0\\1 - x_4
\end{array}
\right]}}.
\]
Checking conditions of Theorem~\ref{thm:main} we have that
$T_0\Gamma^\star + \Gg_{2}(0) = T_0\Real^5 \simeq \Real^5$
so condition (a) holds at $x_0 = 0$. In order to check condition (b)
of Theorem~\ref{thm:main} we first write $\W = \ann{\left(\Sp{\left\{\D h_1, \D h_2\right\}}\right)} =
\Sp{\set{w_1, w_2, w_3}}(x) = \Sp\{e_2, e_3, e_1\}$.
The distribution $\Gg_1 + \W$ is regular in a \nbhd of the origin, it
has dimension four is everywhere. Calculating the Lie brackets between
the vector fields $g$, $ad_fg$, $w_1$, $w_2$ shows that, for all $x
\in \Real^5$, $\Gg_1 + \W(x) = \inv{\left(\Gg_1 + \W\right)}(x)$. This
makes verifying that condition (b) of Theorem~\ref{thm:main} holds
easy to check. We conclude that there exists an observable function
such that Theorem~\ref{thm:rel_iff} holds and hence that LTFLPI is
solvable.
In order to actually find the observable output, in this simple case,
one can follow the semi-constructive procedure of
Theorem~\ref{thm:main}. Construct the maps from the proof of
Theorem~\ref{thm:main} noting that, in a \nbhd of $x_0$,
$T\Gamma^\star/ \Gg^\tang = \{0\}$. We have
\[
\Phi^{\tang}_{S^\tang}(x_0) = \phi^{w_2}_{s^\tang_2} \circ
\phi^{w_1}_{s^\tang_1}(x_0) = \col{\left(x_0, s^\tang_1
+x_0, s^\tang_2 + x_0, x_0, x_0 \right)}
\]
and $\Phi^{\tran}_{S_\tran}(x) = \phi^g_{s^\tran_0} \circ
\phi^{ad_fg}_{s^\tran_1}(x)$. Hence the overall map~\eqref{eq:flows}
is given by $\Phi_s(x_0) =\Phi^\tang_{S^\tang} \circ
\Phi^{\tran}_{S_\tran} \circ \phi^{ad^2_fg}_{s^\tran_2}(x_0)$ where
\[
\begin{aligned}
s^\tran_0 &\mapsto \phi^{g}_{s^\tran_0}(x) =
\col{\left(\e^{s^\tran_0}x_1, x_2, x_3, s^\tran_0 + x_4, \e^{s^\tran_0}x_5\right)}\\
s^\tran_1 &\mapsto \phi^{ad_fg}_{s^\tran_1}(x) = \col{\left((x_4
-1)s^\tran_{1} + x_1, x_2, x_3, x_4, x_5\right)}\\
s^\tran_2 &\mapsto \phi^{ad^2_fg}_{s^\tran_2}(x) = \col{\left(x_1,
x_2, x_3, x_4, (1 - x_4)s^\tran_2 + x_5\right)}.
\end{aligned}
\]
and therefore, with $x_0 = 0$,
\[
\Phi_s(x_0) =\Phi^\tang_{S^\tang} \circ \Phi^{\tran}_{S_\tran}
\circ \phi^{ad^2_fg}_{s^\tran_2}(x_0) =
\col{\left(-s_1^\tran\e^{s^\tran_0}, s^\tang_1, s^\tang_2, s^\tran_0, s^\tran_2\e^{s^\tran_0}\right)}.
\]
Inverting this map yields
\[
\col{\left(s^\tang_1(x), s^\tang_2(x),
s_0^\tran(x), s^\tran_1(x), s^\tran_2(x)\right)} =
\col{\left(x_2,x_3, x_4, -x_1\e^{-x_4}, x_5\e^{-x_4}\right)}.
\]
The observable transverse output is the function $\lambda(x) =
s^\tran_2(x) = x_5\e^{-x_4}$. This function allows one to follow the
design procedure in Section~\ref{sec:motivation}.
\section{Extensions and applications}
\label{sec:extensions}
In this section we present extensions and applications of
Theorem~\ref{thm:main}. In Section~\ref{sec:global} we pose the global
transverse feedback linearization problem with partial information in
which, roughly speaking, one seeks a single coordinate transformation
such that~\eqref{eq:system} is equivalent to the normal
form~\eqref{eq:partial_info} in a \nbhd of the entire set
$\Gamma^\star$. We present sufficient conditions for the global
problem to be solvable by restricting the geometry of the target
set. In Section~\ref{sec:disturbances} we illustrate how these results
can be applied to systems affected by disturbances that cannot be
measured.
\subsection{Global transverse feedback linearization with partial information}
\label{sec:global}
The results of Section~\ref{sec:main} are local, valid in a \nbhd of a
point on the target manifold $\Gamma^\star$. In this section we seek a
global solution. By global we mean a solution valid in a \nbhd of
$\Gamma^\star$, not necessarily all of $\Real^n$. The following is a
global version of LTFLPI.
\begin{problem*}{Global Transverse Feedback Linearization with Partial
Information (GTFLPI) Problem }
Given a smooth single-input system~\eqref{eq:system} with smooth
output~\eqref{eq:output}, a closed, connected, embedded,
$n^\star$-dimensional controlled invariant submanifold $\Gamma^\star
\subset \Real^n$, find, if possible, a diffeomorphism $\Xi \in
\Diff{\mathcal{N}}$
\begin{equation}
\begin{aligned}
\Xi : \mathcal{N} &\to \Xi(\mathcal{N}) \subseteq \Gamma^\star \times
\Real^{n -
n^\star}\\
x &\mapsto (\eta, \xi)
\end{aligned}
\end{equation}
where $\mathcal{N}$ is a \nbhd of $\Gamma^\star$, such that
\begin{itemize}
\item[(i)] The restriction of $\Xi$ to $\Gamma^\star$ is
\[
\left.\Xi\right|_{\Gamma^\star} : x \mapsto (\eta, 0).
\]
\item[(ii)] The dynamics of system~\eqref{eq:system} in $(\eta,
\xi)$-coordinates is given by~\eqref{eq:partial_info} where the pair
$(A,b)$ is in Brunovsk\'{y} normal form (one chain of integrators)
and $a_2(\eta, \xi) \neq 0$ in $\Xi(\mathcal{N})$.
\item[(iii)] The first component of $\xi$, denoted $\xi_1$ is
observable, i.e., there exists a function $\tilde{\lambda} :
h(\mathcal{N}) \subseteq \Real^p \to \Real$ such that
\[
\xi_1(x) = \tilde{\lambda} \circ h(x).
\]
\end{itemize}
\end{problem*}
In order to solve GTFLPI we restrict the class of allowable target
sets $\Gamma^\star$.
\begin{assumption}
The set $\Gamma^\star$ is diffeomorphic to a generalized cylinder,
i.e., $\Gamma^\star \simeq \mathbb{T}^k \times \Real^{n^\star - k}$, $k
\in \{0, \ldots, n^\star\}$ where $\mathbb{T}^k$ is the $k$-torus.
\label{ass:cylinder}
\end{assumption}
Assumption~\ref{ass:cylinder} restricts the class of set considered in
the global problem. However, there are many applications, most notably
path following, where the set to be stabilized is a generalized
cylinder~\cite{ConMagNieTos10}. This type of set arises when solving a
path following problem for closed curves.
\begin{proposition}
\label{prop:global_normal_form}
Under Assumption~\ref{ass:cylinder}, GTFLPI is solvable if
\begin{itemize}
\item[(a)] $(\forall x \in \Gamma^\star)$ $T_{x}\Gamma \oplus \Gg_{n-n^\star-1}(x) = T_{x}\Real^{{n}}$
\item[(b)] The distribution $\Gg_{n - n^\star -2}$ is non-singular and
involutive in a \nbhd of $\Gamma^\star$
\item[(c)] The distribution $\inv{\left(\Gg_{{n}-n^\star -2}+
\W\right)}$ is non-singular and, $(\forall x \in \Gamma^\star)$,
\[
\dim{\left(T_x\gstar \oplus \Gg_{{n}-n^\star -2}(x)\right)} =
\dim{\left(T_x\gstar \oplus \inv{\left(\Gg_{{n}-n^\star -2}+
\W\right)}(x)\right)}.
\]
\end{itemize}
\end{proposition}
\begin{proof}
The proof of this result is based on the proof of~\cite[Theorem
4.4]{NieMag06}. Assumption~\ref{ass:cylinder} and hypotheses (a) and
(b) imply that the conditions of~\cite[Theorem 4.4]{NieMag06}
hold. This guarantees the existence of a diffeomorphism with
properties (i) and (ii) in the GTFLPI problem. In particular it
ensures the existence of a function $\lambda : \mathcal{N} \subseteq
\Real^{{n}} \To \Real$ such that $\Gamma^\star \subset
\lambda^{-1}(0) = \set{x \in \mathcal{N} : \lambda(x) = 0}$ and
which yields a uniform relative degree of ${n} - n^\star$ over
$\Gamma^\star$.
Now, assume that (c) holds, then using similar arguments to those in
Theorem~\ref{thm:main} we have
\[
\begin{aligned}
\D\lambda \in \ann{(T\Gamma^\star)} \cap \ann{({\Gg}_{n - n^\star
-2})} &= \ann{(T\Gamma^\star + {\Gg}_{n - n^\star -2})}\\ &=
\ann{(T\Gamma^\star + \inv{\left(\Gg_{{n} - n^\star - 2} + W\right)})}
\\&\subseteq \ann{(T\Gamma^\star + \Gg_{{n} - n^\star - 2} + W)}\\
&= \ann{(T\Gamma^\star)}\cap \ann{(\Gg_{{n} - n^\star - 2} + W)}.
\end{aligned}
\]
Therefore,
\[
\begin{aligned}
\D\lambda &\in \ann{(\Gg_{{n} - n^\star - 2})} \cap \ann{(W)}\\&=
\ann{(\Gg_{{n} - n^\star - 2})} \cap \Sp\{\D h_1, \ldots,
\D h_p\}
\end{aligned}
\]
so
\[
\D\lambda = \sum^{p}_{i=1}\sigma_i(x)\D h_i(x)
\]
which implies that $\lambda = \tilde{\lambda}(h(x))$ with $\sigma_i(x)
= \left.\DER{\tilde{\lambda}}{y_i}\right|_{y = h(x)}$ and therefore
that $\xi_1(x) = \lambda(x)$ is observable.
\end{proof}
The next result is partly motivated by the results
in~\cite{MarPraIsi07} as discussed in
Section~\ref{sec:problem}. Recall that once a system is represented in
the normal form~\eqref{eq:partial_info}, the approach one uses to
stabilize $\Gamma^\star$ using output feedback depends on the dynamics
of the tangential subsystem~\eqref{eq:tang}. When the tangential
dynamics do not necessarily converge to zero but remain bounded, one
may use the results in~\cite{MarPraIsi07}. However,
in~\cite{MarPraIsi07} the $\eta$-dynamics in~\eqref{eq:partial_info}
must only depend on the tangential states $\eta$ and the observable
transversal state $\xi_1$. The next result gives sufficient conditions
for this to be the case.
\begin{corollary}
If, in addition to the hypothesis of
Proposition~\ref{prop:global_normal_form},
\begin{itemize}
\item[(d)] For all $i, j \in \set{0, \ldots, n - n^\star - 1}$,
$\liebr{ad^i_fg}{ad^j_fg} = 0$
\end{itemize}
then there exists a diffeomorphism solving the GTFLPI problem such
that the dynamics of system~\eqref{eq:system} in $(\eta,
\xi)$-coordinates is given by
\begin{equation}
\begin{aligned}
\dot{\eta} &= f_0(\eta, \xi_1)\\
\dot{\xi} &= A\xi + b(a_1(\eta, \xi) + a_2(\eta, \xi)u),
\end{aligned}
\label{eq:partial_info_special}
\end{equation}
with $(A, b)$ a controllable pair in Brunovsk\'{y} normal
form.
\label{cor:simple_tang}
\end{corollary}
\begin{proof}
By Proposition~\ref{prop:global_normal_form} there exists a
real-valued function $\lambda : \mathcal{N} \subseteq \Real^{{n}} \To
\Real$ with $\Gamma^\star \subset \lambda^{-1}(0)$ and yields a
uniform relative degree of ${n} - n^\star$ over $\Gamma^\star$. The
remainder of the claim follows from the proof of~\cite[Proposition
9.1.1]{Isi95} and the assumption that (d) holds. In this case, since
the vector fields $ad^i_fg$, $i \in \set{1, \ldots, n - n^\star
- 1}$ are not assumed to be complete, we are only assured to
obtain a diffeomorphism with the required properties in a \nbhd of
$\Gamma^\star$.
\end{proof}
\subsection{Systems with unobservable disturbances}
\label{sec:disturbances}
Theorem~\ref{thm:main} finds applications on systems affected by
disturbance signals that cannot be measured. Consider the single-input
smooth control system
\begin{equation}
\label{eq:SISOexo}
\begin{aligned}
\dot{x} &= f(x, w) + g(x, w)u\\
\dot{w} &= s(w),
\end{aligned}
\end{equation}
with $x \in \Real^n$, $w \in \Real^k$ and control input $u \in
\Real$. The state $w$ can be thought of as an unobservable disturbance
generated by the dynamical system modeled by $\dot{w} = s(w)$. The
variable $x$ is assumed measured. Denote $\bar{n} \coloneqq n + k$ and
let $q \coloneqq \col{(x, w)}$. Re-writing~\eqref{eq:SISOexo} as
\begin{equation}
\dot{q} = F(q) + G(q)u
\label{eq:q_model}
\end{equation}
with $F(q) \coloneqq \col{(f(x,w), s(w))}$, $G(q) \coloneqq
\col{(g(x,w), 0)}$ and with output
\[
y = H(q), \qquad H(q) = \left[\begin{array}{cc}I_{n \times n} & 0_{n \times
k}\end{array}\right]q,
\]
we immediately have the following corollary.
\begin{corollary}
\label{cor:disturbances}
Suppose that $\Gamma^\star \subset \Real^{\bar{n}}$ is a closed,
connected, embedded, $n^\star$-dimensional controlled invariant
submanifold for~\eqref{eq:SISOexo}. Let $q_0 = (x_0, w_0) \in
\Gamma^\star$ and
\[
\W \coloneqq \ann{\left(\Sp{\left\{\D H_1, \ldots , \D H_n\right\}}\right)}
= \Sp{\left\{\DER{}{w_1}, \ldots , \DER{}{w_{k}}\right\}}.
\]
If the hypotheses of Theorem~\ref{thm:main} hold, then there exists a
diffeomorphism
\[
\begin{aligned}
\Xi : &U \To \Xi(U)\\
&q \mapsto (\eta, \xi)
\end{aligned}
\]
where $U$ is a \nbhd of $q_0$ in $\Real^{\bar{n}}$, such that
system~\eqref{eq:SISOexo} is locally diffeomorphic
to~\eqref{eq:partial_info}. Moreover, the set $\Gamma^\star \cap U$ in
$(\eta, \xi)$-coordinates is given by
\[
\Xi(\Gamma^\star \cap U) = \set{(\eta, \xi) \in \Xi(U): \xi = 0}.
\]
Finally, the first component of $\xi$, denoted $\xi_1$ is observable.
\end{corollary}
As an application of these ideas, consider the path following problem
in presence of disturbances. Path following problems can naturally be
cast as set stabilization problems~\cite{NieMag06} and transverse
feedback linearization has been effectively used to implement path
following controllers~\cite{NieFulMag10}. When the control system is
affected by unobservable disturbances, the results in this paper are
more suitable to solving the path following problem.
\subsection*{Example}
Consider a kinematic unicycle with unit translational velocity
affected by an unobservable disturbance $w$ described by the model
\begin{equation}
\left[\begin{array}{c}\dot{x}_1 \\ \dot{x}_2 \\
\dot{x}_3\end{array}\right] =
\left[\begin{array}{c}\cos{(x_3)}\\ \sin{(x_3)}\\
0\end{array}\right] +
\left[\begin{array}{c}0\\0\\1\end{array}\right]u +
\left[\begin{array}{cc}0 & 0\\ 1 & 0 \\ 0 & 0\end{array}\right]w.
\label{eq:unicycle}
\end{equation}
Suppose that the disturbance $w$ is generated by the exosystem
\begin{equation}
\left[\begin{array}{c}\dot{w}_1\\\dot{w}_2\end{array}\right] =
\left[\begin{array}{r}w_2\\-w_1\end{array}\right].
\label{eq:disturbance}
\end{equation}
We assume that the disturbance satisfies, for all $t \in \Real$,
$\|w(t)\| < 1$. Our objective is to design a control law that makes
the position $(x_1, x_2)$ of the unicycle converge to the unit circle
in the $(x_1, x_2)$-plane. The measured states are $x_1$, $x_2$,
$x_3$.
In the state space of~\eqref{eq:unicycle},~\eqref{eq:disturbance} we
view the goal set as the maximal controlled invariant set contained in
$\{(x,w) \in \Real^3 \times \Real^2: x_1^2 + x_2^2 - 1 = 0\}$ which is
given by
\begin{equation}
\Gamma^\star = \{(x, w) \in \Real^{3} \times \Real^2: x_1^2 + x_2^2 - 1 =
x_1\cos{(x_3)} + x_2\sin{(x_3)} + x_2w_1= 0\}.
\label{eq:gstar_ex}
\end{equation}
Here, $n^\star = \dim{(\gstar)} = 3$. Writing $\Gamma^\star =
\gamma^{-1}(0)$ for the function $\gamma : \Real^3 \times \Real^2 \to
\Real^2$ defined by~\eqref{eq:gstar_ex}, one can verify that, under the
assumption that $\|w\| < 1$, zero is a regular value of $\gamma$. The
set~\eqref{eq:gstar_ex} has two connected components, one
corresponding to clockwise motion along the path, the other
corresponding to counterclockwise motion. Since our results are local,
a choice of $(x_0, w_0) \in \Gamma^\star$ selects one of these two
components.
To apply Corollary~\ref{cor:disturbances}, set $q \coloneqq \col{(x,
w)}$, and define $F(q) = \col{\left(\cos{(x_3)}\right.}$,
$\left.\sin{(x_3)} + w_1,0,w_2,-w_1\right)$, $G(q) =
\col{\left(0,0,1,0,0\right)}$, $H(q) = \col{\left(x_1, x_2,
x_3\right)}$. The question we ask is: can transverse feedback
linearization be used to stabilize $\Gamma^\star$ using only $y$ for
feedback? The answer, of course, is yes. For, one can check that the
observable function $\lambda(q) = x_1^2 + x_2^2 - 1$ yields a
well-defined relative degree of $2 = n - \ns$ at each $q \in
\Gamma^\star$. If we set $\xi_1(q) = \lambda(q)$, $\xi_2(q) =
L_F\lambda(q)$ and let $\eta_1, \eta_2, \eta_3$ be any\footnote{If the
functions $\eta_1$, $\eta_2$, $\eta_3$ are chosen so that $\D\eta_1,
\D\eta_2, \D\eta_3 \in \ann{(G(q))}$, then we obtain the normal
form~\eqref{eq:partial_info}. It is always possible to do this. If,
on the other hand, $\D\eta_1, \D\eta_2, \D\eta_3 \not \in
\ann{(G(q))}$, then the control $u$ will appear in the $\eta$
subsystem.} three additional linearly independent functions, then,
the local diffeomorphism $\Xi(q) = \col{(\eta(q), \xi(q))}$ solves
LTFLPI.
Next we confirm this observation using
Corollary~\ref{cor:disturbances}. In this case, with $q_0 = (x_0,
w_0)$, conditions $(a)$ and $(b)$ of Theorem~\ref{thm:main} become
\begin{itemize}
\item[$(a)$] $\dim{(T_{q_0}\Gamma^\star + \Gg_{1}(q_0))} = 5$
\item[$(b)$] there exists an open \nbhd $U$ of $q_0$ in
$\Real^5$ such that
\[
(\forall \; q \in \gstar\cap U) \; \dim(T_{q}\Gamma^\star +
\Gg_{0}(q)) = \dim(T_{q}\Gamma^\star + \inv{\left({\Gg}_{0} + \W\right)}(q))=
\text{ constant,}
\]
\end{itemize}
where $\W$, in this case, is given by $\W = \ann{\left(\Sp\{\D H_1, \D
H_2, \D H_3\}\right)} = \Sp\{e_4, e_5\}$
where $e_4$ and $e_5$ are the fourth and fifth natural basis vectors
for $\Real^5$. At each $q \in \Gamma^\star$ we have that
\[
T_q\Gamma^\star =
\Sp{\left\{\left[\begin{array}{c}0\\0\\0\\0\\1\end{array}\right],
\left[\begin{array}{r}x_2^2\\-x_1x_2\\0\\\phi(q)\\0\end{array}\right],
\left[\begin{array}{r}-x_2\varphi(q)\\x_1\varphi(q)\\\phi(q)\\0\\0\end{array}\right]
\right\}}
\]
where $\phi(q) \coloneqq x_1\left(\sin{(x_3)} + w_1\right) -
x_2\cos{(x_3)}$, $\varphi(q) \coloneqq x_1\sin{(x_3)} - x_2\cos{(x_3)}$
are both non-zero on $\Gamma^\star$ if $\|w\| < 1$. Therefore
\[
T_q\Gamma^\star + \Gg_1(q)=
\Sp{\left\{\left[\begin{array}{c}0\\0\\0\\0\\1\end{array}\right],
\left[\begin{array}{r}x_2^2\\-x_1x_2\\0\\\phi(q)\\0\end{array}\right],
\left[\begin{array}{r}-x_2\varphi(q)\\x_1\varphi(q)\\\phi(q)\\0\\0\end{array}\right],
\left[\begin{array}{c}0\\0\\1\\0\\0\end{array}\right],
\left[\begin{array}{c}-\sin{(x_3)}\\\cos{(x_3)}\\0\\0\\0\end{array}\right]
\right\}}.
\]
The determinant of the matrix whose columns are the above basis
vectors is given by $\phi(q)\varphi^2(q)$.
If this determinant is non-zero then condition (a) holds. To this end,
note that on the set~\eqref{eq:gstar_ex}
\[
\begin{aligned}
&x_1\cos{(x_3)} + x_2\sin{(x_3)} + x_2w_1= 0\\
\Rightarrow & x_2\left(\sin{(x_3)} + w_1\right) =
-x_1\cos{(x_3)}\\
\Rightarrow & x^2_2\left(\sin{(x_3)} + w_1\right)^2 =
x_1^2\cos^2{(x_3)}\\
\Rightarrow & x^2_2\left(\sin{(x_3)} + w_1\right)^2 =
(1 - x_2^2)\cos^2{(x_3)} \qquad (\text{since $x_1^2 + x_2^2 - 1 =
0$})\\
\Rightarrow & x_2 = \pm \frac{\cos{(x_3)}}{\sqrt{\left(\sin{(x_3)} +
w_1\right)^2 + \cos^2{(x_3)}}}.
\end{aligned}
\]
Similarly, using $x_1^2 = 1 - x_2^2$,
\[
x_1 = \mp \frac{\sin{(x_3)} + w_1}{\sqrt{\left(\sin{(x_3)} +
w_1\right)^2 + \cos^2{(x_3)}}}.
\]
Therefore, we have that
\[
\phi(q) = \mp \frac{1 + 2w_1\sin{(x_3)} + w_1^2}{\sqrt{\left(\sin{(x_3)} + w_1\right)^2 +
\cos^2{(x_3)}}}, \qquad \varphi(q) = \mp \frac{1 +
w_1\sin{(x_3)}}{\sqrt{\left(\sin{(x_3)} + w_1\right)^2 + \cos^2{(x_3)}}}
\]
which, under the assumption that $\|w\| < 1$, are both non-zero.
To check condition (b) we first note that since $\Gg_0$ and $W$ are
constant distributions, $\Gg_0 + \W = \inv{\left(\Gg_0 +
\W\right)}$. We have that, for any $q \in \Gamma^\star$,
\[
T_q\Gamma^\star + \Gg_0(q)=
\Sp{\left\{\left[\begin{array}{c}0\\0\\0\\0\\1\end{array}\right],
\left[\begin{array}{r}x_2^2\\-x_1x_2\\0\\\phi(q)\\0\end{array}\right],
\left[\begin{array}{r}-x_2\varphi(q)\\x_1\varphi(q)\\\phi(q)\\0\\0\end{array}\right],
\left[\begin{array}{c}0\\0\\1\\0\\0\end{array}\right]\right\}},
\]
and
\[
T_q\Gamma^\star + \inv{(\Gg_0 + \W)}(q)=
\Sp{\left\{\left[\begin{array}{c}0\\0\\0\\0\\1\end{array}\right],
\left[\begin{array}{r}x_2^2\\-x_1x_2\\0\\\phi(q)\\0\end{array}\right],
\left[\begin{array}{r}-x_2\varphi(q)\\x_1\varphi(q)\\\phi(q)\\0\\0\end{array}\right],
\left[\begin{array}{c}0\\0\\1\\0\\0\end{array}\right],
\left[\begin{array}{c}0\\0\\0\\1\\0\end{array}\right]\right\}}.
\]
Since, as already shown, when $\|w\| < 1$ the functions $\varphi$ and
$\phi$ are non-zero on $\Gamma^\star$ and since $x_1$ and $x_2$ are
not equal to zero simultaneously, condition (b) is satisfied.
In $(\eta, \xi)$-coordinates the unicycle has the normal
form~\eqref{eq:partial_info}. At this point a ``high-gain'' output
feedback controller can be used to make the target set $\Gamma^\star$
attractive and have the unicycle traverse the desired path in the
presence of unobservable disturbances.
\section{Conclusions}
In this paper we studied the problem of stabilizing a controlled
invariant embedded submanifold in the state space of autonomous
nonlinear control systems using output feedback. We studied the most
natural approach to solving this problem : given a controlled
invariant manifold (the target set), find an observable output
function yielding a well-defined relative degree whose associated zero
dynamics manifold locally coincides with the target set. We call this
the local transverse feedback linearization problem with partial
information. Necessary and sufficient conditions were presented under
which this problem is solvable. We also presented a global solution to
this problem in the case when the target set is a generalized
cylinder. Finally we illustrated how this work may find applications
in system affected by unobservable disturbances and to path following
problems.
\bibliographystyle{siam}
|
\section{Introduction}
\label{sec:intro}
There are many possible definitions for a Smart Grid (SG). The SG European Technology Platform defines \cite{smartgrids_ETP} a SG as an ``electricity network that can intelligently integrate the actions of all the connected users, generators, consumers and those that do both, in order to efficiently deliver sustainable economic and secure electricity supply.''
A SG employs innovative products and services together with intelligent monitoring, control, communication, and self-healing technologies in order to (i) facilitate the connection and operation of generators of all sizes and technologies; (ii) allow consumers to play an active role in optimizing the operation of the system; (iii) significantly reduce the environmental impact of the whole electricity supply system; (iv) preserve or improve the level of system reliability, quality of service, and security.
SGs can be considered as an ``evolution'' rather than a ``revolution'' of the existing energy networks \cite{Energy_Information_Administration}. The evolution is leaded by the symbiotic exchange between power grid technologies and the Information and Communication Technologies (ICT). ICT provide instruments, such as \textit{Smart Sensors} (SS), to monitor the network status, wired and wireless communication network to collect and transport data, and powerful computational architectures for data processing.
A SG can be framed as a complex non-linear and time-varying system \cite{dorfler2013synchronization,amin2005toward,5535240,Venayagamoorthy__2009,mei_power_2011,machowski2011power}, where heterogeneous elements, including exogenous factors, are extremely interconnected through the exchange of both energy and information.
Computational Intelligence (CI) techniques offer sound modeling and algorithmic solutions in the SG context \cite{5179088,5952102,abido2014computational}.
Well-known CI techniques adopted in the SG context include approximate dynamic programming \cite{5768096}, neural networks and fuzzy inference systems for prediction and control \cite{6111641,molderink2010management}, and swarm intelligence and evolutionary computation for optimization problems \cite{de2013genetic,6608435,abdelaziz2009distribution}.
An important key issue of SGs is the design of a Decision Support System (DSS), which is an expert system that provides decision support for the commanding and dispatching systems of the power grid.
Such a system analyzes the risk for damage of crucial equipments, assesses the power grid security, forecasts and provides warnings about the magnitude and location of possible faults, and timely broadcasts the early-warning signals through suitable communication networks \cite{mei_power_2011}.
The information provided by the DSS can be used for Condition Based Maintenance (CBM) in the power grid \cite{Raheja2006-J-IJPR}.
CBM is defined as ``a philosophy that posits repair or replacement decisions on the current or future condition of assets''.
The objective of CBM is thus to minimize the total cost of inspection and repair by collecting and interpreting (heterogeneous) data related to the operating condition of critical components.
Through the use of CBM, advanced SS technology has the potential to help utilities to improve the power grid reliability by avoiding unexpected outages.
A discussion on how the changes in modern power grids have affected the maintenance procedures can be found in \cite{1600559}; the importance of modern diagnostic techniques is treated in \cite{6039785}.
Collecting heterogeneous measurements in modern SG systems is of paramount importance. As an instance, the available measurements can be used for dealing with various important pattern recognition and data mining problems on SGs, such as fault recognition \cite{6175733,Zhang2011791,Saha2011887}.
On the basis of the data at hand, different problem types could be formulated.
In \cite{1645199} the authors have established a relationship between environmental features and fault causes. A fault cause classifier based on the linear discriminant analysis (LDA) is proposed in \cite{5275689}. Information regarding weather conditions, longitude-latitude information, and measurements of physical quantities (e.g., currents and voltages) related to the power grid have been taken into account. In \cite{5156572}, the authors proposed a system based on LDA, which processes phasor measurement unit data, with the aim of recognizing and locating faults on power lines.
As concerns fault diagnosis in power grids, in \cite{5234528} is proposed a Support Vector Machine (SVM) based method to perform the recognition of faults related to high-voltage transmission lines.
The One-Class Quarter-Sphere SVM algorithm is proposed \cite{6477812} for faults classification in the power grid. The reported experimental evaluation is however performed on synthetically generated data only.
ACEA is the electricity distribution company managing the electrical network feeding the entire city of Rome. In this paper, we extend our previous work \cite{enrico_occ} on the problem of modeling and recognizing faults in the real-world SG system of ACEA by introducing several improvements.
Initially, we introduce the application's context and the approach followed to implement the one-class classification system used to recognize conditions of fault.
Since the available ACEA data is highly structured (i.e., it is formed by several heterogeneous information), we designed a dedicated one-class classifier (OCC) that is suitable for the specific application at hand.
The first herein presented improvement consists in equipping the designed OCC with the capability of producing also \textit{soft} output decisions. This is implemented by interpreting the decision regions synthesized by the classifier as fuzzy sets with suitable membership functions \cite{Livi_ga_2013,eocc__arxiv,pedrycz1998introduction}.
This fact allows us to provide also a measure of reliability concerning the already implemented hard classification mechanism.
As concerns the experiments, we provide (i) several evaluations of the recognition systems on either ad-hoc synthetic and ACEA datasets, (ii) a comparison on some well-known UCI datasets \cite{Bache+Lichman:2013} with other state of the art OCCs, and finally (iii) a more in depth analysis of the informativeness of the solutions found by the proposed OCC on the ACEA data.
The paper is structured as follows.
We offer a brief review on the one-class classification setting in Sec. \ref{sec:occ_prob}.
In Sec. \ref{sec:SG_proj} we provide a short overview about the main project where this study is collocated.
In Sec. \ref{sec:SG}, we describe the technical details of the considered SG.
Sec. \ref{sec:system} introduces the fault recognition system that we designed for the specific application at hand.
In this section, we describe (i) the representation of a fault pattern instance and (ii) the computational system as a whole highlighting also the new contributions introduced here in this paper.
In Sec. \ref{sec:experiments} we discuss the experiments.
Finally, in Sec. \ref{sec:conclusions} we draw our conclusions.
\section{Brief Overview on the One-class Classification Problem}
\label{sec:occ_prob}
The one-class classification problem can be considered as a particular instance of a standard $n$-class classification problem, where, during the training stage, patterns belonging only to a specific class are available. Such patterns are usually termed \textit{target} or \textit{positive} patterns.
This particular scenario covers several interesting real-world situations \cite{Ding2014313,one-class_survey__2010,oilspill__2010,Kemmler201329,enrico_occ,Utkin:2012:FOC:2213741.2433967,NIPS2002_2163}.
Practically, OCCs define a decision rule on the basis of a model that is able to describe suitable \textit{boundaries} pertaining the target patterns. Such boundaries define the decision regions/surface of the classifier. The aim is to synthesize effective models such that target patterns are recognized while non-target patterns are rejected.
\citet{one-class_survey__2010} provided a recent survey on the subject of one-class classification.
One important class of OCCs has been elaborated from the well-known SVM \cite{Tax19991191,SchWilSmoShaetal00,Wang2013875}.
\citet{Tax19991191} defined well-known system called Support Vector Data Description (SVDD).
The classification model is defined in terms of hyper-spheres, which are placed over the training set through an SVM-like optimization problem (the minimization of the sphere radiuses is enforced).
SVDD can be extended to different input domains by defining suitable positive definite kernel functions.
\citet{SchWilSmoShaetal00} proposed an alternative approach to SVDD that employs a hyperplane, like in the conventional SVM case.
The hyperplane is synthesized towards the aim of separating the region of the input space containing (target) patterns form the region containing no data. Also this approach has the capability of using kernel functions.
Other more recent approaches include algorithms based on the minimum spanning tree \cite{Juszczak20091859} and on Gaussian processes \cite{Kemmler2013}.
\section{An Overview on the ACEA Smart Grid Project}
\label{sec:SG_proj}
The following work represents a branch of a larger project, namely ``ACEA Smart Grid Pilot Project'' \cite{ACEA_SG_Pilot_Proj}.
The project aims to develop an automated recognition tool of fault states in the ACEA power grid.
In addition, the tool is designed to offer also diagnostic features, allowing the characterization of the power grid status during fault events.
The process flow diagram depicted in Fig. \ref{fig:Process_flow} shows the overall system and how raw data coming from the SS are transformed into meaningful information in order to support business strategies.
To obtain the dataset herein elaborated, a preliminary preprocessing stage, operated together with the ACEA experts, is performed. The dataset is then used as input for the herein presented OCC, which by means of an evolutionary strategy is in charge of learning typical situations of faults.
Clustering techniques are used to define the model of the proposed OCC. The synthesized partition is used also for post-processing purposes, such as data analysis and visualization. Those last two post-processing stages, belonging to the work packages set of the overall project, are not discussed in this paper.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 855 347,scale=0.4,keepaspectratio=true]{./Process_flow}
\caption{Process flow diagram describing the``ACEA Smart Grid Project''.}
\label{fig:Process_flow}
\end{figure}
\section{The Considered Power Grid}
\label{sec:SG}
The considered grid is constituted of backbones of uniform section, exerting radially with the possibility of counter-supply if a branch is out of order. Each backbone of the power grid is supplied by two distinct Primary Stations (PS) and each half-line is protected against faults through the breakers.
The Medium Voltage (MV) power grid consists in lines (feeders) in which the nominal voltage is 20 kV, with the presence of few ``legacy'' lines that still work at 8.4 kV. The MV part of the network covers 10,490 km while the Low Voltage (LV) section covers 11.120 km.
Cables can be on air or underground and their sections can vary along the backbone with the presence of bottlenecks.
The MV section has 1.565 lines in service and it is supplied with 76 PSs, while LV section is supplied with 13.292 Secondary Stations (SS). Each MV line feeds a given number of secondary stations (MV-LV transformers), each one provided of two breakers, so that the substation can be isolated from the feeder in case of fault. These breakers on the medium voltage busbar are used to assure the radiality condition of the network, stating that each substation must be fed by only one PS.
We deal with the problem of modeling and recognizing a particular type of faults, which are commonly termed as localized faults (LFs) \cite{Acea_Road_mapr_2010}. Before providing a precise definition for a LF, it is important to discriminate among \textit{outages} and \textit{faults}, according to the CEI 5160 normative \cite{Doe:2013:Online}.
An outage (i.e., an interruption of the service) is the condition in which the voltage on the access point to the electrical energy of a user is less than 5\% of the declared voltage on all phases of supply \cite{Acea_Road_mapr_2010}.
Three types of outage are considered: (i) long, if the duration is more than three minutes (long outages); (ii) short, if the duration is more than one second and less than three minutes (short outages); (iii) transient, if the duration does not exceed one second (transient outages).
A fault, instead, is related to the failure of the electrical insulation (e.g., insulation of cables) that compromises the correct functioning of the grid.
Therefore, a LF is effectively a fault in which a physical element of the grid is permanently damaged causing long outages.
\section{The Proposed One-class Classification System for Smart Grid Fault Detection}
\label{sec:system}
\subsection{Representation of a Fault Pattern}
\label{sec:pattern_representation}
Instances of fault patterns (FPs) describing LFs occurred in the SG have been elaborated from a historical database provided by ACEA. The considered period spans across 2009--2012. During this time period the electrical network was not provided of any ICT infrastructure to record in a proper data base faults events, neither normal working conditions. Faults were recorded by manual entry following a given protocol. As a consequence these fault records are the only available data for this study.
Faults are characterized by heterogeneous data, including weather conditions, spatio-temporal data (i.e., longitude-latitude pairs and timestamps), physical data related to the state of power grid and its electric equipments (e.g., measured currents and voltages), and finally meteorological data.
As a consequence, a FP is actually defined by features of different nature, containing categorical (nominal) data, quantitative data (i.e., data having a well-defined metric), and also Time Series (TS) describing short outages occurred before a LF.
A detailed description of the considered features characterizing a FP is provided in Tab. \ref{tab:features_description}.
\begin{table*}[htbph!]\scriptsize
\centering
\caption{Description of the considered features describing a FP instance.}
\begin{tabular} {|p{4.5cm}|p{3.5cm}|p{5cm}|}
\hline
\rowcolor{lgray} \bf{Feature} & \bf{Data typology and features space label} & \bf{Description}\\
\hline
\bf{Day start} & Quantitative (Integer) $\mathcal{F}_{D}$ & Day in which the LF was detected \\
\hline
\bf{Time start} & Quantitative (Integer) $\mathcal{F}_{T}$ & Time stamp (minutes) in which the LF was detected\\
\hline
\bf{Location element}& Categorical (String) $\mathcal{F}_{1}^{C}$ & Element positioning (aerial or underground) \\
\hline
\bf{Material} & Categorical (String) $\mathcal{F}_{2}^{C}$ & Constituent material element (CU, AL) \\
\hline
\bf{Primary station fault distance} & Quantitative (Real) $\mathcal{F}_{1}^{Q}$ & Distance between the primary station and the geographical location of the LF \\
\hline
\bf{Median point} & Quantitative (Real) $\mathcal{F}_{2}^{Q}$ & Fault location calculated as median point between two secondary stations \\
\hline
\bf{\# Secondary Stations (SS)} & Quantitative (Real) $\mathcal{F}_{3}^{Q}$ & Number of out of service secondary stations after the LF \\
\hline
\bf{Current out of bounds} & Quantitative (Integer) $\mathcal{F}_{4}^{Q}$ & The maximum operating current of the backbone is less than or equal to 60\% of the threshold ``out of bounds'', typically established at 90\% of capacity \\
\hline
\bf{Max. temperature} & Quantitative (Real) $\mathcal{F}_{5}^{Q}$ & Maximum registered temperature \\
\hline
\bf{Min. temperature} & Quantitative (Real) $\mathcal{F}_{6}^{Q}$ & Minimum registered temperature \\
\hline
\bf{Delta temperature section} & Quantitative (Real) $\mathcal{F}_{7}^{Q}$ & Difference between the maximum and minimum temperature \\
\hline
\bf{Rain} & Quantitative (Real) $\mathcal{F}_{8}^{Q}$ & Millimeters of rain calculated as the average two hours preceding to the LF \\
\hline
\bf{Cable section} & Quantitative (Real) $\mathcal{F}^{S}$ & Section of the cable, if applicable \\
\hline
\bf{Backbone electric current} & Quantitative (Real) $\mathcal{F}^{EC}$ & Function of electric current that flows in a backbone \\
\hline
\bf{Interruption (breaker)} & TS (Integer) $\mathcal{F}_{1}^{TS}$ & Outages caused by the opening of the \textit{breakers} in the primary station \\
\hline
\bf{Petersen alarms} & TS (Integer) $\mathcal{F}_{2}^{TS}$ & Alarms detected by the device called ``Petersen's coil'' due to loss of electrical insulation on the power line \\
\hline
\bf{Saving intervention} & TS (Integer) $\mathcal{F}_{3}^{TS}$ & Decisive interventions of the Petersen's coil which have prevented the LF \\
\hline
\end{tabular}
\label{tab:features_description}
\end{table*}
\subsection{Data Description and Preprocessing}
\label{sec:Data_Preprocessing}
Data normalization is a universally important aspect in pattern analysis, which becomes even more crucial when processing patterns characterized by many heterogeneous features.
The numerical data provided by ACEA have been normalized using the affine normalization technique:
\begin{equation}
\label{eq:Affine_normalization}
v=\frac{c-m}{\left (M-m \right )} \in [0, 1].
\end{equation}
\noindent where $c$ is the original (non-normalized) value; $m$ and $M$ are, respectively, the minimum/maximum values for the specific feature in the considered dataset.
\subsubsection{Temporal Data}
\label{sec:Temporal_Data}
The ``Day start'' and ``Time start'' features (Tab. \ref{tab:features_description}) have been encoded as integer values. The former ranges in $\{0, 1, ..., 364\}$, while the latter in $\{0, 1, ..., 1439\}$, corresponding to the number of minutes in a year.
Normalization of such data follows straightforwardly.
\subsubsection{Spatial Data}
\label{sec:Spatial_Data}
Three types of information regarding the geographical position of a LF are available: the absolute position of the PS where the LF has occurred, and the absolute position of the two SSs delimiting the section of power line where the LF has been detected.
The original coordinates of the geographical position of the LF have been expressed in WGS84 (decimal degrees), the same that it is used in the GPS geolocalization system.
It is reasonable that the information regarding the PSs positions and the absolute locations of the LFs can provide indirectly the information about the amount of electric current flowing in the power line. The main hypotheses that led us to that statement are: (i) the MV lines have a radial distribution with respect to the PSs and their extension is of the order of kilometers, (ii) the portion of power line between the two SSs has an extension of hundreds of meters.
In addition, the power grid has a meshed structure and it is ``radially'' distributed, so that every MV line is supplied through only one PS.
In Fig. \ref{fig:Backbone_Scheme} is depicted a typical scenario: a MV backbone composed by two lines supplied through two distinct PSs; the cutting point is situated in (roughly) the middle point.
From the Kirchhoff's second law, the current $I_{PS\ A-6}$ that the PS denoted by A provides to the MV lines to feed nodes from 2 to 6 is equal to the sum of the currents provided by each substation before the cutting point:
\begin{equation}
\label{eq:Kirchoff}
I_{PS\ A-6}=I_{2}+I_{3}+I_{4}+I_{5}+I_{6}.
\end{equation}
The intensity of the electric current in the MV feeder decreases as we move away from the PS, until the open breaker is reached, given the radial topology of the network, we can roughly evaluate the fault location as the median point between of the line segment connecting the two SSs soon before and soon after the localized fault. Moreover, the distance between the PS location and the fault location can be considered as an indirect measure of the total current provided by the PS in the fault condition. -- see Fig. \ref{fig:Hyp_PS_SS} for a graphical representation.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 564 137,scale=0.55,keepaspectratio=true]{./Backbone_Scheme1}
\caption{Scheme of a MV backbone.}
\label{fig:Backbone_Scheme}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 443 296,scale=0.55,keepaspectratio=true]{./Hyp_PS_SS}
\caption{Radial structure of the ACEA grid from the PS and the approximated distance computation (length of red arrow) adopted to reduce the number of available features.}
\label{fig:Hyp_PS_SS}
\end{figure}
The maximum spatial resolution of the geographical localization of the LF is therefore defined by the two SS positions.
The distance between two geographical locations is calculated through the Vincenty's algorithm \cite{vincenty_direct_1975}.
The normalization process of the position data is based on the calculation of the largest rectangle including all the PS and SS stations. (see Fig. \ref{fig:Rome_area} for an example).
Hence, applying the affine normalization (\ref{eq:Affine_normalization}), the spatial positions of the LFs are normalized in $[0, 1]$.
The affine normalization is applied also for the distance values among PSs and the positions of the LFs.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 919 483,scale=0.35,keepaspectratio=true]{./Rome_map}
\caption{Covered are of Rome and the quadrilateral normalization.}
\label{fig:Rome_area}
\end{figure}
\subsubsection{Physical Data}
\label{sec:Physical_Data}
The data describing the physical elements of the power grid is defined by both categorical and quantitative information.
The normalization of quantitative data (i.e., ``\# Secondary station'', ``Current out of bounds'', and ``Cable section'') is implemented by means of (\ref{eq:Affine_normalization}).
It is well known that a possible cause of faults in distribution systems affecting cables and joints is the abrupt change in the current loads, more than the actual amount of electrical power flowing in the devices. To define a feature taking this effect into account, current measures sampled every 10 minutes have been considered in a main window of 24 hours before the LF occurrence. This time window is divided in two non overlapping sub-windows, $w_{1}$ and $w_{2}$, each of 12 hours. The feature ``Backbone Electric Current'' is finally computed as the absolute difference between the average of current values in each sub-window.
This value is normalized with respect to the possible minimum and maximum values.
Clearly, this single real value is a lossy compression of the information conveyed by the whole TS in a single numerical value. However, we performed this simplification in order to capture the average information about the fluctuations of the values of the electric current observed before the LF.
Future dedicated research works will be focused to the study of the TSs of electric current and their relation/causation with the observed LFs.
\subsubsection{Meteorological Data}
\label{sec:Meteorological_Data}
The meteorological data are acquired by suitable stations located in different areas of Rome. The ``Rain'' feature is calculated as the average millimeters of rain observed in the 2 hours soon before the LF occurrence.
\subsubsection{Short Outages Data}
\label{sec:Time_Series}
Here we describe the data related to the short outages observed before a LF.
With an abuse of notation, we will refer to such sequences of events as time series, although formally such sequences are not sampled with a predefined constant period (those events are registered as they occur).
We consider three types of events that can be associated to the ``short outages'' type (see Sec. \ref{sec:SG}). The considered TS of events are: ``Interruption (breaker)'', the ``Petersen alarms'', and the ``Saving intervention'' (see Tab. \ref{tab:features_description} for details).
The short outages events are represented as variable-length sequences, which contain the temporal distances (expressed in seconds) from the subsequent LF (see Fig. \ref{fig:Time_Series_rapres} for a graphical example).
The time window in which those events fall spans across three months (i.e., we search in the three months preceding a LF).
A TS $S^{i}$ of $K_{i}$ outage events is defined as follows:
\begin{equation}
\label{eq:TS_set}
S^{i}=\left [ \xi _{1}^{i},\xi _{2}^{i}, ..., \xi _{K_{i}}^{i} \right ],
\end{equation}
where $\xi$ is the temporal distance from the LF (considered as the origin), $i\in \{1,2,3\}$ is the index distinguishing the three aforementioned types of outages, and $K_{i}$ is the number of events for the $i$-th type of outage.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 387 119,scale=0.7,keepaspectratio=true]{./Time_Series_rapres}
\caption{Representation of TSs of outages happened before a LF.}
\label{fig:Time_Series_rapres}
\end{figure}
Normalization of such data is performed as follows. Given a dissimilarity measure for TSs (see Sec. \ref{sec:Time_Series_Data}), we pre-compute, for each type of short outage feature, the dissimilarity matrix, i.e. a symmetric square matrix containing all the possible pairwise dissimilarity values between patterns in the considered dataset. Successively, we apply the normalization dividing the dissimilarity value between two TSs with the maximum value in the dissimilarity matrix.
\subsection{The Proposed One-class Classifier}
As a consequence of the difficulty of modeling useful (and meaningful) instances of non-faults in the considered SG, we designed a OCC for the purpose of recognizing LFs.
Such a goal is implemented by building a OCC relying on clustering techniques. The idea of using clusters for modeling a region of the ``fault representation space'', $\mathcal{F}$, containing target patterns representing LFs, is reasonable and also intuitive. The underlying assumption is that, similar statuses of the SG have similar chances of generating a LF, assumption that is reflected by the cluster model.
A dataset of FPs is partitioned in $k$ (disjoint) clusters, where each cluster contains faults having similar features.
Accordingly, the most important component of the OCC system is the core dissimilarity measure $d: \mathcal{F}\times\mathcal{F}\rightarrow\mathbb{R}^+$, which assigns a dissimilarity value to a pair of FPs.
The partition, as well as other parameters that will be described in the following, constitute the model of the OCC.
We would like to remark the need of designing an ad-hoc OCC. The application that we face in this study deals with highly structured data, which are processed by means of a weighted aggregation of several heterogeneous dissimilarity measures. This results in a FP representation space $(\mathcal{F}$, $d(\cdot, \cdot))$ that is not Euclidean (nor it is metric). Therefore, using more consolidated methods, such as the SVDD cited in Sec. \ref{sec:occ_prob}, is not straightforward and it would require the definition of particular (positive definite) kernel functions tailored to the problem at hand.
\subsubsection{The Dissimilarity Measure Among FPs}
\label{sec:dissimilarity_measure}
Let $\mathcal{S}\subset\mathcal{F}$ be a dataset of FPs. A FP $x\in\mathcal{S}$ is described as
\begin{equation}
\label{eq:data_obj}
x=\left \{ \mathbf{F}_{1},\mathbf{F}_{2}, ...,\mathbf{F}_{m} \right \},
\end{equation}
where the $l$-th feature, $\mathbf{F}_{l}$, $1 \leq l \leq m$, lies in its specific feature space $\mathcal{F}_{l}$. Hence, each pattern $x$ lies on the $m$-fold product feature space $\mathcal{F}=\mathcal{F}_{1}\times \mathcal{F}_{2}\times ...\times \mathcal{F}_{m}$.
Given two FPs $x, y \in \mathcal{S}$, the proposed weighted dissimilarity measure reads as:
\begin{equation}
\label{eq:Custom_Distance}
d(x, y; \underline{\mathbf{w}})= \sqrt{\sum_{j=1}^{m} w_{j}\times \left(x_{j} \circleddash y_{j} \right)^2},
\end{equation}
where the $\circleddash$ operator represents a generic dissimilarity measure, and $w_j\in[0, 1]$ is the weight related to the $j$-th feature. In practice, Eq. \ref{eq:Custom_Distance} computes the weighted $l_2$ norm of the vector containing the dissimilarity values calculated feature-wise. However, since not all dissimilarity measures implementing the $\circleddash$ operator are metrics, the resulting dissimilarity measure (\ref{eq:Custom_Distance}) is not metric and hence it does not induce a metric space. This aspect is carefully taken in consideration into the design of our OCC system.
The weights $\underline{\mathbf{w}}\in[0, 1]^m$ are suitably optimized during the training phase of the OCC by means of a Genetic Algorithm (GA). In other words, the overall system is in charge to synthesize a classification model by a clustering technique, learning the parameters of the dissimilarity measure yielding the most appropriate clusters able to define the faults decision regions. For this reason the proposed OCC system fully belongs to the Metric Learning framework \cite{Metric_learning}.
In the following paragraphs, we describe the implementations of $\circleddash$, i.e., the specific dissimilarity measures tailored for each specific FP component.
The nature of the feature, $\mathcal{F}_{i}$, will be denoted using the same notation of Tab. \ref{tab:features_description}.
\paragraph{Categorical Data}
\label{sec:Categorical_Data}
Categorical attributes, also referred to as nominal attributes, are data without an ``meaningful'' ordering (see Tab. \ref{tab:features_description} for the data treated as nominal).
Let $\mathcal{F}^{c}=\left \{ \eta_{1} ,\eta_{2},..., \eta_{n} \right \}$ be the set of all categorical features of the entire dataset, each described by $d$ categorical attributes: $\nu_{1} ,\nu_{2},..., \nu_{d}$. Let us define the domain of the attribute $\nu_{j}$, $\mathrm{DOM}( \nu_{j} )= \left \{ A_{j_1},A_{j_2},...,A_{j_{n(j)}} \right \}$, where $ A_{j_l}$ $(1\leqslant l \leqslant n(j) )$ is the set of possible values for the categorical attribute $\nu_{j}$, and $n(j)$ is its cardinality. We consider the well-known simple matching distance, defined as follows:
\begin{equation}
\label{eq:SimpleMatch_Distance}
\delta(x,y)=\left\{\begin{matrix}
0 & x=y, \\
1 & x\neq y.
\end{matrix}\right.
\end{equation}
Let $x^{c}$ and $y^{c}$ be the projections on the categorical feature space $\mathcal{F}^{c}$ of two generic patterns $x, y$. The dissimilarity measure between the two categorical objects described by $d$ categorical attributes is implemented as:
\begin{equation}
\label{eq:SimpleMatch_Distance_complete}
d^{c}(x^{c}, y^{c})=\frac{1}{d}\sum_{j=1}^{d}\delta (x_{j}^{c},y_{j}^{c}).
\end{equation}
\paragraph{Quantitative Data}
\label{sec:Quantitative_Data}
As concerns the quantitative data (see Tab. \ref{tab:features_description}) we distinguish between (i) ``Normal'' quantitative data and (ii) ``Special'' quantitative data. The former type includes both numerical and integer values (normalized in $[0, 1]$); the operator $\circleddash$ is implemented by the absolute difference: $d^{N} =\left | x-y\right |$.
For integer values describing information related to timestamps temporal information, such as the day in which the LF happened and the time of day, we defined a particular dissimilarity measure implementing $\circleddash$, called \textit{circular difference}.
Given an ordered set of integer numbers $\{0, 1, ..., a\}$, the circular difference among any $x,y$ in this set is given by
\begin{equation}
\label{eq:Circular_diff}
d^{CD}(x, y; a)=\min(\left |x-y \right |,a-\left |x-y \right |),
\end{equation}
where $a$ is considered as a parameter.
For ``Day start'' and ``Time start'' the maximum value for $a$ in (\ref{eq:Circular_diff}) is 364 and 1439, respectively.
The implementation of the circular difference is designed to avoid that pairs of close days or timestamps give raise to high dissimilarity values.
``Special'' quantitative data are normalized in the range $[0, 1]$, but can assume also a special symbol, $\epsilon$, indicating the ``not applicable'' condition.
It is the case for the ``Cable section'' feature, since for LFs not related to cables this field is undefined. The dissimilarity measure $d^{S}: \{[0, 1] \cup \epsilon\} \times \{[0, 1] \cup \epsilon\} \rightarrow [0, 1]$ for two special quantitative values $x,y \in \mathcal{F}^{S}$ is defined as follows:
\begin{equation}
\label{eq:Special_diff}
d^{S}=\left\{\begin{matrix}
|x-y| & x \neq \epsilon \wedge y \neq \epsilon, \\
1 & x =\epsilon \vee y=\epsilon, \\
0 & x=\epsilon \wedge y=\epsilon.
\end{matrix}\right.
\end{equation}
\paragraph{Time Series Data}
\label{sec:Time_Series_Data}
The Dynamic Time Warping (DTW) is a well-known algorithm to find an optimal alignment between two sequences of objects (i. e. Time Series) of variable length.
The use of DTW as dissimilarity measure for sequences of generic objects is well-established in many applications, such as biology, finance, multimedia, and image analysis \cite{PiyushShanker20071407,t2vsdiss__ifsanafips2013}.
An in-depth description of DTW algorithm can be found in \cite{muller_dtw}.
Following the notation introduced in Sec. \ref{sec:Time_Series}, the dataset consists in three types of TSs, $S^{i}$, with $i\in\{1,2,3\}$. Each one represents a vector belonging to the TSs feature vector subspace, $\mathcal{F}_{i}^{TS}, i\in\{1, 2,3\}$.
Let $x, y \in \mathcal{F}_{i}^{TS}$ be two TSs, the dissimilarity measure between them is computed as as a function $d^{TS}: \mathbb{R}^{m}\times \mathbb{R}^{n}\rightarrow \mathbb{R}$:
\begin{equation}
\label{eq:DTW}
d^{TS}(x, y)=\mathrm{DTW}(x,y),
\end{equation}
where $m$ and $n$ are the lengths of $x$ and $y$, respectively.
\subsubsection{Model Definition and the Classifier Decision Rule}
\label{sec:model}
The most important component of the OCC model is the partition $P$ (i.e., a set of clusters), determined on the training set $\mathcal{S}_{tr}$. The partition is obtained through a clustering algorithm -- see Fig. \ref{fig:Clustering_model} for an overview.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 502 200,scale=0.6,keepaspectratio=true]{./Clustering_system}
\caption{The OCC model is defined by a partition of the training set $\mathcal{S}_{tr}$.}
\label{fig:Clustering_model}
\end{figure}
A hard partition of order $k$ is a collection of $k$ disjoint and non-empty clusters, $P=\{\mathcal{C}_{1}, \mathcal{C}_2, ..., \mathcal{C}_k\}$.
Each cluster $\mathcal{C}_i\in P$ is synthetically described by a \textit{representative} element, which we denote as $c_i=R(\mathcal{C}_i)$; let $R(P)=\{c_1, c_2, ..., c_k\}$ be the set of representatives of the partition $P$.
The representative of $\mathcal{C}_i$ is computed as the element $c_i$ that minimizes the sum of distances (MinSOD) \cite{delvescovo+livi+rizzi+frattalemascioli2011}:
\begin{equation}
\label{eq:minsod}
c_i = \operatornamewithlimits{arg\ min}_{x_j\in\mathcal{C}_i} \sum_{x_k\in\mathcal{C}_i} d(x_j, x_k).
\end{equation}
A cluster representative $c_i$ is, in a sense, the prototype of a \textit{typical fault scenario} individuated in $\mathcal{S}_{tr}$.
As a consequence, the information provided by the cluster $\mathcal{C}_i$ as a whole is useful to conceive a region of the pattern space ``around'' $c_i$, which describes similar fault scenarios.
By defining $\delta(\mathcal{C}_i)\geq0$ as a measure of cluster extent, we can construct the decision region associated to each cluster $\mathcal{C}_j$, used to implement the classification rule.
The cluster extent can be computed as the average/maximum intra-cluster dissimilarity value or by considering their standard deviation, for instance.
In the case of the average, the expression reads as:
\begin{equation}
\delta(\mathcal{C}_i) = \frac{1}{|\mathcal{C}_i|-1}\sum_{x_k\in\mathcal{C}_i} d(c_i, x_k)).
\end{equation}
In addition to $\delta(\mathcal{C}_i)$, we consider also a tolerance parameter, $\sigma_i\geq0$, for defining the decision region.
The decision region derived from a cluster $\mathcal{C}_i$ is hence defined by the quantity $B(\mathcal{C}_i)=\delta(\mathcal{C}_i)+\sigma_i$, which actually defines the neighborhood of $c_i$.
Fig. \ref{fig:Region_model} provides a schematic overview of a cluster model and its use in the process of classifying a test pattern $\bar{x}$.
The classification rule for a test pattern $\bar{x}$ operates in two stages. First, the nearest cluster representative $c^*\in R(P)$ is individuated according to the following expression:
\begin{equation}
\label{eq:1}
c^* = \mathop {\arg \min }\limits_{{c_j} \in R(P)} d(\bar x, {c_j}).
\end{equation}
The second step consists in comparing the dissimilarity value $d(\bar{x}, c^*)$ with $B(\mathcal{C}^*)$.
We define a binary-valued function $h(\cdot)$ that performs the hard classification:
\begin{equation}
\label{eq:class_rule}
h(x)=
\begin{cases}
1 & \textrm{if}\ d(\bar x, c^*) \le B(\mathcal{C}^*), \\
0 & \textrm{otherwise}.
\end{cases}
\end{equation}
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 631 224,scale=0.5,keepaspectratio=true]{./DR_description}
\caption{The cluster decision region and its characterizing parameters.}
\label{fig:Region_model}
\end{figure}
Along with the hard classification (\ref{eq:class_rule}), we developed a mechanism based on fuzzy sets to provide the user with a measure of ``reliability'' associated to the decisions.
This topic is discussed in the following dedicated section.
\subsubsection{Evaluating the Reliability of the Classification}
\label{sec:class_reliability}
A Boolean decision regarding if a new test pattern (i.e., a given SG status) is a fault or not, is operatively reasonable. However, it is important to provide the user also with an additional measure that quantifies the reliability of such a decision. This becomes even more appropriate in the particular OCC setting.
For this purpose, we equip each cluster $\mathcal{C}_i$ with a suitable membership function, denoted in the following as $\mu_{\mathcal{C}_i}(\cdot)$.
In practice, we generate a fuzzy set over $\mathcal{C}_i$.
The membership function allows us to quantify the uncertainty (expressed by the membership degree in $[0, 1]$) of a decision about the recognition of a test pattern.
Fig. \ref{fig:Reliability} depicts this idea by an intuitive illustration.
Membership values close to either 0 or 1 denote ``certain'' and hence reliable decisions. When the membership degree assigned to a test pattern is close to 0.5, there is no clear distinction about the fact that such a test pattern is really a fault or not (regardless of the correctness of the Boolean decision).
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 728 319,scale=0.4,keepaspectratio=true]{./Reliability_f}
\caption{Sigmoidal membership function associated to the decision region.}
\label{fig:Reliability}
\end{figure}
For this purpose, we used a parametric sigmoid model for $\mu_{\mathcal{C}_i}(\cdot)$, which is defined as follows:
\begin{equation}
\label{eq:membership}
\mu_{\mathcal{C}_i}(x) = \frac{1}{1+\exp((d(c_i, x)-b_i)/a_i)},
\end{equation}
where $a_i,b_i\geq 0$ are two parameters specific to $\mathcal{C}_i$, and $d(\cdot, \cdot)$ is the dissimilarity measure (\ref{eq:Custom_Distance}). Notably, $a_i$ is used to control the steepness of the sigmoid (the lower the value, the faster the rate of change), and $b_i$ is used to translate the function in the input domain.
If a cluster (that models a typical fault situation found in the training set) is very compact, then it describes a very specific fault scenario. Therefore, no significant variations should be accepted to consider test patterns as members of this cluster.
The converse is also true. If a cluster is characterized by a wide extent, then we might be more tolerant in the evaluation of the membership.
Accordingly, the parameter $a_i$ is set equal to $\delta(\mathcal{C}_i)$.
On the other hand, we define $b_i = \delta(\mathcal{C}_i) + \sigma_i/2$. This allows us to position the part of the sigmoid that changes faster right in-between the area of the decision region determined by the dissimilarity values falling in $[B(\mathcal{C}_i)-\sigma_i, B(\mathcal{C}_i)]$.
Finally, the soft decision function, $s(\cdot)$, is defined as
\begin{equation}
\label{eq:soft_decision}
s(\bar x) = \mu_{\mathcal{C}^*}(\bar x),
\end{equation}
where $\mathcal{C}^*$ is the cluster satisfying Eq. \ref{eq:1}.
The evaluation of Eq. \ref{eq:soft_decision} over a test set $\mathcal{S}_{ts}, n=|\mathcal{S}_{ts}|$, yields $n$ membership degrees, each assigned to a specific pattern of $\mathcal{S}_{ts}$.
We can evaluate the overall reliability of the decisions taken on $\mathcal{S}_{ts}$ by considering a fuzzy set, say $\mathcal{M}$, characterized by those $n$ membership degrees.
We can implement such an evaluation measure by calculating the fuzzy entropy \cite{Livi_ga_2013} of $\mathcal{M}$. A possible expression for the fuzzy entropy reads as
\begin{equation}
\label{eq:fe}
\frac{\mathrm{card}(\mathcal{M}\cap \mathcal{M}^c)}{\mathrm{card}(\mathcal{M}\cup \mathcal{M}^c)}\in[0, 1],
\end{equation}
where $\cap$ and $\cup$ are defined as the minimum and maximum operators, respectively, and the cardinality of the resulting fuzzy set is taken as the sum of the membership degrees.
Fuzzy entropy values close to zero would indicate that, overall, the decisions are reliable. Conversely, if the fuzzy entropy is close to one then the decisions are highly unreliable.
It is worth stressing that the reliability measures herein discussed should not be confused with the measures of correctness of the recognition (i.e., the evaluation of the correctness of the discrimination among target and non-target patterns).
\subsubsection{Training of the OCC by the \textit{k}-means Algorithm}
\label{sec:kmeans_training}
We propose a learning strategy to synthesize the OCC model that is based on the well-known \textit{k}-means \cite{Jain:2010:DCY:1755267.1755654}. Since our feature space is non-metric and we represent each cluster by the MinSOD pattern, in the technical literature this algorithm is usually referred to $k$-medoids \cite{Park20093336}. This clustering procedure depends on an integer parameter, $k$, defining a priori the partition order.
The dissimilarity measure described in Sec. \ref{sec:dissimilarity_measure} depends on a vector of weights, $\underline{\mathbf{w}}$. Moreover, the decision regions -- Sec. \ref{sec:model} -- are defined by the thresholds $\sigma_i$.
Setting those parameters, denoted $p_j=[\underline{\mathbf{w}}_j, \underline{\sigma}_j]$, is of utmost importance, and of course it has a significative influence on the results yielded by \textit{k}-means.
For this reason, a GA is employed to find the best-performing values for $p_j$, i.e., those maximizing the following objective function:
\begin{equation}
\label{eq:obj_kmeans}
f(p_j) = \alpha A(\mathcal{S}_{vs}) + (1-\alpha) \sum_{i=1}^{k} 1-\sigma_i.
\end{equation}
In (\ref{eq:obj_kmeans}), $A(\mathcal{S}_{vs})$ is the performance measure achieved on $\mathcal{S}_{vs}$ (we specify the nature of such a measure in the experiments section).
The second term in (\ref{eq:obj_kmeans}) defines a constrain for the (average) cluster extent. The GA is in charge to find the parameters, $p_j$, that minimize the $l_1$ norm of the tolerances used to define the decision regions, while at the same time providing an effective performance in terms of recognition.
Fig. \ref{fig:System_model} shows a diagram illustrating the optimization stage as a whole.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 704 330,scale=0.5,keepaspectratio=true]{./OCC_System}
\caption{Block diagram depicting the classification model synthesis procedure.}
\label{fig:System_model}
\end{figure}
As concerns the learning phase, it is well-known that the \textit{k}-means algorithm is sensitive to the adopted cluster initialization strategy; here we used a fast randomized initialization.
To compensate this fact, the current version of the OCC takes as external parameter the $k$ value and a classification model is synthesized for each $k$ in a given user-defined range $k_{\rm{min}}, k_{\rm{max}}$.
For each $k$ in this range, we synthesize three models with different random initializations. The fitness (\ref{eq:obj_kmeans}) associated to a candidate solution $p_j$ is hence the average of the fitness calculated for those three models.
In the test phase, we use a majority voting scheme to decide if a given test pattern falls in one of the synthesized decision regions or not.
\section{Experimental Evaluation}
\label{sec:experiments}
In Sec. \ref{sec:exp_setting} we introduce the main experimental setting adopted in this paper.
Successively, we discuss the obtained results in terms of quality of the recognition, respectively on synthetically generated data (Sec. \ref{sec:synth}), on some well-known UCI datasets (Sec. \ref{sec:comparison}), and on the data provided by ACEA (Sec. \ref{sec:acea}).
Although the goal of this paper is to present and discuss the results on the ACEA data, we performed experiments also on some UCI datasets to offer a comparison with respect to well-known OCCs on a more established groundwork.
\subsection{Experimental setting}
\label{sec:exp_setting}
A (one-class) classification problem instance is defined as a triple of disjoint sets, namely training set ($\mathcal{S}_{tr}$), validation set ($\mathcal{S}_{vs}$), and test set ($\mathcal{S}_{ts}$).
Given a specific parameters setting, a classification model is synthesized on $\mathcal{S}_{tr}$ and it is validated on $\mathcal{S}_{vs}$. The generalization capability of the optimized model is computed on $\mathcal{S}_{ts}$.
The proposed OCC produces both hard (\ref{eq:class_rule}) and soft decisions (\ref{eq:soft_decision}) on each test pattern.
In the hard decision case, we evaluate the recognition performance of the classifier by exploiting the confusion matrix.
In particular, we consider the false positive rate (FPR), recall, precision, and accuracy \cite{Fawcett:2006:IRA:1159473.1159475}.
On the other hand, in the soft decision case we quantify the correctness of the classifier by computing the area under the ROC curve (AUC) \cite{Fawcett:2006:IRA:1159473.1159475} generated by interpreting the membership degrees (\ref{eq:soft_decision}) as suitable ``scores'' assigned by the classifier to the test patterns.
The OCC parameters defining the model are optimized by means of a GA, which is guided by the objective shown in Eq. \ref{eq:obj_kmeans}; $\alpha=0.8$ has been used as a suitable setting for the problem at hand.
In (\ref{eq:obj_kmeans}), we implement the accuracy term as the accuracy elaborated from the confusion matrix.
In the voting scheme used during the cross-validation, it is possible to obtain more than one model scoring the same highest value of accuracy on $\mathcal{S}_{vs}$. In such a case, we chose the model characterized by the soft decisions denoting the lowest fuzzy entropy (\ref{eq:fe}).
This would help in choosing a model that able to correctly discriminate fault patterns, maximizing its reliability.
The adopted GA performs stochastic uniform selection, Gaussian mutation, and scattered crossover (with crossover fraction of 0.8).
It implements a form of elitism that imports the two fittest individuals in the next generation; the population size is is kept constant throughout the generations and equal to 50 individuals.
The stop criterion is defined by considering a maximum number of iterations (250) and checking the variations of the best individual fitness.
\subsection{Tests on synthetic data}
\label{sec:synth}
Fig. \ref{fig:test1} shows the first synthetic test that we have conceived just to show the functioning of the proposed OCC.
The target patterns used for training the OCC are distributed in three Gaussian shaped, well-separated, clusters.
Patterns used for testing are clearly highly recognizable.
The \textit{k}-means is executed with $k=3$ and by synthesizing three different models to be used for the majority voting mechanism.
The accuracy obtained on the test set with the \textit{k}-means is equal to one for all three models.
In Fig. \ref{fig:test1_results}, we report the soft decisions on the test set of the three best-performing models. While the hard decisions are all correct, we note that with the first model we obtain more reliable decisions. In fact, the computed fuzzy entropy is much lower than the other two (the fuzzy entropy is almost zero). As a consequence, the OCC chooses the results of this model as the final output.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 407 344,scale=0.5,keepaspectratio=true]{./Test1}
\caption{Synthetic problem.}
\label{fig:test1}
\end{figure}
\begin{figure}[ht!]
\centering
\subfigure[First model. Fuzzy entropy is 0.0007.]{
\includegraphics[viewport=0 0 365 284,scale=0.45,keepaspectratio=true]{./Test1_k3_M1}
\label{fig:test1_k3_m1}}
\quad
\subfigure[Second model. Fuzzy entropy is 0.0599.]{
\includegraphics[viewport=0 0 365 284,scale=0.45,keepaspectratio=true]{./Test1_k3_M2}
\label{fig:test1_k3_m2}}
\subfigure[Third model. Fuzzy entropy is 0.0518.]{
\includegraphics[viewport=0 0 365 284,scale=0.45,keepaspectratio=true]{./Test1_k3_M3}
\label{fig:test1_k3_m3}}
\caption{Results of the soft decisions obtained by the three models synthesized with the \textit{k}-means using the best-performing solution.}
\label{fig:test1_results}
\end{figure}
We now move to a synthetic test designed to provide a justification for the generation of the non-fault (in the following, non-target and non-fault will be treated as synonyms) patterns that we used within the ACEA data (discussed later in Sec. \ref{sec:acea}).
Fig. \ref{fig:test2} illustrates the considered setting. Training, validation, and test target patterns are still grouped in three well-separated Gaussian clusters. Non-target patterns used for validating and testing a model are distributed uniformly over the $[0, 1]^2$ domain; those patterns outnumber the target patterns, since some non-target patterns will fall in the fault decision region.
By testing the OCC over such data, we expect to observe an ``implicit'' FPR that is proportional to the number of (training set) target patterns. In other terms, we expect to observe $FPR\simeq |\mathcal{S}_{tr}|/|\mathcal{\hat{S}}_{ts}|$, where $\mathcal{\hat{S}}_{ts}$ is the subset of test patterns belonging to the non-target class.
We considered 150 target patterns for the training, validation, and test sets, while we used 1500 non-target patterns in the validation and test sets.
As expected, the FPR of the best model is $\simeq 0.1082$, with an overall accuracy of 0.8953.
By means of this interpretation, we could safely affirm that the ``true'' FPR is only $\simeq 0.0082$, since $|\mathcal{S}_{tr}|/|\mathcal{\hat{S}}_{ts}|=0.1$.
The AUC is 0.9884; Fig. \ref{fig:test2_mf} shows the calculated membership values for the test patterns. Although the discrimination is very good, the OCC necessarily commits some mistakes, due the uniform distribution of the non-target patterns.
To demonstrate the reliability of this interpretation of the test, we repeated this experiment by increasing the ratio $|\mathcal{S}_{tr}|/|\mathcal{\hat{S}}_{ts}|$ from 0.1 to 0.475, and, accordingly, increasing also the spread of the target patterns over the domain.
In Fig. \ref{fig:regression} we show the linear correlation between the increments of the $|\mathcal{S}_{tr}|/|\mathcal{\hat{S}}_{ts}|$ ratio and the calculated FPR over the respective test set.
As it was expected, there is a strong linear relationship among those two quantities (correlation coefficient is $\simeq 0.96$), which demonstrates that the implicit FPR obtained with this method of generation of non-target patterns is predictable.
\begin{figure}[ht!]
\centering
\subfigure[Dataset distribution.]{
\includegraphics[viewport=0 0 411 378,scale=0.45,keepaspectratio=true]{./Test2}
\label{fig:test2}}
~
\subfigure[Fuzzy entropy of the best model is 0.0352.]{
\includegraphics[viewport=0 0 417 323,scale=0.45,keepaspectratio=true]{./Test2_mf}
\label{fig:test2_mf}}
\caption{Dataset distribution and related fuzzy membership values calculated by the best model.}
\label{fig:test2_results}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 461 261,scale=0.5,keepaspectratio=true]{./regression}
\caption{Correlation among $|\mathcal{S}_{tr}|/|\mathcal{\hat{S}}_{ts}|$ and the computed FPR on the test set.}
\label{fig:regression}
\end{figure}
\subsection{Comparison on UCI datasets}
\label{sec:comparison}
In Tab. \ref{tab:uci_ds} we show the UCI datasets considered in this study; all datasets are standardized with zero mean and unitary variance.
In Tab. \ref{tab:uci_ds_results} we report the results in terms of AUC obtained by our system, denoted as ``OCC{\_}System'', together with those retrieved from the literature (see Ref. \cite{eocc__arxiv} and references therein).
Results in Tab. \ref{tab:uci_ds_results} show how the OCC{\_}System reaches satisfactory performances on the considered five UCI datasets.
Notably, considering \textbf{D} and \textbf{L} datasets the OCC{\_}System achieves the highest AUC.
It is worth pointing out that the proposed OCC achieves a good AUC on the dataset with the highest number of features (i.e., eight for dataset \textbf{D}).
In Fig. \ref{fig:Ecoli_k_search} we show the trend of the AUC achieved for the \textbf{E} dataset with respect to the partition value $k$ ranging in $\{3, ..., 10\}$.
The figure shows how the OCC{\_}System, although it reaches slightly different AUC values varying the $k$ parameter, for some of that, i.e., $k=4$ and $k=10$, the performance is still comparable.
This fact suggests that, at least on the considered dataset, the proposed OCC does not overfit the model reaching good performances also with partitions of lower order.
\begin{table*}[th!]\scriptsize
\begin{center}
\caption{UCI datasets considered in this study.}
\label{tab:uci_ds}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{UCI Dataset} & \textbf{Acronym} & \textbf{Target class} & \textbf{\# Target} & \textbf{\# Non-target} & \textbf{\# Params} \\
\hline
Biomed & BI & normal & 127 & 67 & 5 \\
Breast Wisconsin & BW & benign & 458 & 241 & 9 \\
Diabetes (prima indians) & D & present & 500 & 268 & 8 \\
Ecoli & E & pp & 52 & 284 & 7 \\
Iris & I & Iris-setosa & 50 & 100 & 4 \\
Liver & L & healthy & 200 & 145 & 6 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[thp!]\scriptsize
\begin{center}
\caption{Test set results showing the average AUC values together with the standard deviations. Best results are reported in bold. The ``-'' symbol indicates that the result is not available.}
\label{tab:uci_ds_results}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{System/Dataset} & \textbf{BI} & \textbf{BW} & \textbf{D} & \textbf{E} & \textbf{I} & \textbf{L} \\
\hline
\rowcolor{lgray} OCC{\_}System & 0.904(0.013) & \textbf{0.996(0.002)} & \textbf{0.756(0.003)} & 0.949(0.008) & \textbf{1.000(0.000)} & \textbf{0.652(0.011)} \\
\hline
EOCC-1 & 0.867(0.005) & 0.853(0.020) & 0.670(0.024) & 0.928(0.011) & \textbf{1.000(0.000)} & 0.396(0.016) \\
EOCC-2 & 0.878(0.006) & 0.995(0.001) & 0.751(0.012) & \textbf{0.957(0.004)} & \textbf{1.000(0.000)} & 0.460(0.026) \\
EOCC-2\_10\% & 0.862(0.016) & 0.995(0.002) & 0.709(0.023) & \textbf{0.954(0.007)} & \textbf{1.000(0.000)} & 0.452(0.026) \\
Gauss & 0.899(0.005) & 0.985(0.001) & 0.721(0.003) & 0.929(0.003) & \textbf{1.000(0.000)} & 0.509(0.005) \\
MoG & 0.911(0.008) & 0.984(0.002) & 0.738(0.003) & 0.929(0.003) & \textbf{1.000(0.000)} & 0.494(0.006) \\
Na\"{\i}ve Parzen & \textbf{0.931(0.002)} & 0.987(0.001) & 0.678(0.003) & 0.930(0.008) & \textbf{1.000(0.000)} & 0.484(0.008) \\
Parzen & 0.915(0.009) & 0.991(0.001) & 0.756(0.002) & 0.929(0.005) & \textbf{1.000(0.000)} & 0.469(0.008) \\
\textit{k}-Means & 0.902(0.009) & 0.984(0.001) & 0.712(0.010) & 0.878(0.015) & \textbf{1.000(0.000)} & 0.469(0.014) \\
1-NN & 0.914(0.012) & 0.991(0.001) & 0.721(0.002) & 0.906(0.008) & \textbf{1.000(0.000)} & 0.511(0.007) \\
\textit{k}-NN & 0.914(0.012) & 0.991(0.001) & 0.721(0.002) & 0.906(0.008) & \textbf{1.000(0.000)} & 0.511(0.007) \\
Auto-encoder & 0.890(0.013) & 0.960(0.002) & 0.658(0.005) & 0.888(0.023) & \textbf{1.000(0.000)} & 0.608(0.008) \\
PCA & 0.776(0.031) & 0.920(0.004) & 0.640(0.006) & 0.655(0.013) & 0.920(0.008) & 0.608(0.008) \\
SOM & 0.908(0.006) & 0.990(0.002) & 0.709(0.009) & 0.898(0.004) & \textbf{1.000(0.000)} & 0.487(0.017) \\
MST\_CD & 0.914(0.012) & 0.992(0.001) & 0.715(0.003) & 0.899(0.009) & \textbf{1.000(0.000)} & - \\
\textit{k}-Centres & 0.906(0.015) & 0.984(0.002) & 0.678(0.009) & 0.870(0.023) & \textbf{1.000(0.000)} & 0.483(0.006) \\
SVDD & 0.915(0.009) & 0.988(0.001) & 0.732(0.005) & 0.922(0.010) & \textbf{1.000(0.000)} & 0.490(0.010) \\
MPM & 0.909(0.010) & 0.991(0.001) & 0.729(0.003) & 0.922(0.007) & \textbf{1.000(0.000)} & 0.521(0.011) \\
LPDD & 0.889(0.008) & 0.989(0.001) & 0.634(0.005) & 0.947(0.004) & \textbf{1.000(0.000)} & 0.506(0.005) \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 601 451,scale=0.40,keepaspectratio=true]{./Ecoli_k_search}
\caption{AUC values for different partition orders achieved on the Ecoli (\textbf{E}) dataset.}
\label{fig:Ecoli_k_search}
\end{figure}
\subsection{Results and discussion on ACEA data}
\label{sec:acea}
The ACEA dataset available for our experiments does not contain instances of non-fault situations (i.e., normal functioning of the system).
This fact creates some difficulty in evaluating any data-driven inference mechanism.
To generate instances of non-target patterns, we use the method discussed in the second experiment presented in Sec. \ref{sec:synth}.
Non-target patterns are formed by randomly generating, with a uniform distribution, each feature value characterizing a FP (see Sec. \ref{sec:pattern_representation} for details on the features).
Since the dataset contains heterogeneous data types, such a uniform generation mechanism seems to be the most appropriate one.
In Fig. \ref{fig:pca_dissmatrix_faults} we show the first two components of the PCA calculated over the dissimilarity matrix, $D_{ij}=d(x_i, x_j), \forall x_i,x_j\in\mathcal{S}$, generated for the entire available ACEA dataset $\mathcal{S}$, containing either fault and non-fault instances; $d(\cdot, \cdot)$ in Eq. \ref{eq:Custom_Distance} is computed by using unitary weights.
The herein considered dataset is divided in training, validation, and test sets according to the following splits.
The training set is composed of 532 fault patterns; in the validation set we have 470 fault and 500 non-fault patterns; finally for the test we have 82 fault and 500 non-fault patterns.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 477 375,scale=0.5,keepaspectratio=true]{./PCA_NORM_DM}
\caption{First two components of the PCA elaborated over the dissimilarity matrix constructed from the ACEA dataset containing either fault and non-fault (uniformly generated) patterns.}
\label{fig:pca_dissmatrix_faults}
\end{figure}
Tab. \ref{tab:results_acea} shows the obtained results. We tested the OCC for five different values of \textit{k} in the \textit{k}-means algorithm. For each $k$, we repeated the test five different times by changing the random seed of the pseudo-random number generator driving the stochastic behaviors of the \textit{k}-means initialization procedure and of the genetic algorithm. Hence the results are intended as averages with related standard deviations.
In the table we show the false positive rate (FPR), the recall (R), the accuracy (A), the area under the ROC curve (AUC), the fuzzy entropy (FE) of the winning model, and finally the mutual information (MI) among the fitness of the best individual and the estimated entropy of the related weights. Let us focus now on the first five columns.
Results in terms of recognition, considering both hard and soft decisions, are in general very good.
The best overall result is obtained with $k=7$ (please note a somewhat clear 7-cluster structure of the fault patterns appearing from the first two components of the PCA in Fig. \ref{fig:pca_dissmatrix_faults}). Notably, FPRs are always very low, demonstrating the capability of the proposed OCC of synthesizing well-defined and effective decision regions.
Notwithstanding non-fault patterns are generated by using a uniform distribution over each feature describing a FP (please see Fig. \ref{fig:pca_dissmatrix_faults}), the system is able to isolate such patterns correctly.
On the same line of thoughts, both accuracy and AUC are nearly one, denoting an almost perfect recognition of faults.
The FE of the best models (see Sec. \ref{sec:class_reliability}) is almost negligible (in accord with the very high AUC), which tells us that the soft decisions are also highly reliable.
To offer an argument to demonstrate the validity of the herein shown results, and accordingly with the generation mechanism for the non-fault patterns, in Fig. \ref{fig:auc_a_fpr} we show the performances obtained by progressively increasing the number of non-fault patterns. The generation mechanism of course remains the same. In the figure we show the best performing configuration of the system for each non-fault patterns set size.
From Fig. \ref{fig:auc_a} it is possible to deduce that the AUC does not vary significantly by increasing the number of non-fault patterns.
On the other hand, the accuracy (A) is slightly more affected, although still denoting good results. This could be interpreted as a sign of robustness of the fuzzy set based soft decision mechanism herein proposed.
Finally, Fig. \ref{fig:fpr} reports the FPR that denotes a trend correlated with the one of the AUC.
\begin{table}[thp!]\scriptsize
\begin{center}
\caption{Average test sets results on the ACEA data. Results are reported for five different values of \textit{k} for the \textit{k}-means algorithm.}
\label{tab:results_acea}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{K} & \textbf{FPR} & \textbf{R} & \textbf{A} & \textbf{AUC} & \textbf{FE} & \textbf{MI} \\
\hline
4 & 0.00449\textpm0.003 & 0.94200\textpm0.027 & 0.98800\textpm0.005 & 0.99600\textpm0.002 & 0.00511\textpm0.002 & 0.70529\textpm0.059 \\
\hline
5 & 0.00599\textpm0.001 & 0.95700\textpm0.012 & 0.98900\textpm0.001 & 0.99500\textpm0.001 & 0.00483\textpm0.002 & 0.61681\textpm0.082 \\
\hline
6 & 0.00649\textpm0.002 & 0.93600\textpm0.020 & 0.98500\textpm0.002 & 0.99500\textpm0.001 & 0.00497\textpm0.001 & 0.65527\textpm0.029 \\
\hline
7 & 0.00449\textpm0.003 & 0.95700\textpm0.015 & 0.99000\textpm0.003 & 0.99600\textpm0.002 & 0.00155\textpm0.018 & 0.78783\textpm0.072 \\
\hline
8 & 0.00499\textpm0.003 & 0.95700\textpm0.007 & 0.99000\textpm0.003 & 0.99500\textpm0.003 & 0.00165\textpm0.014 & 0.72686\textpm0.131 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[ht!]
\centering
\subfigure[AUC and Accuracy (A).]{
\includegraphics[viewport=0 0 352 243,scale=0.6,keepaspectratio=true]{./AUC_A}
\label{fig:auc_a}}
~
\subfigure[FPR.]{
\includegraphics[viewport=0 0 352 243,scale=0.6,keepaspectratio=true]{./FPR}
\label{fig:fpr}}
\caption{Test set results by increasing the number of considered non-fault patterns in the test set. Fig. \ref{fig:auc_a} shows the AUC and A while Fig. \ref{fig:fpr} shows the FPR.}
\label{fig:auc_a_fpr}
\end{figure}
Let us now go back to discuss the information conveyed by the MI column in Tab. \ref{tab:results_acea}.
As it is clear from Eq. \ref{eq:Custom_Distance}, each feature is weighted by a specific $w_j\in[0, 1]$. Such weights, $\underline{\mathbf{w}}$, are calculated by exploiting the fitness function in Eq. \ref{eq:obj_kmeans}, which searches for the best-performing parameters configuration on a suitable validation set.
It is important to assess the importance of the considered features describing a fault pattern with respect to the classification problem at hand. Notably, we expect to find features that are, in average, more relevant than others, thus contributing with more impact in the discrimination process.
To verify such a hypothesis, in Fig. \ref{fig:density_weigth_k7} we show the estimated density (kernel-based estimator) of the weights related to the best-performing individual; $k=7$ is considered here.
From the figure it is possible to deduce that the features are not uniformly weighted.
It is worth noting that, regardless of the value of $k$, there is a subset of features that is always associated with high weights (details not shown). Such features are: ``Time start'', ``Primary station fault distance'', ``Median point'', ``Max temperature'', ``Rain'', ``Interruption (breaker)'', and ``Petersen alarms''.
Such features confirm what the expertise of the ACEA company indicates as most important factors congruent to a LF.
The time of the day (``Time start'') is an important variable due to the changing on energy demand that normally presents two peaks, one at the middle of the day and one in the later evening.
The distance from the PS (``Primary station fault distance'') and the absolute position of the LF (``Median point'') is also a characterizing property, confirming the hypothesis on the amount of electric current that flow along a backbone -- see Sec. \ref{sec:Spatial_Data}.
Other important indicators advised by the ACEA company experts are the weather conditions, specially the millimeters of rains and the maximum temperature in a day. In fact, it is well-known (and it is also reasonable) that the events of heavy rain are strongly correlated with grid black-outs.
Accordingly, a strong discrimination is conveyed by the sequences of automatically registered events. In particular, the interruptions registered in the PSs due to the opening of the prevention breakers caused by short circuits on the power line and the Petersen alarms, which are registered as soon as a loss of the dielectric capacity of the power equipments is detected.
This phenomena can be physically characterized by overheating (due to the high currents) in which the dielectric of power equipments changes his properties for a short period.
The presence of bursts of these events could be indicative of an imminent LF. This last perspective will be studied in detail in future research studies.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 842 500,scale=0.35,keepaspectratio=true]{./density_weigth_k7}
\caption{Density reconstructed for the best solution -- the weights of the dissimilarity measure (\ref{eq:Custom_Distance}) -- found by the OCC with $k=7$. The density is not uniform over the $[0, 1]$ range.}
\label{fig:density_weigth_k7}
\end{figure}
The OCC system is trained by cross-validation, exploiting the fitness function in Eq. \ref{eq:obj_kmeans}.
Here we aim to demonstrate that the solutions found by the OCC become more informative as the system finds better solutions, i.e., with solutions that improve the discrimination of faults/non-faults.
This result helps us in justifying the claim that the final solutions found by the proposed OCC are informative (the considered features have different importance in the discrimination process).
As previously described, the synthesis of the OCC consists in performing a certain number of iterations until the stop criterion in reached.
An important observation is that, during the iterations characterizing the optimization, the fitness of the best-performing individual has a non-decreasing trend (this is obtained since we use a form of elitism in our GA). Accordingly, we expect to observe a non-decreasing trend also for what concerns the estimated entropy of the distribution underlying the weights $\underline{\mathbf{w}}$ related to the best-performing individual of each iteration.
To demonstrate such a claim, we calculated the (non-linear) correlation among the sequence of fitness values and the sequence of estimated entropy values, both related to the best-performing solutions found at each iteration/evolution.
Fig. \ref{fig:mi} shows those two sequences for a specific test performed with $k=7$. Although at the beginning the two series are not very correlated, they stabilize to a similar trend that is captured via the estimation of the mutual information \cite{Moon_MI__2002}.
The MI column in Tab. \ref{tab:results_acea} reports hence the average mutual information estimated between those two series for each $k$; MI values fall within the $[0, 1]$ range.
As it is possible to observe, the non-linear correlation is in general good, especially in the $k=7$ case, where also the MI reaches its maximum score.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 410 302,scale=0.7,keepaspectratio=true]{./fitness-entropy_affnorm}
\caption{Sequences of fitness values and entropy estimations on the weights corresponding to the best individual solution found at each evolution of the OCC model optimization.}
\label{fig:mi}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
Predicting faults in real-world smart grid systems is a challenging task. This is due to the high variability of those types of systems and also to the heterogeneity of the data that actually characterize a fault situation.
In our study, we modeled localized faults in the ACEA smart grid by means of several heterogeneous features.
The proposed one-class classifier is based on an interplay among clustering and dissimilarity measure learning techniques, where specialized measure have been designed to deal with each feature type (e.g., categorical, metric, and time series). The classifier synthesis is guided by a genetic algorithm, which is in charge to optimize the weighting parameters of the dissimilarity measure adopted in the input domain, as well as the tolerances defining decision region boundaries.
The proposed system is able to provide both hard (i.e., Boolean) and soft decisions regarding the recognition of a test pattern.
Soft decisions are introduced also to offer an overall measure of reliability concerning the decisions; non reliable decisions can be individuated by calculating the fuzzy entropy of the resulting membership function.
Experimental evaluations performed on the ACEA data demonstrate the effectiveness of the proposed solution in a real-world smart grid setting.
Future directions include the possibility to evaluate other clustering algorithms and different global optimization schemes.
Particular interest will be given to the study of the time series of electrical current values measured in different areas of the ACEA power grid, aiming to characterize the underlying (dynamic) system generating such data with the aim to find correlation/causation rules with the observed faults.
\bibliographystyle{abbrvnat}
|
\section{Introduction}
Since its launch occurred on February 11 2010, the {\em{Atmospheric Imaging Assembly}} in the {\em{Solar Dynamics Observatory (SDO/AIA)}} has been contributing to our knowledge of solar flare emission, permitting full disk imaging in $10$ wavebands (seven of which are centered at EUV wavelengths), at a spatial resolution of about $1$ arcsec and with a temporal cadence of $12$ sec in each of the EUV bands. However, the scientific potential of {\em{SDO/AIA}} images is limited by the presence of significant saturation, which flattens and muddles the brightest, physically most interesting core of such images for even relatively modest flares.
{\em{Saturation}} refers to the condition where, in a CCD-based imaging system, pixels lose their ability to accommodate additional charge ({\em{primary saturation}}), causing charge to spill into their neighbors. This secondary effect, named {\em{blooming}}, also induces bright rays typically along the north-south (for AIA) axis in the image.
We now introduce a computational method that utilizes the diffraction component of the Point Spread Function (PSF) of {\em{SDO/AIA}} \citep{boetal12,poetal13,grsuwe11} to recover the signal in the primary saturation region. The PSF of each {\em{AIA}} telescope models the system response to a point source and describes blurring, dispersion, and diffraction. The various {\em{AIA}} PSFs depend on the wavelength of the incoming photons and present a two-component structure which is similar to those of the {\em{TRACE}} passbands \citep{gbsyma06}. The main idea of our approach is to exploit, in particular, the diffraction pattern showing up {\em{outside}} the core of the saturated image to recover information {\em{inside}} the saturated region. The feasibility of this approach was demonstrated by previous analyses of {\em{TRACE}} and {\em{AIA}} diffraction patterns \citep{rakrli11,linita01,grsuwe11} where, however, the inverse diffraction problem is addressed by means of non-automatic, laborious and semi-heuristic procedures.
Our approach is based on three image-processing steps. In the first step ({\em{correlation}}) we identify the region of primary saturation distinguishing it from blooming on the basis of a data-dependent correlation procedure. In the second step ({\em{convolution}}) the diffraction fringes in the image are localized by convolving the diffraction PSF with the pixels in the region of primary saturation. The third step ({\em{back-projection}}) is used to reconstruct the true image in the primary saturation region using the information contained in the diffracted images. Details about these steps are given in Section 2. Section 3 illustrates the performances of the method in the case of {\em{SDO/AIA}} data recorded during the July 8 2013 event. Our conclusions are offered in Section 4.
\section{The reconstruction method}
The general model equation mimicking the signal formation process in an optical system is
\begin{equation}\label{general}
g = K * f + b,
\end{equation}
where $g$ is the measured data, $K$ is the global PSF, $f$ is the true flux distribution, and $b$ is the background. This equation applies to all points in the data image. However, in the case of the {\em{SDO/AIA}} image reconstruction problem the use of this equation should account for the fact that
\begin{itemize}
\item the true image can be recovered just in a sub-domain of the Field Of View (FOV) corresponding to the primary saturation region;
\item since the total diffracted image is very intense, the usable data, $g$, is just the one containing the diffraction pattern;
\item just the diffraction component of the PSF is useful for this operation.
\end{itemize}
In fact, the PSF associated with each {\em{SDO/AIA}} wavelength can be realistically modeled as
\begin{equation}\label{psf-components}
K(x,y) = K_C(x,y) + K_D(x,y)
\end{equation}
where $K_C(x,y)$ is the core, Gaussian-like component mimicking diffusion, and $K_D(x,y)$ is the diffraction component wherein about $20\%$ of the total flux is coherently scattered due to the two square wire grid meshes supporting the thin EUV filters. Figure \ref{fig:psf} shows the PSF of the $131~ \AA$ passband with, superimposed and rotated of $44$ deg, the one of the $304 ~\AA$ passband (we applied a threshold in order to show the diffraction patterns over the full range of the FOV). The figure points out the presence of 8 arms of diffraction for each passband, how diffraction scales with wavelength and how dispersion in the diffraction peaks scales with respect to dispersion in the core.
\begin{figure}[pht]
\includegraphics[width=0.5\textwidth]{fig1}
\caption{The {\em{SDO/AIA}} Point Spread Function. The $304~ \AA$ PSF is superimposed to the $131~ \AA$ PSF and rotated of $44$ deg. A threshold has been applied in order to show the diffraction patterns over the full range of the FOV.}
\label{fig:psf}
\end{figure}
The diffraction component of the PSF plays a crucial role in the de-saturation process described in this Section. The workflow of such a process is illustrated in Figure \ref{fig:ingredients} for a flare occurring on July 8 2013, measured at $131~ \AA$ (more details on this event will be given in Section 3). The computational steps at the basis of its realization are:
\begin{description}
\item{\bf{Correlation.}} Let us denote with $S$ the overall saturation domain in the {\em{AIA}} image in the top left panel of Figure \ref{fig:ingredients}. $S$ is the sum of the primary saturation region $PS$ and the blooming region $B$ and all its pixels have intensity of around $16000$ DN pixel$^{-1}$. Such pixels are masked away to obtain the image $g^{\prime}$ represented in the top right panel of Figure \ref{fig:ingredients}. The back-projection matrix $K_D^T$ is the transpose of the diffraction matrix $K_D$ and has size equal to the number of unsaturated pixels times the number of saturated pixels. We computed the correlation map
\begin{equation}\label{correlation-map}
C = K_D^T g^{\prime}
\end{equation}
for all pixels in $S$, and show it in the middle left panel of Figure \ref{fig:ingredients}. Since primary saturation is identified by those pixels in $S$ for which $C$ is high, a thresholding procedure allowed the identification of $PS$ in the correlation map.
\item{\bf{Convolution.}} The white curve in Figure \ref{fig:ingredients}, middle right panel, bounds the domain $PS$ in the observed saturated image where information can be recovered by means of image reconstruction. We computed the convolution between $PS$ and $K_D$ to localize the diffraction fringe in $g^{\prime}$ and utilized just that information, indicated with $g$ in the reconstruction procedure. In fact $g$ contains pixels where the diffraction effect is most significant.
\item{\bf{Back-projection.}} The reconstruction of information in the region of primary saturation $PS$ required the computation of the back-projection
\begin{equation}\label{back-projection}
f = K_D^T g
\end{equation}
where $f$ is defined in $PS$ and here $K_D^T$ has size equal to the number of pixels in the diffraction fringes times the number of pixels in the primary saturation region. The result is the de-saturated image at the bottom left panel of Figure \ref{fig:ingredients} where the blooming effects are still present, the diffraction pattern is clearly visible and the recovery of information in the region originally affected by primary saturation is evident. Such a recovery of information is even more evident in the last, bottom right panel of the figure, a zoom on the de-saturated domain. We note that the use of the log scale better shows the presence of the diffraction fringes with respect to the reconstructed signal.
\end{description}
From a computational viewpoint, the correlation map $C$ in (\ref{correlation-map}) and the back-projected map $f$ in (\ref{back-projection}) have been obtained by applying Expectation Maximization (EM) \citep{deetal77} to $g^{\prime}$ and $g$ respectively.
EM (also known as the Lucy-Richardson method in this context \citep{lu74}) realizes maximum likelihood in the case of Poisson data and imposing positivity on the solution space (CCD data are more correctly quasi-Poisson but EM effectiveness relies more on positivity than on the way data-prediction discrepancy is computed). EM iteration in the back-projection case is
\begin{equation}\label{EM}
f^{k+1} = \frac{f^{k}}{K_D^T I} K^T_D \left( \frac{g}{K_Df^k + b}\right),
\end{equation}
where $I$ is the unit vector and $b$ is an estimate of the background. In our implementation, the stopping rule that avoids numerical instability is based on the concept of constrained residual \citep{bepi14} and imposes that the empirical value of this random variable coincides with its expectation value \citep{beetal13}. We point out that for bright saturated images which are intrinsically small, diffraction places almost perfect copies over the fringes of the eight separate arms. In this case, the back-projection side-lobes are unambiguously removed and the formulation becomes a version of 2-D epoch-folding completely analogous to 1-D pulse synthesis \citep{leelwe83}.
\begin{figure*}[pht]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=6.cm]{fig2a} &
\includegraphics[width=6.cm]{fig2b} \\
\includegraphics[width=6.cm]{fig2c} &
\includegraphics[width=6.cm]{fig2d} \\
\includegraphics[width=6.cm]{fig2e} &
\includegraphics[width=6.cm]{fig2f}
\end{tabular}
\caption{From a saturated to a de-saturated {\em{AIA}} image using correlation/reconstruction. All maps are log-scaled and the green-white linear color table is used for their representation. Top left panel: July 8 2013 event at 01:21:32 UT, observed at $131~ \AA$ (the white artifact around the left side of the saturation region was present in the original data and may be due to the pre-processing step; it has been left untouched by our analysis). Top right panel: the same image but with the saturation region masked. Middle left panel: representation of the correlation between the diffraction PSF and the masked image. Middle right panel: the primary saturation region is bounded by the white contour in the observed image. Bottom left panel: the reconstruction offered by EM. Bottom right panel: zoom on the reconstructed core region.}
\label{fig:ingredients}
\end{center}
\end{figure*}
\begin{figure*}[pht]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=6.3cm]{fig3a} &
\includegraphics[width=6.3cm]{fig3b} \\
\includegraphics[width=6.3cm]{fig3c} &
\includegraphics[width=6.3cm]{fig3d} \\
\includegraphics[width=6.3cm]{fig3e} &
\includegraphics[width=6.3cm]{fig3f}
\end{tabular}
\caption{The de-saturation method at work. The event is the July 8 2013 one. The first row corresponds to data in the $131 ~\AA$ wavelength at times 01:21:56 UT; the second row corresponds to data in the $211~ \AA$ wavelength at times 01:22:14 UT; the third row corresponds to data in the $304~ \AA$ wavelength at times 01:21:31 UT. For each dataset, we provided the saturated image and the corresponding de-saturated image, both in linear scale.}
\label{fig:de-saturated}
\end{center}
\end{figure*}
\section{Applications}
On July 8 2013 the Sun experienced a C9.7 GOES class flare in the time range 01:13 - 01:23 UT with peak at around 01:22 UT. This event was very well observed by {\em{RHESSI}} but, despite the moderate peak flux, the corresponding {\em{AIA}} images saturated at most wavelengths and times. In principle, simultaneous solar flare observations with {\em{SDO}} and {\em{RHESSI}} provide spatially resolved information on hot plasma and energetic particles during flares and the unavailability of EUV maps due to saturation clearly limits this kind of analysis. As a test of our de-saturation approach we considered this event and, specifically, we applied the method to a number of saturated images collected at three different wavelengths ($131 ~\AA$, $211~ \AA$ and $304~ \AA$) and at two times for each wavelength. We downloaded {\em{AIA}} level 1.5 data from http://www.lmsal.com/get$\_$aia$\_$data and used a slightly modified version of the routine "aia$\_$calc$\_$psf.pro" in Solar SoftWare (SSW) for the generation of the diffraction PSF \citep{grsuwe11} (since our method strongly relies on the characteristics of the applied PSF, it would be interesting to run the reconstruction algorithm also in the case of the semi-empirical PSF described by \citet{poetal13}). The results of our analysis are given in Table \ref{table:c-stat} that reports the Cash statistic values predicted by the de-saturated images. Specifically, we have computed the C-statistic of a set made of $100$ random samples of the fringes, repeated this process $10$ times (every time changing the set of random samples) and reported the average values over these ten times (we point out that the C-statistic values may suffer because our computation of the PSF does not allow for an emission line distribution within the AIA passbands). Further, Figure \ref{fig:de-saturated} visually compares saturated and de-saturated maps corresponding to one of the two observation times for each wavelength. We point out that the linear scale softens the effect of diffraction fringes and blooming (blooming cannot be eliminated by means of reconstruction but its visual effect could be further reduced by means of interpolation) .
The potential of this method for a more extensive use of EUV data is illustrated in the example of Figure \ref{fig:RHESSI-AIA}. To construct it, we first considered a {\em{RHESSI}} image of hard X-ray flux reconstructed by Pixon \citep{pugo05} in the time interval 01:22:14 - 01:22:26 UT for the energy channel $12-25$ keV. In the figure, the contours of this X-ray image are superimposed on the de-saturated $131 \AA$ image corresponding to the {\em{SDO/AIA}} observation occurred at 01:22:20 UT (an offset of around $2$ arcsec is visible between the {\em{RHESSI}} and {\em{AIA}} images).
\begin{table}
\begin{center}
\begin{tabular}{c|c|c}
wavelength & time (UT) & C-stat \\
\hline \hline
& 01:21:56 & 2.085 \\
$131 \AA$ & 01.22:20 & 3.739 \\
\hline
& 01:22:14 & 1.243 \\
$ 211 \AA$ & 01:22:23 & 2.046 \\
\hline
& 01:21:31 & 2.275 \\
$304 \AA$ & 01:21:43 & 2.796 \\
\end{tabular}
\end{center}
\caption{Cash statistic for SDO/AIA datasets acquired by three different wavebands. For each line, two acquisition times are considered.}
\label{table:c-stat}
\end{table}
\section{Conclusions}
The EUV analysis of flaring events by means of {\em{SDO/AIA}} is notably limited by saturation and blooming, occurring at all wavelengths and times even in the case of moderate peak flux. This Letter illustrates a method that addresses, for the first time in a systematic way, the problem of eliminating primary saturation from these maps, utilizing standard image processing techniques like correlation, convolution and statistical deconvolution (the problems of recovering information in the blooming region and removing the multiple diffraction components will be addressed in future research). We applied our approach to data recorded during the July 8 2013 flare at three different EUV wavelengths and showed its effectiveness to reduce saturation in a variety of artifact configurations and intensities. Once fully automated, these de-saturation procedures will be included in Solar SoftWare. Further we will be extending this technique to include the wavelength dependent dispersion of the PSF to integrate the method of \citet{rakrli11} into this schema.
The impact of this method on the exploitation of EUV information in solar flares can be notable, for example in imaging spectroscopy studies of the dynamic response of solar corona and chromosphere to solar flares. Another broad and significant applicability domain involves multi-spectral studies. For example, combined analysis of {\em{SDO/AIA}} and {\em{RHESSI}} data may provide information on crucial plasma parameters like differential emission measure, temperature and density, which would notably contribute to unveil still mysterious processes concerned with energy transport mechanisms in flares.
\begin{figure}[pht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=7.cm]{fig4}
\end{tabular}
\caption{Integration of {\em{RHESSI}} data and SDO/AIA de-saturated images in the case of the July 8 2013 event. Contours of a RHESSI image reconstructed by Pixon in the time interval 01:22:14 - 01:22:26 UT in the energy channel $12-25$ keV are superimposed on the de-saturated linear-scale AIA map acquired by the $131~ \AA$ telescope at 01:22:20 UT.}
\label{fig:RHESSI-AIA}
\end{center}
\end{figure}
\begin{acknowledgments}
Richard Schwartz was supported by NASA grant NNX14AG06G. Gabriele Torre was supported during his stay at Goddard by The Catholic University of America in Washington, DC, using {\em{RHESSI}} funds under grant number NNX11AB37G. Anna Maria Massone and Federico Benvenuto are kindly acknowledged for fruitful discussions.
\end{acknowledgments}
|
\section{#1}\setcounter{equation}{0}}
\def\thesection{\arabic{section}}
\newcommand{\bi}[1]{\bibitem{#1}}
\newcommand{\fr}[2]{\frac{#1}{#2}}
\def\kk{\vec{k}}
\def\m{\rm{m}}
\def\D{\rm{D}}
\def\Omo{\Omega_{\rm{m}0}}
\def\Oma{\Omega_{\rm{m}}(a)}
\def\Omz{\Omega_{\rm{m}}(z)}
\def\w{\omega}
\def\obs{\rm{obs}}
\def\Ref{\rm{ref}}
\def\res{\rm{res}}
\def\red{\rm{red}}
\def\EdS{\rm{EdS}}
\def\Geff{G_{\rm{eff}}}
\def\GN{G_{\rm{N}}}
\newcommand{\oo}{\omega_{0}}
\newcommand{\ob}{\omega_{b}}
\newcommand{\oa}{\omega_{a}}
\newcommand{\Ob}{{\cal O}}
\newcommand{\rhom}{\rho_{\rm{m}}}
\newcommand{\rhode}{\rho_{\rm{DE}}}
\newcommand{\rhodeo}{\rho_{\rm{DE}0}}
\newcommand{\Odeo}{\Omega_{\rm{DE}0}}
\newcommand{\der}{\rm{der}}
\newcommand\Perms[2]{\tensor[_{#2}]P{_{#1}}}
\newcommand{\Fapp}{\widetilde{F}}
\newcommand{\PRD}{Phys.\ Rev.\ D}
\newcommand{\PRL}{Phys.\ Rev.\ Lett.}
\newcommand{\JCAP}{J.\ Cosmol.\ Astropart.\ Phys.}
\newcommand{\MNRAS}{Mon.\ Not.\ Roy.\ Astron.\ Soc.}
\newcommand{\AST}{astro-ph}
\newcommand{\ApJ}{Astrophys.\ J.\ }
\newcommand{\DE}{\rm{DE}}
\def\ga{\mathrel{\raise.3ex\hbox{$>$\kern-.75em\lower1ex\hbox{$\sim$}}}}
\def\la{\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}}
\begin{document}
\leftline{KIAS-P14014}
\title{Exact Third-Order Density Perturbation and One-Loop Power Spectrum in General Dark Energy Models}
\author{Seokcheon Lee$^{1}$, Changbom Park$^{1}$, and Sang Gyu Biern$^{2}$}
\affiliation{
$^1$School of Physics, Korea Institute for Advanced Study, Heogiro 85, Seoul 130-722, Korea\\
$^2$Department of Physics and Astronomy, Seoul National University, Seoul 151-747, Korea}
\begin{abstract}
Under the standard perturbation theory (SPT), we obtain the fully consistent third-order density fluctuation and kernels for the general dark energy models without using the Einstein-de Sitter (EdS) universe assumption for the first time. We also show that even though the temporal and spatial components of the SPT solutions can not be separable, one can find the exact solutions to any order in general dark energy models. With these exact solutions, we obtain the less than \% error correction of one-loop matter power spectrum compared to that obtained from the EdS assumption for $k = 0.1 {\rm h\, Mpc}^{-1}$ mode at $z = 0$ (1, 1.5). Thus, the EdS assumption works very well at this scale. However, if one considers the correction for $P_{13}$, the error is about 6 (9, 11) \% for the same mode at $z = 0$ (1, 1.5). One absorbs $P_{13}$ into the linear power spectrum in the renormalized perturbation theory (RPT) and thus one should use the exact solution instead of the approximation one. The error on the resummed propagator $N$ of RPT is about 14 (8, 6) \% at $z =0$ (1, 1.5) for $k = 0.4 {\rm h\, Mpc}^{-1}$. For $k = 1 {\rm h\, Mpc}^{-1}$, the error correction of the total matter power spectrum is about 3.6 (4.6, 4.5) \% at $z = 0$ (1, 1.5). Upcoming observation is required to archive the sub-percent accuracy to provide the strong constraint on the dark energy and this consistent solution is prerequisite for the model comparison.
\end{abstract}
\pacs{95.36.+x, 98.65.-r, 98.80.-k
\maketitle
\setcounter{equation}{0}
The standard perturbation theory (SPT) has been widely used to investigate the correction to the linear power spectrum in a quasi-nonlinear regime. The recent progress and the development of alternative analytical methods have been made \cite{0112551, 13112724}. The approximate recursion relations for the Fourier components of the $n$-th order matter density fluctuation $\hat{\delta}_{n}(\tau,\kk)$ and the divergence of the peculiar velocity $\hat{\theta}_{n}(\tau,\kk)$ has been obtained for the Einstein-de Sitter (EdS) universe \cite{Goroff, Makino}. When one extends the SPT to the general background universe, one uses the assumption that the dependence of the SPT solutions on the cosmological parameters is encoded in the linear growth factor, $D_{1}(a)$ \cite{0112551}. This is also confirmed for the dark energy models \cite{9807211, 08061437}. However, this argument is partly correct because one also needs to investigate the error on the power spectrum induced from EdS assumption ({\it i.e.} the value of the linear growth rate is equal to that of the square root of the matter energy density contrast, $f_1 \equiv \fr{d \ln D_{1}}{d \ln a} = \sqrt{\Omega_{m}}$ ). We obtain the exact kernels for $\hat{\delta}_{n}$ and $\hat{\theta}_{n}$ without using EdS assumption and study its effect on the power spectrum.
The renormalized perturbation theory (RPT) tries to reorganize the perturbative series expansion of SPT and resums some of the terms into a function that can be factorized out of the series \cite{0509418, 0509419}. This function is called as the resummed propagator and referred as $N$. All the kernels of the higher order power spectrum terms must be expressed as a product of kernels that correspond to full mode coupling terms and full propagator terms in order to make the resummation possible. If the kernels are approximated as a product of one-loop propagator kernels, then the resummed propagator is given by $N(k) \equiv \exp [P_{13}(k)/P_{{\rm lin}}(k)]$. We find that $P_{13}(k)$ using EdS assumption causes 6 $\sim$ 11 \% errors for $k = 0.5 {\rm h\, Mpc}^{-1}$ mode at $z = 0 \sim 1.5$ and these induce errors on $N$ about 11 $\sim$ 20 \%.
In addition to SPT, the Lagrangian perturbation theory (LPT) is an another widely used analytic technique for the quasi-linear perturbative expansion. There also have been studies to investigate the dark energy dependence on the linear growth factor in LPT \cite{12034260, 14012226}. Recently, we also obtain the kernels in the recursion relations without using EdS assumption in the LPT and investigate its consequences on the one-loop power spectrum \cite{14043813}.
In this {\em Letter}, we obtain the exact relations for the temporal and spatial components of the SPT solutions in general dark energy models up to third order. When we obtain the kernels, we remove the EdS assumption in the derivation and investigate the its effects on the observable quantities.
The equations of motion of $\hat{\delta}(\tau,\kk)$ and $\hat{\theta}(\tau,\kk)$ in the Fourier space are given by
\ba \fr{\partial \hat{\delta}}{\partial \tau} + \hat{\theta} &=& - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \hat{\theta} (\tau,\vec{k}_1) \hat{\delta} (\tau,\vec{k}_2) \label{massFT} \, , \\
\fr{\partial \hat{\theta}}{\partial \tau} + {\cal H} \hat{\theta} + \fr{3}{2} \Omega_{m} {\cal H}^2 \hat{\delta} &=& - \fr{1}{2} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \beta(\vec{k}_1, \vec{k}_2) \hat{\theta} (\tau,\vec{k}_1) \hat{\theta} (\tau,\vec{k}_2) \label{EulerFT} \, ,
\ea
where $\tau$ is the conformal time, $\vec{k}_{12} \equiv \vec{k}_1 + \vec{k}_2$, $\delta_{\D}$ is the Dirac delta function, ${\cal H} \equiv \fr{1}{a} \fr{\partial a}{\partial \tau}$, $\Omega_{m}$ is the matter energy density contrast, $\alpha(\vec{k}_1, \vec{k}_2) \equiv \fr{\vec{k}_{12} \cdot \vec{k}_1}{k_1^2}$, and $\beta(\vec{k}_1, \vec{k}_2) \equiv \fr{k_{12}^2 (\vec{k}_1 \cdot \vec{k}_2)}{k_1^2 k_2^2}$.
Due to the mode coupling of the nonlinear terms shown in the right hand side of Eqs. (\ref{massFT}) - (\ref{EulerFT}), one needs to make a perturbative expansion in $\hat{\delta}$ and $\hat{\theta}$ \cite{0112551}. One can introduce the proper perturbative series of solutions for the fastest growing mode $D_{n}$
\ba \hat{\delta}(\tau,\vec{k}) &\equiv& \sum_{n=1}^{\infty} \hat{\delta}^{(n)} (\tau,\vec{k}) \label{hatdeltaS} \, , \\
\hat{\theta}(\tau,\vec{k}) &\equiv& \sum_{n=1}^{\infty} \hat{\theta}^{(n)} (\tau,\vec{k}) \label{hatthetaS} \, , \ea
where one can define the each order solution as
\ba \hat{\delta}^{(1)} (a,\vec{k}) &\equiv& D_{1}(a) \delta_{1}(\vec{k}) \, , \label{delta1} \\
\hat{\theta}^{(1)} (a,\vec{k}) &\equiv& D_{\theta 1}(a) \theta_{1}(\vec{k}) \equiv -a {\cal H} \fr{d D_1}{da} \delta_{1}(\vec{k}) \, , \label{theta1} \\
\hat{\delta}^{(2)} (a,\vec{k}) &\equiv& \sum_{i=1}^{2} D_{2i}(a) K_{2i}(\vec{k}) \equiv D_{1}^2(a) \sum_{i=1}^{2} c_{2i}(a) K_{2i}(\vec{k}) \, , \label{delta2} \\
\hat{\theta}^{(2)} (a,\vec{k}) &\equiv& \sum_{i=1}^{2} D_{\theta 2i}(a) K_{2i}(\vec{k}) \equiv a {\cal H} D_{1} \fr{d D_1}{da} \sum_{i=1}^{2} c_{\theta 2i}(a) K_{2i}(\vec{k}) \, , \label{theta2} \\
\hat{\delta}^{(3)} (a,\vec{k}) &\equiv& \sum_{i=1}^{6} D_{3i}(a) K_{3i}(\vec{k}) \equiv D_{1}^3(a) \sum_{i=1}^{6} c_{3i}(a) K_{3i}(\vec{k}) \, , \label{delta3} \\
\hat{\theta}^{(3)} (a,\vec{k}) &\equiv& \sum_{i=1}^{6} D_{\theta 3i}(a) K_{3i}(\vec{k}) \equiv a {\cal H} D_{1}^2 \fr{d D_1}{da} \sum_{i=1}^{6} c_{\theta 3i}(a) K_{3i}(\vec{k}) \, , \label{theta3} \ea
To be consistent with the current observation, we consider the dark energy dominated flat universe as a background model. It has been known that the $n$-th order fastest growing mode solutions are proportional to the $n$-th power of the linear growth factor $D_{1}$ ({\it i.e.} $D^{n} \propto D_{1}^n$) for the EdS universe. And this is not true for the general background models. There have been the investigations of the validity of these ansatz (\ref{hatdeltaS}) and (\ref{hatthetaS}) by using the different growth rates for $\hat{\delta}$ and $\hat{\theta}$ \cite{9807211, 08061437}. However, the improper decomposition of fastest mode solutions and the incorrect initial conditions are used for the $n$-th order growth rate in both cases (see Appendix).
If one takes a derivatives of Eq. (\ref{massFT}) and replace Eq. (\ref{EulerFT}) into it, then one obtains
\ba \fr{\partial^2 \hat{\delta}}{\partial \tau^2} + {\cal H} \fr{\partial \hat{\delta}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 \hat{\delta} &=& - {\cal H} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \hat{\theta} (\tau,\vec{k}_1) \hat{\delta} (\tau,\vec{k}_2) \nonumber \\ &&- \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \Biggl[ \fr{\partial \hat{\theta} (\tau,\vec{k}_1)}{\partial \tau} \hat{\delta} (\tau,\vec{k}_2) + \hat{\theta} (\tau,\vec{k}_1) \fr{\partial \hat{\delta} (\tau,\vec{k}_2)}{\partial \tau} \Biggr] \nonumber \\ && + \fr{1}{2} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \beta(\vec{k}_1, \vec{k}_2) \hat{\theta} (\tau,\vec{k}_1) \hat{\theta} (\tau,\vec{k}_2) \label{massEulerFT} \, ,
\ea
From the Eqs.(\ref{massFT}) and (\ref{massEulerFT}) , one obtains the expressions for the higher order solutions of $\hat{\delta}^{(2)}$, $\hat{\theta}^{(2)}$, and $\hat{\delta}^{(3)}$ as
\ba \hat{\delta}^{(2)}(a,\vec{k}) &\equiv& D_{21}(a) K_{21}(\vec{k}) + D_{22}(a) K_{22}(\vec{k}) \equiv D_{1}^2 \Biggl[ c_{21}(a) K_{21}(\vec{k}) + c_{22}(a) K_{22}(\vec{k}) \Biggr] \equiv D_{1}^2(a) \delta_{2}(a,\vec{k}) \nonumber \\
&\equiv& D_{1}^2 \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) F_{2}^{(s)}(a, \vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \, , \label{delta22} \\
\hat{\theta}^{(2)}(a,\vec{k}) &\equiv& D_{\theta 21}(a) K_{21}(\vec{k}) + D_{\theta 22}(a) K_{22}(\vec{k}) \equiv D_1 \fr{\partial D_{1}}{\partial \tau} \Biggl[ c_{\theta 21}(a) K_{21}(\vec{k}) + c_{\theta 22}(a) K_{22}(\vec{k}) \Biggr] \equiv D_1 \fr{\partial D_{1}}{\partial \tau} \theta_{2}(a,\vec{k}) \nonumber \\
&\equiv& -D_1 \fr{\partial D_{1}}{\partial \tau} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) G_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \, , \label{theta22} \\
\hat{\delta}^{(3)}(a,\vec{k}) &\equiv& D_{31}(a) K_{31}(\vec{k}) + \cdots +D_{36}(a) K_{36}(\vec{k}) \equiv D_1^3(a) \Biggl[ c_{31}(a) K_{31}(\vec{k}) + \cdots + c_{36}(a) K_{36}(\vec{k}) \Biggr] \nonumber \\ &\equiv& D_1^3(a) \int d^3 k_1 d^3 k_2 d^3 k_3 \delta_{\D}(\vec{k}_{123} - \vec{k}) F_{3}^{(s)}(a,\vec{k}_1,\vec{k}_2,\vec{k}_3) \delta_{1}(\vec{k}_1) \delta_{1}(\vec{k}_2) \delta_{1}(\vec{k}_3) \, , \label{delta32} \ea
where
\ba c_{2i} &=& \fr{D_{2i}}{D_1^2}\, , \,\,\,\, c_{\theta 2i} = \fr{D_{\theta 2i}}{D_1} \Bigl( \fr{\partial D_{1}}{\partial \tau} \Bigr)^{-1}\, , \,\,\,\, c_{3i} = \fr{D_{3i}}{D_1^3} \, , \label{ci} \\
K_{21}(\vec{k}) &=& -\int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) \delta_1 (\vec{k}_2) \, , \label{K21} \\
K_{22}(\vec{k}) &=& - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \beta(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) \theta_1 (\vec{k}_2) \, , \label{K22} \\
F_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) &=& \fr{1}{2} \Biggl[ c_{21} \Bigl( \fr{\vec{k}_{12} \cdot \vec{k}_1}{k_1^2} + \fr{\vec{k}_{12} \cdot \vec{k}_2}{k_2^2} \Bigr) - 2 c_{22} \fr{k_{12}^2 (\vec{k}_1 \cdot \vec{k}_2)}{k_1^2 k_2^2} \Biggr] \label{F2s} \\
&=& c_{21} -2 c_{22} \Biggl(\fr{\vec{k}_1 \cdot \vec{k}_2}{k_1 k_2} \Biggr)^2 + \fr{1}{2} \Bigl(c_{21} -2 c_{22} \Bigr) \vec{k}_1 \cdot \vec{k}_2 \Biggl(\fr{1}{k_1^2} + \fr{1}{k_2^2} \Biggr) \, , \nonumber \\
G_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) &=& \fr{1}{2} \Biggl[ -c_{\theta 21} \Bigl( \fr{\vec{k}_{12} \cdot \vec{k}_1}{k_1^2} + \fr{\vec{k}_{12} \cdot \vec{k}_2}{k_2^2} \Bigr) + 2 c_{\theta 22} \fr{k_{12}^2 (\vec{k}_1 \cdot \vec{k}_2)}{k_1^2 k_2^2} \Biggr] \label{G2s} \\
&=& - c_{\theta 21} +2 c_{\theta 22} \Biggl(\fr{\vec{k}_1 \cdot \vec{k}_2}{k_1 k_2} \Biggr)^2 - \fr{1}{2} \Bigl(c_{\theta 21} -2 c_{\theta 22} \Bigr) \vec{k}_1 \cdot \vec{k}_2 \Biggl(\fr{1}{k_1^2} + \fr{1}{k_2^2} \Biggr) \, , \nonumber \\
F_{3}^{(s)}(a,\vec{k}_1,\vec{k}_2,\vec{k}_3) &=& \sum_{i=1}^{6} F_{3i}^{(s)}(a,\vec{k}_1,\vec{k}_2,\vec{k}_3) \, , \label{F3s}
\ea
where explicit forms of $F_{3i}^{(s)}$ are given in the appendix.
One can use the above equations to compute the power spectrum at any order in perturbation theory
\ba \langle \hat{\delta}(a,\vec{k}) \hat{\delta}(a,\vec{k}') \rangle &=&
D_{1}^2(a) \langle \delta_1(\vec{k}) \delta_1(\vec{k}') \rangle + D_1^4(a) \langle \delta_2(a, \vec{k}) \delta_2(a, \vec{k}') \rangle + D_1^4(a) \langle \delta_1(\vec{k}) \delta_3(a, \vec{k}') + \delta_3(a, \vec{k}) \delta_1(\vec{k}') \rangle \nonumber \\ &\equiv& \Bigl( D_1^2 P_{11}(k) + D_1^4 P_{22}(a,k) + 2 D_1^4 P_{13}(a,k) + \cdots \Bigr) \delta_{\D}(\vec{k}+\vec{k}') \equiv P(a,k) \delta_{\D}(\vec{k}+\vec{k}') \label{Pktau} \ea
The one-loop power spectrum is defined as
\be P_{2}(a,k) = D_1^4(a) \Biggl[ P_{22}(a,k) + 2 P_{13}(a,k) \Biggr] \, , \label{P2} \ee where $P_{22}$ and $P_{13}$ are obtained as
\ba P_{22}(a,k) &=& 2 \int d^3 q P_{11}(q) P_{11} (|\vec{k}-\vec{q}\,|) \Bigl[ F_{2}^{(s)}(a,\vec{q}, \vec{k}-\vec{q}\,) \Bigr]^2 =
\fr{(2 \pi)^{-2} k^3}{2} \int_{0}^{\infty} dr P_{11}(kr) \nonumber \\ &\times& \int_{-1}^{1} dx P_{11} \Bigl( k\sqrt{1+r^2-2rx} \Bigr) \Biggl[ \fr{(c_{21}+2c_{22}) r + (c_{21}-2c_{22}) x -2 c_{21} r x^2}{(1 + r^2 - 2rx)} \Biggr]^2 \, , \label{P22k} \\
2 P_{13}(k) &=& 6 P_{11}(k) \int d^3 q P_{11}(q) F_{3}^{(s)} (a, \vec{q}, \, -\vec{q}, \, \vec{k}\,) \label{P13k} \\
&=& (2\pi)^{-2} k^3 P_{11}(k) \int_{0}^{\infty} dr P_{11}(kr) \Biggl[ 2c_{35} r^{-2} -\fr{1}{3} \Bigl( 4 c_{31} -8 c_{32} +3c_{33} +24c_{35} - 16 c_{36} \Bigr) \nonumber \\ &-& \fr{1}{3}\Bigl(4 c_{31} -8c_{32} +12c_{33}-8c_{34}+6c_{35} \Bigr)r^2 + c_{33} r^4 + \Bigl(\fr{r^2-1}{r} \Bigr)^3 \ln \Bigl|\fr{1+r}{1-r} \Bigr| \Bigl(c_{35} - \fr{1}{2}c_{33}r^2 \Bigr) \Biggr] \nonumber \, ,\ea
where $r = \fr{q}{k}$ and $x = \fr{\vec{q} \cdot \vec{k}}{q k}$. The above equations (\ref{P22k}) and (\ref{P13k}) are identical to Eqs. (2.24)
and (2.25) of \cite{Makino} when one replace the coefficients of higher solutions $c_{2i}$ and $c_{3i}$ with those given in Eqs.(\ref{c2iEdS}) and (\ref{c3iEdS}). Thus, the terms with $c_{2i}$ and $c_{3i}$ represent the dark energy effect on the one-loop power spectrum.
Now we obtain the one-loop power spectrum for $\Lambda$CDM model. We run the camb to obtain the linear power spectrum \cite{camb} using $\Omega_{b0} = 0.044$, $\Omega_{m0} = 0.26$, $h = 0.72$, $n_{s} = 0.96$, and the numerical integration range for $q$ in Eqs. (\ref{P22k}) and (\ref{P13k}) is $10^{-6} \leq q \leq 10^{2}$.
\begin{center}
\begin{figure}
\vspace{1.5cm}
\centerline{\psfig{file=Fig1.eps, width=14cm}}
\vspace{-0.1cm}
\caption{Both the linear matter power spectra (thin lines) and the nonlinear matter power spectra with one-loop correction (thick lines) at $z = 0, 1.0,$ and 1.5 (solid, dotted, and dotdashed lines) for $\Omega_{m0} = 0.26$ $\Lambda$CDM model.} \label{fig1}
\end{figure}
\end{center}
In Fig. \ref{fig1}, we show both the linear power spectra $P_{\rm{L}}$ (thin lines) and the nonlinear power spectra $P_{\rm{NL}} = P_{\rm{L}} + P_{2}$ (thick lines) at the different redshift $z =$ 0 (solid), 1.0 (dotted), and 1.5 (dotdashed), respectively. We demonstrate the $\Lambda$CDM model with $\Omega_{m0} = 0.26$ in this figure. As one expects, the nonlinear power spectra are not simply enhanced by multiplying the differences of the square of the growth factor $D_1^2$ at the different redshifts. One also needs to emphasize that the exact kernels Eqs.(\ref{F2s}), (\ref{G2s}) and (\ref{F31s})-(\ref{F36s}) also depend on time. The coefficient of each kernel changes at the different observational epoch.
Now, we investigate the corrections in $P_{22}$ and $P_{13}$ compared to those using the EdS assumption.
As one expects, the effect of the removing EdS assumption on $P_{22}$ and $P_{13}$ becomes larger as $z$ increases.
This is due to the fact that we use the Gaussianity initial conditions for the perturbed quantities. The coefficients $c_{21}$-$c_{36}$ approach to those of $EdS$ models as $z$ decreases. This causes the fact that the kernels based on the EdS assumption deviate from the exact ones as $z$ decreases. Thus, the exact $P_{22}$ and $P_{13}$ show the larger deviations from the EdS assumed $P{22}$ and $P_{13}$ as $z$ increases. This is shown in Fig. \ref{fig2}. $P_{22}^{\rm{\Lambda CDM}}$ and $P_{13}^{\rm{\Lambda CDM}}$ mean the exact one loop corrections based on the $\Lambda$CDM models using the exact solution. While $P_{22}^{\rm{EdS}}$ and $P_{13}^{\rm{EdS}}$ mean the one loop corrections based on the EdS assumed kernels. In the left panel of Fig. \ref{fig2}, we show the errors in $P_{22}$ at the different redshift. The solid, dotted, and dotdashed lines correspond to errors of $P_{22}$ at $z = 0$, 1.0, and 1.5, respectively. The differences are about 5 (9, 11) \% for $k = 0.1 {\rm h\, Mpc}^{-1}$ mode at $z =$ 0 (1.0, 1.5). In the right panel of Fig. \ref{fig2}, we show the errors in $P_{13}$ at the different redshift. We use the same notation as the left panel. The differences between the exact and EdS assumed $P_{13}$ are about 6 (9, 11) \% for $k = 0.1 {\rm h\, Mpc}^{-1}$ mode at $z =$ 0 (1.0, 1.5).
\begin{figure}
\centering
\vspace{1.5cm}
\begin{tabular}{cc}
\epsfig{file=Fig2a.eps,width=0.5\linewidth,clip=} &
\epsfig{file=Fig2b.eps,width=0.5\linewidth,clip=} \\
\end{tabular}
\vspace{-0.5cm}
\caption{Errors in $P_{22}$ and $P_{13}$ a) Differences between the correct $P_{22}$ and the one with EdS assumption at the different epoches. The solid, dashed, and dotdashed lines correspond to $z = 0$, 1.0, and 1.5, respectively. a) Differences between the correct $P_{13}$ and the EdS assumed $P_{13}$ at different epoches.} \label{fig2}
\end{figure}
We show the corrections on $P_{\rm{NL}} \equiv P_{\rm{total}}$ and the resummed propagator $N$.
The one loop correction is sum of the $P_{22}$ and $P_{13}$. However, $P_{22}$ and $P_{13}$ have the different signs. Thus, if one considers the nonlinear power spectrum with the one loop correction, then the correction due to using the exact solution is very small compared to the nonlinear power spectrum based on EdS assumption. $P_{\rm{NL}} = P_{\rm{L}} + P_{2}$ where $P_{2} = P_{22} + P_{13}$. As we show in the Fig. \ref{fig2}, each correction at each mode is about same at any epoch. Thus, the corrections on $P_{2}$ are canceled each other. This is shown in the left panel of Fig. \ref{fig3}. $P_{\rm{total}}^{\rm{\Lambda CDM}}$ means the exact nonlinear matter power spectrum based on the $\Lambda$CDM models using the exact solution. While $P_{\rm{total}}^{\rm{EdS}}$ means the nonlinear matter power spectrum based on the EdS assumption. The solid, dotted, and dotdashed lines correspond to errors of $P_{\rm{total}}$ at $z = 0$, 1.0, and 1.5, respectively. The present nonlinear matter power spectrum is dominated by the one loop power spectrum at small scale $k \geq 0.1$. The correction for the total matter power spectrum is about 2 \% for $k = 0.4$ h/Mpc at any epoch. Thus, the EdS assumed nonlinear power spectrum is not a bad approximation. However, if one expands the SPT into RPT, then one needs to use the exact solution. This is shown in the right panel of Fig. \ref{fig3} by using the resummed propagator $N$. For the same mode, the deviations of $N$ from the exact values are about 14 (8, 6) \% at $z =$ 0 (1.0, 1.5). Thus, if one uses the EdS assumed nonlinear $P_{13}$, then one is not able to avoid these amount of errors on the $N$.
\begin{figure}
\centering
\vspace{1.5cm}
\begin{tabular}{cc}
\epsfig{file=Fig3a.eps,width=0.5\linewidth,clip=} &
\epsfig{file=Fig3b.eps,width=0.5\linewidth,clip=} \\
\end{tabular}
\vspace{-0.5cm}
\caption{Errors in $P_{\rm{total}}$ and $N$ a) Differences between the correct $P_{\rm{total}}$ and the one with $\lambda =1$ (EdS) assumption at the different epoches. The solid, dashed, and dotdashed lines correspond to $z = 0$, 0.5, and 1, respectively. a) Differences between the correct $P_{NL}$ and the $\lambda =1$ (EdS) assumed $P_{NL}$ at different epoches.} \label{fig3}
\end{figure}
The upcoming redshift surveys of galaxies such as BOSS, eBOSS, PFS, EUCLID, and MS-DESI will provide observational data of large scale structure
of the universe in larger volume with higher density. The analysis of these observational data requires very accurate theoretical modeling down to the quasi-linear regime. In this {\em Letter}, we present an accurate perturbation theory without adopting the EdS assumption. The obtained results are general for any background universe model including time varying dark energy models, and will be useful for studies of future surveys.
\section*{Acknowledgments}
This work were carried out using computing resources of KIAS Center for Advanced Computation. S.L would like to thank for the hospitality at APCTP during the program TRP.
\renewcommand{\theequation}{A-\arabic{equation}}
\setcounter{equation}{0}
\section*{APPENDIX}
In this section, we show the spatial and temporal solutions of the each order by using Eqs. (\ref{EulerFT}) and (\ref{massEulerFT}).
The equations for the first order solution of $\delta^{(1)}(\tau,\kk)$ and $\theta^{(1)}(\tau,\kk)$ are given by
\ba \Biggl[ \fr{\partial^2 D_{1}}{\partial \tau^2} + {\cal H} \fr{\partial D_{1}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{1} \Biggr] \delta_{1}(\vec{k}) &=& 0 \, , \label{delta12} \\
D_{\theta 1} \theta_{1}(\kk) &=& - \fr{\partial D_1}{\partial \tau} \delta_{1}(\kk) \, . \label{theta12} \ea
From the above Eqs. (\ref{delta12}) and (\ref{theta12}), one obtains
\ba && \fr{d^2 D_{1}}{da^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{1}}{da} - \fr{3}{2a^2} \Omega_{m} D_{1} = 0 \, , \label{D1eq} \\
&& D_{\theta 1}(a) = a {\cal H} \fr{\partial D_1}{\partial a}, \,\,\,\,
\theta_{1}(\kk) = - \delta_{1}(\kk) \, , \label{theta1k} \ea
where we use
\ba \fr{d D_1}{d \tau} &=& a {\cal H} \fr{d D_1}{da}, \,\,\,\,
\fr{d^2 D_1}{d \tau^2} = (a {\cal H})^2 \fr{d^2 D_1}{d a^2} + \Bigl( a {\cal H}^2 + a \fr{d {\cal H}}{d \tau} \Bigr) \fr{d D_1}{d a} \label{D1tau2} \, , \\
{\cal H} &=& a H, \,\,\,
\fr{d {\cal H}}{d \tau} = {\cal H}^2 + a^2 \fr{d H}{d t}, \,\,\,
\fr{1}{{\cal H}^2} \fr{d{\cal H}}{d \tau} = 1 + \fr{1}{H^2} \fr{d H}{d t} = 1 -\fr{3}{2} \Bigl( 1 + w \Omega_{de} \Bigr) \label{calHtau2} \, . \ea
If one uses the fact that the dark energy is dominated only at the late universe, then one can adopt the EdS conditions ({\it i.e.} $\Omega_{m} = 1$) for $D_1$ at early time ({it i.e.} $a_i$),
\be D_{1}(a_i) = a_i\, , \,\,\,\,\, {\rm and} \,\,\,\, \fr{d D_1}{da} \Bigl|_{a=a_i} = 1 \, . \label{D1EdS} \ee
Thus, one can obtain the exact solution for $D_1(a)$ for any dark energy model from Eqs.(\ref{delta12}) and (\ref{D1EdS}) except for the early dark energy one \cite{9408025,0012139,0504650}.
One can repeat the same process for $\hat{\delta}^{(2)}(\tau,\vec{k})$ to get
\ba \Biggl[ \fr{\partial^2 D_{21}}{\partial \tau^2} + {\cal H} \fr{\partial D_{21}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{21} \Biggr] K_{21}(\vec{k}) &=& \Biggl[{\cal H} D_{1} \fr{\partial D_{1}}{\partial \tau} + D_1 \fr{\partial^2 D_{1}}{\partial \tau^2} + \Bigl(\fr{\partial D_1}{\partial \tau} \Bigr)^2 \Biggr] \label{delta21} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) \delta_1 (\vec{k}_2) \Biggr] \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{22}}{\partial \tau^2} + {\cal H} \fr{\partial D_{22}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{22} \Biggr] K_{22}(\vec{k}) &=& -\fr{1}{2} \Bigl(\fr{\partial D_1}{\partial \tau} \Bigr)^2 \label{delta22} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \beta(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) \theta_1 (\vec{k}_2) \Biggr] \nonumber \, . \ea
If we adopt the initial zero non-Gaussianity of the higher order solutions ($\delta^{(n)} = 0$), then one can obtain the equations for the fastest growing mode solutions with the initial Gaussianity and the EdS initial conditions
\ba \fr{d^2 D_{21}}{da^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{21}}{da} - \fr{3}{2a^2} \Omega_{m} D_{21} &=& \fr{3}{2a^2} \Omega_{m} D_1^2 + \Bigl(\fr{d D_1}{da} \Bigr)^2 \label{D21eq} \\
&{\rm with}& D_{21}(a_i) = 0, \hspace{0.1in} \fr{d D_{21}}{da} \Bigl|_{a_i} = \fr{5}{7} a_i \nonumber \, , \\
\fr{d^2 D_{22}}{da^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{22}}{da} - \fr{3}{2a^2} \Omega_{m} D_{22} &=& -\fr{1}{2} \Bigl(\fr{d D_1}{da} \Bigr)^2 \label{D22eq} \\
&{\rm with}& D_{22}(a_i) = 0, \hspace{0.1in} \fr{d D_{22}}{da} \Bigl|_{a_i} = -\fr{1}{7} a_i \nonumber \, ,\ea
where we use the fastest growing mode solutions for the EdS universe
\be D_{21}^{(\EdS)} = \fr{5}{7} a^2 - \fr{5}{7} a_i a, \,\,\,\,
D_{22}^{(\EdS)} = -\fr{1}{7} a^2 + \fr{1}{7} a_i a \, . \label{D22EdS} \ee
Often it is knows as the EdS coefficient as
\be c_{21} = \fr{5}{7}\, , \,\,\,\, c_{22} = - \fr{1}{7} \, . \label{c2iEdS} \ee
However, this is not the coefficients for the fastest growing mode solutions because of the existence of the second terms in Eq.(\ref{D22EdS}).
From Eq.(\ref{EulerFT}), one can obtain equations for $\hat{\theta}^{(2)}$ by using other solutions
\ba D_{\theta 21}(a) K_{21}(\kk) + D_{\theta 22}(a) K_{22}(\kk) &=& a {\cal H} \Bigl[-\fr{d D_{21}}{d a} + D_1 \fr{d D_1}{d a} \Bigr] K_{21}(\kk) - a {\cal H} \fr{d D_{22}}{d a} K_{22}(\kk) \nonumber \, , \\
&\equiv& a {\cal H} D_{1} \fr{d D_{1}}{d a} \Bigl[ c_{\theta 21} K_{21}(\kk) + c_{\theta 22} K_{22}(\kk) \Bigr] \, . \label{theta22} \ea
From Eqs.(\ref{delta22})-(\ref{theta22})
\ba \hat{\delta}^{(2)}(a,\vec{k}) &\equiv& D_{21}(a) K_{21}(\vec{k}) + D_{22}(a) K_{22}(\vec{k}) \equiv D_{1}^2 \Biggl[ c_{21} K_{21}(\vec{k}) + c_{22} K_{22}(\vec{k}) \Biggr] \equiv D_{1}^2(a) \delta_{2}(a,\vec{k}) \nonumber \\
&=& D_{1}^2 \int d^3 k_{1} d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \Bigl[ c_{21}(a) \alpha(\vec{k}_1, \vec{k}_2) - c_{22}(a) \beta(\vec{k}_1, \vec{k}_2) \Bigr] \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \nonumber \\
&\equiv& D_{1}^2 \int d^3 k_{1} d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) F_{2}(a, \vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \label{delta23} \\
&\equiv& D_{1}^2 \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) F_{2}^{(s)}(a, \vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \nonumber \, , \\
F_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) &=& \fr{1}{2} \Biggl[ c_{21} \Bigl( \fr{\vec{k}_{12} \cdot \vec{k}_1}{k_1^2} + \fr{\vec{k}_{12} \cdot \vec{k}_2}{k_2^2} \Bigr) - 2 c_{22} \fr{k_{12}^2 (\vec{k}_1 \cdot \vec{k}_2)}{k_1^2 k_2^2} \Biggr] \nonumber \\
&=& c_{21} -2 c_{22} \Biggl(\fr{\vec{k}_1 \cdot \vec{k}_2}{k_1 k_2} \Biggr)^2 + \fr{1}{2} \Bigl(c_{21} -2 c_{22} \Bigr) \vec{k}_1 \cdot \vec{k}_2 \Biggl(\fr{1}{k_1^2} + \fr{1}{k_2^2} \Biggr) \label{F2s} \, , \\
\hat{\theta}^{(2)}(a,\vec{k}) &\equiv& D_{21}(a) K_{21}(\vec{k}) + D_{22}(a) K_{22}(\vec{k}) \equiv D_1 \fr{\partial D_{1}}{\partial \tau} \Biggl[ c_{\theta 21}(a) K_{21}(\vec{k}) + c_{\theta 22}(a) K_{22}(\vec{k}) \Biggr] \nonumber \\
&=& a {\cal H} \Bigl[-\fr{d D_{21}}{d a} + D_1 \fr{d D_{1}}{d a} \Bigr] K_{21}(\vec{k}) - a {\cal H} \fr{d D_{22}}{d a} K_{22}(\vec{k}) \equiv a {\cal H} D_1 \fr{d D_{1}}{d a} \theta_{2}(a,\vec{k}) \nonumber \\
&=& -a {\cal H} D_1 \fr{d D_{1}}{d a} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \Bigl[ -c_{\theta 21}(a) \alpha(\vec{k}_1, \vec{k}_2) + c_{\theta 22}(a) \beta(\vec{k}_1, \vec{k}_2) \Bigr] \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \nonumber \\
&\equiv& -a {\cal H} D_1 \fr{d D_{1}}{d a} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) G_{2}(a,\vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \label{theta23} \\
&\equiv& -a {\cal H} D_1 \fr{d D_{1}}{d a} \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) G_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) \delta_1(\vec{k}_1) \delta_1(\vec{k}_2) \nonumber \, , \\
G_{2}^{(s)}(a,\vec{k}_1,\vec{k}_2) &=& \fr{1}{2} \Biggl[ -c_{\theta 21} \Bigl( \fr{\vec{k}_{12} \cdot \vec{k}_1}{k_1^2} + \fr{\vec{k}_{12} \cdot \vec{k}_2}{k_2^2} \Bigr) + 2 c_{\theta 22} \fr{k_{12}^2 (\vec{k}_1 \cdot \vec{k}_2)}{k_1^2 k_2^2} \Biggr] \nonumber \\
&=& - c_{\theta 21} +2 c_{\theta 22} \Biggl(\fr{\vec{k}_1 \cdot \vec{k}_2}{k_1 k_2} \Biggr)^2 - \fr{1}{2} \Bigl(c_{\theta 21} -2 c_{\theta 22} \Bigr) \vec{k}_1 \cdot \vec{k}_2 \Biggl(\fr{1}{k_1^2} + \fr{1}{k_2^2} \Biggr) \label{G2s} \, .
\ea
Now one can obtain the third order solutions from the previous solutions up to the second order. One can write the third order solution
\ba \hat{\delta}^{(3)}(a,\vec{k}) &\equiv& \sum_{i=1}^{6} D_{3i}(a) K_{3i}(\vec{k}) = D_{31}(a) K_{31}(\vec{k}) + \cdots +D_{36}(a) K_{36}(\vec{k}) \label{delta3}\\ &\equiv& D_1^3(a) \Biggl[ c_{31}(a) K_{31}(\vec{k}) + \cdots + c_{36}(a) K_{36}(\vec{k}) \Biggr] \nonumber \\ &\equiv& D_1^3(a) \int d^3 k_1 d^3 k_2 d^3 k_3 \delta_{\D}(\vec{k}_{123} - \vec{k}) F_{3}(a,\vec{k}_1,\vec{k}_2,\vec{k}_3) \delta_{1}(\vec{k}_1) \delta_{1}(\vec{k}_2) \delta_{1}(\vec{k}_3) \nonumber \\
&\equiv& D_1^3(a) \int d^3 k_1 d^3 k_2 d^3 k_3 \delta_{\D}(\vec{k}_{123} - \vec{k}) F_{3}^{(s)}(a,\vec{k}_1,\vec{k}_2,\vec{k}_3) \delta_{1}(\vec{k}_1) \delta_{1}(\vec{k}_2) \delta_{1}(\vec{k}_3) \, . \nonumber \ea
If one replaces Eq.(\ref{delta3}) into Eq.(\ref{massEulerFT}), then one obtains
\ba \Biggl[ \fr{\partial^2 D_{31}}{\partial \tau^2} + {\cal H} \fr{\partial D_{31}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{31} \Biggr] K_{31}(\vec{k}) &=& \Biggl[\Bigl(\fr{\partial^2 D_{1}}{\partial \tau^2} + {\cal H} \fr{\partial D_{1}}{\partial \tau} \Bigr) D_{21} + \fr{\partial D_1}{\partial \tau} \fr{\partial D_{21}}{\partial \tau} \Biggr] \label{delta31} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) K_{21} (\vec{k}_2) \Biggr] \nonumber \, , \\
&\equiv& \Biggl[\fr{3}{2} \Omega_{m} {\cal H}^2 D_1 D_{21} + \fr{\partial D_1}{\partial \tau} \fr{\partial D_{21}}{\partial \tau} \Biggr] K_{31}(\vec{k}) \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{32}}{\partial \tau^2} + {\cal H} \fr{\partial D_{32}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{32} \Biggr] K_{32}(\vec{k}) &=& \Biggl[\Bigl(\fr{\partial^2 D_{1}}{\partial \tau^2} + {\cal H} \fr{\partial D_{1}}{\partial \tau} \Bigr) D_{22} + \fr{\partial D_1}{\partial \tau} \fr{\partial D_{22}}{\partial \tau} \Biggr] \label{delta32} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) K_{22} (\vec{k}_2) \Biggr] \nonumber \, , \\
&\equiv& \Biggl[\fr{3}{2} \Omega_{m} {\cal H}^2 D_1 D_{22} + \fr{\partial D_1}{\partial \tau} \fr{\partial D_{22}}{\partial \tau} \Biggr] K_{32}(\vec{k}) \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{33}}{\partial \tau^2} + {\cal H} \fr{\partial D_{33}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{33} \Biggr] K_{33}(\vec{k}) &=& \Biggl[-\Bigl(\fr{\partial^2 D_{21}}{\partial \tau^2} + {\cal H} \fr{\partial D_{21}}{\partial \tau} \Bigr) D_{1} + \Bigl(\fr{\partial^2 D_{1}}{\partial \tau^2} + {\cal H} \fr{\partial D_{1}}{\partial \tau} \Bigr) D_{1}^2 \nonumber \\ &+& 2 D_1 \Bigl( \fr{\partial D_1}{\partial \tau} \Bigr)^2 - \fr{\partial D_1}{\partial \tau} \fr{\partial D_{21}}{\partial \tau} \Biggr] \label{delta33} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) K_{21} (\vec{k}_1) \delta_1 (\vec{k}_2) \Biggr] \nonumber \, , \\
&\equiv& \Biggl[ -\fr{3}{2} \Omega_{m} {\cal H}^2 D_1 D_{21} - \fr{\partial D_1}{\partial \tau} \fr{\partial D_{21}}{\partial \tau} + D_1 \Bigl( \fr{\partial D_1}{\partial \tau} \Bigr)^2 \Biggr] K_{33}(\vec{k}) \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{34}}{\partial \tau^2} + {\cal H} \fr{\partial D_{34}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{34} \Biggr] K_{34}(\vec{k}) &=& \Biggl[-\Bigl(\fr{\partial^2 D_{22}}{\partial \tau^2} + {\cal H} \fr{\partial D_{22}}{\partial \tau} \Bigr) D_{1} - \fr{\partial D_1}{\partial \tau} \fr{\partial D_{22}}{\partial \tau} \Biggr] \label{delta34} \\
&\times& \Biggl[ - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) B_2 (\vec{k}_1) \delta_1 (\vec{k}_2) \Biggr] \nonumber \, , \\
&\equiv& \Biggl[-\fr{3}{2} \Omega_{m} {\cal H}^2 D_1 D_{22} - \fr{\partial D_1}{\partial \tau} \fr{\partial D_{22}}{\partial \tau}+\fr{1}{2}D_1 \Bigl( \fr{\partial D_1}{\partial \tau} \Bigr)^2 \Biggr] K_{34}(\vec{k}) \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{35}}{\partial \tau^2} + {\cal H} \fr{\partial D_{35}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{35} \Biggr] K_{35}(\vec{k}) &=& \fr{1}{2} \Biggl[\fr{\partial D_1}{\partial \tau} \fr{\partial D_{21}}{\partial \tau} -D_1\Bigl(\fr{\partial D_1}{\partial \tau} \Bigr)^2 \Biggr] \label{delta35} \\ &\times& \Biggl[-\int d^3 k_{1} d^3 k_2 \delta_{\D}(\vec{k}_{12}-\vec{k}) \beta(\vec{k}_1, \vec{k}_2) \Bigl(\theta_1 (\vec{k}_1) K_{21} (\vec{k}_2) + K_{21} (\vec{k}_1) \theta_1 (\vec{k}_2) \Bigr) \Biggr] \nonumber \, , \\
\Biggl[ \fr{\partial^2 D_{36}}{\partial \tau^2} + {\cal H} \fr{\partial D_{36}}{\partial \tau} - \fr{3}{2} \Omega_{m} {\cal H}^2 D_{36} \Biggr] K_{36}(\vec{k}) &=& \fr{1}{2} \fr{\partial D_1}{\partial \tau} \fr{\partial D_{22}}{\partial \tau} \label{delta36} \\
&\times& \Biggl[-\int d^3 k_{1} d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \beta(\vec{k}_1, \vec{k}_2) \Bigl(\theta_1 (\vec{k}_1) K_{22} (\vec{k}_2) + K_{22} (\vec{k}_1) \theta_1 (\vec{k}_2) \Bigr) \Biggr] \nonumber \, , \ea
One can rewrite the temporal parts of the above Eqs. (\ref{delta31})-(\ref{delta36}) with the proper initial conditions obtained from the EdS solutions at the early epoch to get the fastest growing mode solutions,
\ba \fr{d^2 D_{31}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{31}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{31} &=& \fr{3}{2a^2} \Omega_{m} D_1 D_{21} + \fr{d D_1}{d a} \fr{d D_{21}}{d a} \label{delta312} \\
&{\rm with}& D_{31}(a_i) = 0, \hspace{0.1in} \fr{d D_{31}}{da} \Bigl|_{a_i} = \fr{20}{441} a_i^2 \nonumber \, , \\
\fr{d^2 D_{32}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{32}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{32} &=& \fr{3}{2a^2} \Omega_{m} D_1 D_{22} + \fr{d D_1}{d a} \fr{d D_{22}}{d a} \label{delta322} \\
&{\rm with}& D_{32}(a_i) = 0, \hspace{0.1in} \fr{d D_{32}}{da} \Bigl|_{a_i} = -\fr{4}{441} a_i^2 \nonumber \, , \\
\fr{d^2 D_{33}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{33}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{33} &=& -\fr{3}{2a^2} \Omega_{m} D_1 D_{21} - \fr{d D_1}{d a} \fr{d D_{21}}{d a} + D_1 \Bigl( \fr{d D_1}{d a} \Bigr)^2 \label{delta332} \\
&{\rm with}& D_{33}(a_i) = 0, \hspace{0.1in} \fr{d D_{33}}{da} \Bigl|_{a_i} = \fr{78}{441} a_i^2 \nonumber \, , \\
\fr{d^2 D_{34}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{34}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{34} &=& -\fr{3}{2a^2} \Omega_{m} D_1 D_{22} -\fr{d D_1}{d a} \fr{d D_{22}}{d a} +\fr{1}{2} D_1\Bigl( \fr{d D_1}{d a} \Bigr)^2 \label{delta342} \\
&{\rm with}& D_{34}(a_i) = 0, \hspace{0.1in} \fr{d D_{34}}{da} \Bigl|_{a_i} = \fr{53}{441} a_i^2 \nonumber \, , \\
\fr{d^2 D_{35}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{35}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{35} &=& \fr{1}{2} \Biggl[ -D_1 \Bigl(\fr{d D_1}{d a} \Bigr)^2 + \fr{d D_1}{d a} \fr{d D_{21}}{d a}\Biggr] \label{delta352} \\
&{\rm with}& D_{35}(a_i) = 0, \hspace{0.1in} \fr{d D_{35}}{da} \Bigl|_{a_i} = -\fr{24}{441} a_i^2 \nonumber \, , \\
\fr{d^2 D_{36}}{d a^2} + \fr{3}{2a} \Bigl(1 - w \Omega_{de} \Bigr) \fr{d D_{36}}{d a} - \fr{3}{2a^2} \Omega_{m} D_{36} &=& \fr{1}{2} \fr{d D_1}{d a} \fr{d D_{22}}{d a} \label{delta362} \\
&{\rm with}& D_{36}(a_i) = 0, \hspace{0.1in} \fr{d D_{36}}{da} \Bigl|_{a_i} = -\fr{5}{441} a_i^2 \nonumber \, , \ea
where we use the fastest growing mode solutions for the EdS universe
\ba D_{31}^{(\EdS)} &=& \fr{5}{882} \Bigl(49 a^3 - 90 a_i a^2 + 41 a_i^2 a \Bigr), \,\,\,\,
D_{32}^{(\EdS)} = \fr{-1}{882} \Bigl(49 a^3 - 90 a_i a^2 + 41 a_i^2 a \Bigr) \, , \nonumber \\
D_{33}^{(\EdS)} &=& \fr{-3}{882} \Bigl(49 a^3 - 45 a_i a^2 - 4 a_i^2 a \Bigr), \,\,\,\,
D_{34}^{(\EdS)} = \fr{2}{882} \Bigl(49 a^3 - 90 a_i a^2 + 41 a_i^2 a \Bigr) \, , \label{D316EdS} \\
D_{35}^{(\EdS)} &=& \fr{3}{882} \Bigl(7 a^3 - 30 a_i a^2 + 23 a_i^2 a \Bigr), \,\,\,\,
D_{36}^{(\EdS)} = \fr{-2}{882} \Bigl(7 a^3 - 9 a_i a^2 + 2 a_i^2 a \Bigr) \, . \nonumber \ea
In the above equation (\ref{D316EdS}), we use the initial Gaussinity condition of $\delta^{(3)}$ ({\it i.e.} $D_{3i}(a_i) = 0$) to obtain the coefficients for the last terms of $D_{3i}$.
One can find the third order kernels ($F_{3i}(\kk)$) from the above Eqs. (\ref{delta31})-(\ref{delta36}). For example, one obtain $F_{31}^{(s)}$ as
\ba K_{31}(\vec{k}) &=& - \int d^3 k_{1} \int d^3 k_2 \delta_{\D}(\vec{k}_{12} - \vec{k}) \alpha(\vec{k}_1, \vec{k}_2) \theta_1 (\vec{k}_1) K_{21} (\vec{k}_2) \nonumber \\
&=& \int d^3 q_1 d^3 q_2 d^3 q_{3} \delta_{\D}(\vec{q}_{1} +\vec{q}_2+\vec{q}_3 - \vec{k}) \alpha(\vec{q}_3, \vec{k}-\vec{q}_3) \alpha(\vec{q}_1, \vec{q}_2) \delta_1(\vec{q}_1) \delta_1(\vec{q}_2) \delta_1 (\vec{q}_3) \, , \nonumber \\
F_{31}(a,\vec{q}_1,\vec{q}_2,\vec{q}_3) &\equiv& c_{31}(a) \alpha(\vec{q}_3, \vec{k}-\vec{q}_3) \alpha(\vec{q}_1, \vec{q}_2), \,\,\,
F_{31}^{(s)}(a,\vec{q}_1, \vec{q}_2, \vec{q}_3) = \fr{c_{31}}{3!} \Biggl[F_{31}(a, \vec{q}_1, \vec{q}_2, \vec{q}_3) + {\rm perm} \Biggr] \nonumber \, , \\
F_{31}^{(s)}(a,\vec{q},-\vec{q},\vec{k}) &=& \fr{c_{31}}{3!} \Biggl[-2 x^2 \Bigl(\fr{1+r^2}{r^2} \Bigr) \Biggr] \label{F31s} \, . \ea
One can repeat the above process to obtain
\ba F_{32}(a, \vec{q}_1, \vec{q}_2, \vec{q}_3) &=& -c_{32} \alpha(\vec{q}_3, \vec{k}-\vec{q}_3) \beta(\vec{q}_1, \vec{q}_2) \, , \nonumber \\ F_{32}^{(s)}(a, \vec{q}, -\vec{q}, \vec{k}) &=& \fr{c_{32}}{3!} 4x^2 \Bigl(\fr{1+r^2}{r^2} \Bigr) \label{F32s} \, , \\
F_{33}(a,\vec{q}_1, \vec{q}_2, \vec{q}_3) &=& -c_{33} \alpha(\vec{k}-\vec{q}_3,\vec{q}_3) \alpha(\vec{q}_1, \vec{q}_2) \, ,\nonumber \\
F_{33}^{(s)}(a,\vec{q}, -\vec{q}, \vec{k}) &=& \fr{c_{33}}{3!} \fr{-4(1+r^2)+2(1+4r^2-r^4)x^2}{(1+r^2+2rx)(1+r^2-2rx)} \label{F33s} \, , \\
F_{34}(a,\vec{q}_1, \vec{q}_2, \vec{q}_3) &=& c_{34} \alpha(\vec{k}-\vec{q}_3,\vec{q}_3) \beta(\vec{q}_1, \vec{q}_2) \, , \nonumber \\
F_{34}^{(s)}(a,\vec{q}, -\vec{q}, \vec{k}) &=& \fr{c_{34}}{3!} 4x^2 \label{F34s} \, , \\
F_{35}(a,\vec{q}_1, \vec{q}_2, \vec{q}_3) &=& 2c_{35} \beta(\vec{k}-\vec{q}_3,\vec{q}_3) \alpha(\vec{q}_1, \vec{q}_2) \, , \nonumber \\
F_{35}^{(s)}(a, \vec{q}, -\vec{q}, \vec{k}) &=& \fr{c_{35}}{3!} \fr{-8r^2(1+r^2) + 4(-1+4r^2+r^4)x^2}{r^2(1+r^2+2rx)(1+r^2-2rx)} \label{F35s} \, , \\
F_{36}(a,\vec{q}_1, \vec{q}_2, \vec{q}_3) &=& -2c_{36} \beta(\vec{k}-\vec{q}_3,\vec{q}_3) \beta(\vec{q}_1, \vec{q}_2) \, , \nonumber \\
F_{36}^{(s)}(a,\vec{q}, -\vec{q}, \vec{k}) &=& \fr{c_{36}}{3!} \fr{ 8x^2}{r^2} \label{F36s} \ea
One can obtain $c_{3i}(a)$ numerically at any epoch by solving the above Eqs.(\ref{delta312})-(\ref{delta362}).
In EdS case, $c_{3i}$ have been known as
\be c_{31} = \fr{5}{18} \,\, , c_{32} = -\fr{1}{18} \,\, , c_{33} = -\fr{1}{6} \,\, , c_{34} = \fr{1}{9} \,\, , c_{35} = \fr{1}{42} \,\, , c_{36} = -\fr{1}{63} \label{c3iEdS} \, . \ee
However, the above values are not exact because they are not the coefficients for the fastest growing mode solutions as shown in Eq. (\ref{D316EdS}). The above values given in Eq.(\ref{c3iEdS}) used in the kernels in the reference \cite{08061437}. For the second order, this is a good approximation but not for the third order.
Now one can explicitly write $\hat{\delta}^{(3)}(a,\vec{k})$ as
\be \hat{\delta}^{(3)}(a,\vec{k}) \equiv D_{1}^3(a) \sum_{i=1}^{6} c_{3i}(a) K_{3i}(\vec{k})= D_{1}^3(a) \Biggl[ c_{31}(a) K_{31} (\vec{k}) + \cdots + c_{36}(a) K_{36}(\vec{k}) \Biggr] \label{delta32} \ee
Thus, one can calculate $P_{22}(a,k)$ and $P_{13}(a,k)$ at any epoch.
\ba P_{22}(a,k) &=& D_{1}^4(a) 2 \int d^3 q \Bigl[ F_{2}^{(s)}(\vec{k}-\vec{q},\vec{q}) \Bigr]^2 P_{11}(k-q) P_{11}(q) \nonumber \\ &=& D_{1}^4(a) 2 \int (2 \pi) q^2 d q \int_{-1}^{1} dx \Bigl[ F_{2}^{(s)}(\vec{k}-\vec{q},\vec{q}) \Bigr]^2 P_{11}(kr) P_{11}(k\sqrt{1+r^2-2rx}) \nonumber \\ &=& \fr{(2\pi)^{-2} k^3}{2} \int_{0}^{\infty} dr P_{11}(kr) \int_{-1}^{1} dx P_{11}(k\sqrt{1+r^2-2rx}) \nonumber \\ &\times& \Biggl[\fr{(c_{21} + 2c_{22})r + (c_{21} - 2c_{22})x - 2c_{21} rx^2 }{(1+r^2-2rx)} \Biggr]^2 \label{P22new2} \, , \\
P_{22}^{({\rm EdS})}(a,k) &=& D_{1}^4(a) \fr{(2\pi)^{-2} k^3}{2} \int_{0}^{\infty} dr P_{11}(kr) \int_{-1}^{1} dx P_{11}(k\sqrt{1+r^2-2rx}) \Biggl[\fr{3r + x - 10 rx^2 }{7(1+r^2-2rx)} \Biggr]^2 \label{P22new2Eds} \, , \\
P_{13}(a,k) &=& 6 D_{1}^4(a) P_{11}(k) \int d^3 q P_{11}(q) \Bigl[ F_{3}^{(s)}(a,\vec{q},-\vec{q},\vec{k}) \Bigr] \label{P13new2} \\ &=& D_{1}^4 (2\pi)^{-2} k^3 P_{11}(k) \int_{0}^{\infty} dr P_{11}(kr) \Biggl[ 2c_{35} r^{-2} -\fr{1}{3} \Bigl( 4 c_{31} -8 c_{32} +3c_{33} +24c_{35} - 16 c_{36} \Bigr) \nonumber \\ &-& \fr{1}{3}\Bigl(4 c_{31} -8c_{32} +12c_{33}-8c_{34}+6c_{35} \Bigr)r^2 + c_{33} r^4 + \Bigl(\fr{r^2-1}{r} \Bigr)^3 \ln \Bigl|\fr{1+r}{1-r} \Bigr| \Bigl(c_{35} - \fr{1}{2}c_{33}r^2 \Bigr) \Biggr] \nonumber \, , \\
P_{13}^{({\rm EdS})}(a,k) &=& D_{1}^4(a) (2\pi)^{-2} k^3 P_{11}(k) \int_{0}^{\infty} dr P_{11}(kr) \Biggl[ \fr{1}{21} r^{-2} -\fr{79}{126} + \fr{25}{63}r^2 -\fr{1}{6}r^4 \nonumber \\ &+& \Bigl(\fr{r^2-1}{r} \Bigr)^3 \ln \Bigl|\fr{1+r}{1-r} \Bigr| \Bigl( \fr{1}{42}+\fr{1}{12}r^2 \Bigr) \Biggr] \label{P13new2EdS} \, , \ea
where we use
\be |\vec{q}| = r |\vec{k}|, \,\, \vec{q} \cdot \vec{k} = x |\vec{q}| |\vec{k}|, \,\, k_1^2 = |\vec{q}|^2 = r^2 k^2, \,\, k_2^2 = |\vec{k}-\vec{q}|^2 = k^2 + q^2 -2 \vec{k} \cdot \vec{q} = k^2 ( 1+ r^2 -2rx), \,\, \vec{k}_1 \cdot \vec{k}_2 = k^2 r (x-r) \label{qk} \ee
We also check the dependence of $c_{3i}$ on $a_i$. When we compare $a_i=\fr{1}{50}$ to $a_i = \fr{1}{1000}$, there are sub percent level differences.
|
\section{Introduction\label{intro}}
The Reissner-Nordstr\"om spacetime $({\mathcal M},g)$ is a fundamental 2-parameter family of solutions to the Einstein field equations coupled to electromagnetism, cf.~ Figure \ref{RN} for the conformal representation of the subextremal case, \mbox{$M>|e|\neq 0$}, with $e$ the charge and $M$ the mass of the black hole.
The problem of analysing the scalar wave equation
\begin{eqnarray}
\label{wave}
\Box_g \phi=0
\end{eqnarray}
on a Reissner-Nordstr\"om background is intimately related to the stability properties of the spacetime itself and to the celebrated Strong Cosmic Censorship Conjecture.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{RN.jpg}
\caption[]{Maximal development of Cauchy hypersurface $\Sigma$ in Reissner-Nordstr\"om spacetime $({\mathcal M},g)$.}
\label{RN}\end{figure}}
The analysis of \eqref{wave} in the exterior region $J^-({\mathcal I}^+)$ has been accomplished already, cf.~ \cite{blue} and \cite{m_lec} for an overview and references therein for more details, as well as Section \ref{horizon_estimates}. The purpose of the present work is to extend the investigation to the interior of the black hole, up to and including the Cauchy horizon ${\mathcal C}{\mathcal H}^+$.
\subsection{Main result}
The main result of this paper can be stated as follows.
\begin{thm}
\label{main}
On subextremal Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, let $\phi$ be
a solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on a two-ended asymptotically flat Cauchy surface $\Sigma$. Then
\begin{eqnarray}
\label{maineq}
|\phi|\leq C
\end{eqnarray}
globally in the black hole interior, in particular up to and including the Cauchy horizon
${\mathcal C}{\mathcal H}^+$, to which $\phi$ extends in fact continuously.
\end{thm}
The constant $C$ is explicitly computable in terms of parameters $e$ and $M$ and a suitable norm on initial data.
The above theorem will follow, after commuting \eqref{wave} with angular momentum operators and applying Sobolev embedding, from the following theorem, expressing weighted energy boundedness.
\begin{thm}
\label{energythm}
On subextremal Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, let $\phi$ be
a solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on a two-ended asymptotically flat Cauchy surface $\Sigma$. Then
\begin{eqnarray}
\label{energy1}
\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ v^p (\partial_v \phi)^2(u, v, \theta, \varphi) + |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& E,\quad \mbox{for $v_{fix} \geq 1$, $u > -\infty$}\\
\label{energy1u}
\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{u_{fix}}\left[ u^p (\partial_u \phi)^2 (u, v, \theta, \varphi)+ |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u, v, \theta, \varphi) \right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& E,\quad \mbox{for $u_{fix} \geq 1$, $v > -\infty$}
\end{eqnarray}
where $p>1$ is an appropriately chosen constant, and $(u, v)$ denote Eddington-Finkelstein coordinates in the black hole interior, where by ${\mathrm{d}} \sigma_{\mathbb S^2}$ we denote the volume element of the unit two-sphere and \mbox{$|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2=\frac{1}{r^2}\left[(\partial_{\theta} \phi)^2+ \frac{1}{\sin^2 {\theta}}(\partial_{\varphi} \phi)^2\right]$}.
\end{thm}
\subsection{A first look at the analysis}
\label{firstlook}
The proof of Theorem \ref{main} and \ref{energythm} involves first considering a characteristic rectangle $\Xi$ within the black hole interior, whose future {\it right} boundary coincides with the Cauchy horizon ${\mathcal C}{\mathcal H}^+$ in the vicinity of $i^+$, cf.~ Figure \ref{alle}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{alle.jpg}
\caption[]{Penrose diagram depicting all regions considered in the entire proof.}
\label{alle}\end{figure}}
Establishing boundedness of weighted energy norms in $\Xi$ is the crux of the entire proof.
Once that is done, analogous results hold for a characteristic rectangle $\tilde{\Xi}$ to the {\it left} depicted in Figure \ref{alle}. Hereafter, boundedness of the energy is easily propagated to regions ${\mathcal R}_V$, $\tilde{{\mathcal R}}_V$ and ${\mathcal R}_{VI}$ as depicted, giving Theorem \ref{energythm}. Commutation by angular momentum operators and application of Sobolev embedding then yields Theorem \ref{main}.
Let us return to the discussion of $\Xi$ since that is the most involved part of the proof.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{bnr_region_am_aktuellsten.jpg}
\caption[]{Conformal representation of a characteristic regtangle $\Xi$ with redshift ${\mathcal R}$, noshift ${\mathcal N}$ and blueshift regions ${\mathcal B}$.}
\label{bnr1}\end{figure}}
In order to prove Theorem \ref{energythm} (and hence Theorem \ref{main}) restricted to $\Xi$ we will begin with an upper decay bound for $|\phi|$ and its derivatives on the event horizon ${\mathcal H}^+$, which can be deduced by putting together preceding work of Blue-Soffer, cf.~ \cite{blue}, Dafermos-Rodnianski, cf.~ \cite{m_price} and Schlue, cf.~ \cite{volker}. The precise result from previous work that we shall need will be stated in Section \ref{horizon_estimates}.
In $\Xi$ the proof involves distinguishing redshift ${\mathcal R}$, noshift ${\mathcal N}$ and blueshift ${\mathcal B}$ regions, as shown in Figure \ref{bnr1}.
Some of these regions have appeared in previous analysis of the wave equation, especially \mbox{${\mathcal R}=\left\{r_{red}\leq r \leq r_+\right\}$}. Region \mbox{${\mathcal N}=\left\{r_{blue}\leq r\leq r_{red}\right\}$} and region \mbox{${\mathcal B}=\left\{r_{-}\leq r\leq r_{blue}\right\}$} were studied in \cite{m_interior} in the {\it spherically symmetric self-gravitating} case, but using techniques which are very special to $1+1$ dimensional hyperbolic equations.\footnote{Let us note that the result of Theorem \ref{main} for \emph{spherically symmetric} solutions $\phi$ can be obtained by specializing \cite{m_interior, m_price, m_bh} to the uncoupled case. Restricted results for fixed spherical harmonics can be in principle also inferred from \cite{mc_b}.}
We will discuss this separation into ${\mathcal R}$, ${\mathcal N}$ and ${\mathcal B}$ regions further in Section \ref{outline}.
One of the main analytic novelties of this paper is the introduction of a new vector field energy identity
constructed for analyses in region ${\mathcal B}$.
In particular, the weighted vector field is given in Eddington-Finkelstein coordinates $(u,v)$ by
\begin{eqnarray*}
S=|u|^p\partial_u+v^p\partial_v,
\end{eqnarray*}
for $p>1$ as appearing in Theorem \ref{energythm}. This vector field associated to region ${\mathcal B}$ will allow to prove uniform boundedness despite the blueshift instability.
\subsection{Outline of the paper}
The paper is organized as follows.
In the remaining Section \ref{motivation} of the introduction we will elaborate on Strong Cosmic Censorship and its relation to this work.
In Section \ref{preliminaries} we introduce the basic tools needed to derive energy estimates from the energy momentum tensor associated to \eqref{wave} and an appropriate vector field. A review of the Reissner-Nordstr\"om solution and the coordinates used in this paper will be given. Moreover, we will discuss further features of Reissner-Nordstr\"om geometry.
In Section \ref{horizon_estimates} we give a brief review of estimates obtained along ${\mathcal H}^+$ from previous work, \cite{blue}, \cite{m_price} and \cite{volker}, for $\phi$ arising from sufficiently regular initial data on a Cauchy hypersurface. This is stated as Theorem \ref{anfang}.
Section \ref{outline} states our main result specialized to the rectangle $\Xi$ (see Theorem \ref{dashier}) and gives an outline of its proof. The investigation is divided into considerations within the redshift ${\mathcal R}$, noshift ${\mathcal N}$, and blueshift ${\mathcal B}$ regions.
The decay bound for the energy flux of $\phi$ given on the event horizon ${\mathcal H}^+$, cf.~ Theorem \ref{anfang}, will be propagated through the redshift region ${\mathcal R}$ up to the hypersurface $r=r_{red}$ in Section \ref{first_section}.
Thereafter, in Section \ref{zweite_region} we propagate the decay bound further into the black hole interior through the noshift region ${\mathcal N}$ up to the hypersurface $r=r_{blue}$.
In Section \ref{blueshift1} a decay bound for the energy flux of $\phi$ is proven on a well chosen hypersurface $\gamma$ that separates the blueshift region into a region in the past of $\gamma$, $J^-(\gamma) \cap {\mathcal B}$, and a region to the future of $\gamma$, $J^+(\gamma) \cap {\mathcal B}$.
In Section \ref{innerhorizon} we will derive pointwise estimates on $\Omega^2$ to the future of $\gamma$ (in particular implying finiteness of the spacetime volume, $\operatorname{Vol}(J^+(\gamma))<C$). This will allow us to propagate our estimates into $J^+(\gamma) \cap {\mathcal B}$ up to ${\mathcal C}{\mathcal H}^+$, yielding finally Corollary \ref{endeacht}.
Section \ref{nineten} reveals how commutation with angular momentum operators and applying Sobolev embedding will return us pointwise boundedness for $|\phi|$.
The necessary higher order boundedness statement is given in Theorem \ref{energythm2}. This completes the proof of Theorem \ref{dashier}.
We now must extend our result to the full interior region.
In Section \ref{leftinterior} we will state the analog of Theorem \ref{dashier} restricted to the rectangle $\tilde{\Xi}$ to the {\it left}. In Section \ref{region5_proof} and Section \ref{region_tilde5_proof} we propagate the energy estimates further along ${\mathcal C}{\mathcal H}^+$ in the depicted regions ${\mathcal R}_V$ and $\tilde{{\mathcal R}}_V$. Eventually, in Section \ref{bifurcate} we propagate the estimate to the region ${\mathcal R}_{VI}$ up to the bifurcation two-sphere, and thus obtain a bound for the energy flux globally in the black hole interior (see Corollary \ref{letzteprop}) completing the proof of Theorem \ref{energythm}.
In Section \ref{global} we prove Theorem \ref{energythm3}, stating boundedness of the weighted higher order energies. Using the conclusion of this theorem, we apply again Sobolev embedding as before (using also the result of Section \ref{uni_bounded}) and thus obtain the boundedness statement of Theorem \ref{main}.
Finally, in Section \ref{continuity} we show continuous extendibility of $\phi$ to the Cauchy horizon.
An Outlook of open problems will be given in Section \ref{outlook}.
We first state an analogous result to our Theorem \ref{main} for general subextremal Kerr black holes (to appear as Theorem 1.1 of \cite{ich}). The conjectured blow up of the transverse derivatives\footnote{Note in contrast that the {\it tangential} derivatives of $\phi$ can be shown to be uniformly bounded up to ${\mathcal C}{\mathcal H}^+$ (away from the bifurcation sphere) from the energy estimates proven in this paper together with commuting with angular momentum operators, cf.~ Theorem \ref{energythm}.} along the Cauchy horizon for generic solutions of \eqref{wave} will also be discussed, as well as the peculiar extremal case.
Finally, we will discuss what our results suggest about the nonlinear dynamics of the Einstein equations themselves.
\subsection{Motivation and Strong Cosmic Censorship}
\label{motivation}
Our motivation for proving Theorem \ref{main} is the Strong Cosmic Censorship Conjecture.
The mathematical formulation of this conjecture, here applied to electrovacuum, is given in \cite{christo_sing} by Christodoulou as
\begin{eqnarray}
\label{sccc}
\begin{minipage}[c]{0.65\textwidth}
``\underline{Generic} asymptotically flat initial data for Einstein-Maxwell spacetimes
have a maximal future development which is
inextendible as a suitably regular Lorentzian manifold.''
\end{minipage}
\end{eqnarray}
Reissner-Nordstr\"om spacetime serves as a counterexample to the inextendibility statement since it is (in fact smoothly) extendable beyond the Cauchy horizon ${\mathcal C}{\mathcal H}^+$.\footnote{Outside the future maximal domain of dependence ${\mathcal D}^+({\mathcal M})$ in the future of the Cauchy horizon $J^+({\mathcal C}{\mathcal H}^+)$ the spacetime shows the peculiar feature that uniqueness of the solutions of the initial value problem is lost {\it without loss of regularity}.
It is precisely the undesirability of this feature that motivates the conjecture.}
Thus, for the above conjecture to be true, this property of Reissner-Nordstr\"om must in particular be unstable.
Originally it was suggested by Penrose and Simpson that small perturbations of Reissner-Nordstr\"om would lead to a spacetime whose boundary would be a spacelike singularity as in Schwarzschild and such that the spacetime would be inextendable as a $C^0$ metric, cf.~ \cite{simpson}.
On the other hand, a heuristic study of a spherically symmetric but fully nonlinear toy model by Israel and Poisson, cf.~ \cite{poisson}, led to an alternative scenario, which suggested
that spacetimes resulting from small perturbations would exist up to a Cauchy horizon, which however would be singular in a weaker sense, see also \cite{amos} by Ori. Considering the {\it spherically symmetric} Einstein-Maxwell-scalar field equations as a toy model, Dafermos proved that the solution indeed exists up to a Cauchy horizon and moreover is extendible as a $C^0$ metric but generically fails to be extendible as a $C^1$ metric beyond ${\mathcal C}{\mathcal H}^+$, cf.~ \cite{m_stab, m_interior}. For more recent extensions see \cite{kommemi, joao1, joao2, joao3}.
In this work, as a first attempt towards investigation of the stability of the Cauchy horizon under perturbations {\it without symmetry}, we employ \eqref{wave} on a fixed Reissner-Nordstr\"om background $({\mathcal M},g)$ as a toy model for the full nonlinear Einstein field equations, cf.~ \eqref{EF}. The result of uniform pointwise boundedness of $\phi$ and continuous extension to ${\mathcal C}{\mathcal H}^+$ is concordant with the work of Dafermos \cite{m_stab}. This suggests that the non-spherically symmetric perturbations of the astrophysically more realistic Kerr spacetime may indeed exist up to ${\mathcal C}{\mathcal H}^+$.
See Section \ref{EFvacuum}.
\section{Preliminaries}
\label{preliminaries}
\subsection{Energy currents and vector fields}
\label{en_cur}
The essential tool used throughout this work is the so called vector field method.
Let $({\mathcal M}, g)$ be a Lorentzian manifold. Let $\phi$ be a solution to the wave equation $\Box_g \phi=0$.
A symmetric stress-energy tensor can be identified from variation of the massless scalar field action by
\begin{eqnarray*}
\label{energymomentum}
T_{\mu\nu}(\phi)=\partial_\mu\phi\partial_\nu\phi-\frac12g_{\mu\nu}g^{\alpha\beta}\partial_\alpha\phi
\partial_\beta\phi,
\end{eqnarray*}
and this satisfies the energy-momentum conservation law
\begin{equation}
\label{divfree}
\nabla^\mu T_{\mu\nu}=0.
\end{equation}
By contracting the energy-momentum tensor with a vector field $V$, we define the current
\begin{equation}
\label{J}
J_\mu^V(\phi)\doteq T_{\mu\nu}(\phi) V^\nu.
\end{equation}
In this context we call $V$ a multiplier vector field. If the vector field $V$ is timelike, then the one-form $J_\mu^V$ can be interpreted as the energy-momentum density.
When we integrate $J_\mu^V$ contracted with the normal vector field over an associated hypersurface we will often refer to the integral as energy flux.
Note that $J^V_\mu (\phi)n^\mu_{\Sigma} \ge 0$ if $V$ is future directed timelike and $\Sigma$ spacelike, where $n^\mu_{\Sigma}$ is the future directed normal vector on the hypersurface $\Sigma$.
Since we will frequently use versions of the divergence theorem, we are interested in the divergence of the current \eqref{J}.
Defining
\begin{equation}
\label{K}
K^V(\phi)\doteq T(\phi)(\nabla V)=(\pi^V)^{\mu\nu}T_{\mu\nu}(\phi),
\end{equation}
by \eqref{energymomentum} it follows that
\begin{eqnarray}
\nabla^{\mu}J^V_{\mu}(\phi)=K^V(\phi).
\end{eqnarray}
Further, $(\pi^V)^{\mu\nu} \doteq \frac{1}{2} ({\mathcal L}_V g)^{\mu\nu}$ is the so called deformation tensor
of $V$.
Therefore, $\nabla^\mu J^V_\mu(\phi) =0$ if $V$ is Killing.
For a Killing vector field $W$ we have in addition the commutation relation $[\Box_g, W]=0$. In that context $W$ is called a commutation vector field.
In particular, we note already that
in Reissner-Nordstr\"om spacetime we have $\Box_g T \phi=0$ and $\Box_g \leo_i \phi=0$, where $T$ and $\leo_i$ with $i=1,2,3$ are Killing vector fields that will be defined in Section \ref{rtcoords} and \ref{angular}, respectively.\\
For a more detailed discussion see \cite{m_lec} by Dafermos and Rodnianski, \cite{sergiu} by Klainerman and \cite{christo_action} by Christodoulou.
\subsection{The Reissner-Nordstr\"om solution}
\label{RNsection}
In the following we will briefly recall the Reissner-Nordstr\"om solution\footnote{The reader unfamiliar with this solution may for example consult \cite{haw_ellis} for a more detailed review.} which is a family of solutions to the
Einstein-Maxwell field equations
\begin{eqnarray}
\label{EF}
R_{\mu \nu}-\frac12 g_{\mu \nu}R=2 T^{EM}_{\mu \nu},
\end{eqnarray}
with $R_{\mu \nu}$ the Ricci tensor, $R$ the Ricci scalar and the units chosen such that $\frac{8 \pi G}{c^4}=2$. The Maxwell equations are given by
\begin{eqnarray}
\label{maxwell}
\nabla^{\alpha} F_{\alpha \beta}=0, \qquad \nabla_{[\lambda} F_{\alpha \beta]}=0,
\end{eqnarray}
and the energy-momentum tensor by
\begin{eqnarray}
\label{tee1}
T^{EM}_{\mu \nu}&=& F^\alpha_{\mu}F_{\alpha \nu}-\frac{1}{4} g_{\mu \nu}F^{\alpha \beta}F_{\alpha \beta}
\end{eqnarray}
The system \eqref{EF}-\eqref{tee1} describes the interaction of a gravitational field with a source free electromagnetic field.
The Reissner-Nordstr\"om solution represents a charged black hole as an isolated system in an asymptotically Minkowski spacetime.
The causal
structure is similar to the structure of the astrophysically more realistic axisymmetric Kerr
black holes.
Since spherical symmetry can often simplify first investigations,
Reissner-Nordstr\"om spacetime is a popular proxy for
Kerr.
\subsubsection{The metric and ambient differential structure}
To set the semantic convention, whenever we refer to the Reissner-Nordstr\"om solution $({\mathcal M},g)$ we mean the maximal domain of dependence \mbox{${\mathcal D} (\Sigma)={\mathcal M}$} of complete two-ended asymptotically flat data $\Sigma$.
The manifold ${\mathcal M}$ can be expressed by \mbox{${\mathcal M}={\mathcal Q}\times {\mathbb S}^2$}, and \mbox{${\mathcal Q}=(-1, 1) \times (-1,1)$} with coordinates $U, V \in (-1,1)$ and thus
\begin{eqnarray}
{\mathcal M}=(-1, 1) \times (-1,1) \times {\mathbb S}^2.
\end{eqnarray}
The metric in global double null coordinates then takes the form
\begin{eqnarray}
\label{metric}
g=-\Upomega^2(U,V){\mathrm{d}} U {\mathrm{d}} V+{\mathpzc{r}}^2(U,V)\left[{\mathrm{d}} \theta^2+\sin^2\theta {\mathrm{d}} \varphi^2\right],
\end{eqnarray}
where $\Upomega^2$ and $\mathpzc{r}$ will be described below.
As a gauge condition we choose the hypersurface $U=0$ and $V=0$ to coincide with what will be the event horizons and we set
\begin{eqnarray}
\label{UVrange}
\Upomega^2(0,V)&=&\frac{1}{1-V^2},\\
\Upomega^2(U,0)&=&\frac{1}{1-U^2},
\end{eqnarray}
consistent with the fact that these hypersurfaces are to have infinite affine length. Fix parameters \mbox{$M>|e|\neq 0$}. The Reissner-Nordstr\"om metric \eqref{metric} in our gauge is uniquely determined from \eqref{EF}-\eqref{tee1} by setting
\begin{eqnarray}
{\mathpzc{r}}(0,V)&=&{\mathpzc{r}}|_{{{\mathcal H}_A}^+}=M+\sqrt{M^2-e^2}=r_+,\\
{\mathpzc{r}}(U,0)&=&{\mathpzc{r}}|_{{{\mathcal H}_B}^+}=M+\sqrt{M^2-e^2}=r_+.
\end{eqnarray}
Rearranging the Einstein-Maxwell equations \eqref{EF} using \eqref{metric} we obtain the following Hessian equation
\begin{eqnarray}
\label{uvr}
\partial_U \partial_V {\mathpzc{r}} &=&\frac{e^2 \Upomega^2}{4{\mathpzc{r}}^3}-\frac{\Upomega^2}{4{\mathpzc{r}}}-\frac{\partial_U {\mathpzc{r}} \partial_V {\mathpzc{r}}}{{\mathpzc{r}}},
\end{eqnarray}
from the $U,V$ component.
From the $\theta,\theta$ or equivalently $\phi,\phi$ component we obtain
\begin{eqnarray}
\label{omegalogrelation}
\partial_U \partial_V \log \Upomega^2&=&-\frac{2\partial_U \partial_V {\mathpzc{r}}}{{\mathpzc{r}}}\\
&\stackrel{\eqref{uvr}}{=}&-\frac{e^2 \Upomega^2}{2{\mathpzc{r}}^4}+\frac{\Upomega^2}{2{\mathpzc{r}}^2}+\frac{2\partial_U {\mathpzc{r}} \partial_V {\mathpzc{r}}}{{\mathpzc{r}}^2},
\end{eqnarray}
In fact, all relevant information about Reissner-Nordstr\"om geometry can be understood directly from \eqref{UVrange} to \eqref{omegalogrelation} without explicit expressions for $\Upomega^2(U,V)$ and ${\mathpzc{r}}(U,V)$. In particular, one can derive the Raychaudhuri equations
\begin{eqnarray}
\label{ray1}
\partial_U\left(\frac{\partial_U {\mathpzc{r}}}{\Upomega^2}\right)=0,\\
\label{ray2}
\partial_V\left(\frac{\partial_V {\mathpzc{r}}}{\Upomega^2}\right)=0,
\end{eqnarray}
from the above.
We can illustrate the 2-dimensional quotient spacetime $\mathcal{Q}$ as a subset of an ambient $\mathbb R^{1+1}$:
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{RN_ganz.jpg}
\caption[Text der im Bilderverzeichnis auftaucht]{Conformal diagram of the maximal domain of dependence of Reissner-Nordstr\"om spacetime.}
\label{RN_ganz}\end{figure}}
Identifying $U$, $V$ with ambient null coordinates of $\mathbb R^{1+1}$, the boundary of ${\mathcal Q} \subset {\mathbb R}^{1+1}$ is given by \mbox{$\pm 1\times [-1,1] \cup [-1,1]\times\pm 1$}. Let us further define the darker shaded region $II$ of Figure \ref{RN_ganz} by \mbox{${\mathcal Q}|_{II}=[0,1) \times [0,1)$}. Particularly important is \mbox{${\mathcal C}{\mathcal H}^+={\mathcal C}{\mathcal H}^+_A \cup {\mathcal C}{\mathcal H}^+_B=1\times (0,1] \cup (0,1]\times 1$}, which is the future boundary of the interior of region $II$.
We define \mbox{${\mathcal M}|_{II}=\pi^{-1}({\mathcal Q}|_{II})$}, where $\pi$ is the projection \mbox{$\pi: {\mathcal M}\rightarrow {\mathcal Q}$}.
\subsubsection{Eddington-Finkelstein coordinates}
It will be convenient to rescale the global double null coordinates and
define
\begin{eqnarray}
\label{u_v_Edd}
u=f(U)=\frac{2r_+}{{r_+}^2-e^2}\ln\left|\ln\left|\frac{1+U}{1-U}\right|\right|,\qquad v=h(V)=\frac{2r_+}{{r_+}^2-e^2}\ln\left|\ln\left|\frac{1+V}{1-V}\right|\right|.
\end{eqnarray}
Note that
$u$ is the retarded and $v$ is the advanced Eddington-Finkelstein coordinate. These coordinates are both regular in the interior of ${\mathcal Q}|_{II}$, cf.~ Figure \ref{RN_ganz}.
Nonetheless, we can view the whole of ${\mathcal Q}|_{II}$ as
\begin{eqnarray}
\label{QII}
{\mathcal Q}|_{II}=[-\infty, \infty) \times [-\infty,\infty),
\end{eqnarray}
where we have formally parametrized by
\begin{eqnarray*}
{{{\mathcal H}_A}^+}&=&\left\{-\infty\right\}\times [-\infty, \infty),\\
{{{\mathcal H}_B}^+}&=&[-\infty, \infty)\times\left\{-\infty\right\},
\end{eqnarray*}
as depicted in Figure \ref{range}, see also \eqref{u_v_Edd}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{range.jpg}
\caption[Text der im Bilderverzeichnis auftaucht]{Conformal diagram of darker shaded region $II$, compare Figure \ref{RN_ganz}, with the ranges of $(u,v)$ depicted.}
\label{range}\end{figure}}\\
In $u$, $v$ coordinates the metric is given by
\begin{eqnarray}
g=-\Omega^2(u,v){\mathrm{d}} u {\mathrm{d}} v+r^2(u,v)\left[{\mathrm{d}} \theta^2+\sin^2\theta {\mathrm{d}} \varphi^2\right],
\end{eqnarray}
with
\begin{eqnarray}
\label{omeganull}
{\Omega}^2(u,v)=\frac{\Upomega^2(U,V)}{\partial_U f \partial_V h}=-\left(1-\frac{2M}{r}+\frac{e^2}{r^2}\right),
\end{eqnarray}
where the unfamiliar minus sign on the right hand side arises since all definitions have been made suitable for the interior.
We will often make use of the fact that by the choice of Eddington-Finkelstein coordinates \eqref{u_v_Edd} for the interior we have scaled our coordinates such that
\begin{eqnarray}
\label{def_l_n}
\frac{\partial_u r}{\Omega^2}=-\frac12, \qquad \frac{\partial_v r}{\Omega^2}=-\frac12.
\end{eqnarray}
(The fact that the above expressions are constants follows from the Raychaudhuri equations \eqref{ray1} and \eqref{ray2}.)
Taking the derivatives of \eqref{omeganull} with respect to $u$ and $v$ and using \eqref{def_l_n} it follows that
\begin{eqnarray}
\label{u-neg}
\frac{\partial_u \Omega}{\Omega}(u,v)&=&\frac{1}{2r^2}\left(M-\frac{e^2}{r}\right),\\
\label{v-neg}
\frac{\partial_v \Omega}{\Omega}(u,v)&=&\frac{1}{2r^2}\left(M-\frac{e^2}{r}\right).
\end{eqnarray}
\subsection{Further properties of Reissner-Nordstr\"om geometry}
\label{geometry}
\subsubsection{$(t,r^{\star})$ and $(t,r)$ coordinates}
\label{rtcoords}
It is useful to define the function \mbox{$t:\mathring{{\mathcal M}}|_{II} \rightarrow {\mathbb R}$} by
\begin{eqnarray}
t(u,v)=\frac{v-u}{2},
\end{eqnarray}
where $\mathring{{\mathcal M}}|_{II}={\mathcal M}|_{II}\setminus \partial {\mathcal M}|_{II}$ is the interior of ${\mathcal M}|_{II}$.
Moreover, we define the function \mbox{$r^{\star}:\mathring{{\mathcal M}}|_{II}\rightarrow {\mathbb R}$} by
\begin{eqnarray}
\label{regge}
r^{\star}(u,v)=\frac{v+u}{2},
\end{eqnarray}
where $r^{\star}$ is usually referred to as the Regge-Wheeler coordinate.
Note that for coordinates $(t, r^{\star}, \varphi, \theta)$ defined in $\mathring{{\mathcal M}}|_{II}$ we have that $\frac{\partial}{
\partial t}$ is a spacelike Killing vector field which extends to the globally defined Killing vector field $T$ on ${\mathcal M}$.
By $\varphi_{\tau}$ we denote a 1-parameter group of diffeomorphisms generated by the Killing field $T$.
We can moreover relate the functions $r$ and $r^{\star}$ by
\begin{eqnarray}
\label{rstar1}
{\mathrm{d}} r^{\star}&=&\frac{{\mathrm{d}} r}{1-\frac{2M}{r}+\frac{e^2}{r^2}}\\
\label{rstar2}
\Rightarrow r^{\star}&=&r+\frac{1}{\kappa_+}\ln{|\frac{r-r_+}{r_+}|}+\frac{1}{\kappa_-}\ln{|\frac{r-r_-}{r_-}|}+C,
\end{eqnarray}
where $C$ is constant which is implicitly fixed by previous definitions,
\begin{eqnarray}
r_-=M-\sqrt{M^2-e^2},
\end{eqnarray}
and the surface gravities are given by
\begin{eqnarray}
\label{kappa}
\kappa_{\pm}=\frac{r_{\pm}-r_{\mp}}{2r_{\pm}^2}.
\end{eqnarray}
Note that $\kappa_{+}$ is the surface gravity at ${\mathcal H}^+$ and $\kappa_{-}$ is the surface gravity at ${\mathcal C}{\mathcal H}^+$.
The function $r(u,v)$ extends continuously and is monotonically decreasing in both $u$ and $v$ towards ${\mathcal C}{\mathcal H}^+$ such that we have
\begin{eqnarray}
r(u, \infty)&=&{r}|_{{{\mathcal C}{\mathcal H}_A}^+}=r_-,\\
r(\infty, v)&=&{r}|_{{{\mathcal C}{\mathcal H}_B}^+}=r_-.
\end{eqnarray}
\subsubsection{Angular momentum operators}
\label{angular}
We have already mentioned the generators of spherical symmetry $\leo_i$, $i=1,2,3$, in Section \ref{en_cur}. They are explicitly given by
\begin{eqnarray}
\label{angmom1}
\leo_{1}&=& \sin \varphi \partial_{\theta} + \cot \theta \cos \varphi \partial_{\varphi},\\
\leo_{2}&=& -\cos \varphi \partial_{\theta} + \cot \theta \sin \varphi \partial_{\varphi},\\
\label{angmom2}
\leo_{3}&=& - \partial_{\varphi},
\end{eqnarray}
which satisfy
\begin{eqnarray}
\label{sum1}
\sum_{i=1}^{3} \left(\leo_i \phi\right)^2&=&r^2|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,\\
\label{sum2}
\sum_{i=1}^{3} \sum_{j=1}^{3}\left(\leo_i \leo_j\phi\right)^2&=&r^4|\mbox{$\nabla \mkern-13mu /$\,}^2 \phi|^2,
\end{eqnarray}
where we define
\begin{eqnarray}
\label{nabb}
|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2=\frac{1}{r^2}\left[(\partial_{\theta} \phi)^2+ \frac{1}{\sin^2 {\theta}}(\partial_{\varphi} \phi)^2\right].
\end{eqnarray}
\subsubsection{The redshift, noshift and blueshift region}
\label{bnrsection}
As we have already mentioned in the introduction, in the interior we can distinguish
\begin{eqnarray}
\mbox{redshift} &\mbox{${\mathcal R}=\left\{r_{red}\leq r \leq r_+\right\}$}&,\\
\label{noshiftdef}
\mbox{noshift} &\mbox{${\mathcal N}=\left\{r_{blue}\leq r\leq r_{red}\right\}$}&,\\
\mbox{and blueshift} &\mbox{${\mathcal B}=\left\{r_{-}\leq r\leq r_{blue}\right\}$}&
\end{eqnarray}
subregions, as shown in Figure \ref{IIbnr}, for values $r_{red}$, $r_{blue}$ to be defined immediately below.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{region_IIbnr.jpg}
\caption[]{Region $II$ with distinction into redshift ${\mathcal R}$, noshift ${\mathcal N}$ and blueshift ${\mathcal B}$ regions.}
\label{IIbnr}\end{figure}}
In the redshift region ${\mathcal R}$ we make use of the fact that the surface gravity $\kappa_+$ of the event horizon is positive. The region is then characterized by the fact that there exists a vector field $N$ such that its associated current $J^N_{\mu}n^{\mu}_{v=const}$ on a $v=const$ hypersurface can be controlled by the related bulk term $K^N$, cf.~ Proposition \ref{mi}. This positivity of the bulk term $K^N$ is only possible sufficiently close to ${\mathcal H}^+$. In particular we shall define
\begin{eqnarray}
\label{rred}
r_{red}=r_+-\epsilon,
\end{eqnarray}
with $\epsilon>0$ and small enough such that Proposition \ref{mi} is applicable. (Furthermore, note that the quantity $M-\frac{e^2}{r}$ is always positive in ${\mathcal R}$.)
As defined in \eqref{noshiftdef} the $r$ coordinate in the noshift region ${\mathcal N}$ ranges between $r_{red}$ defined by \eqref{rred} and $r_{blue}$, defined below, strictly bigger than $r_-$.
In ${\mathcal N}$ we exploit the fact that $J^{-\partial_r}$ and $K^{-\partial_r}$ are invariant under translations along $\partial_t$. For that reason we can uniformly control the bulk by the current along a constant $r$ hypersurface. This will be explained further in Section \ref{zweite_region}.
The blueshift region ${\mathcal B}$ is characterized by the fact that the bulk term $K^{S_0}$ associated to the vector field $S_0$ to be defined in \eqref{Nstar} is positive. We define
\begin{eqnarray}
\label{rblue}
r_{blue}=r_-+\tilde{\epsilon},
\end{eqnarray}
with $\tilde{\epsilon}>0$ for an $\tilde{\epsilon}$ such that $M-\frac{e^2}{r}$ carries a negative sign
and such that (for convenience)
\begin{eqnarray}
\label{rchoice}
r^{\star}(r_{blue})>0.
\end{eqnarray}
In particular, in view of \eqref{u-neg} and \eqref{v-neg}
for $\tilde{\epsilon}$ sufficiently small the following lower bound holds in ${\mathcal B}$
\begin{eqnarray}
\label{lowerboundu}
0<{\beta} &\leq& -\frac{\partial_u \Omega}{\Omega},\\
\label{lowerboundv}
0<{\beta} &\leq& -\frac{\partial_v \Omega}{\Omega},
\end{eqnarray}
with $\beta$ a positive constant.
\subsection{Notation}
\label{nota}
We will describe certain regions derived from the hypersurfaces $r=r_{red}$, $r=r_{blue}$ and in addition the hypersurface $\gamma$ which will be defined in Section \ref{gamma_curve}.
For example given the hypersurface $r=r_{red}$ and the hypersurface $u=\tilde{u}$ we define the $v$ value at which these two hypersurfaces intersect by a function $v_{red}(\tilde{u})$ evaluated for $\tilde{u}$. Let us therefore introduce the following notation:
\begin{eqnarray}
\label{notation_neu}
v_{red}(\tilde{u}) \quad &\mbox{is determined by}& \quad r(v_{red}(\tilde{u}), \tilde{u})=r_{red},\nonumber\\
v_{\gamma}(\tilde{u}) \quad &\mbox{is determined by}& \quad (v_{\gamma}(\tilde{u}), \tilde{u}) \in \gamma,\nonumber\\
v_{blue}(\tilde{u}) \quad &\mbox{is determined by}& \quad r(v_{blue}(\tilde{u}), \tilde{u})=r_{blue},\nonumber\\
\mbox{and similarly we will also use}&&\nonumber\\
u_{red}(\tilde{v}) \quad &\mbox{is determined by}& \quad r(u_{red}(\tilde{v}), \tilde{v})=r_{red},\nonumber\\
u_{\gamma}(\tilde{v}) \quad &\mbox{is determined by}& \quad (u_{\gamma}(\tilde{v}), \tilde{v}) \in \gamma,\nonumber\\
u_{blue}(\tilde{v}) \quad &\mbox{is determined by}& \quad r(u_{blue}(\tilde{v}), \tilde{v})=r_{blue}.
\end{eqnarray}
For a better understanding the reader may also refer to Figure \ref{integralbild} and Figure \ref{integralbild2}.
{\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{integralbild.jpg}
\caption[]{Sketch of blueshift region ${\mathcal B}$ with quantities depicted dependent on $\tilde{u}$.}
\label{integralbild}\end{figure}}
{\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{integralbild2.jpg}
\caption[]{Sketch of blueshift region ${\mathcal B}$ with quantities depicted dependent on $\tilde{v}$.}
\label{integralbild2}\end{figure}}
Note that the above functions are well defined since $r=r_{red}$, $r=r_{blue}$ and $\gamma$ are spacelike hypersurfaces terminating at $i^+$.
\section{Horizon estimates and Cauchy stability}
\label{horizon_estimates}
Our starting point will be previously proven decay bounds for $\phi$ and its derivatives in the black hole {\it exterior} up to and including the event horizon; in particular we can state:
\begin{thm}
\label{anfang}
Let $\phi$ be a solution of the wave equation \eqref{wave} on a subextremal Reissner-Nordstr\"om background $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, arising from smooth compactly supported initial data on an arbitrary Cauchy hypersurface $\Sigma$, cf.~ Figure \ref{RN2}. Then, there exists $\delta>0$ such that
\begin{eqnarray}
\label{thpur}
\int\limits_{{\mathbb S}^2}\int\limits^{v+1}_v \left[(\partial_v \phi)^2 (-\infty, v)+|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(-\infty, v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& C_{0}v^{-2-2\delta},\\
\label{thpur1}
\int\limits_{{\mathbb S}^2}\int\limits^{v+1}_v \left[(\partial_v \leo\phi)^2(-\infty, v) +|\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(-\infty, v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& C_{1}v^{-2-2\delta},\\
\label{thpur2}
\int\limits_{{\mathbb S}^2}\int\limits^{v+1}_v\left[ (\partial_v \leo^2\phi)^2(-\infty, v) +|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(-\infty, v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& C_{2}v^{-2-2\delta},
\end{eqnarray}
on ${{\mathcal H}_A}^+$, for all $v$ and some positive constants $C_{0}$, $C_{1}$ and $C_{2}$ depending on the initial data.\footnote{The notation $\leo$ and $\leo^2$ is explained in Section \ref{leonotationsec} and simply denotes summation over angular momentum operators $\leo_i$ and $\leo_i\leo_j$.}
\end{thm}
\begin{proof}
The Theorem follows by putting together work of P.~ Blue and A.~ Soffer \cite{blue} on integrated local energy decay, M.~ Dafermos and I.~ Rodnianski \cite{m_price} on the redshift and V.~ Schlue \cite{volker} on improved decay using the method of \cite{m_new} in the exterior region.
The assumption of smoothness and compact support can be weakened. Moreover, we can in fact take $\delta$ arbitrarily close to $\frac12$, but $\delta>0$ is sufficient for our purposes and allows in principle for a larger class of data on $\Sigma$.
\end{proof}
On the other hand, trivially from Cauchy stability, boundedness of the energy along the second component of the past boundary of the characteristic rectangle $\Xi$, cf.~ Section \ref{firstlook}, which we have picked to be $v=1$, can be derived. More generally we can state the following proposition.
\begin{prop}
\label{initialdataprop}
Let \mbox{$u_{\diamond}, v_{\diamond} \in (-\infty, \infty)$}.
Under the assumption of Theorem \ref{anfang}, the energy at advanced Eddington-Finkelstein coordinate \mbox{$\left\{v=v_{\diamond}\right\}\cap\left\{{-\infty}\leq u \leq {u_{\diamond}}\right\}$}
is bounded from the initial data
\begin{eqnarray}
\label{proppur}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{u_{\diamond}}\left[ \Omega^{-2}(\partial_u \phi)^2(u,v_{\diamond})+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u,v_{\diamond})\right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2} &\leq& D_{0}(u_{\diamond},v_{\diamond}),\\
\label{proppur1}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{u_{\diamond}}\left[ \Omega^{-2}(\partial_u \leo\phi)^2(u,v_{\diamond})+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(u,v_{\diamond})\right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& D_{1}(u_{\diamond},v_{\diamond}),\\
\label{proppur2}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{u_{\diamond}} \left[\Omega^{-2}(\partial_u \leo^2\phi)^2(u,v_{\diamond})+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u,v_{\diamond})\right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& D_{2}(u_{\diamond},v_{\diamond}),
\end{eqnarray}
and further
\begin{eqnarray}
\label{aufeins1}
\sup_{-\infty\leq u \leq {u_{\diamond}}}\int\limits_{{\mathbb S}^2} (\phi)^2(u,v_{\diamond}){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& D_{0}(u_{\diamond},v_{\diamond}),\\
\sup_{-\infty\leq u \leq {u_{\diamond}}}\int\limits_{{\mathbb S}^2} (\leo\phi)^2(u,v_{\diamond}){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& D_{1}(u_{\diamond},v_{\diamond}),\\
\label{aufeins3}
\sup_{-\infty\leq u \leq {u_{\diamond}}}\int\limits_{{\mathbb S}^2} (\leo^2\phi)^2(u,v_{\diamond}){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& D_{2}(u_{\diamond},v_{\diamond}),
\end{eqnarray}
with $D_{0}(u_{\diamond},v_{\diamond})$, $D_{1}(u_{\diamond},v_{\diamond})$ and $D_{2}(u_{\diamond},v_{\diamond})$ positive constants depending on the initial data on $\Sigma$.
\end{prop}
\begin{proof}
This follows immediately from local energy estimates in a compact spacetime region. Note the $\Omega^{-2}$ and $\Omega^{2}$ weights which arise since $u$ is not regular at ${\mathcal H}^+_A$.
\end{proof}
\section{Statement of the theorem and outline of the proof in the neighbourhood of $i^+$}
\label{outline}
The most difficult result of this paper can now be stated in the following theorem.
\begin{thm}
\label{dashier}
On subextremal Reissner-Nordstr\"om spacetime with $M>|e|\neq 0$, let $\phi$ be as in Theorem \ref{anfang}, then
\begin{eqnarray*}
|\phi|\leq C
\end{eqnarray*}
locally in the black hole interior up to
${\mathcal C}{\mathcal H}^+$ in a ``small neighbourhood'' of timelike infinity $i^+$,
that is in \mbox{$(-\infty, u_{\text{\ding{34}}}] \times [1, \infty)$} for some $u_{\text{\ding{34}}}>-\infty$.
\end{thm}
{\em Remark.} We will see that $C$ depends only on the initial data.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{RN.jpg}
\caption[]{Maximal development of Cauchy hypersurface $\Sigma$ in Reissner-Nordstr\"om spacetime $({\mathcal M},g)$.}
\label{RN2}\end{figure}}
We will consider a characteristic rectangle $\Xi$ extending from ${{\mathcal H}_A}^+$ as shown in Figure \ref{RN_character}.
We pick the characteristic rectangle to be defined by \mbox{$\Xi=\left\{(-\infty\leq u\leq u_{\text{\ding{34}}}), (1\leq v <\infty)\right\}$}, where $u_{\text{\ding{34}}}$ is sufficiently close to $-\infty$ for reasons that will become clear later on, cf.~ Proposition \ref{kastle}.
As described in Section \ref{horizon_estimates}, from bounds of data on $\Sigma$
bounds on the solution on the lower segments follow according to Theorem \ref{anfang} and Proposition \ref{initialdataprop}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{RN_character.jpg}
\caption[]{Characteristic rectangle $\Xi$ in the interior of Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, for $\Xi$ zoomed in see Figure \ref{bnr}.}
\label{RN_character}\end{figure}}
In order to prove Theorem \ref{dashier}
we distinguish the redshift ${\mathcal R}$, the noshift ${\mathcal N}$ and the blueshift ${\mathcal B}$ region, with the properties as explained in Section \ref{bnrsection},
cf.~ Figure \ref{bnr}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{bnr_region_mit_gamma1.jpg}
\caption[]{Characteristic rectangle $\Xi$ with redshift ${\mathcal R}$, noshift ${\mathcal N}$ and blueshift ${\mathcal B}$ regions.}
\label{bnr}\end{figure}}
This distinction is made since different vector fields have to be employed in the different regions\footnote{The reader may wonder why the noshift region ${\mathcal N}$ is introduced instead of just separating the red- and the blueshift regions along the $r$ hypersurface whose value renders the quantity $M-\frac{e^2}{r}$ equal zero. This was to ensure strict positivity/negativity of the quantity in the redshift/blueshift region.}.
In the redshift region ${\mathcal R}$ we will make use of the redshift vector field $N$ of \cite{m_lec} on which we will elaborate more in Section \ref{first_section}.
Proposition \ref{mi} gives the positivity of the bulk $K^N$ which thus bounds the current $J^N_{\mu}n^{\mu}_{v=const}$ from above.
Applying the divergence theorem, decay up to $r=r_{red}$ will be proven.
In the noshift region ${\mathcal N}$ we can simply appeal to the fact that the future directed timelike vector field $-\partial_r$ is invariant under the flow of the spacelike Killing vector field $\partial_t$.
It is for that reason that the bulk term $K^{-\partial_r}$
can be uniformly controlled by the energy flux $J_{\mu}^{-\partial_r}n^{\mu}_{r=\bar{r}}$ through the $r=\bar{r}$ hypersurface.
Decay up to $r=r_{blue}$ will be proven by making use of this together with the uniform boundedness of the $v$ length of ${\mathcal N}$.
To understand the blueshift region ${\mathcal B}$, we will partition it by the hypersurface $\gamma$ admitting logarithmic distance in $v$ from $r=r_{blue}$, cf.~ Section \ref{gamma_curve}. We will then separately consider the region to the past of $\gamma$, \mbox{$J^-(\gamma)\cap {\mathcal B}$} and the region to the future of $\gamma$, \mbox{$J^+(\gamma)\cap {\mathcal B}$}.
The region to the future of $\gamma$ is characterized by good decay bounds on $\Omega^2$
(implying for instance that the
spacetime volume is finite, $\operatorname{Vol}(J^+(\gamma))<C$).
In \mbox{$J^-(\gamma)\cap {\mathcal B}$} we use a vector field
\begin{eqnarray}
\label{Nstar}
{S_0}=r^q\partial_{r^{\star}}=r^q(\partial_u+\partial_v),
\end{eqnarray}
where $q$ is sufficiently large, cf.~ Section \ref{blueshift1}. We will see that for the right choice of $q$ we can render the associated bulk term $K^{{S_0}}$ positive which is the ``good'' sign when using the divergence theorem.
In order to complete the proof, we consider finally the region \mbox{$J^+(\gamma)\cap {\mathcal B}$} and propagate the decay further from the hypersurface $\gamma$ up to the Cauchy horizon in a neighbourhood of $i^+$. For this, we introduce a new timelike vector field ${S}$ defined by
\begin{eqnarray}
\label{N1}
{S}=|u|^p\partial_u+v^p\partial_v,
\end{eqnarray}
for an arbitrary $p$ such that
\begin{eqnarray}
\label{waspist}
1<p\leq 1+2\delta,
\end{eqnarray}
where $\delta$ is as in Theorem \ref{anfang}.
We use pointwise estimates on $\Omega^2$ in $J^+(\gamma)$ as a crucial step, cf.~ Section \ref{finiteness}.
Putting everything together, in view of the geometry and the weights of $S$, we finally obtain for all $v_*\geq 1$
\begin{eqnarray}
\label{fluxscetch}
\int\limits_{{\mathbb S}^2} \int\limits_1^{v_*} v^p(\partial_v \phi)^2 r^2{\mathrm{d}} v {\mathrm{d}} \sigma_{\mathbb S^2}\leq \mbox{Data},
\end{eqnarray}
for the weighted flux.
Using the above, the uniform boundedness for $\phi$ stated in Theorem \ref{dashier} then follows from an argument that can be sketched as follows.
Let us first see how we get an integrated bound on the spheres of symmetry.
By the fundamental theorem of calculus and the Cauchy-Schwarz inequality one obtains
\begin{eqnarray}
\int\limits_{{\mathbb S}^2} \phi^2(u, v_*, \theta, \varphi){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq& C\int\limits_{{\mathbb S}^2}\left(\int\limits_1^{v_*} v^p(\partial_v \phi)^2{\mathrm{d}} v\right)\left(\int\limits_1^{v_*} v^{-p}{\mathrm{d}} v\right)r^2{\mathrm{d}} \sigma_{\mathbb S^2}+\mbox{data},\nonumber
\end{eqnarray}
where the first factor of the first term is controlled by \eqref{fluxscetch}.
Therefore, we further get
\begin{eqnarray}
\label{fundcauchy11}
\int\limits_{{\mathbb S}^2} \phi^2{\mathrm{d}} \sigma_{\mathbb S^2}&\stackrel{\eqref{fluxscetch}}{\leq}&\mbox{Data}\int\limits_{{\mathbb S}^2}\int\limits_1^{v_*} v^{-p}{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}+\mbox{data}\nonumber\\
&\leq&\mbox{Data}+\mbox{data},
\end{eqnarray}
where we have used \mbox{$\int\limits^{\infty}_{1} v^{-p}{\mathrm{d}} v< \infty$} which followed from the first inequality of \eqref{waspist}.
Obtaining a pointwise statement from the above will be achieved by commuting \eqref{wave} with symmetries as well as applying Sobolev embedding.
As outlined in Section \ref{en_cur} in Reissner-Nordstr\"om geometry we have \mbox{$\Box_g \leo_i \phi=0$}, where $\leo_i$ with $i=1,2,3$ are the 3 spacelike Killing vector fields resulting from the spherical symmetry.
Thus one obtains the analogue of \eqref{fundcauchy11} but with $\leo_i \phi$ and \mbox{$\leo_i \leo_j \phi$} in place of $\phi$.
Using Sobolev embedding on ${\mathbb S}^2$ thus leads immediately to the desired bounds. See Section \ref{uni_bounded}.
This will close the proof of Theorem \ref{dashier}.\\
\section{Propagating through ${\mathcal R}$ from ${\mathcal H}^+$ to $r=r_{red}$}
\label{first_section}
The estimates in this and the following section are motivated by work of Luk \cite{luk}.
He proves that any polynomial decay estimate that holds along the event horizon of Schwarzschild black holes can be propagated to any constant $r$ hypersurface in the black hole interior.
This followed a previous spherically symmetric argument of \cite{m_interior}. See also Dyatlov \cite{dyatlov}.
As outlined in Section \ref{horizon_estimates}, we will first propagate energy decay from ${\mathcal H}^+$ up to the $r=r_{red}$ hypersurface.
The rough idea can be understood with the help of Figure \ref{tilde_r_decay}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{region1.jpg}
\caption[Text der im Bilderverzeichnis auftaucht]{Regions ${\mathcal R}_I$ and $\tilde{{\mathcal R}_I}$.}
\label{tilde_r_decay}\end{figure}
By Theorem \ref{anfang} we are given energy decay on the event horizon ${\mathcal H}^+$, see dash-dotted line.
By using the energy identity for the vector field $N$ in region \mbox{${\mathcal R}_I=\left\{r_{red}\leq r \leq r_+\right\}\cap \left\{1<v\leq v_*\right\}$}, the coarea formula etc., we obtain decay of the flux through constant $v$ hypersurfaces throughout the entire region. Using this result and considering the energy identity once again in region \mbox{$\tilde{{\mathcal R}}_I=\left\{r_{red}\leq r \leq r_+\right\}\cap \left\{v_*\leq v\leq v_*+1\right\}$} we eventually obtain decay on the $r=r_{red}$ hypersurface, note the dashed line.
The redshift vector field was already introduced by Dafermos and Rodnianski in \cite{m_red} and elaborated on again in \cite{m_lec}.
The existence of such a vector field in the neighbourhood of a Killing horizon ${\mathcal H}^+$ depends only on the positivity of the surface gravity, in this case $\kappa_+$. Thus by \eqref{kappa} the following proposition follows by Theorem 7.1 of \cite{m_lec}.
\begin{prop}
\label{mi}
(M. Dafermos and I. Rodnianski) For $r_{red}$ sufficiently close to $r_+$ there exists a $\varphi_{\tau}$-invariant\footnote{cf.~ Section \ref{geometry}} smooth future directed timelike vector field $N$ on \mbox{$\left\{r_{red}\leq r \leq r_+\right\}\cap \left\{v\geq 1\right\}$} and a positive constant $b_1$ such that
\begin{eqnarray}
\label{energy_controll}
b_1J_{\mu}^N(\phi) n^{\mu}_{{v}} \leq K^N(\phi) ,
\end{eqnarray}
for all solutions $\phi$ of $\Box_g \phi=0$.
\end{prop}
The decay bound along $r=r_{red}$ can now be stated in the following proposition.
\begin{prop}
\label{rtildedecay}
Let $\phi$ be as in Theorem \ref{anfang}.
Then, for all $\tilde{r} \in [r_{red}, r_+)$, with $r_{red}$ as in Proposition \ref{mi} and for all $v_*>1$,
\begin{eqnarray*}
\label{decay_*}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{r=\tilde{r}} {\mbox{dVol}}_{r=\tilde{r}} \leq C v_*^{-2-2\delta} ,
\end{eqnarray*}
with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{prop}
{\em Remark 1.} The decay in Proposition \ref{rtildedecay} matches the decay on ${\mathcal H}^+$ of Theorem \ref{anfang}.
{\em Remark 2.}
$n^{\mu}_{r=r_{red}}$ denotes the normal to the $r=r_{red}$ hypersurface oriented according to Lorentzian geometry convention. ${\mbox{dVol}}$ denotes the volume element over the entire spacetime region and ${\mbox{dVol}}_{r=r_{red}}$ denotes the volume element on the $r=r_{red}$ hypersurface. Similarly for all other subscripts.\footnote{Refer to Appendix \ref{Jcurrents} for further discussion of the volume elements.}
\begin{proof}
Applying the divergence theorem, see e.g.~ \cite{m_lec} or \cite{taylor},
in region \mbox{${\mathcal R}_I=\left\{r_{red}\leq r \leq r_+\right\}\cap \left\{v_0\leq v\leq v_*\right\}$}, with $v_0\geq 1$, we obtain
\begin{eqnarray*}
& &\int\limits_ {{\mathcal R}_I} K^N(\phi) {\mbox{dVol}}
+ \int\limits_ {\left\lbrace v_0 \leq v \leq v_*\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}
+ \int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*} \\
&=&
\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_0} {\mbox{dVol}}_{v=v_0}
+\int\limits_ {\left\lbrace v_0 \leq v \leq v_*\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{{\mathcal H}^+} {\mbox{dVol}}_{{\mathcal H}^+}.
\end{eqnarray*}
We immediately see that the second term on the left hand side is positive since $r=r_{red}$ is a spacelike hypersurface and $N$ is a timelike vector field. Therefore, we write
\begin{eqnarray}
\label{R_I}
& & \int\limits_ {{\mathcal R}_I} K^N(\phi) {\mbox{dVol}}
+ \int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*} \nonumber \\
&\leq&
\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_0} {\mbox{dVol}}_{v=v_0}
+\int\limits_ {\left\lbrace v_0 \leq v \leq v_*\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{{\mathcal H}^+} {\mbox{dVol}}_{{\mathcal H}^+}.
\end{eqnarray}
By Theorem \ref{anfang} we have
\begin{eqnarray*}
\int\limits_ {\left\lbrace v_0 \leq v \leq v_*\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{{\mathcal H}^+} {\mbox{dVol}}_{{\mathcal H}^+} \leq C_{0} \mbox{max}\left\lbrace v_*-v_0,1\right\rbrace v_0^{-2-2\delta}.
\end{eqnarray*}
Using that the energy current associated to the timelike vector field $N$ is controlled by the deformation $K^N$ as shown in \eqref{energy_controll}
and substituting
\begin{eqnarray}
\label{def_E}
E(\phi;\tilde{v})=\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=\tilde{v}} {\mbox{dVol}}_{v=\tilde{v}}
\end{eqnarray}
into \eqref{R_I} as well as using the coarea formula
\begin{eqnarray}
\label{coaerea2}
\int\limits_ {{{\mathcal R}_I} } J_{\mu}^N(\phi) n^{\mu}_{v=\tilde{v}} {\mbox{dVol}}
\sim \int\limits_{v_0}^{v_*} \int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}} {\mathrm{d}} {\bar{v}},
\end{eqnarray}
for the bulk term,\footnote{where $f\sim g$ means that there exist constants $0<b<B$ with $bf<g<Bf$} we obtain for all $v_0\geq1$ and $v_*>v_0$, the relation
\begin{eqnarray}
\label{decay_statement}
E(\phi;v_*) + \tilde{b}_1\int\limits_{v_0}^{v_*} E(\phi;{\bar{v}}){\mathrm{d}} {\bar{v}} \leq E(\phi;v_0)+C_{0} \mbox{max}\left\lbrace v_*-v_0,1\right\rbrace v_0^{-2-2\delta}.
\end{eqnarray}
Note by Proposition \ref{initialdataprop}, applied to $u_{\diamond}$ defined through the relation $r_{red}=r(u_{\diamond},1)$, we have
\begin{eqnarray}
\label{estineu}
E(\phi;1) \leq C D_0(u_{\diamond},1),
\end{eqnarray}
since the vector field $N$ is regular at ${\mathcal H}^+$ and thus $E(\phi;1)$ is comparable to the left hand side of \eqref{proppur}.
In order to obtain estimates from \eqref{decay_statement} we appeal to the following lemma.
\begin{lem}
Let $f:[1, \infty)\rightarrow {\mathbb R}^+$,
\label{decay_lemma2}
\begin{eqnarray}
\label{starter3}
f(t) + b\int\limits_{\tilde{t}}^{t} f(\bar{t}){\mathrm{d}} {\bar{t}} \leq f(\tilde{t}) +C_0(t-\tilde{t}+1)\tilde{t}^{-\tilde{p}},
\end{eqnarray}
for all $\tilde{t}\geq 1$, where $C_0$, $\tilde{p}$ are positive constants.
Then for any $t\geq 1$ we have
\begin{eqnarray}
\label{ftwant}
f(t) \leq \tilde{C}t^{-\tilde{p}},
\end{eqnarray}
where $\tilde{C}$ depends only on $f(1)$, $b$ and $C_0$.
\end{lem}
\begin{proof}
For $t>t_0$, we will show \eqref{ftwant} by a continuity argument. It suffices to show that
\begin{eqnarray}
\label{ft}
f(\tilde{t})\leq 2\tilde{C} \tilde{t}^{-\tilde{p}}, \quad \mbox{for $\tilde{t}\leq t$},
\end{eqnarray}
leads to
\begin{eqnarray}
\label{wantedft}
\Rightarrow f(\tilde{t}) \leq \tilde{C}\tilde{t}^{-\tilde{p}}, \quad \mbox{for $\tilde{t}\leq t$},
\end{eqnarray}
for some large enough constant $\tilde{C}$.
We note first that given any $t_0$, from \eqref{starter3} we obtain, \mbox{$\forall \, 1\leq t\leq t_0$}
\begin{eqnarray}
\label{startert0}
f(t) &\leq& f(1) +C_0\, t ,\nonumber\\
&\leq& \left[ f(1){t_0}^{\tilde{p}} +C_0\, {t_0}^{+\tilde{p}+1}\right]{t}^{-\tilde{p}}.
\end{eqnarray}
Given $t\geq t_0$, choose $\tilde{t}=t-L$ for an $L$ to be determined later.
Moreover, $t_0$ will have to be chosen large enough so that $\forall \, t \geq t_0$,
\begin{eqnarray}
\label{tverhaeltnis}
(t-L)^{-\tilde{p}}=\tilde{t}^{-\tilde{p}}<2t^{-\tilde{p}}.
\end{eqnarray}
Given a $t$ satisfying \eqref{ft} applying \eqref{starter3} yields
\begin{eqnarray}
\label{starter5}
f(t) + b\int\limits_{\tilde{t}}^{t} f(\bar{t}){\mathrm{d}} {\bar{t}} &\leq& \left[2\tilde{C} +C_0(L+1)\right]\tilde{t}^{-\tilde{p}}\nonumber\\
&\stackrel{\eqref{tverhaeltnis}}{\leq}&\left[4\tilde{C} +2C_0(L+1)\right]t^{-\tilde{p}}.
\end{eqnarray}
Further, by the pigeonhole principle, there exists $t_{in} \in \left[\tilde{t},t\right]$ such that
\begin{eqnarray}
\label{pigeon1}
f(t_{in})\leq \frac{1}{L} \int\limits_{\tilde{t}}^{t} f(\bar{t}){\mathrm{d}} {\bar{t}}.
\end{eqnarray}
Since $f(t)$ is a positive function \eqref{starter5} also leads to
\begin{eqnarray}
\label{starter4}
b\int\limits_{\tilde{t}}^{t} f(\bar{t}){\mathrm{d}} {\bar{t}} \leq \left[4\tilde{C} +2C_0(L+1)\right]t^{-\tilde{p}}.
\end{eqnarray}
Thus, \eqref{pigeon1} and \eqref{starter4} yield
\begin{eqnarray}
\label{vier}
f(t_{in})\leq \frac{1}{bL}\left[4\tilde{C} +2C_0(L+1)\right]t^{-\tilde{p}}.
\end{eqnarray}
Now let $\tilde{t}=t_{in}$ and use \eqref{vier} in \eqref{starter3}, then
\begin{eqnarray}
\label{schluss}
f(t)\leq f(t) + b\int\limits_{t_{in}}^{t} f(\bar{t}){\mathrm{d}} {\bar{t}} &\leq& \left(\frac{1}{bL}\left[2\tilde{C} +C_0(L+1)\right]+C_0(L+1)\right)t_{in}^{-\tilde{p}},\nonumber\\
&\stackrel{\eqref{tverhaeltnis}}{\leq}& \left[\frac{4\tilde{C}}{bL}+\frac{2C_0(L+1)}{bL}+2C_0(L+1)\right]t^{-\tilde{p}}.
\end{eqnarray}
If $1-\frac{4}{bL}>0$
and
\begin{eqnarray}
\tilde{C}\geq \left(1-\frac{4}{bL}\right)^{-1}\left[\frac{2C_0(L+1)}{bL}+2C_0(L+1)\right]
\end{eqnarray}
then \eqref{wantedft} follows.
Thus picking first $L$ such that $1-\frac{4}{bL}>0$, and then $t_0$ such that $t_0\geq L+1$ and satisfying \eqref{tverhaeltnis}, and finally choosing $\tilde{C}$ as \mbox{$\tilde{C}=\mbox{max}\left\{\left[ f(1) +C_0{t_0}^{-1-2\delta}\right],\left(1-\frac{4}{bL}\right)^{-1}\left[\frac{2C_0(L+1)}{bL}+2C_0(L+1)\right]\right\}$} \eqref{wantedft} and thus
\eqref{ftwant} follows by continuity.
\end{proof}
By Lemma \ref{decay_lemma2} we obtain from \eqref{decay_statement} together with \eqref{estineu}
\begin{eqnarray}
\label{decay_1}
E(\phi; v_*)=\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
\leq \tilde{\tilde{C}} {v_*}^{-2-2\delta},
\end{eqnarray}
with $\tilde{\tilde{C}}$ depending on $\tilde{b}_1$ and $D_{0}(u_{\diamond}, 1)$.
Finally, in order to close the proof of Proposition \ref{rtildedecay} we perform again the divergence theorem but for region \mbox{$\tilde{{\mathcal R}}_I=\left\{r_{red}\leq r \leq r_+\right\}\cap \left\{v_*\leq v\leq v_*+1\right\}$}:
\begin{eqnarray}
\label{nochmaldiv}
& &\int\limits_{\tilde{{\mathcal R}_I}} K^N(\phi) {\mbox{dVol}}
+ \int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}
+ \int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*
+1} {\mbox{dVol}}_{v=v_*+1}\nonumber\\
&=&
\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{{\mathcal H}^+} {\mbox{dVol}}_{{\mathcal H}^+}.
\end{eqnarray}
In view of the signs we obtain
\begin{eqnarray*}
\Rightarrow && \int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}
\leq
\int\limits_ {\left\lbrace r_{red}\leq r \leq r_+\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^N(\phi) n^{\mu}_{{\mathcal H}^+} {\mbox{dVol}}_{{\mathcal H}^+}.
\end{eqnarray*}
Due to \eqref{decay_1} and Theorem \ref{anfang} we are left with the conclusion of Proposition \ref{rtildedecay}.
\end{proof}
Note that the above also implies the following statement.
\begin{cor}
Let $\phi$ be as in Theorem \ref{anfang} and for $r_{red}$ as in Proposition \ref{mi}. Then, for all $v_*\geq 1$, $v_*+1\leq v_{red}(\tilde{u})$ and for all $\tilde{u}$ such that $r(\tilde{u},v_*+1) \in [r_{red}, r_+)$, we have
\label{cor5.2}
\begin{eqnarray}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J^N_\mu(\phi)n^\mu_{u=\tilde{u}}dVol_{u=\tilde{u}} \le C v_*^{-2-2\delta},
\end{eqnarray}
with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$ and $v_{red}(\tilde{u})$ as in \eqref{notation_neu}.
\end{cor}
\begin{proof}
The conclusion of the statement follows by applying again the divergence theorem
and using the results of the proof of Proposition \ref{rtildedecay}.
\end{proof}
\section{Propagating through ${\mathcal N}$ from $r=r_{red}$ to $r=r_{blue}$}
\label{zweite_region}
Now that we have obtained a decay bound along the $r=r_{red}$ hypersurface in the previous section, we propagate the estimate further inside the black hole through the noshift region ${\mathcal N}$ up to the $r=r_{blue}$ hypersurface.
In order to do that we will use the future directed timelike vector field
\begin{eqnarray}
\label{partial_r}
-\partial_r=\frac{1}{{\Omega^2}}(\partial_u+\partial_v).
\end{eqnarray}
Using \eqref{partial_r} in \eqref{Kplug} of Appendix \ref{Kcurrents} we obtain
\begin{eqnarray}
\label{partialrbulk}
K^{-\partial_r}&=&\frac{4}{\Omega^3}\left[\frac{\partial_u \Omega}{\Omega}(\partial_v\phi)^2 +\frac{\partial_v \Omega}{\Omega}(\partial_u\phi)^2 \right]\nonumber\\
&&- \frac{4}{r\sqrt{\Omega^2}}(\partial_u \phi \partial_v \phi)\nonumber\\
&&-\left(\frac{\partial_u \Omega}{\Omega^2}+\frac{\partial_v \Omega}{\Omega^2}\right)\left(\frac{1}{\Omega}-1\right)|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,
\end{eqnarray}
for the bulk current. It has the property that it can be estimated by
\begin{eqnarray}
\label{controll_bulk_partial_r}
|K^{-\partial_r}(\phi)| \leq B_1 J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r={\bar{r}}},
\end{eqnarray}
where $B_1$ is independent of $v_*$.
Validity of the estimate can in fact be seen without computation from the fact that timelike currents, such as $J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r={\bar{r}}}$ contain all derivatives.
The uniformity of $B_1$ is given by the fact that $K^{-\partial_r}$ and $J^{-\partial_r}$ are invariant under translations along $\partial_t$, cf.~ Section \ref{rtcoords} for definition of the $t$ coordinate. Therefore, we can just look at the maximal deformation on a compact \mbox{$\left\{t=const \right\} \cap \left\{r_{blue} \leq r \leq r_{red}\right\}$} hypersurface and get an estimate for the deformation everywhere.
\begin{prop}
\label{r_{red}}
Let $\phi$ be as in Theorem \ref{anfang}, $r_{blue}$ as in \eqref{rblue} and $r_{red}$ as in Proposition \ref{mi}. Then, for all $v_*>1$ and $\tilde{r} \in [r_{blue},r_{red})$, we have
\begin{eqnarray*}
\int\limits_ {\left\lbrace v_* \leq v \leq {v_*}+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=\tilde{r}} {\mbox{dVol}}_{r=\tilde{r}}
\leq
C{v_*}^{-2-2\delta},
\end{eqnarray*}
with $C$ depending on the initial data or more precisely depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{prop}
\begin{proof}
Given $v_*$, we define regions ${\mathcal R}_{II}$ and $\tilde{{\mathcal R}}_{II}$ as in Figure \ref{r_{red}_decay}, where we use \eqref{notation_neu} and
\begin{eqnarray}
\label{notation_v_*}
v(\tilde{r}, v_*) \quad &\mbox{is determined by}& \quad r(u_{blue}(v_*), v(\tilde{r}, v_*))=\tilde{r}.
\end{eqnarray}
Thus the depicted regions are given by \mbox{${\mathcal R}_{II}\cup \tilde{{\mathcal R}}_{II}= {\mathcal D}^+(\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace \cap \left\lbrace r=r_{red}\right\rbrace )\cap {\mathcal N} $}, where region \mbox{${\mathcal R}_{II}$} is given by \mbox{${\mathcal R}_{II}= {\mathcal D}^+(\left\lbrace v_1 \leq v \leq v_*\right\rbrace \cap \left\lbrace r=r_{red}\right\rbrace) $}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{region2.jpg}
\caption[Text der im Bilderverzeichnis auftaucht]{Region ${\mathcal R}_{II}\cup \tilde{{\mathcal R}}_{II}$ represented as the hatched area.}
\label{r_{red}_decay}\end{figure}}
In the following we will apply the divergence theorem in region ${\mathcal R}_{II}\cup \tilde{{\mathcal R}}_{II}$ to obtain decay on an arbitrary $r=\tilde{r}$ hypersurface, dash-dotted line, for $\tilde{r} \in [r_{blue},r_{red})$, from the derived decay on the $r=r_{red}$ hypersurface.
\begin{eqnarray*}
\int\limits_ {{\mathcal R}_{II}\cup \tilde{{\mathcal R}}_{II}} K^{-\partial_r}(\phi) {\mbox{dVol}}
&+& \int\limits_ {\left\lbrace r_{blue}\leq r \leq r_{red}\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{u=u_{blue}(v_*)} {\mbox{dVol}}_{u=u_{blue}(v_*)}\\
+ \int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}}
&+& \int\limits_ {\left\lbrace r_{blue}\leq r \leq r_{red}\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{v=v_*+1} {\mbox{dVol}}_{v=v_*+1} \\
=
\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}.
\end{eqnarray*}
The second integral of the left hand side represents the current through the $u=u_{blue}(v_*)$ hypersurface, defined by \eqref{notation_neu}.
As $u=u_{blue}(v_*)$ is a null hypersurface and $-\partial_r$ is timelike, the positivity of that second term is immediate. Similarly, the fourth term of the left hand side of our equation is positive and we obtain
\begin{eqnarray*}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}}
\leq
\int\limits_ {{\mathcal R}_{II}\cup \tilde{{\mathcal R}}_{II}} |K^{-\partial_r}(\phi)| {\mbox{dVol}}
+\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}.
\end{eqnarray*}
Further, we use that the deformation $K^{-\partial_r}$ is controlled by the energy associated to the timelike vector field $-\partial_r$ as stated in \eqref{controll_bulk_partial_r}.
Thus we obtain
\begin{eqnarray}
\label{oberes}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}}
&\leq&
B_1 \int\limits_{{\mathcal R}_{II}\cup{\tilde{{\mathcal R}}}_{II}} J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=\bar{r}} {\mbox{dVol}}\nonumber\\
&&+\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}.
\end{eqnarray}
By the coarea formula we obtain
\begin{eqnarray}
\label{94}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}}
&\leq&
\tilde{B}_1 \int\limits_ {r_{blue}}^{r_{red}}\int\limits_{\left\lbrace v(\bar{r},v_*) \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=\bar{r}} {\mbox{dVol}}_{r=\bar{r}}{\mathrm{d}} \bar{r}\nonumber\\
&&+\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}.
\end{eqnarray}
Now let
\begin{eqnarray}
E(\phi;\tilde{r},\tilde{v})=\int\limits_ {\left\lbrace \tilde{v} \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=\tilde{r}} {\mbox{dVol}}_{r=\tilde{r}},
\end{eqnarray}
with $\tilde{r} \in [r_{blue},r_{red})$.
Replacing $r_{blue}$ with $\tilde{r}$ in the above, considering the future domain of dependence of $\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace \cap \left\lbrace r=r_{red}\right\rbrace$ up to the $r=\tilde{r}$ hypersurface we obtain similarly to \eqref{94}
\begin{equation}
\label{E_energy2}
E(\phi;\tilde{r},v(\tilde{r}, v_*))\leq \tilde{B}_1\int\limits_ {\tilde{r}}^{r_{red}}E(\phi;\bar{r},v(\bar{r}, v_*)){\mathrm{d}} \bar{r} +E(\phi;r_{red},v_1).
\end{equation}
Using Gr\"onwall's inequality in \eqref{E_energy2} yields
\begin{eqnarray}
E(\phi;\tilde{r},v(\tilde{r}, v_*))&\leq& E(\phi;r_{red},v_1)\left[ 1+\tilde{B}_1(r_{red}-\tilde{r})e^{\tilde{B}_1(r_{red}-\tilde{r})}\right]\nonumber \\
\Rightarrow E(\phi;\tilde{r},v(\tilde{r}, v_*))&\leq& \tilde{C}E(\phi;r_{red},v_1).
\end{eqnarray}
Finally, note that
\begin{eqnarray}
\label{vconstk}
\left[v_*+1\right]-v(r_{red}, v_*) =\left[v_*+1\right]-v_1= k< \infty,
\end{eqnarray}
where $k=2\left[r^{\star}(r_{blue})-r^{\star}(r_{red})\right]+1$.
This can be seen since \eqref{def_l_n} and \eqref{rstar1} yields
\begin{eqnarray*}
&&\frac{\partial_v r}{\Omega^2}=-\partial_v r^{\star}=-\frac12\\
\Rightarrow&&r^{\star}(u_{blue}(v_*), v_*)-r^{\star}(u_{blue}(v_*), v_1)=r^{\star}(r_{blue})-r^{\star}(r_{red})\stackrel{\eqref{regge}}{=}\frac12 (v_*-v_1).
\end{eqnarray*}
Further, by using the conclusion of Proposition \ref{rtildedecay} and \eqref{vconstk} we have
\begin{eqnarray}
\label{v_1_decay}
E(\phi; r_{red}, v_1)=\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}
&=&\int\limits_ {\left\lbrace v_1 \leq v \leq v_1+k\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}\nonumber\\
&\leq&
C\mbox{max}\left\lbrace {k,1}\right\rbrace {v_1}^{-2-2\delta}\sim C{v_*}^{-2-2\delta}.
\end{eqnarray}
We thus infer
\begin{eqnarray*}
\int\limits_ {\left\lbrace v_* \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=\tilde{r}} {\mbox{dVol}}_{r=\tilde{r}}
\leq
\tilde{C}\int\limits_ {\left\lbrace v_1 \leq v \leq v_*+1\right\rbrace } J_{\mu}^{-\partial_r}(\phi) n^{\mu}_{r=r_{red}} {\mbox{dVol}}_{r=r_{red}}
\leq
\tilde{C} C{v_*}^{-2-2\delta}.
\end{eqnarray*}
\end{proof}
The above now also implies the following statement.
\begin{cor}
\label{cor6.1}
Let $\phi$ be as in Theorem \ref{anfang}, $r_{blue}$ as in \eqref{rblue} and $r_{red}$ as in Proposition \ref{mi}. Then, for all $v_*>1$ and all $\tilde{u}$ such that $r(\tilde{u}, v_*) \in [r_{blue}, r_{red})$
\begin{eqnarray}
\int\limits_ {\left\lbrace v_{red}(\tilde{u}) \leq v \leq v_{blue}(\tilde{u})\right\rbrace } J^N_\mu(\phi)n^\mu_{u=\tilde{u}}dVol_{u=\tilde{u}} \le C v_*^{-2-2\delta},
\end{eqnarray}
with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$ and $v_{red}(\tilde{u})$, $v_{blue}(\tilde{u})$ are as in \eqref{notation_neu}.
\end{cor}
\begin{proof}
The conclusion of the statement follows by considering the divergence theorem for a triangular region \mbox{$J^-(x)\cap {\mathcal N}$} with \mbox{$x=(\tilde{u}, v_{blue}(\tilde{u}))$}, $x \in J^-(r=r_{blue})$ and using the results of Proposition \ref{r_{red}}. Note that \mbox{$v_*\sim v_{blue}(\tilde{u})\sim v_{red}(\tilde{u})$}.
\end{proof}
By the previous proposition we have successfully propagated the energy estimate further inside the black hole, up to $r=r_{blue}$.
To go even further will be more difficult and we will address this in the next section.
\section{Propagating through ${\mathcal B}$ from $r=r_{blue}$ to the hypersurface $\gamma$}
\label{blueshift1}
In the following we want to propagate the estimates from the $r=r_{blue}$ hypersurface further into the blueshift region to a hypersurface $\gamma$ which is located a logarithmic distance in $v$ from the $r=r_{blue}$ hypersurface, cf.~ Figure \ref{region3_neu}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{region3_neu.jpg}
\caption[Text der im Bilderverzeichnis auftaucht]{Logarithmic distance of hypersurface $r=r_{blue}$ and hypersurface $\gamma$ depicted in a Penrose diagram.}
\label{region3_neu}\end{figure}}
We will define the hypersurface $\gamma$ and its most basic properties in Section \ref{gamma_curve} and propagate the decay bound to $\gamma$ in Section \ref{togamma}.
\subsection{The hypersurface $\gamma$}
\label{gamma_curve}
The idea of the hypersurface $\gamma$ was already entertained in \cite{m_interior} by Dafermos and basically locates $\gamma$ a logarithmic distance in $v$ from a constant $r$ hypersurface living in the blueshift region.
Let $\alpha$ be a fixed constant satisfying
\begin{eqnarray}
\label{alpha}
\alpha>\frac{p+1}{\beta},
\end{eqnarray}
with $\beta$ as in \eqref{lowerboundu} and \eqref{lowerboundv}.
(The significance of the bound \eqref{alpha} will become clear later.) Let us for convenience also assume that
\begin{eqnarray}
\alpha&>&1,
\end{eqnarray}
and
\begin{eqnarray}
\alpha(2-\log 2\alpha) &>&2r^{\star}_{blue}+1.
\end{eqnarray}
We define the function $H(u,v)$ by
\begin{eqnarray}
\label{gross_h}
H(u,v)=u+v-\alpha \log v-2r^{\star}(r_{blue})=u+v-\alpha \log v-2r^{\star}_{blue},
\end{eqnarray}
were $r^{\star}(r_{blue})=r^{\star}_{blue}$ is the $r^{\star}$ value evaluated at $r_{blue}$ according to \eqref{rstar2}, and $r^{\star}_{blue}>0$ according to the choice \eqref{rchoice}.
We then define the hypersurface $\gamma$ as the levelset
\begin{eqnarray}
\label{gammadefine}
\gamma=\left\{H(u,v)=0\right\}\cap \{v> 2\alpha\}.
\end{eqnarray}
Since
\begin{eqnarray}
\label{ableit}
\frac{\partial H}{\partial u}=1, \qquad \frac{\partial H}{\partial v}=1-\frac{\alpha}{v},
\end{eqnarray}
we see that $\gamma$ is a spacelike hypersurface and terminates at $i^+$, cf.~ Appendix \ref{Jcurrents}. (In the notation \eqref{notation_neu}, \mbox{$u_{\gamma}(v)\rightarrow -\infty$} as \mbox{$v\rightarrow \infty$}.)
Note that by our choices $u<-1$ and $v>|u|$ in ${\mathcal D}^+(\gamma)$.
Recall that in Section \ref{rtcoords} we have defined the Regge-Wheeler tortoise coordinate $r^{\star}$ depending on $u$, $v$ by \eqref{regge}. Using this for $r^{\star}_{blue}$ we have
\begin{eqnarray}
r^{\star}(r_{blue})=\frac{v_{blue}(u)+u}{2},
\end{eqnarray}
with $v_{blue}(u)$ as in \eqref{notation_neu}. Plugging this into \eqref{gammadefine} recalling $v_{\gamma}(u)$ defined in \eqref{notation_neu}, we obtain
the relation
\begin{eqnarray}
\label{gamma}
v_{\gamma}(u)-v_{blue}(u)=\alpha \log v_{\gamma}(u).
\end{eqnarray}
As we shall see in Section \ref{finiteness} the above properties of $\gamma$ will allow us to derive pointwise estimates of $\Omega^2$ in $J^+(\gamma)\cap {\mathcal B}$. We first turn however to the region $J^-(\gamma)\cap {\mathcal B}$.
\subsection{Energy estimates from $r=r_{blue}$ to the hypersurface $\gamma$}
\label{togamma}
Now we are ready to propagate the energy estimates further into the blueshift region ${\mathcal B}$ up to the hypersurface $\gamma$.
We will in this part of the proof use the vector field
\begin{eqnarray*}
{S_0}=r^q\partial_{r^{\star}}=r^q(\partial_u+\partial_v),
\end{eqnarray*}
which we have defined in \eqref{Nstar}.
Let us now consider the bulk term and derive positivity properties which are needed later on.
Plugging \eqref{Nstar} in \eqref{Kplug} of Appendix \ref{Kcurrents} yields
\begin{eqnarray}
\label{KNstar}
K^{{S_0}}=&+& qr^{q-1}\left[(\partial_v \phi)^2+(\partial_u \phi)^2\right]\nonumber\\
&-&\left[ \frac{qr^{q-1}}{2}\left[\partial_v r +\partial_u r\right]+{r^q}\left(\frac{\partial_u \Omega}{\Omega}+\frac{ \partial_v \Omega}{\Omega}\right)\right] |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2\nonumber\\
&-&4r^{q-1}(\partial_u \phi\partial_v \phi).
\end{eqnarray}
Our aim is to show that $K^{{S_0}}$ is positive. All terms multiplied by the angular derivatives are manifestly positive in ${\mathcal B}$, cf.~ \eqref{lowerboundu}, \eqref{lowerboundv} together with \eqref{def_l_n} to \eqref{v-neg}.
Therefore, it is only left to show that the first term on the right hand side dominates the last term.
Since by the Cauchy-Schwarz inequality
\begin{eqnarray}
\label{positivityproperty}
2qr^{q-1}(\partial_u \phi\partial_v \phi)\leq qr^{q-1}\left[(\partial_v \phi)^2+(\partial_u \phi)^2\right],
\end{eqnarray}
$K^{{S_0}}$ is positive in ${\mathcal B}$ for all $q \geq 2$.
We show now that at the expense of one polynomial power, we can extend the local energy estimate on $r=r_{blue}$ to an energy estimate along $\gamma$ which is valid for a dyadic length.
\begin{prop}
\label{to_gamma}
Let $\phi$ be as in Theorem \ref{anfang}.
Then, for all $v_*> 2\alpha$
\begin{eqnarray}
\label{sieben}
\int\limits_ {\left\lbrace v_* \leq v \leq 2v_* \right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}
\leq
C{v_*}^{-1-2\delta},
\end{eqnarray}
on the hypersurface $\gamma$, with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{prop}
{\em Remark.} $n^{\mu}_{\gamma}$ denotes the normal vector on the hypersurface $\gamma$ which is a levelset $\gamma=\left\{H(u,v)=0\right\}$ of the function $H(u,v)$ defined in
\eqref{gross_h}. For calculation of $n^{\mu}_{\gamma}$ and $J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma}$ refer to \eqref{ngamma} and \eqref{jgamma} of Appendix \ref{Jcurrents}.
\begin{proof}
In the following we will again make use of notation \eqref{notation_neu}.
Let $v_*> 2\alpha$, such that $\gamma$ is spacelike for $v>v_*$, cf.~ Section \ref{gamma_curve}.
Define $u_3$ by \mbox{$(u_3, v_*) \in \gamma$}, i.e.~ \mbox{$u_{\gamma}(v_*)=u_3$} and define $v_{blue}$ as the intersection of $u_3$ with $r_{blue}$, i.e.~ $v_{blue}(u_3)=v_{blue}$.
And similarly the hypersurfaces $u=u_1$ and $u=u_2$ as shown in Figure \ref{region3_neu} are given by $u_{blue}(2v_*)=u_1$ and $u_{\gamma}(2v_*)=u_2$.
Having defined the relations between all these quantities we can now carry out the divergence theorem for region ${{\mathcal R}}_{III}$:
\begin{eqnarray}
\label{div_region3}
&&\int\limits_ {{{\mathcal R}}_{III}} K^{{S_0}}(\phi) {\mbox{dVol}}
+ \int\limits_ {\left\lbrace v_{blue} \leq v \leq v_{*}\right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}
+ \int\limits_ {\left\lbrace v_* \leq v \leq 2v_*\right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}\nonumber\\
&+& \int\limits_ {\left\lbrace u_1 \leq u \leq u_2 \right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{v=2v_*} {\mbox{dVol}}_{v=2v_*}
= \int\limits_ {\left\lbrace v_{blue} \leq v \leq 2v_*\right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}}.
\end{eqnarray}
Positivity of the flux along the $u=u_3$ segment and the flux along the $v=2v_*$ segment, as well as positivity of $K^{{S_0}}$ for the choice $q\geq 2$, which was derived in \eqref{KNstar} and \eqref{positivityproperty}, leads to
\begin{eqnarray}
\label{r3esti}
\int\limits_ {\left\lbrace v_* \leq v \leq 2v_*\right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}
&\leq& \int\limits_ {\left\lbrace v_{blue} \leq v \leq 2v_*\right\rbrace } J_{\mu}^{{S_0}}(\phi) n^{\mu}_{r=r_{blue}} {\mbox{dVol}}_{r=r_{blue}},\nonumber\\
&{\leq}& C \mbox{max}\left\{2v_*-v_{blue},1 \right\} v_{blue}^{-2-2\delta},\nonumber\\
&\stackrel{\eqref{gamma}}{\leq}& C \left(v_*+\alpha\log{v_*}\right) v_{blue}^{-2-2\delta},\nonumber\\
&{\leq}& \tilde{C}C v_{*}^{-1-2\delta},
\end{eqnarray}
where the second step is implied by Proposition \ref{r_{red}} and the last step follows from the inequality $v_*\leq C v_{blue}$ which is implied by \eqref{gamma}.
\end{proof}
We have already mentioned in the introduction that we will use the vector field ${S}$, cf.~ \eqref{N1} in the region $J^+(\gamma)\cap{\mathcal B}$.
To control the initial flux term of $S$ we require a weighted energy estimate along the hypersurface $\gamma$.
\begin{cor}
\label{cor2}
Let $\phi$ be as in Theorem \ref{anfang}.
Then, for all $v_*>2\alpha$
\begin{eqnarray}
\label{Nstargammaesti}
\int\limits_ {\left\lbrace v_* \leq v <\infty \right\rbrace }v^p J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}
&\leq& Cv_*^{-1-2\delta+p},
\end{eqnarray}
on the hypersurface $\gamma$, with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$ and $p$ as in \eqref{waspist}.
\end{cor}
\begin{proof}
This follows by weighting \eqref{sieben} with $v_*^p$ and summing dyadically.
\end{proof}
Further, we can state the following.
\begin{cor}
\label{cor7.1}
Let $\phi$ be as in Theorem \ref{anfang}, $r_{blue}$ as in \eqref{rblue} and $\gamma$ as in \eqref{gammadefine}. Then, for all $v_*>2\alpha$ and for all $\tilde{u} \in [u_{blue}(v_*), u_{\gamma}(v_*))$
\begin{eqnarray}
\int\limits_ {\left\lbrace v_{blue}(\tilde{u}) \leq v \leq v_{\gamma}(\tilde{u}) \right\rbrace } v^pJ^{S_0}_\mu(\phi)n^\mu_{u=\tilde{u}}dVol_{u=\tilde{u}} \le C v_*^{-1-2\delta+p},
\end{eqnarray}
with $C$ depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$ and $v_{\gamma}(\tilde{u})$, $v_{blue}(\tilde{u})$ as in \eqref{notation_neu}.
\end{cor}
\begin{proof}
The proof is similar to the proof of Corollary \ref{cor6.1} by considering the divergence theorem for a triangular region \mbox{$J^-(x)\cap {\mathcal B}$} with \mbox{$x=(\tilde{u}, v_{\gamma}(\tilde{u}))$}, $x \in J^-(\gamma)$ and using the results of the proof of Proposition \ref{to_gamma}.
\end{proof}
\section{Propagating through ${\mathcal B}$ from the hypersurface $\gamma$ to ${\mathcal C}{\mathcal H}^+$ in the neighbourhood of $i^+$}
\label{innerhorizon}
In order to prove our Theorem \ref{dashier} and close our estimates up to the Cauchy horizon in the neighbourhood of $i^+$ we are interested in considering a region ${{\mathcal R}_{IV}}$ within the trapped region whose boundaries are made up of the hypersurface $\gamma$, a constant $u$ and a constant $v$ segment, which can reach up to the Cauchy horizon, cf.~ Figure \ref{RN_mit_u_v}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{RN_mit_v_u.jpg}
\caption[]{Blueshift region of Reissner-Nordstr\"om spacetime from hypersurface $\gamma$ onwards.}
\label{RN_mit_u_v}\end{figure}}
Let $v_*>2\alpha$ and let $\hat{v}>v_*$.
We may write \mbox{${\mathcal R}_{IV}=J^+(\gamma)\cap J^-(x)$} with
\mbox{$x=(u_{\gamma}(v_*), \hat{v})$}, $x \in J^+(\gamma)\cap{\mathcal B}$.
Note that ${\mathcal R}_{IV}$ lies entirely in the blueshift region, which was characterized by the fact that the quantity $M-\frac{e^2}{r}$ takes the negative sign, cf.~ \eqref{lowerboundu}, \eqref{lowerboundv} and \eqref{def_l_n} to \eqref{v-neg}.
In Section \ref{finiteness} we will derive pointwise estimates for $\Omega^2$ in the future of the hypersurface $\gamma$.
With this estimate, the bulk term will be bounded in terms of the currents through the null hypersurfaces. Consequently, we will be able to absorb the bulk term and to show that the currents through the null hypersurfaces can be bounded by the current along the hypersurface $\gamma$, cf.~ Section \ref{cauchy}.
\subsection{Pointwise estimates on $\Omega^2$ in $J^+(\gamma)$}
\label{finiteness}
In the following we will derive pointwise estimates on $\Omega^2$ in $J^+(\gamma)$. We note that these will imply that the
spacetime volume to the future of the hypersurface $\gamma$ is finite, $\operatorname{Vol}(J^+(\gamma))<C$.
We first derive a future decay bound along a constant $u$ hypersurface for the function $\Omega^2(u,v)$ for $(u, v) \in {\mathcal B}$.
Let \mbox{$x=(u_{fix}, v_{fix})$, $x \in {\mathcal B}$},
then, from \eqref{v-neg} we can immediately see that
\begin{eqnarray}
\label{lambda_comp}
\left.\log\left({\Omega^2(u_{fix},{v})}\right)\right|^{\bar{v}}_{v_{fix}}&=&\int\limits^{\bar{v}}_{v_{fix}} \frac{1}{2r^2}\left(M-\frac{e^2}{r}\right) {\mathrm{d}} {v},\nonumber\\
&\stackrel{\eqref{lowerboundv}}{\leq}&- \beta[\bar{v}-v_{fix}].
\end{eqnarray}
It then immediately follows that
\begin{eqnarray}
\label{omegaufix}
\Omega^2(\bar{u},\bar{v}){\leq} \Omega^2(\bar{u}, v_{fix}) e^{-\beta[\bar{v}-v_{fix}]}, \quad \mbox{for all $(\bar{u}, v_{fix}) \in {\mathcal B}$ and $\bar{v}>v_{fix}$}.
\end{eqnarray}
Analogously, we obtain
\begin{eqnarray}
\label{omegavfix}
\Omega^2(\bar{u},v_{fix}){\leq} \Omega^2(u_{fix}, v_{fix}) e^{-\beta[\bar{u}-u_{fix}]}, \quad \mbox{for all $(\bar{u},v_{fix}) \in J^+(x)$},
\end{eqnarray}
and plugging \eqref{omegaufix} into \eqref{omegavfix} it yields
\begin{eqnarray}
\label{omegafix}
\Omega^2(\bar{u},\bar{v}){\leq} \Omega^2(u_{fix}, v_{fix}) e^{-\beta[\bar{u}-u_{fix}+\bar{v}-v_{fix}]}, \quad \mbox{for all $(\bar{u},\bar{v}) \in J^+(x)$}.
\end{eqnarray}
From \eqref{omegaufix} and \eqref{gamma} we obtain a relation for $\Omega^2(u,v)$ on the hypersurface $\gamma$ as follows
\begin{eqnarray}
\label{lambdalog}
\Omega^2(\bar{u}, \bar{v}) \leq \Omega^2(\bar{u}, v_{blue}(\bar{u})) e^{-\beta\alpha \log v_{\gamma}(\bar{u})}=\Omega^2(\bar{u}, v_{blue}(\bar{u})) {v_{\gamma}(\bar{u})}^{-\beta \alpha},\quad \mbox{for $(\bar{u}, \bar{v}) \in \gamma$}.
\end{eqnarray}
For $J^+(\gamma)$, using \eqref{omegaufix} we further get
\begin{eqnarray}
\label{spacevolumedecay}
\Omega^2(\bar{u}, \bar{v}) \leq C{v_{\gamma}(\bar{u})}^{-\beta \alpha} e^{-\beta\left[\bar{v}-v_{\gamma}(\bar{u})\right]}, \quad \mbox{for $(\bar{u}, \bar{v}) \in J^+(\gamma)$},
\end{eqnarray}
where we have used $\Omega^2(\bar{u}, v_{blue}(\bar{u}))\leq C$.
Moreover, we may think of a parameter $\bar{v}$ which determines the associated $u$ value via intersection with $\gamma$, we denote this value by the evaluation the function $u_{\gamma}(\bar{v})$ which was introduced in \eqref{notation_neu}, cf.~ Figure \ref{integralbild2}.
Moreover, by \eqref{u-neg} we can also state
\begin{eqnarray}
\label{spacevolumedecay_u}
\Omega^2(\bar{u}, \bar{v}) \leq C {|u_{\gamma}(\bar{v})|}^{-\beta \alpha} e^{\beta\left[u_{\gamma}(\bar{v})-\bar{u}\right]} \quad \mbox{for $(\bar{u}, \bar{v}) \in J^+(\gamma)$}.
\end{eqnarray}
Note that the choice \eqref{alpha} of $\alpha$ implies that $\beta \alpha>1$. From \eqref{spacevolumedecay_u}, the fact that
$|u_{\gamma}(\bar{v})| \sim \bar{v}$, and the extra exponential factor, finiteness of the spacetime volume to the future of $\gamma$ follows,
\begin{eqnarray}
\label{finitevol}
\operatorname{Vol}(J^+(\gamma))<C.
\end{eqnarray}
See also \cite{kommemi}.
\subsection{Bounding the bulk term $K^{S}$}
\label{knsec}
To derive energy estimates in ${{\mathcal R}_{IV}}$ we use the timelike vector field multiplier
\begin{eqnarray*}
\label{{S}}
{S}=|u|^p\partial_u+v^p\partial_v,
\end{eqnarray*}
which we have given before in \eqref{N1}.
The weights of ${S}$ are chosen such that they will allow us to derive pointwise estimates from energy estimates; see Section \ref{uni_bounded}.
In order to obtain our desired estimates
first of all we need a bound on the scalar current $K^{S}$, in terms of $J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}}$ and $J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}$.
In the following we will bound the occurring $(u,v)$-dependent weight functions by functions that depend on either $u$ or $v$, respectively.
Plugging the vector field ${S}$, cf.~ \eqref{N1}, into \eqref{Kplug} of Appendix \ref{Kcurrents} we obtain
\begin{eqnarray}
\label{KN}
K^{S}=&-&\frac{2}{r}\left[ v^p +|u|^p\right] (\partial_u \phi \partial_v \phi)\nonumber\\
&-&\left[ \frac{\partial_u \Omega}{\Omega}|u|^p+\frac{ \partial_v \Omega}{\Omega}v^p+\frac{p}{2}(v^{p-1}+|u|^{p-1})\right] |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2.
\end{eqnarray}
Recall \eqref{lowerboundu} and \eqref{lowerboundv}. For large absolute values of $v$ and $u$ the first two terms multiplying the angular derivatives of $\phi$ dominate the last two terms, so in total the term multiplying the angular derivatives is always positive in ${\mathcal D}^+(\gamma)$. Consequently we will be able to use this property to derive an inequality by using the divergence theorem in the proof of Proposition \ref{kastle}.
Let us therefore define
\begin{eqnarray}
\label{KNtilde}
\tilde{K}^{S}=&-&\frac{2}{r}\left[ v^p +|u|^p\right] (\partial_u \phi \partial_v \phi),
\end{eqnarray}
and state
\begin{eqnarray}
\label{knrelation}
-{K}^{S}\leq |\tilde{K}^{S}| \quad \mbox{for $v>\frac{p}{2\beta}$ and $|u|>\frac{p}{2\beta}$, cf.~ \eqref{lowerboundv}, \eqref{lowerboundu}}.
\end{eqnarray}
(Note that $\tilde{K}^{S}$ coincides with the bulk term for spherically symmetric $\phi$.) We have the following
\begin{lem}
\label{K_spher}
Let $\phi$ be an arbitrary function. Then, for all $v_*>2\alpha$
and all $\hat{v}>v_*$,
the integral over region \mbox{${\mathcal R}_{IV}=J^+(\gamma)\cap J^-(x)$} with \mbox{$x=(u_{\gamma}(v_*), \hat{v})$}, $x \in {\mathcal B}$, cf.~ Figure \ref{RN_mit_u_v}, of the current $\tilde{K}^{S}$, defined by \eqref{KNtilde}, can be estimated by
\begin{eqnarray}
\label{Prop6.1}
\int\limits_{{\mathcal R}_{IV}} |\tilde{K}^{S}| {\mbox{dVol}} \leq &\delta_1& \sup_{u_{\gamma}(\hat{v})\leq \bar{u}\leq u_{\gamma}(v_*)}\int\limits_{\left\lbrace v_{\gamma}(\bar{u}) \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}\nonumber \\
+ &\delta_2& \sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_{\left\lbrace u_{\gamma}(\hat{v}) \leq u \leq u_{\gamma}(\bar{v}) \right\rbrace }
J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}},
\end{eqnarray}
where $\delta_1$ and $\delta_2$ are positive constants, with $\delta_1\rightarrow 0$ and $\delta_2\rightarrow 0$ as $v_*\rightarrow \infty$.
\end{lem}
{\em Remark.} In the proof of Proposition \ref{kastle} we will see that the above proposition determines $u_{\text{\ding{34}}}$ of Theorem \ref{dashier}, depicted in Figure \ref{bnr}. We have to choose $u_{\text{\ding{34}}}=u_{\gamma}(v_*)$, with $v_*$ such that $\delta_1$ is small.
\begin{proof}
Using the Cauchy-Schwarz inequality twice for the remaining part of the bulk term we obtain
\begin{eqnarray}
|\tilde{K}^{S}|&\leq&\frac{1}{r}\left[ \left(1 +\frac{|u|^p}{v^p}\right){v^p}(\partial_v \phi)^2+ \left(1 +\frac{v^p}{|u|^p}\right){|u|^p}(\partial_u \phi)^2\right],
\end{eqnarray}
with the related volume element
\begin{eqnarray}
{\mbox{dVol}}&=&r^2{\frac{\Omega^2}{2}} {\mathrm{d}} u{\mathrm{d}} v{\mathrm{d}} \sigma^2_{{\mathbb S}}.
\end{eqnarray}
Note that the currents related to the vector field $S$ with their related volume elements are given by
\begin{eqnarray}
J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} &=&{\frac{2}{\Omega^2}}\left[|u|^p(\partial_u \phi)^2+ \frac{\Omega^2}{4}\bar{v}^p|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2\right], \quad {\mbox{dVol}}_{v=\bar{v}}=r^2{\frac{\Omega^2}{2}} {\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} u,\\
J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}} &=&{\frac{2}{\Omega^2}}\left[v^p(\partial_v \phi)^2+ \frac{\Omega^2}{4}|\bar{u}|^p|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2\right],\quad {\mbox{dVol}}_{u=\bar{u}}=r^2{\frac{\Omega^2}{2}} {\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} v,
\end{eqnarray}
cf.~ Appendix \ref{Jcurrents}.
Taking the integral over the spacetime region yields
\begin{eqnarray}
\label{knboundhier}
\int\limits_ {{\mathcal R}_{IV}} |\tilde{K}^{S}(\phi)| {\mbox{dVol}} &\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\bar{v})}\int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace }\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1 +\frac{|{\bar{u}}|^p}{{\bar{v}}^p}\right)
J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}} {\mathrm{d}} \bar{u}\nonumber \\
&&+ \int\limits_{\bar{v}}^{\hat{v}}\int\limits_{\left\lbrace u_{\gamma}(\hat{v}) \leq u \leq u_{\gamma}(v_*)\right\rbrace } \frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left(1 +\frac{{\bar{v}}^p}{|{\bar{u}}|^p}\right) J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}}{\mbox{dVol}}_{v=\bar{v}}{\mathrm{d}} \bar{v},
\end{eqnarray}
with $u_{\gamma}(v)$ in the integration limits as defined in \eqref{notation_neu}.
Note the following general relation for positive functions $f(\bar{u},\bar{v})$ and $g(\bar{u},\bar{v})$
\begin{eqnarray}
\label{uintegral}
\int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})}\int\limits_{\bar{v}}^{\hat{v}} f(\bar{u},\bar{v})g(\bar{u},\bar{v}){\mathrm{d}} \bar{v}{\mathrm{d}} \bar{u}
&\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})}\int\limits_{\bar{v}}^{\hat{v}} \left[\sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}f(\bar{u},\bar{v})\right]g(\bar{u},\bar{v}){\mathrm{d}} \bar{v}{\mathrm{d}} \bar{u}\nonumber\\
&\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})}\left[ \sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}f(\bar{u},\bar{v})\right]\int\limits_{\bar{v}}^{\hat{v}}g(\bar{u},\bar{v}){\mathrm{d}} \bar{v}{\mathrm{d}} \bar{u}\nonumber\\
&\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})} \sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}f(\bar{u},\bar{v}){\mathrm{d}} \bar{u}\sup_{u_{\gamma}(\hat{v})\leq \bar{u}\leq u_{\gamma}(v_*)}\int\limits_{\bar{v}}^{\hat{v}}g(\bar{u},\bar{v}){\mathrm{d}} \bar{v}\nonumber\\
&\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})} \sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}f(\bar{u},\bar{v}){\mathrm{d}} \bar{u}\sup_{u_{\gamma}(\hat{v})\leq \bar{u}\leq u_{\gamma}(v_*)}\int\limits_{v_*}^{\hat{v}}g(\bar{u},\bar{v}){\mathrm{d}} \bar{v}.
\end{eqnarray}
Similarly, it immediately follows that
\begin{eqnarray}
\label{vintegral}
\int\limits^{\hat{v}}_{v_*}\int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\bar{v})} f(\bar{u},\bar{v})g(\bar{u},\bar{v}){\mathrm{d}} \bar{u}{\mathrm{d}} \bar{v}
&\leq&\int\limits^{\hat{v}}_{v_*}\sup_{u_{\gamma}(\bar{v}) \leq \bar{u}\leq u_{\gamma}(v_*)}f(\bar{u},\bar{v}){\mathrm{d}} \bar{v}\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\bar{v})}g(\bar{u},\bar{v}){\mathrm{d}} \bar{u}.
\end{eqnarray}
Using \eqref{uintegral} and \eqref{vintegral} in \eqref{knboundhier} we obtain
\begin{eqnarray}
\label{131}
\int\limits_ {{\mathcal R}_{IV}} |\tilde{K}^{S}(\phi)| {\mbox{dVol}}&\leq& \int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})} \sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ \frac{|{\bar{u}}|^p}{{\bar{v}}^p}\right)\right]{\mathrm{d}} \bar{u} \sup_{u_{\gamma}(\hat{v})\leq \bar{u}\leq u_{\gamma}(v_*)}\int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace }J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
\nonumber\\
&&+ \int\limits^{\hat{v}}_{v_*}\sup_{u_{\gamma}(\bar{v}) \leq \bar{u}\leq u_{\gamma}(v_*)}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{{\bar{v}}^p}{|{\bar{u}}|^p}\right) \right]{\mathrm{d}} \bar{v}\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_{\left\lbrace u_{\gamma}(\bar{v}) \leq u \leq u_{\gamma}(v_*)\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}}{\mbox{dVol}}_{v=\bar{v}}.\nonumber\\
&&
\end{eqnarray}
It remains to show finiteness and smallness of \mbox{$\int\limits^{u_{\gamma}(v_*)}_{u_{\gamma}(\hat{v})} \sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ \frac{|{\bar{u}}|^p}{{\bar{v}}^p}\right)\right]{\mathrm{d}} \bar{u} $} and \mbox{$\int\limits^{\hat{v}}_{v_{\gamma}(\bar{u})}\sup_{u_{\gamma}(\bar{v}) \leq \bar{u}\leq u_{\gamma}(v_*)}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{{\bar{v}}^p}{|{\bar{u}}|^p}\right) \right]{\mathrm{d}} \bar{v}$}.
Earlier we obtained the relation \eqref{spacevolumedecay_u} for $\Omega^2$ in region ${\mathcal R}_{IV}$. Therefore, we can write
\begin{eqnarray}
\int\limits_{u_{\gamma}(\hat{v})}^{u_{\gamma}(v_*)}
\sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{|{\bar{u}}|^p}{{\bar{v}}^p}\right) \right]{\mathrm{d}} \bar{u}
&\leq& C \int\limits_{u_{\gamma}(\hat{v})}^{u_{\gamma}(v_*)}\sup_{v_{\gamma}(\bar{u})\leq \bar{v} \leq \hat{v}}\left[{|u_{\gamma}(\bar{v})|}^{-\beta \alpha} e^{\beta\left[u_{\gamma}(\bar{v})-\bar{u}\right]}\left( 1 + \frac{|{\bar{u}}|^p}{{\bar{v}}^p}\right)\right] {\mathrm{d}} \bar{u}\nonumber\\
&\leq& \tilde{C} \int\limits_{u_{\gamma}(\hat{v})}^{u_{\gamma}(v_*)}{|\bar{u}|}^{-\beta \alpha} \left( 1 + \frac{{|\bar{u}|}^p}{v_*^p}\right) {\mathrm{d}} \bar{u}\nonumber\\
\label{integral}
&\leq& \tilde{\tilde{C}} \int\limits_{u_{\gamma}(\hat{v})}^{u_{\gamma}(v_*)}{|\bar{u}|}^{-\beta \alpha+p}{\mathrm{d}} \bar{u}\nonumber\\
&\leq& \frac{\tilde{\tilde{C}}}{|-\beta\alpha+p+1|}\left[{|\bar{u}|}^{-\beta\alpha+p+1}\right]_{u_{\gamma}(\hat{v})}^{u_{\gamma}(v_*)}\nonumber\\
&\leq& \delta_1,
\end{eqnarray}
where $\delta_1\rightarrow 0$ for $|u_{\gamma}(v_*)|\rightarrow -\infty$ and thus for $v_*\rightarrow \infty$. Note that we have here used \eqref{alpha}.
For finiteness of the second term in \eqref{131} we follow the same strategy and use \eqref{spacevolumedecay} for the second term to obtain
\begin{eqnarray}
\int\limits^{\hat{v}}_{v_*}\sup_{u_{\gamma}(\bar{v}) \leq \bar{u}\leq u_{\gamma}(v_*)}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{{\bar{v}}^p}{|{\bar{u}}|^p}\right) \right]{\mathrm{d}} \bar{v}
&\leq& C \int\limits^{\hat{v}}_{v_*}\sup_{u_{\gamma}(\bar{v}) \leq \bar{u}\leq u_{\gamma}(v_*)}
\left[\bar{v}^{-\beta\alpha}\left( 1 + \frac{{\bar{v}}^p}{|{\bar{u}}|^p}\right)\right]
{\mathrm{d}} \bar{v}\nonumber\\
&\leq& \tilde{C}\int\limits^{\hat{v}}_{v_*}{\bar{v}}^{-\beta\alpha}\left( 1 + \frac{{\bar{v}}^p}{|u_{\gamma}(v_*)|^p}\right){\mathrm{d}} \bar{v}\nonumber\\
&\leq& \frac{\tilde{\tilde{C}}}{|-\beta\alpha+p+1|}\left[{|\bar{v}|}^{-\beta\alpha+p+1}\right]^{\hat{v}}_{v_*}\nonumber\\
&\leq& \delta_2,
\end{eqnarray}
where $\delta_2 \rightarrow 0$ for $v_* \rightarrow \infty$.
Therefore, we obtain the statement of Lemma \ref{K_spher}.
\end{proof}
\subsection{Energy estimates from $\gamma$ up to ${\mathcal C}{\mathcal H}^+$ in the neighbourhood of $i^+$}
\label{cauchy}
Now we come to the actual proof of weighted energy boundedness up to the Cauchy horizon.
\begin{prop}
\label{kastle}
Let $\phi$ be as in Theorem \ref{anfang} and $p$ as in \eqref{waspist}. Then, for $u_{\text{\ding{34}}}$ sufficiently close to $-\infty$, for all $v_*\geq v_{\gamma}(u_{\text{\ding{34}}})$ and $\hat{v}>v_*$
\begin{eqnarray}
\label{propgamma}
\int\limits_ {\left\lbrace u_{\gamma}(\hat{v})\leq u \leq u_{\gamma}(v_*)\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
&+&\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=u_{\gamma}(v_*)} {\mbox{dVol}}_{u=u_{\gamma}(v_*)}\leq C {v_*}^{-1-2\delta+p},
\end{eqnarray}
where $C$ is a positive constant depending on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond}, 1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{prop}
{\em Remark.} Refer to \eqref{notation_neu} for the definition of $u_{\gamma}(v)$ and see Figure \ref{integralbild2} for further clarification.
\begin{proof}
In Section \ref{blueshift1}, Corollary \ref{cor2}, we have obtained the global estimate \eqref{Nstargammaesti} for the weighted ${S_0}$ current which follows from Proposition \ref{to_gamma}. Recall that in ${\mathcal D}^+(\gamma)$ we have $|u|^p\leq v^p$, cf.~ Section \ref{gamma_curve}, which immediately leads to
\begin{eqnarray}
\label{vergleichN_Nstar}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v} \right\rbrace } J_{\mu}^{{S}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}\leq \tilde{C}\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v} \right\rbrace }v^p J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma},
\end{eqnarray}
cf.~ Appendix \ref{Jcurrents} for explicit expressions of $J_{\mu}^{{S_0}}(\phi) n^{\mu}_{\gamma}$ and $J_{\mu}^{{S}}(\phi) n^{\mu}_{\gamma} $.
From \eqref{vergleichN_Nstar} we see that
\begin{eqnarray}
\label{hypo}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v} \right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma} \leq C {v_*}^{-1-2\delta+p} \quad \mbox{for all $v_*>\alpha$ and $p$ as in \eqref{waspist},}
\end{eqnarray}
is implied by Corollary \ref{cor2}.
Let $v_*>2\alpha$ and $\hat{v}>v_*$.
In order to obtain \eqref{propgamma} we consider a region \mbox{${\mathcal R}_{IV}=J^+(\gamma)\cap J^-(x)$} with \mbox{$x=(u_{\gamma}(v_*), \hat{v})$}, $x \in {\mathcal B}$, as shown in
Figure \ref{RN_mit_u_v}. Applying the divergence theorem we obtain
\begin{eqnarray}
&&\int\limits_ {{\mathcal R}_{IV}} K^{S}(\phi) {\mbox{dVol}}
+\int\limits_ {\left\lbrace u_{\gamma}(\hat{v})\leq u \leq u_{\gamma}(v_*)\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=u_{\gamma}(v_*)} {\mbox{dVol}}_{u=u_{\gamma}(v_*)}\nonumber\\
&=&
\int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}.
\end{eqnarray}
In Section \ref{knsec} we found that the angular part of $K^{S}(\phi)$ is positive in ${\mathcal R}_{IV}$ and we called the remaining part $\tilde{K}^{S}(\phi) $ given in \eqref{KNtilde}. Using \eqref{knrelation} we can therefore write
\begin{eqnarray}
\label{cons_law}
&&\int\limits_ {\left\lbrace u_{\gamma}(\hat{v})\leq u \leq u_{\gamma}(v_*)\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=u_{\gamma}(v_*)} {\mbox{dVol}}_{u=u_{\gamma}(v_*)}\nonumber\\
&\leq& \int\limits_ {{\mathcal R}_{IV}} |\tilde{K}^{S}(\phi)| {\mbox{dVol}}
+ \int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}.
\end{eqnarray}
Using Lemma \ref{K_spher} we obtain
\begin{eqnarray}
\label{start}
&&\int\limits_ {\left\lbrace u_{\gamma}(\hat{v})\leq u \leq u_{\gamma}(v_*)\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=u_{\gamma}(v_*)} {\mbox{dVol}}_{u=u_{\gamma}(v_*)}\nonumber\\
&\leq&\delta_1 \sup_{ u_{\gamma}(\hat{v})\leq \bar{u} \leq u_{\gamma}(v_*)}\int\limits_ {\left\lbrace v_{\gamma}(\bar{u}) \leq v \leq \hat{v} \right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}+ \delta_2 \sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{\gamma}(\hat{v}) \leq u \leq u_{\gamma}(\bar{v}) \right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&&+ \int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}.
\end{eqnarray}
Repeating estimate \eqref{start} with $\bar{u}$, $\bar{v}$ in place of $u_{\gamma}(v_*)$, $\hat{v}$ and taking the supremum we have
\begin{eqnarray}
\label{140}
&&\sup_{u_{\gamma}(\hat{v})\leq \bar{u}\leq u_{\gamma}(v_*)}\int\limits_ {\left\lbrace v_{\gamma}(\bar{u}) \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
+ \sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{\gamma}(\hat{v})\leq u \leq u_{\gamma}(\bar{v})\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&\leq&\delta_1 \sup_{ u_{\gamma}(\hat{v})\leq \bar{u} \leq u_{\gamma}(v_*)}\int\limits_ {\left\lbrace v_{\gamma}(\bar{u}) \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}} + \delta_2 \sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{\gamma}(\hat{v}) \leq u \leq u_{\gamma}(\bar{v}) \right\rbrace }J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&&+ \int\limits_{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{\gamma} {\mbox{dVol}}_{\gamma}.
\end{eqnarray}
Recalling $\delta_1\rightarrow 0$, $\delta_2\rightarrow 0$ as $v_*\rightarrow \infty$, choose $u_{\text{\ding{34}}}$ sufficiently close to $-\infty$, such that for $v_*>v_{\gamma}(u_{\text{\ding{34}}})$, say
\begin{eqnarray}
\delta_1, \delta_2 \leq \frac12
\end{eqnarray}
holds.
The conclusion of Proposition \ref{kastle} then follows by absorbing the first two terms of the right hand side of \eqref{140} by the two terms on the left and estimating the third from \eqref{hypo}.
\end{proof}
\subsection{Energy estimates globally in the rectangle $\Xi$ up to ${\mathcal C}{\mathcal H}^+$ in the neighbourhood of $i^+$}
\label{allxi}
In the previous Sections \ref{first_section} to \ref{cauchy} we have proven energy estimates for each region with specific properties separately. Putting all results together we can state the following proposition.
\begin{prop}
\label{gesamtesti}
Let $\phi$ be as in Theorem \ref{anfang} and $p$ as in \eqref{waspist}. Then, for $u_{\text{\ding{34}}}$ sufficiently close to $-\infty$, for all $v_*>1$, $\hat{v}>v_*$ and $\tilde{u} \in (-\infty, u_{\text{\ding{34}}})$.
\begin{eqnarray}
\label{propgamma2}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}\leq {C} {v_*}^{-1-2\delta+p},
\end{eqnarray}
where ${C}$ is a positive constant depending on $C_0$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond},1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{prop}
\begin{proof}
First of all we partition the integral of the statement into a sum of integrals of the different regions. That is to say
\begin{eqnarray*}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}&=&\int\limits_ {{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } \cap {\mathcal R}} J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}
+\int\limits_ {{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } \cap {\mathcal N}} J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}\\
&&+\int\limits_ {{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } \cap J^-(\gamma)\cap{\mathcal B}} J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}
+\int\limits_ {{\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } \cap J^+(\gamma)\cap{\mathcal B}} J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}.
\end{eqnarray*}
For the integral in ${\mathcal R}$ and the integral in ${\mathcal N}$ we use Corollaries \ref{cor5.2} and \ref{cor6.1}. (Note that the former has to be summed resulting in the loss of one polynomial power.)
Further, for the integral in region $J^-(\gamma)\cap{\mathcal B}$ we apply Corollary \ref{cor7.1} and for the integral in region $J^+(\gamma)\cap{\mathcal B}$ we use Proposition \ref{kastle}.
Putting all this together we arrive at the conclusion of Proposition \ref{gesamtesti}.
\end{proof}
In particular, we have
\begin{cor}
\label{endeacht}
Let $\phi$ be as in Theorem \ref{anfang} and $p$ as in \eqref{waspist}. Then, for $u_{\text{\ding{34}}}$ sufficiently close to $-\infty$, for all \mbox{$v_{fix}\geq1$}, and $\tilde{u} \in (-\infty, u_{\text{\ding{34}}})$,
\begin{eqnarray}
\label{propgamma2}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ v^p (\partial_v \phi)^2(\tilde{u}, v) + |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(\tilde{u}, v) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq C,
\end{eqnarray}
where $C$ is a positive constant dependent on $C_{0}$ of Theorem \ref{anfang} and $D_{0}(u_{\diamond},1)$ of Proposition \ref{initialdataprop}, where $u_{\diamond}$ is defined by $r_{red}=r(u_{\diamond},1)$.
\end{cor}
\begin{proof}
The conclusion of the proposition follows immediately examining the weights in Proposition \ref{gesamtesti}.
\end{proof}
\section{Pointwise estimates from higher order energies}
\label{nineten}
\subsection{The $\leo$ notation and Sobolev inequality on spheres}
\label{leonotationsec}
Recall that we had stated the expressions for the generators of spherical symmetry $\leo_i$, $i=1,2,3$, in Section \ref{angular}. They were explicitly given by \eqref{angmom1} to \eqref{angmom2}. Further, having expressions \eqref{sum1} and \eqref{sum2} in mind we introduce the following notation
\begin{eqnarray}
\label{leonotation}
\sum_{k=0}^{2} \left(\leo^k \phi\right)^2=|\phi|^2+\sum_{i=1}^{3} \left(\leo_i \phi\right)^2+\sum_{i=1}^{3} \sum_{j=1}^{3}\left(\leo_i \leo_j\phi\right)^2,
\end{eqnarray}
where $k$ has to be understood as the order of an exponent and not as an index.
By Sobolev embedding on the standard spheres we have in this notation
\begin{eqnarray}
\label{sobo_embed}
\sup_{\theta,\varphi \in {\mathbb S}^2}|\phi(u,v,\theta,\varphi)|^2\leq \tilde{C} \sum_{k=0}^{2} \int\limits_{{\mathbb S}^2} \left(\leo^k \phi\right)^2(u,v,\theta,\varphi){\mathrm{d}} \sigma_{\mathbb S^2},
\end{eqnarray}
which means that we can derive a pointwise estimate from an estimate of the integrals on the spheres, see e.g.~ \cite{m_red}.
More generally, in the following we will also use the notation
\begin{eqnarray}
J_\mu^X(\leo\phi) = \sum_{i=1}^{3} J_\mu^X(\leo_i\phi),
\end{eqnarray}
for any $J$-current related to an arbitrary vector field $X$, and similarly for other quadratic expressions, e.g.~ \eqref{energy3_hier}, \eqref{energy3_hier2}.
\subsection{Higher order energy estimates in the neighbourhood of $i^+$}
\label{higherorder_neu}
We will need the following extension of Corollary \ref{endeacht} for higher order energies.
\begin{thm}
\label{energythm2}
On subextremal Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, let $\phi$ be
a solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on a two-ended asymptotically flat Cauchy surface $\Sigma$. Then, for $v_{fix} \geq 1$ and $u_{fix} > -\infty$
\begin{eqnarray}
\label{energy2}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[v^p (\partial_v \phi)^2(u_{fix}, v, \theta, \varphi) + |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq E_{0},\\
\label{energy3_hier}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}} \left[v^p (\partial_v \leo\phi)^2(u_{fix}, v, \theta, \varphi) + |\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq E_{1},\\
\label{energy3_hier2}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ v^p (\partial_v \leo^2\phi)^2 (u_{fix}, v, \theta, \varphi)+ |\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq E_{2},
\end{eqnarray}
where $p$ is as in \eqref{waspist}.
\end{thm}
\begin{proof}
Statement \eqref{energy2} was already derived in Corollary \ref{endeacht}.
Recall that $\leo_i\phi$, $\leo_i\leo_j\phi$ also satisfy the massless scalar wave equation, cf.~ Section \ref{en_cur}. Summing over all angular momentum operators, keeping in mind notation \eqref{leonotation}, etc., we therefore obtain \eqref{energy3_hier} and \eqref{energy3_hier2}.
\end{proof}
\subsection{Pointwise boundedness in the neighbourhood of $i^+$}
\label{uni_bounded}
We turn the discussion to the derivation of pointwise boundedness from energy estimates. In particular we prove Theorem \ref{dashier} from Theorem \ref{energythm2}.
By the fundamental theorem of calculus it follows for all $v_*>1$, $\hat{v} >v_*$ and $u \in (-\infty, u_{\text{\ding{34}}})$ that
\begin{eqnarray}
\phi(u,\hat{v}, \theta, \varphi)&=& \int\limits_{v_*}^{\hat{v}} \left(\partial_v \phi\right)(u, v, \theta, \varphi) {\mathrm{d}} v +\phi(u, v_*, \theta, \varphi),\nonumber\\
&\leq& \int\limits_{v_*}^{\hat{v}} (\partial_v \phi)(u, v, \theta, \varphi)v^{\frac{p}{2}}v^{-\frac{p}{2}}{\mathrm{d}} v+\phi(u, v_*, \theta, \varphi),\nonumber\\
&\leq& \left(\int\limits_{v_*}^{\hat{v}} v^p(\partial_v \phi)^2(u, v, \theta, \varphi){\mathrm{d}} v\right)^{\frac12}\left(\int\limits_{v_*}^{\hat{v}} v^{-{p}}{\mathrm{d}} v\right)^{\frac12}+\phi(u, v_*, \theta, \varphi),\nonumber\\
\end{eqnarray}
where we have used the Cauchy-Schwarz inequality in the last step.
Squaring the entire expression, using Cauchy-Schwarz again and integrating over ${\mathbb S}^2$ we obtain the expression that we had sketched in Section \ref{outline} already
\begin{eqnarray}
\int\limits_{{\mathbb S}^2} \phi^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq& \tilde{C}\left[\int\limits_{{\mathbb S}^2}\left(\int\limits_{v_*}^{\hat{v}} v^p(\partial_v \phi)^2(u,v){\mathrm{d}} v\int\limits_{v_*}^{\hat{v}} v^{-{p}}{\mathrm{d}} v\right)r^2{\mathrm{d}} \sigma_{\mathbb S^2}+\int\limits_{{\mathbb S}^2} \phi^2(u,v_*){\mathrm{d}} \sigma_{\mathbb S^2}\right],
\end{eqnarray}
with $p$ as in \eqref{waspist} and the first term on the right hand side controlled by the flux for which we derived boundedness in Section \ref{innerhorizon}.
Therefore, by using Theorem \ref{energythm2} we obtain
\begin{eqnarray}
\label{fundcauchy}
\int\limits_{{\mathbb S}^2} \phi^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2}&\leq&\tilde{C}\left[ E_{0}\int\limits_{{\mathbb S}^2}\int\limits_{v_*}^{\hat{v}} v^{-p}{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}+\int\limits_{{\mathbb S}^2} \phi^2(u,v_*){\mathrm{d}} \sigma_{\mathbb S^2}\right]\nonumber\\
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{0}+\int\limits_{{\mathbb S}^2} \phi^2(u,v_*){\mathrm{d}} \sigma_{\mathbb S^2}\right].
\end{eqnarray}
It is here that we have used the requirement $p>1$ of \eqref{waspist}.
Applying all our estimates to $\leo_i \phi$, $\leo_i\leo_j \phi$ and summing, we obtain in the notation of Section \ref{leonotationsec} the following:
\begin{eqnarray}
\label{fundcauchy1}
\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{1}+\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(u,v_*){\mathrm{d}} \sigma_{\mathbb S^2}\right],\\
\label{fundcauchy2}
\int\limits_{{\mathbb S}^2} ({\leo}^2 \phi)^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{2}+\int\limits_{{\mathbb S}^2} ({\leo}^2 \phi) ^2(u,v_*){\mathrm{d}} \sigma_{\mathbb S^2}\right].
\end{eqnarray}
Let us now use \eqref{aufeins1} to \eqref{aufeins3} of Proposition \ref{initialdataprop} to estimate the right hand sides of \eqref{fundcauchy} to \eqref{fundcauchy2} with $v_*=1$.
Adding all equations up we derive pointwise boundedness according to \eqref{sobo_embed}
\begin{eqnarray}
\label{supr}
\sup_{{\mathbb S}^2}|\phi(u,\hat{v},\theta,\varphi)|^2&\leq& \tilde{C} \left[\int\limits_{{\mathbb S}^2} ( \phi) ^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2} +\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2} +\int\limits_{{\mathbb S}^2} (\leo^2 \phi) ^2(u,\hat{v}){\mathrm{d}} \sigma_{\mathbb S^2} \right],\\
\label{supr2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}}\left( E_{0}+ E_{1}+ E_{2}\right)+D_{0}(u_{\diamond}, 1)+D_{1}(u_{\diamond}, 1)+D_{2}(u_{\diamond}, 1)\right],\\
\label{supr3}
&\leq&C,
\end{eqnarray}
with $C$ depending on the initial data on $\Sigma$.
We therefore arrive at the statement given in Theorem \ref{dashier}.
\begin{trivlist}
\item[\hskip \labelsep ]\qed\end{trivlist}
\section{{\it Left} neighbourhood of $i^+$}
\label{leftinterior}
We now turn to establish boundedness in the neighbourhood of the {\em left} timelike infinity $i^+$.
For this we simply repeat the entire proof carried out in the characteristic rectangle $\Xi$ in Section \ref{first_section} to Section \ref{innerhorizon} but this time for the region $\tilde{\Xi}$ at the {\it left end} of ${\mathcal Q}|_{II}$, cf.~ \eqref{QII}, as shown in the Penrose diagram \ref{RN_character_linkes}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{RN_character_links.jpg}
\caption[]{Penrose diagram with characteristic rectangle $\tilde{\Xi}$ depicted on the {\it left} side.}
\label{RN_character_linkes}\end{figure}}
Since everything is completely analogous to the derivation for the {\it right} side, we will merely state the main Theorems and Propositions and remind the reader that $u$ and $v$ are interchanged now. Recall that $u\rightarrow \infty$ on ${\mathcal C}{\mathcal H}_B^+$ and $v\rightarrow -\infty$ on ${\mathcal H}_B^+$ for the {\it left end}.
The rectangle under consideration is now $\tilde{\Xi}=\left\{(1\leq u< \infty), (-\infty\leq v\leq v_{\text{\ding{34}}}) \right\}$
\begin{thm}
\label{anfangl}
Let $\phi$ be a solution of the wave equation \eqref{wave} on a subextremal Reissner-Nordstr\"om background $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, arising from smooth compactly supported initial data on an arbitrary Cauchy hypersurface $\Sigma$, cf.~ Figure \ref{RN_character_linkes}. Then, there exists $\delta>0$ such that
\begin{eqnarray}
\label{thpurl}
\int\limits_{{\mathbb S}^2}\int\limits^{u+1}_u \left[(\partial_u \phi)^2(u, -\infty)+|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u, -\infty)\right]r^2 {\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {C_{0}}u^{-2-2\delta},\\
\label{thpur1l}
\int\limits_{{\mathbb S}^2}\int\limits^{u+1}_u \left[(\partial_u \leo\phi)^2(u, -\infty) +|\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(u, -\infty)\right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {C_{1}}u^{-2-2\delta},\\
\label{thpur2l}
\int\limits_{{\mathbb S}^2}\int\limits^{u+1}_u \left[(\partial_u \leo^2\phi)^2(u, -\infty) +|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u, -\infty)\right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {C_{2}}u^{-2-2\delta},
\end{eqnarray}
on ${{\mathcal H}_B}^+$, for all $u$ and some positive constants ${C_{0}}$, ${C_{1}}$ and ${C_{2}}$ depending on the initial data.
\end{thm}
\begin{prop}
\label{initialdatapropl}
Let \mbox{$u_{\diamond}, v_{\diamond} \in (-\infty, \infty)$}.
Under the assumption of Theorem \ref{anfangl}, the energy at retarded Eddington-Finkelstein coordinate \mbox{$\left\{u=u_{\diamond}\right\}\cap\left\{{-\infty}\leq v \leq {v_{\diamond}}\right\}$} is bounded from the initial data
\begin{eqnarray}
\label{proppurl}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{v_{\diamond}}\left[ \Omega^{-2}(\partial_v \phi)^2(u_{\diamond},v)+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u_{\diamond},v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{0}(u_{\diamond}, v_{\diamond})},\\
\label{proppur1l}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{v_{\diamond}}\left[\Omega^{-2}(\partial_v \leo\phi)^2(u_{\diamond},v)+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(u_{\diamond},v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{1}(u_{\diamond}, v_{\diamond})},\\
\label{proppur2l}
\int\limits_{{\mathbb S}^2}\int\limits_{-\infty}^{v_{\diamond}}\left[\Omega^{-2}(\partial_v \leo^2\phi)^2(u_{\diamond},v)+\frac{\Omega^{2}}{2}|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u_{\diamond},v)\right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{2}(u_{\diamond}, v_{\diamond})},
\end{eqnarray}
and further
\begin{eqnarray}
\label{aufeins1l}
\sup_{-\infty\leq v \leq {v_{\diamond}}}\int\limits_{{\mathbb S}^2} (\phi)^2(u_{\diamond},v){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{0}(u_{\diamond}, v_{\diamond})},\\
\sup_{-\infty\leq v \leq {v_{\diamond}}}\int\limits_{{\mathbb S}^2} (\leo\phi)^2(u_{\diamond},v){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{1}(u_{\diamond}, v_{\diamond})},\\
\label{aufeins3l}
\sup_{-\infty\leq v \leq {v_{\diamond}}}\int\limits_{{\mathbb S}^2} (\leo^2\phi)^2(u_{\diamond},v){\mathrm{d}} \sigma_{\mathbb S^2}&\leq& {D_{2}(u_{\diamond}, v_{\diamond})},
\end{eqnarray}
with ${D_{0}(u_{\diamond}, v_{\diamond})}$, $D_{1}(u_{\diamond}, v_{\diamond})$ and $D_{2}(u_{\diamond}, v_{\diamond})$ positive constants depending on the initial data.
\end{prop}
Note the $\Omega^{-2}$ weights which arise since $v$ is not regular at ${\mathcal H}^+_B$.
Analogous to the Proposition \ref{gesamtesti} obtained for the {\it right} side we can state the following for the {\it left}.
\begin{prop}
\label{gesamtestil}
Let $\phi$ be as in Theorem \ref{anfangl} and $p$ as in \eqref{waspist}. Then, for $v_{\text{\ding{34}}}$ sufficiently close to $-\infty$, for $u_*>1$, $\hat{u}>u_*$ and $\tilde{v} \in (-\infty, v_{\text{\ding{34}}})$.
\begin{eqnarray}
\label{propgamma2l}
\int\limits_ {\left\lbrace u_* \leq u \leq \hat{u}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\tilde{v}} {\mbox{dVol}}_{v=\tilde{v}}\leq {C} {u_*}^{-1-2\delta+p},
\end{eqnarray}
where ${C}$ is a positive constant depending on $C_0$ of Theorem \ref{anfangl} and $D_{0}(u_{\diamond},1)$ of Proposition \ref{initialdatapropl}, where $v_{\diamond}$ is defined by $r_{red}=r(1, v_{\diamond})$.
\end{prop}
\begin{proof}
The proof is analogous to the proof of Proposition \ref{gesamtesti} with $u$ and $v$ interchanged.
\end{proof}
Having obtained Proposition \eqref{gesamtestil} we can derive higher order estimates analogous to Section \ref{higherorder_neu}. The pointwise estimate is then obtained via the same strategy as in Section \ref{uni_bounded} but integrated in $u$ and not in $v$, and can be stated as follows.
\begin{thm}
\label{dashierl}
Let $\phi$ be as in Theorem \ref{anfangl}, then
\begin{eqnarray*}
|\phi|\leq C
\end{eqnarray*}
locally in the black hole interior up to
${\mathcal C}{\mathcal H}^+$ in a ``small neighbourhood'' of {\em left} timelike infinity $i^+$,
that is in \mbox{$[1, \infty)\times (-\infty, v_{\text{\ding{34}}}] $} for some $v_{\text{\ding{34}}}>-\infty$.
\end{thm}
\section{Energy along the future boundaries of ${\mathcal R}_{V}$}
\label{region5_proof}
Let $u_{\diamond}>u_{\text{\ding{34}}}$ and $v_*\geq v_{\gamma}(u_{\text{\ding{34}}})$. Define \mbox{${\mathcal R}_{V}=\left\{u_{\text{\ding{34}}}\leq u \leq u_{\diamond}\right\}\cap\left\{ v_*\leq v\leq \hat{v}\right\}$}, cf.~ Figure \ref{region5}, and note that \mbox{${\mathcal R}_{V} \subset {\mathcal B}$}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{region5.jpg}
\caption[]{Penrose diagram depicting region ${\mathcal R}_{V}$.}
\label{region5}\end{figure}}
We will apply the vector field
\begin{eqnarray}
\label{V_feld}
W=v^p \partial_v+\partial_u
\end{eqnarray}
as a multiplier.
The bulk can be calculated as
\begin{eqnarray}
\label{kv}
K^{W}=&-&\frac{2}{r}[v^p+1](\partial_v\phi\partial_u\phi)\nonumber\\
&-&\left[\frac12 p v^{p-1}+\frac{\partial_v \Omega}{\Omega}v^p+\frac{\partial_u \Omega}{\Omega}\right]|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2.
\end{eqnarray}
Let us define
\begin{eqnarray}
\label{tildekv}
\tilde{K}^{W}=&-&\frac{2}{r}[v^p+1](\partial_v\phi\partial_u\phi),
\end{eqnarray}
and
\begin{eqnarray}
\label{Ksphaerisch}
K_{\mbox{$\nabla \mkern-13mu /$\,}}^{W}=&-&\left[\frac12 p v^{p-1}+\frac{\partial_v \Omega}{\Omega}v^p+\frac{\partial_u \Omega}{\Omega}\right]|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,
\end{eqnarray}
with $K_{\mbox{$\nabla \mkern-13mu /$\,}}^{W}$ positive since the second term in \eqref{Ksphaerisch} dominates over the first for $v_*>2\alpha$ and $\frac{\partial_u \Omega}{\Omega}$, $\frac{\partial_v \Omega}{\Omega}$ are negative in the blueshift region.
We have therefore
\begin{eqnarray}
\label{167}
-K^{W}\leq|\tilde{K}^{W}|
\end{eqnarray}
in ${\mathcal R}_V$.
We aim for estimating it via the currents along $v=constant$ and $u=constant$ hypersurfaces.
\begin{lem}
\label{K_spherV}
Let $\phi$ be an arbitrary function. Then,
for all \mbox{$v_*\geq v_{\gamma}(u_{\text{\ding{34}}})$},
$\hat{v} >v_*$,
for $u_{\diamond}\geq u_2>u_1\geq u_{\text{\ding{34}}}$ and $\epsilon\geq u_2-u_1>0$
\begin{eqnarray}
\label{Propdelta}
\int\limits_{{\mathcal R}_{V_1}} |\tilde{K}^{W}| {\mbox{dVol}} \leq &\delta_1& \sup_{u_{1}\leq \bar{u}\leq u_{2}}\int\limits_ {\left\lbrace v_*\leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}\nonumber \\
+ &\delta_2& \sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_1\leq u \leq u_2\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}},
\end{eqnarray}
where ${\mathcal R}_{V_1}=\left\{u_1\leq {u} \leq u_2\right\}\cap {\mathcal R}_{V}$ and $\delta_1$, $\delta_2$
are positive constants, depending only on $v_*$ and $\epsilon$ such that $\delta_1\rightarrow 0$ for $\epsilon\rightarrow 0$ and $\delta_2\rightarrow 0$ as ${v}_*\rightarrow \infty$.
\end{lem}
\begin{proof}
Using the Cauchy-Schwarz inequality for equation \eqref{tildekv} we
obtain
\begin{eqnarray}
|\tilde{K}^{W}|&\leq&\frac{1}{r}\left[ \left(1 +{v^{-p}}\right){v^p}(\partial_v \phi)^2+ \left(1 +{v^p}\right)(\partial_u \phi)^2\right],
\end{eqnarray}
with the related volume element
\begin{eqnarray}
{\mbox{dVol}}&=&r^2{\frac{\Omega^2}{2}} {\mathrm{d}} u{\mathrm{d}} v{\mathrm{d}} \sigma^2_{{\mathbb S}}.
\end{eqnarray}
Note that the currents related to the vector field $W$ with their related volume elements are given by
\begin{eqnarray}
J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}} &=&{\frac{2}{\Omega^2}}\left[(\partial_u \phi)^2+ \frac{\Omega^2}{4}\bar{v}^p|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2\right], \quad {\mbox{dVol}}_{v=\bar{v}}=r^2{\frac{\Omega^2}{2}} {\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} u,\\
J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}} &=&{\frac{2}{\Omega^2}}\left[v^p(\partial_v \phi)^2+ \frac{\Omega^2}{4}|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2\right],\quad {\mbox{dVol}}_{u=\bar{u}}=r^2{\frac{\Omega^2}{2}} {\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} v,
\end{eqnarray}
cf.~ Appendix \ref{Jcurrents}.
Taking the integral over the spacetime region therefore yields
\begin{eqnarray}
\label{knbound}
\int\limits_ {{\mathcal R}_{V_1}} |\tilde{K}^{W}(\phi)| {\mbox{dVol}} &\leq& \int\limits^{u_{2}}_{u_{1}}\int\limits_ {\left\lbrace v_*\leq v \leq \hat{v}\right\rbrace }\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1 +{\bar{v}}^{-p}\right)
J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}} {\mathrm{d}} \bar{u}\nonumber \\
&&+ \int\limits^{\hat{v}}_{v_*} \int\limits_ {\left\lbrace u_1\leq u \leq u_2\right\rbrace }\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left(1 +{\bar{v}}^p\right) J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}}{\mbox{dVol}}_{v=\bar{v}}{\mathrm{d}} \bar{v},\nonumber \\
&\leq& \int\limits^{u_{2}}_{u_1} \sup_{v_*\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ {{\bar{v}}^{-p}}\right)\right]{\mathrm{d}} \bar{u} \sup_{u_{1}\leq \bar{u}\leq u_{2}}\int\limits_ {\left\lbrace v_*\leq v \leq \hat{v}\right\rbrace }J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
\nonumber\\
&&+ \int\limits^{\hat{v}}_{v_*}\sup_{u_{1}\leq \bar{u}\leq u_{2}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + {\bar{v}}^{p}\right) \right]{\mathrm{d}} \bar{v}\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_1\leq u \leq u_2\right\rbrace }J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}}{\mbox{dVol}}_{v=\bar{v}}.\nonumber\\
&&
\end{eqnarray}
It remains to show finiteness and smallness of \mbox{$\int\limits^{u_{2}}_{u_{1}} \sup_{v_*\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ {{\bar{v}}^{-p}}\right)\right]{\mathrm{d}} \bar{u} $} and \mbox{$\int\limits^{\hat{v}}_{v_*}\sup_{u_{1}\leq \bar{u}\leq u_{2}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + {\bar{v}}^{p}\right) \right]{\mathrm{d}} \bar{v}$}.
Recall the properties of the hypersurface $\gamma$ shown in Section \ref{gamma_curve}. Since $v_*> v_{\gamma}(u_{\text{\ding{34}}})$, \eqref{spacevolumedecay_u} implies that
\begin{eqnarray}
\label{zukunftu}
\Omega^2(\bar{u}, \bar{v})\leq C \Omega^2(u_{\text{\ding{34}}}, v_*), \quad \mbox{for any $(\bar{u}, \bar{v}) \in J^+(x)$, with \mbox{$x=(u_{\text{\ding{34}}}, v_*)$}, $x \in {\mathcal B}$,}
\end{eqnarray}
so that we obtain
\begin{eqnarray}
\int\limits^{u_{2}}_{u_1} \sup_{v_*\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ {{\bar{v}}^{-p}}\right)\right]{\mathrm{d}} \bar{u}
&\stackrel{\eqref{zukunftu}}{\leq}&
C \int\limits^{u_{2}}_{u_1} \sup_{v_*\leq \bar{v} \leq \hat{v}}
\Omega^2(u_{\text{\ding{34}}}, v_*)\left(1+ {{\bar{v}}^{-p}}\right){\mathrm{d}} \bar{u},\nonumber\\
&\leq& \tilde{C} \int\limits^{u_{2}}_{u_1}
|u_{\text{\ding{34}}}|^{-\beta\alpha}\left(1+ {{v_*}^{-p}}\right){\mathrm{d}} \bar{u},\nonumber\\
&\leq& \tilde{\tilde{C}}\left|{u_{2}}-{u_1}\right| \nonumber\\
&\leq&\delta_1,
\end{eqnarray}
and moreover $\delta_1 \rightarrow 0$ for $\epsilon \rightarrow 0$.
Further, in Section \ref{finiteness} we derived that similarly
\begin{eqnarray}
\label{zukunft}
\Omega^2(\bar{u}, \bar{v})\leq C \bar{v}^{-\beta\alpha}, \quad \mbox{for any $(\bar{u}, \bar{v}) \in J^+(x)$, with \mbox{$x=(u_{\text{\ding{34}}}, v_*)$}, $x \in {\mathcal B}$,}
\end{eqnarray}
where $v_*> v_{\gamma}(u_{\text{\ding{34}}})$.
\begin{eqnarray}
\int\limits^{\hat{v}}_{v_*}\sup_{u_{1}\leq \bar{u}\leq u_{2}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + {\bar{v}}^{p}\right) \right]{\mathrm{d}} \bar{v}&\stackrel{\eqref{zukunft}}{\leq}&
C\int\limits^{\hat{v}}_{v_*}\bar{v}^{-\beta\alpha}
\left( 1 + {\bar{v}}^{p}\right) {\mathrm{d}} \bar{v}\nonumber\\
&\leq& \frac{\tilde{C}}{|-\beta\alpha+p+1|}\left[\bar{v}^{-\beta\alpha+p+1}\right]^{\hat{v}}_{v_*}\nonumber\\
&\leq&\delta_2,
\end{eqnarray}
where $\delta_2 \rightarrow 0$ for $v_*\rightarrow \infty$. Thus the conclusion of Lemma \ref{K_spherV} is obtained.
\end{proof}
From the above we obtain
\begin{prop}
\label{sequenceregion}
Let $\phi$ be as in Theorem \ref{anfang} and $p$ as in \eqref{waspist}. For all \mbox{$v_*>v_{\gamma}(u_{\text{\ding{34}}})$} sufficiently large,
\mbox{$\hat{v} \in (v_*, \infty)$},
for \mbox{$u_{\diamond}\geq u_2>u_1\geq u_{\text{\ding{34}}}$} and $\epsilon\geq u_2-u_1>0$. Then for $\epsilon$ sufficiently small, the
following is true. If
\begin{eqnarray}
\label{istart2}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{1}} {\mbox{dVol}}_{u=u_{1}}&\leq& {{\tilde{C}}_1},
\end{eqnarray}
then
\begin{eqnarray}
\label{idiamond}
\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u_{2}} {\mbox{dVol}}_{u_{\diamond}}
+\int\limits_ {\left\lbrace u_{1} \leq u \leq u_{2}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
&\leq& {{\tilde{C}}_{2}}(\tilde{C}_1, u_\diamond, v_*),
\end{eqnarray}
where ${{\tilde{C}}_{2}}$ depends on $\tilde{C}_1$, ${C_{0}}$ of Theorem \ref{anfang} and ${D_{0}(u_{\diamond}, v_*)}$ of Proposition \ref{initialdataprop}.
\end{prop}
{\em Remark.} Note already that the hypothesis \eqref{istart2} is implied by the conclusion of Proposition \ref{gesamtesti} for $u_1=u_{\text{\ding{34}}}$.
\begin{proof}
By the divergence theorem and \eqref{167} we can state
\begin{eqnarray}
\label{divv}
&&\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{2}} {\mbox{dVol}}_{u=u_{2}}
+ \int\limits_ {\left\lbrace u_{1}\leq u \leq u_{2}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
\nonumber\\
&\leq&\int\limits_ {{\mathcal R}_{V_1}} |\tilde{K}^{W}| {\mbox{dVol}}+\int\limits_ {\left\lbrace u_{1}\leq u \leq u_{2}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{1}} {\mbox{dVol}}_{u=u_{1}}.
\end{eqnarray}
We can replace the hypersurfaces $u=u_2$ and $v=\hat{v}$ with $u=\bar{u}$ and $v=\bar{v}$ hypersurfaces and therefore obtain
\begin{eqnarray}
&&\sup_{u_{1}\leq \bar{u}\leq u_{2}}\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
+\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{1}\leq u \leq u_{2}\right\rbrace }J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&\leq&\int\limits_ {{\mathcal R}_{V_i}} |\tilde{K}^{W}| {\mbox{dVol}}+\int\limits_ {\left\lbrace u_{\text{\ding{34}}}\leq u \leq u_{\diamond}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{1}} {\mbox{dVol}}_{u=u_{1}},\nonumber\\
&\stackrel{\mbox{Lem.} \ref{K_spherV}}{\leq}&\delta_1\sup_{u_{1}\leq \bar{u}\leq u_{2}}\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
+\delta_2\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{1}\leq u \leq u_{2}\right\rbrace }J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&&+\int\limits_ {\left\lbrace u_{\text{\ding{34}}}\leq u \leq u_{\diamond}\right\rbrace }J_{\mu}^{W}(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{1}} {\mbox{dVol}}_{u=u_{1}}.
\end{eqnarray}
Thus, we have
\begin{eqnarray}
\label{deltamax}
&&\Rightarrow \sup_{u_{1}\leq \bar{u}\leq u_{2}}\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}
+\sup_{v_*\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{1}\leq u \leq u_{2}\right\rbrace }J_{\mu}^{W}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}}\nonumber\\
&\leq& \frac{1}{1-\mbox{max}\left\{\delta_1,\delta_2\right\}}\left[\int\limits_ {\left\lbrace u_{\text{\ding{34}}}\leq u \leq u_{\diamond}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=v_*} {\mbox{dVol}}_{v=v_*}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=u_{1}} {\mbox{dVol}}_{u=u_{1}}\right].\nonumber\\
&\leq&\tilde{C}{D_{0}(u_{\diamond}, v_*)}+{{\tilde{C}}_1},
\end{eqnarray}
where the last step
follows by statement \eqref{proppur} of Proposition \ref{initialdataprop} and by \eqref{istart2}, and where we have chosen $\epsilon$ sufficiently small and $v_*$ sufficiently close to $\infty$, such that
$\delta_1$ and $\delta_2$ satisfy say
\begin{eqnarray}
\label{deltacondition}
\delta_1, \delta_2\leq\frac12.
\end{eqnarray}
The conclusion of Proposition \ref{sequenceregion} is obtained.
\end{proof}
We are now ready to make a statement for the entire region ${\mathcal R}_V$.
\begin{prop}
\label{rechtes}
Let $\phi$ be as in Theorem \ref{anfang} and $p$ as in \eqref{waspist}.
Then,
for all \mbox{$v_*> v_{\gamma}(u_{\text{\ding{34}}})$} sufficiently large,
$\hat{v}>v_*$,
and \mbox{$u_{\diamond}>\hat{u}>u_{\text{\ding{34}}}$},
\begin{eqnarray}
\label{diamond}
\int\limits_ {\left\lbrace u_{\text{\ding{34}}} \leq u \leq u_{\diamond}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{v=\hat{v}} {\mbox{dVol}}_{v=\hat{v}}
+\int\limits_ {\left\lbrace v_* \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{W}(\phi) n^{\mu}_{u=\hat{u}} {\mbox{dVol}}_{u=\hat{u}}&\leq& {C(u_{\diamond}, v_*) },
\end{eqnarray}
where $C$ depends on ${C_{0}}$ of Theorem \ref{anfang} and ${D_{0}(u_{\diamond}, v_*)}$ of Proposition \ref{initialdataprop}.
\end{prop}
\begin{proof}
Let $\epsilon$ be as in Proposition \ref{sequenceregion}.
We choose a sequence \mbox{$u_{i+1}-u_i\leq\epsilon$} and $i=\left\{1,2,..,n\right\}$ such that $u_1=u_{\text{\ding{34}}}$ and $u_n=\hat{u}$.
Denote \mbox{${\mathcal R}_{V_i}=\left\{u_{i}\leq u \leq u_{i+1}\right\}\cap\left\{ v_*\leq v\leq \hat{v}\right\}$}, cf.~ Figure \ref{region5_sequence}.
Iterating the conclusion of Proposition \ref{sequenceregion} from $u_1$ up to $u_n$ then completes the proof.
Note that $n$ depends only on the smallness condition on $\epsilon$ from Proposition \ref{sequenceregion},
since $n\lesssim \frac{u_{\diamond}-u_{\text{\ding{34}}}}{\epsilon}$.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{region5_sequence.jpg}
\caption[]{Penrose diagram depicting regions ${\mathcal R}_{V_i}$.}
\label{region5_sequence}\end{figure}}
\end{proof}
\section{Energy along the future boundaries of $\tilde{{\mathcal R}}_{V}$}
\label{region_tilde5_proof}
Again we also need the estimates on the {\it left} side, therefore, we repeat the derivation of Section \ref{region5_proof} for region
\mbox{$\tilde{{\mathcal R}}_{V}=\left\{u_*\leq u\leq \hat{u}\right\}\cap\left\{ v_{\text{\ding{34}}}\leq v \leq v_{\diamond} \right\}$}, according to Figure \ref{region5_links}, which is located in the blueshift region \mbox{$\tilde{{\mathcal R}}_{V}\subset {\mathcal B}$}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{region5_links.jpg}
\caption[]{Penrose diagram depicting region $\tilde{{\mathcal R}}_{V}$.}
\label{region5_links}\end{figure}}
Note that for region $\tilde{{\mathcal R}}_{V}$ not only do we have to interchange $u$ and $v$, we also have to use the vector field
\begin{eqnarray}
\label{U_feld}
Z=\partial_v+u^p\partial_u
\end{eqnarray}
instead of $W$, cf.~ \eqref{V_feld}. Therefore, we can immediately state the following proposition about the energy along the horizon away from $v_{\text{\ding{34}}}$.
\begin{prop}
\label{linkes}
Let $\phi$ be as in Theorem \ref{anfangl} and $p$ as in \eqref{waspist}.
Then,
for $v_{\text{\ding{34}}}$ sufficiently close to $-\infty$,
for all \mbox{$u_*> u_{\gamma}(v_{\text{\ding{34}}})$} sufficiently large,
$\hat{u} >u_*$,
and for $v_{\diamond}>v_{\text{\ding{34}}}$
\begin{eqnarray}
\label{diamondl}
\int\limits_ {\left\lbrace v_{\text{\ding{34}}} \leq v \leq v_{\diamond}\right\rbrace } J_{\mu}^{Z}(\phi) n^{\mu}_{u=\hat{u}} {\mbox{dVol}}_{u=\hat{u}}
+\int\limits_ {\left\lbrace u_* \leq u \leq \hat{u}\right\rbrace } J_{\mu}^{Z}(\phi) n^{\mu}_{v=v_{\diamond}} {\mbox{dVol}}_{v=v_{\diamond}}&\leq& {C}(u_*, v_{\diamond}),
\end{eqnarray}
where $C$ depends on ${C_{0}}$ of Theorem \ref{anfangl} and ${D_{0}(u_*, v_{\diamond})}$ of Proposition \ref{initialdatapropl}.
\end{prop}
\begin{proof}
The proof is analogous to the proof of Proposition \ref{rechtes}.
\end{proof}
\section{Propagating the energy estimate up to the bifurcation sphere}
\label{bifurcate}
In this section we will use both results from the right and left side on ${\mathcal C}{\mathcal H}^+$.
Fix $u_{\diamond}=v_{\diamond}$, such that moreover Proposition \ref{rechtes} holds with $v_{\diamond}=v_*$,
and such that Proposition \ref{linkes} holds with $u_{\diamond}=u_*$.
We will consider a region ${\mathcal R}_{VI}=\left\{u_{\diamond}\leq u \leq \hat{u}, v_{\diamond}\leq v \leq \hat{v}\right\}$, with $\hat{u} \in (u_{\diamond}, \infty)$ and $\hat{v} \in (v_{\diamond}, \infty)$, cf.~ Figure \ref{diamant}.
{\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{diamant.jpg}
\caption[]{Penrose diagram depicting region ${\mathcal R}_{VI}$.}
\label{diamant}\end{figure}}
Recall, that in Section \ref{knsec} we have defined the weighted vector field\footnote{Since $u$ is always positive in the remaining region under consideration ${\mathcal R}_{VI}$, we have omitted the absolute value in the $u$-weight.}
\begin{eqnarray}
\label{{S2}}
{S}=v^p\partial_v+u^p\partial_u,
\end{eqnarray}
which we are going to use again to obtain an energy estimate up to the bifurcate two-sphere.
Recall $K^{S}$ given in \eqref{KN},
where the terms multiplying the angular derivatives are positive since ${\mathcal R}_{VI}$ is located in the blueshift region.
We further defined $\tilde{K}^{S}$ in \eqref{KNtilde}
and stated \eqref{knrelation} which will be useful to state the following proposition.
\begin{lem}
\label{K_spher_ende}
Let $\phi$ be an arbitrary function. Then, for all $(u_{\diamond},v_{\diamond})\in J^+(\gamma)\cap{\mathcal B}$ and all $\hat{u}>u_{\diamond}$, all $\hat{v}>v_{\diamond}$,
the integral over \mbox{${\mathcal R}_{VI}$}, cf.~ Figure \ref{diamant} of the current $\tilde{K}^{S}$, defined by \eqref{KNtilde}, can be estimated by
\begin{eqnarray}
\label{deltaende}
\int\limits_{{\mathcal R}_{VI}} |\tilde{K}^{S}| {\mbox{dVol}} \leq &\delta_1& \sup_{u_{\diamond}\leq \bar{u}\leq \hat{u}}\int\limits_ {\left\lbrace v_{\diamond} \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\bar{u}}{\mbox{dVol}}_{u=\bar{u}}\nonumber \\
+ &\delta_2& \sup_{v_{\diamond}\leq \bar{v} \leq \hat{v}}\int\limits_ {\left\lbrace u_{\diamond} \leq u \leq \hat{u}\right\rbrace }J_{\mu}^{S}(\phi) n^{\mu}_{v=\bar{v}} {\mbox{dVol}}_{v=\bar{v}},
\end{eqnarray}
where $\delta_1$ and $\delta_2$ are positive constants, with $\delta_1\rightarrow 0$ as $u_{\diamond}\rightarrow \infty$ and $\delta_2\rightarrow 0$ as $v_{\diamond}\rightarrow \infty$.
\end{lem}
\begin{proof}
The proof is similar to the proof of Lemma \ref{K_spher} of Section \ref{knsec} and Lemma \ref{K_spherV} of Section \ref{region5_proof}.
We still need to show finiteness and smallness of \mbox{$\int\limits_{u_{\diamond}}^{\hat{u}} \sup_{v_{\diamond}\leq \bar{v} \leq \hat{v}}
\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left(1+ \frac{{\bar{u}}^p}{{\bar{v}}^p}\right)\right]{\mathrm{d}} \bar{u} $} and \mbox{$\int\limits^{\hat{v}}_{v_{\diamond}}\sup_{u_{\diamond} \leq \bar{u}\leq \hat{u}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{{\bar{v}}^p}{{\bar{u}}^p}\right) \right]{\mathrm{d}} \bar{v}$}.
In Section \ref{finiteness} we derived \eqref{omegafix} which we will use now
for all $\bar{u},\bar{v} \in J^+(u_{\diamond}, v_{\diamond})$
Therefore, we can write
\begin{eqnarray}
\int\limits_{u_{\diamond}}^{\hat{u}} \sup_{v_{\diamond}\leq \bar{v} \leq \hat{v}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}
\left( 1 + \frac{{\bar{u}}^p}{{\bar{v}}^p}\right) \right]{\mathrm{d}} \bar{u}
&\leq& C\int\limits_{u_{\diamond}}^{\hat{u}} \sup_{v_{\diamond}\leq \bar{v} \leq \hat{v}}\left[\Omega^2(u_{\diamond}, v_{\diamond})e^{-\beta\left[\bar{v}-v_{\diamond}+\bar{u}-u_{\diamond}\right]}\left( 1 + \frac{{\bar{u}}^p}{{\bar{v}}^p}\right)\right] {\mathrm{d}} \bar{u},\nonumber\\
&\leq& \tilde{C}\int\limits_{u_{\diamond}}^{\hat{u}} \Omega^2(u_{\diamond}, v_{\diamond})e^{-\beta\left[\bar{u}-u_{\diamond}\right]}
\left( 1 + \frac{{\bar{u}}^p}{v_{\diamond}^p}\right) {\mathrm{d}} \bar{u},\nonumber\\
&\leq& \delta_1,
\end{eqnarray}
where $\delta_1\rightarrow 0$ as $u_{\diamond}= v_{\diamond}\rightarrow \infty$ (since $\Omega^2(u_{\diamond}, v_{\diamond})\rightarrow 0$, cf.~ \eqref{omeganull}).
Similarly, for finiteness of the second term we obtain
\begin{eqnarray}
\int\limits^{\hat{v}}_{v_{\diamond}}\sup_{u_{\diamond} \leq \bar{u}\leq \hat{u}}\left[\frac{ \Omega^2(\bar{u},\bar{v})}{2r}\left( 1 + \frac{{\bar{v}}^p}{{\bar{u}}^p}\right) \right]{\mathrm{d}} \bar{v}
&\leq& C\int\limits_{v_{\diamond}}^{\hat{v}} \sup_{u_{\diamond}\leq \bar{u} \leq \hat{u}}\left[\Omega^2(u_{\diamond}, v_{\diamond})e^{-\beta\left[\bar{v}-v_{\diamond}+\bar{u}-u_{\diamond}\right]}\left( 1 + \frac{{\bar{v}}^p}{{\bar{u}}^p}\right)\right] {\mathrm{d}} \bar{v},\nonumber\\
&\leq& \tilde{\tilde{C}}\int\limits_{v_{\diamond}}^{\hat{v}} \Omega^2(u_{\diamond}, v_{\diamond})e^{-\beta\left[\bar{v}-v_{\diamond}\right]}
\left( 1 + \frac{{\bar{v}}^p}{u_{\diamond}^p}\right) {\mathrm{d}} \bar{v},\nonumber\\
&\leq& \delta_2,
\end{eqnarray}
where $\delta_2\rightarrow 0$ as $u_{\diamond}= v_{\diamond}\rightarrow \infty$.
Thus we obtain the statement of Lemma \ref{K_spher} by fixing $u_{\diamond}=v_{\diamond}$ sufficiently large.
\end{proof}
\begin{prop}
\label{bifu}
Let $\phi$ be as in Theorem \ref{anfang} and Theorem \ref{anfangl}.
Then, for $u_{\diamond}=v_{\diamond}$ sufficiently close to $\infty$ and $\hat{u} >u_{\diamond}$, $\hat{v} >v_{\diamond}$
\begin{eqnarray}
\int\limits_ {\left\lbrace v_{\diamond} \leq v \leq \hat{v}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{u=\tilde{u}} {\mbox{dVol}}_{u=\tilde{u}}+
\int\limits_ {\left\lbrace u_{\diamond} \leq u \leq \hat{u}\right\rbrace } J_{\mu}^{S}(\phi) n^{\mu}_{v=\tilde{v}} {\mbox{dVol}}_{v=\tilde{v}}&\leq& C(u_{\diamond}, v_{\diamond}),
\end{eqnarray}
where $C$ depends on ${C_{0}}$ of Theorems \ref{anfang}, \ref{anfangl} and ${D_{0}(u_{\diamond}, v_{\diamond})}$ of Propositions \ref{initialdataprop}, \ref{initialdatapropl}.
\end{prop}
\begin{proof}
The proof follows from applying the divergence theorem for the current $J_{\mu}^{S}(\phi)$ in the region ${\mathcal R}_{VI}$. The past boundary terms are estimated by Proposition \ref{rechtes} and Proposition \ref{linkes}. Note that the weights of $J_{\mu}^{S}(\phi)$ are comparable to the weights of $J_{\mu}^{W}(\phi)$ for fixed $u_{\diamond}$, and similarly the weights of $J_{\mu}^{S}(\phi)$ are comparable to the weights of $J_{\mu}^{Z}(\phi)$ for fixed $v_{\diamond}$. The bulk term is absorbed by Lemma \ref{K_spher_ende}.
\end{proof}
Now that we have shown boundedness for different subregions of the interior we can state the following proposition for the entire interior region \mbox{$\mathring{{\mathcal M}}|_{II}$}, cf.~ Section \ref{rtcoords}
\begin{cor}
\label{letzteprop}
Let $\phi$ be as in Theorem \ref{anfang} and Theorem \ref{anfangl}.
Then
\begin{eqnarray}
\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ (|v|+1)^p (\partial_v \phi)^2(u_{fix}, v, \theta, \varphi) +|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}&\leq C, \quad \mbox{for $v_{fix} \geq v_{\text{\ding{34}}}$, $u_{fix} > -\infty$ }, \\
\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{u_{fix}} \left[(|u|+1)^p (\partial_u \phi)^2 (u, v_{fix}, \theta, \varphi)+|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u, v_{fix}, \theta, \varphi) \right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2} &\leq C,\quad \mbox{for $u_{fix} \geq u_{\text{\ding{34}}}$, $v_{fix} > -\infty$},
\end{eqnarray}
where $p$ is as in \eqref{waspist} and
$C$ depends on ${C_{0}}$ of Theorems \ref{anfang}, \ref{anfangl} and ${D_{0}(u_{\diamond}, v_{\diamond})}$ of Propositions \ref{initialdataprop}, \ref{initialdatapropl}, where $u_{\diamond}=v_{\diamond}$ is as in \ref{bifu}.
\end{cor}
\begin{proof}
This follows by examining the weights in Propositions \ref{rechtes}, \ref{linkes} and \ref{bifu} together with Theorem \ref{energythm2} and its analog for the {\it left} side.
\end{proof}
\section{Global higher order energy estimates and pointwise boundedness}
\label{global}
To obtain pointwise bounds in analogy to Section \ref{nineten} we first have to extend Corollary \ref{letzteprop} to a higher order statement.
\begin{thm}
\label{energythm3}
On subextremal Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M>|e|\neq 0$, let $\phi$ be
a solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on a two-ended asymptotically flat Cauchy surface $\Sigma$. Then, for $v_{fix}\geq v_{\text{\ding{34}}}$, $u_{fix}>-\infty$
\begin{eqnarray}
\label{array1}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ (|v|+1)^p (\partial_v \phi)^2(u_{fix}, v, \theta, \varphi) +|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq {E_{0}},\\
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}} \left[(|v|+1)^p (\partial_v \leo\phi)^2(u_{fix}, v, \theta, \varphi) +|\mbox{$\nabla \mkern-13mu /$\,} \leo\phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq {E_{1}},\\
\label{array1c}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{v_{fix}}\left[ (|v|+1)^p (\partial_v \leo^2\phi)^2 (u_{fix}, v, \theta, \varphi)+|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u_{fix}, v, \theta, \varphi) \right]r^2{\mathrm{d}} v{\mathrm{d}} \sigma_{\mathbb S^2}\leq {E_{2}};
\end{eqnarray}
and for $u_{fix}\geq u_{\text{\ding{34}}}$, $v_{fix}>-\infty$
\begin{eqnarray}
\label{array2}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{u_{fix}} \left[(|u|+1)^p (\partial_u \phi)^2 (u, v_{fix}, \theta, \varphi)+|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2(u, v_{fix}, \theta, \varphi) \right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2} \leq {E_{0}},\\
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{u_{fix}}\left[ (|u|+1)^p (\partial_u \leo\phi)^2 (u, v_{fix}, \theta, \varphi)+|\mbox{$\nabla \mkern-13mu /$\,} \leo \phi|^2(u, v_{fix}, \theta, \varphi) \right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}\leq{E_{1}},\\
\label{array2c}
&&\int\limits_{{\mathbb S}^2}\int\limits^{\infty}_{u_{fix}}\left[ (|u|+1)^p (\partial_u \leo^2\phi)^2(u, v_{fix}, \theta, \varphi) +|\mbox{$\nabla \mkern-13mu /$\,} \leo^2\phi|^2(u, v_{fix}, \theta, \varphi) \right]r^2{\mathrm{d}} u{\mathrm{d}} \sigma_{\mathbb S^2}\leq {E_{2}},
\end{eqnarray}
where $p$ is as in \eqref{waspist}.
\end{thm}
\begin{proof}
This follows immediately from Corollary \ref{letzteprop} by commutation.
\end{proof}
Having proven Theorem \ref{energythm3}, the pointwise boundedness of $|\phi|$ in all of $\mathring{{\mathcal M}}|_{II}$ follows analogously to Section \ref{uni_bounded}.
We estimate
\begin{eqnarray}
\label{uglobalfundcauchy}
\int\limits_{{\mathbb S}^2} \phi^2(\hat{u}, v){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq& \tilde{C}\left[\int\limits_{{\mathbb S}^2}\left(\int\limits_{u_*}^{\hat{u}} (|u|+1)^p(\partial_u \phi)^2(u,v){\mathrm{d}} u\int\limits_{u_*}^{\hat{u}} (|u|+1)^{-{p}}{\mathrm{d}} v\right)r^2{\mathrm{d}} \sigma_{\mathbb S^2}+\int\limits_{{\mathbb S}^2} \phi^2(u_*,v){\mathrm{d}} \sigma_{\mathbb S^2}\right],\nonumber \\
\int\limits_{{\mathbb S}^2} \phi^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{0}+\int\limits_{{\mathbb S}^2} \phi^2(u_*,v){\mathrm{d}} \sigma_{\mathbb S^2}\right],
\end{eqnarray}
where $u_*\geq u_{\text{\ding{34}}}$, $\hat{u} \in (u_*, \infty)$ and $v \in (1, \infty)$.
Commuting by angular momentum operators $\leo_i$ and summing over them we obtain
\begin{eqnarray}
\label{uglobalfundcauchy1}
\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{1}+\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(u_*,v){\mathrm{d}} \sigma_{\mathbb S^2}\right],\\
\label{uglobalfundcauchy2}
\int\limits_{{\mathbb S}^2} ({\leo}^2 \phi)^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2}
&\leq&\tilde{C}\left[\tilde{\tilde{C}} E_{2}+\int\limits_{{\mathbb S}^2} ({\leo}^2 \phi) ^2(u_*,v){\mathrm{d}} \sigma_{\mathbb S^2}\right].
\end{eqnarray}
By using the result \eqref{supr2} in \eqref{uglobalfundcauchy} we derive pointwise boundedness according to \eqref{sobo_embed}
\begin{eqnarray}
\label{usupr}
\sup_{{\mathbb S}^2}|\phi(\hat{u},{v},\theta,\varphi)|^2&\leq& \tilde{C} \left[\int\limits_{{\mathbb S}^2} ( \phi) ^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2} +\int\limits_{{\mathbb S}^2} (\leo \phi) ^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2} +\int\limits_{{\mathbb S}^2} (\leo^2 \phi) ^2(\hat{u},{v}){\mathrm{d}} \sigma_{\mathbb S^2} \right],\nonumber\\
&\leq&\tilde{C}\left[\tilde{\tilde{C}}\left( E_{0}+ E_{1}+ E_{2}\right)+C )\right],\nonumber\\
&\leq&C,
\end{eqnarray}
with $C$ depending on the initial data.
Inequalities \eqref{usupr} and \eqref{supr3} gives the desired \eqref{maineq} for all $v\geq1$.
Interchanging the roles of $u$ and $v$,
likewise \eqref{maineq} follows for all $u\geq 1$.
The remaining subset of the interior has compact closure in spacetime for which \eqref{maineq} thus follows by Cauchy stability. We have thus shown \eqref{maineq} globally in the interior.
As we will see in the next section, the continuity statement of Theorem \ref{main}
follows easily by revisiting the Sobolev estimates.
\section{Continuity statement of Theorem \ref{main}}
\label{continuity}
In the previous section we have shown pointwise boundedness, $|\phi(u,v,\varphi,\theta)|\leq C$. In the following we prove
that $\phi$ extends continuously to ${\mathcal C}{\mathcal H}^+$, that is to say
$\phi$ extends to \mbox{$\left\{\infty\right\}\times(-\infty, \infty] \cup (-\infty, \infty]\times\left\{\infty\right\}$} so that $\phi$ is continuous as a function on \mbox{$(-\infty, \infty] \times (-\infty, \infty] \times {\mathbb S}^2$}.
Showing the extension closes the proof of Theorem \ref{main}.
In order to first show continuous extendibility of $\phi$ to \mbox{$(-\infty, \infty)\times\left\{\infty\right\}$}, it suffices to show:
Given $-\infty <u<\infty$ and $\varphi, \theta \in {\mathbb S}^2$, \mbox{$\forall \epsilon>0 \quad \exists \delta, v_*$}, such that
\begin{eqnarray}
\label{entire_proof}
|\phi(u,v,\varphi, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})|<4\epsilon,\quad \mbox{for all} \left\{\begin{aligned}
&v>\tilde{v},\quad \tilde{v}\geq v_*& \\
&u-\tilde{u}<\delta&\\
&\varphi-\tilde{\varphi}<\delta&\\
&\theta-\tilde{\theta}<\delta.&
\end{aligned} \right.
\end{eqnarray}
By the triangle inequality we obtain
\begin{eqnarray}
\label{triangle}
|\phi(u,v,\varphi, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})| &\leq& |\phi(u,v,\varphi, \theta)- \phi(\tilde{u},{v},\varphi, \theta)|+|\phi(\tilde{u},{v},\varphi, \theta)- \phi(\tilde{u},\tilde{v},\varphi, \theta)|\nonumber\\
&+&|\phi(\tilde{u},\tilde{v},{\varphi}, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \theta)|+|\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})|.\nonumber\\
\end{eqnarray}
We will show that each term can be bounded by $\epsilon$.
Considering first the $u$ direction by the fundamental theorem of calculus we have
\begin{eqnarray}
\label{fund_part}
\phi(u,v,\varphi, \theta)- \phi(\tilde{u},{v},\varphi, \theta)&=& \int\limits_{\tilde{u}}^{u} \partial_{\bar{u}}\phi(u,v,\varphi, \theta){\mathrm{d}} \bar{u}.
\end{eqnarray}
Applying Cauchy Schwarz, we obtain for fixed $v, \varphi, \theta$
\begin{eqnarray}
|\phi(u,v,\varphi, \theta)- \phi(\tilde{u}, {v},\varphi, \theta)|^2 &\leq& \left(\int\limits_{\tilde{u}}^{u}|\partial_{\bar{u}}\phi(\bar{u}, {v},\varphi, \theta)|{\mathrm{d}} \bar{u}\right)^2\nonumber\\
&\leq&
\left(\int\limits_{\tilde{u}}^{u} (|{\bar{u}}|+1)^p\left(\partial_{\bar{u}}\phi(\bar{u}, {v},\varphi, \theta)\right)^2{\mathrm{d}} \bar{u}\right)\left(\int\limits_{\tilde{u}}^{u} (|{\bar{u}}|+1)^p{\mathrm{d}} \bar{u}\right)\nonumber\\
&\leq&
\left(\int\limits_{\tilde{u}}^{u} (|{\bar{u}}|+1)^p\left(\partial_{\bar{u}}\phi(\bar{u}, {v},\varphi, \theta)\right)^2{\mathrm{d}} \bar{u}\right)
\nonumber\\
&&\times\left(\frac{1}{-p+1} (|{u}|+1)^{-p+1}-\frac{1}{-p+1} (|{\tilde{u}}|+1)^{-p+1}\right)\nonumber\\
\label{sobo}
&\leq&
\tilde{C}\left(\int\limits_{\tilde{u}}^{u} \sum_{k=0}^{2} \int\limits_{{\mathbb S}^2} (|{\bar{u}}|+1)^p\left(\leo^k \partial_{\bar{u}}\phi\right)^2{\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} \bar{u}\right)\nonumber\\
&&\times\frac{1}{p-1}\left((|{\tilde{u}}|+1)^{-p+1}-{(|{u}|+1)}^{-p+1}\right)\\
\label{fund_partu}
&\leq& {\epsilon}.
\end{eqnarray}
In the above $p$ is as in \eqref{waspist} and \eqref{sobo} follows from \eqref{sobo_embed} applied to $\partial_{\bar{u}}\phi$,
\begin{eqnarray}
\label{sobo_embedup}
\sup_{\theta,\varphi \in {\mathbb S}^2}|(|{\bar{u}}|+1)^p\partial_{\bar{u}}\phi(\bar{u},v,\theta,\varphi)|^2\leq \tilde{C} \sum_{k=0}^{2} \int\limits_{{\mathbb S}^2} (|{\bar{u}}|+1)^p\left(\leo^k \partial_{\bar{u}}\phi\right)^2(\bar{u},v,\theta,\varphi){\mathrm{d}} \sigma_{\mathbb S^2}.
\end{eqnarray}
Further, the last step, \eqref{fund_partu}, then follows from \eqref{array2}-\eqref{array2c} for a suitable chosen $\delta$ in \eqref{entire_proof}.
For the second term in \eqref{triangle}, again by the fundamental theorem of calculus and the Cauchy Schwarz inequality we obtain
\begin{eqnarray}
|\phi(\tilde{u},v,\varphi, \theta)- \phi(\tilde{u},\tilde{v},\varphi, \theta)|^2 &\leq& \left(\int\limits_{\tilde{v}}^{v}|\partial_{\bar{v}}\phi(\tilde{u},\bar{v},\varphi, \theta)|{\mathrm{d}} \bar{v}\right)^2\nonumber\\
&\leq&
\left(\int\limits_{\tilde{v}}^{v} {\bar{v}}^{p}\left(\partial_{\bar{v}}\phi(u,\bar{v},\varphi, \theta)\right)^2{\mathrm{d}} \bar{v}\right)\left(\int\limits_{\tilde{v}}^{v} {\bar{v}}^{-p}{\mathrm{d}} \bar{v}\right)\nonumber\\
&\leq&
\label{dritter}
\tilde{C}\left(\int\limits_{\tilde{v}}^{v} \sum_{k=0}^{2} \int\limits_{{\mathbb S}^2} \bar{v}^p\left(\leo^k \partial_{\bar{v}}\phi\right)^2{\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} \bar{v}\right)\frac{1}{p-1}\left(\tilde{v}^{-p+1}-{v}^{-p+1}\right)\\
\label{dasda}
&\leq&
\tilde{C}\left(\int\limits_{\tilde{v}}^{v} \sum_{k=0}^{2} \int\limits_{{\mathbb S}^2} \bar{v}^p\left(\leo^k \partial_{\bar{v}}\phi\right)^2{\mathrm{d}} \sigma_{\mathbb S^2}{\mathrm{d}} \bar{v}\right)\frac{v_*^{-p+1}}{p-1}\\
\label{fund_partv}
&\leq& {\epsilon},
\end{eqnarray}
where in the third step, \eqref{dritter}, we have used \eqref{sobo_embed} applied to $\partial_{\bar{v}}\phi$. Equation \eqref{dasda} follows since $v>\tilde{v}$, $\tilde{v}\geq v_*$ and the last step follows by using \eqref{array1}-\eqref{array1c} and for $v_*$ large enough.
In the $\varphi$ direction for fixed $\tilde{u},\tilde{v}$ and $\theta$ we can state
\begin{eqnarray}
|\phi(\tilde{u},\tilde{v},\varphi, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \theta)|^2 &\leq& \left(\int\limits_{\tilde{\varphi}}^{\varphi}|\partial_{\bar{\varphi}}\phi(\tilde{u},\tilde{v},\bar{\varphi}, \theta)|{\mathrm{d}} \bar{{\varphi}}\right)^2\nonumber\\
&\leq&
C\left(\int\limits_{\tilde{\varphi}}^{\varphi} \int\limits_{0}^{\pi}\left[ \left|\partial_{\bar{\varphi}}\phi(\tilde{u},\tilde{v},\bar{\varphi}, \theta)\right|+\left|\partial_{\theta}\partial_{\bar{\varphi}}\phi(\tilde{u},\tilde{v},\bar{\varphi}, \theta)\right|\right]{\mathrm{d}} \sigma_{\mathbb S^2}\right)^2\nonumber\\
&\leq&
\tilde{C}\left(\int\limits_{{\mathbb S}^2}\left[ \left|\leo_3\phi(\tilde{u},\tilde{v},\bar{\varphi}, \theta)\right|^2+\left|\sum_i a_i \leo_i \leo_3\phi(\tilde{u},\tilde{v},\bar{\varphi}, \theta)\right|^2\right]{\mathrm{d}} \sigma_{\mathbb S^2}\right)\left(\int\limits_{\tilde{\varphi}}^{\varphi}\int\limits_{0}^{\pi}{\mathrm{d}} \sigma_{\mathbb S^2}\right)\nonumber\\
\label{fund_partphi}
&\leq& {\epsilon},
\end{eqnarray}
where in the second step we have used one dimensional Sobolev embedding.
In the third step we have used \eqref{angmom2} for $\partial_{\bar{\varphi}}$ and \eqref{angmom1} to \eqref{angmom2} for $\partial_{\theta}$. Further, we applied the Cauchy Schwarz inequality twice. The last step then follows by using \eqref{uglobalfundcauchy}-\eqref{uglobalfundcauchy2} for $\tilde{u}\geq u_{\text{\ding{34}}}$ and \eqref{fundcauchy}-\eqref{fundcauchy2} for $\tilde{u}\leq u_{\text{\ding{34}}}$ and since the second integral term is arbitrarily small by suitable choice of $\delta$.
Similarly, in $\theta$ direction for fixed $\tilde{u},\tilde{v}$ and $\tilde{\varphi}$ we obtain
\begin{eqnarray}
\label{fund_parttheta}
|\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})|^2 &\leq& \left(\int\limits_{\tilde{\theta}}^{\theta}|\partial_{\bar{\theta}}\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \bar{\theta})|{\mathrm{d}} \bar{{\theta}}\right)^2\nonumber\\
&\leq&
C\left(\int\limits_{\tilde{\theta}}^{\theta} \int\limits_{0}^{2\pi}\left[ \left|\partial_{\bar{\theta}}\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \bar{\theta})\right|+\left|\partial_{\varphi}\partial_{\bar{\theta}}\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \bar{\theta})\right|\right]{\mathrm{d}} \sigma_{\mathbb S^2}\right)^2\nonumber\\
&\leq&
\tilde{C}\left(\int\limits_{{\mathbb S}^2}\left[ \left|\sum_i a_i \leo_i\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \bar{\theta})\right|^2+\left| \leo_3\sum_i a_i \leo_i\phi(\tilde{u},\tilde{v},\tilde{\varphi}, \bar{\theta})\right|^2\right]{\mathrm{d}} \sigma_{\mathbb S^2}\right)\left(\int\limits_{\tilde{\varphi}}^{\varphi}\int\limits_{0}^{2\pi}{\mathrm{d}} \sigma_{\mathbb S^2}\right)\nonumber\\&\leq& {\epsilon}.
\end{eqnarray}
The second step follows by one dimensional Sobolev embedding and the third from \eqref{angmom1}-\eqref{angmom2} and using the Cauchy Schwarz inequality twice.
In the last step we used \eqref{uglobalfundcauchy}-\eqref{uglobalfundcauchy2} for $\tilde{u}\geq u_{\text{\ding{34}}}$ and \eqref{fundcauchy}-\eqref{fundcauchy2} for $\tilde{u}\leq u_{\text{\ding{34}}}$ and
a suitable choice of $\delta$.
Using the above results \eqref{fund_partu}, \eqref{fund_partv}, \eqref{fund_partphi} and \eqref{fund_parttheta} in \eqref{triangle} yields the desired result \eqref{entire_proof}.
To show continuous extendibility of $\phi$ to \mbox{$\left\{\infty\right\}\times(-\infty, \infty)$}, it suffices to show:
Given $-\infty <v<\infty$ and $\varphi, \theta \in {\mathbb S}^2$, \mbox{$\forall \epsilon>0 \quad \exists \delta, u_*$}, such that
\begin{eqnarray}
\label{entire_proofu}
|\phi(u,v,\varphi, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})|<4\epsilon,\quad \mbox{for} \left\{\begin{aligned}
&u>\tilde{u},\quad \tilde{u}\geq u_*&\\
&v-\tilde{v}<\delta&\\
&\varphi-\tilde{\varphi}<\delta&\\
&\theta-\tilde{\theta}<\delta.&
\end{aligned} \right.
\end{eqnarray}
This can be proven by substituting $v$ with $u$ and $\tilde{v}$ with $\tilde{u}$ and repeating all above steps.
Similarly, to show continuous extendibility to \mbox{$\left\{\infty\right\}\times\left\{\infty\right\}$}, it suffices to show:
Given $\varphi, \theta \in {\mathbb S}^2$, \mbox{$\forall \epsilon>0 \quad \exists \delta, u_*, v_*$}, such that
\begin{eqnarray}
\label{entire_proofuv}
|\phi(u,v,\varphi, \theta)- \phi(\tilde{u},\tilde{v},\tilde{\varphi}, \tilde{\theta})|<4\epsilon,\quad \mbox{for} \left\{\begin{aligned}
&u>\tilde{u},\quad \tilde{u}\geq u_*&\\
&v>\tilde{v},\quad \tilde{v}\geq v_*&\\
&\varphi-\tilde{\varphi}<\delta&\\
&\theta-\tilde{\theta}<\delta.&
\end{aligned} \right.
\end{eqnarray}
This follows as in \eqref{dasda} and completes the proof of Theorem \ref{main}.
\begin{trivlist}
\item[\hskip \labelsep ]\qed\end{trivlist}
\section{Outlook}
\label{outlook}
In the following section we give a brief overview of further related open problems.
\subsection{Scalar waves on subextremal Kerr interior backgound}
All constructions of this paper have direct generalizations to the subextremal Kerr case.\footnote{This is in contrast to the exterior region for which the
analysis of the Kerr case is significantly harder than for spherically symmetric black hole spacetimes in view of the difficulties of superradiance and the more complicated trapping. See the references below.}
The following theorem will appear in a subsequent companion paper, cf.~ \cite{ich}.
\begin{thm}
\label{main2}
\cite{ich}
On a fixed subextremal Kerr background $0\neq|a|<M$, with angular momentum per unit mass $a$ and mass $M$, let $\phi$ be a solution of the wave equation arising from sufficiently regular initial data on a Cauchy hypersurface $\Sigma$.
Then
\begin{eqnarray*}
|\phi|\leq C
\end{eqnarray*}
globally in the black hole interior up to
${\mathcal C}{\mathcal H}^+$, to which $\phi$ in fact extends continuously.
\end{thm}
The above Theorem of course depends on the fact that the analog of Theorem \ref{anfang} has been proven in the full subextremal range $|a|<M$ on Kerr backgrounds by Dafermos, Rodnianski and Shlapentokh-Rothman, cf.~ \cite{m_scalar, m_yakov}; see also \cite{anderson, tataru, tataru2, luk2, m_I_kerr, m_lec, m_bound} for the $|a|\ll M$ case and \cite{bernard, yakov} for mode stability.
\subsection{Mass inflation}
In view of our stability result, what remains of the ``blueshift instability''?
The result of Theorem \ref{main} is still compatible with the expectation that the {\it transverse} derivatives (with respect to regular coordinates on ${\mathcal C}{\mathcal H}^+$) of $\phi$ will blow up along ${\mathcal C}{\mathcal H}^+$, cf.~ the work of Simpson and Penrose \cite{simpson}.
In fact given a lower bound, say
\begin{eqnarray}
\label{lowerb}
|\partial_{v} \phi|\geq c{v}^{-4},
\end{eqnarray}
for some constant $c>0$ and all sufficiently large $v$
on ${\mathcal H}^+$, Dafermos has shown for the {\it spherically symmetric} Einstein-Maxwell-scalar field model that transverse derivatives blow up, cf.~ \cite{m_interior}.
This blow up result of \cite{m_bh} would also apply to our setting here {\it if the above lower bound \eqref{lowerb} is assumed on the spherical mean of $\phi$ on ${\mathcal H}^+$}.
See also \cite{mc_i}.
Such a lower bound however has not been proven yet for solutions arising from generic data on $\Sigma$; see also Sbierski \cite{jan}.
One might therefore aim for proving the following conjecture.
\begin{con}
Let $\phi$ be as in Theorem \ref{main} or \ref{main2}.
For \underline{generic} data on $\Sigma$,
transverse derivatives of $\phi$ blow up on ${\mathcal C}{\mathcal H}^+$, in fact $\phi$ is not $H^1_{\rm loc}$.
\end{con}
\subsection{Extremal black holes}
For a complete geometric understanding of black hole interiors one must also consider extremal black holes.
Aretakis proved stability and instability properties for the evolution of a massless scalar field on a fixed extremal Reissner-Nordstr\"om exterior background. For data on a spacelike hypersurface $\Sigma$ intersecting the event horizon and extending to infinity he has proven decay of $\phi$ up to and including the horizon, cf.~ \cite{stef1}.
In subsequent work \cite{stef2} Aretakis showed that first transverse derivatives of $\phi$ generically {\it do not decay} along the event horizon for late times. He further proved that higher derivatives {\it blow up} along the event horizon.
The analysis of the evolution of the scalar wave in the region beyond the event horizon for the extreme case remains to be shown. Motivated by heuristics and numerics of Murata,
Reall and Tanahashi \cite{reall3}, which suggest stronger stability results in the interior than in the subextremal case, we conjecture
\begin{con}
\label{outlookext}
(See \cite{reall3}.)
In extremal Reissner-Nordstr\"om spacetime $({\mathcal M},g)$, with mass $M$ and charge $e$ and $M=|e|\neq 0$, let $\phi$ be
a solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on $\Sigma$. Then, for $V_-$ regular at ${\mathcal C}{\mathcal H}^+$
\begin{eqnarray*}
|\phi|\leq C, \qquad |\partial_{V_-} \phi| \leq C,
\end{eqnarray*}
globally in the maximal domain of dependence ${\mathcal D}^+(\Sigma)$ of the hypersurface $\Sigma$ as shown in Figure \ref{exRN}.
\end{con}
{\begin{figure}[!ht]
\centering
\includegraphics[width=0.4\textwidth]{extremalRN.jpg}
\caption[]{Extremal Reissner-Nordstr\"om spacetime spacetime with Cauchy hypersurface ending in the non trapped interior region.}
\label{exRN}\end{figure}}
See upcoming work of Gajic \cite{dejan}.
Aretakis has further considered $\Box_g \phi$ on a fixed extremal Kerr background, c.f \cite{stef3}. Analogously to the extremal Reissner-Nordstr\"om case, he shows decay up to and including the event horizon for axisymmetric solutions $\phi$ and in \cite{stef4} shows instability properties when considering transverse derivatives.
This has been further generalised in \cite{reall1, reall2}.
In analogy with Conjecture \ref{outlookext}, we thus also make the following conjecture:
\begin{con}
In extremal Kerr spacetime $({\mathcal M},g)$, with mass $M$ and angular momentum per unit mass $a$ and $M=|a|\neq 0$, let $\phi$ be
an axisymmetric solution of the wave equation $\Box_g \phi=0$ arising from
sufficiently regular Cauchy data on $\Sigma$. Then, for $V_-$ regular at ${\mathcal C}{\mathcal H}^+$
\begin{eqnarray*}
|\phi|\leq C, \qquad |\partial_{V_-} \phi| \leq C
\end{eqnarray*}
globally in the maximal domain of dependence ${\mathcal D}^+(\Sigma)$ of the hypersurface $\Sigma$.
\end{con}
The case of non axisymmetric $\phi$ seems to be rather complicated and we thus do not venture a conjecture here.
\subsection{Einstein vacuum equations}
\label{EFvacuum}
We now return to the problem that originally motivated our work, namely the dynamics of the Einstein vacuum equations.
Our result and the forthcoming extension to subextremal Kerr backgrounds further support the expectation that the spherically symmetric toy models \cite{m_stab, m_interior, m_bh, amos, poisson, penrose}
are indeed indicative of what happens for the full nonlinear Einstein vacuum equations {\it without symmetry}.
In particular our results support the following conjecture given in \cite{m_bh}.
\begin{con}
{\em (A. Ori)} Let $({\mathcal M},g)$ be the maximal vacuum Cauchy development of sufficiently small perturbations of asymptotically flat two-ended Kerr data corresponding to parameters $0<|a|<M$. Then;
(a) there exist both a future and past extension $(\tilde{{\mathcal M}},\tilde{g})$ of ${\mathcal M}$ with $C^0$ metric $\tilde{g}$ such that $\partial{\mathcal M}$ is a bifurcate null cone in $\tilde{{\mathcal M}}$ and {\bf all} future (past) incomplete geodesics in $\gamma$ pass into $\tilde{{\mathcal M}}/{\mathcal M}$.
(b) Moreover, for generic such perturbations, any $C^0$ extension $\tilde{{\mathcal M}}$ will fail to have $L^2$ Christoffel symbols in a neighbourhood of any point of $\partial {\mathcal M}$.
\end{con}
A proof of part (a) has recently been announced by Dafermos and Luk, {\it given the conjectured stability of Kerr exterior} (i.e.~ given the analog of Theorem \ref{anfang} for the full nonlinear Einstein vacuum equations). See the upcoming \cite{m_luk}.
Specific examples of vacuum spacetimes with null singularities as in (b) have been constructed by Luk \cite{luk3}.
For a discussion of what all this means to Strong Cosmic Censorship see \cite{m_bh}.
\section*{Acknowledgements}
I am very grateful to Mihalis Dafermos for suggesting this problem to me and guiding me towards its completion.
Further, I would like to thank Gerard 't Hooft for his continuous encouragement and comments on the manuscript.
Parts of this work were conducted at Cambridge University, ETH Z\"urich and Princeton University, whose hospitality is gratefully acknowledged.
This work is part of the D-ITP consortium, a program of the Netherlands
Organisation for Scientific Research (NWO) that is funded by the Dutch
Ministry of Education, Culture and Science (OCW).
\begin{appendix}
\section{The $J$-currents and normal vectors}
\label{Jcurrents}
In the following we will derive the $J$-currents on constant $r$, $u$ and $v$ hypersurfaces as well as the hypersurface $\gamma$, defined in \eqref{gross_h}.
We consider an arbitrary function
$F(u,v)$ independent of the angular coordinates. Let $\zeta$ be a levelset \mbox{$\zeta=\left\{F(u,v)=0\right\}$}. Then, the normal vector to the hypersurface $\zeta$ is given by
\begin{eqnarray}
\label{ngamma}
n^{\mu}_{\zeta}&=&\frac1{\sqrt{{\Omega^2}|{{\partial_u F \partial_v F}}}|}(\partial_v F\partial_u+\partial_u F\partial_v).
\end{eqnarray}
In particular, for the future directed normal vector of an $r(u,v)=const$ hypersurface we obtain
\begin{eqnarray}
\label{n_r}
n^{\mu}_{r=const}&=&\frac{1}{\sqrt{\Omega^2}}(\partial_u+\partial_v),
\end{eqnarray}
and on constant $u$ and $v$ null hypersurfaces with their related volume elements we have
\begin{eqnarray}
\label{n_u}
n^{\mu}_{u=const}&=&\frac{2}{\Omega^2}\partial_v,\quad {\mbox{dVol}}_{u=const}=r^2\frac{\Omega^2}{2}{\mathrm{d}} \sigma_{\mathbb S^2} {\mathrm{d}} v,\\
\label{n_v}
n^{\mu}_{v=const}&=&\frac{2}{\Omega^2}\partial_u, \quad {\mbox{dVol}}_{v=const}=r^2\frac{\Omega^2}{2}{\mathrm{d}} \sigma_{\mathbb S^2} {\mathrm{d}} u.
\end{eqnarray}
For \eqref{n_u} and \eqref{n_v},
note that since vectors orthogonal to null hypersurfaces cannot be normalized, their proportionality has to be chosen consistent with an associated volume form in the application of the divergence theorem. Further, the volume element of the 4 dimensional spacetime is given by
\begin{eqnarray}
{\mbox{dVol}}=r^2\frac{\Omega^2}{2}{\mathrm{d}} \sigma_{\mathbb S^2} {\mathrm{d}} u {\mathrm{d}} v.
\end{eqnarray}
According to \eqref{J}, using an arbitrary vector field \mbox{$X=X^u \partial_u+X^v \partial_v$} we then obtain:
\begin{eqnarray}
\label{jgamma}
J_{\mu}^X(\phi) n^{\mu}_{\zeta}
&=&\frac{1}{\sqrt{\Omega^2}}\left[X^v\sqrt{\frac{\partial_u F}{\partial_v F}}
(\partial_v \phi)^2+X^u\sqrt{\frac{\partial_v F}{\partial_u F}}
(\partial_u \phi)^2\right]\nonumber\\
&&+\frac{\sqrt{\Omega^2}}{4}
\left[X^v\sqrt{\frac{\partial_v F}{\partial_u F}}+X^u\sqrt{\frac{\partial_u F}{\partial_v F}}\right]|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,\\
\label{jr}
J^{X}_{\mu}(\phi)n^{\mu}_{r=const}&=&\frac{1}{\sqrt{\Omega^2}}\left[X^v
(\partial_v \phi)^2+X^u
(\partial_u \phi)^2\right]
+\frac{\sqrt{\Omega^2}}{4}
\left[X^v+X^u\right]|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2, \\
\label{jv}
J^{X}_{\mu}(\phi)n^{\mu}_{v=const}&=&\frac{2}{\Omega^2}X^u(\partial_u \phi)^2+\frac12 X^v|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2 ,\\
\label{ju}
J^{X}_{\mu}(\phi)n^{\mu}_{u=const}&=&\frac{2}{\Omega^2}X^v(\partial_v \phi)^2+\frac12 X^u|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2 .
\end{eqnarray}
\section{The $K$-current}
\label{Kcurrents}
In order to compute all scalar currents according to \eqref{K}
in $(u,v)$ coordinates we first derive the components of the deformation tensor which is given by
\begin{eqnarray}
\label{deftensor}
(\pi^X)^{\mu \nu}=\frac12(g^{\mu \lambda}\partial_{\lambda}X^{\nu}+g^{\nu \sigma}\partial_{\sigma}X^{\mu}+g^{\mu \lambda}g^{\nu \sigma}g_{\lambda \sigma, \delta}X^{\delta}),
\end{eqnarray}
where $X$ is an arbitrary vector field, \mbox{$X=X^u \partial_u+X^v \partial_v$} without angular components.\footnote{Recall that all our multipliers $N$, $-\partial_r$, ${S_0}$ and ${S}$ only contain $u$ and $v$ components.}
From this we obtain
\begin{eqnarray*}
(\pi^X)^{v v}&=&-\frac{2}{\Omega^2}\partial_u X^v,\\
(\pi^X)^{u u}&=&-\frac{2}{\Omega^2}\partial_v X^u,\\
(\pi^X)^{u v}&=&-\frac{1}{\Omega^2}(\partial_vX^v+\partial_uX^u)-\frac{2}{\Omega^2}\left( \frac{\partial_v \Omega }{\Omega}X^v+\frac{\partial_u \Omega }{\Omega}X^u\right) ,\\
(\pi^X)^{\theta \theta}&=&\frac1{r^2}\left( \frac{\partial_v r }{r}X^v+\frac{\partial_u r }{r}X^u\right) ,\\
(\pi^X)^{\phi \phi}&=&\frac1{r^2\sin^2 \theta}\left( \frac{\partial_v r }{r}X^v+\frac{\partial_u r }{r}X^u\right) .
\end{eqnarray*}
From \eqref{energymomentum} we calculate the components of the energy momentum tensor in $(u,v)$ coordinates as
\begin{eqnarray*}
T_{v v}&=&(\partial_v \phi)^2,\\
T_{u u}&=&(\partial_u \phi)^2,\\
T_{u v}&=&T_{v u}=\frac{\Omega^2}{4}|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,\\
T_{\theta \theta}&=&(\partial_\theta \phi)^2+\frac{2 r^2}{\Omega^2} (\partial_u \phi \partial_v \phi)-\frac12r^2 |\mbox{$\nabla \mkern-13mu /$\,} \phi|^2,\\
T_{\phi \phi}&=&(\partial_\phi \phi)^2+\frac{2 r^2 \sin^2 \theta}{\Omega^2} (\partial_u \phi \partial_v \phi) -\frac12r^2 \sin^2 \theta|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2.
\end{eqnarray*}
Multiplying the components according to \eqref{K} and using the relations \eqref{def_l_n} we obtain
\begin{eqnarray}
\label{Kplug}
K^X&=& -\frac{2}{\Omega^2}\left[\partial_u X^v (\partial_v \phi)^2+\partial_v X^u (\partial_u \phi)^2\right]\nonumber\\
&&-\frac{2}{r} \left[X^v+X^u\right](\partial_u \phi\partial_v \phi)\nonumber\\
&&-\left[\frac12 \left(\partial_v X^v+\partial_u X^u\right)+\left(\frac{\partial_v \Omega}{\Omega}X^v+\frac{\partial_u \Omega}{\Omega}X^u\right)\right]|\mbox{$\nabla \mkern-13mu /$\,} \phi|^2.
\end{eqnarray}
\end{appendix}
|
\section{Introduction\label{intro}}
The {\it hydrodynamic formalism} was pioneered by Landau~\cite{LL6} and Khalatnikov~\cite{Khal} in the context of superfluid helium, and introduced to complex fluids by de Gennes~\cite{deGennes}. (Most physicists take {\em hydrodynamics} to mean the long-wave-length continuum theory of any condensed system, while engineers typically use it as a synonym for the Navier-Stokes equations.) The formalism considers energy and momentum conservation simultaneously, and has a tailored set of state variables for each condensed system. In contrast, the usual approach via {\it constitutive relations} typically leaves out energy conservation and considers the same set of variables for all systems.
In deriving a constitutive relation for a complex fluid, one usually focuses on its rheology, and postulates a quantity $\mathfrak{C}_{ij}$, as a function of the stress $\sigma _{k\ell}$, strain rate $v_{mn}$, and density $\rho $, such that the constitutive relation ${\partial}_{t}\sigma _{ij}=\mathfrak{C}_{ij}$ holds. (${\partial}_{t}\equiv\partial/\partial t$ needs to be replaced by an objective derivative more generally.)
Together with the continuity equation $\partial _{t}\rho +\nabla _{i}(\rho v_{i})=0$, momentum conservation, $\partial _{t}(\rho v_{i})+\nabla_{j}(\sigma _{ij}+\rho v_iv_j)=0$, it forms a closed set of equations for $\rho $, $\sigma _{k\ell}$ and the velocity $v_{i}$, which one may take as the universal set of variables.
The function $\mathfrak{C}_{ij}$ is specified employing experimental data, of which, usually, only a subset is employed, say elasto-plastic motion in granular media, but then not fast dense flow or elastic waves. Even though, in complex systems, it is a difficult approach requiring many arbitrary steps.
This is far better in the hydrodynamic approach.
In deriving a theory, one first identifies the basic physics of a system, with the help of which a set of state variables is specified. Different complex systems with different underlying physics therefore have different state variables. Then, by considering energy and momentum conservation, in addition to entropy balance, the energy flux and the stress are derived (not postulated), as functions of the state variables and their spatial derivatives, with a clear distinction between dissipative and reactive terms. Given the stringency of derivation, therefore, the continuum-mechanical theory thus obtained is the appropriate one for the given system if the input in physics is adequate.
Being a subject of practical importance, elasto-plastic deformation of dense granular media has been under the focus of engineering research for many decades if not
centuries~\cite{schofield,nedderman,wood1990,kolymbas1,kolymbas2,gudehus2010}.
The state of the art, however, is confusing: A large number of constitutive models compete, employing strikingly different
expressions, with none accepted as authoritative. In his recent book, {\it Physical Soil Mechanics}~\cite{gudehus2010}, Gudehus uses phrases such as {\em morass of equations} and {\em jungle of data} as metaphors.
Moreover, this competition is among theories applicable only to elasto-plastic deformation, while rapid dense flow is taken to obey yet rather different equations~\cite{hutter2007}.
It took us a while to understand, but now we realize that although these theories achieve considerable realism, they are in essence clever renditions of complex data, not reflections of the underlying physics.
This is the reason it appears worthwhile to us trying out the hydrodynamic approach, by focusing on the physics first, leaving the rich and subtle granular phenomenology aside while constructing the theory. Our hope is to arrive at one that, though not necessarily accurate in every aspect, is firmly based in physics, applicable over the complete range of shear rates, and affords a well founded understanding.
Hydrodynamic theories~\cite{hydro-1,hydro-2} have been derived for many condensed systems, including liquid crystals
~\cite{liqCryst-1,liqCryst-2,liqCryst-3,liqCryst-4,liqCryst-5,liqCryst-6,liqCryst-7},
superfluid $^3$He~\cite{he3-1,he3-2,he3-3,he3-4,he3-5,he3-6},
superconductors~\cite{SC-1,SC-2,SC-3}, macroscopic
electro-magnetism~\cite{hymax-1,hymax-2,hymax-3,hymax-4},
ferrofluids~\cite{FF-1,FF-2,FF-3,FF-4,FF-5,FF-6,FF-7,FF-8,FF-9}, and
polymers~\cite{polymer-1,polymer-2,polymer-3,polymer-4}. It is useful and possible also for granular media: Useful, because it should help to illuminate and order their
complex behavior; possible, because a significant portion has
already been done. We call it {\sc gsh}, for ``{\it granular solid hydrodynamics}.''
Remarkably, it divides granular behavior into three regimes, with the granular temperature $T_g$, or jiggling of the grains, serving as a dial:
\begin{enumerate
\item At $T_g\to0$, grains hardly jiggle, implying vanishing shear rates $\dot\gamma$: Static stress distribution and the propagation of elastic waves are phenomena of this regime. We call it {\em quasi-elastic} because the stress stems from deformed grains and is elastic in origin, and because the terms responsible for plastic behavior are quadratically small.
\item At slightly elevated $T_g$ and slow rates: The
stress is still predominantly elastic, but it may now relax: When the grains loose (or loosen) contact with one another, both granular deformation and the associated stress will decrease. The plastic terms are now comparable to the elastic ones. Typical phenomena are the {\em critical state}~\cite{schofield} and {\em incremental nonlinearity} (ie. the strikingly different loading and unloading curves), not seen in the quasi-elastic regime. The hypoplastic model~\cite{kolymbas1} and other rate-independent constitutive relations hold here. We call it the {\em hypoplastic regime} -- without implying the lack of potentials as originally thought.
\item At large $T_g$ and high shear rates, we have the rapid dense flow behavior covered by the $\mu(I)$-model~\cite{midi} and Bagnold flow. The jiggling is so strong that it exerts a pressure, and viscosities are important. They compete with the elastic stress, becoming dominant at high rates and low densities.
\end{enumerate
Finally, some words on the difference between the structure and parameters of a theory: The first concerns the part that is derived from general principles, the second is a material-dependent input, typically an assumption. This difference is especially clear-cut in constitutive models, where the structure is given by the laws of mass and momentum conservation, and the parameter is given by $\mathfrak{C}_{ij}$. In a hydrodynamic theory, the structure consists of (1)~the conservation laws for energy, momentum and mass, and (2) their respective fluxes, given as functions of the state and conjugate variables (eg. entropy $s$ and temperature $T\equiv\partial w/\partial s$), including their spacial derivatives (eg. $\nabla_iT$). In this form, the fluxes are generally valid. The most important parameter is the expressions for the energy $w$. Once given, the conjugate variables, and with them also the fluxes, are explicit functions of the state variables. Transport coefficients (eg. the viscosity) are also parameters. Ideally, one would like to obtained the parameters from a microscopic theory, though this is a tall order accomplished mainly in dilute systems (or formally dilute ones).
For denser ones, the realistic approach is to engage in a trial-and-error iteration, in which the ramification of postulated dependences are compared to experiments and simulations.
The structure of {\sc gsh} is, we believe, complete and adequate, but the parameters are not as yet specified with complete confidence. The main reason is, although the dependency of the transport coefficients on the granular temperature $T_g$ seems fairly universal, obtainable from more general considerations, that on the density varies with the type of grains. And this system-specific information needs to be obtained from data on only one type of grains. (In a sense, the $T_g$-dependence is more {\it structural}.)
In what follows, we shall first discuss the basic physics of granular media in Sec~\ref{intro-3}, then present the equations of {\sc gsh} in Sec~\ref{GSH}. The next two sections are respectively devoted to granular behavior in the quasi-elastic and hypoplastic regime, as defined above. (Application of {\sc gsh} to fast dense flow of Regime 3 is in~\cite{p&g2013}.) The manuscript ends with a conclusion and a list of symbols.
\section{The Basic Physics of Granular Media\label{intro-3}}
\subsection{Two-Stage Irreversibility}
To derive the
hydrodynamic theory for granular media, one needs the input of what the
essence of granular physics is. We believe it is
encapsulated by two notions: {\em two-stage irreversibility} and {\em variable transient elasticity}. The first is related to the three spatial scales of any
granular media: (a)~the macroscopic, (b)~the mesoscopic, granular, and (c)~the
microscopic, inner granular. Dividing all degrees of freedom (DoF) into these three
categories, we treat those of (a) differently from (b,c). Macroscopic
DoF: the slowly varying stress, flow and density fields, are
employed as state variables, but inter- and
inner granular DoF are treated summarily:
Only their contributions to the energy is considered and taken,
respectively, as granular and true heat. So we do not account for the motion of a jiggling grain, only include its fluctuating kinetic and elastic energy as contributions to the granular
heat, $\int T_g{\rm d}S_g$, characterized by the granular entropy $S_g$ and temperature $T_g$. Similarly, phonons are taken as part of true heat, $\int T{\rm d}S$. There are a handful of macroscopic DoF~(a), a large number of granular ones (b), and yet many orders of magnitude more inner granular ones (c). So the statistical tendency to equally distribute the energy among all DoF implies an energy decay: (a) $\to$ (b,c) and (b) $\to$ (c).
This is what we call {\em two-stage irreversibility}, see Fig~\ref{2stageIrr}
\begin{figure}[b]
\begin{center}
\vspace{-2cm}
\includegraphics[scale=0.18]{two-stage-irr}
\end{center}
\vspace{-1.5cm}
\caption{\label{2stageIrr}
{\em Two-stage irreversibility}. Dissipative processeses produce either granular entropy $S_g$, or directly thermal entropy $S$. Eventually, $S_g$ is also converted to $S$.}
\end{figure}
The system is in equilibrium if the true entropy $S$ is maximal.
Maximal $S_g$ would characterize a quasi-equilibrium only
if there were no energy decay from (a,b) to (c), or when it is slow enough to be neglected. As the ubiquitous inelasticity of granular collisions demonstrate, this is never the case.
A division into three scales works well when they are clearly separated -- though this is a problem of accuracy, not viability. Scale separation is well satisfied in large-scaled, engineering-type experiments, but less so in small-scaled ones. Using glass or steel beads aggravates the problem. The same is true of 2D experiments employing less and larger disks. On the other hand, when there is too little space for spatial averaging, one may still average over time and runs, to get rid of fluctuations not contained in a hydrodyanmic theory.
\subsection{Significance and Measurement of $\boldmath{T_g}$\label{significanceT_g}}
Thermodynamic variables are, strictly speaking, either conserved ones, such as energy and mass, or broken symmetry ones, such as the preferred direction in nematic liquid crystals. But one can conveniently include slowly relaxing variables in an appropriately generalized thermodynamic theory. A well known example for such variables is the magnitude of the order parameter $\psi$ in a Ginzburg-Landau theory, such as the superfluid density $\rho_s\sim\psi\psi^*$~\cite{LL6,LL5}. Expanding the free energy in $\psi$ at the phase transition, one obtains $f=\alpha|\psi|^2+\beta|\psi|^4+\cdots$, with equilibrium given for minimal $f$, or $\partial f/\partial\psi=0$. Circumstances are especially simple for $\alpha>0$, when the term $\sim\beta$ is negligible, and $\psi=0$ the equilibrium condition.
The idea behind this is a generalized notion of equilibrium -- call it {\it quasi-equilibrium} -- from one in which $\psi$ vanishes to one with a given value of $\psi$. The associated statistical ensemble includes only the micro-states compatible with this value. These are not as numerous as those with $\psi=0$, but more than sufficient for the consideration to hold, that quasi-equilibrium is the macro-state with the largest number of compatible micro-states. Any thermodynamic consideration that derives from it remains well founded, especially the principle of maximal entropy. As a result, the conjugate variables retain their thermodynamic significance: For instance, $P\equiv-\left.{\partial W}/{\partial V}\right|_{\psi}$ (with $W$ the energy and $V$ the volume) is the equilibrium force for given $\psi$.
Conserved and relaxing variables have different equilibrium conditions. The latter assumes a specific value (frequently zero, and more generally given by the vanishing of the conjugate variable, $\partial W/\partial\psi=0$). The former does not, though the associated conjugate variable is constant. For instance, the equilibrium condition with respect to energy exchange between two systems is equal temperatures, $T_1=T_2$, and to mass exchange equal chemical potentials, $\mu_1=\mu_2$. Curiously, $T_g$ alternates between both types of behavior.
Before we enter into its discussion, a caveat and a note.
Writing $\int T_g{\rm d}S_g$ for the energy contained in the granular DoF is useful only if they are in equilibrium with one another. This may not always be the case, eg. in a granular gas maintained by vibrating walls. (The system therefore needs additional state variables to characterize the velocity distribution, see~\cite{2-p&g2013}.) But grains are increasingly better equilibrated for higher densities and more frequent collisions. To keep the discussion simple, we assume that they are always in equilibrium. The note: Static granular ensembles are always in (a properly understood) thermodynamic equilibrium. Both the force equilibrium and $T_g=T$ are results of maximizing the total entropy, see Sec~\ref{equicond}.
In a rarefied granular gas, the energy $W$ has only kinetic contributions. Equipartion implies we have $\frac12T_g$ per DoF, or $W=\frac32T_gN$ (with $N$ the number of grains).
Assuming that the inner granular DoF may be modeled as a phonon gas, we take its energy as $3TN_a$ (for $T\gg T_D$, the Debye temperature, and $N_a$ the number of atoms in all the grains). If the grains maintained their mechanical integrity at arbitrarily high $T$, they will heat up during a collision for $T_g>T$, but cool down for $T_g<T$ (by amplifying the Brownian motion), until they are in equilibrium, for $T_g=T$. Clearly, all this results from $T_g$ being associated with the conserved energy (or equivalently, with the ``conserved'' entropy while minimizing the energy). Usually, of course, because $T_g\gg T$, the heat transfer is taken as independent of $T$, given by a constant restitution coefficient. It then relaxes, like $\psi$, until it vanishes.
Since the grains collide more frequently at higher densities, increasing the elastic contribution to the energy, this simple picture gets blurred, breaking down completely when the contacts become enduring.
Given the complicated interaction between grains at higher densities, including dissipation and friction, a valid $W(S_g)$ with $T_g\equiv\partial W/\partial S_g$ seems difficult to obtain head-on. Therefore, we choose the following pragmatic approach. Starting from the energy density as a function of the two entropy densities, $w(s,s_g)$, we write
${\rm d}w=T{\rm d}s+T_g{\rm d}s_g=T{\rm d}(s+s_g)+(T_g-T){\rm d}s_g$,
identifying the first term as the equilibrium energy for $T_g=T$, and the second as the additional contribution $\Delta w$ if $T_g\not=T$. Written this way, $T$ is
associated with the conserved total entropy $s_{tot}\equiv s_g+s$. It does not have a definite equilibrium value, but will equalize with the temperature of another system if heat exchange is allowed. Associated with $s_g$ at given $s_{tot}$, $T_g-T$ is an internal excitation of a non-optimal energy distribution, akin to mass nonuniformity (that will relax if uniform mass is the equilibrium state, because mass conservation does not come in here). Therefore, $T_g-T$ relaxes until $T_g-T=0$.
Using the same arguments as employed for the Ginzburg-Landau free energy, that the energy $\Delta w(\rho,T,T_g-T)$ has a minimum for $T_g-T=0$, we expand $\Delta w$
to find
\begin{equation}\label{2b-1}
\Delta w={s_g^2}/{2\rho b}, \quad T_g-T\equiv \left.{\partial w}/{\partial s_g}\right|_{s_g+s}={s_g}/{\rho b},
\end{equation}
with $b=b(\rho,T)>0, \Delta w=\frac12\rho b (T_g-T)^2$.
As this consideration assumes only analyticity of $w$ and does not depend on the interaction, it is quite general. So Eq~(\ref{2b-1}) should hold for $T_g$ sufficiently small,
and the remaining question is what $b$ is.
First, we note that taking the dimension of $T_g, s_g$ as energy and inverse volume, respectively, that of $1/b$ is volume$\times$energy. Next, having established the quadratic dependence (surprising as we are used to $w\sim T_g$), we may now take the realistic limit, $T_g\gg T$, $s\gg s_g$, to realize that the above rewriting of ${\rm d}w$ did not change much, since $s\approx s_{tot}$, $T_g\approx T_g-T$. But we now do know that $T$ is associated with a conserved variable, while $T_g$ is, {\it cum grano salis}, a relaxing one. And we may consider $b=b(\rho)\equiv b(\rho,T\to0)$.
As remarked at the end of the introduction, density dependence is a system-specific property, that needs to be obtained from experiments or simulation. We note that a gas of light-weight and completely elastic beads may serve as a {\it granular thermometer} in a DEM-experiment for a regular granular system comprising of dissipative, heavy grains, if both are separated by a massless membrane (that will transmit momentum but no particles). Maintaining the regular system at $\rho,\langle v\rangle=$ const,
where $\langle v\rangle^2\equiv\langle\vec v_i\cdot\vec v_i\rangle$, one can measure its energy $w(\rho,\langle v\rangle)$, both the kinetic and the elastic contributions, and read its temperature $T_g(\rho,\langle v\rangle)$ off the thermometer. Combining both yields $w=w(\rho,T_g)$, or $b(\rho)$ via an expansion of $w$ in $T_g$.
This procedure works mainly because waiting long enough, the thermometer will equilibrate with the grains, irrespective how dissipative they are. The many reasons a real experiment would not work is of course related to the conclusion we draw above that $T_g$ is primarily a relaxing quantity that seeks to attain its local equilibrium value, irrespective whether there is another system of a different $T_g$, with which energy may be exchanged. Given this lack of circumstances in which the thermodynamic significance of $T_g$ plays any robust role, it seems futile to insist on it.
So, instead of measuring $b$ via $T_g$'s thermodynamic significance, we may employ Eqs~(\ref{2b-1}), with a postulate $b(\rho)$ [such as given in Eqs~(\ref{2b-5}) below], to define $T_g$, such that it holds for all $T_g,\rho$. Although $T_g$ is then a true temperature only at $T_g=T$, this strategy works surprisingly well, not only for elasto-plastic motion in dense media, but also for gases.
In {\sc gsh}, the pressure exerted by jiggling grains is $\sim b T_g^2$, the viscosity $\sim T_g$, see Eq~(\ref{2b-5}, \ref{2c-3}) below. Equating $\Delta w=\frac12\rho bT_g^2$ to $\frac32 T_k\rho/m$ for
granular gases, we obtain
\begin{equation}\label{2b-2c}
3T_k/m=b(\rho) T_g^2.
\end{equation}
And indeed, the pressure is found $\sim T_k$, the viscosity $\sim\sqrt{T_k}$, in an approach combining the kinetic theory and DEM results~\cite{luding2009,Bocquet}.
Finally, it seems useful to probe whether one may identify $T_g$ with the average velocity $\langle v\rangle$. Taking the energy density of the granular DoF as $w=\rho c(\rho)\langle v\rangle^2/2$, we have $c=1$ in the dilute limit, and may plausibly take $c(\rho)$ approaching 2 in the dense limit, where enduring contacts dominate, hence any kinetic energy is converted into elastic one at the next instance, and back again. However, this conjecture, conveniently linking a macroscopic to a mesoscopic quantity, needs to be thoroughly validated.
\subsection{Variable Transient Elasticity}
Our second notion, {\em variable transient elasticity}, addresses granular plasticity.
The free surface of a granular system at rest is frequently tilted. When
perturbed, when the grains jiggle and $T_g\not=0$, the tilted surface will
decay and become horizontal. The stronger the grains jiggle
and slide, the faster the decay is. We take this as indicative of a system
that is elastic for $T_g=0$, transiently elastic for $T_g\not=0$,
with a stress relaxation rate $\sim T_g$.
A relaxing stress is typical of any viscous-elastic system such as polymers~\cite{polymer-1}.
The unique circumstance here is that the relaxation rate is not a material
constant, but a function of the state variable $T_g$. As we shall see, it
is this {\em variable transient elasticity} -- a simple fact at
heart -- that underlies the complex behavior of granular plasticity.
Realizing it yields a most economic way to capture granular rheology at elasto-plastic rates.
Employing a strain field rather than the
stress as a state variable usually yields a simpler description, because the former is in
essence a geometric quantity, the latter a physical one -- compare the evolution equations for both. Yet one cannot use the
standard strain field $\epsilon_{ij}$ as a granular state variable, because the relation between stress and $\epsilon_{ij}$ lacks uniqueness when the system is plastic.
Engineering theories frequently divide the strain into two fields, elastic $u_{ij}$ and plastic
$\epsilon^{p}_{ij}$, with the first accounting for
the reversible and second for the irreversible part. They then employ $\epsilon_{ij}$ and $\epsilon^{p}_{ij}$ as two independent
strain fields to account for elasto-plastic motion of granular media~\cite{Houlsby,Houlsby2}.
We believe that, on the contrary, the elastic strain $u_{ij}$
is the sole state variable, as there is a unique relation between the elastic stress
$\pi_{ij}$ and $u_{ij}$, as convincingly argued by Rubin~\cite{rubin}. We take $u_{ij}$ as the portion of the strain that deforms the grains and changes the energy $w=w(u_{ij})$. And since an elastic stress $\pi_{ij}$ only exists when the grains are deformed, it is also a function of $u_{ij}$.
Employing $u_{ij}$ as the sole state variable preserves many useful
features of elasticity, especially the (so-called hyper-elastic) relation,
\begin{equation}\label{1-1} \pi_{ij}=-\partial w(u_{ij})/\partial u_{ij}.
\end{equation}
This is derived in~\cite{granR2} but easy to understand via
an analogy. The wheels of a car driving up a snowy hill will grip the ground part of the time, slipping otherwise. When the wheels grip, the car moves and its gravitational energy $w$ is increased. Dividing the wheel's rotation $\theta$ into a gripping $\theta^{(e)}$ and a slipping $\theta^{(p)}$ portion, we may compute the torque on the wheel as $\partial w/\partial\theta^{(e)}$ (if the wheel turns sufficiently slowly), same as in Eq~(\ref{1-1}). How much the wheel turns or slips, how large $\theta$ or $\theta^{(p)}$ are, is irrelevant for the torque.
Although $\epsilon_{ij}$ is not a state variable, there are effects of the rates that need to be included: Given a shear rate $\dot\gamma$, the grains will jiggle and slide, producing a finite $T_g$. This effect is included in the balance equation for $T_g$, which accounts for the energy decay from (a) to (b). It has two steady-state limits, $T_g\sim\dot\gamma^2$ for low shear rates, and $T_g\sim\dot\gamma$ for higher ones.
The only way to find out whether {\it two-stage irreversibility} and {\it variable transient elasticity} are appropriate and sufficient, is to derived the theory and compare its ramifications with experiments. The structure of the theory (that we call {\sc gsh}) has already been derived, see~\cite{granR2,granR3,gudehus-jl}, though it was written such that the formal derivation is stressed, not the results, making them less accessible. And there were some blanks left, especially the density dependence of the transport coefficients, and the dependence of the elastic energy on the third invariant. All results are presented here in a readable way, and with the blanks filled in as far as possible. The second step, finding the ramifications, is a more lengthy process, in the midst of which we are.
\subsection{Validity of General Principles\label{intro-2}}
{\sc gsh} is derived employing conventional methods of theoretical physics, assuming thermodynamic considerations and associated general principles, especially the Onsager relation, are valid in granular media. As some in the community do not subscribe to it (possibly following Kadanoff~\cite{kadanoff}, who conjectured early on that granular media, being unique, may not have a hydrodynamic theory), we lay out our reasons why we believe granular media are not different to the extend as to actually violate general principles.
First, we distinguish between a {\it general principle} and an analogy. The first has been proven to hold under general conditions, hence the name; the second may be good or bad, though no basic result of theoretical physics is imperiled whatever its validity. For instance,
there are two versions of the {\it fluctuation-dissipation theorem} in granular media, one in terms of the true temperature $T$, the other in terms of the granular temperature $T_g$. The former is a {general principle} that is equally applicable to gas, a block of copper and a pile of sand, quantifying how much, eg. the volume of each fluctuates. The latter derives from the analogy between $T_g$ and $T$, and has been shown to be invalid at times -- hardly surprising, since the analogy is far from perfect. Similarly, while energy is frequently deemed not to be conserved in granular media, it is in fact only the kinetic energy of the grains that is not conserved. The total energy, including the heat in the grains, of course is.
Then there is the argument~\cite{onsager} that since grains collide inelastically and execute irreversible motion, and since the validity of the Onsager relation depends on the {\it time reversal invariance} of the underlying microscopic dynamics, the Onsager relation does not hold in granular media. This argument is not convincing -- the fact that {\it granular kinetic theory is irreversible because it is mesoscopic} has been overlooked here. The true microscopic dynamics in sand is, as everywhere else, the reversible Schr\"{o}dinger equation for the constituent atoms.
It is a deeply held belief in theoretical physics that all systems obey CPT-invariance. In condensed matter, with only electromagnetic interaction, T-invariance holds. This is the foundation of the Onsager relation, a general principle, see eg. the proof in~\cite{LL5}. The specificity of the system, or the theory one happens to employ, are irrelevant for its validity.
Another argument states that, since a sand pile has much more gravitational energy than a monolayer of grains, only the latter, the minimal energy state, is in equilibrium. The former, being ``{\it jammed}'' and prevented to reach the latter,
is too far off equilibrium for thermodynamics to hold. We contend that one needs to first also include in the consideration the elastic energy; and second, to realize that a stuck piston,
positioned between two chambers of air, is also ``jammed.'' Yet this is a system in equilibrium because all its many degrees of freedom are except one: the position of the piston that upholds a {\it constraint} on the volumes of the two subsystems. Thermodynamics is routinely applied to such a system.
In a macroscopic body, all elastic DoF are in equilibrium if the force balance holds, implying the sum of gravitational and elastic energy is minimal, see Sec.\ref{equicond}. Two elastic bodies, one on top of another, are also in equilibrium if the sum of their energy is minimal -- though there is the constraint that the upper body must not slide with respect to the lower one. A sand pile is many little elastic bodies on top of one another. If they are constrained to stay put, and their total energy is minimal, the pile is in equilibrium and amenable to thermodynamic considerations.
\section{The Expressions of {GSH} \label{GSH} }
The expressions of {\sc gsh} are divided into the static and dynamic parts.
{\bf Statics} includes the state variables, the formal equilibrium conditions in terms of them, and the expression for the thermodynamic energy. We also discuss the convexity transition of the energy, how it accounts for yield surfaces in the variable space, beyond which no elastic solutions remain stable. We note that it is qualitatively different from the yield-like critical state, one being a static, the other a dynamic, phenomenon.
{\bf Dynamics} includes conservation laws, balance equations for $s_g$ and $s$, and evolution equations for the rest of the state variables. Explicit expressions for the transport coefficients, the energy flux and the Cauchy stress are given.
\subsection{Granular Statics\label{gsh-sta}}
\subsubsection{Complete Set of State Variables}
In accordance to the above stated understanding of granular media's basic physics, the state variables are: the granular entropy $s_g$ and the elastic strain $u_{ij}$, in addition to the usual variables: the density $\rho$, the momentum density $\rho v_i$, the true entropy $s$. Denoting the
energy density (in the rest frame, $v_i=0$) as $w=w(\rho, s, s_g, u_{ij})$,
the conjugate variables are:
\begin{equation}\label{2-2} \mu\equiv\frac{\partial w}{\partial\rho},\quad
T\equiv\frac{\partial w}{\partial s},\quad T_g\equiv\frac{\partial
w}{\partial s_g},\quad \pi_{ij}\equiv-\frac{\partial w}{\partial u_{ij}},
\end{equation}
where $\mu$ is the chemical potential, $T$ the temperature, $T_g$ the
granular temperature, and $\pi_{ij}$ the elastic stress. These are given once the energy $w$ is.
Next, in Sec~\ref{equicond}, equilibrium conditions will be
derived formally, in terms of the energy and its conjugate variables, whatever $w$ is. Then, in Sec~\ref{granEn}, an example for $w$ will be given, and the conjugate variables calculated -- with the help of which the equilibrium conditions are rendered explicit.
A complete set of state variables is one that determines a
unique macroscopic state of the system. If a set is given, there is no room
for ambiguity, for ``history-" or ``preparation-dependence." Conversely, any such
dependence indicates that the set is incomplete. In the hydrodynamic
approach, a physical quantity is a state variable if (and only if) the
energy $w$ depends on it. Our assumption is that the above set is
complete.
\subsubsection{Formal Equilibrium Conditions\label{equicond}}
Equilibrium conditions for the state variables, usually in terms of
their conjugate variables, are obtained by requiring the entropy $\int s\,{\rm d}^3r$
to be maximal with appropriate constraints: constant energy $\int w\,{\rm d}^3r$, constant mass $\int \rho\, {\rm d}^3r, \cdots$. In granular media, remarkably, this universally
valid procedure leads to two distinct sets of equilibrium conditions, the
solid- and the fluid-like one. Maximizing the entropy (see~\cite{granR2,granR3} for details), we first
obtain the condition of uniform true temperature $\nabla_iT=0$, and the
requirement that the granular temperature vanishes, $T_g=0$. Usually, $T_g$
vanishes quickly, and if it does, the density is not independent from
the elastic strain, $d\rho/\rho=-du_{\ell\ell}$. They share a
common condition that we identify as the solid one,
\begin{eqnarray}
\label{2a-1}
\nabla_i(\pi_{ij}+P_T\delta_{ij})=\rho\, {\rm g}_i,\,\,
\quad
P_T\equiv-\partial(wV)/\partial V
\stackrel{T_g\to0}{\longrightarrow}0,
\end{eqnarray}
where ${\rm g}_i$ is the gravitational constant, $\pi_{ij}$ the elastic stress,
$P_T$ the usual expression for the fluid pressure, and $V$ the volume. (The derivative is taken at constant $\rho V$, $sV$ and $s_gV$.)
With the energy expression $w$ of the next Sec~\ref{granEn}, $P_T\sim T_g^2$ is the pressure exerted by jiggling grains. We therefore call it the {\em seismic pressure}~\cite{gudehus-jl}. Clearly, equilibrium condition Eq~(\ref{2a-1}), expressing force balance, is logically the result of maximal true entropy.
If $T_g$ is kept finite by external perturbations, the system may further increase its entropy by independently varying $\rho$ and $u_{ij}$, to arrive at the fluid equilibrium. It is characterized by two conditions, the first with respect to $u_{ij}$, and the second with respect to $\rho$:
\begin{equation}\label{2a-2} \pi_{ij}=0, \quad
\nabla_iP_T=\rho\, {\rm g}_i. \end{equation}
The first condition requires shear stresses to vanish in equilibrium, and free surfaces to be horizontal. The second governs reversible compaction.
\subsubsection{Granular Energy\label{granEn}}
Interested in stiff grains with small $u_{ij}$, we look for the lowest order terms in the elastic energy $w_\Delta$. Denoting $\Delta\equiv -u_{\ell\ell}$, $P_\Delta\equiv\pi_{\ell\ell}/3$, $u_s^2\equiv
u^*_{ij}u^*_{ij}$, $\pi_s^2\equiv \pi^*_{ij}\pi^*_{ij}$, where
$u^*_{ij},\pi^*_{ij}$ are the respective traceless tensors, we take it as
\begin{eqnarray}\label{2b-2} w=w_T+w_\Delta,\quad
&w_T&={s_g^2}/({2\rho b}), \quad
w_\Delta=\sqrt{\Delta }(2 {\mathcal B} \Delta^2/5+ {\mathcal A}u_s^2),
\\\label{2b-2a}
\pi_{ij}=\sqrt\Delta({\cal B}\Delta&+&{\cal A}
{u_s^2}/{2\Delta})\delta _{ij}-2{\cal A}\sqrt\Delta\, u_{ij}^*,
\\\label{2b-2b}
P_\Delta=\sqrt\Delta({\cal B}\Delta&+&{\cal A}
{u_s^2}/{2\Delta}),\quad \pi_s=-2{\cal A}\sqrt\Delta\, u_s.
\end{eqnarray}
Note $u_{ij}$ and $\pi_{ij}$ are collinear and have the same principal axes. The contribution
$w_T$ is an expansion in $s_g$, as discussed in detail around Eq~(\ref{2b-1}).
Fixing the density-dependence of the coefficient $b$ yields a
contribution for the seismic pressure $P_T\equiv-\partial(wV)/\partial(V)$. (There is also one from $w_\Delta\sim\Delta^{2.5}$ that is always much
smaller than $P_\Delta\sim\Delta^{1.5}$ for small $\Delta$, and hence neglected.) With $\rho_{cp}$ the random close density, we take
\begin{equation}\label{2b-5}
b=b_0\left(1-\rho/{\rho_{cp}}\right)^a,\quad
P_T={\rho^2\,a b\, T_g^2}/{2(\rho_{cp}-\rho)}, \end{equation}
where $b_0$ and $a$ being positive numbers. Given Eq~(\ref{2b-2c}) (noting the density dependence of $b$), this is essentially the familiar pressure expression $\sim
T_G/(\rho_{cp}-\rho)$, see eg.~\cite{Bocquet}.
Fast dense flow experiments appear to point to a small $a$, say
$a\approx0.1$, see~\cite{denseFlow}. (As we are not, at present, interested in effects such as thermal expansion, the dependence on $s$ is not discussed.)
The second term $w_\Delta$ of Eq~(\ref{2b-1}), with ${\cal A,B}>0$, is the elastic contribution. Its order of 2.5 is important for many granular features, especially {\em stress-induced anisotropy} (see below) and the {\em convexity transition}, discussed in the next section, Sec~\ref{yield surfaces}. The associated stress expression $\pi_{ij}$ has been validated for the following circumstances, achieving good agreement:
\begin{itemize}
\item Static stress distribution in three classic geometries: silo, sand pile,
point load on a granular sheet, calculated using the equilibrium condition, Eq~(\ref{2a-1}), see~\cite{ge1,ge2}.
\item Small-amplitude stress-strain relation, see~\cite{kuwano2002,ge3}.
\item Anisotropic propagation of elastic waves, see~\cite{jia2009,ge4}.
\end{itemize}
An explanation of ``{\it stress-induced anisotropy}'': In linear elasticity $w\sim u_s^2$, we have constant second derivatives $\partial^2w/\partial u_s^2$, and the velocity of a elastic wave $\sim\sqrt{\partial^2w/\partial u_s^2}\,$ does not depend on the elastic strain, or equivalently, the stress. For any exponent other than 2, the velocity depends on the stress, and is anisotropic if the stress is.
Note that the energy $w=w_T+w_\Delta$ vanishes when the grains are neither deformed nor jiggling: $w\to0$ for $s_g, u_{ij}\to0$, implying the lack of any longer-ranged
interaction among the grains. If there were one, there would be a density-dependent term in
$w$ that remains finite for $s_g, u_{ij}\to0$.
\subsection{The Yield Surfaces\label{yield surfaces}}
In a space spanned by stress components and the density, there is a surface that
divides two regions in any granular media, one in which the grains necessarily move,
another in which they may be at rest. We shall refer to this surface as {\em the
yield surface} -- though we emphasize that it is unrelated to, and different from, any yield associated with the critical state, see the next paragraph. To make its definition precise, we take the yield surface
to be the divide between two regions, one in which elastic solutions may be stable, and another in which they never are. Clearly, the medium may be at rest for a given stress only if an appropriate elastic solution is stable.
Since the elastic energy of any solution satisfying Eq~(\ref{2a-1}) is extremal, the energy is convex and minimal in the stable region, concave and maximal in the unstable one ---in which infinitesimal perturbations suffice to destroy the solution.
The yield surface defines a yield stress [such as given by Eq~(\ref{2b-3}) below]. Many textbooks identify it with the highest shear stress achieved in an approach to the {\em critical state}, with the justification that the accompanying shear rate is so low that one may consider the motion {\em quasi-static}. And since the critical state is a form of yield, the physics behind it must be static, energetic. We believe this argument overlooks the following point:
A {quasi-static motion} is one that visits a series of static, equilibrium states, so slowly that the dissipation is negligible, implying $T_g\to0$. This is what was defined as {\em quasi-elastic motion} above, see also Sec~\ref{3regimes}. The rate-independent, hypoplastic motion, taking place eg. during an approach to the critical state, is different. It does visit a series of elastic states, but at an elevated $T_g$, and is therefore highly dissipative.
The energetic instability and the critical state are two distinct concepts, static versus dynamic.
The first is a convexity transition of the elastic energy, the second a stationary solution of the
evolution equation for the elastic strain $u_{ij}$, see Sec~\ref{uri-regime}, comparable to the stationary solution of any diffusion equation.
The two yield stresses are frequently similar in magnitude, which is probably related
to the fact that both account for the clearance with the
profile of the underlying layer, though one with granular jiggling, $T_g\not= 0$, and hence a little easier. But the yield stress given by Eq~(\ref{2b-3}) below needs to be larger than the highest shear stress achieved during an approach to the critical state, because a series of elastic states is being visited during the approach. Otherwise, the system will abandon it, in search for a stable but nonuniform configuration, typically shear bands.
\subsubsection{The Coulomb Yield Surface\label{Druck-Prager}}
The elastic energy of Eq~(\ref{2b-2}) is convex only for
\begin{equation}\label{2b-3} u_s/\Delta\le\sqrt{2{\cal B}/{\cal A}} \quad
\text{or}\quad \pi_s/P_\Delta\le\sqrt{2{\cal A}/{\cal B}},
\end{equation}
turning concave if the condition is violated. The second constraint may be derived by rewriting Eq~(\ref{2b-2b}) as
${4P_\Delta}/{\pi_s}={2{\cal B}{\Delta}}/{{\cal
A}}{u_s}+{u_s}/{\Delta}$,
which shows
$P_\Delta/\pi_s=\sqrt{{\cal B}/2{\cal A}}$ is minimal for
$u_s/\Delta=\sqrt{2{\cal B}/{\cal A}}$. This corroborates the behavior that no
granular system stays static if the shear stress is too large for given
pressure. We call it the Coulomb yield surface, although technically, it is the Drucker-Prager relation, see Sec.\ref{MYS}. And again, nothing in connection to the critical state is meant here.
Taking ${\cal B}/{\cal A}$ as density independent, typically
${\cal B}/{\cal A}\approx5/3$,
we only need to specify the density dependence of $\cal B(\rho)$, which we require should account for the following three important characteristics of granular media:
\begin{itemize}
\item The energy should be concave for $\rho<\rho_{\ell p}$, the
random loose density, as no elastic solution exists when the grains loose contacts with one another.
\item The energy must be convex for larger densities, $\rho_{\ell p}<\rho<\rho_{cp}$, to ensure the stability of elastic solutions in this region.
\item The density dependence of sound
velocities as measured by Harding and Richart~\cite{hardin} should be well
rendered by $\sqrt{\partial^2w/\partial u_s^2}\sim\sqrt{\cal B}$.
\end{itemize}
The simplest expression we could find [see~\cite{granR2} for details of the struggle] is
\begin{eqnarray}\label{2b-4} {\cal B}={\cal B}_0
[(\rho-\bar\rho)/(\rho_{cp}-\rho)]^{0.15},\qua
\bar\rho\equiv(20\rho_{\ell p}-11\rho_{cp})/9,
\end{eqnarray}
with ${\cal B}_0>0$ a material constant.
The small exponent of 0.15 does not imply an accuracy over a few orders of magnitude for $\rho\to\bar\rho$. Since $\cal B$ loses its convexity at $\rho_{\ell p}$, the density is never close to $\bar\rho$. (Note $\bar\rho<\rho_{\ell p}<\rho_{cp}$, with $\rho_{cp}-\rho_{\ell p}\approx\rho_{\ell p} -\bar\rho$.) And although $\rho$ may be close to $\rho_{cp}$, the slow divergence only expresses, qualitatively and very tentatively, that the system becomes orders of magnitude stiffer there.
\subsubsection{More and Different Yield Surfaces\label{MYS}}
\begin{figure}[t] \begin{center}
\includegraphics[scale=0.28]{yield}
\end{center}
\caption{\label{fig1}
Granular yield surfaces for $T_g=0$, as a function of the pressure $P$, shear stress $\sigma_s=\pi_s$, and the void ratio $e$, as calculated from the convexity transition of the energy given in Eqs.(\ref{2b-2},\ref{2b-4},\ref{2b-5a}), with ${\cal B}/{\cal A} =5/3$, ${\cal D}_{1}=1$, ${\cal D}_{2},{\cal D}_3=2$, and $\rho _{lp}=0.85\rho _{cp}$.
The plots (a,b) are at $\sigma _{s}=0$, with the inset having a logarithmic scale; the curves of (c) are at the indicated densities. [The dashed straight lines in (b,c) are, respectively, the formula $e=e_{0}-k\ln P$ and the Coulomb yield line.] The curves of (d) are the same as in (a,c), though now in 3D-space, spanned by $e,P,\sigma_s$. }
\end{figure}
As depicted in Fig.~\ref{fig1}, granular media possess more yield surfaces. We consider
the space spanned by the pressure $P_\Delta$, shear stress $\sigma_s=\pi_s$, and the void ratio $e$, where $e\equiv1/\phi-1$. ($\phi\equiv\rho/\rho_g$ is the packing fraction, and $\rho_g$ the bulk density of the grains.) First,
for given $e$, there should be a maximal pressure that a granular system can
sustain before it collapses, implying a yield surface as depicted in (a) of Fig~\ref{fig1}.
Sand at rest will not cross this boundary when compressed. Instead, it will
collapse, becoming more compact, with a smaller $e$. The curve $e(P)$ in (a) holds for vanishing shear stress $\sigma_s=0$. If there were no dependence of $\sigma_s$, we would have vertical lines in (c), connecting the $P$-axis and the Coulomb yield line, the position of which depends on $e$. More plausible, however, would be a bending of these lines, as depicted, because a shear stress should render a static granular ensemble less stable. All this may be accounted for in {\sc gsh} by higher order terms in the elastic energy.
Although the qualitative aspects of the above described behavior must be correct, it is difficult to make them more quantitative. For lack of better data, we tentatively identify the behavior of (a) with what in textbooks on soil-mechanics~\cite{schofield,nedderman} is frequently referred to as the {\em virgin consolidation line}, and that of (c) with ``{\it caps}.'' This may not be appropriate, because both are usually associated with clay, and there are indications that with sand the consolidation line is associated with grain crushing~\cite{einaf}. [The inset, (b) of Fig~\ref{fig1}, has a logarithmic scale. It serves to demonstrate that the standard formula $e=e_{0}-k\ln P$ do not go to $\rho _{lp}$ and $\rho _{cp}$, for $P\rightarrow 0$ and $\infty $, respectively.]
We include the following higher- order terms, with ${\cal D}_1,{\cal D}_2, {\cal D}_3>0$,
\begin{equation}\label{2b-5a} -{\cal B}_0({\cal D}_1\Delta^3
+{\cal D}_2\Delta u_s^2+{\cal D}_3 u_s^4),
\end{equation}
to be added to $w_\Delta$,
Eq~(\ref{2b-2}). Consider first $u_s^2=0$. If $\Delta$ is large enough, the
term $-{\cal D}_1\Delta^3$, with a negative second derivative, will
work against ${\cal B}\Delta^{2.5}$ and turn $w_\Delta$ concave. The value
$\Delta_c$ at which this happens is given by
$\sqrt{\Delta_c}=5{\cal B}(\rho)/8{\cal D}_1(\rho)$. As ${\cal B}$ diverges
at $\rho_{cp}$, so does $\Delta_c$. If $\Delta_c(\rho)=0$ for
$\rho=\rho_{\ell p}$, ${\cal D}_1(\rho)$ will have to diverge there.
Next consider $u_s^2\not=0$. If ${\cal D}_2,{\cal D}_3=0$, the yield lines in the
space spanned by $P_ \Delta,\pi_s$ for given density would be vertical
lines. The presence of $-{\cal D}_2\Delta u_s^2$ and $-{\cal D}_3 u_s^4$ reduce the value of
$\Delta$ (or $P_ \Delta$) for growing $u_s$ (or $\pi_s$), bending the lines
to the left. We did not find enough data that we could have used to fix the values of ${\cal D}_1,{\cal D}_2,{\cal D}_3$.
Next we address varying forms of yield laws, of which there are many. That of Eq~(\ref{2b-3}) is usually referred to as the
Drucker-Prager approximation of the Coulomb yield surface. The actual Coulomb
law is anisotropic. And there are those referred to as
{\em Lade-Duncan}~\cite{lade-duncan} or {\em Matsuoka-Nakai}~\cite{matsuoka}.
Defining the friction angle as
$\varphi\equiv\arcsin\sqrt{3/(6P_\Delta^2/\pi_s^2-1)}$, the Coulomb,
Drucker-Prager, and Lade-Duncan yield laws are respectively given as
\begin{eqnarray}\nonumber \frac{\pi_3-\pi_1}{\pi_3+\pi_1
=\sin\varphi,
\quad
\frac{\pi_s}{P_\Delta}=\frac{\sqrt{6}\sin\varphi}{\sqrt{3+\sin^2\varphi}},
\quad\frac{\pi_1\pi_2\pi_3}{27P^3_\Delta
=\frac{(1-\sin\varphi)\cos^2\varphi}{(3-\sin\varphi)^3}. \end{eqnarray}
Engineers choose among them depending on the system, personal preferences
and experiences, apparently without a commonly accepted rule.
We discovered that, by including
the third strain invariant $u_t^3\equiv u^*_{ij}u^*_{jk}u^*_{ki}$ into
Eq~(\ref{2b-2}),
\begin{equation}\label{2b-6} w_\Delta= {\mathcal B}
\sqrt{\Delta }\left(2\Delta^2/5+ {\mathcal A} u_s^2/ {\mathcal B}
-{\mathcal C} {u^3_t}/ {\mathcal B}\Delta\right), \end{equation}
with ${\cal A,B,C}>0$, it is possible to account for all these laws
simultaneously. (Note the new term is also of order 2.5.) Tuning $\cal C$ is, the yield surface an be made numerically
indistinguishable from all these yield laws. Because a single expression
is employed, and because intermediate yield laws are also possible, this is
a simplifying and unifying step, see~\cite{3inv} for details (including how $u_{ij},\pi_{ij}$ remain collinear).
\subsection{Dynamics \label{dynamics}}
\subsubsection{Structure of the Dynamics}
Next, we specify the evolution equations for the state variables. The equation for the elastic strain, assuming both $u_{ij},v_{ij}$ are uniform, is~\cite{granR2}
\begin{equation}\label{2c-6a}
\partial_tu_{ij}-v_{ij}+\alpha_{ijk\ell}v_{k\ell}=-(\lambda_{ijk\ell}T_g)\,u_{k\ell}
\end{equation}
(where $v_{ij}\equiv\frac12(\nabla_iv_j+\nabla_jv_i)$ is the shear rate,
$v^*_{ij}$ its traceless part, and $v_s^2\equiv v^*_{ij}v^*_{ij}$).
If $T_g$ is finite, grains jiggle and briefly
lose contact with one another, during which their deformation is
partially lost. (More realistically, the grains only loosen contact with one another. But this suffices to mobilize them, and free them briefly from the ``elastic corset'' of neighboring grains that maintains the deformation.) Macroscopically, this shows up as a relaxation of
$u_{ij}$, with a rate that grows with $T_g$, and vanishes for
$T_g=0$. So the lowest order term in a $T_g$-expansion is
$\lambda_{ijk\ell}T_g$. With the elastic energy a convex function, the (negative) elastic stress $-\pi_{ij}=\partial w/\partial u_{ij}$ is a monotonically increasing function of $u_{ij}$. Therefore, $-\pi_{ij}, u_{ij}$
decrease at the same time. And Eq~(\ref{2c-6a}) accounts for the stress relaxation discussed in the introduction.
The Onsager coefficient $\alpha_{ijk\ell}$ is an off-diagonal element. Dividing $u_{ij}$ into $\Delta\equiv-u_{\ell\ell}$, $u_{ij}^*$, and specifying the matrices $\alpha_{ijk\ell},\lambda_{ijk\ell}$ with two elements each, Eq~(\ref{2c-6a}) is written as
\begin{eqnarray}
\label{2c-7xx}
\partial_t\Delta+(1-\alpha )v_{\ell\ell} -\alpha_1u^*_{ij}v^*_{ij}
=-\lambda_1T_g\Delta,
\\\label{2c-8}
\partial_tu^*_{ij}-(1-\alpha )v^*_{ij}
= -\lambda T_gu^*_{ij},
\\\label{2c-9xx}
\partial_tu_s-(1-\alpha )v_s= -\lambda T_gu_s.
\end{eqnarray}
The third equation is valid only if strain and rate are collinear, $u^*_{ij}/|u_s|=v^*_{ij}/|v_s|$. This is frequently the case for steady rates, because any
component of $u_{ij}$ not collinear with $v_{ij}$ relaxes to zero.
The coefficient $\alpha$ (assuming $0<\alpha <1$) describes a reduced gear ratio: The same shear rate yields a smaller deformation, $\partial_tu_{ij}=(1-\alpha)v_{ij}\cdots$, but
acts also at a smaller stress, $\sigma_{ij}=(1-\alpha)\pi_{ij}\cdots$, see below. $\alpha_1$ accounts for the fact that shearing granular media will change the compression $\Delta$, implying {\em dilatancy} and {\em contractancy}. (Although more Onsager coefficients are permitted by symmetry, they have been excluded to keep the equations as simple as possible.)
Next are the continuity equations for mass and momentum
density,
\begin{equation}\label{2c-1} \partial_t\rho+\nabla_i(\rho
v_i)=0,\qquad \partial_t(\rho v_i)+\nabla_j(\sigma_{ij} +\rho v_iv_j)=0,
\end{equation}
where the stress tensor $\sigma_{ij}=P\delta_{ij}+\sigma^*_{ij}$ (with $\sigma^*_{ij}$ the traceless part) is determined by general principles~\cite{granR2,granR3} as
\begin{eqnarray} \label{2c-2}
P\equiv\sigma_{\ell\ell}/3=(1-\alpha )P_\Delta+P_T-\zeta_gv_{\ell\ell},
\\\label{2c-2a}
\sigma^*_{ij}=(1-\alpha)\pi_{ij}^*-\alpha_1u^*_{ij}P_\Delta -\eta_gv^*_{ij},
\\ \label{2c-2b}
\sigma_{s}=(1-\alpha )\pi_s-\alpha_1u_sP_\Delta
-\eta_gv_s.
\end{eqnarray}
Again, the third equation (with $\sigma_s^2\equiv\sigma_{ij}^*\sigma_{ij}^*$) is
valid only if $\pi_{ij}^*$ and $v^*_{ij}$ are collinear,
$\pi_{ij}^*/|\pi_s|=v^*_{ij}/|v_s|$.
The pressure $P$ and shear stress $\sigma_s$ contain elastic contributions
$\sim\pi_s,P_\Delta$ from Eq~(\ref{2b-2b}), the seismic pressure $P_T\sim T_g^2$ from Eq~(\ref{2b-5}), and viscous contributions $\sim\eta_g,\zeta_g$. The off-diagonal Onsager coefficients $\alpha ,\alpha_1$ (introduced in the equation for the elastic strain $u_{ij}$) soften and mix the elastic stress components. The term preceded by $\alpha_1$ is smaller by an order in the elastic strain, and may frequently be neglected.
The balance equation for the granular entropy $s_g=b\rho T_g$ is
\begin{eqnarray}\label{2c-4}
\partial_ts_g+\nabla_i(s_gv_i-\kappa\nabla_iT_g)=
(\eta_g v_s^2+\zeta_gv^2_{\ell\ell}-\gamma T_g^2)/T_g.
\end{eqnarray}
Here, $s_gv_i$ is the convective, and $-\kappa\nabla_iT_g$ the diffusive flux. $\eta_g v_s^2$ accounts for viscous heating, for the increase of $T_g$ because
macroscopic shear rates jiggle the grains. A compressional rate
$\zeta_gv^2_{\ell\ell}$ does the same, though not as efficiently~\cite{granL3}. The term $-\gamma T_g^2$ accounts for the relaxation of $T_g$, ie., for the conversion of granular energy into inner granular one. Frequently, this equation may be simplified, first
by linearizing in $\nabla_iT_g$, assuming it to be small; then by taking all other variables to be uniform, the convective term and $v_{\ell\ell}$ as negligible. Finally, an extra source term $\gamma_1 h^2T_a^2$ may be added, to account for an ``ambient temperature'' $T_a$ -- external perturbations such as given by a sound field or by tapping.
(Generally speaking, any source mechanism contributing to $T_g$ is already included in the expression without $T_a$. For instance, given a sound field -- generated either by loudspeakers or tapping -- there is the term on the right hand side
of Eq~(\ref{2c-4} ), $\zeta_1 (v_{\ell\ell}^{sound})^2$, where $v_{\ell\ell}^{sound}$ is the fast varying compressional rate of the sound field. Coarse-graining it, we may set
$\langle\zeta_1 (v_{\ell\ell}^{sound})^2\rangle \equiv\gamma_1h^2T_a^2\equiv \eta_1 v_a^2$,
to quantify this contribution, either in terms of $T_a$, or the shear rate $v_a$ needed to produce this $T_a$. Adding such a term is a convenient short cut to account for a general perturbation without specifying the cause.) The result is
\begin{eqnarray}
\label{sum T_g}
b\rho\partial_tT_g-\kappa_1T_g\nabla^2T_g=
\eta_1 v_s^2-\gamma_1 h^2(T_g^2-T_a^2).
\end{eqnarray}
An rather similar equation holds for the true entropy $s$, see~\cite{granR2}.
\subsubsection{Transport Coefficients\label{transport coefficients}}
All coefficients $\alpha,\alpha_1,\eta_g,\zeta_g$ are functions of the state variables, $u_{ij}$, $T_g$ and $\rho$. As the hydrodynamic formalism only delivers
the structure of the dynamics, not the functional
dependence of the transport coefficients, these are to be obtained (same as the energy) from experiments or simulations, in an iterative trial-and-error process. And the specification below is what we at present believe is appropriate. Generally speaking, we find
strain dependence to be weak -- plausibly so because the strain is a small
quantity. We expand in it, keeping only the constant terms. We also expand in $T_g$, but eliminate the constant terms, because we assume granular media are fully elastic for $T_g\to0$, implying the force balance $\nabla_j\sigma_{ij}=\rho{\rm g}_i$ should reduce to the equilibrium condition, Eq~(\ref{2a-1}). Therefore we take
$\alpha,\alpha_1,\eta_g,\zeta_g,\kappa_g$ to vanish for $T_g\to0$. In addition, we
also need $\alpha,\alpha_1$ to saturate at an elevated $T_g$, such that rate-independence may be established in the hypoplastic regime. Hence
\begin{eqnarray}\label{2c-3}
\eta_g=\eta_1T_g,\,\, \zeta_g=\zeta_1T_g,\,\,
\kappa=\kappa_1T_g,
\\\nonumber
\alpha/\bar\alpha =
\alpha_1/\bar\alpha_1={T_g}/({T_\alpha+T_g}),
\end{eqnarray}
with $\bar\alpha,\bar\alpha_1,\eta_1,\zeta_1,\kappa_1,T_\alpha$ functions of $\rho$ only, or the packing fraction $\phi$. Expanding $\gamma$ in $T_g$,
\begin{equation}\label{2c-5} \gamma=\gamma_0+\gamma_1 T_g,
\end{equation}
we keep $\gamma_0$, because the reason that led to Eqs~(\ref{2c-3}) does not apply. More importantly, $\gamma_0$ ensures a smooth transition from the hypoplastic to the quasi-elastic regime, see Eq~(\ref{TgVs2}) below. (Although $\gamma_0=0$ in rarefied systems~\cite{luding2009}, we do not see any reasons for this to hold for denser ones.)
The transport coefficients are also functions of $\rho$, containing especially a divergent/vanishing part $\sim(\rho_{cp}-\rho)$. Assuming that, at $\rho=\rho_{cp}$, the plastic phenomena of stress relaxation, softening and dilatancy vanish, $T_g$ relaxes instantly, and the system is infinitely viscous, we take
\begin{equation}\label{density-dependence}
\lambda,\,\lambda_1,\,\alpha,\,\alpha_1,\,\gamma_1^{-1},\,\eta_1^{-1}\sim\rho_{cp}-\rho.
\end{equation}
We need to stress that we stand behind the temperature dependence with much more confidence than that of the density, for two reasons: First, $\rho$ is not a small quantity that one may expand in, and we lack the general arguments employed to extract the $T_g$-dependence. Second, not coincidentally, the $\rho$ dependence does not appear universal: The above dependence of $\gamma_1,\,\eta_1$ seems to fit glass beads data, while $ \gamma_1\sim(\rho_{cp}-\rho)^{-0.5}$, $\eta_1\sim(\rho_{cp}-\rho)^{-1.5}$ appear more suitable for polystyrene beads, see~\cite{denseFlow}.
At given shear rates, $v_s=$ const, the stationary state of Eq~(\ref{2c-4}) -- characterized by $\partial_ts_g=0$, with viscous heating balancing $T_g$-relaxation -- is quickly
arrived at, say within $10^{-3}$ s in dense granular media, implying
\begin{eqnarray}\label{2c-6}
{\gamma_1} \,h^2\, T_g^2=v_s^2\,{\eta_1}+v^2_{\ell\ell}\,{\zeta_1},
\qua
h^2\equiv1+\gamma_o/(\gamma_1T_g).
\end{eqnarray}
Taking the density for simplicity as either constant or slowly changing, ie. $v^2_{\ell\ell}\approx0$,
we have a quadratic regime for small $T_g$ and low $v_s$, and a linear one at elevated $T_g,v_s$:
\begin{eqnarray}
\label{TgVs}
T_g=|v_s|\sqrt{\eta_1/\gamma_1}\quad\,\,\text{for}\quad \gamma_1T_g\gg\gamma_0,
\\\label{TgVs2}
T_g=v_s^2\,\,({\eta_1/\gamma_0})\quad\text{for}\quad \gamma_1T_g\ll\gamma_0.
\end{eqnarray}
As mentioned above and discussed in the next section, the linear regime is hypoplastic, in which the system displays rate-independent elasto-plastic behavior and the hypoplastic model holds. In the quadratic regime, because $T_g\sim v_s^2\approx0$ is quadratically small, the behavior is quasi-elastic, quasi-static, with slow, consecutive visit of static stress distributions. Note that we have $h=1$ in the hypoplastic regime, and $h\to\infty$ in the quasi-static one.
Eqs~(\ref{2c-7xx},\ref{2c-8},\ref{2c-9xx}) also have a stationary solution,
$\partial_t\Delta, \partial_tu_s=0$, in which the shear rate $v_s=$ const
is compensated by the relaxation $\sim T_g$. As a result,
$\Delta=\Delta_c,u_s=u_c$ remain constant, and with them also the
pressure and shear stress, $P=P_c, \sigma_s=\sigma_c$. This ideally plastic behavior is the {\em critical state}. In the linear regime,
$T_g\sim|v_s|$, both $P_c$ and $\sigma_c$ are rate-independent. Since the rate-independent critical state is a motion in the linear regime, and since it is irreversible and strongly dissipative, it is not quasi-static.
\subsection{Summary\label{sum}}
With the above set of equations derived, the expressions for energy density and transport coefficients in large part specified, {\sc gsh} is a fairly well-defined theory. It contains clear ramifications and provides little
leeway for retrospective adaptation to observations. As a first step to coming to terms with its ramifications, we examine its basic features.
Granular rheology as observed may be divided into three shear rate regimes: {\em Bagnold} for high, {\em hypoplastic} for low, and {\em quasi-elastic} for even lower
ones. Fast dense flow is in the first regime, in which pressure and shear stress are proportional to shear rate squared, $p,\sigma_s\sim v_s^2$. Various elasto-plastic motions, observed especially
in triaxial apparatuses, are in the second, rate-independent regime. The third regime is elastic -- no difference between load and unload, and no critical state.
Static stress distribution and elastic waves belong here.
This third regime is again rate-independent.
Although textbooks, taking the hypoplastic regime as quasi-static, do not acknowledge the existence of a third rate regime, we note that elastoplastic motion cannot be quasi-static, because it is plastic and irreversible, see the discussion in Sec~\ref{yield surfaces} and~\ref{3regimes}. In {\sc gsh}, the static, equilibrium state with $T_g=0$ is fully elastic. If quasi-static motion exists, it must be quasi-elastic. On the other hand, it is admittedly difficult to observe. Some possible reasons are discussed in Sec~\ref{aeipt}, with suggestions in~\ref{soft springs} on how to overcome them.
{\sc gsh} is constructed such that any deviation from elasticity -- encapsulated in the
coefficients $\alpha,\alpha_1,\eta_g,\zeta_g,\kappa_g,\lambda
T_g,\lambda_1T_g$ -- vanishes with $T_g$.
For $T_g=0$, we have $\partial_t u_{ij}=v_{ij}\equiv\partial_t\epsilon_{ij}$, or
$u_{ij}=\epsilon_{ij}$
$\sigma_{ij}=\pi_{ij}$,
implying perfect elasticity. At very low shear rates, $T_g\sim v_s^2$, deviations from elasticity are quadratically small. The system is then {quasi-elastic} -- though only as long as no yield surface (as discussed in Sec~\ref{yield surfaces}) is breached.
When $T_g$ is more elevated, we are in the linear regime, $T_g\sim|v_s|$, see Eq~(\ref{TgVs}). Here, the full complexity of granular media emerges. Nevertheless, three scalar equations, derived starting from {\em two-stage irreversibility} and {\em variable transient elasticity}, suffice to account for most phenomena. Two account for transient elasticity,
Eqs~(\ref{2c-7xx},\ref{2c-9xx}), and one for $T_g$, Eq~(\ref{2c-4}) or Eq~(\ref{sum T_g})
In the hypoplastic regime, the stress is still elastic, though softened by $\bar\alpha$.
Noting $\pi^*_{ij}$, $u^*_{ij}, \sigma^*_{ij}$ are
collinear, and assuming the higher order term $\bar\alpha_1u_sP_\Delta$ may be neglected, we have the rate-independent expressions
\begin{equation}
P=(1-\bar\alpha)P_\Delta, \quad \sigma_{s}=(1-\bar\alpha)\pi_s.
\label{sum stress hp}\end{equation}
As we shall see, these simple expressions are well capable of accounting for elasto-plastic motion generally, including especially load-unload behavior, Sec~\ref{Load and Unload}, and the approach to the critical state, Sec~\ref{critical state}. They were also used for a successful comparison to the hypoplastic and barodesy model, in Sec~\ref{conrel}, and for the damping of elastic waves, Sec~\ref{elastic waves}.
For yet larger rates, the total stress includes the seismic pressure $P_T$ and the viscousity (of which the compressional one is neglected), see Eqs~(\ref{2c-2},\ref{2c-2b}),
\begin{eqnarray}\label{sum stress df}
P=(1-\bar\alpha)P_\Delta+{\textstyle\frac12} T_g^2\, a\, \rho^2 \,b/(\rho_{cp}-\rho),
\quad
\sigma_{s}=(1-\bar\alpha)\pi_s-\eta_1T_gv_s.
\end{eqnarray}
Since $T_g\sim|v_s|$, we have $P_T\sim T_g^2\sim v_s^2$ and $\eta_gv_s=\eta_1T_gv_s\sim v_s^2$.
So both may be written as $e_1+e_2v_s^2$, implying a quadratic
dependence on the rate for $e_2v_s^2\gg e_1$, and rate-independence for $e_2v_s^2\ll e_1$. Rapid dense flow is considered in~\cite{p&g2013}.
This ends the brief presentation of {\sc gsh}.
\section{The Quasi-Elastic Regime\label{quasi elastic motion}}
\subsection{Quasi-Elastic versus Hypoplastic Regime\label{3regimes}}
Many in soil mechanics call the slow granular motion in the hypoplastic regime -- say the approach to the critical state -- {\it quasi-static}. We do not think this is the right term, because, as discussed at the beginning of Sec.\ref{yield surfaces}, the motion occurs at an elevated $T_g$, is dissipative and irreversible.
Quasi-static motion is never dissipative. Consider sound propagation in any system, say Newtonian liquid, elastic media or liquid crystals. The velocity is a constant, and the damping $\sim\omega$, the frequency. Therefore, sound waves are less damped the smaller the frequency is. This is a rather generic feature: Changing the state variable $A$ slowly, dissipation vanishes with $\partial_tA$, the rate of change. At the very slow limit, the dissipation may be neglected, and the motion is rate-independent. It is then called {\em quasi-static}, because the system is at this rate visiting static, equilibrium states consecutively.
Granular systems are both dissipative and rate-independent in the hypoplastic regime. As we shall see in Sec~\ref{uri-regime}, this rate-independence is a reflection of the fact that reactive and dissipative terms have the same frequency dependence, and are comparable in size -- they are exactly equal in the critical state. If there were only the hypoplastic regime, elastic waves would always be overdamped. Since this is not the case, there must be a different rate-independent but dissipation-free regime.
Faced with this dilemma, a frequent suggestion is to take a small incremental strain (such as given in an elastic wave) to be elastic and free of dissipation, but a large one as elasto-plastic and dissipative. For the following reason, we believe this is incompatible with the notion of a quasi-static motion, and the wrong way out: Starting from a static state of given stress, and applying a small incremental strain that is elastic, the system is again in a static state and an equally valid starting point. The next small increment must therefore also be purely elastic. Many consecutive small increments yield a large change in strain, and if the small ones are not dissipative, neither can their sum be.
In {\sc gsh}, it is the strain rate, not its amplitude, that decides whether the system is elastic or hypoplastic. Small strain increments achieved with
a high but short lasting shear rate will provoke an elastic response, if
$T_g$ does not have time to get to a sufficiently high value to induce
any plastic responses. Furthermore, the mere existence of a quasi-static, quasi-elastic regime does not imply that it is also easily observable, though see Sec~\ref{soft springs}.
Finally, we note that backtracing of the stress curve $\hat\sigma(t)$ when reversing the strain, $\hat\epsilon(t)\to\hat\epsilon(-t)$, occurs only in
the quasi-elastic regime, not the hypoplastic one. (We use a hat
to indicate a tensor.) The stress is a function of the
elastic strain, $\hat\sigma=\hat\sigma(\hat u)$. Reversing $\hat u(t)$ will
always backtrace $\hat\sigma(t)$. But only in the quasi-elastic regime may
we identify $\hat u(t)=\hat\epsilon(t)$. Failure to backtrace at hypoplastic rates are not evidence of a ``history dependence."
\subsection{An Elastic-Ideally-Plastic Trajectory\label{aeipt}}
In the quadratic regime of very slow shear rates, $T_g\sim |v_s|^2\to0$, the granular temperature is so small that the system is essential elastic, moving from one elastic, equilibrium state to a slightly different one. This is the reason we call it {\em quasi-elastic}. Because $\hat\sigma=\hat\pi$ and $\partial_t\hat u\to\partial_t\hat\epsilon=\hat v$, the change of the the shear stress $\sigma_s$ is well approximated by the (hyper-) elastic relation,
\begin{equation}\label{3a-1}
\partial_t\sigma_{ij}=\frac{\partial\sigma_{ij}}{\partial
u_{k\ell}}\partial_t u_{k\ell} =\frac{\partial\pi_{ij}}{\partial
u_{k\ell}}\partial_t\epsilon_{k\ell}=-\frac{\partial^2 w}{\partial
u_{ij}\partial u_{k\ell}}v_{k\ell}.
\end{equation}
Shearing a granular medium at quasi-elastic rates, the result will be a trajectory $\hat\sigma(\hat\epsilon)$ that is much steeper than in experiments at hypoplastic rates, such as observed during an approaching to the critical state. The gradient is given directly by the stiffness constant ${\partial^2 w}/{\partial \hat u^2}$, and possibly three to four times as large as the average between loading and unloading at hypoplastic rates [because Eq~(\ref{2c-8}) lacks the factor of $(1-\alpha)$]. This goes on until the system reaches a yield surface of the elastic energy, the convexity transitions discussed in Sec~\ref{yield surfaces}. The system becomes ideally plastic at this point, abruptly, by forming shear bands. The critical state will not be reached. Reversing the shear rate in between will retrace the function $\hat\sigma(t)$.
\subsection{Soft Springs versus Step Motors\label{soft springs}}
\begin{figure}[b] \begin{center}
\includegraphics[scale=0.35]{stepmotor}
\end{center}
\caption{Why observing the quasi-elastic regime is hard if step motors are used.\label{StepMotor}}
\end{figure}
Quasi-elastic behavior has not been observed in triaxial apparatus. This may simply be because even the lowest rates are not slow enough. Or because step motors are widely used in these appliances. Plotting the shear
rate versus time, $\dot\gamma(t)$, different shear rates are approximately given as depicted in Fig~\ref{StepMotor}. Although the curves have different average rates $\langle\dot\gamma\rangle$, the time-resolved, maximal rates $\dot\gamma_M$ are identical. And if the time span of $\dot\gamma_M$ is long enough for $T_g$ to respond, and $\dot\gamma_M$ is high enough for the system to be in the linear regime, $T_g\sim \dot\gamma_M$, the system will display consecutive hypoplastic behavior in both cases, irrespective of the average rate $\langle\dot\gamma\rangle$.
Here, we suggest two ways to observe quasi-elastic behavior, both by fixing the stress rate at low $T_g$, because a given stress rate corresponds to two different shear rates, a high one at elevated $T_g$ and a low one at vanishing $T_g$. The first method is slowly incline a plane supporting a layer of grains. In such a situation, the shear rate remains very small, and the system starts flowing only when a yield surface is breached. In contrast, employing a feedback loop in a triaxial apparatus to maintain a stress rate would not work well, because the correcting motion typically has strain rates that are too high.
A second method is to insert a very {soft spring}, even a rubber band, between the granular
medium and the device moving at a given velocity $v$ to deform it. If the spring is softer by a large factor $a$ than the granular medium (which is itself rather soft), it will absorb most of the displacement, leaving the granular medium deforming at a rate smaller by
the same factor $a$ than without the spring. In other words, the soft spring serves as a ``stress reservoir'' for the granular medium. The same physics applies when the feedback loop is connected via a soft spring, as then only little $T_g$ is excited.
\section{The hypoplastic regime\label{uri-regime}}
Hypoplastic motion occurs at an elevated $T_g\sim|v_s|$, in what we have named the linear regime. It is {\em rate-independent} for given, constant strain rates, in the sense that the increase in the stress $\Delta\sigma_{ij}$ depends only on the increase in the strain,
$\Delta\epsilon_{ij}=\int v_{ij}{\rm d}t$, not how fast it takes place. We also call this regime hypoplastic because this is where the {\em hypoplastic model} holds, a state-of-the-art engineering theory~\cite{kolymbas1} that we shall consider in Sec~\ref{Hypoplasticity}.
\subsection{Load and Unload\label{Load and Unload}}
In the hypoplastic regime, for given shear rate $v_s$, the granular temperature relaxes quickly to its stationary value $T_g=|v_s|\,\sqrt{\eta_1/\gamma_1}$. Inserting this into
Eqs~(\ref{2c-7xx}, \ref{2c-9xx}), we arrive at
\begin{eqnarray}
\partial_t\Delta=v_s\,\alpha_1u_s&-&|v_s|\,\Lambda_1\Delta,
\quad \label{3b-3}
\partial_tu_s=v_s\,(1-\alpha )-|v_s|\,\Lambda u_s,
\\\label{3b-4}
\Lambda&\equiv&{\lambda}/{h}\sqrt{{\eta_1}/{\gamma_1}}\equiv\Lambda_1{\lambda}/{\lambda_1}\sim
(\rho_{cp}-\rho), \end{eqnarray}
which are explicitly rate-independent for $\alpha=\bar\alpha,\alpha_1=\bar\alpha_1$, see Eq~(\ref{2c-3}).
The last equation is a result of inserting the density dependence of Eqs~(\ref{density-dependence}) and indicates that relaxation of the elastic strain becomes slower at higher density, and stops at the close-packed density $\rho_{cp}$, where the system is essentially elastic. We take $\Lambda\approx3.3\Lambda_1$, as compressional relaxation is typically slower than shear relaxation~\cite{granL3}.
\begin{figure}[tbh] \begin{center}
\includegraphics[scale=0.3]{fig2}
\end{center}
\caption{\label{fig2} The hysteretic change of the shear stress ($\sim u_{s}$) with the strain, as accounted for by Eq~(\ref{3b-3}). The sign of the shear rate $v_{s}(t)$ is given in (b), the shear deformation $\varepsilon _{s}=\int_{0}^{t}v_{s}(t^{\prime })dt^{\prime }$ in
(c). Inset (d) is the the temporal evolution of $u_{s}$. }
\end{figure}
In this form, it is obvious that loading ($v_s=|v_s|>0$) and unloading ($v_s=- |v_s|<0$) have different slopes: $\partial_tu_s/v_s=(1-\alpha)\mp(\Lambda u_s/h)$.
This phenomenon is referred to as {\em
incremental nonlinearity} in soil mechanics, and the reason why no backtracking takes place under reversal of shear rate: Starting from isotropic stress, $u_s=0$, see Fig~\ref{fig2}, the gradient is at first $(1-\alpha)$, becoming smaller as $u_s$ grows, until it
is zero, in the stationary case $\partial_tu_s/v_s=0$. Unloading now, the
slope is $(1-\alpha)+(\Lambda u_s/h)$, steeper than it has ever been. It is again
$(1-\alpha)$ for $u_s=0$, and vanishes for $u_s$ sufficiently negative, see Fig~\ref{fig2}. Same scenario holds for $\partial_t\Delta/v_s$.
Clearly, only the stress $P,\sigma_s$ are measurable, not
$\Delta, u_s$. The former is calculated employing Eq~(\ref{sum stress hp}) when
the latter is given. The resultant expressions can be complicated
(especially if the pressure is held constant instead of the density), but the basic physics remains the same -- an illustration of why $u_{ij}$ is the better state variable.
In systematic studies employing discrete numerical simulation, Roux and coworkers have
accumulated great knowledge about the mesoscopic physics on granular scales, see
eg.~\cite{roux}. And they were especially able to distinguish between two types of strain, I
and II, complete with two regimes in which either dominates. However, although
type I strain may clearly be identified as our state variable $u_{ij}$, one needs to be aware that regime~I is not necessarily {\em quasi-static}, or {quasi-elastic} as considered in Sec~\ref{quasi elastic motion}. The difference is: The relaxation term may be temporarily small at hypoplastic shear rates, say because $u_s$ or $\rho_{cp}-\rho$ are, see Eqs~(\ref{3b-3}, \ref{3b-4}), they do not stay small if one wanders in the variable space. At quasi-elastic rates, deformation are always free of dissipation.
\subsection{Stationary Elastic Solution, or the Critical State\label{critical state}}
\begin{figure}[b] \begin{center}
\includegraphics[scale=0.27]{pi-plane1}
\includegraphics[scale=0.2]{yield-cs-pai.eps}
\end{center}
\caption{\label{yield-cs-pai} Loci of static yield surface and the critical states, calculated employing the more general energy of Eq~(\ref{2b-6}). Left: in the space spanned by the three stress eigenvalues, $\sigma_1, \sigma_2, \sigma_3$; right: in the $\pi$-plane of constant pressure, $P\equiv\sigma_1+\sigma_2+\sigma_3$, where $\sqrt{2}\pi _{1} \equiv(\sigma _{3}-\sigma _{2})/P$,
$\sqrt{6}\pi _{2} \equiv(2\sigma _{1}-\sigma _{2}-\sigma _{3})/P$. }
\end{figure}
When there is complete compensation of the shear rate $\sim v_s$ and the relaxation
$\sim T_g$, the stationary solution of Eqs~(\ref{3b-3}) for the elastic strain
$u_{ij}$ holds. It is generally called the {\em critical state}, see~\cite{critState}, and may be considered ideally plastic, because a shear rate does not lead to a stress
increase. Setting $\partial_t\Delta, \partial_tu_s=0$ in Eqs~(\ref{2c-7xx},\ref{2c-9xx}), we obtain the expressions,
\begin{equation}\label{3b-3a} u_c=\frac{1-\alpha}{\lambda}\frac{v_s}{T_g}=\pm\frac{1-\alpha}{\Lambda},
\quad \frac{\Delta_c}{|u_c|}=\frac{\alpha_1}{\lambda_1}\frac{|v_s|}{T_g}=\frac{\alpha_1}{\Lambda_1}.
\end{equation}
From Eq~(\ref{2c-8}), the collinearity of the critical strain and rate, $u^*_{ij}|_c/|u_c|=v^*_{ij}/|v_s|$, is easy to see.
In the hypoplastic regime (for $h=1, \alpha=\bar\alpha,\alpha_1=\bar\alpha_1$), $u_c,\Delta_c$ depend only on the density and is rate-independent.
The critical stress is given by inserting $u_c,\Delta_c$ into Eqs~(\ref{sum stress hp}),
\begin{eqnarray}\label{3b-3b} P^c_\Delta&\equiv&P_\Delta(\Delta_c,
u_c)=\sqrt\Delta_c({\cal B}\Delta_c+{\cal A} {u_c^2}/{2\Delta_c}),
\\\label{3b-3c}
\pi_c&\equiv&\pi_s(\Delta_c, u_c)=-2{\cal A}\sqrt\Delta_c\, u_c,
\\\label{3b-3cA}
{P^c_\Delta}/{\pi_c}&=&({{\cal B}}/{2{\cal
A}}){\Delta_c}/{u_c}+{u_c}/{4\Delta_c},
\\\label{3b-4a}
P_c&=&(1-\bar\alpha)P^c_\Delta,\quad
\sigma_c=(1-\bar\alpha)\pi_c.
\end{eqnarray}
The critical ratio $\sigma_c/P_c$ -- same as the Coulomb yield of Eq~(\ref{2b-3}) -- is also frequently associated with a friction angle. Since one is relevant for vanishing $T_g\sim v_s^2\to0$, while the other requires an elevated $T_g\sim |v_s|$, it is appropriate to
identify one as the static friction angle, and the other as the
dynamic one. The dynamic friction angle is always smaller than
the static one, see Fig~\ref{yield-cs-pai}, because the critical state is elastic,
and must stay below Coulomb yield,
\begin{equation}\label{3b-5} \Lambda_1/\bar\alpha_1<\sqrt{2{\cal B/A}}.
\end{equation}
Textbooks on soil mechanics frequently state that the friction angle is
essentially independent of the density -- although they do not, as a rule,
distinguish between the dynamic and the static one, cf. Sec~\ref{yield
surfaces}. We assume, for lack of more discriminating information, that both are. Therefore, we take $\alpha_1\sim(\rho_{cp}-\rho)$, because $\Lambda_1$ also does, see Eq~(\ref{3b-4}).
Quite generally, we note that accepting the density dependence of Eqs~(\ref{density-dependence}), we have $\Delta_c, u_c$ being monotonically increasing functions of $1/(\rho_{cp}-\rho)$. The same holds for $P_c,\sigma_c\sim{\cal B}$, though ${\cal B}$'s density dependence make the increase slightly faster.
\subsection{Constitutive Relations\label{conrel}}
As discussed in the introduction, granular dynamics is frequently modeled employing the strategy of
{\em rational mechanics}, by postulating a function $\mathfrak{C}_{ij}$ -- of the stress $\sigma _{ij}$, strain rate $v_{k\ell}$, and density $\rho $ -- such that the constitutive relation, ${\partial}_{t}\sigma _{ij}=\mathfrak{C}_{ij}(\sigma_{ij}, v_{k\ell}, \rho)$ holds (where ${\partial}_{t}$ is to be replaced by an appropriate objective derivative more generally). It forms, together with the continuity equation $\partial _{t}\rho +\nabla _{i}\rho v_{i}=0$, momentum conservation, $\partial _{t}(\rho v_{i})+\nabla_{j}(\sigma _{ij}+\rho v_iv_j)=0$, a closed set of equations for $\sigma _{ij}$, the velocity $v_{i}$, and the density $\rho $ (or the void ratio $e$). Both hypoplasticity and barodesy considered below belong to this category. These models yield, in circumstances where they hold, a realistic account of the complex elasto-plastic motion, providing us with highly condensed and intelligently organized empirical data. This enables us to validate {\sc gsh} and reduce the latitude in specifying the energy and transport coefficients.
At the same time, one needs to be aware of their drawbacks, especially the hidden ones. First of all is the apparent freedom in fixing $\mathfrak C$ -- constrained only by the data one considers, not by energy conservation or entropy production that were crucial in deriving {\sc gsh}. This is what we believe the main reason why there are so many competing engineering models. As this liberty explodes when one includes gradient terms, most models refrain from the attempt to account for nonuniform situations, say elastic waves. Second, in dispensing with the variables $T_g$ and $u_{ij}$, and restricting the variables to $\sigma_{ij}, v_{k\ell}, \rho$, one reduces the model's range of validity and looses the benefit of $u_{ij}$'s simple behavior: First, the models of hypoplasticity and barodesy are valid only for $T_g\sim|v_s|$, so a $T_g$ that is either too small or oscillates too fast will invalidate them. Second, as the analytical solution of the approach to the critical state~\cite{critState} shows, considering $u_{ij}$ -- though it is not directly measurable -- is a highly simplifying intermediate step. The case for $u_{ij}$ is even stronger, when considering proportional paths and the barodesy model, see below.
\subsubsection{The Hypoplastic Model\label{Hypoplasticity}}
The {\em hypoplastic model} starts from the rate-independent
constitutive relation,
\begin{equation}\label{3b-1}
\partial_t\sigma_{ij}=H_{ijk\ell}v_{k\ell}+
\Lambda_{ij}\sqrt{v_s^2+\epsilon v_{\ell\ell}^2},
\end{equation}
postulated by Kolymbas~\cite{kolymbas1}, where
$H_{ijk\ell},\Lambda_{ij},\epsilon$ are (fairly involved) functions of the
stress and packing fraction. Incremental nonlinearity as discussed in Sec~\ref{Load and Unload} is also part of the postulate. The simulated granular response is realistic for deformations at constant or slowly changing rates.
{\sc gsh} reduces to the hypoplastic model in the hypoplastic regime, for $T_g\sim|v_s|$, $\alpha =\bar \alpha,\alpha_1=\bar\alpha_1$, $P_T, \eta_1T_g
v^0_{ij}\to0$. This is because $\sigma_{ij}=(1-\bar\alpha)\pi_{ij}$ of Eq~(\ref{sum stress hp}) is then, same as $\pi_{ij}$, a function of $u_{ij}, \rho$, and we may write $\partial_t\sigma_{mn}=({\partial\sigma_{mn}}/{\partial u_{ij}})\partial_tu_{ij}+
({\partial\sigma_{mn}}/{\partial\rho})\partial_t\rho$.
Replacing $\partial_t\rho$ with the first of Eq~(\ref{2c-1}),
$\partial_tu_{ij}$ with Eq~(\ref{2c-8}), using Eq~(\ref{TgVs}) to eliminate $T_g$, we arrive at an equation with the same structure
as Eq~(\ref{3b-1}). Our derived result for $H_{ijk\ell},\Lambda_{ij}$ is
different from the postulated engineering expressions, and somewhat simpler, but they yield very
similar {\em response ellipses}, see~\cite{granL3}. (Response ellipses are
the strain increments as the response of the system, given unit stress
increments in all directions starting from an arbitrary point in the stress
space, or vice versa, stress increments as the response for unit strain increments.)
\subsubsection{Proportional Paths and Barodesy}
Barodesy is a very recent model, again proposed by Kolymbas~\cite{barodesy}. As compared to hypoplasticity, it is more modular and better organized, with different parts in $\mathfrak C_{ij}$ taking care of specific aspects of granular deformation, especially that of {\em proportional paths}. We take {\sc p}$\varepsilon${\sc p} and {\sc p}$\sigma${\sc p} to denote, respectively, proportional strain and stress path. Their behavior is summed up by the Goldscheider rule ({\sc gr}): (1)~A {\sc p}$\varepsilon${\sc p} starting from the stress $\sigma_{ij}=0$ is associated with a {\sc p}$\sigma${\sc p}.
(2)~A {\sc p}$\varepsilon${\sc p} starting from $\sigma_{ij}\not=0$
leads asymptotically to the corresponding {\sc p}$\sigma${\sc p} obtained when starting at $\sigma_{ij}=0$.
(The initial value $\sigma_{ij}=0$ is a mathematical idealization, neither easily realized nor part of the empirical data that went into {\sc gr}. We take it {\em cum grano salis}.)
Explanation: Any constant strain rate $v_{ij}$ is a {\sc p}$\varepsilon${\sc p}. In the principal strain axes $(\varepsilon_1,\varepsilon_2,\varepsilon_3)$, a constant $v_{ij}$ means the system moves with a constant rate along its direction, with $\varepsilon_1/\varepsilon_2=v_1/v_2,\, \varepsilon_2/\varepsilon_3=v_2/v_3$ independent of time. What {\sc gr} states is that there exists an associated stress path that is also proportional, also a straight line in the principal stress space, that there are pairs of strain and stress path which are linked, and if the initial stress value is not on the right line, it will converge onto it.
If {\sc gsh} is as claimed a broad-ranged theory on granular behavior, we should be able to understand {\sc gr} with it, which is indeed the case. Given any constant rate $v_{ij}$, the elastic strain will -- irrespective of its initial value, relax into the stationary state of Eqs~(\ref{2c-7xx},\ref{2c-9xx}),
\begin{equation}
\label{eq74}
u_c=\frac{1-\alpha}{\Lambda},\quad \frac{\Delta_c}{u_c}=\frac{\alpha _1}{\Lambda _{1}}+\frac{1-\alpha}{u_c\Lambda_1}\frac{v_{\ell\ell}}{v_s},
\end{equation}
with ${u_{ij}^{\ast }|_c}/{u_c}={v_{ij}^{\ast }}/{v_s}$. Adding in the information from Eqs~(\ref{2b-2a},\ref{2b-2b}), we also find
\begin{equation}
{\sigma_{ij}^{\ast }}/{\sigma_s}(\rho)={v_{ij}^{\ast }}/{v_s}.
\end{equation}
If the strain path is isochoric, with $v_{\ell\ell}=0$ and $\rho=$ const, both the deviatoric strain and stress are dots that remain stationary and do not walk down a path as time progresses. Clearly, these are simply the ideally plastic, stationary, critical state~\cite{critState}. If $v_{\ell\ell}\not=0$ with the density $\rho[t]$ changing accordingly, ${u_{ij}^{\ast }|_c}$ and ${\sigma_{ij}^{\ast }}$ will walk down a straight line along ${v_{ij}^{\ast }}/{v_s}$, with a velocity determined, respectively, by $u_c(\rho[t])$ and $\sigma_s(\rho[t])$.
Given an initial strain deviating from that prescribed by Eq~(\ref{eq74}), $u_0\not=u_c,\Delta_0\not=\Delta_c$, Eqs~(\ref{2c-7xx},\ref{2c-9xx}) clearly state that the deviation will exponentially relax, until they vanish -- ie., the strain and the associated stress will converge onto the prescribed line. All this is very well, but {\sc gr} states that it is the total stress that walks down a straight line. With
\begin{equation}
\pi_{ij}=P_\Delta(\rho)[\delta_{ij}+(\pi_s/P_\Delta)v_{ij}^*/v_s],
\end{equation}
this fact clearly hinges on $(\pi_s/P_\Delta)$ -- a function of $\Delta/u_s$ [see Eq~(\ref{2b-3b})] -- not depending on the density. As long as $v_{\ell\ell}\ll v_s$, we have $\Delta_c/u_c\approx{\alpha _1}/{\Lambda _{1}}$, a combination that we did assume is density independent, see Eq~(\ref{density-dependence}), partially in anticipation of the fact that the friction angle of the critical state, a function of $(\pi_s/P)$, is independent of the density. And $v_{\ell\ell}/v_s$ must indeed remain small to avoid hitting either $\rho_{cp}$ or $\rho_{lp}$ too quickly.
In~\cite{GSH&Barodesy}, the results of {\sc gsh} are compared to that of barodesy, with mostly quantitative agreement. (The energy of Eq~(\ref{2b-6}) was employed there. So the results are more realistic.) When looking at $\mathfrak C_{ij}$, it is easy to grasp that the construction of a constitutive relation is only for someone with vast experience about granular media. That we could substitute this deep knowledge with the notions of {\em variable transient elasticity}, giving rise to a theory just as capable of accounting for elasto-plastic motion, is eye-opening. It suggests that sand, in its qualitative behavior, may be after all neither overly complicated, nor such a rebel against general principles.
\subsection{Elastic Waves\label{elastic waves}}
That elastic waves propagate in granular media~\cite{jia1999,jia2004} is an important fact, because it is an unambiguous proof that granular media possess an
elastic regime, and behave as elastic media in certain parameter ranges. Experimental exploration of the elastic to plastic transition would be equally crucial, and elastic waves remain a useful tool for this purpose.
There is a wide-spread believe that small, quasi-static increments from any
equilibrium stress state is elastic, but large ones are plastic. As discussed in
Sec~\ref{3regimes}, this assumption is illogical, because a large increment is
the sum of small ones. In {\sc gsh}, the parameter that sets the boundary between elastic and plastic regime is the granular temperature $T_g$. We have
quasi-elastic regime for vanishing $T_g\sim v_s^2$, and the
hypoplastic one for elevated $T_g\sim v_s$.
A perturbation in the elastic strain or stress propagate as a wave only in the quasi-elastic regime, while it diffuses in the hypoplastic one.
More specifically, we shall derive a telegraph equation from {\sc gsh}, with a quantity $\sim T_g$ taking on the role of the electric resistance~\cite{zhang2012}. It defines a characteristic frequency $\omega_0\sim T_g$, such that elastic perturbations of the frequency $\omega$ diffuse for $\omega\ll\omega_0$, and propagate for $\omega\gg\omega_0$. In the quasi-elastic regime, $\omega_0\to0$, and all perturbations propagate. In the hypoplastic regime, when $T_g$ gets elevated, so does $\omega_0$, pushing the propagating range to ever higher frequencies. Eventually, the associated wave length become comparable to the granular diameter, exceeding {\sc gsh}'s range of validity.
To derive the telegraph equation, we start with two basic equations of {\sc gsh}, Eqs~(\ref{2c-8},\ref{2c-1}),
\begin{eqnarray}
\rho\partial_tv_i-(1-\alpha)\nabla_mK_{imkl}u^*_{kl}=0, \quad
\partial_tu^*_{ij}-(1-\alpha)v^0_{ij}=-\lambda T_gu^*_{ij}, \label{ew2}
\end{eqnarray}
where $K_{imkl}\equiv-\partial^2w/\partial u_{im}\partial u_{kl}$. (For simplicity, we concentrate on shear waves, assuming $v_{\ell\ell}\equiv0 $.)
For $T_g\to0$, both plastic terms $\lambda T_gu^*_{ij}$ and $\alpha\sim T_g$ are negligibly small, such that these two equations reproduce conventional elasticity theory. The variation of wave velocities $c$ with stress and density is then easily calculated, because $c^2$ is given by the eigenvalues of the matrix $K_{imnj}q_mq_n/(\rho q^2)$ ($q_m$ is the wave vector). The result~\cite{ge4} agrees well with observations~\cite{jia2009}.
There are two ways to crank up $T_g$. First is to introduce an ambient temperature, such as by tapping or a remote shear band, second is to increase the amplitude of the wave mode, because its own shear rate also creates $T_g$. The granular temperature has a characteristic time $\tau_T=b\rho/\gamma_1$, see Eq~(\ref{sum T_g}), that is of order $10^{-3}$~s in dense media. For simplicity, we assume that the wave mode's frequency is much larger than $1/\tau_T$ , such that $T_g$ and $\alpha(T_g)$ are essentially constant. This implies
\begin{eqnarray}\label{ew3}
(\partial^2_t+\lambda T_g\partial_t)\,u^*_{ij}={\textstyle\frac12}(1-\alpha)^2\times
\nabla_m[K_{imkl}\nabla_ju^*_{kl}+K_{jmkl}\nabla_iu^*_{kl}].
\nonumber
\end{eqnarray}
Concentrating on one wave mode propagating along $x$, with $c_{\rm qs}$ the quasi-elastic velocity and $\bar u\sim e^{iqx-i\omega t}$ the amplitude of the associated eigenvector, we obtain the telegraph equation,
\begin{equation}\label{ew4}
(\partial^2_t+\lambda T_g\partial_t)\,\bar u= (1-\alpha)^2c_{\rm qs}^2\nabla_x^2\,\bar u
\equiv c^2\nabla_x^2\,\bar u.
\end{equation}
(The coefficient $\alpha$ accounts for the fact that granular contacts soften with $T_g$, and the effective elastic stiffness decreases by $(1-\alpha)^2$. In the language of electromagnetism,
$(1-\alpha)^{-2}$ is a dielectric permeability.)
Inserting $\bar u\sim e^{iqx-i\omega t}$ into Eq~(\ref{ew4}), we find
\begin{equation}
c^2 q^2={\omega^2+i\omega\lambda T_g},
\end{equation}
implying diffusion for the low frequency limit, $\omega\ll\lambda T_g$,
\begin{equation}
q\approx\pm\frac{\sqrt{\omega\lambda T_g}}c \, \frac{1+i}{\sqrt2},
\end{equation}
and propagation for the high-frequency limit, $\omega\gg\lambda T_g$,
\begin{eqnarray}
q\approx\pm\frac\omega c\left(1+i\,\frac{\lambda T_g}{2\omega}
\right),
\quad
\bar u\sim\exp{\left[-i\omega\left(t\mp\frac xc\right)
t\mp x\frac{\lambda T_g}{2 c}\right]}.
\end{eqnarray}
The first term in the square bracket accounts for wave propagation, the second a decay length $2c/\lambda T_g$, which is frequency-independent if $T_g$ is an ambient temperature. If $T_g$ is produced by the elastic wave itself, it varies between $T_g\sim v_s^2\sim\omega^2q^2\sim\omega^4$ and $T_g\sim v_s\sim\omega^2$ depending on the amplitude, and the decay length is strongly frequency dependent.
A brief wave pulse, arbitrarily strong, can always propagate through granular media if its duration is too brief to excite sufficient $T_g$ for the system to enter the hypoplastic regime. The duration must be much smaller than the characteristic time $b\rho/\gamma$ of $T_g$, see Eq~(\ref{sum T_g}).
\section{Conclusions}
{\sc gsh} is derived employing the hydrodynamic approach, starting from two assumptions about granular media's basic physics: {\it variable transient elasticity} and {\it two-stage irreversibility}. Because of the many constraints this approach provides, {\sc gsh} should be a valid broad-ranged theory, from statics to fast dense flow, if these assumptions are appropriate
{\it Variable transient elasticity} prescribes relaxation of the elastic stress with a rate $\sim T_g$, with full elasticity restored at equilibrium, for $T_g=0$. And indeed, this leads to results in several static geometries, including silos, sand pile and a layer subject to a point load, that agree well with data. Same is true for the incremental stress-strain relation and velocities of sound waves, both calculated setting $T_g=0$. Finally, we conclude that the quasi-static motion in {\sc gsh} is quasi-elastic, a visit of consecutive elastic states at $T_g=0$.
{\it Two-stage irreversibility} defines what the granular temperature $T_g$ is, and provides a relation between $T_g$ and the shear rate $\dot\gamma$, in the stationary state in which $T_g$ is a constant. Given by transport coefficients (the functional dependence of which is an input), this relation is not completely fixed. Assuming the simplest dependence, we have $T_g\sim\dot\gamma$, implying more stress relaxation the faster the shear motion is. This is the physics of rate-independence: Because the same shear motion also deforms the grains and builds up the elastic stress, a motion double as fast with twice the relaxation rate leads to the same stress. At given $\dot\gamma$, the elastic stress relaxes toward its stationary solution, in which the elastic deformation and dissipative relaxation cancel, implying a constant stress. This is the critical state.
However, a rate-independent ratio between the elastic deformation and dissipative relaxation means that the latter does not vanish more quickly than the former for $\dot\gamma\to0$, implying the complete lack of quasi-static motion, which is, of course, a fairly general phenomenon. Although there is as yet not much direct experimental evidence for its existence, we note that it may be restored by changing the transport coefficients slightly, such that $T_g\sim\dot\gamma$ goes smoothly over to $T_g\sim\dot\gamma^2$ for $\dot\gamma\to0$, because a quadratically small $T_g$ implies a dissipation that vanishes more quickly than the elastic motion.
As a result, {\sc gsh} has three rate regimes, given by:
\begin{itemize}
\item The {\em quasi-elastic regime} of vanishing shear rates $\dot\gamma$, possibly below $10^{-5}$~s$^{-1}$, with a quadratically vanishing granular temperature, $T_g\sim\dot\gamma^2\to0$. The stress $\sigma_{ij}=\pi_{ij}$ is purely elastic, as given by Eq~(\ref{1-1}).
This regime is admittedly difficult to observe. Some possible reasons are discussed in Sec~\ref{aeipt}, with suggestions in~\ref{soft springs} on how to overcome them.
\item The {\em hypoplastic regime} of low shear rates, possibly between $10^{-3}$ and 1~s$^{-1}$, where the engineering theories such as the hypoplastic model~\cite{kolymbas1} holds. The stress $\sigma_{ij}=(1-\alpha)\pi_{ij}$ is still elastic, but softer
by the factor $1-\alpha$, typically between 0.2 and 0.3. Granular temperature is more elevated, allowing stress relaxation. Rate-independence is a result of $T_g\sim\dot\gamma$.
As we have seen, three simple scalar equations are sufficient to account for general elasto-plastic motion, including especially load-unload behavior, Sec~\ref{Load and Unload}; and the approach to the critical state, Sec~\ref{critical state}. They were also used for a successful comparison to the hypoplastic and barodesy model, Sec~\ref{conrel}; and for the damping of elastic waves, Sec~\ref{elastic waves}.
This regime is frequently termed the quasi-static one, because it lacks inertial effects, is rate-independent, and the even slower quasi-elastic regime is hard to
observe. We note in Sec~\ref{yield surfaces} and~\ref{3regimes} that the hypoplastic regime, characterized by stress relaxation, is strongly dissipative.
\item The {\em rapid flow regime}, for shear rates well above 1 s$^{-1}$. We still have $T_g\sim\dot\gamma$, but it is no longer small. Therefore, the $T_g$-generated, seismic pressure $P_T\sim T_g^2\sim\dot\gamma^2$ and the viscous shear stress $\sigma_s\sim T_g\dot\gamma\sim\dot\gamma^2$ become significant and compete with the elastic contribution $\pi_{ij}$. This is where the MIDI model and Bagnold flow hold. As both the pressure and the shear stress may be written as $e_1+e_2\dot\gamma^2$, where $e_1$ is the elastic, and $e_2$ the seismic, or viscous, contributions, we have a quadratic dependence of the Bagnold flow for $e_2\dot\gamma^2\gg e_1$, and hypoplastic rate-independence for $e_2\dot\gamma^2\ll e_1$.
This rate regime has already been considered in~\cite{p&g2013}.
\end{itemize}
Finally, a summary of some frequently used quantities, for which physics and engineering textbooks employ rather different notations,
first a general tensor, say $v_{ij}$. We take $v_{\ell\ell}$ as its trace, $v_{ij}^*$ as its traceless part, with $v_s\equiv v_{ij}^*v_{ij}^*$ as the second invariant.
Then stresses and strains: total or Cauchy stress: $\sigma_{ij}$, the elastic stress: $\pi_{ij}$, with $\sigma_s,\pi_s$ as defined above. The elastic strain: $u_{ij}$, with $\Delta\equiv- u_{\ell\ell}$ and $u_s\equiv u_{ij}^*u_{ij}^*$, is defined around Eq~(\ref{1-1}). The strain rate (frequently denoted as $\dot\epsilon_{ij}$) is taken as $v_{ij}\equiv\frac12(\nabla_iv_j+\nabla_jv_i)$, and the scalar shear rate $v_s\equiv v_{ij}^*v_{ij}^*$ (or interchangeably, $\dot\gamma$).
The granular temperature is $T_g$, note the energy is $\sim T_g^2$, see Sec~\ref{significanceT_g}.
\noindent{\bf Acknowledgment:} We thank Itai Einaf for a critical reading of the manuscript and many help- and insightful comments.
|
\section{Introduction and main results}
Let $K$ be a closed subset of the complex plane $\mathbb{C}$ and $\nu$ a measure on $K$. For $k=1,2,...$, we will be concerned with the following ensemble of probability measures on $K^{k+1}$:
$$\frac{1}{Z_k}|VDM(z_0,...,z_k)|^{2\beta}\exp \Big(-k[Q(z_0)+\cdots + Q(z_k)]\Big)d\nu (z_0)...d\nu (z_k).$$
Here \begin{itemize}
\item $Z_k$ is a normalization constant;
\item $VDM(z_0,...,z_k)=\prod_{0\leq i<j\leq k}(z_j-z_i)$
is the usual Vandermonde determinant;
\item $Q:K\rightarrow (-\infty,+\infty]$ is a lower semicontinuous function; and
\item $\beta >0$.
\end{itemize}
\noindent These probability measures occur in random matrix theory as the joint probability of eigenvalues and also in the theory of Coulomb gases, where $z_0,...,z_k$ are the positions of particles. They have been extensively studied (cf., \cite{AGZ} or \cite{hpbook}).
We will deal with the global behavior as $k\rightarrow \infty$. In particular, we study the almost sure convergence of the empirical measure of a random point $\frac{1}{k+1}\sum_{i=0}^k\delta_{z_i}$ to the equilibrium measure given by the unique minimizer of the weighted energy functional; i.e.,
$$\inf \{I^Q(\mu): \ \mu\in \mathcal{M}(K)\}$$
where $\mathcal{M}(K)$ are the probability measures on $K$ and
$$I^Q(\mu)=\int_K\int_K\log \frac{1}{|z-t|w(z)w(t)}d\mu(z)d\mu(t),$$ with $w(z)=\exp(-Q(z))$. We will also establish a large deviation principle (LDP).
Ben Arous and Guionnet \cite{BG}, building on work of Voiculescu, first proved a large deviation principle for the Gaussian Unitary Ensemble. This was subsequently extended to general unitary invariant ensembles. Hiai and Petz \cite{hpbook} extended these methods to
the complex plane and {\it strongly admissible} (see Definition \ref{admit}) continuous weights $Q$. In these settings, $\nu$ was taken to be Lebesgue measure.
More recently, the case of {\it weakly admissible weights} (see Definition \ref{admit}) on unbounded subsets of the plane was studied in \cite{H} and the existence of a unique minimizer of the weighted energy functional (which in this case may not have compact support) was established. In \cite{[ESS]} a large deviation principle was established for $Q$ weakly admissible, continuous on $\mathbb{R}$ or $\mathbb{C}$ and $\nu$ the Lebesgue measure. Such weights occur in certain ensembles (see \cite{[ESS]}, the Cauchy ensemble) and in certain vector energy problems (see \cite{H}).
In this paper we will systematically develop the case when $Q$ is lower semicontinuous, weakly admissible and $\nu$ is more general than Lebesgue measure. We will use the methods of \cite{VELD} which first of all give the almost sure convergence of the empirical measure of a random point, and, subsequently, we obtain a large deviation principle.
The paper is organized as follows: in the next section, we give some basic results on logarithmic potential theory in $\mathbb{R}^n$ valid for $n\geq 2$. Using the results in $\mathbb{R}^{3}$ together with inverse stereographic projection from a two-dimensional sphere ${\bf S}$ to the complex plane, in section 3 we readily extend some classical potential-theoretic results valid for compact subsets of $\mathbb{C}$ to closed, unbounded sets with weakly admissible weights.
In sections 4-7, we return to the setting of compact sets $K$ in $\mathbb{R}^n$ and admissible weights (see Definition \ref{realad}; such weights need only be lower semicontinuous). Corollary \ref{416} establishes the almost sure convergence of the empirical measure of a random point to the equilibrium measure in this setting, for appropriate measures $\nu$.
Our next goal is to show that two functionals $\underline J$ and $\overline J$ on the space $\mathcal M(K)$ of probability measures coincide. These functionals are defined as asymptotic $L^2(\nu)-$averages of Vandermonde determinants with respect to a {\it Bernstein-Markov} measure $\nu$ on $K$. As in previous work (cf., \cite{VELD} and \cite{PELD}), weighted versions of these functionals are of essential use (Theorem \ref{rel-J-E}). This equality immediately yields a large deviation principle in this $\mathbb{R}^n$ setting, Theorem \ref{ldp}, in which the rate function is given in terms of the weighted energy functional.
In section 8, we deal with compact subsets of the sphere in $\mathbb{R}^3$ and measures of infinite mass, again establishing a LDP (Theorem \ref{ldp-S}). Measures of infinite mass arise as the push-forward of measures on unbounded subsets of the plane under stereographic projection.
Our ultimate goal, achieved in sections 9 and 10, is to utilize the $\mathbb{R}^n$ result to prove the analogous equality of the appropriate $J-$functionals for probability measures on closed, unbounded sets in $\mathbb{C}$ allowing weakly admissible weights and very general measures of infinite mass (Theorem \ref{85}). Then, via a contraction principle, we obtain an LDP (Theorem \ref{ldp2}):
\begin{theorem} Let $K\subset \mathbb{C}$ be closed, and let $Q$ be a weakly admissible weight on $K$. Assume $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property (\ref{unboundedBM}).
If $\nu$ has finite mass, assume that $(K,\nu)$ satisfies a strong Bernstein-Markov property while if $\nu$ has infinite mass in a neighborhood of infinity, assume that (\ref{cond-nu}) and (\ref{cond-eps}) are satisfied for some function $\epsilon(z)$.
Define a sequence $\{\sigma_k\}$ of probability measures on $\mathcal M(K)$ by
$$\sigma_k(G) = \frac{1}{Z_k}\int_{\tilde G_k} |VDM(z_0,..., z_k)|^2\prod_{i=0}^k e^{-2kQ(z_i)}\prod_{i=0}^k d\nu(z_i)$$
where $\tilde G _k= \{ (z_0,..., z_k)\in K^{k+1} : \frac{1}{k+1}\sum \delta_{z_i}\in G\}$. Then $\{\sigma_k\}$ satisfies a {\bf large deviation principle} with speed $k^{2}$ and good rate function $\mathcal I:=\mathcal I_{K,Q}$ where, for $\mu \in \mathcal M(K)$,
$$
\mathcal I(\mu)=I^Q(\mu)-I^Q(\mu_{K,Q}).
$$
\end{theorem}
In section 11 we extend this result to the case of general $\beta$ (Theorem \ref{ldp-free}). Our results include the LDP for a number of ensembles occurring in the literature (see Remark \ref{appldp}) and also the results of Hardy \cite{H} for Lebesgue measure in $\mathbb{R}$ or $\mathbb{C}$ (see the discussion after Theorem \ref{ldp-free}). The idea of using inverse stereographic projection and working in $\mathbb{R}^3$ to obtain an LDP for unbounded sets in $\mathbb{C}$ comes from this work.
\section{Logarithmic Potential Theory in $\mathbb{R}^n$}
Let $K\subset \mathbb{R}^n$ be compact and let $\mathcal M(K)$ be the set of probability measures on $K$ endowed with the topology of weak convergence from duality with continuous functions. We consider the logarithmic energy minimization problem:
$$\inf_{\mu\in \mathcal M(K)}I(\mu)$$
where
$$I(\mu):=\int_K \int_K \log \frac{1}{|x-y|}d\mu(x) d\mu(y)$$
is the logarithmic energy of $\mu$. We will say that $K$ is {\it log-polar} if $I(\mu)=\infty$ for all $\mu \in \mathcal M(K)$. It is known that any compact set of positive Hausdorff dimension is non log-polar \cite{Ca}. For a Borel set $E\subset \mathbb{R}^n$ we will say $E$ is log-polar if every compact subset of $E$ is log-polar. We write
$$U^{\mu}(x):= \int_K \log \frac{1}{|x-y|}d\mu(y)$$
for the logarithmic potential of $\mu$. It is locally integrable and superharmonic in all of $\mathbb{R}^{n}$.
We gather known results about logarithmic potentials in $\mathbb{R}^{n}$ in the next theorem.
\begin{theorem}\label{basics}
The following results, whose precise statements can be found in \cite{ST} for logarithmic potentials in $\mathbb{C}=\mathbb{R}^{2}$, hold true for logarithmic potentials in $\mathbb{R}^n$, $n\geq2$:
\begin{enumerate}[noitemsep,nolistsep]
\item
for $\mu=\mu_{1}-\mu_{2}$ a signed measure with compact support and total mass zero, with $\mu_{1}$ and $\mu_{2}$ of finite energies, $I(\mu)$ is nonnegative
and is zero if and only if $\mu_{1}=\mu_{2}$.
\item
principle of descent and lower envelope theorem (with ``q.e.'' in the latter replaced by ``off of a log-polar set'');
\item
maximum principle;
\item
continuity principle.
\end{enumerate}
\end{theorem}
\begin{proof}
The version of Item 1. in $\mathbb{C}$ is \cite[Lemma 1.8]{ST}. In $\mathbb{R}^{n}$, it follows from \cite[Theorem 2.5]{CKL}. One checks that the proofs of the principle of descent and lower envelope theorem in $\mathbb{C}$, Theorems I.6.8. and I.6.9. of \cite{ST}, are valid in $\mathbb{R}^{n}$. Items 3. and 4. are Theorems 5.2 and 5.1 of \cite{HaK}. A maximum principle restricted to the two dimensional sphere also follows as a particular case of \cite[Theorem 5]{BDS}.
\end{proof}
We will need to work in a weighted setting. We caution the reader that, unlike the setting of compact sets in $\mathbb{R}^n$ where we have a single notion of admissibility for a weight function, when we work on unbounded sets in $\mathbb{C}$ in the next section we will have several different notions.
\begin{definition}\label{realad}
Given a compact set $K\subset \mathbb{R}^n$ which is not log-polar, let $Q$ be a lower semicontinuous function on $K$ with $\{x\in K: Q(x) <\infty\}$ not log-polar. We call such $Q$ {\it admissible} and write $Q\in \AA(K)$. We define $w(x):=e^{-Q(x)}$.
\end{definition}
\noindent We refer to either $Q$ or $w$ as the {\it weight}; in \cite{ST} this terminology is reserved for $w$.
We consider now the weighted logarithmic energy minimization problem:
$$\inf I^Q(\mu),\quad\mu\in \mathcal M(K),$$
where
$$I^Q(\mu):=\int_K \int_K \log \frac{1}{|x-y|w(x)w(y)}d\mu(x) d\mu(y)=I(\mu)+2\int_KQ(x)d\mu(x).$$ Following the arguments on pp. 27-33 in \cite{ST}, we have the following.
\begin{theorem}\label{frost}
For $K\subset \mathbb{R}^n$ compact and not log-polar, and $Q\in \AA(K)$,
\begin{enumerate}[noitemsep,nolistsep]
\item $V_w:=\inf_{\mu\in \mathcal M(K)}I^Q(\mu)$ is finite;
\item there exists a unique weighted equilibrium measure $\mu_{K,Q} \in \mathcal M(K)$ with $I^Q(\mu_{K,Q})=V_w$ and the logarithmic energy $I(\mu_{K,Q})$ is finite;
\item the support $S_w:=$supp$(\mu_{K,Q})$ is contained in $\{x\in K: Q(x) <\infty\}$ and $S_w$ is not log-polar;
\item Let $F_w:=V_w -\int_K Q(x)d\mu_{K,Q}(x)$ denote the (finite) Robin constant. Then
\begin{align*}
U^{\mu_{K,Q}}(x)+Q(x) & \geq F_w\text{ on }K\setminus P\text{ where }P
\text{ is log-polar (possibly empty);}\\
U^{\mu_{K,Q}}(x)+Q(x) & \leq F_w\text{ for all }x\in S_w.
\end{align*}
\end{enumerate}
\end{theorem}
\begin{remark} In the proof of the Frostman-type property 4. in \cite{ST}, one simply replaces ``q.e.'' -- off of a set of positive logarithmic capacity in $\mathbb{C}$ -- by ``off of a log-polar set'' as the essential property used is the existence of a measure of finite logarithmic energy on a compact subset of a set of positive logarithmic capacity in $\mathbb{C}$. We should mention that, in the unweighted case, the existence portion of 2. and property 4. can be found in \cite{HaK}, Theorems 5.4 and 5.8.
\end{remark}
Next we discretize: for $k\geq 2$, let the $k$-th weighted diameter $\delta_{k}^{Q}(K)$ be defined by
$$\delta_k^Q(K):=\sup_{x_1,...,x_k\in K}
|VDM_k^Q(x_1,...,x_k)|^{2/k(k-1)},$$
where $|VDM_k^Q(x_1,...,x_k)|$ denotes the weighted Vandermonde:
\begin{align}\notag
|VDM_k^Q(x_1,...,x_k)| & :=\prod_{i<j} |x_i-x_j|w(x_i)w(x_j)= \prod_{i<j} |x_i-x_j|\prod_{j=1}^kw(x_j)^{k-1}\\\label{VDM-Q}
& =:|VDM_k(x_1,...,x_k)|\cdot \prod_{j=1}^kw(x_j)^{k-1}.
\end{align}
By the uppersemicontinuity of $(x_1,...,x_k)\to \prod_{i<j} |x_i-x_j|w(x_i)w(x_j)$ on $K^k$ the supremum is attained; we call any collection of $k$ points of K at which the maximum is
attained {\it weighted Fekete points} of order $k$ for $K,Q$. Following the proofs of Propositions 3.1--3.3 of \cite[Section 3]{VELD} we may derive similar results in $\mathbb{R}^{n}$.
\begin{theorem} \label{sec3} Given $K\subset \mathbb{R}^n$ compact and not log-polar, and $Q\in \AA(K)$,
\begin{enumerate}
\item if $\{\mu_k=\frac{1}{k}\sum_{j=1}^k\delta_{x_j^{(k)}}\}\subset \mathcal M(K)$ converge weakly to $\mu\in \mathcal M(K)$, then
\begin{equation}\label{upboundVDM}
\limsup_{k\to \infty} |VDM_k^Q(x_1^{(k)},...,x_k^{(k)})|^{2/k(k-1)}\leq \exp{(-I^Q(\mu))};
\end{equation}
\item we have
$$\delta^Q(K):=\lim_{k\to \infty} \delta_k^Q(K)=\exp{(-V_w)};$$
\item if $\{x_j^{(k)}\}_{j=1,...,k; \ k=2,3,...}\subset K$ and
$$\lim_{k\to \infty} |VDM_k^Q(x_1^{(k)},...,x_k^{(k)})|^{2/k(k-1)}= \exp{(-V_w)}$$
then
$$\mu_k=\frac{1}{k}\sum_{j=1}^k\delta_{x_j^{(k)}}\to \mu_{K,Q} \ \hbox{weakly}.$$
\end{enumerate}
\end{theorem}
\section{Weighted potential theory on unbounded sets in $\mathbb{C}$}
We use the previous results in $\mathbb{R}^{3}$ and the inverse stereographic projection from the two-dimensional sphere to $\mathbb{C}$ to extend classical results concerning potential theory on compact subsets of $\mathbb{C}$ to unbounded closed sets with weakly admissible weights. Some of these results already appeared in the literature, see, e.g., \cite{HK,Sim}.
Thus let $K\subset \mathbb{C}$ be closed and unbounded. We consider three types of admissibility for weight functions on $K$.
\begin{definition} \label{admit} Let $Q$ be a lower semicontinuous function on $K$ with $\{z\in K: Q(z)< \infty\}$ a nonpolar subset of $\mathbb{C}$ (equivalently a non log-polar subset of $\mathbb{R}^2$). We say $Q$ is
\begin{enumerate}
\item {\sl weakly admissible} if there exists $M\in(-\infty,\infty)$ such that
\begin{equation}\label{weak-adm}
\liminf_{z\in K, \ |z|\to \infty}\bigl(Q(z)-\log |z|\bigr) =M.
\end{equation}
\item {\sl admissible} if $\liminf_{z\in K, \ |z|\to +\infty}\bigl(Q(z)-\log |z|\bigr) = +\infty$.
\item {\sl strongly admissible} if for some $\epsilon >0$, there exists $R> 0$ with $Q(z) > (1+\epsilon)\log |z|$ for $z\in K$ and $|z|>R$.
\end{enumerate}
\end{definition}
\noindent Examples of weakly admissible weights arise from logarithmic potentials: if $\mu$ is a probability measure on $\mathbb{C}$ such that
$U^{\mu}$ is continuous, then $Q=-U^{\mu}$ is weakly admissible on $K=\mathbb{C}$.
We assume now that $Q$ is weakly admissible. We consider the inverse stereographic projection $T:\mathbb{C} \cup \{\infty\}\to {\bf S}$ where ${\bf S}$ is the sphere in $\mathbb{R}^3$ centered in $(0,0,1/2)$ of radius $1/2$. It is defined by
\begin{equation} \label{stereo} T(z)=\left(\frac{\Re(z)}{1+|z|^{2}},\frac{\Im(z)}{1+|z|^{2}},\frac{|z|^{2}}{1+|z|^{2}}\right),\quad z\in\mathbb{C}\end{equation}
and $T(\infty)=P_0,$ where $P_0=(0,0,1)$ denotes the ``north pole'' of ${\bf S}$.
The map $T$ is a homeomorphism with
\begin{equation}\label{rel-dist}
|T(z)-T(u)|=\frac{|z-u|}{\sqrt{1+|z|^{2}}\sqrt{1+|u|^{2}}},\quad z,u\in\mathbb{C},
\end{equation}
where $|\cdot|$ denotes the Euclidean distance.
For $\nu$ a positive Borel measure supported on $K$, not necessarily finite, we denote by $T_{*}\nu$ its push-forward by $T$, that is, the measure on $T(K)$ such that
$$\int_{T(K)}f(x)dT_{*}\nu(x)=\int_{K}f(T(z))d\nu(z),$$
for any Borel function $f$ on $T(K)$.
Lemma 2.1 in \cite{H} shows that the map
$$T_*:~\mathcal M(K)\to\mathcal M(T(K)),$$
is a homeomorphism from ${\mathcal M}(K)$ to the subset of ${\mathcal M}(T(K))$ of measures which put no mass at the north pole $P_0$ of ${\bf S}$.
Here, $\mathcal M(K)$ and $\mathcal M(T(K))$ are endowed with the topology of weak convergence. This is the topology coming from duality with bounded, continuous functions. On $K$, it suffices to consider bounded, continuous functions $f:K\to \mathbb{C}$ such that $\lim_{|z|\to \infty} f(z)$ exists. This follows from the correspondence of ${\mathcal M}(K)$ with the measures in ${\mathcal M}(T(K))$ putting no mass at $P_0$.
When the support of a measure $\mu\in\mathcal M(K)$ is unbounded, its potential
$$U^{\mu}(z)=\int\log\frac{1}{|z-t|}d\mu(t),\qquad z\in\mathbb{C}$$
is not always well-defined. However, the following lemma
holds true.
\begin{lemma}\label{pot-unbdd}
If there exists a $z_{0}\in\mathbb{C}$ with $U^{\mu}(z_{0})>-\infty$ then
\begin{equation}\label{cond-log-mu}
\int\log(1+|t|)d\mu(t)<\infty,
\end{equation}
which implies that $U^{\mu}(z)$ is well-defined as a function on $\mathbb{C}$ with values in $(-\infty,\infty]$. Moreover, $U^{\mu}(z)$ is then superharmonic and
$$-U^{\mu}(z)\leq\log(1+|z|)+\int\log(1+|t|)d\mu(t).$$
\end{lemma}
\begin{proof}
We only prove that, under assumption (\ref{cond-log-mu}), the potential $U^{\mu}(z)$ is superharmonic. This follows e.g. from the fact that
$$U^{\mu}(z)=\int\log\frac{1+|t|}{|z-t|}d\mu(t)-\int\log(1+|t|)d\mu(t),$$
and the first integral on the right-hand side is superharmonic with respect to $z$, see \cite[Theorem 2.4.8]{R}.
\end{proof}
Logarithmic potentials on $\mathbb{C}$ and on the sphere ${\bf S}$ correspond by the relation
\begin{equation}\label{pot-C-S}
U^{\mu}(z)=U^{T_{*}\mu}(Tz)-\frac12\log(1+|z|^{2})-\frac12\int\log(1+|t|^{2})d\mu(t),\qquad z\in\mathbb{C}.
\end{equation}
The weighted logarithmic energy of a measure $\mu\in\mathcal M(K)$ is defined as
\begin{equation}\label{enerK}
I^Q(\mu)=\int_{K}\int_{K}\log\frac{1}{|z-t|w(z)w(t)}d\mu(z)d\mu(t)= I(\mu) +2\int_KQd\mu,
\end{equation}
where $w=e^{-Q}$. The double integral is always well-defined. Indeed it follows from the upper semicontinuity of $w$ and (\ref{weak-adm}) that the integrand is bounded below. On the contrary, the second expression has a meaning only if $I(\mu)>-\infty$ which is not necessarily true. Another equivalent way to define $I^{Q}(\mu)$, which is always valid, is by using the map $T$ as was done in \cite{HK}. Here, one identifies $I^{Q}(\mu)$ with the weighted logarithmic energy of the measure $T_{*}\mu\in\mathcal M(T(K))$,
\begin{equation}\label{enerTK}
I^{\tilde Q}(T_* \mu)= \int_{T(K)}\int_{T(K)}\log {1\over |x-y|}dT_* \mu(x) d T_* \mu(y) + 2\int_{T(K)} \tilde Q(x)dT_*\mu(x),
\end{equation}
where
\begin{equation}\label{tildeq} \tilde Q(T(z))=Q(z) -{1\over 2} \log (1+|z|^2).\end{equation}
To define $\tilde Q$ on the whole of $T(K)$ we set $\tilde Q(P_0)=M$, so that $\tilde Q$ becomes lower semicontinuous, and we get a correspondence between {\it weakly admissible weights on the closed set $K$ in $\mathbb{C}$} and {\it admissible weights on the compact set $T(K)\subset {\bf S}$ in $\mathbb{R}^{3}$}.
\begin{lemma}
A closed subset $K\subset\mathbb{C}$ is polar if and only if $T(K)\subset {\bf S}$ is log-polar.
\end{lemma}
\begin{proof}
We have
$$\int_{T(K)}\int_{T(K)}\log {1\over |x-y|}dT_* \mu(x) d T_* \mu(y)=
\int_{K}\int_K\log\frac{1}{|z-t|}d\mu(z)d\mu(t)+\int_{K}\log(1+|t|^{2})d\mu(t).$$
Recall that a closed subset $K\subset\mathbb{C}$ is polar if $K_{r}=K\cap B(0,r)$ is polar for all $r>0$. Thus, if $K$ is nonpolar, there exists $r>0$ with $K_{r}$ nonpolar, that is, there is a measure $\mu_{r}$ supported on $K_{r}$ of finite energy. By the above equality, $T_{*}\mu_{r}$ is a measure on $T(K_{r})\subset T(K)$ of finite energy, so $T(K)$ is not log-polar. Conversely, if $K\subset\mathbb{C}$ is polar, for any finite measure $\mu$ of compact support in $K$ we have $I(\mu)=\infty$ in $\mathbb{C}$ (cf. \cite{R}) and thus $I(T_{*}\mu)=\infty$ in ${\bf S}$. Since any measure on $T(K)$ charging the north pole $P_{0}$ has infinite energy, it follows that $T(K)$ is log-polar.
\end{proof}
Theorem \ref{frost} asserts the existence and uniqueness of a weighted energy minimizing measure on a non log-polar compact subset of $\mathbb{R}^{n}$ with an admissible weight. Obviously, this minimizing measure does not charge any point of the set, in particular the north pole $P_0$ if it belongs to the set. Hence the above correspondence implies the following.
\begin{theorem} \label{Frost-C}
Let $K$ be a nonpolar closed subset of $\mathbb{C}$ and $Q$ a weakly admissible weight on $K$. Then,
\begin{enumerate}[noitemsep,nolistsep]
\item $V_w:=\inf_{\mu\in \mathcal M(K)}I^Q(\mu)$ is finite;
\item there exists a unique weighted equilibrium measure $\mu_{K,Q} \in \mathcal M(K)$ with $I^Q(\mu_{K,Q})=V_w$ and the logarithmic energy $I(\mu_{K,Q})$ is finite (hence $-\Delta U^{\mu_{K,Q}}=2\pi\mu_{K,Q}$);
\item the support $S_w:=$supp$(\mu_{K,Q})$ is contained in $\{x\in K: Q(z) <\infty\}$ and $S_w$ is not polar;
\item Let $F_w:=V_w -\int_K Q(z)d\mu_{K,Q}(z)$ denote the (finite) Robin constant. Then
\begin{align*}
U^{\mu_{K,Q}}(z)+Q(z) & \geq F_w\text{ on }K\setminus P\text{ where }P
\text{ is polar (possibly empty);}\\
U^{\mu_{K,Q}}(z)+Q(z) & \leq F_w\text{ for all }z\in S_w.
\end{align*}
\end{enumerate}
\end{theorem}
\begin{proof}
The above assertions correspond by the map $T$ to the similar assertions from Theorem \ref{frost} applied with a non log-polar compact subset of the sphere ${\bf S}$. Note that
$$T_{*}\mu_{K,Q}=\mu_{T(K),\tilde Q},\qquad V_{w}=V_{\tilde w}, \qquad F_{w}=F_{\tilde w}-
\frac12\int_{K}\log(1+|t|^{2})d\mu_{K,Q}(t),$$
where we have set $\tilde w:=e^{-\tilde Q}$.
The fact that $I(\mu_{K,Q})<\infty$ follows from
$$I(\mu_{K,Q})=I^{Q}(\mu_{K,Q})-2\int_{K}Qd\mu_{K,Q}
=I^{\tilde Q}(T_{*}\mu_{K,Q})-2\int_{K}Qd\mu_{K,Q},$$
where we know that $I^{\tilde Q}(T_{*}\mu_{K,Q})=V_{\tilde w}$ is finite and $Q$ is bounded below. If $S_{w}$ is compact, it is clear that the other inequality $-\infty<I(\mu_{K,Q})$ is satisfied. If $S_{w}$ is not compact,
$$I(\mu_{K,Q})=I(T_{*}\mu_{K,Q})-\int_{K}\log(1+|t|^{2})d\mu_{K,Q}(t),$$
so to verify $-\infty<I(\mu_{K,Q})$ it suffices to show that
$$\int_{K}\log(1+|t|^{2})d\mu_{K,Q}(t)<\infty,$$
which holds true since the equilibrium potential satisfies $U^{\mu_{K,Q}}>-\infty$, see Lemma \ref{pot-unbdd}.
\end{proof}
In particular, if $\mu$ is a probability measure on $\mathbb{C}$ such that
$U^{\mu}$ is continuous, taking $Q=-U^{\mu}$ on $K=\mathbb{C}$ we have $\mu=\mu_{K,Q}$ so that, in general, $\mu_{K,Q}$ need not have compact support. As specific examples, if $K =\mathbb{C}$ and $Q(z)=\frac{1}{2}\log (1+|z|^2)$, then $d\mu_{K,Q}=
\pi^{-1}(1+|z|^2)^{-2}dm(z)$ where $dm$ is Lebesgue measure, cf., Example 1.4 of \cite{H}. If $K=\mathbb{R}$ and $Q(x)=\frac{1}{2}\log (1+x^2)$, then $d\mu_{K,Q}=\pi^{-1}(1+x^2)^{-1}dx$, cf., Example 1.3 of \cite{H}. We mention that in \cite{HK}, existence and uniqueness of a minimizing measure was proven in the more general context of weakly admissible vector equilibrium problems.
Let $L(\mathbb{C})$ be the set of all subharmonic functions $u$ on $\mathbb{C}$ with the property that $$u(z)-\log |z|\text{ is bounded above as }|z| \to \infty.$$
We will need the following version of the domination principle, see \cite[Corollary A.2]{BL}.
\begin{proposition}
Let $u,v\in L(\mathbb{C})$ with $u(z)-v(z)$ bounded above as $|z|\to \infty$ and suppose $I(\Delta v)<\infty$. If $u\leq v$ a.e.-$\Delta v$, then
$u\leq v$ on $\mathbb{C}$.
\end{proposition}
\noindent Here, $\Delta v$ need not have compact support.
We can now state a weighted version of the Bernstein-Walsh lemma with a weakly admissible weight (see \cite[Theorem III.2.1]{ST} for the case of an admissible weight). This will be used in section 8 to get a version for appropriate polynomials on the sphere (Theorem \ref{BW-S}).
\begin{theorem}\label{BW-C}
Let $K$ be a closed nonpolar subset of $\mathbb{C}$ and $Q$ a weakly admissible weight on $K$. If $p_{k}$ is a polynomial of degree at most $k$ and
$$|p_{k}(z)e^{-kQ(z)}|\leq M\qquad\text{ for q.e. }z\in S_{w},$$
then
$$|p_{k}(z)|\leq M\exp(k(-U^{\mu_{K,Q}}(z)+F_{w})),\qquad z\in\mathbb{C},$$
and
$$|p_{k}(z)e^{-kQ(z)}|\leq M,\qquad \text{for q.e. }z\in K.$$
\end{theorem}
\begin{proof}
The function $g:=\log(|p_{k}|/M)/k$ belongs to $L(\mathbb{C})$ and
$$g(z)\leq Q(z)\leq -U^{\mu_{K,Q}}(z)+F_{w}\quad\text{for q.e. }z\in S_{w}.$$
By Lemma \ref{pot-unbdd}, $-U^{\mu_{K,Q}}+F_{w}$ also belongs to $L(\mathbb{C})$ and
$-\Delta U^{\mu_{K,Q}}=2\pi\mu_{K,Q}$ is of finite energy. Hence, by the above domination principle,
$$g(z)\leq -U^{\mu_{K,Q}}(z)+F_{w},\qquad z\in\mathbb{C},$$
which, together with the first inequality in item 4. of Theorem \ref{Frost-C}, proves our contention.
\end{proof}
We proceed with properties of the weighted Vandermonde. We have the relation
$$|VDM_k^Q(z_1,...,z_k)|=|VDM_k^{\tilde Q}(T(z_1),...,T(z_k))|,$$
from which it follows that the assertions of Theorem \ref{sec3} about the Vandermonde can be carried over to $\mathbb{C}$. Since the result may be of interest on its own, we state it as a theorem.
\begin{theorem}
Let $K$ be a closed nonpolar subset of $\mathbb{C}$ and $Q$ a weakly admissible weight on $K$. The $k$-th weighted diameters $\delta_{k}^{Q}(K)$, $k\geq 2$, are finite and
\begin{enumerate}
\item if $\{\mu_k=\frac{1}{k}\sum_{j=1}^k\delta_{x_j^{(k)}}\}\subset \mathcal M(K)$ converge weakly to $\mu\in \mathcal M(K)$, then
\begin{equation}\label{upboundVDM-C}
\limsup_{k\to \infty} |VDM_k^Q(x_1^{(k)},...,x_k^{(k)})|^{2/k(k-1)}\leq \exp{(-I^Q(\mu))};
\end{equation}
\item we have
$\delta^Q(K):=\lim_{k\to \infty} \delta_k^Q(K)=\exp{(-V_w)};$
\item if $\{x_j^{(k)}\}_{j=1,...,k; \ k=2,3,...}\subset K$ and
$$\lim_{k\to \infty} |VDM_k^Q(x_1^{(k)},...,x_k^{(k)})|^{2/k(k-1)}= \exp{(-V_w)}$$
then
$$\mu_k=\frac{1}{k}\sum_{j=1}^k\delta_{x_j^{(k)}}\to \mu_{K,Q} \ \hbox{weakly}.$$
\end{enumerate}
\end{theorem}
\section{Bernstein-Markov properties in $\mathbb{R}^n$}\label{Sec-BM}
In sections 4-8, we return to the setting of compact sets in $\mathbb{R}^n$. In particular, admissible weights will be in the sense of Definition \ref{realad}. For $k=1,2,...$, let $\mathcal P_k=\mathcal P_k^{(n)}$ denote the \emph{real} polynomials in $n$ variables of degree at most $k$. Given a compact set $K\subset \mathbb{R}^n$ and a positive measure $\nu$ on $K$, we say that $(K,\nu)$ satisfies the Bernstein-Markov property (or $\nu$ is a Bernstein-Markov measure for $K$) if for all $p_k\in \mathcal P_k$,
$$||p_k||_K:=\sup_{z\in K} |p_k(z)|\leq M_k||p_k||_{L^2(\nu)} \ \hbox{with} \ \limsup_{k\to \infty} M_k^{1/k} =1.$$
It was shown in \cite{PELD} that any compact set in $\mathbb{C}^n$ admits a Bernstein-Markov measure for {\it holomorphic} polynomials; indeed, the following stronger statement is true.
\begin{proposition} [\cite{PELD}] \label{allbm} Let $K\subset \mathbb{R}^n$. There exists a measure $\nu \in \mathcal M(K)$ such that for all complex-valued polynomials $p$ of degree at most $k$ in the (real) coordinates $x=(x_1,...,x_n)$ we have
$$||p||_K\leq M_k ||p||_{L^2(\nu)}$$
where $\limsup_{k\to \infty}M_k^{1/k}=1$.
\end{proposition}
More generally, for $K\subset \mathbb{R}^n$ compact, $Q\in \AA(K)$, and $\nu$ a measure on $K$, we say that the triple $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property if for all $p_k\in \mathcal P_k$,
$$||e^{-kQ}p_k||_K \leq M_k ||e^{-kQ}p_k||_{L^2(\nu)} \ \hbox{with} \ \limsup_{k\to \infty} M_k^{1/k} =1.$$
\begin{remark}\label{lp} These properties can be stated with $L^p$ norms for any $0<p< \infty$. The proof of Theorem 3.4.3 in \cite{StTo} in $\mathbb{C}$ that if $(K,\nu)$ satisfies an (weighted) $L^p-$Bernstein-Markov property for some $0<p<\infty$ then $(K,\nu)$ satisfies an (weighted) $L^p-$Bernstein-Markov property for all $0<p<\infty$ just uses H\"older's inequality and remains valid in our setting.
\end{remark}
Now another very important observation: Theorem 3.2 of \cite{bloom} works -- indeed, is even stated -- in $\mathbb{R}^n$ for any $n\geq 2$:
\begin{theorem} \label{blooms} Given $K\subset \mathbb{R}^n$ compact, and $Q$ a {\it continuous} weight, if $\nu$ is a finite measure on $K$ such that $(K,\nu)$ satisfies a Bernstein-Markov property, then the triple $(K,\nu,Q)$ satisfies a weighted Bernstein-Markov property.
\end{theorem}
\begin{definition} Given $K\subset \mathbb{R}^n$ compact, a finite measure $\nu$ on $K$ is called a {\it strong Bernstein-Markov measure} for $K$ if for any continuous weight $Q$ on $K$, the triple $(K,\nu,Q)$ satisfies a weighted Bernstein-Markov property.
\end{definition}
\begin{remark}\label{strongbm} Combining Proposition \ref{allbm} and Theorem \ref{blooms} we see that {\it any compact set $K$ in $\mathbb{R}^n$ admits a strong Bernstein-Markov measure; and any Bernstein-Markov measure on $K$ is automatically a strong Bernstein-Markov measure for $K$}. For $K\subset \mathbb{C}$, there are well-known sufficient mass-density conditions on a measure $\nu$ on $K$ so that $(K,\nu)$ satisfies a Bernstein-Markov property \cite{StTo}. In particular, Lebesgue measure on an interval or Lebesgue planar measure on a compact set in $\mathbb{C}$ having $C^1$ boundary satisfy the Bernstein-Markov property. We remark that if $\mathbb{C}\setminus K$ is regular for the Dirichlet problem, the condition that $(K,\nu)$ satisfies a Bernstein-Markov property is equivalent to the condition that $\nu$ be a regular measure; i.e., $\nu \in {\bf Reg}$ in the terminology of \cite{StTo}. We refer to this book for more details.
Furthermore, for every compact set $K$ in $\mathbb{R}^n$ there exist discrete measures which satisfy the (strong) Bernstein-Markov property \cite{PELD}. If one considers $K\subset \mathbb{R}^n\subset \mathbb{C}^n$, there are sufficient mass-density conditions on a measure $\nu$ on $K$ so that $(K,\nu)$ satisfies a Bernstein-Markov property for polynomials on $\mathbb{C}^n$ and hence on $\mathbb{R}^n$. For more on this, cf., \cite{BLmass} and \cite{PELD}.
\end{remark}
\begin{remark}\label{bmprop} Let $K\subset \mathbb{R}^n$ be compact and not log-polar and let $v\in \mathcal A(K)$. If $\alpha$ is a finite measure on $K$ such that $(K,\alpha,v)$ satisfies a weighted Bernstein-Markov property, then
\begin{enumerate}
\item $(K,c\alpha,v)$ satisfies a weighted Bernstein-Markov property for any $0< c <\infty$ and
\item $(K,\alpha+\beta,v)$ satisfies a weighted Bernstein-Markov property for any finite measure $\beta$ on $K$.
\end{enumerate}
\end{remark}
The importance of a (weighted) Bernstein-Markov property is the following consequence on the asymptotic behavior of the (weighted) {\it $L^{2}$ normalization constants} defined by
\begin{equation}\label{L2}
Z^Q_k=Z^{Q}_k(K,\nu):=\int_{K^{k}}|VDM_k^Q({\bf X_k})|^2d\nu({\bf X_k}),
\end{equation}
where ${\bf X_k}:=(x_1,...,x_k)\in K^{k}$ and $\nu$ is a finite positive measure on $K$.
\begin{remark}\label{L2-Z}
The quantity $Z_{k}^{Q}$ appears as the normalization constant in the law of eigenvalues of random matrix models. It is also referred to as the partition function in the theory of Coulomb gases. See Section \ref{appli} for more details on the link between these notions.
\end{remark}
\begin{proposition} \label{weightedtd}
Given $K\subset \mathbb{R}^n$ compact and not log-polar, $Q\in \AA(K)$, and $\nu$ a finite measure on $K$ such that $(K,\nu,Q)$ satisfies a weighted Bernstein-Markov property, we have
$$\lim_{k\to \infty} (Z_k^Q)^{1/k(k-1)}=\exp{(-V_w)}=\delta^{Q}(K).$$
\end{proposition}
\begin{proof} We clearly have $\limsup_{k\to \infty} (Z_k^Q)^{1/k(k-1)}\leq \exp{(-V_w)}$ from 2. of Theorem \ref{sec3}. For the reverse inequality with $\liminf$, note that
$$|VDM_k(x_1,...,x_k)|^2 =\prod_{i<j} |x_i-x_j|^2=\prod_{i<j} \bigl(\sum_{l=1}^n(x_{i,l}-x_{j,l})^2\bigr)$$
(where we write $x_i =(x_{i,1},...,x_{i,n})$)
is a polynomial of degree $k(k-1)$ in the $nk$ real coordinates $\{x_{i,l}\}_{l=1,...,n; \ i=1,...,k}$.
Now if ${\bf F_k}=(f_1,...,f_k)$ is a set of weighted Fekete points of order $k$ for $K,Q$, then
$$p(x_1):=|VDM_k(x_1,f_2,...,f_k)|^2\prod_{j=2}^ke^{-2(k-1)Q(f_j)}$$
is a (nonnegative) polynomial of degree $2(k-1)$ in (the coordinates of) $x_1$. By definition of weighted Fekete points, for any $x_1\in K$,
$$p(x_1) e^{-2(k-1)Q(x_1)} \leq \max_{x\in K} p(x)e^{-2(k-1)Q(x)}=p(f_1) e^{-2(k-1)Q(f_1)}$$
since this right-hand-side is precisely $|VDM_k^Q({\bf F_k})|^2$. By the weighted Bernstein-Markov property using $L^1$ norm instead of $L^2$ (see Remark \ref{lp}),
$$|VDM_k^Q({\bf F_k})|^2\leq M_{2(k-1)} \int_K |VDM_k(x_1,f_2,...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)}\cdot \prod_{j=2}^ke^{-2(k-1)Q(f_j)}d\nu(x_1).$$
Now for each fixed $x_1\in K$, we consider
$$p_2(x_2):= |VDM_k(x_1,x_2,f_3...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)} \cdot \prod_{j=3}^ke^{-2(k-1)Q(f_j)}$$
which is a (nonnegative) polynomial of degree $2(k-1)$ in (the coordinates of) $x_2$. Then
$$p_2(f_2) e^{-2(k-1)Q(f_2)}\leq \max_{x\in K} p_2(x) e^{-2(k-1)Q(x)}.$$
The left-hand-side is
$$|VDM_k(x_1,f_2,f_3...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)} \cdot \prod_{j=2}^ke^{-2(k-1)Q(f_j)}.$$
The right-hand-side, by the weighted Bernstein-Markov property, is bounded above by
$$M_{2(k-1)} \int_K |VDM_k(x_1,x_2,f_3,...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)}\cdot\prod_{j=3}^ke^{-2(k-1)Q(f_j)}e^{-2(k-1)Q(x_2)}d\nu(x_2).$$
Plugging these into our first estimate, we have
$$|VDM_k^Q({\bf F_k})|^2\leq M_{2(k-1)} \int_K |VDM_k(x_1,f_2,...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)}\cdot \prod_{j=2}^ke^{-2(k-1)Q(f_j)}d\nu(x_1)$$
$$\leq M_{2(k-1)} \int_K\bigl[ M_{2(k-1)} \int_K |VDM_k(x_1,x_2,f_3,...,f_k)|^2\cdot e^{-2(k-1)Q(x_1)}\cdot$$
$$\prod_{j=3}^ke^{-2(k-1)Q(f_j)}e^{-2(k-1)Q(x_2)}d\nu(x_2)\bigr]d\nu(x_1)$$
$$=(M_{2(k-1)} )^2\int_{K^2} |VDM_k(x_1,x_2,f_3,...,f_k)|^2\prod_{j=3}^ke^{-2(k-1)Q(f_j)}\prod_{j=1}^2e^{-2(k-1)Q(x_j)}d\nu(x_2)d\nu(x_1).$$
Continuing the process and using $M_{2(k-1)}^{1/2k}\to 1$ gives the result.
\end{proof}
Given {\it any} $Q\in \AA(K)$, we can always find a finite measure satisfying the important conclusion of Proposition \ref{weightedtd}.
\begin{proposition}\label{7to4} Let $K\subset \mathbb{R}^n$ be compact and not log-polar and let $Q\in \mathcal A(K)$. Then there exists a finite measure $\mu$ on $K$ such that
\begin{equation}\label{in7to4}\lim_{k\to \infty} (Z_k^Q(K,\mu))^{1/k(k-1)}=\exp{(-V_w)}=\delta^{Q}(K).\end{equation}
We can even construct $\mu$ so that, in addition, $\mu$ is a (strong) Bernstein-Markov measure for $K$.
\end{proposition}
\begin{proof} Consider the weighted equilibrium measure $\mu_{K,Q}$. We have $V_w=I(\mu_{K,Q})<\infty$. By Lusin's continuity theorem, for every integer $m> 1$, there exists a compact subset $K_{m}$ of $K$ such that $\mu_{K,Q}(K\setminus K_{m})\leq 1/m$ and $Q$ (considered as a function on $K_{m}$ only) is continuous on $K_{m}$. We may assume that each $K_m$ is not log-polar and that the sets $K_{m}$ are increasing as $m$ tends to infinity. Let $\mu_m\in \mathcal M(K_m)$ be a (strong) Bernstein-Markov measure for $K_m$. We claim that $\mu=\sum_{m=1}^{\infty} \frac{1}{2^m}\mu_m$ satisfies (\ref{in7to4}).
Since $\mu_m\in \mathcal M(K_m)$ is a (strong) Bernstein-Markov measure for $K_m$ and $Q_m:=Q|_{K_m}$ is continuous, it also follows from 1. of Remark \ref{bmprop} that $(K_m,Q_m,\frac{1}{2^m}\mu_m)$ satisfies a weighted Bernstein-Markov property.
Since $Z_k^Q(K,\mu)\leq \max_{{\bf x}\in K^k}|VDM^Q({\bf x})|^2\mu(K)^k$, we have
$$\limsup_{k\to \infty} (Z_k^Q(K,\mu))^{1/k(k-1)}\leq \delta^{Q}(K).$$
To show
$$\liminf_{k\to \infty} (Z_k^Q(K,\mu))^{1/k(k-1)}\geq \delta^{Q}(K),$$
let $\lambda_m:= \mu_{K,Q}(K_m)$ so that $\lambda_m\uparrow 1$. Letting
$\theta_m:=\frac{1}{\lambda_m} \mu_{K,Q}|_{K_m}\in \mathcal M(K_m)$, we have
$$I^{Q_m}(\theta_m)\geq I^{Q_m}(\mu_{K_m,Q_m}).$$
Since $(K_m,Q_m,\frac{1}{2^m}\mu_m)$ satisfies a weighted Bernstein-Markov property,
$$\exp{(- I^{Q_m}(\mu_{K_m,Q_m}))}=\lim_{k\to \infty}(Z_k^{Q_m}(K_m,\frac{1}{2^m}\mu_m))^{1/k(k-1)}.$$
Clearly
$$Z_k^Q(K,\mu)\geq Z_k^{Q_m}(K_m,\mu|_{K_m})\geq Z_k^{Q_m}(K_m,\frac{1}{2^m}\mu_m).$$
Thus
$$\liminf_{k\to \infty} (Z_k^Q(K,\mu))^{1/k(k-1)}\geq \liminf_{k\to \infty} (Z_k^{Q_m}(K_m,\frac{1}{2^m}\mu_m))^{1/k(k-1)}$$
$$=\exp{(- I^{Q_m}(\mu_{K_m,Q_m}))}\geq \exp{(-I^{Q_m}(\theta_m))}.$$
By monotone convergence we have
$$\lim_{m\to \infty} I^{Q_m}(\theta_m)= I^Q(\mu_{K,Q})$$
so that
$$\liminf_{k\to \infty} (Z_k^Q(K,\mu))^{1/k(k-1)}\geq \exp{-I^Q(\mu_{K,Q})}=\delta^{Q}(K),$$
as desired.
For the second part, let $\nu$ be a (strong) Bernstein-Markov measure for $K$ and define
$$\mu:= \sum_{m=1}^{\infty} 2^{-m}\mu_m + \nu.$$
The fact that $\mu$ is a (strong) Bernstein-Markov measure for $K$ follows from the fact that $\nu$ is a (strong) Bernstein-Markov measure for $K$ and 2. of Remark \ref{bmprop}. Finally, $\mu$ satisfies (\ref{in7to4}) from the previous part applied to $\sum_{m=1}^{\infty} 2^{-m}\mu_m$ and the obvious inequality $Z_k^Q(K,\mu) \geq Z_k^Q(K,\sum_{m=1}^{\infty} 2^{-m}\mu_m)$.
\end{proof}
\begin{example} If $\mu$ is a (strong) Bernstein-Markov measure for $K$ and the set of points of discontinuity of $Q\in \mathcal A(K)$ is of $\mu$ measure zero, then (\ref{in7to4}) holds for $\mu$. As a simple but illustrative example, let $K=[-1,1]\subset \mathbb{R}$ and take
$$Q(x)=0 \ \hbox{at all} \ x\in [-1,1]\setminus \{0\}; \ Q(0)=-1.$$
It is easy to see that Lebesgue measure $d\mu$ on $[-1,1]$ satisfies (\ref{in7to4}) but $(K,Q,\mu)$ does not satisfy the weighted Bernstein-Markov property. On the other hand, $(K,Q,\mu+\delta_0)$ does satisfy the weighted Bernstein-Markov property.\end{example}
For $Q\in \mathcal A(K)$ and $\nu$ a finite measure on $K$,
we define a probability measure $Prob_k$ on $K^{k}$: for a Borel set $A\subset K^{k}$,
\begin{equation}\label{probk}Prob_k(A):=\frac{1}{Z^Q_k}\cdot \int_A |VDM_k^Q({\bf X_k})|^2 d\nu({\bf X_k}).
\end{equation}
Directly from Proposition \ref{7to4} and (\ref{probk}) we obtain the following estimate.
\begin{corollary} \label{johansson}
Let $Q\in \mathcal A(K)$ and $\nu$ a finite measure on $K$ satisfying
$$\lim_{k\to \infty} (Z_k^Q(K,\nu))^{1/k(k-1)}=\exp{(-V_w)}=\delta^{Q}(K).$$
Given $\eta >0$, define
\begin{equation}\label{aketa}
A_{k,\eta}:=\{{\bf X_k}\in K^{k}: |VDM_k^Q({\bf X_k})|^2 \geq
( \delta^Q(K) -\eta)^{k(k-1)}\}.
\end{equation}
Then there exists $k^*=k^*(\eta)$ such that for all $k>k^*$,
$$Prob_k(K^{k}\setminus A_{k,\eta})\leq \Big(1-\frac{\eta}{2 \exp{(-V_w)}}\Big)^{k(k-1)}\nu(K^{k}).
$$
\end{corollary}
We get the induced product probability measure ${\bf P}$ on the space of arrays on $K$,
$$\chi:=\{X=\{{\bf X_{k}}\in K^{k}\}_{k\geq 1}\},$$
namely,
$$(\chi,{\bf P}):=\prod_{k=1}^{\infty}(K^{k},Prob_k).$$
As an immediate consequence of the Borel-Cantelli lemma and 3. of Theorem \ref{sec3}, we obtain:
\begin{corollary}\label{416} Let $Q\in \mathcal A(K)$ and $\nu$ a finite measure on $K$ satisfying $$\lim_{k\to \infty} (Z_k^Q(K,\nu))^{1/k(k-1)}=\exp{(-V_w)}=\delta^{Q}(K).$$
For ${\bf P}$-a.e. array $X=\{x_j^{(k)}\}_{j=1,...,k; \ k=2,3,...}\in \chi$,
$
\frac{1}{k}\sum_{j=1}^k\delta_{x_j^{(k)}}\to \mu_{K,Q} \ \hbox{weakly as }
k\to\infty.$$
\end{corollary}
\section{Approximation of probability measures}
For the proof of a large deviation principle (LDP) in $\mathbb{R}^n$, as in \cite{VELD}, we will need to approach general measures in $\mathcal M(K)$ by weighted equilibrium measures. For that, we consider equilibrium problems with weights that are the negatives of potentials. We first verify that the natural candidate solution to such a problem is, indeed, the true solution.
\begin{lemma}\label{lem-non-adm}
Let $\mu\in\mathcal M(K)$, $K\subset \mathbb{R}^n$ compact, $I(\mu)<\infty$. Consider the possibly non-admissible weight $u:=-U^{\mu}$ on $K$. The weighted minimal energy on $K$ is obtained with the measure $\mu$, that is
$$\forall\nu\in\mathcal M(K),\quad I(\mu)+2\int ud\mu\leq I(\nu)+2\int ud\nu,$$
with equality if and only if $\nu=\mu$.
\end{lemma}
\begin{proof}
We may assume that $I(\nu)<\infty$. The inequality may be rewritten as
$$0\leq I(\nu)-2I(\mu,\nu)+I(\mu)=I(\nu-\mu),$$
which is true, and, moreover, the energy $I(\nu-\mu)$ can vanish only when $\nu=\mu$, see item 1. of Theorem \ref{basics}
\end{proof}
The following two approximation results are analogous to \cite[Lemma I.6.10]{ST}.
\begin{lemma}\label{lemma-scal}
Let $K\subset \mathbb{R}^n$ be compact and non log-polar and let $\mu\in\mathcal M(K)$ with $I(\mu)<\infty$. Let $Q\in\AA(K)$ be finite $\mu$-almost everywhere.
There exist an increasing sequence of compact sets $K_m$ in $K$ and a sequence of measures $\mu_m\in \mathcal M(K_m)$ satisfying
\begin{enumerate
\item the measures $\mu_m$ tend weakly to $\mu$, as $m\to\infty$;
\item the functions $Q_{|K_m}\in C(K_m)$ and $\int Qd\mu_m$ tend to $\int Qd\mu$ as $m\to\infty$;
\item the energies $I(\mu_m)$ tend to $I(\mu)$ as $m\to\infty$.
\end{enumerate}
\end{lemma}
\begin{proof}
By Lusin's continuity theorem, for every integer $m\geq 1$, there exists a compact subset $K_m$ of $K$ such that $\mu(K\setminus K_m)\leq 1/m$ and $Q$ (considered as a function on $K_m$ only) is continuous on $K_m$. We may assume that $K_m$ is increasing as $m$ tends to infinity. Then, the measures $\tilde\mu_m:=\mu_{|K_m}$ are increasing and tend weakly to $\mu$.
Since $Q$ is bounded below on $K$, the monotone convergence theorem tells us that
$$\int Qd\tilde\mu_m=\int Q_{|K_m}d\mu\to\int Qd\mu,\quad\text{as }m\to\infty.$$
Moreover, since the measures $\tilde\mu_m$ are increasing to $\mu$ and $I(\mu)<\infty$, we get by the generalized dominated convergence theorem, see e.g. \cite[Theorem 2.2]{HL},
$$I(\tilde\mu_m)\to I(\mu),\quad\text{as }m\to\infty,$$
where we use that the potentials $U^{\tilde\mu_m}$ are increasing (possibly by scaling the compact set $K$ to a set of diameter less than one so that $\log 1/|z-t|>0$ for $z,t\in K$).
Finally, defining $\mu_m:= \tilde \mu_m/\mu(K_m)$ gives the result.
\end{proof}
\begin{corollary} \label{corapprox}
Let $K\subset \mathbb{R}^n$ be compact and non log-polar and let $\mu\in\mathcal M(K)$ with $I(\mu)<\infty$. Let $K_m$ be the sequence of increasing compact sets in $K$ and $\mu_m$ the sequence of measures in $\mathcal M(K_m)$ given by Lemma \ref{lemma-scal} with $Q=U^{\mu}$. There exist a sequence of continuous functions $Q_m$ on $K$ such that
\begin{enumerate}
\item the measures $\mu_m$ tend weakly to $\mu$ and the energies $I(\mu_m)$ tend to $I(\mu)$, as $m\to\infty$;
\item the measures $\mu_m$ are equal to the weighted equilibrium measures $\mu_{K,Q_m}$.
\end{enumerate}
\end{corollary}
\begin{proof} First, note that $U^{\mu}\in\AA(K)$ and is finite $\mu$-almost everywhere since $I(\mu)<\infty$, so that Lemma \ref{lemma-scal} applies with $Q=U^{\mu}$. Now we define $Q_m:=-U^{\mu_m}|_K$. Then,
$$Q_m=-U^{\mu}+U^{\mu-\mu_m}.$$
By definition $Q_m$ is upper semicontinuous on $K$ while $U^{\mu}$ is continuous on $K_m$, see item 2. of Lemma \ref{lemma-scal}. Hence, the right-hand side is lower semicontinuous on $K_m$ which shows that $Q_m$ is continuous on $K_m$. By the continuity principle for logarithmic potentials (4. of Theorem \ref{basics}), $-U^{\mu_m}$ is continuous on $\mathbb{R}^n$ and hence $Q_m$ is continuous
on $K$. Items 1. and 2. follow from Lemmas \ref{lemma-scal} and \ref{lem-non-adm} respectively.
\end{proof}
\section{The $J^{Q}$ functionals on $\mathbb{R}^n$}\label{J-funct}
In this section, we introduce and establish the main properties of the weighted
$L^2$ functionals
$\overline J^{Q},\underline J^{Q}$ as well as the relation with the weighted energy $I^Q$. Our goal is to establish an LDP in the next section.
Fix a compact set $K$ in $\mathbb{R}^{n}$,
a measure
$\nu$ in ${\mathcal M}(K)$ and $Q\in\AA(K)$.
We recall that $\mathcal M(K)$ endowed with the weak topology is a Polish space, i.e., a separable complete metrizable space.
Given $G\subset {\mathcal M}(K)$, for each $k=1,2,...$ we let
\begin{equation}\label{nbhddef}
\tilde G_k:=\{{\bf a} =(a_{1},...,a_{k})\in K^{k},~
\frac{1}{k}\sum_{j=1}^{{k}} \delta_{a_{j}}\in G\},
\end{equation}
and set
\begin{equation}\label{jkqmu}
J^Q_k(G):=\Big[\int_{\tilde G_{k}}|VDM^Q_k({\bf a})|^{2}d\nu ({\bf a})\Big]
^{1/k(k-1)}.
\end{equation}
\begin{definition} \label{jwmuq} For $\mu \in \mathcal M(K)$ we define
$$\overline J^Q(\mu):=\inf_{G \ni \mu} \overline J^Q(G) \ \hbox{where} \ \overline J^Q(G):=\limsup_{k\to \infty} J^Q_k(G);$$
$$\underline J^Q(\mu):=\inf_{G \ni \mu} \underline J^Q(G) \ \hbox{where} \ \underline J^Q(G):=\liminf_{k\to \infty} J^Q_k(G);$$
\end{definition}
\noindent Here the infima are taken over all neighborhoods $G$ of the measure $\mu$ in ${\mathcal M}(K)$.
Note tha
, a priori, $\overline J^{Q},\underline J^{Q}$ depend on $\nu$. For the unweighted case $Q=0$, we simply write $\overline J$ and $\underline J$.
\begin{lemma}
The functionals $\underline J(\mu)$, $\overline J(\mu)$, $\underline J^Q(\mu)$, $\overline J^Q(\mu)$, are upper semicontinuous on ${\mathcal M}(K)$ in the weak topology.
\end{lemma}
\begin{proof}
The proof is similar to the one of \cite[Lemma 3.1]{BL}.
\end{proof}
\begin{lemma}\label{lem-J-JQ}
The following properties hold (and with the $\underline J,\underline J^Q$
functionals as well):
\begin{enumerate}
\item $\overline J^Q(\mu)\leq e^{-I^Q(\mu)}$ for $Q\in \AA(K)$;
\item $
\overline J^{Q}(\mu)\leq\overline J(\mu)\cdot e^{-2\int_K Qd\mu}$ for $Q\in \AA(K)$;
\item
$ \overline J^{Q}(\mu)=\overline J(\mu)\cdot e^{-2\int_K Qd\mu}$ for {\bf $Q$ {continuous}}.
\end{enumerate}
\end{lemma}
\begin{proof}
Item 1. follows from
$$J^Q_{k}(G)\leq \sup_{{\bf a}\in\tilde G_{k}}|VDM_{k}^{Q}({\bf a})|^
{2/k(k-1)}
,$$
and the upper bound (\ref{upboundVDM}) on the limit of the Vandermonde. We prove item 2. and item 3. simultaneously. We first observe that if $\mu \in {\mathcal M}(K)$ and $Q$ is continuous on $K$, given $\epsilon >0$, there exists a neighborhood $G \subset {\mathcal M}(K)$ of $\mu$ with
$$\big|\int_{K} Q\Big(d\mu -\frac{1}{k}\sum_{j=1}^{k}
\delta_{a_{j}}\Big)\big|\leq\epsilon \quad \hbox{for} \ {\bf a} \in \tilde G_k$$
for $k$ sufficiently large. Thus we have
\begin{equation}\label{ineg}
-\epsilon -\int_{K}Qd\mu\leq -\frac{1}{k}\sum_{j=1}^{k}
Q(a_{j})
\leq \epsilon-\int_{K}Qd\mu.
\end{equation}
Note that for $Q\in \AA(K)$, hence lower semicontinuous, we only have the second inequality.
Since
$$|VDM_k^Q({\bf a})|=|VDM_k({\bf a})|\cdot
\prod_{j=1}^{k} e^{-(k-1)Q(a_{j})},$$
we deduce from (\ref{ineg}) that
$$|VDM_k({\bf a})|e^{-k(k-1)
(\epsilon+\int_K Qd\mu)}
\leq |VDM^Q_k({\bf a})| \leq
|VDM_k({\bf a})|e^{k(k-1)(\epsilon-\int_K Qd\mu)}.
$$
Now we take the square, integrate over ${\bf a}\in \tilde G_k$ and take a
$k(k-1)$-th root of each side to get
$$J_k(G)e^{-2(\epsilon+\int_K Q d\mu)} \leq J^Q_k(G) \leq J_k(G)
e^{2(\epsilon-\int_K Q d\mu)}.
$$
Precisely, given $\epsilon >0$, these inequalities are valid for $G$ a sufficiently small neighborhood of $\mu$. Hence we get, upon taking $\limsup_{k\to \infty}$, the infimum over $G\ni \mu$, and noting that $\epsilon >0$ is arbitrary,
$$\overline J(\mu)=\overline J^Q(\mu)\cdot e^{2\int_K Qd\mu}$$
as desired. If $Q$ is only lower semicontinuous, we still have the upper bounds in the above, which gives item 2.
\end{proof}
From item 1. in Lemma \ref{lem-J-JQ}, we know that for $Q\in \AA(K)$
\begin{equation}\label{Jupbound}
\log \underline J^{Q}(\mu)\leq \log \overline J^{Q}(\mu)\leq -I^Q(\mu).
\end{equation}
In the remainder of this section, we show that
when the measure $\nu$ satisfies a {Bernstein-Markov property}, equalities hold in (\ref{Jupbound}).
We first consider the unweighted functionals $\underline J$ and $\overline J$ and the case of an equilibrium measure $\mu=\mu^{K,v}$ where $v\in \AA(K)$
\begin{lemma}\label{lem-eq-case}
Let $K$ be non log-polar, $v\in \AA(K)$, and let $\nu\in {\mathcal M}(K)$ such that $(K,\nu,v)$ satisfy a weighted Bernstein-Markov property. Then,
\begin{equation}\label{jversion}
\log \overline J(\mu^{K,v})= \log\underline J(\mu^{K,v})=-I(\mu^{K,v}).
\end{equation}
\end{lemma}
\begin{proof}
To prove (\ref{jversion}), we first verify the following.
\medskip
\noindent {\sl Claim: Fix a neighborhood $G$ of $\mu_{K,v}$. For $\eta >0$, define
\begin{equation*
A_{k,\eta}:=\{{\bf Z_k}\in K^{k}: |VDM_k^v({\bf Z_k})|^2 \geq
( \delta^v(K) -\eta)^{k(k-1)}\}.
\end{equation*}
Given a sequence $\{\eta_j\}$ with $\eta_j\downarrow 0$, there exists a $j_{0}$ and a $k_{0}$ such that
\begin{equation}\label{setinclu} \forall j\geq j_{0},\quad\forall k\geq k_{0},\quad A_{k,\eta_j} \subset \tilde G_{{k}}.
\end{equation}
}
\medskip
\noindent
We prove (\ref{setinclu}) by contradiction: if false, there are sequences $\{k_l\}$ and $\{j_{l}\}$ tending to infinity such that for all $l$ sufficiently large we can find a point ${\bf Z_{k_l}}=(z_1,...,z_{k_l})$ with ${\bf Z_{k_l}}\in
A_{k_l,\eta_{j_{l}}} \setminus \tilde G_{k_l}$. But
$$\mu^l:=\frac{1}{k_{l}}\sum_{i=1}^{k_{l}}\delta_{z_{i}}\not\in G$$
for $l$ sufficiently large is a contradiction with item 3. of Theorem \ref{sec3} since ${\bf Z_{k_l}}\in A_{k_l,\eta_{j_{l}}}$ and $\eta_{j_l}\to 0$ imply $\mu^l\to \mu_{K,v}$ weakly. This proves the claim.
\medskip
Fix a neighborhood $G$ of $\mu_{K,v}$ and a sequence $\{\eta_j\}$ with $\eta_j\downarrow 0$. For $j\geq j_{0}$, choose $k=k_j$ large enough so that the inclusion in (\ref{setinclu}) holds true as well as
\begin{equation}\label{probk2}Prob_{k_j}(K^{k_j}\setminus A_{k_j,\eta_j})\leq
\Big(1-\frac{\eta_j}{2\delta^v(K)}\Big)^{{k_j}({k_j}-1)}
\end{equation}
and
\begin{equation}\label{seq}\Big(1-\frac{\eta_j}{2\delta^v(K)}\Big)^{{k_j}({k_j}-1)}
\to 0 \quad \hbox{as} \quad j\to \infty,
\end{equation}
which is possible (for (\ref{probk2}) we make use of Corollary \ref{johansson}). In view of (\ref{setinclu}), the definition of $Prob_{k_{j}}$, and (\ref{probk2}), we have
\begin{align}\notag
\frac{1}{Z^v_{k_j}}
\int_{\tilde G_{k_j}} |VDM_{k_j}^{v}({\bf Z_{k_j}})|^2 d\nu({\bf Z_{k_j}})
& \geq \frac{1}{Z^v_{k_j}}
\int_{A_{k_j,\eta_j} } |VDM_{k_j}^{v}({\bf Z_{k_j}})|^2 d\nu({\bf Z_{k_j}}) \\
\label{ineg-for-J}
& \geq 1- \Big(1-\frac{\eta_j}{2\delta^v(K)}\Big)^{{k_j}({k_j}-1)}.
\end{align}
Note that, because of (\ref{seq}), the lower bound in (\ref{ineg-for-J}) tends to 1 as $j\to\infty$. Then,
since $(K,\nu,v)$ satisfy a weighted Bernstein-Markov property, we derive, with
the asymptotics of $Z_{k_{j}}^{v}$ given in Proposition \ref{weightedtd}, that
$$\liminf_{j\to \infty} \frac{1}{k_j(k_j-1)}\log \int_{\tilde G_{k_j}} |VDM_{k_j}^{v}({\bf Z_{k_j}})|^2 d\nu({\bf Z_{k_j}}) \geq \log \delta^v(K).$$
Given any sequence of positive integers $\{k\}$ we can find a subsequence $\{k_j\}$ as above corresponding to some $\eta_j\downarrow 0$; hence
$$\liminf_{k\to \infty} \frac{1}{{k}({k}-1)}\log \int_{\tilde G_{k}} |VDM_{k}^{v}({\bf Z}_k)|^2 d\nu({\bf Z}_{k})\geq \log \delta^v(K).$$
It follows that
$$
\log \underline J^{v}(G)\geq \log \delta^v(K).$$
Taking the infimum over all neighborhoods $G$ of $\mu_{K,v}$ we obtain
$$\log \underline J^{v}(\mu_{K,v})\geq \log \delta^v(K).$$
Using item 2. of Lemma \ref{lem-J-JQ} with $\mu = \mu_{K,v}$, we get
$$
\log \underline J(\mu_{K,v})\geq -I(\mu_{k,v})
$$
and with the unweighted version of item 1., we obtain (\ref{jversion}).
\end{proof}
\begin{remark} We observe that the proof only used the property
$$\lim_{k\to \infty} (Z_k^v(K,\nu))^{1/k(k-1)}=\delta^{v}(K).$$
\end{remark}
\begin{theorem} \label{rel-J-E}
Let $K$ be a non log-polar compact subset of $\mathbb{R}^{n}$
and let $\nu\in {\mathcal M}(K)$ satisfy the (strong) Bernstein-Markov property. \\
(i) For any $\mu\in \mathcal M(K)$,
\begin{equation}\label{minunwtd}
\log \overline J(\mu)= \log\underline J(\mu)=-I(\mu).\end{equation}
(ii) Let $Q\in\AA(K)$. Then
\begin{equation}\label{rel-J-JQ}
\overline J^{Q}(\mu)=\overline J(\mu)\cdot e^{-2\int_K Qd\mu},
\end{equation}
(and with the $\underline J,\underline J^Q$
functionals as well) so that,
\begin{equation}\label{minwtd}\log \overline J^Q(\mu)= \log \underline J^Q(\mu)=-I^{Q}(\mu).
\end{equation}
\end{theorem}
\begin{proof}
We first prove (i). The upper bound
\begin{equation}\label{upper-J}
\log\overline J(\mu)\leq -I(\mu)
\end{equation}
is the unweighted version of (\ref{Jupbound}).
For the lower bound $-I(\mu)\leq\log\underline J(\mu)$ we first assume that
$I(\mu)<\infty$.
Using Corollary \ref{corapprox},
there exists a sequence of (continuous) functions $Q_{m}$ defined on $K$ and measures $\mu_{m}=\mu_{K,Q_{m}}$ tending weakly to $\mu$ such that,
\begin{equation}\label{from2}\lim_{m\to \infty}I(\mu_{m}) =I(\mu).
\end{equation}
Thus we can apply Lemma \ref{lem-eq-case} to conclude
$$\log \overline J(\mu_{m})= \log \underline J(\mu_{m})=-I(\mu_{m}),$$
and from (\ref{from2}) along with the uppersemicontinuity of the functional $\mu \to \underline J(\mu)$, we derive
$$\lim_{m\to \infty} \log \underline J(\mu_{m})=-I(\mu)\leq \log \underline J(\mu).$$
Together with (\ref{upper-J})
we get
$$\log \underline J(\mu)=\log \overline J(\mu)=-I(\mu).$$
If $\mu\in \mathcal M(K)$ satisfies $I(\mu)=\infty$,
Item 1. of Lemma \ref{lem-J-JQ} shows that $\overline J(\mu)=0$.
We next proceed with assertion (ii). For (\ref{rel-J-JQ}), it is sufficient to prove the inequality
\begin{equation}\label{minor-JQ}
\overline J^{Q}(\mu)\geq\overline J(\mu)\cdot e^{-2\int_K Qd\mu}.
\end{equation}
We first assume that $Q$ is finite $\mu$-almost everywhere and $I(\mu)<\infty$ so that Lemma \ref{lemma-scal} can be applied on $K$. Let $K_{m}$ be the sequence of compact subsets of $K$ and $\mu_{m}$ be the sequence of measures in $\mathcal M(K_{m})$ given by that lemma.
By the upper semicontinuity of the functional $\overline J^{Q}$,
$$\overline J^{Q}(\mu)\geq\limsup_{m \to \infty}\overline J^{Q}(\mu_{m}).$$
Also, by item 3. of Lemma \ref{lem-J-JQ} and (\ref{minunwtd}), since $Q|_{K_m}$ is continuous,
$$\overline J^{Q}(\mu_{m})=\overline J(\mu_{m})e^{-2\int Qd\mu_{m}}=
e^{-I(\mu_{m})-2\int Qd\mu_{m}}.$$
Hence, (\ref{minor-JQ}) follows from items 2. and 3. of Lemma \ref{lemma-scal}.
When $I(\mu)=\infty$, both sides of (\ref{minor-JQ}) equal 0, since $\overline J(\mu)=e^{-I(\mu)}$ and, by definition,
$0\leq\overline J^{Q}(\mu)$. If $\mu(\{Q=\infty\})>0$, this is true as well because
$\overline J(\mu)>-\infty$ while the exponential in the right-hand side vanishes.
Finally, (\ref{minwtd}) follows from (\ref{minunwtd}) and (\ref{rel-J-JQ}).
\end{proof}
\begin{remark} \label{fourfour} We note that the Bernstein-Markov property of the measure $\nu$ has only been applied with the sequence of continuous weights $Q_{m}$ that appear when approaching $\mu$ with Corollary \ref{corapprox}. \end{remark}
From now on, we simply use the notation $J,J^Q$, without the overline or underline. It follows from (\ref{minunwtd}) and (\ref{minwtd}) that these functionals are independent of the measure $\nu$; i.e., we have shown: {\it if $\nu\in \mathcal M(K)$ is {\bf any} (strong) Bernstein-Markov measure, for any $Q\in \AA(K)$,
and for any $\mu\in \mathcal M(K)$ we have}
\begin{equation}\label{w=j=i}
\log J^Q(\mu)=-I^Q(\mu).
\end{equation}
\section{Large Deviation Principle in $\mathbb{R}^n$}
Fix a non log-polar compact set $K$ in $\mathbb{R}^{n}$, a measure $\nu$ on $K$
and $Q\in\AA(K)$.
Define
$j_k: K^{k} \to \mathcal M(K)$ via
\begin{equation}\label{jk} j_k(x_1,...,x_{k}
=\frac{1}{k}\sum_{j=1}^{k} \delta_{x_j}.\end{equation}
The push-forward
$\sigma_k:=(j_k)_*(Prob_k) $ (see (\ref{probk}) for the definition of $Prob_k$) is a probability measure on $\mathcal M(K)$: for a Borel set $G\subset \mathcal M(K)$,
\begin{equation}\label{sigmak}
\sigma_k(G)=\frac{1}{Z_k^{Q}} \int_{\tilde G_{k}} |VDM_k^Q(x_1,...,x_{k})|^2 d\nu(x_1) \cdots d\nu(x_{k}),
\end{equation}
recall (\ref{L2}), (\ref{probk}) and (\ref{nbhddef}); here, $Z_k^{Q}$ depends on $K$, $Q$ and $\nu$.
\begin{theorem} \label{ldp} Assume $\nu$ is a (strong) Bernstein-Markov measure on $K$, $Q\in \mathcal A(K)$, and $\nu$ satisfies
\begin{equation}\label{nuas}\lim_{k\to \infty} (Z_k^Q(K,\nu))^{1/k(k-1)}=\exp{(-V_w)}=\delta^{Q}(K).\end{equation} The sequence $\{\sigma_k=(j_k)_*(Prob_k)\}$ of probability measures on $\mathcal M(K)$ satisfies a {\bf large deviation principle} with speed $k^{2}$ and good rate function $\mathcal I:=\mathcal I_{K,Q}$ where, for $\mu \in \mathcal M(K)$,
\begin{equation*
\mathcal I(\mu):=\log J^Q(\mu_{K,Q})-\log J^Q(\mu)=I^Q(\mu)-I^Q(\mu_{K,Q}).
\end{equation*}
\end{theorem}
This means that $\mathcal I:\mathcal M(K)\to [0,\infty]$ is a lower semicontinuous mapping such that the sublevel sets $\{\mu \in \mathcal M(K): \mathcal I(\mu)\leq \alpha\}$ are compact in the weak topology on $\mathcal M(K)$ for all $\alpha \geq 0$ ($\mathcal I$ is ``good'') satisfying (\ref{lowb}) and (\ref{highb}):
\begin{definition} \label{equivform}
The sequence $\{\mu_k\}$ of probability measures on $\mathcal M(K)$ satisfies a {\bf large deviation principle} (LDP) with good rate function $\mathcal I$ and speed
$k^2$ if for all
measurable sets $\Gamma\subset \mathcal M(K)$,
\begin{equation}\label{lowb}-\inf_{\mu \in \Gamma^0}\mathcal I(\mu)\leq \liminf_{k\to \infty} \frac{1}{k^2} \log \mu_k(\Gamma) \ \hbox{and}\end{equation}
\begin{equation}\label{highb} \limsup_{k\to \infty} \frac{1}{k^2} \log \mu_k(\Gamma)\leq -\inf_{\mu \in \bar \Gamma}\mathcal I(\mu).\end{equation}
\end{definition}
In the setting of $\mathcal M(K)$, to prove a LDP it suffices to work with a base for the weak topology. The following is a special case of a basic general existence result for a LDP given in Theorem 4.1.11 in \cite{DZ}.
\begin{proposition} \label{dzprop1} Let $\{\sigma_{\epsilon}\}$ be a family of probability measures on $\mathcal M(K)$. Let $\mathcal B$ be a base for the topology of $\mathcal M(K)$. For $\mu\in \mathcal M(K)$ let
$$\mathcal I(\mu):=-\inf_{\{G \in \mathcal B: \mu \in G\}}\bigl(\liminf_{\epsilon \to 0} \epsilon \log \sigma_{\epsilon}(G)\bigr).$$
Suppose for all $\mu\in \mathcal M(K)$,
$$\mathcal I(\mu)=-\inf_{\{G \in \mathcal B: \mu \in G\}}\bigl(\limsup_{\epsilon \to 0} \epsilon \log \sigma_{\epsilon}(G)\bigr).$$
Then $\{\sigma_{\epsilon}\}$ satisfies a LDP with rate function $\mathcal I(\mu)$ and speed $1/\epsilon$.
\end{proposition}
We give our proof of Theorem \ref{ldp} using Theorem \ref{rel-J-E}.
\begin{proof} As a base $\mathcal B$ for the topology of $\mathcal M(K)$, we
simply take all open sets. For $\{\sigma_{\epsilon}\}$, we take the sequence of probability measures $\{\sigma_k\}$ on $\mathcal M(K)$ and we take $\epsilon =k^{-2}$. For $G\in \mathcal B$,
$$\frac{1}{k^{2}}\log \sigma_k(G)= \frac{k-1}{k}\log J_k^Q(G)-\frac{1}{k^{2}}\log Z_k^{Q}$$
using (\ref{jkqmu}) and (\ref{sigmak}). From Proposition \ref{weightedtd}, and the fact that (\ref{nuas}) holds,
$$\lim_{k\to \infty} \frac{1}{k^{2}}\log Z_k^{Q}=\log \delta^Q(K)= \log J^Q(\mu_{K,Q});$$ and by Theorem \ref{rel-J-E},
$$\inf_{G \ni \mu} \limsup_{k\to \infty} \log J_k^Q(G)=\inf_{G \ni \mu} \liminf_{k\to \infty} \log J_k^Q(G)=\log J^Q(\mu).$$
Thus by Proposition \ref{dzprop1} and Theorem \ref{rel-J-E}, $\{\sigma_k\}$ satisfies an LDP with rate function
$$\mathcal I(\mu):=\log J^Q(\mu_{K,Q})-\log J^Q(\mu)=I^Q(\mu)-I^Q(\mu_{K,Q})$$
and speed $k^{2}$. This rate function is good since $\mathcal M(K)$ is compact.
\end{proof}
\begin{remark} Note that the rate function is independent of the (strong) Bernstein-Markov measure $\nu$ satisfying (\ref{nuas}).
\end{remark}
\section{Measures $\nu$ of infinite mass on $K\subset {\bf S}\subset \mathbb{R}^3$}
In this section, we restrict to the setting of compact subsets of the two-dimensional sphere ${\bf S}$ in $\mathbb{R}^{3}$, of center $(0,0,1/2)$ and radius $1/2$. Then in the following sections, we use stereographic projection from ${\bf S}$ to the complex plane to derive a large deviation principle on unbounded subsets of $\mathbb{C}$. Now typically, on the complex plane, one would like to consider locally finite measures with infinite mass like, e.g., the Lebesgue measure.
We use the stereographic projection $T$ defined in (\ref{stereo}) which sends the north pole $P_{0}=(0,0,1)$ of ${\bf S}$ to the point at infinity in $\mathbb{C}$.
On the sphere ${\bf S}$ we are thus led to consider positive measures $\nu$, locally finite in ${\bf S}\setminus P_{0}$, such that
\begin{equation}\label{cond-nu-inf}
\nu(V_{P_{0}})=\infty,\quad\text{for all neighborhoods }V_{P_{0}}\text{ of }P_{0}.
\end{equation}
The goal of this section is to extend the results from the previous sections to such measures.
Fix a compact subset $K$ of ${\bf S}$ containing $P_{0}$. To ensure the finiteness of the different quantities defined previously, some condition should be satisfied linking the measure $\nu$ and the increase of the weights $Q$ near $P_{0}$. We assume that
\begin{equation}\label{cond-nu-S}
\exists a>0,\quad\int_{K}\epsilon(x)^{a}d\nu(x)<\infty,
\end{equation}
where $\epsilon(x)$ is some nonnegative continuous function that tends to 0 as $x$ tends to $P_{0}$, and that
\begin{equation}\label{cond-eps-S}
Q(x)\geq-\log\epsilon(x),\quad\text{as }x\to P_{0}.
\end{equation}
This implies in particular that $Q(P_{0})=\infty$. We next state a weighted Bernstein-Walsh lemma on the sphere.
\begin{theorem}\label{BW-S}
Let $K$ be a closed non log-polar subset of ${\bf S}$ and $Q\in \AA(K)$. Let
\begin{equation}\label{pol-S}
p_{k}(x)=\prod_{j=1}^{k}|x-x_{j}|,\qquad x\in {\bf S},
\end{equation}
where $x_{1},\ldots,x_{k}\in {\bf S}$ and assume
$$|p_{k}(x)e^{-kQ(x)}|\leq M\quad\text{for }x\in S_{w}\setminus P\text{ where }P
\text{ is log-polar (possibly empty)}.$$
Then
$$|p_{k}(x)|\leq M\exp(k(-U^{\mu_{K,Q}}(x)+F_{w})),\qquad x\in {\bf S},$$
and
$$|p_{k}(z)e^{-kQ(z)}|\leq M \quad\text{for }x\in K\setminus \tilde P\text{ where }
\tilde P\text{ is log-polar (possibly empty)}.$$
\end{theorem}
\begin{proof}
Using the stereographic map $T$ defined in (\ref{stereo}), this theorem is a translation of Theorem \ref{BW-C} on $\mathbb{C}$.
\end{proof}
We will also need a lemma related to where the $L^{p}$ norm of a weighted ``polynomial'' lives, see \cite[III, Theorem 6.1]{ST}, \cite[Theorem 6.1]{B2} for polynomials on $\mathbb{C}$.
For $w=e^{-Q}$, we set
$$S_{w}^{*}=\{x\in K,~U^{\mu_{K,Q}}(x)+Q(x)\leq F_{w}\}.$$
Note that, as $U^{\mu_{K,Q}}(x)+Q(x)$ is lower semicontinuous, $S_{w}^{*}$ is a closed subset of ${\bf S}$ which, moreover, does not contain $P_{0}$. Indeed, $U^{\mu_{K,Q}}(x)$ is bounded below while $Q(x)$ tends to infinity as $x$ tends to $P_{0}$. Moreover, from Theorem \ref{Frost-C}, $S_w\subset S_w^*$.
\begin{lemma}\label{Tom}
Let $p>0$, $K$ a non log-polar compact subset of ${\bf S}$ containing $P_{0}$, $Q\in \AA(K)$ and $\nu$ a positive measure on $K$ satisfying
(\ref{cond-nu-inf})--(\ref{cond-eps-S}). We assume that $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property. Let $N\subset K$ be a closed neighborhood of $S_{w}^{*}$. Then, there exists a constant $c>0$ independent of $k$ and $p$ such that, for all expressions $p_{k}$ of the form (\ref{pol-S}),
$$\int_{K}|p_{k}e^{-kQ}|^{p}d\nu\leq(1+\mathcal{O}(e^{-ck}))\int_{N}|p_{k}e^{-kQ}|^{p}d\nu. $$
\end{lemma}
\begin{proof}
We normalize $p_{k}$ so that $\|p_{k}e^{-kQ}\|_{S_{w}^{*}}=1$. It is sufficient to show that there exists a constant $c>0$ such that for $k$ large,
\begin{equation}\label{ineg-K-N}
\int_{K\setminus N}|p_{k}e^{-kQ}|^{p}d\nu\leq e^{-ck},
\end{equation}
and that, for every $\epsilon>0$ and $k$ large,
$$\int_{K}|p_{k}e^{-kQ}|^{p}d\nu\geq e^{-\epsilon k}.$$
For the second inequality, we use the $L^{p/2}$--Bernstein-Markov property (recall Remark \ref{lp}) which gives
$$\int_{K}|p_{k}e^{-kQ}|^{p}d\nu\geq M_{2k}^{-p/2}\|p_{k}e^{-kQ}\|_{K}^{p}\geq
M_{2k}^{-p/2}\|p_{k}e^{-kQ}\|_{S_{w}^{*}}^{p}=M_{2k}^{-p/2}\geq e^{-\epsilon k},$$
for $k$ large, where we notice that $p_{k}^{2}$ is a real polynomial of degree $2k$.
For the first inequality, we use Theorem \ref{BW-S}
and the fact that $S_{w}\subset S_{w}^{*}$. This implies that, for $x\in K$,
$$|e^{-kQ(z)}p_{k}(x)|\leq\|e^{-kQ}p_{k}\|_{S_{w}}e^{-k(U^{\mu_{K,Q}}+Q-F_{w})}\leq
e^{-k(U^{\mu_{K,Q}}+Q-F_{w})}.$$
Since $U^{\mu_{K,Q}}$ is bounded below on $K$, there exists a constant $b_{0}$ such that
$$-U^{\mu_{K,Q}}(x)-Q(x)+F_{w}\leq\log\epsilon(x)+b_{0},\quad x\in K,$$
and, as $\epsilon(x)$ tends to 0 as $x$ tends to $P_{0}$, there exists
a neighborhood $V_{P_{0}}$ of $P_{0}$ such that $e^{b_{0}}\epsilon(x)^{1/2}<1$ for $x\in V_{P_{0}}$. On the other hand, since $N$ is a closed neighborhood of $S_{w}^{*}$ and $-U^{\mu_{K,Q}}-Q$ is upper semicontinuous, there exists a constant $b_{1}>0$ such that
$$-U^{\mu_{K,Q}}(x)-Q(x)+F_{w}\leq -b_{1}<0,\quad x\in K\setminus N.
$$
From this we deduce that
\begin{align*}\int_{K\setminus N}|p_{k}e^{-kQ}|^{p}d\nu & =
\int_{V_{P_{0}}}|p_{k}e^{-kQ}|^{p}d\nu+
\int_{K\setminus (N\cup V_{P_{0}})}|p_{k}e^{-kQ}|^{p}d\nu\\
& \leq \int_{V_{P_{0}}}e^{kpb_{0}}\epsilon(x)^{kp}d\nu+e^{-kpb_{1}}\nu(K\setminus V_{P_{0}})\\
& \leq \|e^{b_{0}}\epsilon(x)^{1/2}\|^{kp}_{V_{P_{0}}}\int_{K}\epsilon(x)^{a}d\nu
+e^{-kpb_{1}}\nu(K\setminus V_{P_{0}}),
\end{align*}
for $k$ large, which implies (\ref{ineg-K-N}).
\end{proof}
We are now in a position to prove an extended version of Proposition \ref{weightedtd}.
\begin{proposition}\label{ZQ-inf}
Let $K$ be a non log-polar compact subset of ${\bf S}$ containing $P_{0}$, $Q\in \AA(K)$ and $\nu$ a positive measure on $K$ satisfying
(\ref{cond-nu-inf})--(\ref{cond-eps-S}). We assume that $(K,\nu,Q)$ satisfies a weighted Bernstein-Markov property. Then the $L^{2}$ normalization constants $Z_{k}^{Q}$ defined in (\ref{L2}) are finite and
$$\lim_{k\to \infty}(Z_{k}^{Q})^{1/k(k-1)}=\delta^{Q}(K).$$
\end{proposition}
\begin{proof}
In view of (\ref{cond-nu-S}) and (\ref{cond-eps-S}), it is clear that, for $k$ large, the integral defining $Z_{k}^{Q}$ is finite.
The lower bound,
$$ \delta^{Q}(K)\leq \liminf_{k\to \infty}(Z_{k}^{Q})^{1/k(k-1)},$$
is proved as in the proof of Proposition \ref{weightedtd} by making use of the weighted Bernstein-Markov property. For the upper bound,
\begin{equation}\label{up-Z-S}
\limsup_{k\to \infty}(Z_{k}^{Q})^{1/k(k-1)}\leq\delta^{Q}(K),
\end{equation}
we first note that the expression $|VDM_{k}^{Q}({\bf X_{k}})|^{2}$ is, in each variable, of the form $e^{-2(k-1)Q}|q|^{2}$ with $|q|$ as in (\ref{pol-S}) for $k-1$. Hence, by using Lemma \ref{Tom} with $p=2$ for each of the $k$ variables, and with $N\subset K$ a closed neighborhood of $S_{w}^{*}$ as in Lemma \ref{Tom} with $\nu(N)<\infty$ and $P_0\not \in N$, we get
\begin{align*}
Z_{k}^{Q}=\int_{K^{k}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}}) & \leq
(1+\mathcal{O}(e^{-c(k-1)}))^{k}\int_{N^{k}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}})\\
& \leq (1+\mathcal{O}(e^{-c(k-1)}))^{k}(\delta^{Q}_{k}(K))^{k(k-1)}\nu(N)^{k},
\end{align*}
which implies (\ref{up-Z-S}) by taking the $k(k-1)$-th root and letting $k$ go to infinity.
\end{proof}
The next goal is to generalize Corollary \ref{johansson}. \begin{corollary}
We assume that the conditions (\ref{cond-nu-inf})--(\ref{cond-eps-S}) are satisfied and that $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property.
Then, with the notation of Corollary \ref{johansson}, there exist a constant $c>0$ and $k^*=k^*(\eta)$ such that for all $k>k^*$,
\begin{equation}\label{ineg-Prob}
Prob_k(K^{k}\setminus A_{k,\eta})\leq
\Big(1-\frac{\eta}{2 \delta^{Q}(K)}\Big)^{k(k-1)}\nu(N)^{k}+\mathcal{O}(e^{-ck}),
\end{equation}
where $N\subset K$ is a closed neighborhood of $S_{w}^{*}$ as in Lemma \ref{Tom} with $\nu(N)<\infty$ and $P_0\not \in N$.
\end{corollary}
\begin{proof}
We set
$B_{k,\eta}:=K^{k}\setminus A_{k,\eta}$ and decompose the integral in
$$Prob_k(K^{k}\setminus A_{k,\eta})=\frac{1}{Z_{k}^{Q}}\int_{B_{k,\eta}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}})$$
as a sum of two integrals over $B_{k,\eta}\cap N^{k}$ and $B_{k,\eta}\cap (K^{k}\setminus N^{k})$. Recalling the definition of the set $A_{k,\eta}$, the first term is less than
$$\Big(1-\frac{\eta}{2 \delta^{Q}(K)}\Big)^{k(k-1)}\nu(N)^{k},$$
for $k$ large. The second term is less than
$$\frac{1}{Z_{k}^{Q}}\int_{K^{k}\setminus N^{k}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}})\leq\sum_{j=1}^{k}\frac{1}{Z_{k}^{Q}}\int_{U_{j}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}}),$$
where $U_{j}=K\times\cdots\times(K\setminus N)\times\cdots K$ and the subset $K\setminus N$ is in $j$-th position.
As already observed in the previous proof, the expression $|VDM_{k}^{Q}({\bf X_{k}})|^{2}$ is, in each variable, of the form $e^{-2(k-1)Q}|q|^{2}$ with $|q|$ as in (\ref{pol-S}) for $k-1$. Hence, applying Lemma \ref{Tom} with the $j$-th variable to the integral over $U_{j}$, we get the upper bound
$$\mathcal{O}(e^{-c(k-1)})\sum_{j=1}^{k}\frac{1}{Z_{k}^{Q}}\int_{V_{j}}|VDM_{k}^{Q}({\bf X_{k}})|^{2}d\nu({\bf X_{k}}),$$
where $V_{j}=K\times\cdots\times N\times\cdots K$. Replacing $N$ with $K$ we finally get the upper bound $\mathcal{O}(ke^{-c(k-1)})$,
which implies (\ref{ineg-Prob}) with a different $c$.
\end{proof}
The last result that needs to be extended is the first item of Lemma \ref{lem-J-JQ}, namely, that for $Q\in \AA(K)$, and $\nu$ satisfying (\ref{cond-nu-inf})--(\ref{cond-eps-S}),
\begin{equation}\label{ineq-J-infty}
\overline J^{Q}(\mu)\leq e^{-I^Q(\mu)}.
\end{equation}
With the notation of Section \ref{J-funct}, we remark that fixing $a >0$ as in (\ref{cond-nu-S}), we can write
$$\int_{\tilde G_{k}}|VDM^{Q}_{k}({\bf a})|^{2}d\nu({\bf a})=\int_{\tilde G_{k}}|VDM^{Q_{k}}_{k}({\bf a})|^{2}d\tilde\nu({\bf a}),$$
where
$$Q_{k}(x)=Q(x)-\frac{a}{2(k-1)}Q_{+}(x),\quad
\quad\tilde\nu(x)=e^{-a Q_{+}(x)}\nu(x),$$
and $Q_{+}=\max(Q,0)$.
Observe that $\{Q_{k}\}_{k}$ is an increasing sequence of admissible weights that converges pointwise to $Q$ as $k$ tends to infinity. Also, in view of (\ref{cond-nu-S}) and (\ref{cond-eps-S}), $\tilde\nu$ is a finite measure. Since
$$\int_{\tilde G_{k}}|VDM^{Q_{k}}_{k}({\bf a})|^{2}d\tilde\nu({\bf a})\leq
\int_{\tilde G_{k}}|VDM^{Q_{k_{0}}}_{k}({\bf a})|^{2}d\tilde\nu({\bf a}),\quad k\geq k_{0},$$
we have
$$\int_{\tilde G_{k}}|VDM^{Q}_{k}({\bf a})|^{2}d\nu({\bf a})\leq \int_{\tilde G_{k}}|VDM^{Q_{k_{0}}}_{k}({\bf a})|^{2}d\tilde\nu({\bf a}),\quad k\geq k_{0}.$$
By letting $k$ go to infinity in this inequality and taking the infima over all neighborhoods $G$ of a measure $\mu$ in $\mathcal M(K)$, we obtain
$$J^{Q}_{\nu}(\mu)\leq J^{Q_{k_{0}}}_{\tilde\nu}(\mu),$$
where, here, the subscript denotes the measure with respect to which the Vandermonde is integrated. Since $\tilde\nu$ is of finite mass, we derive from item 1. of Lemma \ref{lem-J-JQ} that
$$J^{Q}_{\nu}(\mu)\leq e^{-I^{Q_{k_{0}}}(\mu)}.$$
Letting $k_{0}$ go to infinity, and making use of
$$\int Q_{k_{0}}d\mu\to\int Qd\mu,\quad\text{as }k_{0}\to\infty,$$
which follows from the monotone convergence theorem,
we obtain (\ref{ineq-J-infty}).
From the results above and the proofs of the previous sections, one may check that
Theorem \ref{rel-J-E} extends to the measures $\nu$ considered in this section.
For future reference, we state this as a theorem.
\begin{theorem}\label{rel-JE-S}
Let $K$ be a non log-polar compact subset of the sphere ${\bf S}$ in $\mathbb{R}^{3}$, containing $P_{0}$, $Q\in \AA(K)$ and $\nu$ a positive measure on $K$ satisfying (\ref{cond-nu-inf})--(\ref{cond-eps-S}). Assume $\nu$ satisfies a (strong) Bernstein-Markov property. Then,
$$\log\overline J^{Q}(\mu)=\log\underline J^{Q}(\mu)=-I^{Q}(\mu).$$
\end{theorem}
Also, the large deviation principle asserted in
Theorem \ref{ldp} extends
\begin{theorem}\label{ldp-S}
Let $K$ be a compact subset of the sphere ${\bf S}$ in $\mathbb{R}^{3}$, containing $P_{0}$, $Q\in \AA(K)$ and $\nu$ a positive measure on $K$ satisfying (\ref{cond-nu-inf})--(\ref{cond-eps-S}) and (\ref{nuas}). Then, the large deviation principle from Theorem \ref{ldp} holds.
\end{theorem}
\begin{remark} In Theorems \ref{rel-JE-S} and \ref{ldp-S}, the conclusion is valid for any $\nu$ satisfying a (strong) Bernstein-Markov property and (\ref{cond-nu-inf})--(\ref{cond-eps-S}) (with any appropriate function $\epsilon(x)$). Moreover, the rate function in Theorem \ref{ldp-S} is independent of $\nu$.
\end{remark}
\section{The $J^{Q}$ functionals on unbounded sets in $\mathbb{C}$}
We return to the case of unbounded sets in $\mathbb{C}$; our goal is to use Theorems \ref{rel-JE-S} and \ref{ldp-S} to derive their versions in our current setting. In the sequel we will need the Bernstein-Markov property on $\mathbb{C}$. For $K$ a closed subset of $\mathbb{C}$, $\nu$ a positive measure on $K$, locally finite but possibly of infinite mass in a neighborhood of infinity, and $Q$ a weakly admissible weight as in (\ref{weak-adm}), we say that $(K,\nu,Q)$ satisfies the Bernstein-Markov property if
\begin{equation}\label{unboundedBM}
\forall p_{k}\in\mathcal P_{k}(\mathbb{C}),\quad\|e^{-kQ}p_{k}\|_{K}\leq M_{k}\|e^{-kQ}p_{k}\|_{L^{2}(\nu)},\quad \text{with }\limsup_{k\to \infty}M_{k}^{1/k}=1.
\end{equation}
As in Section \ref{Sec-BM}, if (\ref{unboundedBM}) holds true for any continuous weakly admissible weight $Q$, we will say that $\nu$ satisfies a strong Bernstein-Markov property. Note that the polynomials $p_{k}$ in (\ref{unboundedBM}) are polynomials with respect to the \emph{complex variable}~$z$.
\begin{example}\label{example-BM}
For $Q$ a continuous admissible weight on $\mathbb{R}$, the linear Lebesgue measure $d\lambda$ provides an example of a measure with unbounded support satisfying (\ref{unboundedBM}). Indeed, from Theorem \ref{BW-C}, the sup norm of $e^{-kQ}p_{k}$ is attained on $S_{w}$ which is compact. Hence, it suffices to prove (\ref{unboundedBM}) on $S_{w}$ or any compact set containing $S_{w}$, for instance a finite interval $I$ (see Remark \ref{strongbm}). For $Q$ a continuous admissible weight on $\mathbb{C}$,
similar reasoning shows that planar Lebesgue measure $dm$ on $\mathbb{C}$ satisfies
(\ref{unboundedBM}) as well (in this case one considers the restriction of $dm$ to a closed disk).
Next, let $Q$ be an admissible weight on $K=\{x\in \mathbb{R}: x\geq 0\}$ which is continuous except $Q(0)=+\infty$. In this case, property (c) of Theorem I.1.3 \cite{ST} shows that $S_w$, the support of $\mu_{K,Q}$, will be compact and disjoint from the origin. Thus linear Lebesgue measure similarly satisfies (\ref{unboundedBM}). Specific examples are Laguerre weights $Q(x)=\lambda x -s\log x$ with $\lambda,s>0$ which occur in the Wishart ensemble (see \cite{hpbook}, section 5.5); and $Q(x)=c(\log x)^2$ with $c\geq 0$, occurring in the Stieltjes-Wigert ensemble (see \cite{[T]} and \cite{[ESS]}).
\end{example}
\begin{remark}\label{BM-Simeo}
For future use we observe that, if $\nu$ has infinite mass in a neighborhood of infinity, then the Bernstein-Markov property (\ref{unboundedBM}) is automatically satisfied if we restrict, for each $k$, to polynomials $p_{k}$ of \emph{exact degree} $k$ and the weight $Q$ satisfies a condition slightly stronger than weak admissibility (\ref{weak-adm}), namely
\begin{equation}\label{simeonov}
\lim_{z\in K,~|z|\to\infty}(Q(z)-\log|z|)=M<\infty.
\end{equation}
Indeed, for $p_{k}$ a monic polynomial of degree $k$, $|e^{-kQ(z)}p_{k}(z)|$ behaves like the constant $e^{-kM}>0$ as $z\to\infty$, so that its sup norm on $K$ is finite while its $L^{2}(\nu)$-norm is infinite. Hence we can take $M_{k}=1$ for each $k\geq 0$.
\end{remark}
Next, we define the (weighted) $L^{2}$ normalization constants for a closed subset $K$ of $\mathbb{C}$, $Q$ weakly admissible and $\nu$ a positive measure on $K$,
\begin{equation}\label{L2-C}
Z_{k}^{Q}(K,\nu):=\int_{K^{k}}|VDM_{k}^{Q}({\bf Z_k})|^{2}d\nu({\bf Z_k}),
\end{equation}
where ${\bf Z_k}:=(z_1,...,z_k)\in K^{k}$. Then we have the correspondence
$$Z_{k}^{Q}(K,\nu)=Z_{k}^{\tilde Q}(T(K),T_{*}\nu)$$
where $\tilde Q$ is defined in (\ref{tildeq}).
To ensure the finiteness of $Z_{k}^{Q}(K,\nu)$ in case $\nu$ has infinite mass in a neighborhood of infinity, like, e.g., Lebesgue measure, we assume that
the weight $Q$ and the measure $\nu$ satisfy conditions that correspond via the inverse of $T$ to the conditions (\ref{cond-nu-S}) and (\ref{cond-eps-S}) on the sphere, namely,
\begin{equation}\label{cond-nu}
\exists a>0,\quad\int_{K}\epsilon(z)^{a}d\nu(z)<\infty,
\end{equation}
and
\begin{equation}\label{cond-eps}
Q(z)-\log|z|\geq-\log\epsilon(z),\quad\text{as }z\to\infty,
\end{equation}
where $\epsilon(z)$ is some nonnegative continuous function that tends to 0 as $z$ tends to $\infty$. Note that the weight $Q$ is then admissible in the sense of \cite{ST} or 2. of Definition \ref{admit}, that is
\begin{equation}\label{admST}
Q(z)-\log|z|\to\infty,\quad\text{as }z\to\infty.
\end{equation}
Using the inequality
$$|z_{i}-z_{j}|\leq(1+|z_{i}|)(1+|z_{j}|),$$
one may also check directly that the $Z_{k}^{Q}(K,\nu)$, $k$ large, are, indeed, finite.
In the typical example where $K=\mathbb{R}$ or $K=\mathbb{C}$ and $\nu$ is Lebesgue measure, $\epsilon(z)$ can be chosen as $|z|^{-\epsilon}$, $\epsilon>0$, and (\ref{cond-eps}) becomes the following \emph{strong admissibility} condition (recall 3. of Definition \ref{admit}):
$$
Q(z)-\log|z|\geq\epsilon\log|z|,\quad\text{as }z\to\infty.
$$
Our next result is a version of Propositions \ref{weightedtd} and \ref{ZQ-inf} on the $k(k-1)$-th root asymptotic behavior of the $L^{2}$ normalization constants for $K$ a closed subset of $\mathbb{C}$.
\begin{proposition}\label{ZQ-C}
Let $K$ be a nonpolar closed subset of $\mathbb{C}$ and $\nu$ a positive measure on $K$. Let $Q$ be a weight on $K$ which is weakly admissible if $\nu$ has finite mass and such that (\ref{cond-nu}) and (\ref{cond-eps}) are satisfied for some function $\epsilon(z)$ if $\nu$ has infinite mass in a neighborhood of infinity. We assume that $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property (\ref{unboundedBM}).
Then,
$$
\lim_{k\to \infty}(Z_{k}^{Q})^{1/k(k-1)}=\delta^{Q}(K).$$
\end{proposition}
\begin{proof}
Via the inverse map of $T$, the statement is essentially a simple translation of Propositions \ref{weightedtd} and \ref{ZQ-inf}. The only observation to be made is that,
for the proof of Proposition \ref{weightedtd} on the sphere, it suffices that
a weighted Bernstein-Markov property is satisfied with respect to polynomials $p$ of the particular form
\begin{equation}\label{special-pol}
p(x)=\prod_{j=1}^{k}|x-T(z_{j})|^{2},\quad x\in {\bf S},
\end{equation}
and that this Bernstein-Markov property corresponds to (\ref{unboundedBM}) via $T^{-1}$.
Also, in the proof of Proposition \ref{ZQ-inf}, it is sufficient to use a version of Lemma \ref{Tom} which only assumes the Bernstein-Markov property for polynomials of the form (\ref{special-pol}) (and thus only holds for such polynomials).
\end{proof}
Weighted $J$-functionals $\underline J^{Q}(\mu)$ and $\overline J^{Q}(\mu)$ can be defined on the closed subset $K$ of $\mathbb{C}$, with respect to a positive measure $\nu$ in $K$, as was done on compact subsets of $\mathbb{R}^{n}$, see Definition \ref{jwmuq}. Then,
$$\underline J^{Q}(\mu)=\underline J^{\tilde Q}(T_{*}\mu),\quad
\overline J^{Q}(\mu)=\overline J^{\tilde Q}(T_{*}\mu),$$
where the $J$-functionals on the right-hand sides involve integrals with respect to the measure $T_{*}\nu$. From this correspondence, and Theorems \ref{rel-J-E} and \ref{rel-JE-S}, we derive the following.
\begin{theorem}\label{85}
With the hypotheses of Proposition \ref{ZQ-C} and assuming that $\nu$ satisfies a strong Bernstein-Markov property on $K$, we have
$$\log\underline J^{Q}(\mu)=\log\overline J^{Q}(\mu)=-I^{Q}(\mu).$$
\end{theorem}
\begin{proof}
The statement is a translation of Theorems \ref{rel-J-E} and \ref{rel-JE-S} on the sphere. Again, we observe that the strong Bernstein-Markov property for polynomials of the form (\ref{special-pol}) is sufficient for their proofs. We also use
the equality of the weighted logarithmic energies (\ref{enerK}) and (\ref{enerTK}).
\end{proof}
\noindent The conclusion is valid for all $\nu$ satisfying the hypotheses; in particular, the functional $J^Q\ (=\underline J^{Q}=\overline J^{Q}) \ =e^{-I^Q}$ for any such $\nu$.
\section{Large deviation principle for unbounded sets in $\mathbb{C}$}
A large deviation principle in the spirit of Theorem \ref{ldp} for compact subsets of $\mathbb{C}^m, \ m\geq 1$ has been obtained in \cite{PELD} using the methods of this paper (see also \cite{VELD}). Yattselev \cite{Y} has proved an LDP for a regular compact set $K\subset \mathbb{C}$ with connected complement and $K=\overline K^o$ when $Q$ is a specific type of weight; he uses Lebesgue measure on $K$. The large deviation principle for strongly admissible weights $Q$ on all of $\mathbb{C}$ with Lebesgue measure can be found in the book of Hiai and Petz \cite{hpbook}. There they extend the method of Ben Arous and Guionnet \cite{BG}.
Here, we will utilize the results from the previous sections to establish a LDP in the setting of a closed set $K$ in $\mathbb{C}$, not necessarily bounded, with a weakly admissible weight $Q$ and an appropriate Bernstein-Markov measure.
The proof is based on a standard contraction principle in LDP theory:
\begin{theorem}[{\cite[Theorem 4.2.1]{DZ}}]\label{contract}
If $\{P_n\}$ is a sequence of probability measures on a Polish space $X$ satisfying an LDP with speed $\{a_n\}$ and rate function ${\mathcal I}$, $Y$ is another Polish space and $f:X \to Y$ is a continuous map, then $\{Q_n:=f_*P_n\}$ satisfies an LDP on $Y$ with the same speed and with rate function
\begin{equation}\label{contract} {\mathcal J}(y):=\inf \{{\mathcal I}(x):x\in X,
f(x)=y\}.
\end{equation}
\end{theorem}
For $K$ a closed, possibly unbounded, subset of $\mathbb{C}$, $\nu$ a locally finite measure on $K$, and $Q$ a weakly admissible weight on $K$, we define the measure $Prob_{k}$ on $K^{k}$ as in (\ref{probk}) for $K$ in $\mathbb{R}^{n}$ {and $j_k:K^k\to \mathcal M(K)$ as in (\ref{jk}).
The statement of the large deviation principle is as follows.
\begin{theorem} \label{ldp2}
Assume $(K,\nu,Q)$ satisfies the weighted Bernstein-Markov property (\ref{unboundedBM}).
If $\nu$ has finite mass, we also assume that $(K,\nu)$ satisfies a strong Bernstein-Markov property while if $\nu$ has infinite mass in a neighborhood of infinity, we assume that (\ref{cond-nu}) and (\ref{cond-eps}) are satisfied for some function $\epsilon(z)$.
Then the sequence $\{\sigma_k=(j_k)_*(Prob_k)\}$ of probability measures on $\mathcal M(K)$ satisfies a {\bf large deviation principle} with speed $k^{2}$ and good rate function $\mathcal I:=\mathcal I_{K,Q}$ where, for $\mu \in \mathcal M(K)$,
\begin{equation}\label{rate}
\mathcal I(\mu):=\log J^Q(\mu_{K,Q})-\log J^Q(\mu)=I^Q(\mu)-I^Q(\mu_{K,Q}).
\end{equation}
\end{theorem}
\noindent We emphasize again that, as in Theorem \ref{ldp-S}, the rate function is independent of the measure $\nu$.
\begin{proof}
We apply Theorem \ref{contract} to the homeomorphism $f=(T^{-1})_*$: thus to prove an LDP in the setting of a closed set $K$ in $\mathbb{C}$, not necessarily bounded, with a weakly admissible weight $Q$, it suffices, via this contraction principle, to use an LDP in the setting of a compact set $T(K)$ in ${\bf S}\subset \mathbb{R}^3$ with the admissible weight $\tilde Q$. This we have from Theorems \ref{ldp} and \ref{ldp-S}.
In case $\nu$ is
of infinite mass in a neighborhood of infinity, we observe that the strong Bernstein-Markov property of $(K,\nu)$ is not needed.
Indeed, the corresponding Bernstein-Markov property on $T(K)$ is only needed for polynomials that are Vandermonde expressions, hence of maximal degree, and for the weights $Q_{m}$ appearing in Corollary \ref{corapprox}. These weights are of the form $-U^{\mu_{m}}$ with $\mu_{m}\in\mathcal M(T(K))$ and the corresponding weights on $K$ satisfy (\ref{simeonov}) with
$M=-U^{\mu_{m}}(P_{0})<\infty$,
so that Remark \ref{BM-Simeo} applies.
Finally, the rate function $\mathcal I(\mu)$ is good because $I^{Q}(\mu)=I^{\tilde Q}(T_{*}\mu)$, the energy $I^{\tilde Q}$ is lower semicontinuous on the compact set $\mathcal M({\bf S})$, and $T_{*}$ is a homeomorphism.
\end{proof}
\begin{remark}\label{appldp} In particular (see Example \ref{example-BM}), we have a large deviation principle on $K=\{x\in \mathbb{R}: x\geq 0\}$ with Lebesgue measure for the Laguerre weights as well as for the weights occurring in the Stieltjes-Wigert ensembles.
\end{remark}
\section{Applications: $\beta$ ensembles}\label{appli}
Let $K$ be a closed subset of $\mathbb{C}$, $\nu$ a positive measure on $K$, and $Q$ a weakly admissible weight on $K$.
Classical models in random matrix theory involve probability distributions on $K^{k}$ of the form
\begin{equation}\label{distr-rmt}
\frac{1}{\hat Z_{\beta,k}^{Q}}\prod_{1\leq i<j\leq k}|z_{i}-z_{j}|^{2\beta}\prod_{i=1}^{k}e^{-2kQ(z_{i})} d\nu (z_i),
\end{equation}
where $\beta>0$ and the normalization constant $\hat Z_{\beta,k}^{Q}$ is
\begin{equation}\label{free-ener}
\hat Z_{\beta,k}^{Q}=\int_{K^{k}}\prod_{1\leq i<j\leq k}|z_{i}-z_{j}|^{2\beta}
\prod_{i=1}^{k}e^{-2kQ(z_{i})}d\nu(z_{i}).
\end{equation}
(We caution the reader that in \cite{H} and \cite{AGZ} the $2\beta$ is replaced by $\beta$). The probability distribution (\ref{distr-rmt}) and normalization constant $\hat Z_{\beta,k}^{Q}$ differ from the distribution in (\ref{probk}) and the $L^{2}$ normalization constant $Z_{k}^{Q}$, defined in (\ref{L2-C}), by the exponent $\beta$ and an additional factor
$\prod_{i}e^{-2Q(x_{i})}$ in its integrand.
One may check that all results from the previous sections remain true, with appropriate modifications, when we consider (\ref{distr-rmt}) and (\ref{free-ener}).
Actually, writing the products in (\ref{distr-rmt}) and (\ref{free-ener}) as the square of a weighted Vandermonde to the power $\beta$, the main modification consists in replacing the weight $Q$ with the weight $Q/\beta$ and to
use the Bernstein-Markov property in $L^{\beta}$ instead of $L^{1}$ as was done in Section \ref{Sec-BM}.
To be precise, because of the factor $k$, instead of $k-1$, in the exponential factors of (\ref{distr-rmt}) and (\ref{free-ener}), the Bernstein-Markov property to be satisfied for a given weight $Q$ here is
\begin{equation}\label{BM-free}
\forall p_{k}\in\mathcal P_{k}(\mathbb{C}),\quad\|e^{-(k+1)Q}p_{k}\|_{K}\leq M_{k}\|e^{-(k+1)Q}p_{k}\|_{L^{2}(\nu)},\quad \text{with }\limsup_{k\to \infty}M_{k}^{1/k}=1.
\end{equation}
This property is slightly weaker than (\ref{unboundedBM}) as it concerns only polynomials in $\mathcal P_{k}(\mathbb{C})$ instead of $\mathcal P_{k+1}(\mathbb{C})$, but it will make a minor difference in the assumptions of the large deviation principle because Remark \ref{BM-Simeo} no longer applies.
When $\nu$ has infinite mass in a neighborhood of infinity, the conditions (\ref{cond-nu}) and (\ref{cond-eps}) become
\begin{equation}\label{cond-nu-free}
\exists a>0,\quad\int_{K}\epsilon(z)^{a}d\nu(z)<\infty,
\end{equation}
and
\begin{equation}\label{cond-eps-free}
Q(z)-\beta\log|z|\geq-\log\epsilon(z),\quad\text{as }z\to\infty,
\end{equation}
where $\epsilon(z)$ is some nonnegative continuous function that tends to 0 as $z$ tends to $\infty$.
Based on the above remarks, one may check that we have the following analogue of Proposition \ref{ZQ-C} concerning the asymptotics of $\hat Z_{\beta,k}^{Q}$.
\begin{proposition
Let $K$ be a nonpolar closed subset of $\mathbb{C}$ and $\nu$ a positive measure on $K$. Let $Q$ be a weight on $K$ such that $Q/\beta$ is weakly admissible if $\nu$ has finite mass and such that (\ref{cond-nu-free}) and (\ref{cond-eps-free}) are satisfied for some function $\epsilon(z)$ if $\nu$ has infinite mass in a neighborhood of infinity. We assume that $(K,\nu,Q/\beta)$ satisfies the weighted Bernstein-Markov property (\ref{BM-free}).
Then
$$
\lim_{k\to \infty}(\hat Z_{\beta,k}^{Q})^{1/k(k-1)}=(\delta^{Q/\beta}(K))^{\beta}.$$
\end{proposition}
The following large deviation principle, an analogue of Theorem \ref{ldp2}, also holds true.
\begin{theorem} \label{ldp-free}
Let $Q/\beta$, $\beta>0$, be a weakly admissible weight on $K$ such that $(K,\nu,Q/\beta)$ satisfies the weighted Bernstein-Markov property (\ref{BM-free}).
We assume that $(K,\nu)$ satisfies a strong Bernstein-Markov property and, in addition, if $\nu$ has infinite mass in a neighborhood of infinity, we assume that (\ref{cond-nu-free}) and (\ref{cond-eps-free}) are satisfied for some function $\epsilon(z)$.
Then the sequence
of probability measures $\tilde\sigma_{k}$ on $\mathcal M(K)$, defined so that for a Borel set $G\subset\mathcal M(K)$,
$$\tilde\sigma_{k}(G):=\frac{1}{\hat Z_{\beta,k}^{Q}}\int_{\tilde G_k}
\prod_{1\leq i<j\leq k}|z_{i}-z_{j}|^{2\beta}
\prod_{i=1}^{k}e^{-2kQ(z_{i})}d\nu(z_{i}),$$
satisfies a {\bf large deviation principle} with speed $k^{2}$ and good rate function $\mathcal I_{K,Q}^{\beta}$ defined by
$$
\mathcal I^{\beta}_{K,Q}(\mu):=I_{\beta}^{Q}(\mu)- I_{\beta}^{Q}(\mu_{K,Q/\beta}),
$$
where
$$I_{\beta}^{Q}(\mu)=\int_{K}\int_{K}\log\frac{1}{|z-t|^{\beta}}d\mu(z)d\mu(t)+
2\int_{K}Q(z)d\mu(z)=\beta I^{Q/\beta}(\mu),\qquad\mu\in\mathcal M(K).$$
\end{theorem}
\begin{proof}
One checks that all arguments in the proof of Theorem \ref{ldp2} go through when considering the probability distribution (\ref{distr-rmt}) instead of the one in (\ref{probk}). In particular, this entails verifying the analogue of Theorems \ref{rel-J-E} and \ref{85}, namely that, for appropriate assumptions on $\nu$ and $Q$, one has
$$\log \overline J_{\beta}^{Q}(\mu)=\log \underline J_{\beta}^{Q}(\mu)=-I_{\beta}^{Q}(\mu),$$
where the functionals $\overline J_{\beta}^{Q}$ and $\underline J_{\beta}^{Q}$ are derived from
$$J^Q_{\beta,k}(G):=\Big[\int_{\tilde G_{k}}|VDM^{Q/\beta}_k({\bf a})|^{2\beta}d\nu ({\bf a})\Big]^{1/k(k-1)},\qquad G\subset\mathcal M(K),$$
in the same way as in Definition \ref{jwmuq}. Since Remark \ref{BM-Simeo} does not apply for the Bernstein-Markov property (\ref{BM-free}), the strong Bernstein-Markov property is needed even if $\nu$ has infinite mass.
\end{proof}
As an example, we take $K=\mathbb{R}$, $d\nu=d\lambda=$ Lebesgue measure on $\mathbb{R}$, and $Q$ a continuous weight such that there exists $\beta'>\beta$ with $Q/\beta'$ weakly admissible:
\begin{equation}\label{q-beta}
\exists M>-\infty,\qquad\liminf_{|z|\to\infty, \ z\in K}(Q(z)-\beta'\log|z|)=M.
\end{equation}
Note that this implies that $Q/\beta$ is admissible. Also, (\ref{cond-nu-free}) and (\ref{cond-eps-free}) hold true with $\epsilon(z)=|z|^{\beta-\beta'}$. The triple
$(\mathbb{R},d\lambda,Q/\beta)$ satisfies the weighted Bernstein-Markov property since $Q/\beta$ is admissible, cf., Example \ref{example-BM}. The measure $d\lambda$ likely also satisfies a strong Bernstein-Markov property, but, as already mentioned in the proof of Theorem \ref{ldp2}, for an LDP it is sufficient that this property is satisfied for weights which correspond via $T$ to continuous weights on the sphere ${\bf S}$ of the form $-U^{\mu_{m}}$ where $\mu_{m}\in\mathcal M(T(\mathbb{R}))$ are the measures from Corollary \ref{corapprox}. Moreover, the proof of Lemma \ref{lemma-scal} shows that the supports of the measures $\mu_{m}$ can be chosen to avoid a neighborhood (depending on $m$) of the north pole $P_{0}$ so that the push-backward measures $\nu_{m}=T_{*}^{-1}\mu_{m}$ have compact supports in $\mathbb{C}$. Using the relations (\ref{pot-C-S}) and (\ref{tildeq}), what is then needed is that $d\lambda$ satisfies the Bernstein-Markov property for continuous weights of the form
$$Q_{m}(z)=-U^{\nu_{m}}(z)-\frac12\int\log(1+|t|^{2})d\nu_{m}(t),$$
where $\nu_{m}\in\mathcal M(\mathbb{R})$ has compact support. These weights are weakly admissible. Hence, by Theorem \ref{BW-C} and the continuity of $Q_{m}$, the corresponding weighted polynomials attain their sup norm on $S_{w_{m}}$ (where $w_{m}=e^{-Q_{m}}$), which is equal to the support of $\nu_{m}$, see Lemma \ref{lem-non-adm} (or more precisely its analogue in $\mathbb{C}$). Consequently, we need that $d\lambda$ satisfies the Bernstein-Markov property for continuous weights on a compact set, which we know holds true (cf., the discussion in Example \ref{example-BM}). Thus we conclude that
the large deviation principle asserted in Theorem \ref{ldp-free} applies on the real line for $d\nu=d\lambda$ and any continuous weight $Q$ satisfying (\ref{q-beta}). We note that this includes the large deviation principle for the law of the spectral measure of Gaussian Wigner matrices (\cite{BG}, \cite[Theorem 2.6.1]{AGZ}) as well as the refined version for weakly confining potentials given in \cite{H}.
When $K=\mathbb{C}$, $d\nu=dm=$ planar Lebesgue measure, and $Q$ is a continuous weight on $\mathbb{C}$ satisfying the growth condition (\ref{q-beta}), assumptions (\ref{cond-nu-free}) and (\ref{cond-eps-free}) still hold true with $\epsilon(z)=|z|^{\beta-\beta'}$. Moreover, the measure $dm$ satisfies the required Bernstein-Markov properties. Hence, Theorem \ref{ldp-free} applies when $K=\mathbb{C}$, $d\nu=dm$, and $Q$ is a continuous weight satisfying (\ref{q-beta}).
|
\section{Introduction}
The theory of dynamical systems is used in many fields of study to model the evolution of systems over time. However, most objects of real-world interest are too complicated to be modeled exactly by a dynamical system. The random fluctuations that occur in nature are conventionally disregarded when analysis is performed, a practice that may be dubious in some situations.
In \cite{ourpaper}, we have considered hybrid systems that consist of a set of continuous time dynamical systems over a compact space, where the dynamical system acting at each time is determined by Markov chain. We have taken $\{Z_n\}$ to be a randomly determined sequence of Markov states in some finite state space, and we have used this construction to find invariant probability measures.
In this paper, we change gears and instead treat $\{Z_n\}$ as a preselected, deterministic sequence of states. This approach to Markov chains has been studied in some detail in \cite{discretesystems}. We will adapt the results to our continuous time hybrid system, and we will see that under this new formulation, the hybrid system can be treated as a dynamical system rather than a stochastic one. We begin by trying to understand the space on which this dynamical system lives, showing that it is compact. As a result, we can discuss limit sets and Morse decompositions of the hybrid dynamical system, but we find that the limiting behavior of the system can be highly irregular. We explore three simple examples, illustrating the variety that is possible in the limiting behavior of hybrid dynamical systems.
\section{Directed Graphs and Dynamical Systems on Graphs}
The majority of the definitions, theorems, and notation in this section were borrowed from \cite{discretesystems}. The proof of all of these theorems can be found there as well. We introduce basic definitions for directed graphs associated with symbolic dynamical systems that will be used in the remainder of the paper.
\subsection{Directed Graphs}
\begin{defn}\label{def 1} A finite directed graph $G=(V,E)$ is a pair of sets $V=\{1,...,n\}$ called vertices, and $E\subseteq V \times V$, called edges.
\end{defn}
\begin{defn}\label{def 3} Any graph $G=(V,E)$ has a set ${\mathcal{P}} = \{(x_1,...,x_k),|(x_i,x_{i+1})\in E \ ; \ i,k \in \mathbb{N}\}$. Any element of ${\mathcal{P}}$ is called an admissible path of G.
\end{defn}
\begin{defn}\label{def 6}The out-degree of any ${\alpha}\in V$, denoted $o({\alpha})$, is the number of ${\gamma}\in{\mathcal{P}}$ with length 1 and ${\gamma}_1={\alpha}$.
\end{defn}
\begin{defn}\label{def 7}The in-degree of any ${\alpha}\in V$, denoted $i({\alpha})$, is the number of ${\gamma}\in{\mathcal{P}}$ with length 1 and ${\gamma}_F={\alpha}$.
\end{defn}
\begin{defn}\label{defNgraph} A finite directed graph $G=(V,E)$ with $o({\alpha})\geq1, i({\alpha})\geq1$; for all ${\alpha}\in V$ is called an $N$-graph.
\end{defn}
\begin{defn}\label{def 13} A communicating class in an $N$-graph $G=(V,E)$ is a subset $C\subseteq V$ for which two things are true:
\begin{enumerate}
\item For all ${\alpha},{\beta}\in C$ there exists ${\gamma}\in{\mathcal{P}}$ such that ${\gamma}_1={\alpha}$ and ${\gamma}_F={\beta}$.
\item There exists no $C'\supset C$ where for all ${\alpha}',{\beta}'\in C'$ there exists ${\gamma}'\in{\mathcal{P}}$ such that ${\gamma}'_1={\alpha}'$ and ${\gamma}'_F={\beta}'$. This condition is called maximality.
\end{enumerate}
\end{defn}
\begin{defn} Communicating classes can be classified further in two ways:
\begin{enumerate}
\item A communicating class $C$ is variant if there exists ${\gamma}\in{\mathcal{P}}$ with ${\gamma}_1\in C$ and ${\gamma}_F\not\in C$.
\item A communicating class $C$ is invariant if for all ${\gamma}\in{\mathcal{P}}$ with ${\gamma}_1\in C$, ${\gamma}_F\in C$.
\end{enumerate}
\end{defn}
\begin{remark}
Note that the empty set is not a communicating class because the empty set does not meet the requirements for maximality.
\end{remark}
\begin{thm}\label{thm 3}Every $N$-graph contains an invariant communicating class.
\end{thm}
\subsection{Dynamical Systems}
\begin{defn}
A dynamical system (d.s.) on a metric space $X$ is given by a map $\Phi : \, \mathbb{T} \times X \rightarrow X $ that satisfies $\Phi(0,x)=x$ and $\Phi(t+s,x)=\Phi(t,\Phi(s,x))$ for all $x\in X$ and all $t,s\in \mathbb{T}$. $\Phi$ can be expressed by two different but equivalent notations for $x, \, x' \in X$ and $t\in \mathbb{T}$: $$\Phi(t,x)= x' \, or\, \Phi_t(x)= x' $$ {\rm \cite{Kliemann}}.
\end{defn}
\begin{defn}
A d.s. is 1-sided when $\mathbb{T}=\mathbb{N}$ or $\mathbb{T}=\mathbb{R} ^+$.
A d.s. is 2-sided when $\mathbb{T}=\mathbb{Z}$ or $\mathbb{T}=\mathbb{R}$.
\end{defn}
\begin{lem}
Any 2-sided d.s. with mapping $\Phi_t$ has an inverse mapping $\Phi_{-t}$ where $$\Phi_t \circ \Phi_{-t}(x)=\Phi_{-t} \circ \Phi_{t}(x)=Id(x)=x$$
\end{lem}
\begin{defn}\label{omegalimit1}
The $\omega$-limit set of an element $x\in X$ is
$$\omega (x) = \{ y\in X | (\exists \, t_k \rightarrow\infty, k \in \mathbb{N})(\Phi (t_k,x)\rightarrow y)\}.$$
The $\alpha$-limit set of an element $x\in X$ is
$$\alpha (x) = \{ y \in X | (\exists \, t_k \rightarrow -\infty, k \in \mathbb{N})(\Phi (t_k,x)\rightarrow y)\}.$$
\end{defn}
\begin{defn}
The $\omega$-limit set of a subset $Y \subseteq X$ is given by
$$
\omega(Y) = \{ y \in X | (\exists \, t_k \rightarrow\infty, y_k \in Y, k \in \mathbb{N})(\Phi (t_k,y_k)\rightarrow y)\}.
$$
The $\alpha$-limit set of a subset $Y \subseteq X$ is given by
$$
\alpha(Y) = \{ y \in X | (\exists \, t_k \rightarrow -\infty, y_k \in Y, k \in \mathbb{N})(\Phi (t_k,y_k)\rightarrow y)\}.
$$\end{defn}
\subsection{Shift Spaces}
\begin{defn} Given an $N$-graph $G=(V,E)$, the bi-infinite product space $\Upsilon$ of the set $V=\{1,...,n\}$ is the set of all bi-infinite sequences $x=(...x_{-1},x_{0},x_1,...)$ where $x_i\in V$ for all $i\in \mathbb{Z}$. \end{defn}
\begin{defn} Given an $N$-graph $G=(V,E)$ with $A\subset V$ and ${\alpha}\in V$, we define:
\begin{itemize}
\item $\Omega=\{ \ (...,x_{-2},x_{-1},x_0,x_1,x_2,...) \ | \ (x_i,x_{i+1})\in E \ \}$ to be the shift space of $G$.
\item $\Omega_A=\{ \ x\in\Omega \ | \ x_i\in A $ for all $ i\in\mathbb{Z} \ \}$ to be the lift of $A$.
\end{itemize}
\end{defn}
\noindent The flow on this dynamical system is determined by the left shift mapping $\Phi$, defined in {\rm \cite{discretesystems}. It is in this paper that a metric is defined, and it is shown that this shift operator is continuous, and that $\Omega$ is compact.
\section{Generalization of the Shift Space to Continuous Time}
\subsection{$\bar \Delta$ and $\Delta$}
Note that in {\rm \cite{discretesystems}}, the flow on $\Omega$ is a discrete time dynamical system. However, to insert this behavior into another dynamical systems to create a hybrid system, it requires that we extend this system to a continuous time dynamical system. The obvious extension of a sequence into a function on $\mathbb{R}$ is a piecewise constant function.
\begin{defn}\label{lambda}
Let
\begin{center}
$\bar{\Lambda} = \left\{ x : \mathbb{R} \rightarrow S \left|
\begin{array}{cc}
\{x(ih)\}_{i \in \mathbb{Z}} \in \Upsilon \\
x$ is piecewise constant on $[nh,(n+1)h) \, \forall \, n \in \mathbb{Z}
\end{array} \right\} \right .$
\end{center}
and
$$\Lambda = \{ x(\cdot + t) | x \in \Lambda, t \in \mathbb{R}\} .$$
\end{defn}
\begin{defn}\label{delta}
Let
\begin{center}
$\bar{\Delta} = \left\{ x : \mathbb{R} \rightarrow S \left|
\begin{array}{cc}
\{x(ih)\}_{i \in \mathbb{Z}} \in \Omega \\
x$ is piecewise constant on $[nh,(n+1)h) \, \forall \, n \in \mathbb{Z}
\end{array} \right\} \right.$
\end{center}
and
$$\Delta = \{ x(\cdot + t) | x \in \Delta, t \in \mathbb{R}\} .$$
\end{defn}
\noindent
In other words, $\Lambda$ and $\Delta$ are the sets of functions that result from translating the functions in $\bar\Lambda$ and $\bar\Delta$, respectively, by some $t \in \mathbb{R}$. We allow for all horizontal translations of functions in $\bar\Lambda$ and $\bar\Delta$ in order for the spaces to be closed under shifts by $t$ for all $t\in \mathbb{R}$.
The next definition adapts the shift operator to continuous time by taking functions in $\bar \Delta$ as a generalization of bi-infinite sequences in $\Omega$.
\begin{defn}\label{psi}
Let
\begin{center}
$
\begin{array}{lr}
\psi: \mathbb{R} \times \bar \Delta \rightarrow \bar \Delta \\
\,\,\,\,\,\,\,\,(t, x(\cdot)) \mapsto x(\cdot + t)
\end{array}$
\end{center}
\end{defn}
Note that $\psi$ satisfies the flow property:
$$\psi(s+t, x(k)) = x(k + s + t) = x((k + t) + s) = \psi(s, x(k + t)) = \psi(s, \psi(t, x(k))) .$$
We now impose a metric on the set of functions $\bar{\Delta}$.
\begin{defn}\label{bardeltafunction}
Define the function
\begin{center}
$\begin{array}{lc}
f: \bar{\Delta} \times \bar{\Delta} \times \mathbb{Z} \rightarrow \mathbb{R} \\
\,\,\,\,(x, y, i) \mapsto \frac{1}{h}\displaystyle\int_{ih}^{(i+1)h}{\delta(x,y,t)dt}
\end{array}$
\end{center}
where
$$ \delta(x,y,t) = \left \{ \begin{array}{cc}
1 & \mbox{if } x(t) \neq y(t)\\
0 & \mbox{if } x(t) = y(t)
\end{array} \right.$$
\end{defn}
\begin{thm}\label{bardeltafunctionmetric}
The function
\begin{center}
$
\begin{array}{lc}
d: \bar{\Delta} \times \bar{\Delta} \rightarrow \mathbb{R} \\
\,\,\,\,\,\,\,\,\,(x,y) \mapsto \displaystyle \sum_{i = - \infty}^{\infty}\frac{f(x,y,i)}{4^{|i|}}
\end{array} $
\end{center}
is a metric on $\bar{\Delta}$.
\end{thm}
\begin{proof}
\ \\
\begin{enumerate}
\item (Non-negativity) $f(x, x, i) = 0$, for all $i \in \mathbb{Z}$. Therefore, $d(x,x) = 0$. For $x \neq y$, $f(x,y,i) \neq 0$ for at least one $i \in \mathbb{Z}$, and $f(x, y, i) \geq 0$ for all $i \in \mathbb{Z}$. Therefore, $d(x,y) > 0$ for all $x \neq y$.
\item (Symmetry) Clearly, $\delta(x,y,t) = \delta(y,x,t)$ for all $t \in \mathbb{R}, x, y \in \bar{\Delta}$. So, $f(x,y,i) = f(y,x,i)$, $d(x,y) = d(y,x)$.
\item (Triangle inequality) Choose $x,y,z\in\bar\Delta$. If $x=z$, then as $d(x,z)=0$ and $d$ is nonnegative, then clearly for all $y\in\bar\Delta$, $d(x,z)\leq d(x,y)+d(y,z)$.
If $x\not =z$, then there exists $t\in\mathbb{R}$ such that $x(t)\not = z(t)$. If $x(t) \neq z(t)$, then either $x(t) \neq y(t)$ or $y(t) \neq z(t)$. Therefore, $\delta(x,z,t) = 1$ implies that $\delta(x,y,t) = 1$ and/or $\delta(y,z,t) = 1$, so $f(x,z,i) \leq f(x,y,i) + f(y,z,i)$. Since $d$ is a linear combination of $f$'s, $d(x,z) \leq d(x,y) + d(y,z)$.
\end{enumerate}
\end{proof}
\begin{prop}
The mapping
$\sigma: \Omega \rightarrow \bar\Delta$ where $x\mapsto x(t)$ where $x(i)=x_i$ for all $i\in\mathbb{Z}$ is an isometric isomorphism.
\end{prop}
\begin{proof}
By the construction of $\bar\Delta$, $\sigma$ is clearly bijective.\\
To show that $\sigma$ is an isometry, it suffices to show that $f(x,y,i) = \bar{f}(x_i,y_i)$, where
$$ \bar{f}(x_i,y_i) =
\left\{ \begin{array}{cc}
1 & \mbox{if } x(i) \neq y(i)\\
0 & \mbox{if } x(i) = y(i)
\end{array} \right.$$
since the bi-infinite sums for $d$ and $\bar{d}$ are identical. Note that
$$f(x,y,i) = \frac{1}{h}\int_{ih}^{(i+1)h}{dt} = 1 = \bar{f}(x_i, y_i)$$
for $x(i) \neq y(i)$.
$$ f(x,y,i) = \frac{1}{h}\int_{ih}^{(i+1)h}{0 \cdot dt} = 0 = \bar{f}(x_i, y_i)$$
for $x(i) = y(i)$.
So indeed, $f(x,y,i) = \bar{f}(x_i,y_i)$, and $d(x,y) = \bar{d}(\{x_i\},\{y_i\})$.
\end{proof}
\begin{lem}\label{psicontinuous}
$\psi_t$ is continuous for all $t \in \mathbb{R}$.
\end{lem}
\begin{proof}
Given $x, y \in \bar{\Delta}$, we need to show that for all $\epsilon > 0$, there exists $\delta > 0$ such that
$$d(x,y) < \delta \Rightarrow d(\psi_t(x),\psi_t(y))<\epsilon.$$
Given any $\epsilon >0$, take $\delta = \frac{\epsilon}{4^n}$, where $n = \left\lceil|\frac{t}{h}|\right\rceil$, the least integer greater than the absolute value of $\frac{t}{h}$. It is useful to rewrite $d(x,y)$ in the form
$$ d(x,y) = \frac{1}{h} \int_{-\infty}^{\infty}{\frac{1}{4^{\left |\lfloor \frac{t'}{h}\rfloor \right |}}\delta(x,y,t')dt'}$$
where $\delta$ is as defined above. Given this, we can write
$$ d(\psi_t(x),\psi_t(y)) = \frac{1}{h} \int_{-\infty}^{\infty}{\frac{1}{4^{\left | \lfloor \frac{(t+t')}{h} \rfloor \right |}}\delta(x,y,t')dt'}$$
And,
$$ \frac{1}{4^{\left | \lfloor \frac{(t+t')}{h} \rfloor \right |}} \leq 4^{\left \lceil |\frac{t}{h}| \right \rceil} \frac{1}{4^{\left | \lfloor \frac{t'}{h} \rfloor \right|}}$$
So, $$ d(\psi_t(x),\psi_t(y)) = \frac{1}{h} \int_{-\infty}^{\infty}{\frac{1}{4^{\left |\lfloor \frac{(t+t')}{h} \rfloor \right |}}\delta(x,y,t')dt'}$$
$$ \leq 4^{\left \lceil |\frac{t}{h}| \right \rceil} \frac{1}{h} \int_{-\infty}^{\infty}{\frac{1}{4^{\left |\lfloor \frac{t'}{h} \rfloor \right|}}\delta(x,y,t')dt'}$$
$$ = 4^{\left \lceil |\frac{t}{h}| \right \rceil} d(x,y)$$
$$ < 4^{\left \lceil |\frac{t}{h}| \right \rceil} \delta$$
$$ = \epsilon$$
\end{proof}
\begin{lem}\label{deltabarcompact}
$\Delta$ is compact.
\end{lem}
\begin{proof}
We will show that given any sequence $\{x^n\}, n \in \mathbb{N}$ of functions $x^n \in \Delta$, there exists a subseqence converging to some $x \in \Delta$. To do this, we consider the space $\Delta$ to be the product of a circle of length $h$ with the set of allowable bi-infinite sequences, $S^1 \times \Omega$, where $S^1 \equiv \mathbb{R}$ mod $h$. An element $x^n$ of $\Delta$ identifies with an element of $S^1 \times \Omega$ by taking $y^n \in \Omega$ to be the sequence of constant values of $x^n$, with $x^n(0) \equiv y^n_0, x^n(h) \equiv y^n_1,$ etc., and taking $\tau \in [0,h)$ to be the unique offset so that $x^n(t-\tau) \in \bar{\Delta}$.
$S^1$ is compact. Therefore, given the sequence $\{x^n\} \in S^1 \times \Omega$, there exists a subsequence $\{x^{n_k}\}, k \in \mathbb{N}$ for which the offsets $\{\tau^{n_k}\}$ converge to a value in $[0,h)$. Therefore, there exists a convergent subsequence $\{\tau^{n_k}\}$ of $\{\tau^n\}$.
For this subsequence $\{x^{n_k}\}$, we want to show that there exists a subsequence $\{x^{n_{k_j}}\}, j \in \mathbb{N}$ such that the bi-infinite sequences $\{y^{n_{k_j}}\}$ converge. We do this inductively, beginning with the subsequence $\{y^{n_{k_j}}_0\}$. We know that $\{y^{n_k}_0\}$ is an infinite sequence of finitely many values, since the state space $S$ is finite. Therefore, by the pigeonhole principle, there is one value that is repeated infinitely many times. Take $\{y^{n_{k_j}}_0\}$ to be this value, $s_0$, so that $x^{n_{k_j}}(0) = s_0$ for all $j \in \mathbb{N}$.
Now, we induct. Given a subsequence of $\{x^{n_k}\}$ that converges at $t = 0, h, -h, 2h, -2h, ..., $ $mh, -mh$, we deduce that there must be a subsequence of this subsequence with one value $x^{n_k}((m+1)h) = s_{(m+1)h}$ repeated infinitely many times, and likewise for $x^{n_k}(-(m+1)h) = s_{-(m+1)h}$. In this manner, we get an infinite subsequence $\{y^{n_{k_j}}\}$, hence $\{x^{n_{k_j}}\}$, converging to a function that is piecewise constant on $[\tau + nh, \tau + (n+1)h), n \in \mathbb{Z}, \tau \in [0,h]$, with values in $S$.
Finally, we have to show closure. That is, we need to show that transitions $x(mh) \rightarrow x((m+1)h)$ in our limit function are allowable. Otherwise, all we would have shown is compactness of $\bar{\Lambda}$, rather than compactness of $\Delta$. Suppose that $x \notin \Delta$. Then, there exists some $m \in \mathbb{Z}$ such that the transition $x(mh) \rightarrow x((m+1)h)$ is not allowed. But, since $\{x^{n_{k_j}}\}$ converges to $x$, we can take $N$ large enough so that $j > N \Rightarrow x^{n_{k_j}}(mh) = x(mh), x^{n_{k_j}}((m+1)h) = x((m+1)h)$. And, $x^{n_{k_j}} \in \Delta$, so the transition $x^{n_{k_j}}(mh) \rightarrow x^{n_{k_j}}((m+1)h)$ must be allowable. This is a contradiction. Therefore, $\{x^{n_{k_j}}\} \rightarrow x \in \Delta$.
\end{proof}
\begin{comment}
Consider a dynamical system given by $\varphi : \mathbb{R}\times X \rightarrow X$, where $X$ is compact with $\sigma$-algebra $\vartheta$.
\begin{defn}\label{dynsysinvm}
A measure $\mu$ on $(X,\vartheta )$ is invariant under $\varphi$ if, for all sets $A\in \vartheta$ and all times $t\in\mathbb{R}$,
$$\mu (\varphi_t ^{-1} [A] ) = \mu (A) $$
or, equivalently, since $\varphi_t ^{-1} = \varphi_{-t}$,
$$\mu (\varphi_{-t} [A] ) = \mu (A) .$$
\end{defn}
\noindent
In other words, the set $A$ and its inverse image at time $-t$ should have the same measure.
\begin{defn}\label{borel measure}
Given $X$ compact, let $\vartheta$ be the smallest $\sigma$-algebra containing the open sets of $X$. Any measure $\mu$ defined on $\vartheta$ is called a Borel measure.
\end{defn}
\begin{thm}\label{krylov-bogolyubov}
{\rm \textbf{(Krylov-Bogolyubov)}} Let $(X, T)$ be a compact, metrizable topological space and $f : X \rightarrow X$ a continuous map. Then $f$ admits an invariant Borel probability measure {\rm \cite{sinai}}.
\end{thm}
\begin{thm}\label{tychonoffs theorem}
{\rm \textbf{(Tychonoff)}} Given two compact topological spaces $(X_1, T_1)$ and $(X_2, T_2)$, the product space with the product topology $(X_1 \times X_2,T_1 \times T_2)$ is also compact.
\end{thm}
\begin{thm}\label{invariant measure}
The dynamical system $(\varphi_t,\psi_t)$ yields an invariant Borel probability measure on the Borel $\sigma$-algebra $\vartheta$ of the space $M \times \bar{\Delta}$ with the product topology induced by the standard topology on $M$ and the metric $d$ on $\bar{\Delta}$.
\end{thm}
\begin{proof}
From Theorem \ref{tychonoffs theorem}, Lemma \ref{deltabarcompact}, and the assumption that $M$ is compact, $M \times \bar{\Delta}$ is a compact. From Lemma \ref{psicontinuous} and the fact that $\varphi_t$ is assumed to be the composition of continuous flows, $(\varphi_t, \psi_t): M \times \bar{\Delta} \rightarrow M \times \bar{\Delta}$ is continuous. Therefore, by Theorem \ref{krylov-bogolyubov}, $(\varphi_t, \psi_t)$ induces an invariant Borel probability measure on $\vartheta$.
\end{proof}
The invariant Borel probability measure $\mu$ on $M \times \bar{\Delta}$ induces an invariant marginal probability measure $\tilde{\mu}$ on $M$ given by $$\tilde{\mu}(A) = \mu(A,\bar{\Delta})$$
for $A \subseteq M$. In many circumstances, where the primary space of interest is $M$, $\tilde{\mu}$ may be more useful than $\mu$.
\end{comment}
\subsection{Morse Sets and Topological Chaos}
For the following definitions and Proposition \ref{order}, taken from \cite{Kliemann}, let $X$ be a compact metric space with an associated flow $\Phi$.
\begin{defn}\label{invariant}
A set $K \subseteq X$ is called invariant if $\Phi(t,x) \in K$ for all $x \in K, t \in \mathbb{R}$.
\end{defn}
\begin{defn}\label{isolated}
A set $K \subseteq X$ is called isolated if there exists a neighborhood $N$ of $K$ (i.e. a set $N$ with $K \subset$ int $N$) such that $\Phi(t,x) \in N$ for all $t \in \mathbb{R}$ implies $x \in K$.
\end{defn}
\begin{defn}\label{morsedecomp}
A Morse Decomposition on $X$ is a finite collection $\{\mathcal{M}_i, i = 1,...,n\}$ of non-void, pairwise disjoint, invariant, isolated, compact sets such that
\begin{enumerate}
\item For all $x \in X, \omega(x), \alpha(x) \subseteq \displaystyle \bigcup_{i=1}^{n}\mathcal{M}_i$.
\item If there exist $\mathcal{M}_0, \mathcal{M}_1, ..., \mathcal{M}_l$ and $x_1,...x_l \in X \setminus \displaystyle \bigcup_{i=1}^{n}\mathcal{M}_i$ with $\alpha(x_j) \subseteq \mathcal{M}_{j-1}$ and $\omega(x_j) \subseteq \mathcal{M}_j$ for $j = 1,...,l$, then $\mathcal{M}_0 \neq \mathcal{M}_l$. This condition is equivalent to to the statement that there are no cycles between the sets of the Morse decomposition.
\end{enumerate}
The sets $\mathcal{M}_i$ above are called Morse sets.
\end{defn}
\begin{prop}\label{order}
The relation $\preceq$ given by
$$
\mathcal{M}_i \preceq \mathcal{M}_k \mbox{ if there are } \mathcal{M}_i,\mathcal{M}_{j_1},...,\mathcal{M}_{j_l} =
\mathcal{M}_k \mbox{ and } x_1,...,x_l \in X $$
$$\mbox{ with } \alpha(x_m) \subseteq \mathcal{M}_{j_{m-1}} \mbox{ and } \omega(x_m) \subseteq \mathcal{M}_{j_m} \mbox{ for } m = 1,...,l.
$$
is an order (satisfying reflexivity, transitivity, and antisymmetry) on the Morse sets $\mathcal{M}_j$ of a Morse decomposition.
\end{prop}
The proof of this proposition can be found in \cite{Kliemann}.
\begin{defn}\label{lifts}
The lift $\bar{\Delta}_C \subseteq \bar{\Delta}$ of a communicating class $C$ is defined by
$$
\bar{\Delta}_C \equiv \{ f \in \bar{\Delta} | f(t) \in C \, \forall \, t \in \mathbb{R} \}
$$
$\Delta_C$ is defined as
$$
\Delta_C \equiv \bar{\Delta}_C \cap \Delta
$$
\end{defn}
\begin{comment}
\begin{defn}\label{def 11}A vertex ${\alpha}\in V$ is variant if there exists no ${\gamma}\in{\mathcal{P}}$ containing ${\alpha}$ which has ${\gamma}_1 ={\gamma}_F$.
\end{defn}
\begin{defn}\label{def 12}A vertex ${\beta}\in V$ is invariant if every ${\gamma}\in{\mathcal{P}}$ containing ${\beta}$ has ${\gamma}_F={\beta}$.
\end{defn}
\end{comment}
\begin{thm}\label{morse1}
The lifts of the communicating classes $\bar{\Delta}_C$ are Morse sets for the dynamical system $\psi$.
\end{thm}
\begin{proof}
We check the seven conditions in turn.
\begin{enumerate}
\item \textbf{Non-void}
Since the empty set is not a communicating class, the lift of any communicating class must be non-empty.
\item \textbf{Pairwise disjoint}
Suppose that there exists $f \in \bar{\Delta}_C, \bar{\Delta}_{C'}$ with $C \neq C'$. Then, $f(0) \in C, C'$. But, by the maximality of communicating classes, $f(0) \in C, C'$ implies $C = C'$. So, $\bar{\Delta}_{C} = \bar{\Delta}_{C'}$.
\item \textbf{Invariant}
By construction of $\bar{\Delta}$, $\psi(t,f) \in \bar{\Delta}$ for all $t \in \mathbb{R}, f \in \bar{\Delta}$. And, if $f(s) \in C$ for all $s \in \mathbb{R}$, then $\psi(t,f)(s) = f(t+s) \in C$. So, $\psi(t,f) \in \bar{\Delta}_C$ for all $t \in \mathbb{R}, f \in \bar{\Delta}_C$.
\item \textbf{Isolated}
Pick $\epsilon = 1/4$. Suppose that there exists $g \notin \bar{\Delta}_C$ such that for some $f \in \bar{\Delta}_C$, $d(g,f) < \epsilon$. Since $g \notin \bar{\Delta}_C$, there exists $t_0$ such that $g(t_0) \notin C$. Let $g' = \psi(-t_0,g)$, so that $g'(0) \notin C$. But then, $g'$ differs from any function in $\bar{\Delta}_C$ on at least some interval of length $h$ containing $0$. The distance, therefore, between $g'$ and any function in $\bar{\Delta}_C$ must be greater than $1/4$. So, given any $g \notin \bar{\Delta}_C$ but within $1/4$ of $\bar{\Delta}_C$, there exists $t_0$ such that $d(\psi(t_0, g),f') > \frac{1}{4}$ for any $f' \in \bar{\Delta}_C$. Hence, $\bar{\Delta}_C$ is isolated.
\item \textbf{Compact}
By an argument similar to that for compactness of $\Delta$ and by compactness of $\Omega_C$, $\bar{\Delta}_C$ is compact.
\item \textbf{No cycles}
Again, this is similar to the corresponding proof in \cite{discretesystems}. Suppose that there exist $f, g \in \bar{\Delta}$ such that $\alpha(f) \subseteq \bar{\Delta}_C$, $\alpha(g) \subseteq \bar{\Delta}_{C'}$ and $\omega(g) \subseteq \bar{\Delta}_C$, $\omega(f) \subseteq \bar{\Delta}_{C'}$. Then, since all the transitions in $f,g$ must be allowable, there must exist an admissible path from $C$ to $C'$ as well as one from $C'$ to $C$. But, this contradicts maximality of communicating classes. So, no such cycle exists.
\end{enumerate}
\end{proof}
\begin{defn}\label{topologicallytransitive}
A flow on a metric space $X$ is called topologically transitive if there exists $x \in X$ such that $\omega(x) = X$.
\end{defn}
\begin{lem}\label{omegalimit}
Given any communicating class $C$, there exists $f^* \in \bar{\Delta}_C$ such that $\omega(f^*) = \bar{\Delta}_C$ (i.e. $\psi$ is topologically transitive on lifts of communicating classes).
\end{lem}
\begin{proof}
It has been shown in \cite{discretesystems} that for the discrete system, there exists $x^* \in \Omega_C$ such that $\omega(x^*) = \Omega_C$. By the correspondence between sequences in $\Omega$ and functions in $\Delta$, there exists $f^* \in \Delta_C$ given by $$ f^*(nh) = x^*_n, n \in \mathbb{Z}$$ such that $\Delta_C \subseteq \omega(f^*)$. And, since $\bar{\Delta}_C$ is given by the shifts $\psi(t,\Delta_C)$, it is clear that $\bar{\Delta}_C \subseteq \omega(f^*)$. $\bar{\Delta}_C$ is invariant by Theorem \ref{morse1}, so $\omega(f^*) \subseteq \bar{\Delta}_C$, and $\omega(f^*) = \bar{\Delta}_C$.
\end{proof}
Since the $\omega$-limit sets of a point on a compact space are connected, we get the following corollary.
\begin{cor}\label{connected}
$\bar{\Delta}_C$ is connected.
\end{cor}
\begin{prop}\label{dense}
The set of all functions $f^*$ satisfying $\omega(f^*) = \bar{\Delta}_C$ is dense in $\bar{\Delta}_C$
\end{prop}
\begin{proof}
Given $f \in \bar{\Delta}_C$, there exists $f^*$ such that $f \in \omega(f^*)$, by Lemma \ref{omegalimit}. Therefore, given $\epsilon > 0$, there exists $t \in \mathbb{R}$ such that $d(\psi(t,f^*),f) < \epsilon$. $\omega(f^*) = \bar{\Delta}_C$ implies $\omega(\psi(t,f^*)) = \bar{\Delta}_C$. So, for any $f \in \bar{\Delta}_C$ and any $\epsilon > 0$, there exists a function $\psi(t,f^*) \in \bar{\Delta}_C$ with $d(\psi(t,f^*),f) < \epsilon$ and $\omega(\psi(t,f^*)) = \bar{\Delta}_C$.
\end{proof}
\begin{defn}\label{finer}
A Morse Decomposition $\{\mathcal{M}_1,...,\mathcal{M}_n\}$ is called finer than a Morse Decomposition $\{\mathcal{M}'_1,...,\mathcal{M}'_l\}$ if for all $j \in \{1,...,l\}$ there exists $i \in \{1,...,n\}$ such that $\mathcal{M}_i \subseteq \mathcal{M}'_j$, where containment is proper for at least one $j$.
\end{defn}
\begin{thm}\label{finest}
The lifts of the communicating classes $\bar{\Delta}_C$ form a finest Morse decomposition on $\bar{\Delta}$.
\end{thm}
\begin{proof}
Suppose there exists a finer Morse decomposition. Then, for some $C$, there exists a Morse set $K \subsetneq \bar{\Delta}_C$, a proper containment. Suppose first that there is only one such $K$. By the definition of a Morse set, $K$ must contain the $\omega$-limit sets of $\bar{\Delta}_C$.
But, by Lemma \ref{omegalimit}, there exists $f^* \in \bar{\Delta}_C$ such that $\omega(f^*) = \bar{\Delta}_C$. Therefore, $\bar{\Delta}_C \subseteq K$, so $K$ is not a proper subset of $\bar{\Delta}_C$.
Now, suppose that for this finer Morse decomposition, there exist several Morse sets $K_1,...,K_n$ such that each $K_i \subsetneq \bar{\Delta}_C$. By the definition of a Morse set, $K_1,...K_n$ are pairwise disjoint and compact. But, since $\omega(f^*) = \bar{\Delta}_C$, we must have $\displaystyle \bigcup_{i=1}^nK_i = \bar{\Delta}_C$. By Corollary \ref{connected}, $\bar{\Delta}_C$ is connected. And, the union of finitely many, pairwise disjoint compact sets cannot be connected. Thus, no finer Morse decomposition exists.
\end{proof}
\begin{defn}\label{sensitivedependence}
A flow $\Phi$ on a metric space $X$ has sensitive dependence on initial conditions if there exists $\delta > 0$ such that for every $x \in X$ and every neighborhood $B$ of $x$, there exists $y \in B$ and $t > 0$ such that $d(\Phi_t(x),\Phi_t(y)) > \delta$.
\end{defn}
\begin{defn}\label{chaotic}
A flow on a metric space $X$ is chaotic if it has sensitive dependence on initial conditions and is topologically transitive.
\end{defn}
\begin{lem}\label{sensdep}
Consider a graph $G$ consisting of a single communicating class $C$ for which the out-degree of at least one vertex is at least two. Then, $\psi$ on $\bar{\Delta}$ has sensitive dependence on initial conditions.
\end{lem}
\begin{proof}
Take $\delta = \frac{1}{2}$. Given $x \in \bar{\Delta}$, we construct a function $y \in \bar{\Delta}$ such that $x$ and $y$ are discontinuous at the same times mod $h$. Given $\varepsilon > 0$, take $N$ large enough so that
$$\displaystyle \sum_{i=-\infty}^{-N}{\frac{1}{4^{|i|}}} + \displaystyle \sum_{i=N}^{\infty}{\frac{1}{4^{|i|}}} < \varepsilon.$$
Thus, taking $x(t) = y(t)$ on $t \in [-Nh,Nh]$ ensures that $d(x,y) < \varepsilon$. Now, we just need to show that there exists $m>N \in \mathbb{R}$ so that $y(t) \neq x(t)$ for all $t \in [mh,(m+1)h)$. This would imply that $d(\psi(mh,x),\psi(mh,y)) \geq 1 > \delta$.
To show that such an $m$ exists, let $\gamma_1$ denote the vertex with out-degree greater than one. If there does not exist a $t > Nh$ such that $x(t) = \gamma_1$, then given $x(Nh) = \gamma_N$, let $y(t), t > Nh$ follow a path from $\gamma_N$ to $\gamma_1$. Such a path must exist since $G$ consists of a single communicating class, so there exists a path between any two vertices in $G$. Thus, we would have $x(t_0+t) \neq \gamma_1, y(t_0+t) = \gamma_1$ for some $t_0 > Nh, t \in [0,h)$, so we can take $m = \frac{t_0}{h}$. If there does exist a $t > Nh$ such that $x(t) = \gamma_1$, then define $\gamma_2 = x(t+h)$, and take $t_1$ so that $x(t_1+t)=\gamma_1$ for $t \in [0,h)$. Since the out-degree of $\gamma_1$ is greater than one, there exists an edge from $\gamma_1$ to some other vertex, $\gamma_3$ (note that it is possible that either $\gamma_1 = \gamma_2$ or $\gamma_1 = \gamma_3$, but not that $\gamma_2 = \gamma_3$). Set $y(t_1+h) = \gamma_3$, and take $m = \frac{t_1+h}{h}$.
\end{proof}
Applying Lemmas \ref{omegalimit} and \ref{sensdep} yields the following result.
\begin{thm}\label{chaos}
Consider a graph consisting of a single communicating class for which the out-degree of at least one vertex is greater than one. Then, $\psi$ is chaotic on $\bar{\Delta}$.
\end{thm}
\section{The Deterministic Hybrid System}
Now that the behavior on $\Delta$ has been determined given a natural number $n$ and an $N$-graph on $n$ vertices, we consider the action of a function $f\in\Delta$ on a set of $n$ dynamical systems.
Consider an $N$-graph $G$ with $n$ vertices. Take a collection of $n$ dynamical systems $\{\phi_1,...,\phi_n\}$ on a compact space $M \subset \mathbb{R}^d$, where each vertex of $G$ corresponds to one dynamical system $\phi_i$. Take $f \in \bar{\Delta}$. Define $\varphi(t,x,f):\mathbb{R} \times M \times \bar{\Delta} \rightarrow M$ by
$$
\varphi n(t,x,f) = \varphi_t(x,f)= \phi_{i_m}(\tau_{m}-\tau_{m-1},\phi_{i_{m-1}}(\tau_{m-1}-\tau_{m-2},...\phi_{i_{1}}(\tau_1,x)))
$$
where the $\tau_k$'s satisfy $0 = \tau_0 < \tau_1 <...< \tau_m = t$ and $f(\tau) = i_j$ for $\tau \in [\tau_{j-1},\tau_j)$.
Thus, $\varphi(t,x,f)$ is given by the flow along the dynamical system $\phi_i$ during the period of time for which $f = i$. With this, we can explicitly define our deterministic hybrid system. Consider
\begin{center}
$\Phi_t(x_0,f_0) \equiv \left (
\begin{array}{cc}
\varphi_t(x_0,f_0) \\
\psi_t(f_0)
\end{array} \right ) $
\end{center}
with initial conditions $f_0 \in \bar \Delta$, and $x_0\in M$. Let $\psi_t(f_0) = f_t$ and notice that
\begin{center}
$\Phi_0(x_0,f_0) = \left (
\begin{array}{cc}
x_0 \\
f_0
\end{array} \right ) $
\end{center}
Then,
$$\Phi_{t+s}(x_0,f_0)
=
\left (
\begin{array}{cc}
\varphi_{t+s} (x_0, f_0) \\
f_{t+s}
\end{array} \right )
=
\left (
\begin{array}{cc}
\varphi_t (\varphi_s(x_0, f_0), f_s) \\
f_t \circ f_s
\end{array} \right )
=
\Phi_t \circ \Phi_s(x_0,f_0) .$$
Thus, $\Phi_t$ is in fact a flow, so the deterministic hybrid system is a dynamical system.
\subsection{Limit Sets on $M \times \bar{\Delta}$}
Part 6 of Theorem \ref{morse1} yields the following corollary.
\begin{cor}\label{morse2}
Given $y \in M \times \bar{\Delta}$, the $\bar{\Delta}$ component of $\alpha(y)$ is contained in $\bar{\Delta}_C$ for some some communicating class $C$, and the $\bar{\Delta}$ component of $\omega(y)$ is contained in $\bar{\Delta}_{C'}$ for some some communicating class $C'$.
\end{cor}
The implication of this corollary is that in order to find the $\alpha$/$\omega$-limit sets of $M \times \bar{\Delta}$, we only need study the $\alpha$/$\omega$-limit sets of trajectories whose states are contained entirely within one particular communicating class.
\begin{defn}\label{projection}
Given a set $X \subseteq M \times \bar{\Delta}$, we define the projection of $X$ onto $M$ by
$$
\pi_{M}(X) \equiv \{x \in M | \exists \, f \mbox{such that } (x,f) \in X \}
$$
Define the projection of $X$ onto $\bar{\Delta}$ by
$$
\pi_{\bar{\Delta}}(X) \equiv \{f \in \bar{\Delta} | \exists \, x \mbox{such that } (x,f) \in X \}
$$.
\end{defn}
Given a hybrid system whose graph $G$ contains some vertex $\gamma$, with corresponding dynamical system $\phi_\gamma$ on $M$, define $\omega_\gamma(x), x \in M$ to be the $\omega$-limit set of $x$ for the flow $\phi_\gamma$. Define $\alpha_\gamma(x)$ to be the $\alpha$-limit set of $x$ for $\phi_\gamma$.
\begin{lem}\label{selfloop2}
If, given a vertex $\gamma\in V$ has an edge from and to itself, then there exists $ f \in \bar{\Delta}$ such that
$$\omega _\gamma (x) = \pi_{M}(\omega(x,f))$$
$$\alpha _\gamma (x) = \pi_{M}(\alpha(x,f))$$
for all $x \in M$.
\end{lem}
\begin{proof}
Since there is an edge from $\gamma$ to itself, the function in $f\in\bar{\Lambda}$ given by $f\equiv\gamma$ is an element of $\bar{\Delta}$. This corresponds to running $\phi_\gamma$ for all of time. Clearly, the projection of the limit sets of this system onto $M$ are the precisely those of $\phi_\gamma$.
\end{proof}
\begin{thm}\label{selfloopomega}
Consider a graph $G$ consisting of a single communicating class $C$ for which every vertex has an edge starting and ending at itself and a corresponding finite set of dynamical systems $\{\phi_i\}$, each of which has a Morse decomposition $\{ \mathcal{M}_1^i,\mathcal{M}_2^i,..., \mathcal{M}_{n_i}^i \}$. Then, for all $x \in M$, there exists $f \in \bar{\Delta}$ such that
$$
\pi_{M}(\omega(x,f)) \cap (\displaystyle \bigcup_j \mathcal{M}_j^i) \neq \emptyset \, \forall \, i .
$$
\end{thm}
\begin{proof}
Given any $x_{00}$, any $\epsilon > 0$, and any $\phi_1$, there exists a time $t_{10}$ such that $\phi_1(t_{10},x_{00})$ is within $\varepsilon$ of an $\omega$-limit set of $\phi_1$ by definition of $\omega$-limit sets. Since the Morse sets of $\phi_1$ contain the $\omega$-limit sets of $\phi_1$, being within $\epsilon$ of an $\omega$-limit set implies being within $\varepsilon$ of a Morse set. So, let $f(t) = $ $\gamma_1$ for $t \in [0, t_{10})$, and let $\phi_1(t_{10},x_{00}) = x_{10}$.
Now, consider a second dynamical system $\phi_2$, represented in $G$ by the vertex $\gamma_2$. Since the graph consists of a single communicating class, it is possible to transition from vertex $\gamma_1$ to vertex $\gamma_2$ in a finite time. Again, pick $t_{20}$ so that $\phi_2(t_{20},x_{10}) \equiv x_{20}$ is within $\varepsilon$ of an $\omega$-limit set, hence Morse set, of $\phi_2$. Iterate this process for all $n$ vertices in the communicating class, yielding a point $\phi_n(t_{n0},x_{(n-1)0}) \equiv x_{n0}$.
Next, repeat the above procedure for $\frac{\epsilon}{2}$. That is, find $t_{11}$ such that $\phi_1(t_{11},x_{n0}) \equiv x_{11}$ is within $\frac{\epsilon}{2}$ of a Morse set of $\phi_1$, $t_{21}$ such that $\phi_2(t_{21},x_{11})$ is within $\frac{\epsilon}{2}$ of a Morse set of $\phi_2$, etc. Repeat for $\frac{\epsilon}{4}, \frac{\epsilon}{8}, ...$. In the limit, for each vertex $\gamma_i$, we get a sequence of points $\{x_{ij}\}, j = 0, 1, 2, ...$ such that $x_{ij}$ is within $\frac{\epsilon}{2^j}$ of a Morse set of $\phi_i$. Since there are only finitely many Morse sets for each $\phi_i$, there exists a subsequence $\{x_{ij_k}\}$ of $\{x_{ij}\}$ converging to one particular Morse set $\mathcal{M}_l^i$ of $\phi_i$. Thus, $\pi_{M}(\omega(x_{00},f)) \cap \mathcal{M}_l^i \neq \emptyset .$
\end{proof}
Clearly, the equivalent theorem applies to $\alpha$-limit sets, going backwards in time. The example of section $6.1$ demonstrates that the existence of self loops is necessary for this theorem to hold.
\subsection{Morse Sets on $M \times \bar{\Delta}$}
The following definitions and theorem come from \cite{Kliemann}. Suppose $X$ is a compact space with associated flow $\Phi$.
\begin{defn}\label{epsilontchain}
Given $x,y \in X$ and $\epsilon, T > 0$, an $(\epsilon,T)$-chain from $x$ to $y$ is given by a natural number $n \in \mathbb{N}$ together with points
$$
x_0 = x,x_1,...,x_n=y \in X \mbox{ and times } T_0,...T_{n-1} \geq T
$$
such that $d(\Phi(T_i,x_i),x_{i+1})<\epsilon$ for $i = 0,1,...,n-1$.
\end{defn}
\begin{defn}\label{chainrecurrent}
A point $x \in X$ is chain recurrent if for all $\epsilon, T >0$ there exists an $(\epsilon,T)$-chain from $x$ to $x$. A subset $Y \subseteq X$ is chain transitive if for all $x,y \in Y$ and all $\epsilon, T > 0$, there exists an $(\epsilon,T)$-chain from $x$ to $y$.
\end{defn}
\begin{thm}\label{chainrecurrentmorse}
Suppose $X$ has a finest Morse decomposition $\{\mathcal{M}_1,...,\mathcal{M}_n\}$. Then, $x$ is chain recurrent for all $x \in \mathcal{M}_i$, and each $\mathcal{M}_i$ is connected.
\end{thm}
With this machinery, we are in a position to define a new term and state a theorem regarding the hybrid system.
\begin{defn}\label{attracting region}
Given a compact set $A \subset X$, we call $A$ an attracting region for a flow $\Phi$ if there exists an open neighborhood $N(A)$ of $A$ such that $\omega(N(A)) \subseteq A$. We call a compact set $R$ a repelling region for $\Phi$ if there exists an open neighborhood $N(R)$ such that $\alpha(N(R)) \subseteq R$.
\end{defn}
\begin{thm}\label{alwaysattracting}
Suppose that $A \subsetneq M$ is an attracting region for every individual dynamical system $\phi_i, i \in \{1,...,n\}$ of a hybrid system. Then, there exists a non-trivial Morse decomposition on $M \times \Delta$. Furthermore, if there exists a finest Morse decomposition on $M \times \Delta$, then it contains a Morse set $\mathcal{M}_A$ such that $\pi_{M}(\mathcal{M}_A) \subseteq A$.
\end{thm}
\begin{proof}
Let $N(A) \equiv \displaystyle \bigcap_{i=1}^l{N_i(A)} \subset M$ be the intersection of the neighborhoods $N_i(A)$ around $A$ admitted by each of the flows $\phi_i$, and note that $N(A)$ is also an open neighborhood of $A$. Clearly, if $A \subsetneq M$ is an attracting region for each flow $\phi_i$, then $A \times\Delta$ is an attracting region for the hybrid system flow $\Phi$, as we can take $N(A) \times \Delta$ to be the open neighborhood around $A \times \Delta$.
Given $x \in N(A)$, let $B$ be some closed neighborhood containing $A$ but contained in $N(A)$ so that $x \notin B$. Let $C$ be some closed neighborhood properly containing $B$ but contained in $N(A)$ so that $x \in C$. Let $\partial B$ and $\partial C$ denotes the boundaries of $B$ and $C$, respectively. Note that $\partial B$ and $\partial C$ contain their boundaries and are bounded since $M$ is bounded, so that $\partial B$ and $\partial C$ are both compact.
We want to show that there exists a time $T$ such that $\varphi_T(y,f) \in B$ for all $y \in \partial C, f \in \Delta$. Suppose not. Then, for each $n \in \mathbb{N}$, there exists $t_n > n, x_n \in \partial C, f_n \in \Delta$ such that $\varphi_{t_n}(x_n,f_n) \in \partial B$. This gives rise to sequences $\{t_n\}$, $\{x_n\}$, $\{ f_n\}$, and $\{\varphi_{t_n}(x_n,f_n)\}$ with $\{t_n\} \rightarrow \infty$. By compactness of $\partial C$, $\Delta$, and $\partial B$, the sequences $\{x_n\}$, $\{ f_n\}$, and $\{\varphi_{t_n}(x_n,f_n)\}$ have convergent subsequences, so that there exists $\{ n_j \}$ with $\{t_{n_j}\} \rightarrow \infty$, $\{x_{n_j}\} \rightarrow \hat{x}$, $\{x_{n_j}\} \rightarrow \hat{x} \in \partial C, \{f_{n_j}\} \rightarrow f \in\Delta,$ and $\{\varphi_{t_{n_j}}(x_{n_j},f_{n_j})\} \rightarrow \bar{x} \in \partial B$. But then, $\bar{x} \in \pi_{M}(\omega(N(A) \times \Delta))$. Since $\bar{x} \in \partial B \subset N(A) \setminus A$, this contradicts the definition of an attracting region.
So, such a $T$ exists. Take $\epsilon < $ min($d(x,B),d(C,B)$). Since any trajectory starting in $C \setminus B$ will flow into $B$ by time $T$, and since the distance from any point in $B$ to $x$ is greater than $\epsilon$, no $(\epsilon,T)$-chain from $x$ to itself exists. Therefore, $(x,f)$ is not chain recurrent for any $f \in \bar{\Delta}$. But, if the only Morse decomposition were trivial, then all of $M \times \bar{\Delta}$ would be the only Morse set, hence chain recurrent by Theorem \ref{chainrecurrentmorse}. This is a contradiction. Thus, there exists a non-trivial Morse decomposition.
If it is known that a finest Morse decomposition exists on $M \times \bar{\Delta}$, then the fact that any solution beginning in $N(A)$ flows into $A$ and remains there implies that the intersection of $A$ with the projection of some limit set onto $M$ is nonempty. Since $(N(A) \setminus A) \times \bar{\Delta}$ is not chain recurrent, as shown, $(N(A) \setminus A) \times \bar{\Delta}$ cannot be contained in a Morse set. So, there exists a Morse set $\mathcal{M}_A$ such that $\pi_{M}(\mathcal{M}_A) \cap A \neq \emptyset$, $\pi_{M}(\mathcal{M}_A) \cap (N(A) \setminus A) = \emptyset$. Since the Morse sets of a finest Morse decomposition are connected, this means that $\pi_{M}(\mathcal{M}_A) \subseteq A$.
\end{proof}
By the same argument, going backwards in time, we get the same result for repelling regions.
\begin{cor}\label{alwaysrepelling}
Suppose that $R \subsetneq M$ is a repelling region for every individual dynamical system $\phi_i, i \in \{1,...,l\}$ of a hybrid system. Then, there exists a non-trivial Morse decomposition on $M \times \bar{\Delta}$. Furthermore, if there exists a finest Morse decomposition on $M \times \bar{\Delta}$, then it contains a Morse set $\mathcal{M}_R$ such that $\pi_{M}(\mathcal{M}_R) \subseteq R$.
\end{cor}
Theorem \ref{alwaysattracting} and Corollary \ref{alwaysrepelling} have important ramifications for hybrid systems that consist of several, slightly-perturbed individual dynamical systems. If each of these systems has an attracting or repelling fixed point in some neighborhood, then it is likely that some attracting or repelling region will exist that contains all of them. We have shown that the existence of such a region implies a non-trivial Morse decomposition of the space.
Suppose now that the graph $G$ consists of a single communcating class, $C$, so that $\bar{\Delta} = \bar{\Delta}_C$.
\begin{defn}\label{attractorrepeller}
A compact invariant set $A$ is an attractor if it admits an open neighborhood $N$ such that $\omega(N) = A$. A repeller is a compact invariant set $R$ that has an open neighborhood $N^*$ such that $\alpha(N^*) = R$.
\end{defn}
Note that any attractor or repeller must be contained in a Morse set, since Morse sets contain $\alpha$- and $\omega$-limit sets. Furthermore, it can be shown that for any Morse decomposition, at least one Morse set is an attractor and at least one Morse set is a repeller \cite{Kliemann}.
\begin{thm}\label{allofdeltabar}
Suppose that an attractor $A \subseteq M \times \bar{\Delta}$ is a Morse set for some Morse decomposition. Then, $\pi_{\bar{\Delta}}(A) = \bar{\Delta}$.
\end{thm}
\begin{proof}
Take $(\bar x, \bar f) \in A$. Since $A$ is an attractor, there exists $\epsilon > 0$ and a ball $B$ of radius $\epsilon$ around $(\bar x, \bar f)$ such that the $\omega$ limit set of any point in the ball is contained in $A$. By Lemma \ref{omegalimit}, there exists $f^* \in \bar{\Delta}$ so that $\omega(f^*) = \bar{\Delta}$. Therefore, there exists $t$ such that $d(\psi(t,f^*),\bar f) < \epsilon$. Thus, $(\bar x, \psi(t,f^*)) \in B$, so $\omega(\bar x, \psi(t,f^*)) \subseteq A$. Since $\omega(f^*) = \bar{\Delta}$ implies $\omega(\psi(t,f^*)) = \bar{\Delta}$, we see that $\pi_{\bar{\Delta}}(A) = \bar{\Delta}$.
\end{proof}
The same argument applies for a repeller, yielding the following corollary.
\begin{cor}\label{repellertime}
Suppose that a repeller $R \subseteq M \times \bar{\Delta}$ is a Morse set for some Morse decomposition. Then, $\pi_{\bar{\Delta}}(R) = \bar{\Delta}$.
\end{cor}
\section{Behavior Under Small Perturbations}
We now analyze the behavior of a dynamical system under ``small" perturbations, i.e. we compare the global behavior of%
\begin{equation}
\phi:\mathbb{R}\times M\rightarrow M \label{systunpert}%
\end{equation}
with that of
\begin{align}
\Phi & :\mathbb{R}\times M\times\Delta\rightarrow M\times
\Delta\label{systpert}\\
\Phi(t,x,f) & \mapsto(\varphi(t,x,f),\psi(t,f))\nonumber
\end{align}
by looking at their respective Morse decompositions. Since the Morse
decompositions of (\ref{systunpert}) live in $M$, and those of (\ref{systpert}
) live in $M\times\Delta$, we need to make the decompositions
comparable. This is accomplished by projecting chain recurrent sets from
$M\times\Delta$ down to $M$ in the following way.
\begin{defn}
A set $E\subset M$ is called a chain set of $\Phi$, if
\begin{enumerate}
\item for all $x\in E$ there exists $f\in\Delta$ with
$\varphi(t,x,f)\in E$ for all $t\in\mathbb{R}$,
\item for all $x,y\in E$ and all $\epsilon,T>0$ there exists an
$(\epsilon,T)-$chain from $x$ to $y$,
\item $E$ is maximal with these two properties.
\end{enumerate}
Here an $(\epsilon,T)-$ chain from $x$ to $y$ is defined by $n\in N$,
$x_{0},...,x_{n}\in M$, $f_{0},...,f_{n-1}\in\bar{\Delta}$, and $t_{0}%
,...,t_{n-1}\geq T$ such that $x_{0}=x$, $x_{n}=y$, and $d(\varphi(t_{i}%
,x_{i},f_{i}),x_{i+1})\leq\epsilon$ for all $i=0,...,n-1$.
\end{defn}
\begin{defn}
The lift $\mathcal{L}(A)$ of a set $A \subseteq M$ onto $M \times \Delta$ is defined by
$$
\mathcal{L}(A) = \{(x,f)\in M \times \Delta | \pi_M(\Phi(t,x,f))\in A \, \forall \, t \in \mathbb{R})\} .
$$
\end{defn}
It turns out that chain sets in $M$ correspond exactly to the chain recurrent
components of $(\Phi,M\times\Delta)$.
\begin{thm}
Let $E\subset M$ be a chain set of $\Phi$.
\begin{enumerate}
\item The lift of $E$ to $\mathcal{E}\subset M\times\Delta$ is a
maximal, invariant chain transitive set of $(\Phi,M\times\Delta)$.
\item If $\mathcal{E}\subset M\times\Delta$ is a maximal, invariant
chain transitive set of $(\Phi,M\times\Delta)$ then its projection
$\pi_{M}(\mathcal{E})$ is a chain set.
\end{enumerate}
\end{thm}
\begin{proof}
(ii) Let $\mathcal{E}$ be an invariant, chain transitive set in $M\times
\bar{\Delta}$. For $x\in\pi_{M}(\mathcal{E})$ there exists $f\in
\bar{\Delta}$ such that $\varphi(t,x,u)\in\mathcal{E}$ for all
$t\in\mathbb{R}$ by invariance. Now let $x,\,y\in\pi_{M}(\mathcal{E})$ and
choose $\epsilon>0,\,T>0$. Then by chain transitivity of $\mathcal{E}$, we
can choose $x_{j},u_{j},t_{j}$ such that the corresponding trajectories
satisfy the required condition. Note also that $E$ is maximal if and only if $\mathcal{E}$ is maximal.
(i) Let $(x,f),\;(y,g)\in\mathcal{E}$ and pick $\epsilon>0.$ Choose
$T_{0}>0$ large enough such that
\[
\sum_{j\notin\lbrack-T_{0},T_{0}]}^{{}}4^{-\left\vert j\right\vert
}<\epsilon
\]
and pick $T>T_{0}$. Since $\varphi(2T,x,f)\in E$ and $\varphi(-T,y,g)\in E$,
the fact that $E$ is a chain set yields the existence of $k\in\mathbb{N}$ and
$x_{0},...,x_{k}\in M,\,f_{0},...,f_{k-1}\in\bar{\Delta}%
,\,t_{0},...,t_{k-1}\geq T$ with $x_{0}=\varphi(2T,x,f),\,x_{k}%
=\varphi(-T,y,g),$ and
\[
d(\varphi(t_{j},x_{j},f_{j}),x_{j+1})<\epsilon\text{\ for }j=0,...,k-1.
\]
We now construct an $(\epsilon,T)$-chain from $(x,f)$ to $(y,g)$ in the
following way. Define
\[%
\begin{array}
[c]{cccc}%
t_{-2}=T, & x_{-2}=x, & g_{-2}=f, & \\
t_{-1}=T, & x_{-1}=\varphi(T,x,f), & g_{-1}(t)= & \left\{
\begin{array}
[c]{cc}%
f(t_{-2}+t) & \text{for }t\leq t_{-1}\\
f_{0}(t-t_{-1}) & \text{for }t>t_{-1}%
\end{array}
\right.
\end{array}
\]
and let the times $t_{0},...,t_{k-1}$ and the points $x_{0},...,x_{k}$ be as
given earlier. Furthermore, set
\[%
\begin{array}
[c]{ccc}%
t_{k}=T, & x_{k+1}=y, & g_{k+1}=g,
\end{array}
\]
and define for $j=0,...,k-2$%
\begin{align*}
g_{j}(t) & =\left\{
\begin{array}
[c]{ccc}%
g_{j-1}(t_{j-1}+t) & \text{for} & t\leq0\\
f_{j}(t) & \text{for} & 0<t<t_{j}\\
f_{j+1}(t-t_{j}) & \text{for} & t>t_{j},
\end{array}
\right. \\
g_{k-1}(t) & =\left\{
\begin{array}
[c]{ccc}%
g_{k-2}(t_{k-2}+t) & \text{for} & t\leq0\\
f_{k-1}(t) & \text{for} & 0<t\leq t_{k-1}\\
g(t-t_{k-1}-T) & \text{for} & t>t_{k-1},
\end{array}
\right. \\
g_{k}(t) & =\left\{
\begin{array}
[c]{cc}%
g_{k-1}(t_{k-1}+t) & \text{for }t\leq0\\
g(t-T) & \text{for }t>0.
\end{array}
\right.
\end{align*}
We see that
\[
(x_{-2},g_{-2}),\,(x_{-1},g_{-1}),...,(x_{k+1},g_{k+1})\text{ and }%
t_{-2},\,t_{-1},...,t_{k}\geq T
\]
yield an $(\epsilon,T)$-chain from $(x,f)$ to $(y,g)$ provided that for
$j=-2,\,-1,...,k$%
\[
d(g_{j}(t_{j}+\cdot),g_{j+1})<\epsilon.
\]
By choice of $T$ we have by the definition of the metric on $\bar{\Delta
}$ for all $h_{1},\,h_{2}\in\bar{\Delta}$
\begin{align*}
d(h_{1},h_{2}) & =\sum_{j\in Z}4^{-\left\vert j\right\vert }f(h_{1}%
,h_{2},j)\\
& =\sum_{j\in Z}4^{-\left\vert j\right\vert }\frac{1}{h}\int\limits_{jh}%
^{(j+1)h}\delta(h_{1},h_{2},t)dt\\
& \leq\sum_{j\in\lbrack-T_{0},T_{0}]}4^{-\left\vert j\right\vert }\frac{1}%
{h}\int\limits_{jh}^{(j+1)h}\delta(h_{1},h_{2},t)dt\text{ \ }+\text{
}\epsilon\text{.}%
\end{align*}
Note that for the $g_{j}$, $j=-2,\,-1,...,k$ as defined above the integrands
vanish, giving the desired result.
\end{proof}
With this result, we can look at families of dynamical systems over graphs in
the following way:
Let $\alpha\in I$, where $I$ is some index set, and consider the family%
\begin{equation}
\Phi_{\alpha}:\mathbb{R}\times M\times\bar{\Delta}\rightarrow
M\times\bar{\Delta} \label{systfamdelta}%
\end{equation}
such that $\Phi_{\cdot}$ depends continuously on $\alpha$. We embed the
systems of (\ref{systfamdelta}) into a larger base space by defining:
Let $U\subset\mathbb{R}^{m}$ be compact, convex with $0\in$ int$(U)$, and set
$\mathcal{U}=\{u:\mathbb{R}\rightarrow U$, locally integrable$\}$. The space
$\mathcal{U}$ with the weak$^{\ast}-$topology is a compact space. If $G$ is a finite directed $N$-graph, then $\bar{\Delta}(G)\subset\mathcal{U}$. For the
family of systems%
\[
\Phi_{\alpha}:\mathbb{R}\times M\times\mathcal{U}\rightarrow M\times
\mathcal{U}%
\]
we have the following result.
\begin{lem}
\label{lemDyCon}Let $\alpha_{k}\rightarrow\alpha_{0}$ in $I$, and let
$E^{\alpha_{k}}\subset M$ be chain sets of $\Phi_{\alpha_{k}}$. If
\[
\underset{\alpha_{k}\rightarrow\alpha_{0}}{\lim\sup}E^{\alpha_{k}}\equiv\{x\in
M\text{, there exist }x^{\alpha_{k}}\in E^{\alpha_{k}}\text{ with }%
x^{\alpha_{k}}\rightarrow x\}\neq\emptyset
\]
then it is contained in a chain set $E^{\alpha_{0}}$ of $\Phi_{\alpha_{0}}$.
\end{lem}
For a proof of this result see Theorem 3.4.6 in \cite{DyCon}.
In many applications the parameter $\alpha\in I$ refers to the size of the
perturbation in the following way: Let $U\subset\mathbb{R}^{m}$ as above and
set $U^{\rho}\equiv\rho\cdot U$ for $\rho\geq0$. Assume that the compact state
space $M$ is a $C^{\infty}-$manifold and consider the family of differential
equations
\begin{equation}
\dot{x}=X(x,\rho f)\text{, }\rho\geq0\text{.} \label{systfamrho}%
\end{equation}
where $X$ is a $C^{\infty}-$vector field on $M$ and the parameter $\rho$ is the perturbation size.
Each differential equation (\ref{systfamrho}) gives rise to a parametrized
dynamical system on $M\times\mathcal{U}^{\rho}$ and on $M\times\bar
{\Delta}^{\rho}$. For $\rho=0$ this is the unperturbed system given
by $\dot{x}=X(x,0)$. In this set up, Lemma \ref{lemDyCon} reads:
\begin{cor}
\label{corDyCon}Let $\rho_{k}\rightarrow0$ and let $E^{\rho_{k}}\subset M$ be
chain sets of $\Phi_{\rho}:\mathbb{R}\times M\times\mathcal{U}^{\rho
}\rightarrow M\times\mathcal{U}^{\rho}$. If $\underset{\rho_{k}\rightarrow
0}{\lim\sup}E^{\rho_{k}}\neq\emptyset$ then it is a chain recurrent set of
$\Phi_{0}$.
\end{cor}
For a proof of this result see Corollary 3.4.8 in \cite{DyCon}. This corollary
then gives the following result for ``small" perturbations of dynamical systems
in the context of the systems on $M\times\mathcal{U}^{\rho}$:
\begin{cor}
\label{corDyCon2}Consider the family $\Phi_{\rho}:\mathbb{R}\times
M\times\mathcal{U}^{\rho}\rightarrow M\times\mathcal{U}^{\rho}$ of dynamical
systems given by the differential equations (\ref{systfamrho}). Assume that
$\Phi_{\rho}$ has a finest Morse decomposition in $M\times\mathcal{U}^{\rho}$
for $\rho\geq0$. Then there exists $\epsilon>0$ such that for all
$0\leq\rho\leq\epsilon$ the finest Morse decomposition of $\Phi_{\rho}$
corresponds to the one of $\Phi_{0}$, i.e. the Morse sets are in 1-1
correspondence and the order induced by the Morse decomposition of $\Phi_{0}$
is preserved.
\end{cor}
To obtain a similar result for dynamical systems over graphs, we consider the
systems $\Phi_{\rho}:\mathbb{R}\times M\times\bar{\Delta}^{\rho
}\rightarrow M\times\bar{\Delta}^{\rho}$ where the finite set $S^{\rho
}\subset U^{\rho}$. Note that chain sets over $\bar{\Delta}^{\rho}$ are
contained in those over $\mathcal{U}^{\rho}$. To make the systems $\Phi_{\rho
}$ comparable we require that $0\in S^{\rho}$ for all $\rho\geq0$. The
examples in the next section show limit sets and Morse sets of $\Phi_{\rho
}:\mathbb{R}\times M\times\bar{\Delta}^{\rho}\rightarrow M\times
\bar{\Delta}^{\rho}$ can behave quite differently from those of $\Phi
_{0}$. However, Corollary \ref{corDyCon2} implies the following result on
small perturbations via systems over graphs:
\begin{cor}
\label{corsmalldelta}Consider the family $\Phi_{\rho}:\mathbb{R}\times
M\times\bar{\Delta}^{\rho}\rightarrow M\times\bar{\Delta}^{\rho}$ of
dynamical systems given by the differential equations (\ref{systfamrho}).
Assume that $\Phi_{\rho}$ has a finest Morse decomposition in $M\times
\bar{\Delta}^{\rho}$ for $\rho\geq0$. Assume furthermore that $0\in
S^{\rho}$ for all $\rho\geq0$ and that the vertex of the underlying graph
corresponding to $0\in S^{\rho}$ has a self loop. Then there exists
$\epsilon>0$ such that for all $0\leq\rho\leq\epsilon$ the finest Morse
decomposition of $\Phi_{\rho}$ corresponds to the one of $\Phi_{0}$, i.e. the
Morse sets are in 1-1 correspondence and the order induced by the Morse
decomposition of $\Phi_{0}$ is preserved.
\end{cor}
The proof follows directly from the fact that, under the conditions of the
corollary, the Morse sets of the finest Morse decomposition of $\Phi_{0}$ are
contained in the chain sets over $\bar{\Delta}^{\rho}$ for each $\rho
\geq0$. Hence we obtain the results of Corollary \ref{corDyCon} for the
systems $\Phi_{\rho}:\mathbb{R}\times M\times\bar{\Delta}^{\rho
}\rightarrow M\times\bar{\Delta}^{\rho}$, which implies Corollary
\ref{corsmalldelta}.
\section{Examples}
The examples in this section show that even simple hybrid systems can admit complicated behavior. Lemma \ref{selfloop2} and Theorem \ref{selfloopomega} suggest that it may be possible to define some sort of relationship between the $\omega$-limit sets of the individual flows $\{\phi_i\}$ and those of the hybrid system $\varphi$. One might suspect that the limit sets of the dynamical system should be contained in the projection of the limit sets of the hybrid system onto $M$, or vice versa. However, the example of section $6.1$ shows that these results are not true for a hybrid system on a general graph.
Conversely, the example in section $6.1$ indicates that it may be possible to find solutions of the hybrid system that avoid the limit sets of each of the individual dynamical systems. The example in section $6.2$, however, shows that this need not be the case, even in the event that the graph is complete.
Finally, Theorem \ref{allofdeltabar} and Corollary \ref{repellertime} suggest that the projection of a Morse set of a hybrid system onto $\bar{\Delta}$ might be itself a Morse set of the dynamical system $\psi$ on $\bar{\Delta}$. Section $6.3$ gives a counterexample to this conjecture.
\subsection{A Hybrid System that Introduces New Limit Sets}
Consider the hybrid system given by two flows, $A$, and $B$ on the compact space $M = [-1,1]$, with $\phi_A(t,x)$ given by the solution to
$$
\dot{x} = (1-x)(1+x)(-\frac{1}{2}-x)
$$
and $\phi_B(t,x)$ given by the solution to
$$
\dot{x} = (1-x)(1+x)(\frac{1}{2}-x)
$$
and graph $G$ given by Figure \ref{figure1}.
\begin{figure}[H]
\begin{center}
\includegraphics[width=15mm]{flickergraph.png}
\end{center}
\caption{Graph for the example hybrid system.}
\label{figure1}
\end{figure}
The fixed points for flow $A$ are $x = \pm 1, -\frac{1}{2}$, while the fixed points for flow $B$ are $x = \pm 1, \frac{1}{2}$. Therefore, the $\omega$-limit sets for flows $A$ and $B$ are, respectively,
$$
\omega_A(M) = \{-1,-\frac{1}{2},1\}
$$
$$
\omega_B(M) = \{-1,\frac{1}{2},1\}
$$
As usual, let $h$ be the length of the intervals on which a function in $\bar{\Delta}$ is piecewise constant.
\begin{prop}\label{counterexample}
There exists $h$ so that given any $x \in M$ and any $f \in \bar{\Delta}$, the projection $\pi_{M}(\omega(x,f))$ contains neither $-\frac{1}{2}$ nor $\frac{1}{2}$.
\end{prop}
\begin{proof}
From the graph in Figure \ref{figure1}, we see that any function $f \in \bar{\Delta}$ will flicker back and forth between system $A$ and $B$. During the intervals of time for which $f(t) = A$, solutions in $(-1,1)$ will be drawn toward the point $x = -\frac{1}{2}$. Likewise, during the intervals for which $f(t) = B$, solutions in $(-1,1)$ will be drawn toward the point $x = \frac{1}{2}$. It is clear, therefore, that given any $x \in (-1,1)$ and any $f \in \bar{\Delta}$, there exists a time at which the $M$ component of the solution will enter the region $(-\frac{1}{2},\frac{1}{2})$.
Suppose, then, that $x(0) \in (-\frac{1}{2},\frac{1}{2})$. Whenever $f(t) = A$, $x$ will decrease, bounded below by $-\frac{1}{2}$. Whenever $f(t) = B$, $x$ will increase, bounded above by $\frac{1}{2}$. Thus, no solution that enters the interval $(-\frac{1}{2},\frac{1}{2})$ will ever leave it.
The only way that an $\omega$-limit set could contain either $-\frac{1}{2}$ or $\frac{1}{2}$, therefore, is if a solution approached one of the points asymptotically. To eliminate this possibility, we need to show that there exists an $\epsilon > 0$ such that the distance between the solution $x$ and the points $-\frac{1}{2}$ and $\frac{1}{2}$ remains bounded below by $\epsilon$ in the limit as $t \rightarrow \infty$. To do this, let $h$ be such that
$$\phi_A(h,\frac{1}{2}) < 0 .$$
By symmetry, this also implies that $\phi_B(h,-\frac{1}{2}) > 0$. Now, pick a point $x \in (-\frac{1}{2},\frac{1}{2})$ and a $f \in \bar{\Delta}$, and allow the hybrid system run for a time $\tau$ so that
$$f(\tau + t) = \left\{
\begin{array}{cc}
A & \mbox{if} \, \left \lfloor \frac{t}{h} \right \rfloor \equiv 0 \mbox{ (mod 2)}\\
B & \mbox{otherwise} \\
\end{array} \right\}$$
This corresponds to shifting $f$ by a time $\tau$ such that $f(\tau) = A$ and $f$ is constant for the entire interval of length $h$ after $\tau$. We know that at this time $\tau$, $x(\tau) \in (-\frac{1}{2},\frac{1}{2})$. Thus, by our choice of $h$, $x(\tau + h) \in (-\frac{1}{2},0)$. Let $\epsilon = \frac{1}{2} - \phi_B(h,0)$. Since $x(\tau + h) \in (-\frac{1}{2},0)$, the distance between $x(\tau + 2h)$ and $\frac{1}{2}$ will be strictly greater than $\epsilon$. But then, $x(\tau + 3h) \in (-\frac{1}{2},0)$, so $\frac{1}{2} - x(\tau + 4h) > \epsilon$. More generally, $\frac{1}{2} - x(\tau + 2nh) > \epsilon$ for $n \in \mathbb{N}$. And, since $x$ reaches a local maximum at each $\tau + 2nh$ before decreasing toward $-\frac{1}{2}$, the distance between $x$ and $\frac{1}{2}$ never approaches $0$ for $t > \tau$. By symmetry, the distance between $x$ and $-\frac{1}{2}$ never approaches 0. So, the projection of $\omega((x,f))$ onto $M$ contains neither $-\frac{1}{2}$ nor $\frac{1}{2}$ (c.f. Figure \ref{figure2}).
\end{proof}
\begin{figure}[H]
\begin{center}
\includegraphics[width=100mm]{paper21d.png}
\end{center}
\caption{Three representative trajectories for the example hybrid system.}
\label{figure2}
\end{figure}
This example has important implications. Real-world systems are often modeled with dynamical systems by averaging over all possible states, neglecting the $\bar{\Delta}$ part of the equation. This example shows that the limiting behavior of a hybrid system need not be anything like the limiting behavior of the individual dynamical systems comprising it. The practice of ignoring $\bar{\Delta}$ may not always be valid.
\subsection{A Hybrid System with Limit Sets Containing Morse Sets}
Let $M = S^1$, the circle. Suppose $G$ is a complete graph consisting of two vertices, with corresponding dynamical systems $\phi_1$ and $\phi_2$ shown in Figure \ref{figure3}.
\begin{figure}[H]
\begin{center}
\includegraphics[width=60mm]{paper2ex2.png}
\end{center}
\caption{The dynamical systems in the second example hybrid system.}
\label{figure3}
\end{figure}
In this picture, the letter $A$ denotes an attractor, while the letter $R$ denotes a repeller. The red, filled circles correspond to the fixed points (and Morse sets $\mathcal{M}_j^1$) of $\phi_1$, while the blue, unfilled circles correspond to the fixed points (and Morse sets $\mathcal{M}_j^2$) of $\phi_2$. The Morse sets divide the circle into four regions, which are marked I, II, III, and IV.
\begin{prop}\label{counterexample2}
There exists $h$ such that $\pi_{S^1}(\omega(x,f)) \cap (\displaystyle \bigcup_{i,j=1}^2 \mathcal{M}_i^j) \neq \emptyset$ for all $x \in S^1, f \in \Delta$.
\end{prop}
\begin{proof}
Consider $x\in S^1$ and $f\in\Delta$. Notice that the flow in region II will always be counterclockwise, for both systems $\phi_1$ and $\phi_2$, and similarly the flow in region IV is always clockwise, for both systems $\phi_1$ and $\phi_2$. Thus, trajectories in region II will precess toward region I and trajectories in region IV will precess toward region $III$, independent of the reigning function $f\in\Delta$. Thus, we need only consider if the trajectory $x$ enters regions I or III. \\
\indent Assume there exists a time $\tau_2$ such that the projection of the trajectory of $(x,f)$ onto $S^1$ enters region III, and $f(t)=2$ for all $t>\tau_2$. Thus, since in system $\phi_2$ solutions in region I converge to to $A_2$, $\pi_{S^1}(\omega(x,f))=A_2$.\\
\indent Similarly, if there exists a time $\tau_1$ such that the projection of the trajectory of $(x,f)$ onto $S^1$ enters region I and $f(t)=1$ for all $t>\tau_1$ then $\pi_{S^1}(\omega(x,f))=A_1$.\\
Now assume that the projection of the trajectory of $(x,f)$ enters region I at a time $t_2$, but that $f(t)\not = 2$ for all $t>t_1$. Then there exists $s>t_2$ such that $f(s)$=1; that is, the system will eventually switch to the flow given by the system $\phi_1$, for a minimum of one time interval. Then take $h_2$ large enough such that $\phi_1(A_2, h_2)$ is in region IV (which exists, since $A_2$ is not a fixed point of $\phi_1$ and there is no fixed point of $\phi_1$ between $A_2$ and region IV). Then, even if $f=1$ for only one time interval, the trajectory of $(x,f)$ will be swept out of region I and into region IV. Therefore, if any solution stays in region I indefinitely, it must converge to $A_2$. \\
\indent Similarly, assume that the projection of the trajectory of $(x,f)$ enters region III at a time $t_1$ but that $f(t)\not =1$ for all $t>t_1$. Thus there exists $s>t_1$ such that $f(s)=2$; that is, the system will eventually switch to the flow given by the system $\phi_2$, for a minimum of one time interval. Take $h_1$ large enough such that $\phi_1(A_1,h_2)$ is in region II (which exists, since $A_1$ is not a fixed point of $\phi_2$ and there is no fixed point of $\phi_2$ between $A_1$ and region II). Therefore, if any solution stays in region I indefinitely, it must converge to $A_1$. \\
\indent Take $h=\max (h_1,h_2)$. In this case, no trajectory can remain in regions II or IV indefinitely, and if a trajectory remains in region I or III indefinitely, it must converge to either $A_2$ or $A_1$ respectively. Thus, the only options for $\omega$-limit sets of $(x,f)$ are $\{A_1\}, \{A_2\}$ or all of $S^1$. In every case, $$\pi_{S^1}(\omega(x,f)) \cap (\displaystyle \bigcup_{i,j=1}^2 \mathcal{M}_i^j) \neq \emptyset.$$
\end{proof}
\subsection{Complete Characterization of Morse Sets of a Hybrid System}
Consider the 1-dimensional system on $[-1,1]$ given by
$$\dot{x} = (-x+1)(x+1)(x-(-1)^n1/2)^2$$
where $n$ takes values from the set $\{1,2\}$. Let $G$, the directed graph dictating the changes in the values of $n$, be a complete graph on 2 vertices. Then the dynamical system corresponding to $n=1$ has a repelling fixed point at $x=-1$, a saddle point at $x=-\frac{1}{2}$, and an attracting fixed point at $x=1$. Meanwhile, the dynamical system corresponding to $n=2$ still has a repelling fixed point at $x=-1$ and an attracting fixed point at $x=1$, but its saddle is located at $x=\frac{1}{2}$.
\begin{prop}\label{thirdexample}
\begin{eqnarray*}
\mathcal{M}_1&=&\{-1\}\times\bar{\Delta}\\
\mathcal{M}_2&=&\{-1/2\}\times\{f \equiv 1\}\\
\mathcal{M}_3&=&\{1/2\}\times\{f \equiv 2\}\\
\mathcal{M}_4&=&\{1\}\times\bar{\Delta}
\end{eqnarray*}
is a Morse Decomposition for this hybrid system.
\end{prop}
\begin{proof} We check the seven conditions in turn.
\begin{enumerate}
\item\textbf{Non-void} Clearly, each Morse set described above is non-empty.
\item\textbf{Pairwise disjoint} Since the projections $\pi_M(\mathcal{M}_i)$ of the Morse sets onto $[-1,1]$ are pairwise disjoint, the Morse sets are pairwise disjoint.
\item\textbf{Invariant} Consider a point $(x,f)$ in $\mathcal{M}_1=\{-1\}\times\bar{\Delta}$. Since the point $x = -1$ is invariant for both of the separate deterministic dynamical systems on $[-1,1]$, no matter what values $f$ takes, any trajectory will never leave $x = -1$. Thus, since any translation of $f$ will always be in $\bar{\Delta}$, the trajectory of $(x,f)$ will always remain in $\mathcal{M}_1$ both forwards and backwards in time, and thus $\mathcal{M}_1$ is invariant. By a similar argument, $\mathcal{M}_4$ is invariant. \\
\indent Now consider the point $(-1/2, f)$, where $f \equiv 1$. Since $f$ takes the value of $n=1$ for all time, this trajectory is equivalent to the trajectory that starts at $x=-1/2$ in the system given by $$\dot{x} = (-x+1)(x+1)(x+1/2)^2.$$
Since $x = -1/2$ is invariant in this system, the trajectory in the pair system will never leave $x = -1/2$. Also, any shift of the function $f \equiv 1$ is still $f \equiv 1$, so $f$ is invariant in $\bar{\Delta}$. Thus, $(-1/2, f)$ is invariant, so $\mathcal{M}_2$ is invariant. By a similar argument for the system $$\dot{x} = (-x+1)(x+1)(x-1/2)^2,$$ $\mathcal{M}_3$ is invariant.
\item\textbf{Isolated} A neighborhood of $\mathcal{M}_i$ given by $$N(\mathcal{M}_i)=\displaystyle\bigcup_{(x,f)\in\mathcal{M}_i}B((x,f),1/4)$$ is an open region containing $\mathcal{M}_i$ that does not intersect any other Morse set. Consider a point $(x,f)\in N(\mathcal{M}_i)$ such that $\alpha(x,f)\subseteq N(\mathcal{M}_i)$ and $\omega(x,f)\subseteq N(\mathcal{M}_i)$. Since by part 6 below every $\alpha,\omega$-limit set is contained in some Morse set, and since $\mathcal{M}_i$ is the only Morse set intersecting $N(\mathcal{M}_i)$, we have that $\alpha(x,f)\subseteq \mathcal{M}_i$ and $\omega(x,f)\subseteq \mathcal{M}_i$. By part 7 below there are no cycles, so $(x,f)\in\mathcal{M}_i$.
\item\textbf{Compact} The sets $\{-1\},\{-1/2\},\{1/2\},\{1\}$ are compact in $M=[-1,1]$, and the sets $\bar{\Delta},\{f \equiv 1\},\{f \equiv 2\}$ are compact in $\bar{\Delta}$. By Tychonoff's theorem, since each of the Morse sets $\mathcal{M}_i$ is a product of two of these sets, each is compact in the product space $M\times\bar{\Delta}$.
\item\textbf{Contains $\alpha$/$\omega$-limit sets} By invariance of each Morse set, $(y,f)\in\mathcal{M}_i$ for some $i$ implies $\alpha(y,f)\subseteq\mathcal{M}_i$ and $\omega(y,f)\subseteq\mathcal{M}_i$. Take a point $(y,f)$ with $y\in(-1,-1/2)$ so that there exists some $z\in\mathbb{R}$ such that $f(t)=1$ for all $t\geq z$. Backwards in time, no matter what value $f$ takes, any trajectory moving in $(-1,-1/2)$ will converge towards $x=-1$, so $\alpha(y,z)\subseteq\mathcal{M}_1$. Forwards in time, the function $f$ will converge to the function $g \equiv 1$. So, suppose $f$ becomes constant before the projection of the trajectory into $[-1,1]$ has passed the point $x = -1/2$. Since all trajectories starting in $(-1,1/2)$ will converge to the fixed point at $x = -1/2$ when $n=1$, the projection of the trajectory into $[-1,1]$ will converge to $x=-1/2$ in forwards time. Thus, the trajectory in $M\times\bar{\Delta}$ will converge forwards to $\mathcal{M}_2$. Instead, suppose $f$ becomes constant after the projection of the trajectory into $[-1,1]$ has passed the point $-1/2$. Since all trajectories starting in $(-1/2,1)$ will converge to $x=-1/2$ when $n=1$, the projection of the trajectory into $[-1,1]$ will converge to $x=1$ in forwards time. Thus, the trajectory in $M\times\bar{\Delta}$ will converge forwards to $\mathcal{M}_4$. \\
Take a point $(y,f)$ where $y\in(1/2,1)$ and there exists some $z\in\mathbb{R}$ such that $f(x)=2$ for all $x\leq z$. Since, forwards in time, no matter what value $f$ takes, any trajectory moving in $(1/2,1)$ will converge towards $1$, $\omega(y,z)\subseteq\mathcal{M}_4$. Backwards in time, clearly the function $f$ will converge to the function $g=2$. Meanwhile, if $f$ becomes constant after the projection of the trajectory into $[-1,1]$ has passed the point $1/2$, since all trajectories starting in $(1/2,1)$ will converge backwards in time to the fixed point at $1/2$ when $n=2$, the projection of the trajectory into $[-1,1]$ will converge to the point $1/2$ in backwards time, and thus the trajectory in $[1,1]\times\bar{\Delta}$ will converge backwards to $\mathcal{M}_3$. If $f$ becomes constant before the projection of the trajectory into $[-1,1]$ has passed the point $1/2$, since all trajectories starting in $(-1/2,1)$ will converge to the fixed point at $-1/2$ when $n=1$, the projection of the trajectory into $[-1,1]$ will converge to the point $-1/2$ in backwards time, and thus the trajectory in $[1,1]\times\bar{\Delta}$ will converge forwards to $\mathcal{M}_2$. \\
\indent Now consider a trajectory starting at $(y,f)$ where $y\in(-1,1)$ and there does not exist a $z$ such that $f(t)=1$ or $f(t)=2$ for all $t\geq z$ or $t\leq z$. Then both forwards and backwards in time, the projection of this point's trajectory into $[-1,1]$ will pass the fixed points at $-1/2$ and at $1/2$, and thus will converge backwards to $\mathcal{M}_1$ and forwards to $\mathcal{M}_4$.
Consider a trajectory starting at $(y,f)$, where $y\in(-1/2,1/2)$ where there exist $z_1$ and $z_2$, $z_1<z_2$ where $f(t)=1$ for all $t\leq z_1$ and $f(t)=2$ for all $t\geq z_2$. If $f$ becomes constant at $n = 1$ while the trajectory is in $(-1/2,1/2)$, then backwards in time the trajectory will converge towards $-1/2$ in $[-1,1]$, and the function $f$ will converge backwards to $g_1 \equiv 1$. Thus $\alpha(y,f)=\mathcal{M}_2$. However, if $f$ becomes constant before the trajectory has entered $(-1/2,1/2)$, then it will converge backwards in time to $\mathcal{M}_1$. Meanwhile, forwards in time, if $f$ becomes constant at $n = 2$ while the trajectory is in $(-1/2,1/2)$, then forwards in time the trajectory will converge towards $1/2$ in $[-1,1]$, and the function $f$ will converge to $g_2 \equiv 2$. Thus $\omega(y,f)=\mathcal{M}_3$. However, if $f$ becomes constant at 2 after the trajectory has exited $(-1/2,1/2)$, then forwards in time the trajectory will converge to $\mathcal{M}_4$. Thus, all $\alpha,\omega$-limit sets are contained in some $\mathcal{M}_i$.
\item\textbf{No Cycles} Note from the description above that given $\mathcal{M}_i$ and $\mathcal{M}_j$ where $i \neq j$ there do not exist points $(x,f)$ and $(y,g)$ in $M \times \bar{\Delta}$ such that $\alpha(x,f)\subseteq\mathcal{M}_i$ and $\omega(x,f)\subseteq\mathcal{M}_j$, and $\alpha(y,g)\subseteq\mathcal{M}_j$ and $\omega(y,g)\subseteq\mathcal{M}_j$. Since the possibilities mentioned in part 6 are exhaustive, there are no cycles.
\end{enumerate}
\end{proof}
This example is important because $\pi_{\bar{\Delta}}(\mathcal{M}_{2,3}) \neq \bar{\Delta}$. This implies that the projection of Morse sets on $M \times \bar{\Delta}$ onto $\bar{\Delta}$ do not need to be Morse sets on $\bar{\Delta}$.
\section{Conclusion}
We have seen that under the correct formulation, a hybrid system can be treated as a dynamical system on a compact space. We have studied the Morse sets on $\bar{\Delta}$, showing that the lifts of invariant communicating classes form a finest Morse decomposition. We have studied Morse decompositions of hybrid systems, concentrating on attracting and repelling Morse sets. In the case that the random perturbations are small, we have seen that a hybrid system can be expected to behave similarly to the unperturbed dynamical system. Finally, we have examined three examples of hybrid systems, which show that the limit sets of a hybrid dynamical system can be complicated objects, possibly with little relation to the limit sets of the individual dynamical systems comprising the hybrid system. Future research could further illuminate the characteristics of the limit sets and Morse sets of these systems.
\section{Acknowledgments}
We wish to recognize Chad Vidden for his helpful discussions and Professor Wolfgang Kliemann for his instruction and guidance. We would like to thank the Department of Mathematics at Iowa State University for their hospitality during the completion of this work. In addition, we'd like to thank Iowa State University, Alliance, and the National Science Foundation for their support of this research.
|
\section{Introduction}
The ATLAS~\cite{AtlasDetector} and CMS~\cite{CmsDetector} experiments reported the observation of a new particle in the search for the Standard Model (SM) Higgs boson at the LHC at the level of 5~$\sigma$ after a preliminary analysis of the data from proton-proton collisions at $\sqrt{s}$ = 7 and 8~TeV recorded in the years 2011 and 2012~\cite{HiggsDiscoveryAtlas,HiggsDiscoveryCms}.
The analysis is based on an integrated luminosity of 25~fb$^{-1}$, which represents 90\% of that delivered by the LHC.
In the following the mass and spin measurement is reviewed, as well as the couplings scale factor measurements
and the limits on new physics derived from these results~\cite{ATL-CONF-2013-014,ATLAS-CONF-2014-009,CMS-PAS-HIG-13-002,CMS-PAS-HIG-13-005,CMS-HIG-13-033}.
\section{Higgs mass and spin measurement}
The mass is measured independently by both experiments in the $H\!\rightarrow\!\gamma\gamma$ and $H\!\rightarrow\!ZZ^*\!\rightarrow\!4\ell$ channels. Both experiments combine the data from both channels to measure the mass of the Higgs, using the invariant mass of the two photons and the four leptons respectively, based on a profile likelihood ratio~\cite{ProfileLikelihood} test statistic where the Higgs boson mass $m_{H}$ is the parameter of interest and all other uncertainties are nuisance parameters.
The Higgs mass is $m_H=125.5\pm{0.2}(\text{stat})^{+0.5}_{-0.4}(\text{syst})$ GeV for ATLAS and $m_H=125.7\pm{0.3}(\text{stat})\pm{0.3}(\text{syst})$ GeV for CMS.
The measured mass value is close to 125.5~GeV as can be seen in Figure~\ref{fig:MassMeasurement}, that shows the contour plots for the signal strength as a function of the mass hypothesis.
The combined measurement is in the 68\% confidence interval for both channels in CMS and in the 95\% confidence interval for ATLAS.
Note there is a difference between the individual Higgs mass measurements in ATLAS, which results from the uncertainty in the measurement of electrons and photons and is correlated between the two channels.
However this mass difference ($\Delta m_H = m^{\gamma\gamma}_{H} - m^{4\ell}_{H}$) is compatible with zero at the level of 2.4~$\sigma$.
\begin{figure}[ht!]
\centering
\includegraphics[width=7cm]{fig_01a.pdf}
\includegraphics[width=6.5cm, trim=0 -30 0 0, clip]{fig_01b.pdf}
\caption{\label{fig:MassMeasurement}
Confidence level intervals in the ($\mu$,$m_H$) plane for the $H\!\rightarrow\!\gamma\gamma$ and $H\!\rightarrow\!ZZ^*\!\rightarrow\!4\ell$ channels and their combination, including all systematic uncertainties for the ATLAS~\cite{ATLAS-CONF-2014-009} and CMS~\cite{CMS-PAS-HIG-13-005} experiments.}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[height=7.5cm, trim=15 0 15 0, clip]{fig_02a.pdf}
\includegraphics[height=8.1cm, trim=10 0 5 0, clip]{fig_02b.pdf}
\caption{Higgs signal strength for different decay channels measured by the ATLAS~\cite{ATLAS-CONF-2014-009} and CMS~\cite{CMS-Higgs-Physics-Results} experiments.\label{fig:SignalStrength}}
\end{figure}
The signal strength being the ratio of the measured cross section to the one predicted by the SM ($\mu=\sigma/\sigma_{SM}$) \cite{YellowReport2013} is slightly higher in ATLAS due to an excess of events observed in both channels, but compatible with the one measured by CMS.
Likewise, the signal strength observed in the different single channel searches ($H\!\rightarrow\!\gamma\gamma$, $H\!\rightarrow\!\tau\tau$, $H\!\rightarrow\!WW$, $H\!\rightarrow\!ZZ$, $VH\!\rightarrow\!b\bar{b}$) are compatible across experiments, as shown in Figure~\ref{fig:SignalStrength}, even for the $VH\rightarrow b\bar{b}$ search where the central values are quite different, but the contours overlap.
The signal strength is an important parameter, large deviations of this quantity from unity could indicate evidence for new physics.
\begin{figure}[ht!]
\centering
\includegraphics[height=7cm]{fig_03a.pdf}
\includegraphics[height=7cm]{fig_03b.pdf}
\caption{\label{fig:SpinMeasurement}Test-statistic distributions for spin parity ($J^P$) hypotheses $0^+$ (SM) and $0^-$ in the $H\rightarrow ZZ^* \rightarrow 4\ell$ channel in ATLAS~\cite{CERN-PH-EP-2013-102} and CMS~\cite{CMS-PAS-HIG-13-002}. The CL${_s}$ probability used to reject the $0^-$ hypothesis is measured as the ratio between the upper tail of the $0^-$ distribution divided by the lower tail of the $0^+$ distribution at the point given by the data. In both experiments the $0^-$ hypothesis is discarded above 99\% probability.}
\end{figure}
In the SM the Higgs boson has a spin parity value of $J^P=0^+$
and different alternate hypotheses, namely ($0^-,1^+,1^-,2^+$), are tested against it.
In the SM the spin 1 is largely suppressed because of the Landau-Yang theorem~\cite{LandauYangTheorem} due to the observation of $H\rightarrow\gamma\gamma$ events, but it could still have other spin values.
For this, a CL$_s$ probability, as shown in equation~\ref{eq:Cls}, measures the odds of the alternate spin parity hypothesis to the SM hypothesis.
All hypotheses are rejected at more than 97\% in ATLAS and CMS, which indicates that if the observed boson is not the SM Higgs boson it must deviate very little from it.
Figure~\ref{fig:SpinMeasurement} shows the distributions of the test-statistic for the $0^+$ and $0^-$ hypothesis.
In this case, the alternate hypothesis $0^-$ is rejected with a confidence level above 99\%.
\begin{equation}
CL_{s}(J^P_\text{alt}) = \frac{p_0(J^P_\text{alt})}{1-p_0(0^+)}
\label{eq:Cls}
\end{equation}
It is also possible to probe the existence of Higgs production via electroweak processes in a model independent way by measuring the ratio of the gluon fusion ($ggH$) and top fusion ($ttH$) production signal strength to the vector boson fusion (VBF) and associated production (VH) signal strength. This ratio is $\mu_{VBF+VH}/\mu_{ggF+ttH} = 1.4^{+0.5}_{-0.4}(\text{stat})^{+0.4}_{-0.2}(\text{sys})$ for ATLAS and $\mu_{VBF+VH}/μ_{ggF+ttH} = 1.538^{+1.611}_{-0.743}$ for CMS, which corresponds to an evidence of VBF production at the level of 4.1$\sigma$ in ATLAS and 3.21$\sigma$ in CMS.
Figure~\ref{fig:VbfProduction} shows the 68\% probability contour plots of $\mu_{VBF+VH}$ versus $\mu_{ggF+ttH}$, where the SM Higgs boson point $\mu_{VBF+VH}$ = $\mu_{ggF+ttH}$ = 1 is compatible with all measurements.
\begin{figure}[ht!]
\centering
\includegraphics[height=6.5cm, trim=0 -30 15 0, clip]{fig_04a.pdf}
\includegraphics[height=6.7cm, trim=20 0 25 -10, clip]{fig_04b.pdf}
\caption{\label{fig:VbfProduction}Signal strength for VBF and VH production versus ggH and ttH production contour plots for ATLAS~\cite{ATLAS-CONF-2014-009} and CMS~\cite{CMS-PAS-HIG-13-005}.}
\end{figure}
\section{Higgs couplings}
Following the leading order tree level motivated framework described by the LHC Higgs Cross Section Working Group~\cite{YellowReport2013},
measurements of coupling scale factors are implemented for a set of benchmark scenarios.
In these measurements it is assumed that the signal observed originates from a single narrow resonance with the mass measured by the corresponding experiment, a negligible width, and a CP-even state.
Hence the product of the cross section times the branching ratio for an initial state $i$, a partial decay width into final state $f$ of $\Gamma_f$, and a total decay width of the Higgs boson $\Gamma_H$ as the sum of all visible and invisible final states, is given by
\begin{equation}
\sigma \times \text{BR}(i\rightarrow H \rightarrow f) =
\frac{\sigma_i \cdot \Gamma_f}{\Gamma_{\text{H}}}\text{.}\nonumber
\end{equation}
Scale factors $\kappa_i$ are added to each coupling and are fitted to the data to test for the modification of the magnitude of the coupling but not its tensor structure.
In this framework, each final state can involve more than one coupling.
The results for individual channels and their combination are shown in Figure~\ref{fig:Couplings}.
\begin{figure}[ht!]
\centering
\includegraphics[height=8.7cm,trim=60 0 10 -20,clip]{fig_05a.pdf}
\includegraphics[height=9.6cm,trim=0 0 0 0,clip]{fig_05b.pdf}
\caption{\label{fig:Couplings}Higgs boson couplings scale factors in different couplings parametrizations for ATLAS~\cite{ATLAS-CONF-2014-009} and CMS~\cite{CMS-PAS-HIG-13-005} experiments.}
\end{figure}
Fermion versus vector boson scale factor couplings are tested assuming that only SM particles contribute to the total width.
Figure~\ref{fig:CouplingsFV} shows the preferred values of coupling scaling factors to fermions ($\kappa_F$) and vector bosons ($\kappa_V$) in a two dimensional fit. The best fit values are are found to be compatible with the SM expectation at the level of 10\%, with a largely constrained $\kappa_F$ due to the observation of the coupling to fermions in the $H\rightarrow\tau\tau$ channel at the level of 4$\sigma$.
When no assumption is made to the total width the measurements in ATLAS are consistent with large signal strength in bosonic decays.
The custodial symmetry in SU(2) that keeps $\rho = m^2_W/m^2_Z \cdot cos^2 \theta_W \approx 1$ is tested by measuring the ratio of the coupling scale factors $\lambda_{WZ} = \kappa_W/\kappa_Z$. The measurement is compatible with the SM prediction and even when an effective scale factor ratio is added to $H\rightarrow\gamma\gamma$ to account for possible contributions beyond the SM.
Up and down type fermion symmetry which is of interest for two Higgs doublet models is probed by the ratio of up to down coupling scale factors $\lambda_{ud} = \kappa_u / \kappa_d$. This measurement provides a 3.6~$\sigma$ evidence of the coupling of the Higgs boson to down type fermions mostly coming from the $H\rightarrow\tau\tau$ measurement. Similarly, quark and lepton symmetry is probed with the ratio $\lambda_{l q} = \kappa_l / \kappa_q$, and a vanishing coupling of the Higgs boson to leptons is excluded at the 4~$\sigma$ level due to the $H\rightarrow\tau\tau$ measurement as in the previous case.
In addition, both experiments probe the contribution of beyond SM particles either in loops or in new final states by introducing effective coupling scale factors for the $H\rightarrow gg$ and $H\rightarrow\gamma\gamma$ vertices, $\kappa_g$ and $\kappa_\gamma$, and assuming the rest of the coupling scale factors to be as predicted by the SM. These vertices are very sensitive to unknown heavy particles beyond the SM.
In the first benchmark model it is assumed that there are no sizeable contributions to the total width by non-SM particles and the free parameters are $\kappa_g$ and $\kappa_\gamma$.
These measurements show the lowest compatibility with the SM in ATLAS due to the high signal strength value observed.
Finally, upper limits to the branching ratio to invisible or undetected final states can be set by considering $BR_{BSM} = \Gamma_{BSM}/\Gamma_{H} = 1 - \kappa^2_H \cdot \Gamma^{SM}_H / \Gamma_H$. CMS quotes $BR_{BSM} < $~0.52 at 95\% CL and ATLAS quotes $BR_{BSM} <$~0.41 for the same CL with a noticeable difference over the expected value ($BR_{BSM} < $~0.55) mostly constrained from channels sensitive to VBF and VH production like $H\rightarrow b\bar{b}$ and $H\rightarrow \tau\tau$.
\begin{figure}[ht!]
\centering
\includegraphics[height=6.5cm,trim=0 -30 0 0,clip]{fig_06a.pdf}
\includegraphics[height=6.5cm,trim=0 0 0 -10,clip]{fig_06b.pdf}
\caption{\label{fig:CouplingsFV}Higgs boson couplings scale factors to fermions ($\kappa_F$) versus vector bosons ($\kappa_V$) and 68\% CL contours assuming only SM contributions to the total width derived from the individual channels and their combination for ATLAS~\cite{ATLAS-CONF-2014-009} and CMS~\cite{CMS-Higgs-Physics-Results}.}
\end{figure}
\section{Limits on new physics}
The mass dependence of the couplings is probed by expressing the coupling scale factors to different species of fermions and vector bosons in terms of a mass scaling parameter ($\varepsilon$) and a vacuum expectation value ($M$).
Combined fits to measured rates are performed as a function of $\varepsilon$ and $M$, and the best fit point is compatible with the SM Higgs boson at the level of 1.5~$\sigma$, as it is shown in Figure~\ref{fig:MassScaling}.
The best fit for the vacuum expectation value is below the one for the SM because the overall signal strength is higher than 1.
If the Higgs were a composite particle following Mininal Composite Higgs Models (MCHM), the couplings to fermions and vector bosons could be modified by its compositeness scale parameter ($f$) such as $\xi = v^2/f^2$, and the SM Higgs boson is recovered in the limit $\xi\rightarrow 0$ and $f\rightarrow \infty$.
In particular, in the model MCHM4, the couplings scale factors are $\kappa = \kappa_F = \kappa_V = \sqrt{1-\xi}$ and in the MCHM5 $\kappa_V = \sqrt{1-\xi}$ and $\kappa_F = \frac{1-2\xi}{\sqrt{1-\xi}}$.
Figure~\ref{fig:Mchm} shows the two dimensional likelihood contours in the ($\kappa_V$ , $\kappa_F$ ) coupling plane, and the coupling predictions in the MCHM4 and MCHM5 models as parametric functions of the Higgs boson compositeness parameter $\varepsilon$.
The upper limit of the composinteness scale is $f$ $>$ 710 GeV for MHCM4 and $f$ $>$ 640 GeV for MHCM5.
\begin{figure}[ht!]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[height=5.5cm,trim=30 10 0 0,clip]{fig_07a.pdf}
\caption{\label{fig:MassScaling}}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\textwidth}
\includegraphics[height=5.5cm,trim=0 10 0 0,clip]{fig_07b.pdf}
\caption{\label{fig:Mchm}}
\end{subfigure}
\caption{(a) Contour plot of the vacuum expectation value ($M$) versus the mass scaling paramter ($\varepsilon$) in ATLAS~\cite{ATLAS-CONF-2014-010}. (b) Contour plot of the coupling scale factors $\kappa_F$ versus $\kappa_V$ in ATLAS~\cite{ATLAS-CONF-2014-010}.}
\end{figure}
In the two Higgs Doublet Model (2HDM) where the SM Higgs sector is extended by an additional doublet predicting the existence of five Higgs bosons, two neutral CP-even, one neutral CP-odd and tho charged bosons, with a vacuum expectation value that follows the relation $v^2_1 + v^2_2 = v^2 \approx (246 \text{ GeV})^2$, the type II category is the most interesting among the four that exist depending on their coupling constants.
In the type II 2HDM where couplings are different for up-type quarks and for down-type quarks and leptons, and are completely determined by the mass of the CP-odd scalar ($m_A$) and the ratio between the up and down type fermions ($\tan\beta=v_u/v_d$) results can be interpreted in a simplified MSSM model. In this model the loop corrections from stops in $ggF$ production and $\gamma\gamma$ decays are ignored and the Higgs boson decays to supersymmetric particles are neglected.
The measured production and decay rates are expressed in terms of the corresponding couplings, assuming identical production to the SM but without any assumption on the total Higgs width. The measurements are shown in Figure~\ref{fig:LimitsMssm}.
The observed exclusion limit at 95\% CL is stronger than expected since measured rates in $H\rightarrow \gamma\gamma$ and $H\rightarrow ZZ^*\rightarrow 4\ell$ are higher than predicted.
\begin{figure}[ht!]
\centering
\includegraphics[height=6cm,trim=0 10 0 10,clip]{fig_08.pdf}
\caption{\label{fig:LimitsMssm}Exclusion region at 95\% CL in the ($m_A$,$\tan\beta$) plane in a simplified MSSM model via a fit to the measured rates of the Higgs boson production and decays for ATLAS~\cite{ATLAS-CONF-2014-010}.}
\end{figure}
It is conceivable that the Higgs particle may have other decay channels that are not predicted by the SM.
In a wide context, the Higgs boson could be coupled to the particle that constitutes all or part of the dark matter in the universe. These are so called Higgs portal models.
The upper limit on the branching ratio of the Higgs boson to invisible final states is derived using the combination of rate measurements from the individual channels and the measured upper limit on the rate of the $Zh\rightarrow\ell\ell + E_T^{miss}$ process, proportional to $BR_{BSM}$. The couplings of the Higgs boson to massive particles other than Weak Interacting Massive Particles (WIMPs) are assumed to be equal to the SM predictions, allowing the corresponding partial decay widths and invisible decay width to be inferred.
Limits are considerably more stringent at low mass and degrade as $m_X$ approaches $m_H/2$ as shown on Figure~\ref{fig:LimitsDarkMatter}.
\begin{figure}[ht!]
\centering
\includegraphics[height=6.4cm,trim=8 0 5 0,clip]{fig_09a.pdf}
\includegraphics[height=6.4cm,trim=10 0 15 0,clip]{fig_09b.pdf}
\caption{\label{fig:LimitsDarkMatter}Upper limits at 95\% CL on the WIMP-nucleon scattering cross section as a function of the WIMP mass $m_\xi$ (scalar, vector or fermion) compared with the limits from direct searches for dark matter at the confidence levels indicated for ATLAS~\cite{ATLAS-CONF-2014-010} and CMS~\cite{CMS-PAS-HIG-13-030}.}
\end{figure}
\section{Conclusions and outlook}
ATLAS and CMS discovered a Higgs like particle with a mass close to 125.5~GeV, and measured the spin, parity and signal strength to be compatible with the one from the SM Higgs boson ($J^P=0^+$ , $\mu_{VBF+VH}$ = $\mu_{ggH+ttH}$ = 1).
In a coupling scale factors analysis, compatibility with the SM is found in all the tests performed, with probabilities ranging from 7\% to 21\%.
Consequently, the Higgs physics potential of the LHC Run-I is almost exploited.
Run-II and beyond will offer the possibility to measure the couplings more precisely, further constrain rare decays, and determine a possible CP admixture of the Higgs boson.
|
\section{Introduction}
\label{sec:intro}
In wireless communications,
{\em relay stations} are used to relay radio signals
between radio stations that cannot directly communicate
with each other due to the signal attenuation.
On the other hand,
it is important to efficiently utilize the scarce bandwidth
due to the limitation of frequency resources
\cite[Chap.~1]{Gol}.
For this purpose,
{\em single-frequency network} is often preferable
in which signals with the same carrier frequency are transmitted through communication networks.
Then, a problem of {\em self-interference} caused by coupling waves arises
in a full-duplex relay station in a single-frequency network \cite{Jai+11}.
Fig.~\ref{coupling} illustrates self-interference by coupling waves.
In this picture, radio signals with carrier frequency $f_1$
are transmitted from the base station (denoted by BS).
One terminal (denoted by T1) directly receives the signal from the base station,
but the other terminal (denoted by T2)
is so far from the base station that they cannot communicate directly.
Therefore, a relay station (denoted by RS) is attached between them
to relay radio signals.
Then, radio signals with carrier frequency $f_1$
from the transmission antenna of the relay station
are fed back to the receiving antenna
directly or through reflection objects.
As a result, self-interference is caused in the relay station,
which may deteriorate the quality of communication
and, even worse, may destabilize the system.
For the problem of self-interference,
adaptive methods have been proposed
to cancel the effect of coupling waves:
a least mean square (LMS) adaptive filters
\cite{SakOkaHay06}, and adaptive array antennas
\cite{NogHayKanSak12}.
In these studies, a relay station is modeled by
a discrete-time system, and the performance
is optimized in the discrete-time domain.
However, radio waves are in nature continuous-time
signals and hence the performance should be
discussed in the continuous-time domain.
In other words, one should take account of {\em intersample behavior}
for coupling wave cancelation.
In theory, if the signals are completely band-limited below the
Nyquist frequency,
then the intersample behavior can be restored
from the sampled-data in principle \cite{Sha},
and the discrete-time domain approaches might work well.
However, the assumption of perfect band limitedness is
hardly satisfied in real signals;
real baseband signals are not fully band-limited
(otherwise they must be non-causal \cite[Chap.~1]{Skl}),
pulse-shaping filters, such as raised-cosine filters, do not act perfectly,
and the nonlinearity in electric circuits
adds frequency components beyond the Nyquist frequency.
One might think that if the sampling frequency is fast enough,
the assumption is almost satisfied and there is no problem.
But this is not true; firstly, the sampling frequency cannot be arbitrarily
increased in real systems, and secondly, even though the sampling
is quite fast, intersample oscillations may happen
in feedback systems \cite[Sect.~7]{Yam99}.
\begin{figure}[t]
\centering
\scalebox{0.55}{\includegraphics{coupling.eps}}
\caption{Self-interference}
\label{coupling}
\end{figure}
To solve the problem mentioned above,
we propose a new design method for coupling wave cancelation
based on the {\em sampled-data control theory}
\cite{CheFra,Yam99}.
We model the transmitted radio signals and coupling waves
as continuous-time signals, and optimize the
worst case continuous-time error due to coupling waves
by a {\em digital} canceler.
This is formulated as a sampled-data $H^\infty$ optimal control problem,
which is solved via the fast-sampling fast-hold (FSFH) method
\cite{KelAnd92,YamMadAnd99}.
In this study, we consider two types of digital canceler:
feedforward and feedback cancelers.
For a feedforward canceler%
\footnote{A feedforward canceler was first reported in \cite{SasNagHayYam14}.},
we cancel self-interference by
a discrete-time (virtual) model of the coupling wave path
that is optimized via sampled-data $H^\infty$ optimization
\cite{YamNagKha12,NagYam13}.
For a feedback canceler,
we place a digital controller in the feedback loop
for stabilizing the feedback system as well as canceling
the self-interference.
This is formulated as a standard sampled-data $H^\infty$
control problem except for the time delay in the feedback loop,
which can be solved via FSFH as well.
Design examples are shown to illustrate the proposed methods.
The reminder of this article is organized as follows.
In Section \ref{sec:ff},
we formulate a design problem of feedforward cancelers.
In Section \ref{sec:fb}, we formulate a feedback canceler design problem
as a sampled-data $H^{\infty}$ optimal control problem,
which can be solved via FSFH approximation
described in Section \ref{sec:FSFH}.
In Section \ref{sec:sim}, simulation results are shown to illustrate the effectiveness of the proposed method.
In Section \ref{sec:conc}, we offer concluding remarks.
\subsection*{Notation}
Throughout this article, we use the following notation.
We denote by $L^2$ the Lebesgue space consisting of all square integrable real functions
on $[0, \infty)$ endowed with $L^2$ norm $\|\cdot\|_{L^2}$,
and $\ell^2$ the space consisting of all square summable sequences,
with $\ell^2$ norm $\|\cdot\|_{\ell^2}$.
The symbol $t$ denotes the argument of time, $s$ the argument of Laplace transform and $z$ the argument of $Z$ transform.
These symbols are used to indicate whether a signal or a system is of continuous-time or discrete-time.
The operator $e^{-Ls}$ with nonnegative real number $L$ denotes continuous-time delay operator
with delay time $L$.
A continuous-time (or discrete-time) system $G$ with transfer function
$C(sI-A)^{-1}B+D$ (or $C(zI-A)^{-1}B+D$) is denoted by
\[
G = \ssr{A}{B}{C}{D}.
\]
\section{Feedforward cancelers}
\label{sec:ff}
In this section, we formulate
the design problem of feedforward cancelers.
Fig.~\ref{outside} shows the block diagram of a relay station
using an amplify and forward relaying protocol
with a coupling wave path and a digital feedforward canceler.
\begin{figure}[t]
\scalebox{0.85}{\includegraphics{outside.eps}}
\caption{Feedforward canceler}
\label{outside}
\end{figure}
In Fig.~\ref{outside},
continuous-time signals are represented in solid lines and
discrete-time signals in dotted lines.
We model the relay station by a continuous-time linear time-invariant (LTI) system
with transfer function $G(s)$.
The characteristic of the coupling wave path is also modeled by a
continuous-time LTI system denoted by $P(s)$ with time delay $e^{-Ls}$
where $L$ is a fixed nonnegative real number (i.e. $L$ is a fixed delay time).
Note that $P(s)$ will be time-varying in general due to the Doppler shift caused by the movement of reflection objects, however, the assumption of the LTI system is valid if the movement is slow or the signal components coming from the moving objects are not dominant.
The digital canceler consists of three operators:
ideal sampler ${\mathcal S}_h$ with sampling period $h>0$,
digital filter $K(z)$, and zero-order hold ${\mathcal H}_h$ with the same sampling period $h$.
The ideal sampler is
defined by
\[
\begin{split}
{\mathcal S}_h: \{y(t)\} &\mapsto \{y_d[n]\}:
y_d[n] = y(nh),\\
n &= 0,1,2,\ldots,
\end{split}
\]
and the zero-order hold is defined by
\[
\begin{split}
{\mathcal H}_h: \{u_d[n]\} &\mapsto \{u(t)\}:
u(t) = u_d[n],\\
t &\in [nh,(n+1)h), \quad n=0,1,2,\ldots.
\end{split}
\]
From Fig.~\ref{outside}, we have
\begin{equation}
y = v + (e^{-Ls}PG - {\mathcal H}_hK{\mathcal S}_h)y. \label{eqn1}
\end{equation}
To model the characteristic of the input signal $y$,
we introduce a subset $FL^2\subset L^2$ defined by
\[
FL^2 := \{y=Fw: w \in L^2, ~\| w \|_{L^2}=1\},
\]
where $F$ is a continuous-time LTI system with
real-rational, stable, and strictly proper transfer function $F(s)$.
The transfer function is a frequency domain weighting function
that gives the frequency characteristic of $y$ \footnote{
Although the assumption that $y\in FL^2$ is artificial,
we here consider a much wider class of signals
than the band-limited signal class assumed in
Shannon's theorem.
}.
Note that this signal model allows non band-limited signals
such as rectangular waves.
For any $y\in FL^2$, we try to {\em uniformly} minimize the error
\[
\begin{split}
e&:=(e^{-Ls}PG-{\mathcal H}_hK{\mathcal S}_h)y\\
&=(e^{-Ls}PG-{\mathcal H}_hK{\mathcal S}_h)Fw
\end{split}
\]
In other words, we minimize the $H^\infty$ norm
of the error system (see Fig.~\ref{error})
\[
{\mathcal{E}}(K) := (e^{-Ls}PG-{\mathcal H}_hK{\mathcal S}_h)F,
\]
that is,
\[
\inf_{K: {\rm stable}} \|{\mathcal{E}}(K)\|_\infty
=\inf_{K: {\rm stable}}\sup_{\substack{w\in L^2\\ \|w\|_{L^2}=1}} {\|\mathcal{E}}(K)w\|_{L^2}.
\]
\begin{figure}[t]
\scalebox{0.9}{\includegraphics{error_system.eps}}
\caption{Error system}
\label{error}
\end{figure}
The optimal discrete-time filter $K(z)$ can be obtained via
FSFH discretization \cite{KelAnd92,YamMadAnd99}.
For details of the design procedure, see \cite{NagYam13}.
If we use the $H^\infty$ optimal filter $K(z)$ for the relay station
that achieves sufficiently small $\|{\mathcal{E}}(K)\|_\infty$,
the effect of the coupling wave, $y-v$, is sufficiently
reduced.
In fact, we have the following theorem:
\begin{theorem}
Assume $\|{\mathcal{E}}(K)\|_\infty\leq\gamma$ with $\gamma>0$.
Then for any $y\in FL^2$ we have
$\|y-v\|_{L^2} \leq \gamma$.
\end{theorem}
\begin{proof}
For any $y\in FL^2$, there exists $w\in L^2$ such that $y=Fw$ with $\|w\|_{L^2}=1$.
This and equation (\ref{eqn1}) give
\[
\begin{split}
\|y-v\|_{L^2} &= \|(e^{-Ls}PG-{\mathcal H}_hK{\mathcal S}_h)y\|_{L^2} \\
&= \|(e^{-Ls}PG-{\mathcal H}_hK{\mathcal S}_h)Fw\|_{L^2} \\
&\leq \|{\mathcal{E}}(K)\|_\infty\\
&\leq \gamma.
\end{split}
\]
\end{proof}
This theorem motivates the proposed $H^\infty$ optimal design
for cancelation of coupling waves.
\section{Feedback cancelers}
\label{sec:fb}
The feedforward canceler shown in Fig.~\ref{outside}
works well when the gain of $G(s)$ is low.
If the gain of $G(s)$ is very high, the self-interference feedback loop
including the coupling wave path may become unstable.
Since the feedforward canceler design does not take the stability
into account, it cannot generally stabilize the feedback loop.
Therefore, we here consider a {\em feedback} canceler
to stabilize the feedback loop as well as reducing the effect of self-interference.
Fig.~\ref{model0} shows the block diagram of a relay station
attached with a digital feedback canceler
${\mathcal{H}}_hK(z){\mathcal{S}}_h$.
The difference between this and the feedforward canceler
in Fig.~\ref{outside} is that the canceler is placed in the feedback loop.
\begin{figure}[t]
\scalebox{0.6}{\includegraphics{model0.eps}}
\caption{Feedback canceler}
\label{model0}
\end{figure}
Our problem here is to
design the digital controller, $K(z)$, that stabilizes the feedback loop
and minimize the effect of self-interference, $z:=v-u$,
for any $v$.
We restrict the input continuous-time signal $v$ to the following
set:
\[
WL^2 := \{v = Ww: w \in L^2, \|w\|_{L^2}=1\},
\]
where $W$ is a continuous-time LTI system with real-rational, stable,
and strictly proper transfer function $W(s)$.
The design problem is formulated as follows:
\begin{problem}
Design digital controller (canceler) $K(z)$ that stabilizes
the self-interference feedback loop and uniformly minimizes
the $L^2$ norm of the error $z=v-u$ for any $v\in WL^2$.
\end{problem}
This problem is reducible to a standard sampled-data $H^\infty$ control problem
\cite{CheFra,Yam99}.
To see this, let us consider the block diagram shown in Fig.~\ref{model}.
\begin{figure}[t]
\centering
\scalebox{0.55}{\includegraphics{model.eps}}
\caption{Block diagram for feedback canceler design}
\label{model}
\end{figure}
Let $T_{zw}$ be the system from $w$ to $z$.
Then we have
\[
z = v - u = T_{zw}w
\]
and hence uniformly minimizing $\|z\|_{L^2}$ for any $v\in WL^2$ is equivalent to
minimizing the $H^\infty$ norm of $T_{zw}$,
\begin{equation}
\|T_{zw}\|_\infty = \sup_{w\in L^2, \|w\|_{L^2}=1} \|T_{zw}w\|_{L^2}.
\label{eq:hinf_Tzw}
\end{equation}
Let $\Sigma(s)$ be a generalized plant given by
\[
\Sigma(s) = \begin{bmatrix}W(s)&-1\\ W(s) & e^{-Ls}P(s)G(s)\end{bmatrix}.
\]
By using this, we have
\[
T_{zw}(s) = {\mathcal{F}}(\Sigma(s), {\mathcal{H}}_hK(z){\mathcal{S}}_h),
\]
where ${\mathcal{F}}$ denotes the linear-fractional transformation (LFT)
\cite{CheFra}. Fig.~\ref{geplant} shows the block diagram of this LFT.
Then our problem is to find a digital controller $K(z)$ that minimizes
$\|T_{zw}\|_\infty$.
\begin{figure}[t]
\centering
\scalebox{0.7}{\includegraphics{geplant.eps}}
\caption{LFT $T_{zw} = {\mathcal{F}}(\Sigma, {\mathcal{H}}_hK{\mathcal{S}}_h)$}
\label{geplant}
\end{figure}
This is a standard sampled-data $H^\infty$ control problem,
and can be efficiently solved via
FSFH approximation.
We discuss this in the next section.
Note that if there exists a controller $K(z)$ that minimizes
$\|T_{zw}\|_\infty$, then the feedback system is stable
and the effect of self-interference $z=v-u$ is bounded by
the $H^\infty$ norm.
We summarize this as a theorem.
\begin{theorem}
Assume $\|T_{zw}\|_\infty \leq \gamma$ with $\gamma > 0$.
Then the feedback system shown in Fig.~\ref{model0} is stable,
and for any $v\in WL^2$ we have $\|v-u\|_{L^2}\leq \gamma$.
\end{theorem}
\begin{proof}
First, if the feedback system is unstable, then the $H^\infty$ norm becomes
unbounded.
Next, for $v\in WL^2$ there exists $w\in L^2$ such that
$v=Ww$ and $\|w\|_{L^2}=1$.
Then, inequality $\|T_{zw}\|_\infty \leq \gamma$ gives
\[
\|v-u\|_{L^2} = \|T_{zw}w\|_{L^2} \leq \|T_{zw}\|_\infty \|w\|_{L^2} \leq \gamma.
\]
\end{proof}
\section{Fast-sample fast-hold approximation}
\label{sec:FSFH}
In this section, we review the method of FSFH approximation
for sampled-data $H^\infty$ optimal controller design.
The idea of FSFH is approximating a continuous-time $L^2$ signal
by a piecewise constant signal, which is generated by a fast hold
${\mathcal H}_{h/N}$ where $N$ is an integer greater than 2,
and evaluating the $L^2$ norm of an $L^2$ signal on the sampling points
generated by ${\mathcal S}_{h/N}$.
We call ${\mathcal S}_{h/N}$ and ${\mathcal H}_{h/N}$
a fast sampler and a fast hold, respectively.
That is, we connect the fast sampler and hold to the continuous-time
signals, $z$ and $w$, in Fig.~\ref{geplant} respectively, to make
a generalized plant with discrete-time input/output
as shown in Fig.~\ref{FSFHplant}.
\begin{figure}[t]
\centering
\scalebox{0.65}{\includegraphics{FSFHgeplant.eps}}
\caption{FSFH approximation of $T_{zw}$}
\label{FSFHplant}
\end{figure}
Roughly speaking, if we take $N\rightarrow\infty$, then the frequency response
of the FSFH approximation ${\mathcal S}_{h/N}T_{zw}{\mathcal H}_{h/N}$
uniformly approaches that of sampled-data system $T_{zw}$,
see \cite{YamMadAnd99} for details.
Since the $H^\infty$ norm defined in \eqref{eq:hinf_Tzw} is equivalent to
the maximum gain of the frequency response gain of the sampled-data
system $T_{zw}$ (defined via lifting)~\cite[Chap.~13]{CheFra},
we can obtain an approximated solution of our
$H^\infty$ optimal control problem if we take a sufficiently large $N$.
The FSFH approximation of sampled-data $T_{zw}$
in FIg.~\ref{FSFHplant}
contains two sampling periods, $h$ and $h/N$,
and the whole system is periodically time-varying.
By using discrete-time lifting defined below,
the system can be equivalently converted to a
finite-dimensional discrete-time LTI system.
The discrete-time lifting is defined by
\[
\begin{split}
{\mathbf L}_N:
&\{x[0],x[1],\ldots\}\\
&\quad \mapsto
\left\{ \begin{bmatrix}x[0]\\\vdots\\x[N-1]\end{bmatrix},
\begin{bmatrix}x[N]\\\vdots\\x[2N-1]\end{bmatrix},\ldots
\right\},
\end{split}
\]
and its inverse by
\[
\begin{split}
{\mathbf L}_N^{-1}:
&\left\{ \begin{bmatrix}x_1[0]\\\vdots\\x_N[0]\end{bmatrix},
\begin{bmatrix}x_1[1]\\\vdots\\x_N[1]\end{bmatrix},\ldots
\right\}\\
& \quad \mapsto \{x_1[0],\ldots,x_N[0],x_1[1],\ldots,x_N[1],\ldots\}.
\end{split}
\]
By definition, discrete-time lifting ${\mathbf L}_N$ converts
a one-dimensional signal with sampling period $h/N$
to a $N$-dimensional signal with sampling period $h$.
Also, discrete-time lifting preserves the $\ell^2$ norm.
By using ${\mathbf L}_N$ and ${\mathbf L}_N^{-1}$,
we obtain a norm-equivalent discrete-time LTI system for
time-varying ${\mathcal S}_{h/N}T_{zw}{\mathcal H}_{h/N}$.
Let $\mathbf{c2d}$ denote the step-invariant transformation \cite{CheFra},
that is,
\[
\begin{split}
{\mathbf{c2d}}\left(\ssr{A}{B}{C}{D},h\right)
&:={\mathcal S}_{h}\ssr{A}{B}{C}{D}{\mathcal H}_{h}\\
&=\ssr{e^{Ah}}{\int_0^h e^{At}Bdt}{C}{D},
\end{split}
\]
and $\mathbf{lift}$ denote the discrete-time lifting transformation \cite{CheFra},
that is,
\[
\begin{split}
&{\mathbf{lift}}\left(\ssr{A}{B}{C}{D},N\right)
:= {\mathbf L}_N\ssr{A}{B}{C}{D}{\mathbf L}_N^{-1}\\
&\quad = \left[\begin{array}{c|cccc}
A^N & A^{N-1}B & A^{N-2}B & \ldots & B\\\hline
C & D & 0 & \ldots & 0\\
CA & CB & D & \ddots & \vdots\\
\vdots & \vdots & \vdots & \ddots & 0\\
CA^{N-1} & CA^{N-2}B& CA^{N-3}B & \ldots & D
\end{array}\right].
\end{split}
\]
Then we have the following theorem:
\begin{theorem}
Assume that $L=mh+\frac{k}{N}h$ for some integers $m\geq 0$ and $k\in\{0,\ldots,N-1\}$. Then,
for FSFH approximation ${\mathcal S}_{h/N}T_{zw}{\mathcal H}_{h/N}$,
there exists a discrete-time LTI generalized plant $\Sigma_{dN}$ such that
\[
\|{\mathcal S}_{h/N}T_{zw}{\mathcal H}_{h/N}\|_\infty = \|{\mathcal F}(\Sigma_{dN},K)\|_\infty,
\]
where the $H^\infty$ norm is defined by the $\ell^2$-induced norm.
Moreover, norm-equivalent $\Sigma_{dN}$ is given by
\[
\Sigma_{dN} := \begin{bmatrix}W_{dN}&-H_N\\ S_NW_{dN} & S_{N,k}z^{-m}P_{dN}G_{dN}H_N\end{bmatrix},
\]
where
\[
\begin{split}
W_{dN} &:= {\mathbf{lift}}\bigl({\mathbf{c2d}}(W,h/N),N\bigr),\\
P_{dN} &:= {\mathbf{lift}}\bigl({\mathbf{c2d}}(P,h/N),N\bigr),\\
G_{dN} &:= {\mathbf{lift}}\bigl({\mathbf{c2d}}(G,h/N),N\bigr),\\
H_N &= [\underbrace{1,1,\ldots,1}_N]^\top,\quad
S_N = [1,\underbrace{0,\ldots,0}_{N-1}],\\
S_{N,k} &= [\underbrace{0 \ 0 \ \ldots \ 0}_{k} \ 1 \underbrace{\ 0 \ \ldots \ 0}_{N-k-1}].
\end{split}
\]
\end{theorem}
\begin{proof}
The formulae are obtained by the method described in \cite[Chap.~8]{CheFra}.
\end{proof}
Note that $W_{dN}$, $P_{dN}$, and $G_{dN}$ are LTI.
Finally, our problem is reduced to a standard discrete-time $H^\infty$ control problem
with the LFT shown in Fig.~\ref{final plant}.
\begin{figure}[t]
\centering
\scalebox{0.8}{\includegraphics{finalplant.eps}}
\caption{Norm-equivalent discrete-time LTI system for FSFH approximation}
\label{final plant}
\end{figure}
In this figure,
\[
\underline{z_d}:= {\mathbf L}_Nz_d,\quad
\underline{w_d}:= {\mathbf L}_Nw_d,\quad
y_d := {\mathcal S}_h y,
\]
and $u_d$ is the output of the controller $K(z)$.
Then the optimal controller $K(z)$ for this standard $H^\infty$ control problem
is easily obtained by using \texttt{hinfsyn} function in MATLAB Robust Control Toolbox.
\section{Simulation}
\label{sec:sim}
In this section, we show simulation results to illustrates the effectiveness of the proposed methods.
We assume that
sampling period $h$ is normalized to $1$.
The coupling wave path is modeled by
\[
P(s) = \frac{0.25}{s+1},
\]
with time delay $L=1$, that is, the time delay is equal to the sampling period $h$.
The relay station is modeled by
$G(s) = 2.5$,
that is, the station amplifies input signals by 8 [dB].
For these parameters, we first design a feedforward canceler
proposed in Section \ref{sec:ff}.
We assume the frequency characteristic of input signals is given by
\[
F(s) = \frac{1}{2s+1}.
\]
Note that the magnitude of $F(j\omega)$ represents the envelope of
the spectra of the input signals (e.g. rectangular waves).
The discretization parameter for FSFH is set to $N=16$.
The obtained $H^\infty$-optimal $K(z)$ is of 18-th order.
With this filter, we simulate coupling wave canceling with a
periodic rectangular wave input with period $8h$.
Note that this signal contains frequency components beyond
the Nyquist frequency, $\pi/h=\pi$ [rad/sec], although the frequency
of the periodic wave,
$\pi/8h=\pi/8$ [rad/sec] is much lower than $\pi$.
Fig.~\ref{fig:ff} shows the reconstructed signal $y$ (see Fig.~\ref{outside})
by the proposed feedforward canceler,
the input signal $v$, and the signal $y$ with no canceler.
\begin{figure}[t]
\centering
\includegraphics[width = 0.98\linewidth]{ff.eps}
\caption{Feedforward cancelation: input signal (dash-dot line),
reconstructed signal $y$ by feedforward canceler (solid line),
signal $y$ with no canceler (dashed line)}
\label{fig:ff}
\end{figure}
This result shows that the feedforward canceler works well.
To see this more precisely, we compute the effect of the coupling wave,
$|y(t)-v(t)|$, which is shown in Fig.~\ref{fig:ff_error}.
\begin{figure}[t]
\centering
\includegraphics[width = 0.98\linewidth]{ff_error.eps}
\caption{Coupling wave effect $|y(t)-v(t)|$
by feedforward canceler (solid line) and
with no canceler (dashed line)}
\label{fig:ff_error}
\end{figure}
This result shows the proposed canceler well cancels the self-interference.
A drawback of the feedforward canceler is that it
never works if the gain of $G(s)$ is so high
that the feedback loop is unstable.
For example, if we take
$G(s) = 1000$,
that is, the relay station amplifies input signals by 60 [dB],
then the feedback loop becomes unstable.
For this situation, we adopt a feedback canceler proposed in Section \ref{sec:fb}.
The frequency characteristic $W(s)$ is assumed to be the same as $F(s)$,
that is, $W(s)=F(s)=1/(2s+1)$.
The other parameters are the same as those for the feedforward canceler design.
With FSFH discretization number $N=16$,
we compute the $H^\infty$-optimal $K(z)$ by the method
described in Section \ref{sec:FSFH}.
Fig.~\ref{fig:fb} shows the reconstructed signal $u$ in
the feedback canceler (see Fig.~\ref{model0}).
\begin{figure}[t]
\centering
\includegraphics[width = 0.98\linewidth]{fb2.eps}
\caption{Feedback cancelation: input signal (dash-dot line) and
reconstructed signal $u$ by feedback canceler (solid line).}
\label{fig:fb}
\end{figure}
Note that with any feedforward canceler,
the signal should diverge because the feedback loop
around the relay station itself is unstable.
On the other hand, the feedback canceler shows
small reconstruction error as shown in Fig.~\ref{fig:fb_error}.
\begin{figure}[t]
\centering
\includegraphics[width = 0.98\linewidth]{fb_error2.eps}
\caption{Coupling wave effect $|v(t)-u(t)|$
by feedback canceler shown in Fig.~\ref{model0}.}
\label{fig:fb_error}
\end{figure}
In summary, by the simulation results, sampled-data $H^\infty$ optimal design is
proved to be effective
for coupling wave canceling.
\section{Conclusions}
\label{sec:conc}
In this article, we have proposed feedforward/feedback cancelers
based on the sampled-data $H^\infty$ control theory.
The design of feedforward cancelers is reduced to a
sampled-data $H^\infty$ optimal discretization problem,
while that of feedback cancelers is formulated by
a standard sampled-data $H^\infty$ control problem.
They can be numerically solved by the FSFH method.
Simulation results have been shown to illustrates the effectiveness of the proposed
feedforward/feedback cancelers.
Future work may include FIR (Finite Impulse Response) filter design as
proposed in \cite{NagYam14},
adaptive FIR filtering as proposed in
\cite{NagHamYam13},
robust filter design against uncertainty in the coupling wave path,
and implementation of the designed filter.
\hyphenation{KAKENHI}
\section*{Acknowledgment}
This work was supported in part by JSPS {KAKENHI} Grant Nos.\ 24360163, 24560457, 24560490,
24560543, and 26120521, and an Okawa Foundation Research Grant.
|
\section{Introduction}
\indent
From a theoretical point of view, auctions are modeled as games of
incomplete information in which asymmetric information among players
(seller/buyer and bidders) is one of the key features, \cite{Krishna2002, McAfeeMcMillan1987, Wilson1992}.
From an applied perspective, as auction is a widely used
mechanism to allocate goods and services, many data sets are available for empirical research.
By assuming
that observed bids are the equilibrium outcomes of an underlying
auction model under consideration, the structural approach to
provides a framework analyze auction data in which the theoretical
model and its empirical counterpart are closely related. The main
objective of this approach is then to recover the structural
elements of the auction model. This line of research has been
considerably developed in the last fifteen years. The difficulties
in estimating auction models are many. First, auction models lead to
nonlinear econometric models through the equilibrium
strategies. Second, auction models may not lead to a closed-form solution making the derivation of an
econometric model even more difficult. Third, the estimation of auction models often
requires the numerical computation of the equilibrium strategies.
Some important work in this are documented by \cite{PerrigneVuong1999, PaarschHong2006, AtheyHaile2007, PerrigneVuong2008}, among others.
We distinguish two methods for estimating structural
auction models: direct method and indirect method. Direct methods
were developed first, and they rely on parametric
econometric models. Starting from a specification of the underlying
distribution of private values, the objective of direct methods is
to estimate the parameter vector characterizing such a distribution.
Within this class of methods, there are two major estimation
procedures. The first methodology, introduced by \cite{Paarsch1992,DonaldPaarsch1993},
is a fully parametric setup that uses
Maximum Likelihood based estimation procedures requiring the
computation of the equilibrium strategy.
Since it is computationally demanding, \cite[see][]{DonaldPaarsch1993}, only very simple distributions are considered in
practice.
Moreover, because the support of the bid
distribution depends on the estimated parameter(s), it has a nonstandard limiting distribution, \cite[see][]{HiranoPorter2003}. In view of
this, \cite{DonaldPaarsch1993} develop a piecewise pseudo
maximum-likelihood estimator requiring the computation of the equilibrium strategy
that can be obtained using specific parametric distribution(s).
\cite{LaffontOssardVuong1995} introduced the second methodology,
which is computationally more convenient.
Relying on the revenue
equivalence theorem, the authors propose a simulation-based method
that avoids computation of the equilibrium strategy, and therefore
allows for more general parametric specifications for the
value distribution.
More recently \cite{GPV} (hereafter, GPV (2000)) developed an alternative, fully nonparametric indirect procedure. This methodology relies on a simple but crucial observation that, using the first-order condition of the bidder's optimization problem, the value can be expressed as a function of the (corresponding) bid, and the distribution and density of observed bids.
This function, which is the inverse of the equilibrium strategy, identifies the model nonparametrically.
Therefore, in contrast to the direct method, this method starts from the distribution of observed bids in order to estimate the distribution of unobserved private values without computing the Bayesian Nash equilibrium strategy or its inverse.
This naturally calls for a two-step procedure. In the first step, a sample of pseudo private values is obtained while using (say) kernel estimators for the distribution and density of observed bids.
In the second step, this sample of pseudo values is used to nonparametrically estimate its density.\footnote{ GPV (2000) also establish uniform consistency and, using the minimax theory as developed by \cite{IbragimovHasminskii1981}, determine the optimal rate of convergence of this estimator.}
Though a fully nonparametric (Kernel) estimator is flexible and robust to misspeficiation, it has few drawbacks.
It has slow rate of convergence, which makes it hard to accommodate a multidimensional auction covariates (
curse of dimensionality) and it is ill-behaved at the boundaries of the support.
To address these problems, we we propose a semiparametric procedure where the first step is
fully nonparametric, in that we use Local Polynomial Estimation (LPE) of \cite{FanGijbels1996}, instead of Kernel, to obtain the bid density and distribution, and in the second step we model private values through a set of conditional moment restrictions and estimate the (finite) parameters using generalized method of moments (GMM). We then derive the asymptotic properties of the estimator.
The advantage of using LPE is that it is well-behaved at the boundary, and by using conditional moment restrictions we can accommodate a large number of covariates, making our method useful for applied work. See for example \cite{BonetPesendorfer2003, Rezende2008, LiZheng2009, AtheyLevinSeira2011, KrasnokutskayaSeim2011, AtheyCoeyLevin2013, Groeger2014} who have used similar, either fully-parametric or semi parametric,
indirect moment based procedure to accommodate a large number of covariates.
None of them, however, provide any asymptotic properties for their
estimator.
We contribute to this literature by showing that our procedure is consistent, asymptotically normal and achieves parametric rate of convergence.
For notational tractability and relatively cleaner exposition we focus primarily on symmetric first-price
sealed-bid auction models with independent private value and a non-binding reserve price.
Once the asymptotic properties of this simple case has been characterized, extending the estimation procedure to accommodate more general auction environment is tedious but conceptually straightforward -- only the asymptotic variance will change, not the rate of convergence.
More generally, our method extends to models estimated using a nonparametric indirect procedure including auctions with asymmetric bidders.
Let $V_{p\ell}$, $p=1,\ldots, I_\ell$, $\ell=1,\dots, L$
denote the private value of the $p$th bidder for the $\ell$th
auctioned object. Let $Z_{\ell}\equiv(X_\ell,I_\ell)\in \hbox{\it I\hskip -2pt R}^{d+1}$
denote the vector of exogenous variables, it includes auction covariates $X_\ell$ and the number of
bidders $I_\ell$.
To model the private values, we posit that there is some known and sufficiently smooth function $M(\cdot,\cdot;\theta):\hbox{\it I\hskip -2pt R}^{d+2}\rightarrow \hbox{\it I\hskip -2pt R}^{q}$ and parameter vector $\theta
\in \hbox{\it I\hskip -2pt R}^p$ such that, $q\geq p$ and at some true parameter $\theta_{0}$ the values satisfy the following set of conditional moment
restrictions
\begin{eqnarray}
{\rm E}[M(V,Z;\theta_0)|Z]=0,\label{eq:1}
\end{eqnarray}
where the expectation is with respect to the value distribution $F(\cdot|Z;\theta_0,\gamma_0)$ with $\gamma_0$ as the (possibly infinite dimensional) nuisance parameter.
These moment conditions are, however, infeasible because $V$ are unobserved.
But, in equilibrium, the bid $B=s(V,Z;\theta_0,\gamma_0),$ where $s(\cdot)$ is the bidding strategy that depends on the parameter vector $\theta_0$ both
directly through $B$, since $B\sim G(\cdot|Z;\theta_0,\gamma_0)$ (say), and
indirectly through $V$, since $V\sim F(\cdot|Z;\theta_0,\gamma_0)$.
This means (\ref{eq:1}) can be naturally expressed as
$
{\rm E}\big\{M[s^{-1}(B,Z;\theta_0,\gamma_0),Z;\theta_0]\big|Z\big\}=0,
$
which requires the computation of the equilibrium
strategy as well as of its inverse. This could be computationally demanding for two different reasons. First, such
computation has to be carried out for any trial value of the
parameters $(\theta,\gamma)$.
Second, in a more general class of auction
models, such as when values are affiliated or when bidders are asymmetric, the computation of the equilibrium strategy
$s(\cdot,\cdot;\theta_0,\gamma_0)$ and of its inverse is much more
involved and costly.
Therefore, we propose to replace $V$ in Equation (\ref{eq:1}) by its nonparametric (local polynomial) estimator $\hat V=\hat \xi(B,Z)$ (and not inverse strategy) to make the moment condition feasible and operational. Thus the feasible conditional moment restriction becomes
\begin{eqnarray*}
{\rm E}\big\{M[\hat\xi(B,Z),Z;\theta_0]\big|Z\big\}\approx0.
\end{eqnarray*}
We propose a two-step semiparametric procedure: first, we use LPE to obtain the nonparametric estimator of the value $\hat V=\hat \xi(B,Z)$; second we
use GMM procedure to obtain an estimate for $\theta_0$.
Unlike the most widely used Parzen-Rosenblatt Kernel based estimator, LPE is not ill-behaved close to
the support boundaries \cite[see][]{FanGijbels1996} and hence we do not have to trim any bids.
This provides a remarkable advantage to our procedure, since otherwise we would have to trim bids, which are endogenous variables, which would then
imply an automatic trimming on private values, thereby affecting the moments. In a standard econometric framework only exogenous variables are trimmed, \cite{LavergneVuong1996,Robinson1988}.
We show that our estimator is consistent, asymptotically normal and
converges uniformly at the parametric $\sqrt{L}$ rate.
As it is well
known that nonparametric estimators converge at a slower rate than
$\sqrt{L}$ and their rates are negatively related to the
dimension of the vector of exogenous variables, the so-called curse of dimensionality.
This makes these
estimators less desirable in applications, especially when a limited
number of observations is available and/or when the number of
exogenous variables is relatively large.\footnote{ Examples of semiparametric estimators attaining $\sqrt{L}$ rate can be found in \cite{NeweyMcFadden1994,Powell1994}. An example of a semiparametric estimator
converging at a slower than the parametric rate but not subject to the curse of dimensionality -- its rate is
independent of $d$ -- is given by \cite{CGPV2011}.}
Our estimator does not have this drawback because its
convergence rate is independent of the dimension of
the exogenous variables.
A second major advantage of our estimation
procedure is that, even though we focus on symmetric, inpdendent private value auctions without reserve price, our method provides a framework for a (moment based) semiparametric procedure that can be used to estimate more general auction such as auctions with binding announced or random reserve price, \cite{LiPerrigne2003}, symmetric and asymmetric affiliated values, \cite{LiPerrigneVuong2002,CampoPerrigneVuong2003}, as long as the moment conditions are sufficiently smooth (defined later).
This rules out moment conditions that are based on quantiles.\footnote{ For examples of use of quantiles in empirical auctions see \cite{HaileHongShum2006, MarmerShneyerovXu2013, Gimenes2014}.}
In an short extension, we show how the semiparametric procedure can be applied to these auctions, including auctions with unobserved heterogeneity \cite{Krasnokutskaya2011}.
As it will be clear, allowing these features will affect the asymptotic variance but not the rate of convergence, except when auctions have unobserved heterogeneity.
This is because to accommodate unobserved heterogeneity we need a three-step semiparametric procedure -- the new step is to estimate the density of the unobserved heterogeneity using empirical characteristics function.
So it is not clear whether we can even achieve the $\sqrt{L}$ consistency, but a proper analysis of asymptotic properties of such semiparametric estimator beyond the scope of our paper and is left for future research.
The rest of the paper is organized as follows. In Section \ref{section:model}, we
introduce the theoretical model, from which the
structural econometric model and our semiparametric estimator is derived. Section \ref{section:asymptotic} establishes the asymptotic properties of our estimator, and \ref{section:MCMC} presents some Monte Carlo experiments to illustrate the properties of our procedure. Section \ref{section:extensions} proposes some extension, and we conclude in Section \ref{section:conclusion}. The
Appendix collects the proofs of our results.
\section{The Model\label{section:model}}
\subsection{The Symmetric IPV Model}
\indent
We present the benchmark theoretical model underlying our structural
econometric model, namely the symmetric IPV model with a non-binding
reserve price. Although this is somehow restrictive for
applications, it allows us to develop our
econometric procedure in a more transparent way.
A single and indivisible object is auctioned to $I_\ell$
risk neutral bidders who are assumed to be ex ante identical.
The
total number of bidders may vary across auctions.
Private values are denoted by $V$ and we assume that
each valuation $V_{p\ell}$, $\ell=1,\ldots,L$, $p=1,\ldots,I_\ell$,
is distributed according to $F(\cdot|Z_\ell;\theta_0,\gamma_0)$,
where $\theta_0 \in \hbox{\it I\hskip -2pt R}^p$ is the parameter of interest and
$\gamma_0$ is a nuisance parameter that could be infinite or finite dimensional or
even an empty set. The support of $F(\cdot|\cdot)$ is
$[\underline{V}_\ell, \overline{V}_\ell]$, with
$0\leq\underline{V}_\ell=\underline{V}(Z_\ell)< \overline{V}_\ell=
\overline{V}(Z_\ell)<\infty$.
Among others, \cite{RileySamuelson1981} have shown that for every $\ell$,
$I_\ell\geq2$ the equilibrium bid $B_{p\ell}$ in the $\ell$th
auction is given by
\begin{eqnarray}
B_{p\ell}=s_0(V_{p\ell},Z_\ell)=V_{p\ell}-\frac{1}{F(V_{p\ell}|Z_\ell;\theta_0,\gamma_0)^{I_\ell-1}}\int_{\underline{V}_\ell}^{V_{p\ell}}F(v|Z_\ell;\theta_0,\gamma_0)^{I_\ell-1} dv,
\label{BNE}
\end{eqnarray}
where $s(\cdot,\cdot)$ is the unique symmetric Bayes Nash Equilibrium strategy that is monotonic and differentiable.
Let $G(\cdot|Z_\ell; \theta_0, \gamma_0)\equiv G_0(\cdot|Z_\ell)$ and $g(\cdot|Z_\ell; \theta_0, \gamma_0)\equiv g_0(\cdot|Z_\ell)$ be the distribution and density of observed bids in the $\ell^{th}$ auction, respectively.
From GPV (2000), values $V$ can be identified as
\begin{eqnarray}
V_{p\ell}=\xi_0(B_{p\ell},Z_\ell)=B_{p\ell}+\frac{1}{I_\ell-1}\frac{G_0(B_{p\ell}|Z_\ell)}{g_0(B_{p\ell}|Z_\ell)},\quad p=1,\ldots,I_\ell; \ell=1,\ldots, L.\label{eq:gpv}
\end{eqnarray}
\subsection{The Two Step Estimator}
\indent
Similar to GPV (2000), (\ref{eq:gpv}) forms the basis for our
econometric model. The difference with GPV (2000) is to model
private values as a set of moment conditions. Therefore knowledge of
$G_0(\cdot|\cdot)$ and $g_0(\cdot|\cdot)$ would lead us to a
GMM framework. However, these functions are unknown in practice but
can be easily estimated from observed bids. This suggests the
following two-step procedure.
In the first step
we recover a sample of pseudo private values by using nonparametric
LPE. The second step departs from the nonparametric
second step of GPV (2000) since we use (parametric) GMM procedure to
obtain an estimator for $\theta_0$ instead. Before presenting our two-step
estimator, it is worth mentioning that some of our assumptions are
similar or even identical to those in GPV (2000). This is not
surprising since our methodology follows closely their methodology. In particular we follow GPV (2000) and indicate when some
modifications are necessary. Our first two assumptions deal with the data generating process and the smoothness of the latent
joint distribution of $(V_{p\ell},Z_\ell)$.
\medskip\noindent
{\bf Assumption A1:} {\em
\begin{itemize}
\item[(i)]
$Z_\ell=(X_\ell,I_\ell)\in \hbox{\it I\hskip -2pt R}^{d+1}$, $\ell=1,2,\ldots,L$ are
independently and identically distributed as $F_m(\cdot,\cdot)$ with
density $f_m(\cdot,\cdot)$.
\item[(ii)]
For each $\ell$, $V_{p\ell}$, $p=1,\ldots,I_\ell$ are independently and
identically distributed conditionally on $Z_\ell$ as
$F(\cdot|\cdot;\theta_0,\gamma_0)$ with density
$f(\cdot|\cdot;\theta_0,\gamma_0)$, where $\theta_0 \in \hbox{\it I\hskip -2pt R}^p$
and $\gamma_0$ can be finite or infinite dimensional or empty.
\end{itemize} }
\medskip\noindent
Let $\cal I$ be the set of possible values for $I_\ell$.
We use ${\mathcal S}(*)$ to denote the support of $*$, and use ${\mathcal S}_I(*)$ to denote the support when there are $I$ bidders.\footnote{ We use the notation $I_\ell$ (with the subscript $\ell$) to denote that there are $I_{\ell}$-many bidders in the $\ell^{th}$ auction, and $I$ (without the subscript $\ell$) to denote an auction with $I$-many bidders. For example, suppose there are $L=3$ auctions, with 2, 3 and 2 bidders in auction $\ell=1,2$ and $3$, respectively. Here $\ell \in \{1,2,3\}, I_1=2, I_2=3$ and $I_3=2$ and simply $I=2$ refers to auctions with 2 bidders, which is either auction 1 or 3.}
\medskip\noindent
{\bf Assumption A2:} {\em
${\cal I}$ is a bounded subset of $\{2, 3 ,\ldots \}$, and
\begin{itemize}
\item[(i)]
For each $I \in {\cal I}$, $ {\mathcal S}_i (F) =
\{ (v,x): x \in [\underline{x},\overline{x}],
v \in [\underline{v} (x), \overline{v} (x) ] \} $,
with $\underline{x}<\overline{x}$,
\item[(ii)]
For $(v,x,I) \in {\mathcal S}(F)$, $f(v|x,I;\theta_0,\gamma_0) \geq c_f > 0$, and
for $(x,I) \in {\mathcal S}(F_m)$, $f_m(x,I) \geq c_f > 0$,
\item[(iii)]
For each $I \in {\cal I}$, $F(\cdot | \cdot,I;\theta_0,\gamma_0 )$ and $f_m(\cdot,I)$ admit
up to $R+1$ continuous bounded partial
derivatives on ${\mathcal S}_I(F)$ and ${\mathcal S}_I(F_m)$, with $R > d+1$.
\end{itemize}}
\noindent These assumptions can be found in GPV (2000) as well, though A2-(iii) is stronger in our case. That is, we require $R$ to be sufficiently large with respect to the dimension of $X$, i.e. $R > d+1$, which is commonly used in the semiparametric literature, see \cite{PowellStockStoker1989} among others.
The next two assumptions are on kernels and bandwidths used in the first stage.
\medskip\noindent
{\bf Assumption A3:}{\em
\begin{itemize}
\item [(i)]
The kernels $K_G(\cdot)$, $K_{1g}(\cdot)$ and $K_{2g}(\cdot)$ are
symmetric with bounded hypercube supports and twice continuously
bounded derivatives.
\item [(ii)]
$\int K_G(x) dx =1$,
$\int K_{1g}(x) dx =1$, $\int K_{2g}(b) db=1$.
\item [(iii)]
$K_G(\cdot)$, $K_{1g}(\cdot)$ and $K_{2g}(\cdot)$ are of the order $(R-1)$
\end{itemize}}
\medskip\noindent
{\bf Assumption A4:} {\em The bandwidths $h_{G}$, $h_{1g}$ and $h_{2g}$ satisfy:
\begin{itemize}
\item [(i)]
$h_{G}\rightarrow 0$ and $\displaystyle{\frac{L h_{G}^d}{\log L} }\rightarrow \infty$, as $L\rightarrow\infty$,
\item [(ii)]
$h_{1g}\rightarrow 0$, $ h_{2g}\rightarrow 0$ and $\displaystyle{\frac{L h_{1g}^dh_{2g}}{\log L} }\rightarrow \infty$, as $L\rightarrow\infty$.
\end{itemize}}
\medskip
For simplicity of presentation and tractability of the notations, in the remainder of the paper we will consider only univariate $X$, i.e., $d=1$, except in the Monte Carlo section when we consider $d=2$.
Since we prove that the rate of convergence is independent of $d$ (Proposition \ref{prop:2}) all the asymptotic results will work for $d>0$ except for the form of asymptotic variance, because
when we move from $d=1$ to $d>1$, we only have to adapt the dimension of the regressor, the degree of polynomial and the asymptotic variance. See \cite[][section 3]{RuppertWand1994} for an example of how to specify a polynomial with $d=2$.
In order to describe our two--step estimator, we observe first that, our objective is
to estimate the ratio
$\psi(\cdot|\cdot)=G_0(\cdot|\cdot)/g_0(\cdot|\cdot)$ by
$\hat{\psi}=\hat{G}(\cdot|\cdot)/\hat{g}(\cdot|\cdot)$ (see equation (\ref{eq:gpv})) using
LPE for each function.
From Proposition 1 in GPV (2000) we know that $G_0(\cdot|\cdot)$ is $R+1$ times continuously differentiable on its entire support and
therefore $g_0(\cdot|\cdot)$ is $R$ times continuously differentiable on its entire support as well.\footnote{Observe that by Proposition 1 in GPV (2000) we also know that the conditional density $g_0(\cdot|\cdot)$ is $R+1$
times continuously differentiable on a closed subset of the interior of the support. Thus the degree of smoothness close
to the boundaries and at the boundaries of the support is not $R+1$.} Given the smoothness of each function we propose to
use a LPE($R$), i.e. a LPE of degree $R$, for $G_0(\cdot|\cdot)$ and a LPE($R-1$) for $g_0(\cdot|\cdot)$.
For consistency of the first step
it is possible to choose the optimal bandwidths \`{a} la \cite{Stone1982}.
However, unlike GPV (2000) we do not need to specify a ``boundary bandwidth'' since the local polynomial method does not require knowledge of the location of the endpoints of the support. Therefore, it is not necessary to estimate the boundary of the support of the bid distribution.
Let $P_\rho(X;\beta)$ denote a polynomial of degree $\rho$ in $X$ with parameter $\beta$. Then for each each $I$,
\begin{eqnarray*}
\hat{G}(b|x)&=&\arg\min_{\beta_G} \sum_{\{\ell:I_\ell=I\}}^L\sum_{p=1}^I\Big\{Y_{p\ell}^G-P_R(X_\ell-x;\beta_G)\Big\}^2 \frac{1}{h_{G}}K_G\left(\frac{X_\ell-x}{h_G}\right)\\
\hat{g}(b|x)&=&\arg\min_{\beta_g} \sum_{\{\ell:I_\ell=I\}}^L\sum_{p=1}^I\Big\{Y_{p\ell}^g-P_{R-1}(X_\ell-x;\beta_g)\Big\}^2 \frac{1}{h_{1g}}K_{1g}\left(\frac{X_\ell-x}{h_{1g}}\right),
\label{LPE_G}
\end{eqnarray*}
where $Y_{p\ell}^G=\hbox{\it 1\hskip -3pt I}(B_{pl}\leq b)$ and$Y_{p\ell}^g=\frac{1}{h_{2g}}K_{2g}\left(\frac{B_{p\ell}-b}{h_{2g}}\right).$
More precisely we have,
\begin{eqnarray}
\hat G(b|x,I)&=&\frac{1}{h_{G}}\sum_{\{\ell:I_\ell=I\}}^L\sum_{p=1}^{I}e_1^T (X_{I,R+1}^T W_{x}^GX_{I,R+1})^{-1} X_{R+1,\ell}K_G\left(\frac{X_\ell-x}{h_{G}}\right) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq b)\nonumber\\
&=&\!\!\frac{1}{Lh_{G}}\frac{L}{n_I}\!\!\sum_{\{\ell:I_\ell=I\}}^L\sum_{p=1}^{I}e_1^T \left(\frac{X_{I,R+1}^T W_{x}^GX_{I,R+1}}{n_I}\right)^{-1}\! \!\!X_{R+1,\ell} K_G\left(\frac{X_\ell-x}{h_{G}}\right) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq b);\label{eq:Ghat}\\
\hat g(b|x,I)&=&\frac{1}{h_{1g}h_{2g}}\sum_{\{\ell:I_\ell=I\}}^L\sum_{p=1}^{I}e_1^T (X_{I,R}^T W_{x}^gX_{I,R})^{-1} X_{R,\ell}K_{1g}\left(\frac{X_\ell-x}{h_{1g}}\right) K_{2g}\left(\frac{B_{p\ell}-b}{h_{2g}}\right)\nonumber\\
&=&\frac{1}{L h_{1g}h_{2g}}\frac{L}{n_I}\!\!\!\sum_{\{\ell:I_\ell=I\}}^L\!\sum_{p=1}^{I}e_1^T \left(\frac{X_{I,R}^T W_{x}^gX_{I,R}}{n_I}\right)^{-1}\!\!\!\!\! X_{R,\ell}K_{1g}\left(\frac{X_\ell-x}{h_{1g}}\right) K_{2g}\left(\frac{B_{p\ell}-b}{h_{2g}}\right),\label{eq:ghat}
\end{eqnarray}
where for $\iota\in\{R,R+1\}$,
$e_1$ is the unit vector in $\hbox{\it I\hskip -2pt R}^{\iota}$ containing a 1 in its first entry,
$n_I=I L_I$,
$L_I=\#\{\ell:I_\ell=I\}$,
$X_{\iota,\ell}=[1 \quad (X_\ell-x)\ldots (X_\ell-x)^{\iota-1}]^T$ is a $\iota\times1$ vector,
$$X_{I,\iota}=\left(
\begin{array}{cccc}
1&(X_1-x) &\ldots &(X_1-x)^{\iota-1}\\
\vdots &\vdots &\vdots &\vdots \\
1&(X_{n_I}-x) &\ldots &(X_{n_I}-x)^{\iota-1}
\end{array}\right)$$
is the $n_I\times \iota$ matrix of regressors with the first $I$ rows identical and similarly for the other rows,
\begin{eqnarray*}
W^G_{x}={\rm{diag}}\left\{\frac{1}{h_{G}}K_G\left(\frac{X_\ell-x}{h_{G}}\right)\right\};\quad
W^g_{x}={\rm{diag}}\left\{\frac{1}{h_{1g}}K_{1g}\left(\frac{X_\ell-x}{h_{1g}}\right)\right\},
\end{eqnarray*}
where $K_G(\cdot)$, $K_{1g}(\cdot)$ and $K_{2g}(\cdot)$ are some kernels with
bounded support and $h_{G}$, $h_{1g}$,$h_{2g}$ are some
bandwidths (see Assumptions A3 and A4). Given (\ref{eq:Ghat}) and (\ref{eq:ghat}), the (pseudo) private value is given by
\begin{eqnarray}
\hat V_{p\ell}=B_{p\ell}+\frac{1}{I_\ell-1}
\hat{\psi}(B_{p\ell}|Z_\ell).\label{eq:vhat}
\end{eqnarray}
Unlike in GPV (2000), $\hat{\psi}$ is not subject to the so-called boundary effect, a typical problem encountered in kernel estimation, and hence we do not need to trim out observations that are ``too close'' to the boundary of the support of the joint distribution of $(B_{p\ell},Z_\ell)$.
The second step of our estimation procedure is as follows. We propose to use the sample of pseudo private values in the following conditional moment restrictions, namely
\begin{eqnarray*}
{\rm E}\big[M(\hat V,Z;\theta_0)\big|Z\big]\approx0,
\end{eqnarray*}
\noindent for some \emph{known} function
$M(\cdot,\cdot;\theta):\hbox{\it I\hskip -2pt R}^{3}\rightarrow\hbox{\it I\hskip -2pt R}^{q}$ and $\theta
\in \hbox{\it I\hskip -2pt R}^p$ with $q\geq p$.
For example, we could use
\begin{eqnarray}
{\rm E}[\ln ( V_{p\ell}) \mid Z_{\ell}]&=& \theta_{0,1}'Z_{\ell}.\\
Var[\ln (V_{p\ell}) \mid Z_{\ell}]&=& [\exp(\theta_{0,2}'Z_{\ell})]^{2}.\label{eq:moment-example}
\end{eqnarray}
as the moment conditions, like in \cite{ KrasnokutskayaSeim2011, BajariHoughtonTadelis2014}.
This set of conditional moment restrictions translates into the following set of unconditional moment restrictions,
\begin{eqnarray}
{\rm E}\big[m(\hat V,Z;\theta_0)\big]\approx0,\label{eq:moment}
\end{eqnarray}
\noindent where $m(\cdot,\cdot;\theta):\hbox{\it I\hskip -2pt R}^{3}\rightarrow
\hbox{\it I\hskip -2pt R}^q$ is known.
In view of (\ref{eq:moment}), we propose to estimate $\theta_0$ by $\hat\theta$, where
\begin{eqnarray}
\hat{\theta}=\arg\min_{\theta\in\Theta} \hat S_L^T(\theta)\Omega \hat
S_L(\theta),\label{eq:thetahat}
\end{eqnarray}
\noindent where $\hat S_L(\theta)=1/L\sum_{\ell=1}^L1/I_\ell\sum_{p=1}^{I_\ell} m(\hat V_{p\ell},Z_\ell;\theta)$ and $\Omega$ is a
positive definite matrix of order $q$.
Ideally, one would like to specify the following set of conditional moment restrictions
${\rm E}[M(V,Z;\theta_0)|Z]=0,
$
which would lead to the unconditional moment
restrictions
$
{\rm E}[m(V,Z;\theta_0)]=0.
$
Therefore, if $S_L(\theta)=1/L\sum_{\ell=1}^L1/I_\ell\sum_{p=1}^{I_\ell} m(V_{p\ell},Z_\ell;\theta)$ the infeasible estimator $\tilde \theta$, (say), is such that
\begin{eqnarray*}
\tilde{\theta}=\displaystyle{\arg\min_{\theta\in\Theta}
S^T_L(\theta)\Omega S_L(\theta)}.\end{eqnarray*}
\noindent\textbf{Remark--} The asymptotic distributions of the
feasible estimator $\hat\theta$ and the infeasible estimator $\tilde\theta$ are closely related, but are not the
same, see Proposition \ref{prop:2}.
\section{Asymptotic Properties\label{section:asymptotic}}
\indent
In this section we show that our two-step semiparametric estimator
$\hat\theta$ of $\theta_0$ is consistent and asymptotically normal
distributed. Moreover, we establish that our estimator attains the
parametric uniform rate of convergence given an appropriate choice
of the bandwidths used in the first step to estimate
$G_0(\cdot|\cdot)$ and $g_0(\cdot|\cdot)$. As we will discuss below
the optimal bandwidths, given by \cite{Stone1982}, i.e. the one-step bandwidths, cannot
be chosen, instead our choice implies that in practice one needs to
undersmooth.
We also discuss the assumptions under which our results
hold.
\subsection{Consistency}
\indent
Our first result establishes that $\hat\theta$ is a
(strongly) consistent estimator for $\theta_0$. Moreover this is the case even if one uses
the optimal bandwidths for estimating $G_0(\cdot|\cdot)$ and
$g_0(\cdot|\cdot)$ in the first step, i.e. the bandwidths proposed
by \cite{Stone1982}. To see this, we notice that the ``optimal one-step'' bandwidths satisfy our assumption A4 above (with $d=1$) since they are of the form,
\begin{eqnarray*}
h_{G}=\lambda_{G}\left( \frac{\log L}{L}\right)^{1/(2R+3)};\quad h_{1g}=\lambda_{1g}\left( \frac{\log L}{L}\right)^{1/(2R+1)};\quad h_{2g}=\lambda_{2g}\left( \frac{\log L}{L}\right)^{1/(2R+1)},
\end{eqnarray*}
where $\lambda_{G}$, $\lambda_{1g}$ and $\lambda_{2g}$ are strictly positive constants.
As observed by GPV (2000), $h_{G}$, $h_{1g}$ and $h_{2g}$, as given above are
optimal bandwidth choices to estimate $G_0(\cdot|\cdot)$ and
$g_0(\cdot|\cdot)$ given Proposition \ref{prop:1} and A2-(iii) in that
paper.\footnote{As pointed out before, A2-(iii) in our case is
stronger than A2-(iii) in GPV (2000). Thus their Proposition 1 also
holds in our framework.}
Thus, A4 implies that our consistency
result can be established when using LPE in the first
stage that converge at the best possible rate.
\medskip\noindent
{\bf Assumption A5:} {\em
\begin{itemize}
\item[(i)]
The parameter space $\Theta\subset\hbox{\it I\hskip -2pt R}^p$ is compact and $\theta_0$
is in the interior of $\Theta$,
\item [(ii)]
Identifying assumption: ${\rm E}[m(V,Z;\theta)]=0$ if and only if
$\theta=\theta_0$,
\item[(iii)]
$\displaystyle{\sup_{\theta \in \Theta}\left\vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \left\Vert
m(V_{p\ell},Z_\ell; \theta)\right\Vert- {\rm E}\left\Vert
m(V,Z;\theta)\right\Vert\right\vert}=o_{as}(1)$,
\item[(iv)]
$m(V,Z;\theta)$ is Lipschitz in $V$-- there exists a measurable
function $K_1(Z), {\rm
E}[K_1]<\infty$ such that
\[\forall V,V'\in[\underline{V},\overline{V}],\forall\theta\in\Theta,\quad\!\!\left\Vert m(V,Z;\theta)-m(V',Z;\theta)\right\Vert\leq K_1(Z)
\left\vert V-V'\right\vert.\]
\end{itemize}}
Let $m_{k}(\cdot,\cdot)$ be the partial derivative of $m(\cdot,\cdot)$ with respect to its $k^{th}$ argument.
\medskip\noindent
{\bf Assumption A6:} {\em
\begin{itemize}
\item[(i)]
$m_3(V,Z;\theta)$ is Lipschitz in $V$: there exists a measurable
function $K_3(Z), {\rm
E}[K_3]<\infty$, such that $$\forall V, V' \in [\underline{V},\overline{V}], \forall \theta\in\Theta,\quad\!\!
\left\Vert m_3(V,Z;\theta)-m_3(V',Z;\theta)\right\Vert\leq K_3(Z)\vert V-V'\vert.$$
\item[(ii)]
$m_3({V},Z;\theta)$ is Lipschitz in $\theta$: there exists a
measurable function $K_4(Z), {\rm
E}[K_4]<\infty$ such that
$$\forall\theta,\theta'\in \Theta,V \in [\underline{V},\overline{V}],\quad\!\! \left\Vert m_3({V},Z;\theta)-m_3({V},Z;
\theta')\right\Vert\leq K_4(Z)\left\Vert \theta-
\theta'\right\Vert.$$
\item[(iii)]
$\displaystyle{\sup_{\theta \in\Theta} \left\Vert \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}
m_3(V_{p\ell},Z_\ell;\theta)-{\rm
E}[m_3(V,Z;\theta)]\right\Vert}=o_{as}(1)$
and
${\rm E}[m_3'(V,Z;\theta)]\Omega {\rm E}[m_3(V,Z;\theta)]$ is non
singular.
\item [(iv)]
$\displaystyle{\sup_{\theta \in\Theta}\left\Vert
m_3(V,Z;\theta)\right\Vert\leq K_5(V,Z)}$ with ${\rm
E}[K_5(V,Z)]<\infty$,
\item[(v)]
$m_1(V,Z;\theta)$ is Lipschitz in $V$: there exists a measurable
function $K_6(Z), {\rm
E}[K_6]<\infty$ such that
$$\forall V,V'\in[\underline{V},\overline{V}],\theta \in \Theta,\quad\!\! \left\Vert m_1(V,Z;\theta)-m_1(V',Z;\theta)\right\Vert\leq K_6(Z)\vert V-V'\vert.$$
\item [(vi)]
$\displaystyle{\sup_{\theta \in \Theta}\left\Vert
m_1(V,Z;\theta)\right\Vert \leq K_7(V,Z)}$ with ${\rm
E}[K_7(V,Z)^2]<\infty$.
\item [(vii)] ${\rm{E}}[m_1(V,Z;\theta_0)]<\infty$, where the expectation is with respect to the joint cdf of $(V,Z)$.
\end{itemize}}
\noindent Assumptions 5 and 6 are implied by the regularity conditions used in GMM estimators, \cite[see][]{NeweyMcFadden1994}.
These regularity conditions impose appropriate differentiability restrictions on the moment functions, which rule out certain kinds of moment conditions.
For instance these assumptions rule out moment conditions that define quantiles.
Let $\rho_{\tau} (s)= \mid s\mid +(2\tau - 1)$ where $\tau\in(0,1)$.
Then the conditional and unconditional moment conditions, \cite{Koenker2005}, in this case are respectively given by
\begin{eqnarray*}
{\rm E}[\rho(V,Z;\theta_{0}(\tau))]&=&{\rm E}[\mid V-Z'\theta_{0}(\tau) \mid+ (2\tau-1)]=0,\\
{\rm E}[Z\rho(V,Z;\theta_{0}(\tau))]&=&{\rm E}[Z(\mid V-Z'\theta_{0}(\tau) \mid+ (2\tau-1))].
\end{eqnarray*}
and the corresponding sample moment condition is
\begin{eqnarray*}
\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} Z_{\ell} (\mid V_{p\ell}- Z_{\ell}'\theta(\tau) \mid +(2\tau-1)).
\end{eqnarray*}
Since a function $|s|$ is not everywhere differentiable, our method does not apply because we use Taylor's series expansion.\footnote{ One can, however, use results from empirical processes to allow for non-smooth moment conditions. Typically those conditions impose sufficient differentiability of the distribution function $F(\cdot|\cdot;\cdot)$ and stochastic equicontinuity. Since one of the arguments of the moment conditions is estimated nonparametrically, verifying stochastic equicontinuity in our framework can be difficult. We want to thank one of the referees for this observation.}
Now, we show that the estimator is consistent.
\begin{proposition}\label{prop:1}Let $\hat\theta$ be defined as in (\ref{eq:thetahat}).
Then, under A1-A5 $\hat{\theta}\stackrel{a.s}{\longrightarrow}\theta_0$.
\end{proposition}
The proof is in the Appendix. This is the first step in order to be able to establish the asymptotic distribution of the estimator. Moreover, there is no need to under-smooth the distribution and density functions in the first step in order for $\hat\theta$ to be consistent.
\subsection{Asymptotic Normality}
\indent
Given that $\hat\theta$ is a (strongly) consistent estimator for $\theta_0$, in Proposition \ref{prop:2} we establish its asymptotic distribution and its uniform
convergence rate, under some additional regularity conditions.
Since optimal bandwidth choice requires under-smoothing in semiparametric procedures we have to modify our choice of bandwidths.\footnote{ Another typical property usually encountered
has to do with a sufficiently large degree of smoothness relative to
the dimension of the exogenous variables, as reflected by A2-(iii).}
Thus, for $\hat\theta$
to achieve the parametric uniform rate of convergence we need to
specify bandwidths for our first step that rule out the optimal
choice and moreover that imply under-smoothed estimates for $\hat
G(\cdot|\cdot)$ and $\hat g(\cdot|\cdot)$, as is made clear by A4.AN below.
\medskip\noindent
{\bf Assumption A4.AN:} {\em The bandwidths $h_{G}$, $h_{1g}$ and $h_{2g}$ satisfy
\begin{itemize}
\item[(i)]
$\sqrt{L}h_G^{R+1}\rightarrow 0$ and $\displaystyle{\frac{\log L}{\sqrt{L} h_{G}} }\rightarrow 0$, as $L\rightarrow\infty$,
\item[(ii)]
$\sqrt{L}h_{1g}^{R}\rightarrow 0$, $ \sqrt{L}h_{2g}^{R}\rightarrow 0$ and $\displaystyle{\frac{\log L}{\sqrt{L} h_{1g}h_{2g}} }\rightarrow 0$, as $L\rightarrow\infty$,
\item[(iii)]
$h_{1g}=h_{2g}$.
\end{itemize}}
\noindent The assumption that $h_{1g}$ and $h_{2g}$ vanish at the same rate, is to simplify the notation in the proof. In fact it is enough to choose any pair of bandwidths strictly smaller than their optimal counterparts.\footnote{For the multivariate case ($d>1$), these conditions become: \begin{itemize}
\item[(i)]
$\sqrt{L}h_G^{R+d}\rightarrow 0$ and $\displaystyle{\frac{\log L}{\sqrt{L} h_{G}^d} }\rightarrow 0$,
as $L\rightarrow\infty.$
\item[(ii)]
$\sqrt{L}h_{1g}^{R+d-1}\rightarrow 0$, $ \sqrt{L}h_{2g}^{R}\rightarrow 0$ and $\displaystyle{\frac{\log L}{\sqrt{L} h_{1g}^dh_{2g}} }\rightarrow 0$, as $L\rightarrow\infty$.
\end{itemize}}
\begin{proposition}\label{prop:2}
Let $\hat\theta$ be defined as in (\ref{eq:thetahat}). Then, under A1-A3, A4.AN and A5-A6, we have
\[\sqrt{L}(\hat\theta-\theta_0)\stackrel{d}{\longrightarrow}{\mathcal N}(0,\Sigma),\]
where for each $I\in \mathcal{I}, \Sigma={\rm{Var}}(\psi_1)$, with
\begin{eqnarray*}
\psi_1&=&-{1}/{I}\sum_{p=1}^{I}\Bigg\{(C^T\Omega C)^{-1}C \Omega
m(V_{p1},X_1,I;\theta_0) +2 \Bigg[\sum_{I}\frac{1}{I(I-1)}N(Y_{p1},I)\\
&& f_m^{-1}(X_1,I) g_0(Y_{p1},I)-{\rm{E}}\Big[\sum_{I}\frac{1}{I(I-1)}N(Y_{p1},I)f_m^{-1}(X_1,I)
g_0(Y_{p1},I)\Big]\Bigg]\Bigg\},\\
C&=&{\rm E}[\partial m(V,X,I; \theta_0)/\partial\theta], Y_{p\ell}\equiv (B_{p\ell}, X_{\ell})\\
N(Y_{p1},I)&=& [m_1(V_{p1},X_1,I;\theta_0)/g_0(B_{p1}|X_1,I)^2] G_0(B_{p1}|X_1,I).
\end{eqnarray*}
\end{proposition}
\noindent Proposition \ref{prop:2} is important for several reasons. First it establishes that our semiparametric estimator has a standard limiting distribution. Asymptotic Normality is fundamental since most of the econometric tests rely on it. Second, although slow estimators are used in the first step of our estimation procedure to recover pseudo private values, the estimator of the parameter of interest converges at the best possible rate. Third, our semiparametric estimator is not subject to the curse of dimensionality. Finally, Proposition \ref{prop:2} can be used to conduct inference on $\theta_0$.
There are already some empirical papers in the literature that fit in our framework. For example, \cite{ KrasnokutskayaSeim2011, AtheyLevinSeira2011} estimate auction models using semiparametric or fully parametric procedure, and if we ignore the unobserved heterogeneity, both these papers satisfy all our assumptions.
We conclude this section with few points about our procedure, especially that of the LPE, that deserve mention.
First, LPE regression is more computationally complex than the standard least squares method,
because a model must be fit for each observed data point.
With ``brute force'' methods, it
would take approximately $L\times I_{\ell}$ times longer to fit a local linear regression than it would take to fit a
``global'' linear regression; see \cite{SeifertBrockmannEngelGasser1994}.
This is without factoring all the calculations that go in kernel evaluations and choosing bandwidths.
Many methods for choosing the bandwidth $h$ rely
on cross-validation, \cite{FanGijbels1995,PrewittLohr2006}.
This necessitates solving the LPE minimization repeatedly.
The complexity multiplies as $d$ increases, for then we need higher degree polynomial which are difficult to evaluate.
So care must be given to using a ``quicker'' method for solving the minimization.
\cite{FanMarron1994, HallWand1996} propose to use ``updating'' and linear ``binning'' for this purpose.\footnote{ All of these methods are now easily implemented using statistical programming language like $\textbf{\textsf{R}}$.}
Second since in the equilibrium, and as mentioned in the introduction, $G(b|Z)\equiv F(s^{-1}(b;\theta_{0},Z)|Z)$, the parameter of interest $\theta_{0}$ enters the moment conditions directly and indirectly through the first stage nonparametric estimate of $\psi(\cdot)$.
This makes our estimation procedure different from the widely studied semiparametric method, for example \cite{Chamberlain1992}, where the parameter of interest does not enter the nuisance (nonparametric) first stage estimate, and as a result we lose some efficiency.
Had the first-step also been parametric then for a candidate $\theta$ the nuisance function would be calculated and then in the minimization step, $\theta$ would enter the moment conditions twice.
With nonparametric first-step, however, we did not have to fix $\theta$, but at a cost of higher variance or lower efficiency.\footnote{ We owe this obsevration to one of the referees.}
But determining the exact loss of efficiency would require us to determine semiparametric efficiency for the non-regular case which is considered to be a hard problem, \cite{IbragimovHasminskii1981,Newey1990,Chamberlain1992, BickelKlaassenRitovWellner1993}, and is beyond the scope of this paper.
\section{Monte Carlo Experiments\label{section:MCMC}}
Now, we want to see the performance of the estimator we proposed through two sets of Monte Carlo exercises.
In the first set, we consider one dimensional auction characteristics, i.e., $d=1$ and in the second set we consider $d=2$.
For both cases we fix number of bidders $I_{\ell}=5$ for all $\ell=1,\ldots, L$, where $L=200$ when $d=1$ but when $d=2$, we let $L$ to be either $200, 100$ or $50$.
We use two yardsticks to evaluate the performance of our estimator.
The first is the visual method where we present the estimated density using our procedure, and to facilitate comparison we also present the true density along with the estimated density that uses GPV (2000) method.
The second method is to compare the optimal ex-ante expected revenue for the seller.
To compute the revenue we first use the plug-in method to choose the optimal reserve price, $r=\frac{1-\hat{F}(r)}{\hat{f}(r)}$, \cite{Myerson1981} and then calculate the corresponding (maximized) expected revenue \cite{Krishna2002}
\begin{eqnarray}
\Pi(r)=I\left[r (1-\hat{F})(\hat{F}(r))^{I-1}+\int_{r}^{\overline{v}}(1-\hat{F}(t))t(I-1)(\hat{F}(t))^{(I-2)}\hat{f}(t)dt\right].\label{eq:revenue}
\end{eqnarray}
Like with the figures, we calculate the revenue corresponding to our semiparametric estimate, GPV (2000) estimate and the true density.
Since the final goal of estimating the value density is to choose optimal auction, comparing revenues across different estimators is a good way to judge the performance of the estimators -- the closer the revenue to the truth the better the estimates.
We present all of these results while fixing $X$ at its median value and find that our estimator performs well according to both measures.
\subsection{One Dimensional Covariate}
Let $X\sim \ln {\mathcal N} (0,1)$ truncated at 0.055 and 30 to satisfy A2-(i), and $V|X\sim F(\cdot|Z;\theta_0,\gamma_0)=\ln {\mathcal N}(1+X, 1)$ truncated at 0.055 and 30, so $\theta_0=(1,1)^T$ and $\gamma_0=\{\emptyset\}$.
While estimating, we assume that $R=3$.
In line with assumption A3, we choose the triweight kernel $(35/32)(1-u^2)^3 \hbox{\it 1\hskip -3pt I}(|u|\leq 1)$ for the
three kernels involved in our first step estimators.
We choose the bandwidths according to A4.AN. In particular we use $h_G=2.978\times1.06\hat\sigma_x (IL)^{-1/6.5}$,
$h_{1g}=2.978\times1.06 \hat\sigma_x (IL)^{-1/4.5}$, $h_{2g}=2.978\times1.06 \hat\sigma_b (IL)^{-1/4.5}$, where $\hat\sigma_b$ and $\hat\sigma_x$ are
the estimated standard deviations of observed bids and object heterogeneity, respectively. The factor $2.978\times1.06$ follows from the so-called rule of thumb \cite[see][]{Hardle1991}. The use of $I$ arises because we have $I$ bidders per auction.
To replicate the GPV (2000) estimator we choose the bandwidths according to the optimal rates. Thus, the order of the bandwidths
is $L^{-1/9}$ for $h_G$ and the second step bandwidth $h_x$ and $L^{-1/10}$ for $h_{gb}$ and $h_{gx}$ and the second step bandwidths $h_{fv}$ and $h_{fx}$. Specifically we use $h_G=1.06\hat\sigma_x (IL)^{-1/9}$,
$h_{gx}=1.06 \hat\sigma_x (IL)^{-1/10}$, $h_{gb}=1.06 \hat\sigma_b (IL)^{-1/10}$ where $\hat\sigma_b$ and $\hat\sigma_x$ are as defined above. The second step bandwidths are $h_{fv}=1.06 \hat\sigma_{\hat v} (n_t)^{-1/10}$, $h_{fx}=1.06 \hat\sigma_{x} (n_t)^{-1/10}$ and $h_x=1.06 \hat\sigma_{x} (L)^{-1/9}$, where $n_t$ is the number of observations remaining after trimming. See Table \ref{hD1} for all the bandwidths.\footnote{For GPV(2000) we also need to compute the boundary bandwidths.}
\begin{table}[t!]
\centering
\scalebox{1}{
\begin{tabular}{lccc}
\hline
& Notation & Constant & Rate \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{LPE}} & $h_G$ & $2.978\times 1.06\times \hat\sigma_x$ & $(IL)^{-1/6.5}$ \\
\multicolumn{1}{l}{} & \multicolumn{1}{c}{$h_{1g}$} & $2.978\times 1.06\times \hat\sigma_x$ & $(IL)^{1/4.5}$ \\
\multicolumn{1}{l}{} & \multicolumn{1}{c}{$h_{2g}$} & $2.978\times 1.06\times \hat\sigma_b$ & $(IL)^{-1/4.5}$ \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{GPV 1st step}} & $h_G$ & $1.06 \times \hat\sigma_x$ & $(IL)^{-1/9}$ \\
\multicolumn{1}{l}{} & $h_{gb}$ & $1.06\times \hat\sigma_b$ & $(IL)^{-1/10}$ \\
\multicolumn{1}{l}{} & $h_{gx}$ & $1.06\times \hat\sigma_x$ & $(IL)^{-1/10}$ \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{GPV 2nd step}} & $h_{f_{\hat v}}$ & $1.06\times \sigma_{\hat v}$ & $n_t^{-1/10}$ \\
\multicolumn{1}{l}{} & $h_{f_{x}}$ & $1.06\times \hat\sigma_x$ & $n_t^{-1/10}$ \\
\multicolumn{1}{l}{} & $h_x$ & $1.06\times \hat\sigma_x$ & $L^{-1/9}$ \\
\hline
\multicolumn{1}{l}{Boundary} & $h_\delta$ & $\lambda_{\delta}>0 $ & $n^{-1/2}$ \\
\hline
\end{tabular}}%
\caption{\footnotesize\textbf{Bandwidths when d=1.}}
\label{hD1}%
\end{table}%
We use 1000 replications for estimation where in each replication we: (i) generate randomly $IL$ private values using the truncated normal
distribution; (ii) compute the corresponding bids $B_{p\ell}$ using (\ref{BNE}); (iii) use these bids to estimate the distribution and density functions using (\ref{eq:Ghat}) and (\ref{eq:ghat}); (iii) determine the pseudo private values $\hat V_{p\ell}$ corresponding to $B_{p\ell}$; (iv) use this sample of pseudo private values to obtain $\hat{\theta}$ using
the sample moment condition
\begin{eqnarray}
\frac{1}{IL}\sum_{\ell=1}^L \sum_{p=1}^I \nabla_{\theta}\ln f(\hat V_{p\ell}|I,X;\theta) =0.\label{eq:score}
\end{eqnarray}
We now quickly verify that this data generating process satisfies Assumptions A1 -- A6.
It is immediate to verify that Assumption A1--A4 are satisfied because of the way we have designed the experiment.
In our estimation we will restrict our attention at finding parameters from a compact set, and since log-likelihood is concave and smooth Assumption A5 (i)--(iii) are satisfied. Although we do not show the derivation, we can use the mean-value theorem to bound the slope of the moment conditions with respect to $V$. This slope is highest but bounded when $V=0.055$, which satisfies A5 (iv).
Although tedious, we can still use the mean-value theorem to verify A6 (i), (ii) and (v).
Again, the first part of A6 (iii) follows from the regularity conditions and the law of large numbers, the second part and the rest of A6 are satisfied by design.
We present our estimator (labeled SP) along with GPV (2000) estimated density and the true density, all evaluated at the median $X$ in Figure \ref{fig:1}.
As is evident, our estimator is very close to the true density suggesting that it performs reasonably well.
Next, we calculate the expected revenue given in (\ref{eq:revenue}) with $\overline{v}=30$.
The true expected revenue is $\Pi_{True}=5.9$ and the revenue using our estimate gives $\Pi_{SP}= 5.8$ while using GPV gives $\Pi_{GPV}=4.6$, which means our estimate is much closer to the true value.
\begin{figure}[t!]
\centering
\includegraphics[width=3.4in]{mcpaper.eps}
\caption{{\bf Valuation Densities}. TRUE refers to the true density, while SP and GPV refer to the semiparametric and GPV(2000) estimators, respectively. The vertical lines correspond to the trimming that is required only for the GPV estimator.} \label{fig:1}
\end{figure}
\subsection{ Two Dimensional Covariate}
Let $X=(X_{1}, X_{2})^{T}\sim \log {\mathcal N}\left(\left(\begin{array}{c}1\\1\end{array}\right), \left(\begin{array}{cc}1 & 0.8\\0.8 &1\end{array}\right)\right)$, and $V|X\sim \log{\mathcal N}(\mu_{v}(X), \sigma_{v}^{2}(X))$, both truncated at 0.055 and 30.
As before we fix $I=5$ bidders in all auctions, but consider three different specifications for $(\mu_{v}(X), \sigma_{v}^{2}(X))$: (i) $(1+X_{1}/X_{2}, 1)$; (ii) $(1+X_{1}+X_{2}, 1)$; and (iii) $(1+X_{1}/X_{2}, \exp(0.01(X_{1}+X_{2})))$.
To compare the performance of our estimator we use simulated data from $L=200, 100, 50$ auctions.
In line with the assumption A3, we choose the products of tri-weight kernel $(35/32)(1-u^2)^3 \hbox{\it 1\hskip -3pt I}(|u|\leq 1)$ in our first-step.
Like before, we choose two bandwidths, one for each covariate, according to A4.AN, while ensuring under smoothing when compared to GPV estimator.
To replicate the GPV (2000) estimator we choose the bandwidths according to the optimal rates. Thus, the order of the bandwidths
is $L^{-1/12}$ for $h_G$ and the second step bandwidth $h_{x_j}$ and $L^{-1/13}$ for $h_{gb}$ and $h_{gx_j}$ and the second step bandwidths $h_{fv}$ and $h_{fx_j}$, for $j=1,2$, see Table \ref{hD2}.
\begin{table}[t]
\centering
\scalebox{0.95}{
\begin{tabular}{lccc}
\hline
& Symbol & Constant & Rate \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{LPE}} & $h_{Gj}$ & $2.978\times 1.06\times \hat\sigma_{xj}$ & $(IL)^{-1/8.5}$ \\
\multicolumn{1}{l}{} & \multicolumn{1}{c}{$h_{1gj}$} & $2.978\times 1.06\times \hat\sigma_{xj}$ & $(IL)^{1/9.5}$ \\
\multicolumn{1}{l}{} & \multicolumn{1}{c}{$h_{2g}$} & $2.978\times 1.06\times \hat\sigma_b$ & $(IL)^{-1/9.5}$ \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{GPV 1st step}} & $h_{Gj}$ & $1.06 \times \hat\sigma_xj$ & $(IL)^{-1/12}$ \\
\multicolumn{1}{l}{} & $h_{gb}$ & $1.06\times \hat\sigma_b$ & $(IL)^{-1/13}$ \\
\multicolumn{1}{l}{} & $h_{gxj}$ & $1.06\times \hat\sigma_{xj}$ & $(IL)^{-1/13}$ \\
\hline
\multicolumn{1}{l}{\multirow{3}[2]{*}{GPV 2nd step}} & $h_{f_{\hat v}}$ & $1.06\times \sigma_{\hat v}$ & $n_t^{-1/13}$ \\
\multicolumn{1}{l}{} & $h_{f_{xj}}$ & $1.06\times \hat\sigma_{xj}$ & $n_t^{-1/13}$ \\
\multicolumn{1}{l}{} & $h_{xj}$ & $1.06\times \hat\sigma_{xj}$ & $L^{-1/12}$ \\
\hline
\multicolumn{1}{l}{Boundary} & $h_\delta$ & $\lambda_{\delta}>0 $ & $n^{-1/3}$ \\
\hline
\end{tabular}}%
\caption{\footnotesize\textbf{Bandwidths when d=2}.}
\label{hD2}%
\end{table}%
We follow exactly the same steps as with $d=1$ to estimate the parameters, except that now we have three different sample sizes, $n=L\times I\in\{250,500,1000\}$ and three different specifications for mean and variance, and we use $\Theta= [0.055, 30]\times[0.001,2]$.
In total, there are 9 different cases, and hence 9 different densities.
Figure \ref{fig:2} below shows the true density of private values against our estimator (the dashed line) and the GPV estimator (the dotted line). As can be seen in Figure \ref{fig:2} our estimator performs really well, even when there are only 50 auctions, while GPV (2000) is infeasible because after trimming we had very few observations left.
The estimated revenues for each density is presented in Table \ref{table:revenue}.
And as before, our estimator still performs relatively well.
\begin{table}[h!]
\begin{center}
\scalebox{0.95}{
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|}
\hline
$(\mu(X),\sigma^{2}(X))$ & \multicolumn{3}{l|}{$\qquad$1} & \multicolumn{3}{l|}{$\qquad$2} & \multicolumn{3}{l|}{$\qquad$3} \\ \hline
Profit$\backslash$ L & 200 & 100 & 50 & 200 & 100 & 50 & 200 & 100 & 50 \\ \hline
$\Pi_{True}$ & 6.2 & 6.1 & 6.0 & 5.9 & 5.7 & 5.7 & 6.9 & 7.0 & 7.0 \\ \hline
$\Pi_{SP}$ & 5.5 & 5.2 & 4.8 & 5.4 & 5.5 & 5.5 & 6.6 & 6.5 & 6.5 \\ \hline
$\Pi_{GPV}$ & 4.2 & 3.6 & -- & 4.8 & 2.9 & -- & 4.6 & 4.0 & -- \\ \hline
\end{tabular}}
\caption{\footnotesize \textbf { Optimal Revenue}: Each column corresponds to the three sets of mean and variance, and each cell contains the optimal revenue as defined in (\ref{eq:revenue}), one for each density in Figure \ref{fig:2}.} \label{table:revenue}
\end{center}
\end{table}
\begin{landscape}
\begin{figure}[ht!]
\centering
\includegraphics[width=8.5in]{Sym_Case_d2.eps}
\caption{\footnotesize\textbf{Valuation Densities}. TRUE refers to the true density, while SP and GPV refer to the semiparmaetric and GPV(2000) estimators, respectively. Each row refers to a different DGP while each column refers to different number of auctions. } \label{fig:2}
\end{figure}
\end{landscape}
\section{Extensions\label{section:extensions}}
\indent
In this Section we indicate how to extend our procedure to a more general class of auction models.
\subsection{Binding Reserve Price}
\medskip
The first natural extension of the model considered in Section 2 is the symmetric IPV first-price auction model with a binding reserve price, announced or random.
\subsubsection{Announced Reserve Price}
An announced binding reserve price $(r_{0}>\underline{V})$ constitutes
a screening device for participating in the auction. As pointed out
by GPV (2000) the Bayesian Nash equilibrium strategy is still given
by (2) in this set up, but the number $I$ of potential bidders
becomes unobserved and typically different from the observed number,
$I^*$, of actual bidders who have submitted a bid $(\geq r_{0})$.
Hence the model has a new structural element, namely $I$, in
addition to the latent distribution of bidders' private values. As
shown in GPV (2000), the differential equation defining the
equilibrium strategy can be rewritten as
\begin{eqnarray*}
V_{p}=\xi_0(B_{p},G^*_0,F(r_{0}),I)=B_{p}+\frac{1}{I-1}\left(\frac{G^*_0(B_{p})}{g^*_0(B_{p})}+\frac{F(r_{0})}{1-F(r_{0})}\frac{1}{g^*_0(B_{p})}\right),
\end{eqnarray*}
\noindent for $p=1,\ldots,I^*$ and where $G^*_0(\cdot)$ is the
truncated distribution of an observed bid conditional upon the fact
that the corresponding private value is grater than or equal to
$r_{0}$. Provided one can estimate $I$ and $F(r_{0})$ this equation is
the basis for a two step procedure analogous to that of Section
\ref{section:model}.
In particular, for $\ell=1\ldots,L$ and $p=1\ldots,I$ if $B_{p\ell}\geq r_{0\ell}$ the above equation becomes
\begin{eqnarray*}
V_{p\ell}=B_{p\ell}+\frac{1}{I_{\ell}-1}\left(\frac{G^*_0(B_{p\ell}\vert I_{\ell},Z_{\ell},\theta_{0})}{g^*_0(B_{p\ell}\vert I_{\ell},Z_{\ell},\theta_{0})}+\frac{F(r_{0}\vert Z_{\ell},\theta_{0}))}{1-F(r_{0}\vert I_{\ell},Z_{\ell},\theta_{0}))}\frac{1}{g^*_0(B_{p\ell}\vert I_{\ell},Z_{\ell},\theta_{0}))}\right),
\end{eqnarray*}
where $Z_{\ell}=(r_{0\ell},X_{\ell})$ and $\theta_{0}$ is the unknown true parameter vector. Let $h(\cdot\vert X_{\ell},\gamma_{0})$ be the probability mass function of the number of potential bidders, $I_{\ell}$, which is known up to a finite parameter $\gamma_{0}$.
At the $\ell$-th auction, a potential bidder $p$ will bid if an only if his private valuation is above the reserve price. Thus, at this auction, the number of actual bidders is $I^{*}_{\ell}=1/I_{\ell}\sum_{p=1}^{I_{\ell}} \hbox{\it 1\hskip -3pt I}(V_{p\ell} \geq r_{0\ell})$, which is a binomial random variable with parameters $(I_{\ell}, 1-F(r_{0\ell}\vert X_{\ell},\theta_{0})).$ In view of these, we propose to use the following moment conditions for bids and the observed number of bidders
\begin{eqnarray*}
{\rm E}[V_{p\ell}\vert V_{p\ell}\geq \xi_{0}(r_{0\ell}),X_{\ell};\gamma]&=&\sum_{I_{\ell}\geq2} m(I_{\ell},r_{0\ell},X_{\ell};\theta_{0}) h(I_{\ell} \vert X_{\ell};\gamma),\\
{\rm E}[I_{\ell}^{*}\vert I_{\ell}, Z_{\ell}]&=&\sum_{I_{\ell}\geq2} I_{\ell} [1-F(r_{0\ell}\vert X_{\ell};\theta)] h(I_{\ell} \vert X_{\ell};\gamma) , \\
{\rm E}[I_{\ell}^{*2}\vert I_{\ell},Z_{\ell}]&=&\sum_{I_{\ell}\geq2} [1-F(r_{0\ell}\vert X_{\ell};\theta)] [F(r_{0\ell}\vert X_{\ell};\theta) +I_{\ell} (1-F(r_{0\ell}\vert X_{\ell};\theta))],
\end{eqnarray*}
where the moment function $m(\cdot,\cdot,\cdot;\cdot)$ can be similar to the ones in (\ref{eq:moment-example}).\footnote{ We thank an anonymous referee for suggesting us to give the moment conditions for this model.}
\subsubsection{Random Reserve Price}
In some cases, as in timber and wine auctions, the seller may decide not to announce the reserve price at the time the auction takes place. Hence, the reserve price is said to be secret or random. Since bidders do not know it when submitting their bids, this fact brings into the model a new kind of uncertainty that has to be taken into account. To present the basic equation underlying our two-step procedure in this model we need first to introduce additional notation.
To keep the notation as simple as possible we consider models without observed object heterogeneity. This is not restrictive since relaxing this assumption implies that the distribution and density functions have to be replaced by their conditional counterparts.
Let $V_0$ be the private value of the risk-neutral seller for the auctioned object. Moreover, we assume that $V_0$ is distributed according to $H(\cdot)$ defined on the same support as $F(\cdot)$ and that $H(\cdot)$ is common knowledge. \cite{ElyakimeLaffontLoiselVuong1994} have shown that in a first-price sealed bid auction $r_0=V_0$. In addition the bidders' equilibrium strategy is the solution of a differential equation which in general cannot be solved explicitly. See \cite{LiPerrigne2003}. However, this differential equation can be rewritten as follows
\begin{eqnarray*}
V_p=\xi_0(B_p,H,G_0,I)=B_{p}+\frac{1}{(I-1)\left(\frac{g_0(B_p)}{G_0(B_p)}+\frac{h(B_p)}{H(B_p)}\right)}\:,
\end{eqnarray*}
\noindent for $p=1,\ldots,I$. As mentioned by \cite{PerrigneVuong1999} since the reserve price is kept secret, all potential bidders
submit a bid. Hence $I$ is typically observed. The above equation
can be used as the basis of a two-step procedure similar to the one
described in Section 2. Namely, in a first step observed bids and
reserve prices can be used to estimate nonparametrically the
distribution $G_0(\cdot)$, its density $g_0(\cdot)$ as well as the
distribution $H(\cdot)$ and its density $h(\cdot)$. Next, pseudo
private values can be recovered using the equation above in order to
define a set of moment conditions for estimating the parameter of
interest $\theta_0$ in a second step.
We conclude this part by noting that because of the revenue equivalence principle, \cite{Myerson1981,RileySamuelson1981}, our method is also useful to study other standard auctions such as third-price auctions, \cite{KagelLevin1993}, all-pay auctions.\footnote{ We thank one of the referees for the suggestion. These results are available upon request.}
\subsection{The Symmetric Affiliated Private Value (APV) Model}
To assume independence across private value can be restrictive since one can expect some degree of affiliation or positive correlation among private values. Thus, another natural extension of our framework is to consider the more general class of model encompassed by symmetric APV models. Affiliation means that if one bidder draws a high valuation for the auctioned object, then others bidders are likely to draw higher valuations too. \cite{LaffontVuong1996} study the problem of identification and theoretical restrictions in a general framework, namely in Affiliated Value (AV) models.
In particular they show that any symmetric AV model is observationally equivalent to some symmetric APV model because the utility function is not identified from observed bids only.
Therefore when only data on observed bids are available, the result in \cite{LaffontVuong1996} implies that APV models can be considered, provided that we have identification.
We briefly indicate here how to adapt our estimation procedure to this kind of models, when all bids are observed and the reserve price is non-binding. Let $Y_p=\max_{p'\neq p}V_j$. The differential equation defining the equilibrium strategy in the APV model can be written as follows
\begin{eqnarray*}
V_p=\xi_0(B_p,G_0)\equiv B_p+\frac{G_{0,B_1|B_1}(B_p|B_p)}{g_{0,B_1|B_1}(B_p|B_p)},
\end{eqnarray*}
subject to the boundary condition $s(\underline{V})=\underline{V}$, where $G_{0,B_1|B_1}(X_1|X_1)=F_{Y_1|V_1}(s^{-1}(X_1))$, $B_1=s_{0}(Y_1)$. The index ``1'' refers to any bidder since bidder are assumed to be ex-ante symmetric.This equation is again the basis for the identification result and estimation procedure. See \cite{PerrigneVuong1999} for details.
The theoretical restrictions as shown by \cite{LiPerrigneVuong2002} indicate that the joint distribution of bids $G_0(\cdot)$ can be rationalized by a symmetric APV model if and only if (i) $G_0(\cdot)$ is symmetric and affiliated and (ii) the function $\xi_0(\cdot,G_0)$ is strictly increasing on its support. Moreover, if these two conditions are satisfied, then the joint distribution $F(\cdot)$ of private values is identified.
In view of their results, \cite{LiPerrigneVuong2002} propose a two step fully nonparametric procedure in the same sprite as GPV(2000). Nevertheless, for the affiliated model the procedure has to be performed for each size $I$ (i.e. for each given number of bidders).
Regarding estimation, the equation above suggests a two-step procedure analogous to the one described in Section \ref{section:model} for each size $I$ of bidders. In the first step the ratio $G_{{B_1|B_1}}(\cdot|\cdot)/g_{{B_1|B_1}}(\cdot|\cdot)$ can be estimated nonparametrically and then pseudo private values can be recovered.
In the second step a GMM procedure can be implemented to estimate the parameters of the underlying distribution of private values for a given $I$.
It is known that with affiliation, the rate of convergence is slower than with independence; this follows from Proposition 2 in \cite{LiPerrigneVuong2002}. In particular this Proposition gives explicit forms for the bandwidths that can be used in our framework since these choices satisfy our assumption A4.AN.\footnote{They consider homogenous auctions; see footnote 10 in \cite{LiPerrigneVuong2002}.}
\subsection{Asymmetric Models}
Assuming that bidders are ex ante identical may constitute a limitation, and in some cases one needs to relax this assumption. Asymmetric auction models, however, lead to systems of differential equations without a closed form solution. Hence, the direct approach becomes extremely difficult to implement. Nevertheless, using our indirect procedure, parameters can be structurally estimated without solving for the equilibrium strategy or its inverse.
\subsubsection{The Asymmetric IPV Model}
Following the exposition in \cite{PerrigneVuong2008} we assume that asymmetry is ex ante known to all bidders. Let $F_1(\cdot),\ldots,F_I(\cdot)$ be the private value distributions of the $I$ bidders whose identities are observed and let $G_{01}(\cdot),\ldots,G_{0I}(\cdot)$ be the corresponding bid distributions; see \cite{FlambardPerrigne2006}. We can then express the system of differential equations as
\begin{eqnarray*}
V_p=B_p+\frac{1}{\sum_{p'\neq p}\frac{g_{0p'}(B_p)}{G_{0p'}(B_p)}},\quad p,p'=1,\ldots, I,
\end{eqnarray*}
which lead, naturally to a two-step procedure, similar as before.
\subsubsection{The Asymmetric APV Model}
For simplicity we consider only two types of bidders. That is, the model assumes that the $I$-dimensional vector $(V_{11},\ldots,V_{1I_1},V_{21},\ldots,V_{2I_2})$ is distributed jointly as $F(\cdot)$ which is exchangeable in its first $I_1$ and last $I_2$ arguments. We can interpret this structure as follows. There is symmetry within each subgroup, and since $F(\cdot)$ is affiliated, there is general positive dependence among private values. \cite{PerrigneVuong1999} show that type specific equilibrium bidding strategies are characterized as the solution of the following system of differential equation,
\begin{eqnarray*}
V_{1p}&=&\xi_{1}(B_{1p},G_0)\equiv B_{1p}+\frac{G_{0B_1^*,B_0|B_1}(B_{1p},B_{1p}|B_{1p})}{\partial G_{0B_{1p}^*,B_0|B_1}(B_{1p},B_{1p}|B_{1p})/\partial (B^*_1,B_2)}, \hspace{6pt} p=1,2,\ldots,I_1\\
V_{2p}&=&\xi_{0}(B_{2p},G_0)\equiv B_{2p}+\frac{G_{0B_1,B_2^*|B_2}(B_{2p},B_{2p}|B_{2p})}{\partial G_{0B_1,B_2^*|B_2}(B_{2p},B_{2p}|B_{2p})/\partial (B_1,B_2^*)}, \hspace{6pt} p=1,2,\ldots,I_2,
\end{eqnarray*}
where $B_t^*=\max _{p\neq1,p\in I_t}B_{tp}$, $B_t=\max _{p\in I_t}B_{tp}$, for $t=1,2$. \cite{CampoPerrigneVuong2003} show that this identifies $F(\cdot,\ldots,\cdot)$, and use a nonparametric two-step procedure to estimate the model.
Similar to above, the two-step semiparametric procedure would involve using the above system of equations to recover the pseudo private values after obtaining nonparametric estimates for $G_{0B_1^*,B_2|B_1}(\cdot,\cdot|\cdot)$ and $G_{0B_1,B_2^*|B_2}(\cdot,\cdot|\cdot)$, and then estimating the parameters of $\theta$ through a set of moment conditions.
As \cite{CampoPerrigneVuong2003} have shown, the choice of bandwidths for asymmetric APV is similar to bandwidths for symmetric case, like \cite{LiPerrigneVuong2002}. Which means we can follow the same steps as in symmetric APV to choose our bandwidths.
\subsection{Unobserved Heterogeneity}
In some auctions it is possible that even after conditioning on auction covariates $Z_{\ell}$, the bids are still correlated.
Other than affiliation, such correlation could be a result of an auction characteristic $U_{\ell}\in\mathbb{R}_{++}$ that is missing in the data, but is observed by the bidders. Such auctions are known as auctions with unobserved heterogeneity.
In this subsection, we propose one possible way to adapt our semiparametric procedure to auctions with multiplicative unobserved heterogeneity as studied by \cite{Krasnokutskaya2011}.
As it will be clear later, in this case the rate of convergence and the asymptotic variance derived in our Proposition \ref{prop:2} will not be applicable.
But determining the exact asymptotic properties is beyond the scope of this paper.
We begin by introducing new and relevant notations and assumptions.
\medskip\noindent
{\bf Assumption A8:} {\em
\begin{itemize}
\item[(i)] Let $V_{p\ell}=\tilde{V}_{p\ell}\times U_{\ell}$ be bidder $p$'s value in $\ell^{th}$ auction such that $U_{\ell}\perp \tilde{V}_{p\ell}$ and $U_{\ell}\perp Z_{\ell}$.
\item[(ii)] Given $Z$ the random variables $V_{p\ell},\tilde{V}_{p\ell}$ are independently and identically distributed as $F(\cdot|Z)$ and $\tilde{F}(\cdot|Z)$, respectively.
\item[(iiii)] $U_{\ell}$ is independently and identically distributed as $F_{U}(\cdot)$ across auctions with $E(\ln (U))=0$.
\end{itemize}
}
\medskip\noindent
In summary, $U_{\ell}$ is independent across auctions, and in every auction with covariates $Z_{\ell}$ each bidder draws his/her ``true value'' $\tilde{V}_{p\ell}$ from $\tilde{F}(\cdot|Z_{\ell})$ and bids according to $V_{p\ell}$.
Let $s(\cdot|Z_{\ell},U_{\ell})$ denote the bidding strategy when the observed and unobserved covariates are, $Z_{\ell}$ and $U_{\ell}$, respectively, and let $\tilde{s}(\cdot|Z_{\ell})=s(\cdot|Z_{\ell},U_{\ell}=1)$ be the bidding strategy when the $U_{\ell}=1$, i.e., without unobserved heterogeneity.
\cite{Krasnokutskaya2011} shows that under Assumption A8: $s(V_{p\ell}|Z_{\ell},U_{\ell})=\tilde{s}(\tilde{V}_{p\ell}|Z_{\ell})\times U_{\ell}$ so that the bids satisfy $B_{p\ell}=\tilde{B}_{p\ell}\times U_{\ell}$, where $\tilde{B}_{p\ell}$ is the bid by bidder $p$ in auction $\ell$ when $U_{\ell}=1$.
If we use $G_{0}(\cdot|Z,U)$ and $\tilde{G}_{0}(\cdot|Z)$ to denote the conditional distribution of $B$ given $(Z,U)$ and the conditional distribution of $\tilde{B}$ given $Z$, respectively, then $B_{p\ell}=\tilde{B}_{p\ell}\times U_{\ell}$ implies
\begin{eqnarray*}
\tilde{G}_{0}(b|Z)&=&\Pr(\tilde{B}\leq b|Z)=\Pr(B\leq b\times U|Z,U)= {G}_{0}(b\times U|Z,U);\\
\tilde{g}_{0}(b|Z)&=&{g}_{0}(b\times U|Z,U)\times U.
\end{eqnarray*}
To simplify notation we suppress $Z$, and everything is to be understood as conditional on $Z$, unless stated otherwise.
The first step is to identify $F_{U}(\cdot)$. For auction $\ell=1,\ldots, L$ select any two bids and call them $(B_{1\ell},B_{2\ell})$, and let ${\mathcal F}_{(\ln B_{1},\ln B_{2})}(\cdot,\cdot)$ be the joint characteristic function of $(\ln B_{1}, \ln B_{2})$.
Then under a normalization $E[\ln \tilde{B}_{1}]=0$, \cite{Krasnokutskaya2011} shows that we can use \cite{Kotlarski1966} to identify the characteristic function of $ \ln U$ as
$$
{\mathcal F}_{\ln U}(t)=\exp\left(\int_{0}^{t}\frac{\partial {\mathcal F}_{(\ln B_{1},\ln B_{2})}(0,c)/\partial \ln B_{1}}{{\mathcal F}_{(\ln B_{1},\ln B_{2})}(0,c)}dc\right),
$$
which identifies $F_{\ln U}(\cdot)$ as the Fourier inverse of ${\mathcal F}_{\ln U}(t)$.
So for the remainder we treat $F_{U}(\cdot)$ as known.
The second step consists on estimating $\tilde{G}_{0}(\cdot)$ and $\tilde{g}_{0}(\cdot)$ of $\tilde{B}$ from:
$$
\ln B_{p\ell}=\ln \tilde{B}_{p\ell}+ \ln U_{\ell}, \quad p=1,\ldots, I_{\ell}, \ell=1,\ldots, L.
$$
Since $\ln U_{\ell}$ is unobserved, the estimators defined in (\ref{eq:Ghat}) and (\ref{eq:ghat}) are infeasible.
We can, however, replace the unobserved $(\ln \tilde{B}_{j}-b)^{\rho}K_{h}(\ln \tilde{B}_{j}- b)$ in (\ref{eq:Ghat}) and (\ref{eq:ghat}) with $(\ln {B}_{j}-b)^{\rho}\tilde{K}_{h}(\ln {B}_{j}- b)$, \cite{FanTruong1993, DelaigleFanCarroll2009}, where $\tilde{K}_{h}(b)=h^{-1}\tilde{K}(b/h)$, and satisfies
$$
{\rm E}\left\{(\ln {B}_{j}-b)^{\rho}\tilde{K}_{h}(\ln {B}_{j}- b)|\ln \tilde{B}_{j}\right\}=(\ln \tilde{B}_{j}-b)^{\rho}K_{h}(\ln \tilde{B}- b),\quad \rho=0,1.
$$
\cite{DelaigleFanCarroll2009} propose using the Fourier transformation of the above equation to determine $\tilde{K}_{h}(\cdot)$, and show that the presence of $U_{\ell}$ only affect the
variance of the estimator.\footnote{ The final result also depends on the smoothness of the density of the unobserved heterogeneity, see \cite{DelaigleHallMeister2008,DelaigleFanCarroll2009}.}
This gives us the estimates of the distribution and density of $\tilde{B}$.
Even though we cannot recover $\ln \tilde{B}_{p\ell}$, from $\ln B_{p\ell}$, we simulate the former from ${\tilde{G}}_{0}(\cdot|Z_{\ell})$ and determine
$$
\tilde{V}_{p\ell}^{\iota}=\tilde{B}_{p\ell}^{\iota}+\frac{1}{I_{\ell}-1}\frac{\tilde{G}_{0}(\tilde{B}_{p\ell}^{\iota}|Z_{\ell})}{\tilde{g}_{0}(\tilde{B}_{p\ell}^{\iota}|Z_{\ell})},\quad p=1,\ldots, I_{\ell}, \ell=1,\ldots, L, \iota=1,\ldots, {\mathbf S},
$$
where ${\bf S}$ is large.
Then in the third-step we estimate $\hat{\theta}$ using the appropriate empirical moment conditions:
$$
\frac{1}{{\bf S}}\sum_{\iota=1}^{{\bf S}} \frac{1}{L} \sum_{\ell=1}^{L} \frac{1}{I_{\ell}} \sum_{p=1}^{I_{\ell}} m(\tilde{V}_{p\ell}^{\iota}, Z_{\ell};\theta)\approx0.
$$
We conjecture that the estimator is consistent and asymptotically normal.
Since the rate of convergence of the estimator in the first step depends heavily on the smoothness of $f_{U}(\cdot)$ -- the smoother the density, the slower the convergence \cite{Fan1991} -- and because the error in steps 1 and 2 affect the asymptotic variance of $\hat{\theta}$, Proposition \ref{prop:2} does not apply here.
Full characterization of the asymptotic properties of $\hat{\theta}$ needs careful consideration and is left for future research.
\section{Conclusions\label{section:conclusion}}
\indent
In this paper we develop an indirect procedure to estimate first-price sealed-bid auction models, contributing in this way to the structural analysis of auction data that has been developed in the last fifteen years. Following GPV (2000) our procedure is in two steps. The difference with GPV (2000) is that our second step is implemented using a GMM procedure so that our resulting model is semiparametric. We show that our semiparametric estimator converges uniformly at the parametric $\sqrt{L}$ rate while the nonparametric estimator in GPV (2000) was shown to converge at the best possible rate according to the minimax theory which is slower than the parametric rate. Moreover, our procedure is not subject to the so-called curse of dimensionality or in other words the convergence rate is independent of the dimension of the exogenous variables. We establish consistency and asymptotic normality of our estimator.
Given the nature of our procedure it is not necessary to solve explicitly for the equilibrium strategy or its inverse. This is a valuable advantage with respect to direct methods specially when estimating models that lead to intractable first-order conditions, such as asymmetric auction models. More generally, our method extends to models which have been estimated using a nonparametric indirect procedure. In this respect, we briefly outline how this can be done in models with a binding reserve price (announced or random), affiliated private value models and asymmetric models.
Finally, we conducted a set of Monte Carlo simulations. The main purpose for this was to asses the performance of our estimator in finite samples relative to the nonparametric estimator proposed by GPV (2000). Our semiparametric estimator does a good job in matching the true density. When comparing with the nonparametric GPV (2000) estimator, we can see that the estimator developed in this paper is not subject to boundary effects. Moreover, using our estimator generates optimal revenue that is closer to the revenue, if we had used the true density, than using GPV (2000) estimator.
Since \cite{Krasnokutskaya2011} unobserved auction heterogeneity has become important in empirical auction -- ignoring it can lead to serious misspecification error. Moreover, it is known that the nonparmetric estimation is precarious, more so that the auctions without unobserved heterogeneity. Although we touched on this subject in this paper, we believe that determining asymptotic and efficiency properties of a semiparametric estimator is important.
We hope our paper provides the necessary impetus and motivation for someone to explore this problem.
\newpage
\small
\setcounter{equation}{0}
\renewcommand{\theequation}{A.\arabic{equation}}
\begin{center}
{\bf Appendix \\
Proofs of Asymptotic Properties}
\end{center}
\bigskip
This Appendix gives the proofs of our asymptotic results
(Propositions \ref{prop:1} and \ref{prop:2}).
First, we present two important results.
\medskip
\noindent {\bf{Results:}} Under A4 we have,
\begin{itemize}
\item[(i)]
$\displaystyle{\sup_{(b,x,I)}}\left\vert \hat g(b|x,I)-g_0(b|x,I)\right\vert=O_{as}\left(h_{1g}^R+h_{2g}^R+ \sqrt{\frac{\log L}{L h_{1g}h_{2g}}}\quad\right)$
\item[(ii)]
$\displaystyle{\sup_{(b,x,I)}}\left\vert \hat G(b|x,I)-G_0(b|x,I)\right\vert=O_{as}\left(h_{G}^{R+1}+ \sqrt{\frac{\log L}{L h_{G}}}\quad\right)$
\medskip
\noindent For a proof of the above results we refer the reader to \cite{KorostelevTsybakov1993}.
We observe that the above results imply that $\displaystyle{\sup_{p\ell}}\vert\hat V_{p\ell}-V_{p\ell}\vert=o_{as}(1).$
\end{itemize}
{\bf{Proposition \ref{prop:1}}}
\begin{proof}
It suffices to show that
$\displaystyle{\sup_{\theta \in \Theta}\parallel S_L(\theta)-\hat
S_L(\theta)\parallel}=o_{as}(1)$. From the triangle inequality, A5-(iv) it follows that
\begin{eqnarray}
\sup_{\theta \in \Theta}\Vert S_L(\theta)-\hat S_L(\theta)\Vert&=&\sup_{\theta \in \Theta}\Bigg\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} m(V_{p\ell},Z_\ell;\theta)-\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} m(\hat V_{p\ell},Z_\ell;\theta)\Bigg\Vert\nonumber\\
&=&\sup_{\theta \in \Theta}\Bigg\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} [m(V_{p\ell},Z_\ell;\theta)-m(\hat V_{p\ell},Z_\ell;\theta)]\Bigg\Vert\nonumber\\
&\leq & \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \sup_{\theta \in \Theta}\Big\Vert m(V_{p\ell},Z_\ell;\theta)-m(\hat V_{p\ell},Z_\ell;\theta)\Big\Vert \nonumber\\
&\leq & \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_1(Z_\ell)\vert\hat{V}_{p\ell}-V_{p\ell}\vert \notag\\
&=&\{{\rm E}[K_1(Z)]+o_{as}(1)\} \sup_{p\ell}\vert\hat V_{p\ell}-V_{p\ell}\vert \nonumber\\
&=&o_{as}(1)
\end{eqnarray}
Where we use the fact that $\hat V_{p\ell}$ is a consistent estimator of $V_{p\ell}$, i.e. we make use of the 2 results stated at the beginning of this Appendix.
Therefore, the desired result follows.
\end{proof}
\newpage
{\bf {Proposition \ref{prop:2}}}
\begin{proof}
Recall that $\hat{\theta}$ is the feasible estimator while $\tilde{\theta}$ is the infeasible estimator.
We want to show:
\begin{eqnarray*}
\sqrt{L}(\hat\theta-\theta_0)&=&\sqrt{L}(\tilde{\theta}-\theta_0)-\frac{L(L-1)}{L^2}\frac{2}{\sqrt{L}}
\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I\Bigg\{\sum_{I}\frac{1}{I(I-1)} N(Y_{p\ell},I) f_m^{-1}(X_\ell,I) \\
&&g_0(Y_{p\ell},I)+{\rm{E}}\Bigg[\sum_{I}\frac{1}{I(I-1)} N(Y_{p\ell},I) f_m^{-1}(X_\ell,I) g_0(Y_{p\ell},I)\Bigg]\Bigg\}+o_p(1)\\
&=&-\frac{1}{\sqrt{L}}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I\Bigg\{(C^T\Omega C)^{-1}C \Omega m(V_{p\ell},X_\ell,I;\theta_0)
+2\frac{L(L-1)}{L^2}\\
&&\sum_{I}\frac{1}{I(I-1)} N(Y_{p\ell},I) f_m^{-1}(X_\ell,I) g_0(Y_{p\ell},I)\\
&&-{\rm{E}}\Bigg[\sum_{I}\frac{1}{I(I-1)} N(Y_{p\ell},I) f_m^{-1}(X_\ell,I) g_0(Y_{p\ell},I)\Bigg]\Bigg\}+o_p(1),
\end{eqnarray*}
where $C={\rm{E}}\left[{\partial m(V,X,I;\theta_0)}/{\partial \theta}
\right], Y_{p\ell}\equiv (B_{p\ell}, X_{\ell}), \Omega$ is the p.d weighting matrix and
$$
N(Y_{p\ell},I)=[m_1(V_{p\ell},X_\ell,I;\theta_0)/ g_0(B_{p\ell}|X_\ell,I)^2]
G_0(B_{p\ell}|X_\ell,I).$$
The terms inside $\{\}$ in the second equality is the influence function.\footnote{The idea behind that is the following observation. Suppose after a Taylor expansion, we have:
\begin{eqnarray*}
\left\{\begin{array}{c}\sqrt{n}(\hat{\theta}-\theta_{0})=\frac{1}{\sqrt{n}}\sum_{i=1}^{n}\psi(z_{i})+o_{p}(1)\\E(\psi(z))=0, var(\psi(z))<\infty\end{array}\right.
\end{eqnarray*}
where $\psi(\cdot)$ is the influence function, then the asymptotic variance is $V=var(\psi(z))$. An alternative method would have been to follow \cite{Newey1994} and use the path derivative approach.}
Once we have shown this asymptotic linear representation the result follows from the Central Limit Theorem.
From the FOCs that characterize $\tilde{\theta}$ and $\hat\theta$
respectively, we have
\begin{eqnarray}
\frac{1}{2}\frac{\partial Q_L}{\partial
\theta}(\tilde{\theta})&=&{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega S_L(\tilde{\theta})=0\label{eq:a2}\\
\frac{1}{2}\frac{\partial \hat Q_L}{\partial
\theta}(\hat\theta)&=&{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega \hat S_L(\hat\theta)=0.\label{eq:a3}
\end{eqnarray}
We can use a Taylor expansion around $\theta_0$ to obtain
\begin{eqnarray}
S_L(\tilde{\theta})&=&S_L(\theta_0)+{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})(\tilde{\theta}-\theta_0)\label{eq:a4}\\
\hat S_L(\hat\theta)&=&\hat
S_L(\theta_0)+{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\overline{\theta}^*)(\hat\theta-\theta_0),\label{eq:a5}
\end{eqnarray}
where $\overline{\theta}$ and $ \overline{\theta}^*$ are
vectors between $\tilde{\theta}$ and $\theta_0$, and $\hat\theta$
and $\theta_0$, respectively.
Thus using (\ref{eq:a4}) in (\ref{eq:a2}) we get
\begin{eqnarray*}
{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega\left[S_L(\theta_0)+{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})(\tilde{\theta}-\theta_0)\right]
={\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega S_L(\theta_0)+{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})(\tilde{\theta}-\theta_0)=0.
\end{eqnarray*}
Therefore, we have
\[\sqrt{L}(\tilde{\theta}-\theta_0)=-\left[{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})\right]^{-1}{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega\sqrt{L}S_L(\theta_0)
=-\tilde{A}^{-1}\tilde{B}\sqrt{L}S_L(\theta_0).\label{eq:infeasible}\tag{**}\]
Similarly using (\ref{eq:a5}) in (\ref{eq:a3}) yields
\begin{eqnarray*}
\sqrt{L}(\hat\theta-\theta_0)&=&-\left[{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\overline{\theta}^*)\right]^{-1}
{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega \sqrt{L} \hat S_L(\theta_0)
=-\hat A^{-1} \hat B\sqrt{L} \hat S_L(\theta_0).
\end{eqnarray*}
Next, we show: (i) $\tilde{B}-\hat B=o_{as}(1)$ ( Step \ref{step1});
(ii) $\tilde{A}-\hat A=o_{as}(1)$, which together with A6-(iii) imply $\tilde{A}^{-1}-\hat A^{-1}=o_{as}(1)$(Step \ref{step2});
and finally (iii) $\sqrt{L}[S_L(\theta_0)-\hat S_L(\theta_0)]=O_p(1)$ (Step \ref{step3}).
\setcounter{equation}{5}
\setcounter{section}{0}
\renewcommand{\theequation}{A.\arabic{equation}}
\section{Step 1\label{step1}} We prove $\tilde{B}-\hat B=o_{as}(1)$. The term
$\tilde{B}-\hat B$ can be written as
\begin{eqnarray*}
\tilde{B}-\hat
B&=&{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega-{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega=\left({\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})-{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\right)\Omega\\
&=&\left(\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \left(m_3^T(V_{p\ell},Z_\ell,\tilde{\theta})-m_3^T( \hat
V_{p\ell},Z_\ell;\hat\theta)\right)\right)\Omega.
\end{eqnarray*}
It suffices to show that the norm of the term between brackets is
$o_{as}(1)$ since $\Omega$ is a positive definite matrix. Namely
\begin{eqnarray}
\lefteqn{\left\Vert \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} m_3^T(V_{p\ell},Z_\ell,\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\hat\theta))\right\Vert}\nonumber\\
&=&\Bigg\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Big[(m_3^T(V_{p\ell},Z_\ell,\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\tilde\theta))+m_3^T(\hat V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\hat\theta))\Big]\Bigg\Vert\nonumber\\
&\leq& \left \Vert \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}\left[(m_3^T(V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat
V_{p\ell},Z_\ell;\tilde\theta))\right]\right\Vert+\left\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} m_3^T(\hat V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\hat\theta))\right\Vert\nonumber\\
&=&C+D,\label{eq:a6}
\end{eqnarray}
\noindent where the last line follows from the triangle inequality.
The term $C$ in (\ref{eq:a6}) is
\begin{eqnarray*}
C&=&\left\Vert \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}\left[(m_3^T(V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat
V_{p\ell},Z_\ell;\tilde\theta))\right]\right\Vert\leq\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}\left\Vert\left[m_3^T(V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat
V_{p\ell},Z_\ell;\tilde{\theta})\right]\right\Vert\\
&\leq& \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_3(Z_\ell) \vert V_{p\ell}-\hat V_{p\ell}\vert\leq\{{\rm{E}}[K_3(Z)]+o_{as}(1)\}\sup_{p\ell}\vert \hat V_{p\ell}-V_{p\ell}\vert=o_{as}(1),
\end{eqnarray*}
\noindent where we use assumption A6-(i) and the fact that $\hat V_{p\ell}$ is uniformly consistent-- these results are stated at the beginning of this Appendix.
We consider now the term $D$ in (\ref{eq:a6}):
\begin{eqnarray*}
D&=&\left\Vert \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \left[m_3^T(\hat V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\hat{\theta})\right]\right\Vert\leq \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \left\Vert m_3^T(\hat V_{p\ell},Z_\ell;\tilde{\theta})-m_3^T(\hat V_{p\ell},Z_\ell;\hat{\theta})\right\Vert\\
&\leq &\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_4(Z_\ell) \Vert \tilde\theta-\hat\theta \Vert=\{{\rm{E}}[K_4(Z)]+o_{as}(1)\} \times o_{as}(1),
\end{eqnarray*}
\noindent where we have used A6-(ii) and the fact that
$\tilde{\theta}$ and $\hat\theta$ are consistent estimators for
$\theta_0$.
\section{Step 2\label{step2}} We prove $\tilde{A}-\hat A=o_{as}(1)$. The term
$\tilde{A}-\hat A$ is
\begin{eqnarray}
\tilde{A}-\hat
A&=&\left({\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})\right)-
\left({\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega {\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\overline{\theta}^*) \right)\nonumber\\
&=&\left[{\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde{\theta})\Omega \left({\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\tilde{\theta})+o_{as}(1) \right)\right]
-\left[{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega \left({\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\hat\theta)+o_{as}(1)\right)\right]\nonumber\\
&=&\left({\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde\theta)\Omega {\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\tilde\theta)\right)-\left({\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\Omega{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\hat\theta)\right)+o_{as}(1)\nonumber\\
&=&\left[\left({\displaystyle{\frac{\partial S_L^{T}}{\partial \theta}}}(\tilde\theta)-{\displaystyle{\frac{\partial \hat S_L^T}{\partial \theta}}}(\hat\theta)\right)\Omega\right]\left({\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\tilde\theta)+{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\hat\theta)\right).\label{eq:a7}
\end{eqnarray}
\noindent where the second equality comes from the following
\begin{eqnarray*}
\left\Vert {\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\overline{\theta})-{\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\tilde\theta)\right\Vert&\leq&\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}
\Vert m_3(V_{p\ell},Z_\ell;\overline{\theta})-m_3(V_{p\ell},Z_\ell;\tilde{\theta})\Vert\\
&\leq& \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_4(Z_\ell) \Vert \overline{\theta}-\tilde\theta \Vert
=\{{\rm{E}}[K_4(Z)]+o_{as}(1)\} o_{as}(1)=o_{as}(1),
\end{eqnarray*}
where we use A6-(ii), the fact that $\tilde\theta\leq\overline{\theta}\leq \theta_0$ and that
$\tilde\theta \stackrel{a.s}{\longrightarrow}\theta_0$. Similarly we can show that
\begin{eqnarray*}
{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\overline{\theta}^*)={\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\hat\theta)+o_{as}(1),
\end{eqnarray*}
since $\hat\theta\leq\overline{\theta}^*\leq \theta_0$ and $\hat\theta \stackrel{a.s}{\longrightarrow}\theta_0$.
Now, for the last line in (\ref{eq:a7}) we observe that by Step 1, the first factor in (\ref{eq:a7}) is $o_{as}(1)$ and
the second factor can be expressed as follows
\begin{eqnarray*}
\lefteqn{\left\Vert\left({\displaystyle{\frac{\partial S_L}{\partial \theta^{T}}}}(\tilde{\theta})+{\displaystyle{\frac{\partial \hat S_L}{\partial \theta^{T}}}}(\hat\theta)\right)\right\Vert=\left\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}
[m_3(V_{p\ell},Z_\ell;\tilde\theta)+ m_3(\hat V_{p\ell},Z_\ell;\hat\theta)]\right \Vert}\\
&\leq& \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert [m_3(V_{p\ell},Z_\ell;\tilde\theta)+ m_3(\hat V_{p\ell},Z_\ell;\hat\theta)] \Vert\leq \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(V_{p\ell},Z_\ell;\tilde\theta) \Vert + \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(\hat V_{p\ell},Z_\ell;\hat\theta)\Vert\\
&\leq& \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \sup_{\theta\in\Theta} \Vert m_3(V_{p\ell},Z_\ell;\theta) \Vert + \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(\hat V_{p\ell},Z_\ell;\hat\theta)-
m_3(\hat V_{p\ell},Z_\ell;\theta_0)+ m_3(\hat V_{p\ell},Z_\ell;\theta_0)\Vert \\
&\leq& \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_5(V_{p\ell},Z_\ell)+ \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(\hat V_{p\ell},Z_\ell;\hat\theta)- m_3(\hat V_{p\ell},Z_\ell;\theta_0)\Vert\\
&&+\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(\hat V_{p\ell},Z_\ell;\theta_0)-m_3(V_{p\ell},Z_\ell;\theta_0) +m_3(V_{p\ell},Z_\ell;\theta_0)\Vert\\
&\leq& \{{\rm{E}}[K_5(V,Z)]+o_{as}(1)\}+\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_4(Z_\ell) \Vert\hat\theta-\theta_0 \Vert\\
&&+\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(\hat V_{p\ell},Z_\ell;\theta_0)-m_3(V_{p\ell},Z_\ell;\theta_0) \Vert+\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Vert m_3(V_{p\ell},Z_\ell;\theta_0) \Vert\\
&\leq&\{{\rm{E}}[K_5(V,Z)]+o_{as}(1)\}+\{{\rm{E}}[K_4(Z)]+o_{as}(1)\}o_{as}(1)+ \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_3(Z_\ell) \vert\hat V_{p\ell}-V_{p\ell} \vert\\
&&+\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \sup_{\theta\in\Theta} \Vert m_3(V_{p\ell},Z_\ell;\theta) \Vert\\
&\leq& \{{\rm{E}}[K_5(V,Z)]+o_{as}(1)\}+\{{\rm{E}}[K_4(Z)]+o_{as}(1)\}o_{as}(1)
+ \{{\rm{E}}[K_3(Z)]+o_{as}(1)\}\sup_{p\ell} \vert \hat V_{p\ell}-V_{p\ell}\vert\\
&&+\{{\rm{E}}[K_5(V,Z)]+o_{as}(1)\} =2\{{\rm{E}}[K_5(V,Z)]+o_{as}(1)\} <\infty
\end{eqnarray*}
where we use assumption A6-(ii),(iv),(v) and the two results stated at the beginning of this Appendix.
Therefore the second factor in the last line of (\ref{eq:a7}) converges to a finite limit, and
since the first factor is $o_{as}(1)$ the desired result follows.
\section{Step 3\label{step3}} The final step is to prove $\sqrt{L}(S_L(\theta_0)-\hat
S_L(\theta_0))=O_p(1)$.
Since this step is the longest and the most tedious, to facilitate reading we divide this step further into two sub-steps: Step 3.1 and Step 3.2, and before we provide the formal proof we give a detailed description of all the steps involved.
Let $B=\sqrt{L}[S_L(\theta_0)-\hat S_L(\theta_0)]= B_1+B_2$.
In Step 3.1 we show that $B_1=O_p(1)+o_{as}(1)$ and in Step 3.2 we show $B_2=o_{as}(1)$, see Equation (\ref{B_2}).
Of these two, Step 3.1 is more involved, but we can break down the proof into following steps:
\setcounter{equation}{0}
\renewcommand{\theequation}{B-\arabic{equation}}
\begin{eqnarray*}
B_1&\leq& B_{11}+B_{12} \quad(\textrm{See Equation (\ref{B11+B12})}) \\
&=&B_{111}+B_{112}+B_{12}\quad(\textrm{See Equation (\ref{B11})}) \\
&\leq& C D +B_{112}+B_{12} \quad (\because C < \infty, D=o_p(1)) \\
&\leq& o(1)+B_{1121}+B_{1122}+B_{12} \quad (\textrm{See Equation (\ref{B112}) and $B_{1121}=o_p(1)$}) \\
&\leq& o(1)+o_p(1)+A+B+B_{12} \quad (\textrm{See Equation (\ref{B1122}) and $B_{1122}=A+B$})\\
&\leq& o(1)+o_p(1)+A_1-A_2+B+B_{12}\quad\!\!(\textrm{See Equation (\ref{A1_A2}), $A_{1} \leq A_{11}+ A_{12}$ and $ A_{11}$ and $A_{12}$ are $o_{as}(1/\sqrt{L})$}) \\
&\leq& o(1)+o_p(1)+o_{as}(1/\sqrt{L})-A_2+B+B_{12}\quad\!\!(\because\textrm{$A_{2} \leq A_{21} + A_{22}$ and $ A_{21}$ and $A_{22}$ are $o_{as}(1/\sqrt{L})$})\\
&\leq& o(1)+o_p(1)+o_{as}(1/\sqrt{L})-o_{as}(1/\sqrt{L})+B+B_{12} \quad (\because B_2=B_{121}\times B_{122}) \\
&\leq& o(1)+o_p(1)+o_{as}(1/\sqrt{L})-o_{as}(1/\sqrt{L})+B+B_{121}+B_{122}\quad\!\!(\because B_{121} < \infty, B_{122}=o(1) \Rightarrow B_{12}=o(1)) \\
&\leq& o(1)+o_p(1)+o_{as}(1/\sqrt{L})-o_{as}(1/\sqrt{L})+B+o(1)\quad(\because B=O_p(1)) \\
&=& o(1)+o_p(1)+o_{as}(1/\sqrt{L})-o_{as}(1/\sqrt{L})+O_p(1)+o(1)\quad (\textrm{See equation (\ref{B1122})}).
\end{eqnarray*}
We formalize the proof below.
First we prove that the term $\sqrt{L}(S_L(\theta_0)-\hat S_L(\theta_0))$ is
\begin{eqnarray*}
B=\sqrt{L}(S_L(\theta_0)-\hat S_L(\theta_0))&=&\sqrt{L}\Bigg(\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \mo-\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}\mhato \Bigg)\\
&=&\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \Big[\mo - \mhato \Big]=O_p(1)+o_{as}(1).
\end{eqnarray*}
The above expression can be rewritten as
\begin{eqnarray}
B&=&-\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}\left[\moneo (\hat{V}_{p\ell}- V_{p\ell}) \right]+\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} \left[ \moneo-\moneob\right](\hat{V}_{p\ell}- V_{p\ell})\nonumber\\
&=&B_1+B_2,
\label{B_step3}
\end{eqnarray}
\noindent where the second equality comes from a Taylor expansion of
order one and the following
\begin{eqnarray*}
\mo-\mhato&=&\moneob ({V}_{p\ell}-\hat V_{p\ell})\\
&=&\moneo (\hat{V}_{p\ell}-V_{p\ell})+\moneob({V}_{p\ell}-\hat V_{p\ell})-\moneo(\hat{V}_{p\ell}-V_{p\ell})\\
&=&-\moneo(\hat{V}_{p\ell}-V_{p\ell})+[\moneo-\moneob](\hat{V}_{p\ell}-V_{p\ell}).
\end{eqnarray*}
\subsection*{Step 3.1} We consider $B_1$ in (\ref{B_step3}) and moreover we observe that for each $I$ we can write
\begin{eqnarray}
\Vert B_1\Vert&=&\left\Vert\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0) (\hat V_{p\ell}-V_{p\ell})\right\Vert\nonumber\\
&=&\Bigg \Vert\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0) \frac{1}{I-1}\left[\frac{\Gtildei}{\gtildei}-\frac{\Gi}{\gi}\right]\Bigg\Vert\nonumber\\
&=&\Bigg \Vert\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-I)}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0)\Bigg\{\frac{\Gtildei}{\gi} -\frac{\Gi}{g^2_0(B_{p\ell|X_\ell,I})}\nonumber\\
&&\gtildei+\frac{\Gi}{\gi}\frac{1}{\gtildei \gi} \Big[\gtildei-\gi\Big]^2\nonumber\\
&&-\frac{1}{\gtildei \gi}\left[\Gtildei-\Gi\right]\left[\gtildei-\gi\right]\Bigg\}\Bigg\Vert\nonumber\\
&\leq& \Bigg\Vert \sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-I)}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0) \Bigg[\frac{\Gtildei}{\gtildei}-\frac{\Gi} {g^2_0(B_{p\ell} | X_\ell,I)}\nonumber\\
&&\gtildei\Bigg]\Bigg\Vert+\Bigg\Vert \sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-I)}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0)\nonumber \\
&&\Bigg(\frac{\Gi}{\gi}\frac{1}{\gtildei\gi}\left[\gtildei-\gi\right]^2\nonumber\\
&&-\frac{1}{\gtildei \gi}\left[\Gtildei-\Gi\right]\nonumber\\
&&\Big[\gtildei-\gi\Big]\Bigg)\Bigg\Vert= \Vert B_{11}\Vert+\Vert B_{12}\Vert
\label{B11+B12}
\end{eqnarray}
\noindent where the third line uses the following identity:
$$\displaystyle{\frac{\tilde{a}}{\tilde{b}}-\frac{a}{b}=
\frac{\tilde{a}-\frac{a}{b}\tilde{b}}{b}
+\frac{a}{b}\frac{1}{\tilde{b}b}[\tilde{b}-b]^2-\frac{1}{\tilde{b}b}[\tilde{a}-a][\tilde{b}-b]}.$$
\medskip
The term $B_{11}$ can be written as
\begin{eqnarray}
B_{11}&=&\sqrt{L}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-I)}\sum_{p=1}^I} m_1(V_{p\ell},X_\ell,I;\theta_0)\left[\frac{\Gtildei-\frac{\Gi}{\gi}\gtildei}{\gi}\right]\nonumber\\
&=&\sqrt{L}(R_L+\frac{L(L-1)}{L^2}U_L) =B_{111}+B_{112},
\label{B11}
\end{eqnarray}
where
\begin{eqnarray*}
R_L&=&\displaystyle{\frac{1}{L^2}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\Bigg[\omega_{I,R+1,j}^G K_{G,h_{G}}(0) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq B_{p\ell})\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(0) K_{2g,h_{g}}(0)\Bigg],\\
U_L&=&\tslminus \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\\
&&\Bigg[\omega_{I,R+1,j}^G K_{G,h_{G}}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(X_{j}-X_{\ell}) K_{2g,h_{g}}(B_{qj}-B_{p\ell})\Bigg].
\end{eqnarray*}
To see how to obtain the last line in (\ref{B11}), we observe that the
term within brackets in the first line of (\ref{B11}) can be expressed
as
\begin{eqnarray}
&&\frac{\Gtildei-\frac{\Gi}{\gi}\gtildei}{\gi}=\frac{1}{\gi}\left[\Gtildei-\frac{\Gi}{\gi}\gtildei \right]\nonumber\\
&=&\frac{1}{\gi}\nonumber\\
&&\Bigg[{\displaystyle{\frac{1}{Lh_{G}}\frac{L}{n_I}\sum_{\{j:I_j=I\}}^L\sum_{q=1}^{I}}e_1^T \left(\frac{X_{I,R+1}^T W_x^G X_{I,R+1}}{L_I}\right)^{-1} X_{R+1,j} K_G\left(\frac{X_j-X_\ell}{h_{G}}\right)\Unit(B_{qj}\leq B_{p\ell})}\nonumber\\
&&-\frac{\Gi}{\gi}{\displaystyle{\frac{1}{Lh_{g}^2}\frac{L}{n_I}\sum_{\{j:I_j=I\}}^L\sum_{q=1}^{I}} e_1^T \left(\frac{X_{I,R}^T W_x^g X_{I,R}}{n_I}\right)^{-1} X_{R,j} K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right)K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)}\Bigg]\nonumber\\
&=&\frac{1}{\gi}\Bigg[\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{j:I_j=I\}}^L\sum_{q=1}^{I}} \ikGqj-\frac{\Gi}{\gi}\nonumber\\
&&\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{j:I_j=I\}}^L\sum_{q=1}^{I}} \kgqj\Bigg]
\label{A.11}
\end{eqnarray}
\noindent where we have used the following notations:
\begin{eqnarray}
K_{G,h_{G}}(X_j-X_\ell)&=&\frac{1}{h_{G}}K_G\left(\frac{X_j-X_\ell}{h_G}\right), \label{A.12}\\
K_{1g,h_g}(X_j-X_\ell)&=&\frac{1}{h_g}K_{1g}\left(\frac{X_j-X_\ell}{h_g}\right),\\
K_{2g,h_g}(B_{qj}-B_{p\ell})&=&\frac{1}{h_g}K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_g}\right),\\
\omega^G_{I,R+1,j}&=&e_1^T \left(\frac{X_{I,R+1}^T W_x^G X_{I,R+1}}{n_I}\right)^{-1} X_{R+1,j},\\
\omega^g_{I,R,j}&=&e_1^T \left(\frac{X_{I,R}^T W_x^g X_{I,R}}{n_I}\right)^{-1} X_{R,j}.
\label{A.16}
\end{eqnarray}
Now using (\ref{A.11}) in the first line of (\ref{B11}), we get
\begin{eqnarray}
B_{11}&=&\sqrt{L}\Bigg(\ts\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\Bigg[\omega_{I,R+1,j}^G K_{G,h_G}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\nonumber\\
&& -\frac{\Gi}{\gi}\kgqj\Bigg]\Bigg).
\label{B11}
\end{eqnarray}
The term between parenthesis in (\ref{B11}) can be decomposed as follows:\\
1) Diagonal terms $(\ell=j, p=q)$
\begin{eqnarray}
R_L&=&\displaystyle{\frac{1}{L^2}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\Bigg[\omega_{I,R+1,j}^G K_{G,h_{G}}(0) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq B_{p\ell})\nonumber\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(0) K_{2g,h_{g}}(0)\Bigg],
\label{RL}
\end{eqnarray}
2) Off-diagonal terms $(\ell\neq j)$
\begin{eqnarray}
\frac{L(L-1)}{L^2}U_L&\!\!\!\!=\!\!\!\!&\tsd \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\nonumber\\
&&\Bigg[\omega_{I,R+1,j}^G K_{G,h_{G}}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\nonumber\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(X_j-X_\ell) K_{2g,h_{g}}(B_{qj}-B_{p\ell})\Bigg].
\label{UL}
\end{eqnarray}
From (\ref{RL}) and (\ref{UL}) we have the expression in the last line of
(\ref{B11}). It remains to show that $B_{11}=B_{111}+B_{112}=o_{as}(1)$.
We consider first $B_{111}=\sqrt{L}R_L$ in (\ref{B11}). Specifically,
\begin{eqnarray*}
B_{111}&=&\sqrt{L}\Vert R_L\Vert=\sqrt{L}\Bigg\Vert\Bigg(\displaystyle{\frac{1}{L^2}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\Bigg[\omega_{I,R+1,j}^G K_{G,h_{G}}(0)\\
&& -\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(0) K_{2g,h_{g}}(0)\Bigg]\Bigg)\Bigg\Vert\\
&=&\sqrt{L}\Bigg\Vert\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\Bigg[\omega_{I,R+1,j}^G \frac{ K_{G,h_{G}}(0)}{L}-\frac{\Gi}{\gi}\\
&& \omega_{I,R,j}^g \frac{ K_{1g,h_{g}}(0) K_{2g,h_{g}}(0) }{L}\Bigg]\Bigg\Vert\\
&\leq&\left(\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \Vert m_1(V_{p\ell},X_\ell,I;\theta_0) \Vert^2 \right)^{\frac{1}{2}}\\
&&\sqrt{L}\Bigg(\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{1}{\gi^2}\Bigg[\omega_{I,R+1,j}^G \frac{ K_{G,h_{G}}(0)}{L}- \frac{\Gi}{\gi}\\
&&\omega_{I,R,j}^g \frac{ K_{1g,h_{g}}(0) K_{2g,h_{g}}(0)}{L}\Bigg]^2\Bigg)^{\frac{1}{2}}\\
&=&C D,
\end{eqnarray*}
\noindent where the inequality comes from Cauchy-Schwartz. First we
show that $C^2<\infty$. Using A6-(vi), $0<(1/(I-1))<1$ for each $I \in {\cal I} $ and $L/n_I=L/(I L_I)<\infty$ we get
\begin{eqnarray*}
C^2&=&\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \Vert m_1(V_{p\ell},X_\ell,I;\theta_0) \Vert^2
\leq\frac{L}{n_I}\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I} \frac{1}{I(I-1)} \sup_{\theta \in \Theta}\Vert m_1(V_{p\ell},X_\ell,I;\theta)\Vert^2\\
&<&\displaystyle{\frac{1}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^I} K_7(V_{p\ell},X_\ell,I)^2={\rm E}[K_7(V,X,I)^2]+o_{as}(1)<\infty.
\end{eqnarray*}
It remains to consider the D term above. Namely,
\begin{eqnarray*}
D&\leq&\sqrt{L} \Bigg(\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}} \frac{1}{\gi^2}\Bigg[\omega_{I,R+1,j}^G \frac{K_G(0)}{Lh_{G}}-\frac{\Gi}{\gi}\\
&& \omega_{I,R,j}^g
\frac{K_{1g,h_{g}}(0) K_{2g,h_{g}}(0)}{Lh_{g}^2}\Bigg]^2\Bigg)^{\frac{1}{2}}\\
&=&\sqrt{L}\Bigg(\displaystyle{\frac{1}{L}\frac{L}{n_I}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I(I-1)}\sum_{p=1}^{I}}\frac{1}{\gi^2}\Bigg[O_p(1) O_p\left(\frac{1}{Lh_G}\right)-\frac{\Gi}{\gi} \\
&&O_p(1) O_p\left(\frac{1}{Lh_{g}^2}\right)\Bigg]^2
\Bigg)^{\frac{1}{2}}
<\sqrt{L}\kappa_1\left[O_p\left(\frac{1}{Lh_{G}}\right)- \kappa_2 O_p\left(\frac{1}{Lh_{g}^2}\right)\right]\\
&=&\kappa_1 \left[O_p\left(\frac{1}{\sqrt{L}h_{G}}\right)- \kappa_2 O_p\left(\frac{1}{\sqrt{L}h_{g}^2}\right) \right]=\kappa_1[o_p(1)-\kappa_2 o_p(1)]
=o_p(1),
\end{eqnarray*}
\noindent where after the first equality we use (\ref{A.12})- (\ref{A.16}). The second line follows from observing that
\begin{eqnarray*}
\omega_{I,R+1,j}^G&=&e_1^T \left(\frac{X_{I,R+1}^T W_x^G X_{I,R+1}}{n_I}\right)^{-1} X_{R+1,j}=e_1^T\left[\frac{1}{n_I h_G}\sum_{\iota=1}^{n_I} {\mathbf{x}}_\iota^T {\mathbf{x}}_\iota K_G\left(\frac{x_\iota-x_j}{h_g}\right) \right]^{-1} e_1=O_p(1),
\end{eqnarray*}
\noindent and similarly for $\omega_{I,R,j}^g$.
The third line uses the fact that densities are bounded away from
zero and $0<(1/I(I-1))<1$ for all $I$. The last line follows from Assumption A4.AN. Thus, $B_{111}=C D=o(1)$ as
desired.
Let $Y_{p\ell}=(B_{p\ell},X_\ell)$ and for each $I$ define
$r_L(Y_{p\ell},I)=E[\pl|(Y_{p\ell},I)]$, where
$p_L(\cdot,\cdot)$ is a symmetric function, and
$$\theta_L=E[r_L(Y_{p\ell},I)]=E[\pl]; \quad\!\!\hat U_L=\theta_L+\displaystyle{\frac{2}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^{I}} [r_L(Y_{p\ell},I)-\theta_L].$$
Next, we consider $B_{112}$ in (\ref{B11})
\begin{eqnarray}
B_{112}&=&\frac{L(L-1)}{L^2}\sqrt{L}U_L =
\frac{L(L-1)}{L^2}\sqrt{L}(U_L-\hat U_L)+\frac{L(L-1)}{L^2}\sqrt{L}\hat U_L =B_{1121}+B_{1122},
\label{B112}
\end{eqnarray}
where $U_L$ can be written as a U-statistic.
Namely,
\begin{eqnarray*}
U_L&=&\tslminus \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi} \\
&&\Bigg[\omega_{I,R+1,j}^G K_{G,h_G}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_g}(X_j-X_\ell)K_{2g,h_g}(B_{qj}-B_{p\ell})\Bigg]\\
&=&{\displaystyle{\frac{1}{L (L-1)}\sum_{\{\ell:I_\ell=I\}}^{L-1\Bigg\{\imfl \frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{\gi}\\
&&\Bigg[\omega_{I,R+1,j}^G K_{G,h_G}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_g}(X_j-X_\ell)K_{2g,h_g}(B_{qj}-B_{p\ell})\Bigg]\Bigg\}\\
&=&{\displaystyle{\frac{2}{L (L-1)}\sum_{\{\ell:I_\ell=I\}}^{L-1\Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0) K^{**}(B_{p\ell},B_{qj},X_\ell,X_j,I)}{2}\\
&&+\frac{m_1(V_{qj},X_j,I;\theta_0) K^{**}(B_{qj},B_{p\ell},X_j,X_\ell,I)}{2}\Bigg]\\
&=&\left(\begin{array}{c}L\\2\end{array}\right)^{-1}{\displaystyle{\sum_{\{\ell:I_\ell=I\}}^{L-1
p_L\left((B_{p\ell},X_\ell,I),(B_{qj},X_j,I)\right).
\end{eqnarray*}
Now we prove $B_{1121}=\sqrt{L}(U_L-\hat U_L)=o_p(1)$. By Lemma 3.1 in
\cite{PowellStockStoker1989} it is enough to show that ${\rm E}[\Vert
p_L((Y_{p\ell},I),(Y_{qj},I))\Vert^2]=o(L)$.
We will show that ${\rm E}[\Vert
p_L((Y_{p\ell},I),(Y_{qj},I))\Vert^2|I]=o(L)$, which implies the above condition.
{\footnotesize\begin{eqnarray}
\lefteqn{{\rm E}[\Vert \pl \Vert^2|I]=\int \Vert \pl \Vert^2 \gz \gzj \dypl \dypj} \nonumber\\
&=&\frac{1}{4}\int \Bigg\Vert \frac{L}{n_I(I-1)}\izmone \Bigg[\hGd \kG \nonumber\\
&&- \ratioGg \hgdone \kgz \Bigg]\nonumber\\
&&+\frac{L}{n_I(II-1)}\izjmone\Bigg [\hGd \kGj\nonumber\\
&&- \ratioGgj \hgdone \kgzj\Bigg] \Bigg \Vert^2 \nonumber\\
&&\gz \gzj \dypl \dypj\nonumber\\
&=&\frac{1}{4}\int \Bigg \Vert \Bigg[\frac{L}{n_II(I-1)} \izmone \hGd \kG\nonumber\\
&&+\frac{L}{n_I(I-1)}\izjmone \hGd \kGj \Bigg]\nonumber\\
&&+ \Bigg[\frac{L}{n_I(I-1)}\izmone \ratioGg \hgdone \omega_{I,R,j}^g K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right) \nonumber\\ &&K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)+\frac{L}{n_I(I-1)}\izjmone \ratioGgj \hgdone \omega_{I,R,j}^g K_{1g}\left(\frac{X_\ell-X_j}{h_{g}}\right) \nonumber\\
&&K_{2g}\left(\frac{B_{p\ell}-B_{qj}}{h_{g}}\right) \Bigg]\Bigg \Vert^2\gz \gzj \dypl \dypj\nonumber\\
&=&\frac{1}{4}\int \Bigg \Vert \hGd \frac{L}{n_I}\Bigg[\frac{1}{(I-1)}\izmone \kG \nonumber\\
&&+ \frac{1}{(I-1)}\izjmone \kGj\Bigg]\nonumber\\
&&+ \hgdone \frac{L}{n_I}\Bigg [\frac{1}{(I-1)}\izmone \ratioGg \kgz\nonumber\\
&&+\frac{1}{I-1}\izjmone \ratioGgj \kgzj \Bigg]\Bigg \Vert^2\nonumber\\
&&\gz \gzj \dypl \dypj\nonumber\\
&\leq&\frac{1}{2}\int \Bigg\{\Bigg\Vert \frac{L}{n_I}\hGd \Bigg [\frac{1}{(I-1)}\izmone \kG\nonumber\\
&&+\frac{1}{(I-1)}\izjmone \kGj\Bigg]\Bigg\Vert^2\nonumber\\
&&\!\!+\!\Bigg\Vert\frac{L}{n_II}\hgdone \Bigg[\frac{1}{(I-1)}\izmone \ratioGg \kgz\nonumber\\
&&+\frac{1}{(I-1)}\izjmone \ratioGgj\kgzj \Bigg] \Bigg\Vert^2\Bigg\}\nonumber\\
&&\gz \gzj \dypl \dypj\nonumber\\
&=&C+D,
\label{A.21}
\end{eqnarray}}
where the inequality comes from using $(a+b)^2\leq
2(a^2+b^2)$. Therefore, now we need to show that both C and D are $o(L)$.
We consider first the C term in (\ref{A.21}), and note that we can write
$V_{p\ell}=\xi(B_{p\ell},X_\ell,I)$. It gives
\begin{eqnarray*}
C\!\!\!\!&=&\!\!\!\hGtwod \int \left(\frac{L}{n_I}\frac{1}{(I-1)}\right)^2\Bigg\Vert \izmonexi \kG \\
&&+\izjmone \kGj \Bigg\Vert^2 \gz \gzj \dypl \dypj\\
&=&\frac{1}{2} \int \left(\frac{L}{n_I}\frac{1}{(I-1)}\right)^2\Bigg\Vert \imoneratioKGu \omega_{I,R+1,j}^G K_G(-u_2)\\
&&\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})+\ratioigj \omega_{I,R+1,j}^G K_G(u_2)\hbox{\it 1\hskip -3pt I}(u_1 h_G+B_{qj}\leq B_{qj} )\Bigg \Vert^2 \\
&&\guhG \gzj du \dypj\\
&\leq& \int \Bigg[\Bigg\Vert \imoneratioKGu \omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})\Bigg\Vert^2 \\
&&+\left\Vert \ratioigj \omega_{I,R+1,j}^G K_G(u_2)\hbox{\it 1\hskip -3pt I}(u_1 h_G+B_{qj}\leq B_{qj} ) \right\Vert^2\Bigg]\\
&&\guhG \gzj du\dypj\\
&=&C_1+C_2,
\end{eqnarray*}
\noindent where we have used the change of variable
$u=\displaystyle{\frac{Y_{p\ell}-Y_{qj}}{h_{G}}=\left(\frac{B_{p\ell}-B_{qj}}{h_G},\frac{X_\ell-X_j}{h_G}\right)=(u_1,u_2)}$,
and the inequality comes from using $(a+b)^2\leq 2(a^2+b^2)$ and $\left(\frac{L}{n_I}\frac{1}{(I-1)}\right)^2<\infty$.
Next we consider $C_1$
\begin{eqnarray*}
C_1&=& \int \Bigg\Vert \imoneratioKGu \omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})\Bigg\Vert^2 \\
&&\guhG \gzj du \dypj\\
&\!\!\!=&\!\!\! \!\!\int \left\Vert \imoneratioKGu \right\Vert^2 \left\Vert \omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})\right\Vert^2\\
&&\guhG \gzj du \dypj\\
&\!\!\!=& \!\!\! \!\!\int \left\Vert \moneu\right\Vert^2 \left\Vert \omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})\right\Vert^2\\
&& \frac{\guhG g(Y_{qj}|I)}{g_0(u_1h_{G}+B_{qj}|u_2h_{G}+X_j,I)^2} du \dypj\\
&\leq& \int \Vert\moneu\Vert^2 \left\Vert\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1 h_G+B_{qj})\right\Vert^2\\
&&\gzj du\dypj,
\end{eqnarray*}
\noindent where the last inequality comes from the assumption that densities are
bounded.
By the Lebesgue Dominated Convergence (LDC) Theorem and A6-(vi), the
above integral converges to
\begin{eqnarray*}
\int \left\Vert O_p(1)K_G(-u_2)\right\Vert^2 du \int \Vert
m_1(V_{qj},X_j,I;\theta_0)\Vert^2 \gzj\dypj <\infty.
\end{eqnarray*}
Hence, $C_1=o(L)$ as $L \rightarrow\infty$.
A similar argument can be used to show that $C_2=o(L)$ as $L \rightarrow\infty$. Therefore, $C=C_1+C_2=o(L)$.
\medskip
Next we consider the D term in (\ref{A.21}). Namely
\begin{eqnarray*}
D&=&\frac{1}{2}\int \Bigg\Vert\frac{L}{n_I}\hgdone \Bigg[\frac{1}{(I-1)}\izmone \ratioGg \omega_{I,R,j}^g K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right)\\
&&K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)+\frac{1}{(I-1)}\izjmone \ratioGgj\omega_{I,R,j}^g \\
&&K_{1g}\left(\frac{X_\ell-X_j}{h_{g}}\right) K_{2g}\left(\frac{B_{p\ell}-B_{qj}}{h_{g}}\right) \Bigg] \Bigg\Vert^2\gz \gzj \dypl \dypj\\
&=&\hgtwod \left(\frac{L}{n_I}\frac{1}{(I-1)}\right)^2 \int \Bigg\Vert \izmonexi \ratioGg \omega_{I,R,j}^g \\
&&K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right)K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)\\
&&+\izjmone \ratioGgj\kgzj \Bigg\Vert^2\\
&&\gz \gzj \dypl \dypj\\
&\leq&\hgdonehalf \int \Bigg\Vert \imoneratioKGu \ratioGgju\\
&&\omega_{I,R,j}^g K_{1g}(-u_2) K_{2g}(-u_1)\\
&&\!\!\!\!\! +\izjmone \ratioGgj\omega_{I,R,j}^g K_{1g}(u_2) K_{2g}(u_1) \Bigg\Vert^2\!\! \gju \gzj du \dypj\\
&\leq& \hgdone \int \Bigg[\Bigg\Vert \imoneratioKGu \ratioGgju\omega_{I,R,j}^g \\
&&K_{1g}(-u_2) K_{2g}(-u_1)\Bigg\Vert^2+\left\Vert \izjmone \ratioGgj \omega_{I,R,j}^g K_{1g}(u_2) K_{2g}(u_1)\right\Vert^2\Bigg]\\
&&\gju \gzj du \dypj\\
&=&D_1+D_2,
\end{eqnarray*}
\noindent where we have used the change of variable
$u=\displaystyle{\frac{Y_{p\ell}-Y_{qj}}{h_{g}}=\left(\frac{B_{p\ell}-B_{qj}}{h_g},\frac{X_\ell-X_j}{h_g}\right)=(u_1,u_2)}$,
and the first inequality follows from the fact that $\left(\frac{L}{n_I}\frac{1}{(I-1)}\right)^2<\infty$ and the second inequality uses $(a+b)^2\leq 2 (a^2+b^2)$.
We consider first $D_1$. Specifically,
\begin{eqnarray*}
D_1&=&\hgdone \int \Big\Vert \izjmoneu \Gzju \Big\Vert^2 \\
&&\Big\Vert\omega_{I,R,j}^g K_{1g}(-u_2) K_{2g}(-u_1) \Big\Vert^2\frac{\gju g_0(Y_{qj}|I)}{g_0(u_1h_{g}+B_{qj}|u_2h_{g}+X_j,I)^4} du \dypj\\
&\leq &\hgdone \int \Vert \izjmoneu \Vert^2 \left\Vert \omega_{I,R,j}^g K_{1g}(-u_2) K_{2g}(-u_1)\right\Vert^2 \\
&&\gzj du \dypj,
\end{eqnarray*}
\noindent where the inequality uses the fact that
$G(\cdot|\cdot,I)$ is bounded and that densities are
bounded from above.
By the LDC Theorem and A6-(vi) the above integral converges to
\begin{eqnarray*}
\int \left\Vert O_p(1) K_{1g}(-u_2) K_{2g}(-u_1)\right\Vert^2 du \int \left\Vert
m_1(V_{qj},X_j,I;\theta_0) \right\Vert^2 \gzj \dypj<\infty.
\end{eqnarray*}
Hence, $D_1=o(L)$ if and only if $Lh_{g}^2 \rightarrow\infty$ as
implied by A4.AN-(ii) since,
$
Lh_{g}^2=\sqrt{L}\sqrt{L} h_{g}^{2}\rightarrow\infty.
$
A similar argument can be used to show that $D_2=o(L)$. That is, $D_2=o(L)$ if and only if $Lh_{g}^2 \rightarrow\infty$, as
implied by A4.AN-(ii).
Therefore,
$C+D=C_1+C_2+D_1+D_2=o(L)$ and the desired result follows, i.e by
Lemma 3.1 in Powell, Stock and Stoker (1989) $\sqrt{L}(U_L-\hat
U_L)=o_p(1)$.
\bigskip
Next we consider the second term in (\ref{B112})
\begin{eqnarray}
B_{1122}&=&\frac{L(L-1)}{L^2} \sqrt{L}\hat U_L =\frac{L(L-1)}{L^2} \sqrt{L}\left\{\theta_L+\displaystyle{\frac{2}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^{I}} [r_L(Y_{p\ell},I)-\theta_L]\right\} \notag\\
&=&\frac{L(L-1)}{L^2}\sqrt{L}E[\pl] +\frac{L(L-1)}{L^2}\sqrt{L}\!\!\displaystyle{\frac{2}{L}\sum_{\{\ell:I_\ell=I\}}^L\frac{1}{I}\sum_{p=1}^{I}} [r_L(Y_{p\ell},I)-\theta_L].
\label{B1122}
\end{eqnarray}
Next, we show that the first term in (\ref{B1122}) is $o_{as}(1)$.
Consider only the expectation part in the first term in (\ref{B1122}). Namely
\begin{eqnarray}
\lefteqn{{\rm E}[\pl]}\nonumber\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int\Bigg\{\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}
\Bigg[\omega_{I,R+1,j}^G K_{G,h_G}(X_j-X_\ell)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\nonumber\\
&&-\frac{\Gi}{\gi}\omega_{I,R,j}^g K_{1g,h_{g}}(X_j-X_\ell) K_{2g,h_{g}}(B_{qj}-B_{p\ell})\Bigg]\nonumber\\
&&+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}\Bigg[\omega_{I,R+1,j}^G K_{G,h_G}(X_\ell-X_j)\hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq B_{qj})\nonumber\\
&&-\frac{\Gj}{\gj} \omega_{I,R,j}^g K_{1g,h_{g}}(X_\ell-X_j) K_{2g,h_{g}}(B_{p\ell}-B_{qj})\Bigg]\Bigg\} \nonumber\\
&&\gjoint \gjointj \dypl\dypj \nonumber\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int \hGd \Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\kG\nonumber\\
&&+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}\kGj\Bigg]\nonumber \\
&& \gjoint \gjointj \dypl\dypj \nonumber \\
&&-\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int\hgdone \Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)} \ratioGg \omega_{I,R,j}^g K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right) \nonumber\\
&&K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}\ratioGgj\omega_{I,R,j}^g K_{1g}\left(\frac{X_\ell-X_j}{h_{g}}\right) \nonumber\\
&& K_{2g}\left(\frac{B_{p\ell}-B_{qj}}{h_{g}}\right)\Bigg]\gjoint \gjointj \dypl\dypj \nonumber\\
&=&A_1-A_2.
\label{A1_A2}
\end{eqnarray}
We consider first $A_1$.
\begin{eqnarray*}
\Vert A_1\Vert&=&\Bigg\Vert\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int \hGd \Bigg
[\frac{m_1(\xi(B_{p\ell},X_\ell,I),X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\omega_{I,R+1,j}^G K_G\left(\frac{X_j-X_\ell}{h_{G}}\right) \\
&& \hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}\kGj\Bigg]\\
&&\gjoint \gjointj \dypl\dypj\Bigg\Vert\leq \Vert A_{11}\Vert+\Vert A_{12}\Vert
\end{eqnarray*}
It is enough to show that $\Vert A_{11}\Vert=o_{as}(1/\sqrt{L})$ since the
same argument can be used to show that $\Vert
A_{12}\Vert=o_{as}(1/\sqrt{L})$. We observe the following
\begin{eqnarray}
\Vert A_{11}\Vert&=&\Bigg\Vert\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int \frac{1}{h_G}
\Bigg[\frac{m_1(\xi(B_{p\ell},X_\ell,I),X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\omega_{I,R+1,j}^G K_G\left(\frac{X_j-X_\ell}{h_{G}}\right)\nonumber\\
&&\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{p\ell})\Bigg] \gjoint \gjointj \dypl\dypj\Bigg\Vert\nonumber\\
&\leq&\Bigg\Vert\sum_{I}\int \frac{1}{h_G}
\frac{m_1(\xi(B_{p\ell},X_\ell,I),X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\kG\nonumber\\
&&\gjoint \gjointj \dypl\dypj\Bigg\Vert\nonumber\\
&\leq&\Bigg\Vert\sum_{I}\int \frac{1}{h_G}m_1(\xi(B_{p\ell},X_\ell,I),X_\ell,I;\theta_0)\kG\nonumber\\
&&\gjoint \dypl\dypj\Bigg\Vert\nonumber\\
&=&h_G \Bigg\Vert\sum_{I}\int m_1(\xi(uh_G+Y_{qj},I),u_2h_G+X_j,I;\theta_0)\omega_{I,R+1,j}^G K_G(-u_2)\nonumber \\
&&\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1h_G+B_{qj}) g_0(uh_G+Y_{qj},I) du \dypj\Bigg\Vert\nonumber\\
&\leq&h_G \Bigg\Vert\sum_{I}\int m_1(\xi(uh_G+Y_{qj},I),u_2h_G+X_j,I;\theta_0)K_G(-u_2)\nonumber \\
&&\hbox{\it 1\hskip -3pt I}(B_{qj}\leq u_1h_G+B_{qj}) g_0(uh_G+Y_{qj},I) du \dypj\Bigg\Vert,
\label{A24}
\end{eqnarray}
\noindent where we have used the change of variable $u=\displaystyle{\frac{Y_{p\ell}-Y_{qj}}{h_{G}}=\left(\frac{B_{p\ell}-B_{qj}}{h_G},\frac{X_\ell-X_j}{h_G}\right)=(u_1,u_2)}$ and the fact that densities are bounded. The last
inequality comes from observing that $\omega_{I,R+1,j}^G=O_p(1)$.
Now consider the expectation inside the norm of the term above
evaluated at $h_G=0$, namely
\begin{eqnarray*}
\sum_{I}\iint m_1(\xi(Y_{qj},I),X_j,I;\theta_0)g_0(Y_{qj},I) \dypj
&=& \sum_{I}\int \left[\int m_1(\xi(Y_{qj},I),X_j,I;\theta_0)
g_0(B_{qj}|X_j,I) dB_{qj}\right] \\
&&f_m(X_j,I) dX_j\\
&=& \sum_{I}{\rm{E}}[m_1(V,X,I;\theta_0)],
\end{eqnarray*}
\noindent where we use A3-(ii) and the Law of Iterated Expectations. The last line in the expression above follows from observing that the integral inside can be solved by using
integration by parts twice, as follows:
\begin{eqnarray*}
\lefteqn{\int_{\underline{B}(X_j)}^{\overline{B}(X_j,I)} m_1(\xi(B_{qj},X_j,I),X_j,I;\theta_0) g_0(B_{qj}|X_j,I) dB_{qj}}\\
&=& m_1(\xi(\overline{B}(X_j,I),X_j,I),X_j,I;\theta_0)G_0(\overline{B}(X_j,I)|X_j,I)-
m_1(\xi(\underline{B}(X_j),X_j,I),X_j,I;\theta_0) \\
&&G_0(\underline{B}(X_j)|X_j,I)-\int_{\underline{B}(X_j)}^{\overline{B}(X_j,I)} m_{11}(\xi(B_{qj},X_j,I),X_j,I;\theta_0) G_0(B_{qj}|X_j,I) dB_{qj}\\
&=& m_1(\overline{V},X_j,I;\theta_0)-\int_{\underline{B}(X_j)}^{\overline{B}(X_j,I)} m_{11}(\xi(B_{qj},X_j,I),X_j,I;\theta_0) G_0(B_{qj}|X_j,I) dB_{qj}\\
&=& m_1(\overline{V},X_j,I;\theta_0)-m_1(\xi(\overline{B}(X_j,I),X_j,I),X_j,I;\theta_0) G_0(\overline{B}(X_j,I)|X_j,I)\\
&&+m_1(\xi(\underline{B}(X_j),X_j,I),X_j;\theta_0) G_0(\underline{B}(X_j)|X_j,I)+\int_{\underline{B}(X_j)}^{\overline{B}(X_j,I)} m_{1}(\xi(B_{qj},X_j,I),X_j,I;\theta_0) g_0(B_{qj}|X_j,I) dB_{qj}\\
&=& m_1(\overline{V},X_j,I;\theta_0)-m_1(\overline{V},X_j,I;\theta_0)+\int_{\underline{B}(X_j)}^{\overline{B}(X_j,I)}
m_{1}(\xi(B_{qj},X_j,I),X_j,I;\theta_0)
g_0(B_{qj}|X_j,I) dB_{qj}\\
&=&\int_{\underline{V}(X_j,I)}^{\overline{V}(X_j,I)} m_1(V_{qj},X_j,I;\theta_0) f(V_{qj}|X_{j},I) dV_{qj}
= {\rm{E}}[m_1(V,X,I;\theta_0)|X,I],
\end{eqnarray*}
\noindent where the fifth equality uses
$G_0(B_{qj}|X_j,I)=F(\xi(B_{qj},X_j,I)|X_j,I)$, so that
$g_0(B_{qj}|X_j,I)=f(V_{qj}|X_j,I) \xi_1(B_{qj},X_j,I)$.
Therefore at $h_G=0$ the integral inside the norm in (\ref{A24}) exists by A6-(vii). Thus, we can apply a Taylor expansion of order $R+1$ in the RHS of (\ref{A24})
around $h_G$ to obtain
\begin{eqnarray*}
\Vert A_{11}\Vert&\leq& h_G \sum_{I} \left\Vert d_1 h_G+d_2\frac{h_G^2}{2}+\ldots+d_R \frac{h_G^R}{R!}+O(h_G^{R+1})\right\Vert\\
&=&\sum_{I} \left\Vert d_1 h_G^2 +d_2\frac{h_G^3}{2}+\ldots+d_R \frac{h_G^{R+1}}{R!}+O(h_G^{R+2})\right\Vert.
\end{eqnarray*}
We note that the remainder term vanishes, i.e.
$\sqrt{L}h_G^{R+2}=o(1)$, and also that
$\sqrt{L}h_G^{R+1}=o(1)$, by A4.AN-(I). The
remaining $R-1$ terms also vanish by A3-(iii), i.e, since the kernels are of order $R-1$. To see this observe
that the $k$th coordinate of $d_\rho$, $\rho=1,\ldots,R-1$ is
\begin{eqnarray*}
d_{k_\rho}&=&\frac{\partial^\rho}{\partial h_G^\rho}\int [H_k(uh_G+\overline{Y})-H_k(uh_G+\underline{Y})]K_G(-u_2)du\Big{|}_{h_G=0}\\
&=&\sum_{k_1,\ldots,k_\rho=1}^{2} \int (u_{k_1}\ldots u_{k_\rho})K_G(-u_2)
\frac{\partial^\rho}{\partial Y_{k_1}\ldots \partial Y_{k_\rho}}H_k(\overline{Y}) du\\
&&-\sum_{k_1,\ldots,k_\rho=1}^{2} \int (u_{k_1}\ldots u_{k_\rho})K_G(-u_2)
\frac{\partial^\rho}{\partial Y_{k_1}\ldots \partial Y_{k_\rho}}H_k(\underline{Y})du=0,
\end{eqnarray*}
where $dH_k/dY(y)= m_{1,k}(\xi(y,I),x,I;\theta_0)
g_0(y,I)$. The third equality uses A3-(iii), that is since
$K_G(\cdot)$ is a higher order kernel, all moments of order strictly
smaller than $R-1$ vanish.
It remains to consider now $A_2$ in (\ref{A1_A2}). Namely
\begin{eqnarray*}
\Vert A_2\Vert&=&\Bigg\Vert\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int\hgdone \Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)} \ratioGg \kgz\\
&&+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}\ratioGgj\kgzj\Bigg] \\
&&\gjoint \gjointj \dypl\dypj\Bigg\Vert \leq \Vert A_{21}\Vert +\Vert A_{22}\Vert.
\end{eqnarray*}
We show only that $A_{21}=o_{as}(1/\sqrt{L})$ since
a similar argument can be used to show that $A_{22}=o_{as}(1/\sqrt{L})$.
We observe the following
\begin{eqnarray*}
\Vert A_{21}\Vert&=&\Bigg\Vert\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int \frac{1}{h_g^{2}}\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}
{g_0(B_{p\ell}|X_\ell,I)}\ratioGg\\
&&\kgz\gjoint \gjointj \dypl\dypj\Bigg\Vert\\
&\leq&\Bigg\Vert\sum_{I}\int \frac{1}{h_g^{2}}\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\ratioGg \omega_{I,R,j}^g K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right) \\
&&K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)\gjoint \gjointj \dypl\dypj\Bigg\Vert\\
&\leq&\Bigg\Vert\sum_{I}\int \frac{1}{h_g^{2}}m_1(\xi(B_{p\ell},X_\ell,I),X_\ell,I;\theta_0)\kgz\\
&&\gjoint \gjointj \dypl\dypj\Bigg\Vert\\
&=&\Bigg\Vert\sum_{I}\int m_1(\xi(uh_g+Y_{qj},I),u_2+X_j,I;\theta_0)\omega_{I,R,j}^g K_{1g,h_g}(-u_2)
K_{2g,h_g}(-u_1)\\
&&g_0(uh_g+Y_{qj},I)du dY_{qj}\Bigg\Vert\\
&\leq&\Bigg\Vert\sum_{I}\int m_1(\xi(uh_g+Y_{qj},I),u_2+X_j,I;\theta_0) K_{1g,h_g}(-u_2)
K_{2g,h_g}(-u_1)
g_0(uh_g+Y_{qj},I)du dY_{qj}\Bigg\Vert,
\end{eqnarray*}
\noindent where we have used that $(1/2)(L/n_I)1/(I-1)\leq \infty$ and also that densities are bounded.
The last equality uses the change of variable $u=(Y_{p\ell}-Y_{qj})/h_g$ and the last inequality comes from observing that $\omega_{I,R,j}^g=O_p(1)$.
We observe that $A_{21}$ can be expanded as a Taylor series of order
$R$ in the bandwidth $h_g$. Moreover, $A_{21}|_{h_g=0}<\infty$ by A6-(vii) as
we have already shown above for $A_{11}|_{h_G=0}<\infty$.
Then, we can apply a Taylor expansion around $h_g$ to
obtain
\begin{eqnarray*}
\Vert A_{21}\Vert \leq \sum_{I_j}\left\Vert c_1 h_g+c_2 \frac{h_g^2}{2}+\ldots+c_{R-1} \frac{h_g^{R-1}}{(R-1)!}+O(h_g^{R})\right\Vert.
\end{eqnarray*}
We note that the remainder term vanishes, i.e.
$\sqrt{L}h_g^{R}=o(1)$ by A4.AN-(ii). The remaining $R-1$ terms also vanish by A3-(iii). To
see this observe that the $k$th coordinate of $c_\rho$,
$\rho=1,\ldots,R-1$ is
\begin{eqnarray*}
c_{k_\rho}&=&\frac{\partial^\rho}{\partial h_g^\rho}\int [H_k(uh_g+\overline{Y})-H_k(uh_g+\underline{Y})]K_{1g}(-u_2) K_{2g}(-u_1) du|_{h_g=0}\\
&=&\sum_{k_1,\ldots,k_\rho=1}^{2} \int (u_{k_1}\ldots u_{k_\rho})K_{1g}(-u_2) K_{2g}(-u_1)
\frac{\partial^\rho}{\partial Y_{k_1}\ldots \partial Y_{k_\rho}}H_k(\overline{Y}) du\\
&&-\sum_{k_1,\ldots,k_\rho=1}^{2} \int (u_{k_1}\ldots u_{k_\rho})K_{1g}(-u_2) K_{2g}(-u_1)
\frac{\partial^\rho}{\partial Y_{k_1}\ldots \partial Y_{k_\rho}}H_k(\underline{Y})du=0,
\end{eqnarray*}
where $dH_k/dY(y)= m_{1,k}(\xi(y,I),x,I;\theta_0)
g_0(y,I)$. The third equality uses A3-(iii), that is since
$K_{1g}(\cdot)$ and $K_{2g}(\cdot)$ are higher order kernels, all moments of order strictly
smaller than $R-1$ vanish.
This shows that the first term in (\ref{B1122}) indeed is $o_{as}(1)$.
We still have to show that the second term in (\ref{B1122}) is $O_p(1)$.
In fact we will not only show that we will also provide the asymptotic linear representation, which gives us the asymptotic variance.
From Equation (\ref{B1122}) we have
\begin{eqnarray*}
\sqrt{L}(\hat\theta-\theta_0)-\sqrt{L}(\tilde{\theta}-\theta_0)=\frac{L(L-1)}{L^2}\frac{2}{\sqrt{L}}
\sum_{\{\ell:I_\ell=I\}}^L \frac{1}{I}\sum_{p=1}^{I}
[r_L(Y_{p\ell},I)-\theta_L]
\end{eqnarray*}
where, $Y_{p\ell}=(B_{p\ell},X_\ell)$ and $r_L(Y_{p\ell},I)={\rm{E}}[\pl|(Y_{p\ell},I)]$ and $\theta_L={\rm{E}}[r_L(Y_{p\ell},I)]={\rm{E}}[\pl]$.
First, we show that $$r_L(Y_{p\ell},I)=\displaystyle{-\sum_{I}\frac{1}{I(I-1)}
N(Y_{p\ell},I)f_m^{-1}(X_\ell,I)g_0(Y_{p\ell},I)+t_L(Y_{p\ell},I)}.$$
We observe that
\begin{eqnarray*}
r_L(Y_{p\ell},I)&=&{\rm{E}}[\pl|(Y_{p\ell},I)]\\
\!\!\!&=&\!\! \!\Bigg\{\begin{array}{ll}
\int p_L((B_{p\ell},X_\ell,I),(B_{qj},X_j,I)) g_0(B_{qj},X_j,I) dY_{qj} & \mbox{if $\ell\neq j$} \\
\int p_L((B_{pj},X_j,I),(B_{qj},X_j,I))
g_0((B_{pj},X_j,I),(B_{qj},X_j,I)|(B_{pj},X_j,I)) dY_{qj} &
\mbox{if $\ell= j$.}
\end{array}\Bigg.
\end{eqnarray*}
We consider first the case $\ell\neq j$.
\begin{eqnarray*}
&&r_L(Y_{p\ell},I)=\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \Bigg\{\int\Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\hGd \kG\\
&&+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)} \hGd\kGj\Bigg] g_0(Y_{qj},I) \dypj \Bigg\}\\
&&-\frac{1}{2} \frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \Bigg\{\int\Bigg[\frac{m_1(V_{p\ell},X_\ell,I;\theta_0)}{g_0(B_{p\ell}|X_\ell,I)}\frac{G_0(B_{p\ell}|X_\ell,I)}{g_0(B_{p\ell}|X_\ell,I)}\hgdone \omega_{I,R,j}^g K_{1g}\left(\frac{X_j-X_\ell}{h_{g}}\right) \\
&&K_{2g}\left(\frac{B_{qj}-B_{p\ell}}{h_{g}}\right)+\frac{m_1(V_{qj},X_j,I;\theta_0)}{g_0(B_{qj}|X_j,I)}
\frac{G_0(B_{qj}|X_j,I)}{g_0(B_{qj}|X_j,I)}\hgdone \omega_{I,R,j}^g K_{1g}\left(\frac{X_\ell-X_j}{h_{g}}\right)\\ &&K_{2g}\left(\frac{B_{p\ell}-B_{qj}}{h_{g}}\right)\Bigg]g_0(Y_{qj},I)\dypj \Bigg\}\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \Bigg\{\int\Bigg[ M(Y_{p\ell},I)\hGd \kG\\
&&+M(Y_{qj},I)\hGd\kGj\Bigg] g_0(Y_{qj},I) \dypj \Bigg\}\\
&&-\frac{1}{2}\frac{L}{n_I} {\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \Bigg\{\int \Bigg[N(Y_{p\ell},I)\hgdone \kgz\\
&&+ N(Y_{qj},I)\hgdone \kgzj\Bigg]g_0(Y_{qj},I) \dypj\Bigg\}\\
&\!\!=&\!\!\!\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq u_1h_G+B_{p\ell})+M(uh_G+Y_{p\ell},I)\\
&&\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p\ell}\leq B_{p\ell})\Big]g_0(uh_G+Y_{p\ell},I)du\\
&&-\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int\Big[N(Y_{p\ell},I)\omega_{I,R,j}^g K_{1g}(u_2)K_{2g}(u_1)+N( uh_g+Y_{p\ell},I)\\
&&\omega_{I,R,j}^g K_{1g}(-u_2)K_{2g}(-u_1)\Big]g_0(uh_g+Y_{p\ell},I)du.
\end{eqnarray*}
We note that as $h=(h_G,h_g) \rightarrow 0$ we have
\begin{eqnarray*}
r_L(Y_{p\ell},I) &\longrightarrow& -\frac{1}{2} \frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \Big[N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) K_{1g}(u)] K_{1g}(u_2)K_{2g}(u_1)\\
&&+ N(Y_{p\ell},I_j)f^{-1}_m(X_\ell,I) K_{1g}(-u_2)K_{2g}(-u_1)\Big]g_0(Y_{p\ell},I)du\\
&&=-\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I)
\end{eqnarray*}
\noindent where we have used the following
\begin{eqnarray*}
\omega_{I,R,j}^g&=&e_1^T \left[\frac{1}{n_I h_g}\sum_{\iota=1}^{n_I} {\mathbf{{x_\iota x_\iota}}}^T K_{1g}
\left(\frac{X_\iota-X_\ell}{h_g}\right) \right]^{-1} [1\quad (X_j-X_\ell)\ldots (X_j-X_\ell)^{R-1}]^T\\
&=&e_1^T \left[\frac{1}{n_I h_g}\sum_{\iota=1}^{n_I} {\mathbf{{x_\iota x_\iota}}}^T K_{1g}
\left(\frac{X_\iota-X_\ell}{h_g}\right) \right]^{-1} [1\quad (-u_2h_g)\ldots (-u_2h_g)^{R-1}]^T\\
&\stackrel{p}{\longrightarrow}& e_1^T \left[{\rm{E}\left({\mathbf{{x_\iota x_\iota}}}^T K_{1g}
\left(\frac{X_\iota-X_\ell}{h_g}\right)\right) } \right]^{-1} e_1=f^{-1}_m(X_\ell,I)
\end{eqnarray*}
\noindent therefore we define
\begin{eqnarray*}
r_L(Y_{p\ell},I_j)=-\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)+t_L(Y_{p\ell},I).
\end{eqnarray*}
We consider now the reminder term $t_L(Y_{p\ell},I)$
\begin{eqnarray*}
t_L(Y_{p\ell},I)&=&r_L(Y_{p\ell},I)+\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I)\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq u_1h_G+B_{p\ell})\\
&&+M(uh_G+Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p\ell}\leq B_{p\ell})\Big]g_0(uh_G+Y_{p\ell},I)du\\
&&-\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int\Big[N(Y_{p\ell},I)\omega_{I,R,j}^g K_{1g}(u_2)K_{2g}(u_1)+N( uh_g+Y_{p\ell},I)\\
&&\omega_{I,R,j}^g K_{1g}(-u_2)K_{2g}(-u_1)\Big]g_0(uh_g+Y_{p\ell},I)du\\
&&+\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)
\end{eqnarray*}
Now, using $\int K_{1g}(u_2)K_{2g}(u_1)du =1$, we can write
\begin{eqnarray*}
\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I)
&=&\frac{1}{2}\frac{1}{I} {\displaystyle{\sum_{I}\frac{1}{(I-1)}}}\int K_{1g}(u_2)K_{2g}(u_1) N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I) du\\
&&+\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int K_{1g}(-u_2)K_{2g}(-u_1) N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I) du
\end{eqnarray*}
\noindent therefore we can write the reminder term as follows
\begin{eqnarray*}
t_L(Y_{p\ell},I_j)&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq u_1h_G+B_{p\ell})\\
&&\!\!\!+M(uh_G+Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p\ell}\leq B_{p\ell})\Big]g_0(uh_G+Y_{p\ell},I)du\\
&&-\frac{1}{2}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \frac{L}{iL_i} N(Y_{p\ell},I) \omega_{I,R,j}^g K_{1g}(u_2)K_{2g}(u_1) g_0(uh_g+Y_{p\ell},I)du\\
&&\!\!\!\!-\!\!\!\!\frac{1}{2}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \frac{L}{iL_i} N(uh_g+Y_{p\ell},I)\omega_{I,R,j}^g K_{1g}(-u_2)K_{2g}(-u_1)g_0(uh_g+Y_{p\ell},I)du\\
&&+\frac{1}{2}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \frac{1}{I} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)K_{1g}(u_2)K_{2g}(u_1) g_0(Y_{p\ell},I) du\\
&&+\frac{1}{2}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \frac{1}{I} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I)K_{1g}(-u_2)K_{2g}(-u_1) g_0(Y_{p\ell},I) du\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq u_1h_G+B_{p\ell})\\
&&+M(uh_G+Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p\ell}\leq B_{p\ell})\Big]g_0(uh_G+Y_{p\ell},I)du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int N(Y_{p\ell},I) K_{1g}(u_2)K_{2g}(u_1)\Bigg[\frac{L}{L_I}\omega_{I,R,j}^gg_0(uh_g+Y_{p\ell},I)-f^{-1}_m(X_\ell,I)g_0(Y_{p\ell},I)\Bigg] du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int K_{1g}(-u_2)K_{2g}(-u_1)\\
&&\Bigg[\frac{L}{L_I}\omega_{I,R,j}^g N(uh_g+Y_{p\ell},I)g_0(uh_g+Y_{p\ell},I)-f^{-1}_m(X_\ell,I)N(Y_{p\ell},I)g_0(Y_{p\ell},I)\Bigg] du\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p\ell}\leq u_1h_G+B_{p\ell})\\
&&+M(uh_G+Y_{p\ell},I)\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p\ell}\leq B_{p\ell})\Big]g_0(uh_G+Y_{p\ell},I)du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int N(Y_{p\ell},I) K_{1g}(u_2)K_{2g}(u_1) \left[f^{-1}_m(X_\ell,I)+o_{as}(1)\right][g_0(uh_g+Y_{p\ell},I)-g_0(Y_{p\ell},I)]du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int K_{1g}(-u_2)K_{2g}(-u_1) \left[f^{-1}_m(X_\ell,I)+o_{as}(1)\right] \\
&&[N(uh_g+Y_{p\ell},I)g_0(uh_g+Y_{p\ell},I)-N(Y_{p\ell},I)g_0(Y_{p\ell},I)]du
\end{eqnarray*}
\noindent thus using the above expression we have
\begin{eqnarray*}
\frac{2}{\sqrt{L}} \sum_{\{\ell:I_\ell=I\}}^L \frac{1}{I}\sum_{p=1}^{I}[r_L(Y_{p\ell},I)-\theta_L]&=&\frac{2}{\sqrt{L}} \sum_{\{\ell:I_\ell=I\}}^L \frac{1}{I}\sum_{p=1}^{I} \Bigg\{-\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)\\
&&+{\rm{E}}\Big[\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)\Big]
+t_L(Y_{p\ell},I)-{\rm{E}}[t_L(Y_{p\ell},I)]\Bigg\}\\
&=&-\frac{2}{\sqrt{L}} \sum_{\{\ell:I_\ell=I\}}^L \frac{1}{I}\sum_{p=1}^{I}
\Bigg\{\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)\\
&&-{\rm{E}}\Big[\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{p\ell},I) f^{-1}_m(X_\ell,I) g_0(Y_{p\ell},I)\Big] \Bigg\}\\
&&+\frac{2}{\sqrt{L}} \sum_{\{\ell:I_\ell=I\}}^L \frac{1}{I}\sum_{p=1}^{I}\Big[t_L(Y_{p\ell},I)-{\rm{E}}[t_L(Y_{p\ell},I)]\Big].
\end{eqnarray*}
We denote the second term above by $T_L$ and we observe that
${\rm{E}}[T_L]=0$. We now show that ${\rm{var}}[T_L]=o_{as}(1)$.
\begin{eqnarray}
{\rm{var}}[T_L]&=&4\frac{L_I}{L} {\rm{var}}\left[\frac{1}{I}\sum_{p=1}^{I}
t_1(Y_{p1},I)\right]= 4\frac{L_I}{L}{\rm{E}}\left\{{\rm{var}}\left[\frac{1}{I}\sum_{p=1}^{I}
t_1(Y_{p1},I)\Bigg|I \right]\right\} + 4\frac{L_I}{L}
{\rm{var}}\left\{{\rm{E}}\left[\frac{1}{I}\sum_{p=1}^{I}
t_1(Y_{p1},I)\Bigg|I \right]\right\}\nonumber\\
&=&4\frac{L_I}{L}{\rm{E}}\left\{\frac{1}{I}{\rm{var}}\left[
t_1(Y_{p1},I)\Bigg|I \right]\right\}+4\frac{L_I}{L} {\rm{var}}\Big\{
{\rm{E}}\left[ t_1(Y_{p1},I)|I
\right]\Big\}= A+B.
\label{varTL}
\end{eqnarray}
We consider first the $k$th coordinate of the conditional variance
inside the A term above, namely
\begin{eqnarray*}
{\rm{var}}\left[ t_{1_k}(Y_{p1},I)\Big|I\right]&\leq&{\rm{E}}\left[ t_{1_k}(Y_{p1},I)^2\Big|I\right]
\leq O\left(h_G^{2}\right)+O\left(h_g^{2(R-1)}\right),
\end{eqnarray*}
\noindent where the last inequality comes from observing that
\begin{eqnarray*}
t_{1_k}(Y_{p1},I)&=&\frac{1}{2}\frac{L}{iL_i}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G\Big[ M(Y_{p1},I)\omega_{I,R+1,j}^G K_G(u_2) \hbox{\it 1\hskip -3pt I}(B_{p1}\leq u_1h_G+B_{p1})\\
&&+M(uh_G+Y_{p1},I)\omega_{I,R+1,j}^G K_G(-u_2)\hbox{\it 1\hskip -3pt I}(u_1h_G+B_{p1}\leq B_{p1})\Big]g_0(uh_G+Y_{p1},I)du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int N(Y_{p1},I) K_{1g}(u_2)K_{2g}(u_1) f^{-1}_m(X_1,I)\\
&&[g_0(uh_g+Y_{p1},I)-g_0(Y_{p1},I)]du\\
&&-\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int K_{1g}(-u_2)K_{2g}(-u_1) f^{-1}_m(X_1,I) [N(uh_g+Y_{p1},I)\\
&&g_0(uh_g+Y_{p1},I)-N(Y_{p1},I)g_0(Y_{p1},I)]du+o_{as}(1)
= a+b+c+o_{as}(1).
\end{eqnarray*}
Therefore, applying $(a+b)^2\leq 2 (a^2+b^2)$ twice yields $(a+b+c)^2\leq \kappa(a^2+b^2+c^2)$, and thus
\begin{eqnarray*}
{\rm{E}}\left[ t_{1_k}(Y_{p1},I)^2\Big|I\right]&=&{\rm{E}}[(a+b+c)^2|I]+o_{as}(1)\leq 4 {\rm{E}}[a^2+b^2+c^2|I]+o_{as}(1)\\
&=& O\left(h_G^{2}\right)+O\left(h_g^{2(R-1)}\right)+O\left(h_g^{2(R-1)}\right),
\end{eqnarray*}
where the order of the last two terms after the last equality follows from $(R-1)^{th}$ Taylor Expansion around $Y_{p1}$ and the kernels are of order $R-1$ by A.3-(iii).
Next, we consider $B$ in (\ref{varTL}):
\begin{eqnarray*}
\frac{B}{4}&=&\frac{L_I}{L}{\rm{var}}\Big\{
{\rm{E}}\left[ t_1(Y_{p1},I)|I
\right]\Big\}\leq{\rm{E}}\Big\{
{\rm{E}}\left[ t_1(Y_{p1},I)|I
\right]^2\Big\}\leq{\rm{E}}\Big\{
{\rm{E}}\left[ t_1(Y_{p1},I)^2|I
\right]\Big\}\\
&\leq& O\left(h_G^{2}\right)+O\left(h_g^{2(R-1)}\right)+O\left(h_g^{2(R-1)}\right),
\end{eqnarray*}
where the last inequality follows from the same argument used above. Hence, by Chebyshev Inequality $T_L=o_p(1)$.
We consider now the case $\ell=j$ and observe the following
\begin{eqnarray*}
r_L(Y_{pj},I)&=&{\rm{E}}[\plj|(B_{pj},X_j,I)]\\
&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int
\frac{1}{h_G}\Bigg[ M(Y_{pj},I)
\omega_{I,R+1,j}^G K_G(0)\hbox{\it 1\hskip -3pt I}(B_{qj}\leq B_{pj})\\
&&+ M(Y_{qj},I) \omega_{I,R+1,j}^G K_G(0)\hbox{\it 1\hskip -3pt I}(B_{pj}\leq B_{qj})\Bigg]g_0((Y_{pj},I),(Y_{qj},I)|(Y_{pj},I)) \dypj\\
&&-\frac{1}{2} \frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \frac{1}{h_g^{2}}\Bigg[
N(Y_{pj},I) \omega_{I,R,j}^g K_{1g}(0)K_{2g}\left(\frac{B_{qj}-B_{pj}}{h_g}\right) \\
&&+ N(Y_{qj},I) \omega_{I,R,j}^g K_{1g}(0) K_{2g}\left(\frac{B_{pj}-B_{qj}}{h_g}\right)\Bigg]g_0((Y_{pj},I),(Y_{qj},I)|(Y_{pj},I))\dypj.
\end{eqnarray*}
Making the change of variables $u=(Y_{qj}-Y_{pj})/h_G$ and $\tilde u=(Y_{qj}-Y_{pj})/h_g$ gives
\begin{eqnarray*}
r_L(Y_{pj},I)&=&\frac{1}{2}\frac{L}{n_I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int h_G \Big[ M(Y_{pj},I)\omega_{I,R+1,j}^G K_G(0) \hbox{\it 1\hskip -3pt I}(u_1h_G+B_{pj}\leq B_{pj})+M(uh_G+Y_{pj},I)\\
&&\omega_{I,R+1,j}^G K_G(0) \hbox{\it 1\hskip -3pt I}(B_{pj} \leq u_1h_G+B_{pj})\Big]g_0((Y_{pj},I),(uh_G+Y_{pj},I)|(Y_{pj},I))du\\
&&-\frac{1}{2}\frac{L}{iL_i}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int\Big[N(Y_{pj},I)\omega_{I,R,j}^g K_{1g}(0)K_{2g}(\tilde u_1)+N(\tilde uh_g+Y_{pj},I)\\
&&\omega_{I,R,j}^g K_{1g}(0)K_{2g}(-\tilde u_1)\Big]g_0((Y_{pj},I),(uh_g+Y_{pj},I)|(Y_{pj},I))d\tilde u.
\end{eqnarray*}
Next, we observe that as $h=(h_G,h_g) \rightarrow 0$ we have
\begin{eqnarray*}
r_L(Y_{pj},I) &\rightarrow& -\frac{1}{2}\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} \int \Big[N(Y_{pj},I) f_m^{-1}(X_j,I)K_{1g}(0)K_{2g}(\tilde u_1)\\
&&+N(Y_{pj},I) f_m^{-1}(X_j,I)K_{1g}(0)K_{2g}(-\tilde u_1)\Big]g_0((Y_{pj},I),(Y_{pj},I)|(Y_{pj},I))d\tilde u\\
&=&-\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{pj},I)f_m^{-1}(X_j,I) g_0(Y_{pj},I),
\end{eqnarray*}
and, as before, we define
\begin{eqnarray*}
r_L(Y_{pj},I)=-\frac{1}{I}{\displaystyle{\sum_{I}\frac{1}{(I-1)}}} N(Y_{pj},I)f_m^{-1}(X_j,I) g_0(Y_{pj},I)+t_L(Y_{pj},I).
\end{eqnarray*}
The rest of the proof is analogous to the one for the case $\ell\neq j$.
Next, we consider $B_{12}$ in (\ref{B11+B12}):
\begin{eqnarray*}
\Vert B_{12}\Vert&=&\Bigg\Vert\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} m_1(V_{p\ell,Z_\ell;\theta_0}) \sqrt{L}\Bigg[\frac{1}{\gtilde\g}\Bigg(\frac{\G}{\g}\\
&&\!\!\left[\gtilde-\g\right]^2\!\!\!-\!\!\left[\Gtilde-\G\right]\!\!\left[\gtilde-\g\right]\!\Bigg)\!\Bigg]\!\Bigg\Vert\\
&\leq&\left(\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} \Vert \moneo\Vert^2 \right)^{\frac{1}{2}}\\
&& \!\!\!\!\!\!\Bigg\{\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} L \Bigg[\frac{1}{\gtilde \g}\!\!\Bigg(\frac{\G}{\g}\left[\gtilde-\g\right]\\
&& -\left[\Gtilde-\G\right]\Big[\gtilde-\g\Big]\Bigg)\Bigg]^2\Bigg\}^{\frac{1}{2}}=B_{121} B_{122},
\end{eqnarray*}
\noindent where the inequality comes from Cauchy-Schwartz.
First we show that $ B_{121}^2 <\infty$ then we show $B_{122}=o(1)$.
\begin{eqnarray*}
B_{121}^2&=&\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} \Vert\moneo\Vert^2 \leq\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} \sup_{\theta \in \Theta} \Vert \monet\Vert^2\\
&\leq &\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} K_7(V_{p\ell},Z_\ell)^2\leq\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_7(V_{p\ell},Z_\ell)^2={\rm E}[K_7(V,Z)^2]+o_{as}(1)<\infty
\end{eqnarray*}
\noindent where the second inequality follows from A6-(vi) and $0<1/(I_\ell-1)\leq1$.
Next we show that $B_{122}=o(1)$.
\begin{eqnarray*}
B_{122}\!\!\!\!\!&=&\!\!\!\!\Bigg\{\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} \!\! L \Bigg[\frac{1}{\gtilde \g}\Bigg(\frac{\G}{\g}\Big[\gtilde-\g\Big]^2\\
&&-\left[\Gtilde-\G\right]\Big[\gtilde-\g\Big]\Bigg)\Bigg]^2\Bigg\}^{\frac{1}{2}}\\
&< &\sqrt{L} \left\{ \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell(I_\ell-1)}\sum_{p=1}^{I_\ell}} \left[\kappa_1 \left(\kappa_2 O\left(\frac{1}{{r_g}^2}\right)-O\left(\frac{1}{r_G}\right)O\left(\frac{1}{r_g}\right)\right)\right]^2\right\}^{\frac{1}{2}}\\
&\leq&\sqrt{L}\kappa_1 \left[\kappa_2 O\left(\frac{1}{{r_g}^2}\right)-O\left(\frac{1}{r_G}\right)O\left(\frac{1}{r_g}\right)\right]=\kappa_1\left[\kappa_2
O\left(\frac{\sqrt{L}}{{r_g}^2}\right)-O\left(\frac{\sqrt{L}}{r_G
r_g}\right)\right]=o(1),
\end{eqnarray*}
\noindent where we have used:
$\Bigg\vert\displaystyle{ \frac{1}{\gtilde \g}}\Bigg\vert
<\kappa_1<\infty$, and $\Bigg\vert \displaystyle{\frac{\G}{\g}}\Bigg\vert<\kappa_2<\infty$ since the densities are bounded away from zero and $\gtilde \stackrel{a.s}{\longrightarrow} \g $ from Proposition B2 in GPV (2000);
\begin{eqnarray*}
O\left(\frac{1}{r_g^2}\right)&=&\Big\vert\gtilde-\g\Big
\vert^2=O\left(h_{1g}^{2R}+h_{2g}^{2R}+\frac{\log L}{L h_{1g} h_{2g}}\right),\\
O\left(\frac{1}{r_G}\right)O\left(\frac{1}{r_g}\right)&=&\Big\vert\Gtilde-\G\Big \vert
\Big\vert\displaystyle{\gtilde-\g}\Big
\vert\\
&=&O\left(h_{G}^{R+1}+\sqrt{\frac{\log L}{L h_{G}}}\quad\right)
O\left(h_{1g}^R+h_{2g}^R+\sqrt{\frac{\log L}{L h_{1g}h_{2g}}}\quad\right);
\end{eqnarray*}
and $\forall\ell, 0<1/(I_\ell-1)\leq1.$ Therefore, $B_{12}=o(1)=o(1)$ in (\ref{B11+B12}).
Next, we consider $B_2$ in (\ref{B_step3})
\subsection*{Step 3.2}
\begin{eqnarray}
\Vert B_2\Vert&\leq& \sqrt{L} \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}}
\Vert \moneo-\moneob \Vert\vert\hat{V}_{p\ell}- V_{p\ell}\vert \notag\\
&\leq& \sqrt{L}\displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_6(Z_\ell) \vert V_{p\ell}-V^*_{p\ell}\vert\vert\hat{V}_{p\ell}-V_{p\ell}\vert\leq\sqrt{L} \displaystyle{\frac{1}{L}\sum_{\ell=1}^L\frac{1}{I_\ell}\sum_{p=1}^{I_\ell}} K_6(Z_\ell) (\hat{V}_{p\ell}-V_{p\ell})^2 \notag\\
&\leq&\sqrt{L} \sup_{p,\ell}(\hat{V}_{p\ell}-V_{p\ell})^2 \displaystyle{\frac{1}{L}\sum_{\ell=1}^L} K_6(Z_\ell)=\sqrt{L} O_{as}\left(\frac{1}{r^2}\right) O_{as}(1)\leq O_{as}\left(\frac{\sqrt{L}}{r^2}\right) O_{as}(1)\notag\\
&=&O_{as}\left(\frac{L^{1/4}}{r}\right) O_{as}(1)=o_{as}(1),
\label{B_2}
\end{eqnarray}
\noindent where the second inequality follows from A6-(v), the third
from $\hat{V}_{p\ell}\leq V^*_{p\ell}\leq V_{p\ell}$ and the last from A4.AN.\end{proof}
\newpage
\bibliographystyle{econometrica}
|
\section{Introduction}
\label{intro}
For the physical interpretation of observables in heavy ion, proton/deuteron-heavy ion and (high-multiplicity) proton+proton collisions,
the existence of a sophisticated description of the multi-particle production mechanism and event-by-event fluctuations is essential.
In the high energy limit the color glass condensate (CGC) framework \cite{Gelis:2010nm} is the proper effective theory of quantum chromodynamics (QCD)
that provides such description.
One particular implementation of the CGC is the IP-Glasma model \cite{Schenke:2012wb,Schenke:2012hg}. It combines the IP-Sat dipole model \cite{Kowalski:2003hm,Kowalski:2007rw}, which parametrizes the impact parameter and $x$-dependence of the saturation scale, with the classical dynamics of produced gluon fields \cite{Krasnitz:1999wc,Krasnitz:2000gz,Lappi:2003bi}.
With its parameters constrained by inclusive and diffractive deeply inelastic scattering (DIS) data from e+p scattering at HERA, the IP-Glasma model correctly describes the bulk features of various collision systems over a wide range of energies \cite{Schenke:2013dpa}.
After briefly introducing the IP-Glasma model and the employed relativistic fluid dynamic simulation \textsc{music}, we
present results for the azimuthal anisotropy of charged hadrons produced in peripheral $\sqrt{s}=2.76\,{\rm TeV}$ Pb+Pb collisions, and $\sqrt{s}=5.02\,{\rm TeV}$ p+Pb collisions of comparable multiplicity. We compare to experimental data from the CMS collaboration \cite{Chatrchyan:2013nka}.
We then present predictions for the systematics of the transverse momentum dependent anisotropy coefficients $v_n(p_T)$ in p+Au, d+Au, and $^3$He+Au collisions at top RHIC energies.
\section{IP-Glasma + MUSIC}\label{sec:ipglasma}
The IP-Glasma model \cite{Schenke:2012wb,Schenke:2012hg} relates the DIS constrained nuclear dipole cross-sections to the initial classical dynamics of highly occupied gluon fields produced in a nuclear collision. Given an initial distribution of color charges in the high energy nuclear wave-functions, the strong multiple scatterings of gluon fields are computed by event-by-event solutions of Yang-Mills equations.
Both fluctuating distributions of nucleons in the nuclear wave-functions and intrinsic fluctuations of the color charge distributions are included. This results in ``lumpy'' transverse projections of the gluon field configurations that vary event to event. The scale of this lumpiness is given on average by the nuclear saturation scale $Q_s$ which corresponds to distance scales smaller than the nucleon size \cite{Kowalski:2007rw}.
The IP-Glasma model provides the initial conditions for fluid dynamic calculations at a given time $\tau_0$. The initial energy density $\varepsilon$ and flow velocities $u^\mu$, are extracted from the gluon fields' energy-momentum tensor $T^{\mu\nu}$ at every transverse position via the relation $u_\mu T^{\mu\nu} = \varepsilon u^\nu$.
In the results presented below, the viscous part of the energy momentum tensor is set to zero at the initial time of the fluid dynamic simulation. This is done because the gluon field strength tensor $T^{\mu\nu}$ is very anisotropic (the longitudinal pressure is approximately zero). A full 3+1 dimensional simulation including quantum fluctuations could provide a mechanism for isotropization via instabilities \cite{Berges:2013eia,Gelis:2013rba}.
We employ the viscous relativistic fluid dynamic simulation \textsc{music} \cite{Schenke:2010nt,Schenke:2010rr,Schenke:2011bn}, which is a 3+1 dimensional simulation. However, because the initial conditions from the IP-Glasma model are boost-invariant, it is used in its 2+1 dimensional mode.
\section{p+Pb collisions at the LHC}
\begin{figure}[th]
\begin{center}
\includegraphics[width=0.48\textwidth]{v2Cent-pPb-CMS}~\includegraphics[width=0.48\textwidth]{v3Cent-pPb-CMS}\\
\caption{Multiplicity dependence of the root-mean-square elliptic (left) and triangular (right) flow coefficients in Pb+Pb (open symbols) and p+Pb collisions (filled symbols) from the IP-Glasma+\textsc{music}
model (connected triangles) compared to experimental data by the CMS collaboration \cite{Chatrchyan:2013nka}. Figure from \cite{Schenke:2014zha}. \label{fig:vnCent-pPb-CMS}}\vspace{-0.5cm}
\end{center}
\end{figure}
It was shown in \cite{Schenke:2014zha} that experimental data from heavy ion collisions is well described by the IP-Glasma+\textsc{music} model out to fairly peripheral centrality bins. The natural question that arises is how the model does in describing data from p+A collisions that produce similar multiplicities.
Fig.\,\ref{fig:vnCent-pPb-CMS} shows the multiplicity dependence of $v_2$ and $v_3$ in peripheral Pb+Pb collisions and p+Pb collisions with comparable multiplicity. While the agreement with experimental data from the CMS collaboration \cite{Chatrchyan:2013nka} is fairly good in the Pb+Pb
case, for p+Pb collisions $v_2$ is under-predicted by approximately a factor of 4. In p+Pb $v_3$ agrees for the lower multiplicities studied, but has a rather flat multiplicity dependence and underestimates the experimental data at the higher multiplicities.
One reason for the disagreement could be that all initial state correlations that lead to an elliptic anisotropy\footnote{When including fluctuations initial state correlations should also contribute to odd harmonics.} \cite{Dusling:2012iga,Dusling:2012cg,Dusling:2012wy,Dusling:2013oia,Gyulassy:2014cfa,Dumitru:2014dra,McLerran:2014uka} are neglected.
Additionally, the description of the proton in the IP-Glasma model is over-simplified. The proton shape is approximated by a sphere, and any deviations from that are due to the small scale color charge fluctuations. Since the interaction region in p+Pb collisions is dominated by the shape of the smaller projectile, initial geometries in p+Pb collisions have very small eccentricities. This inevitably leads to small $v_n$ coefficients.
If there is indeed a large contribution to $v_n$ coefficients in p+Pb collisions from collective effects, our result indicates that the shape of the proton fluctuates significantly more than assumed in the IP-Glasma model. One could envision a description where the small $x$ gluon distributions are still concentrated around large $x$ valence quark positions, leading to much larger eccentricities and fluctuations.
In this case p+A collisions could be used to determine the shape and fluctuations of gluon distributions in the proton at high energies.
\section{p+Au, d+Au, and $^3$He+Au collisions at RHIC}
To determine whether final state collective effects provide the dominant contribution to the measured azimuthal anisotropy, RHIC is now
studying $^3$He+Au collisions that on average generate more triangular initial state configurations compared to p+Au or d+Au.
If collectivity is the physical explanation for the observed anisotropies, we expect a larger $v_3$ in $^3$He+Au collisions compared to p+Au and d+Au collisions at the same energy. To make this expectation more quantitative, we present predictions from the IP-Glasma+\textsc{music} framework.
For deuteron-gold collisions (d+Au) we compute the nucleon distribution in the deuteron using the Hulthen form of its wave function \cite{Miller:2007ri,Bzdak:2013zma}. For $^3$He, we use the same nucleon configurations as employed in \cite{Nagle:2013lja}. They are obtained from Green's
function Monte Carlo calculations using the AV18 + UIX model interaction \cite{Carlson:1997qn}.
For this comparative study we do not perform a detailed centrality selection, but instead sample the impact parameter $b$ between 0 and $2\,{\rm fm}$ in all systems. We then compute the initial state distribution of the energy density and flow velocity at time $\tau_0=0.5\,{\rm fm}/c$ and evolve the system using viscous fluid dynamics with $\eta/s=0.12$ until freeze-out at $T=135\,{\rm MeV}$.
We present typical configurations of the initial energy density distribution in the transverse plane and final results for the transverse momentum dependent azimuthal anisotropy coefficients $v_2$ to $v_5$ in Fig.\,\ref{fig:pAu-dAu-He3Au-withFlow}.
While we find very small values for $v_2$ through $v_5$ in p+Au collisions, the additional nucleons and their position fluctuations generate larger $v_2-v_4$ in d+Au and $^3$He+Au collisions. The odd harmonics $v_3$ and $v_5$ are noticeably larger in $^3$He+Au collisions compared to d+Au collisions. This qualitative prediction can be compared to future measurements at RHIC.
\begin{figure}
\begin{center}
\includegraphics*[width=0.96\textwidth]{pAu-dAu-He3Au-withFlow}\\
\caption{Typical configurations of the initial energy density distribution for p+Au, d+Au and $^3$He+Au collisions (upper panel).
Azimuthal anisotropy coefficients $v_2-v_5$ in the three collision systems (lower panel).\label{fig:pAu-dAu-He3Au-withFlow}}
\label{fig:generic}
\end{center}
\end{figure}
\section{Conclusions}
We have demonstrated that experimental results for $v_2$ and $v_3$ in proton-heavy ion collisions at the LHC are not well described by the IP-Glasma+\textsc{music} model. Reasons for this could be the neglected initial state correlations and/or the lack of a detailed description of the fluctuating subnucleonic structure of the proton. Our results for p+A collisions differ significantly from those in \cite{Bozek:2012gr,Werner:2013ipa,Kozlov:2014fqa,Bzdak:2014dia}, suggesting that the details of the initial shape in small systems are of paramount importance.
We predict an increase of both $v_3$ and $v_5$ in $^3$He+Au collisions compared to d+Au collisions, while the even harmonics are comparable in both systems. The detailed substructure of the nucleon is expected to be less important for the initial state geometry in these collisions compared to p+A collisions.
\vspace{0.25cm}
\emph{Acknowledgments}
We thank Jamie Nagle for providing us with the $^3$He nucleon configurations. This work was supported under DOE Contract No. DE-AC02-98CH10886. BPS acknowledges support from a DOE Office of Science Early Career Award. This research used resources of the National Energy Research Scientific Computing Center, which is supported by the DOE Office of Science under Contract No. DE-AC02-05CH11231.
\bibliographystyle{elsarticle-num-names}
|
\section*{}
The goal of statistical mechanics is to elucidate the relation between the microscopic world and the macroscopic world. {\textit{Equilibrium statistical mechanics}} assume the microscopic systems studied to be in equilibrium. In this course we will be concerned with {\textit{non-equilibrium statistical mechanics}} where time evolution is taken into account: our interest will not only be in the relation between the microscopic and the macroscopic scales in space but also in time.
By microscopic system we refer to molecules or atoms governed by the classical Newton's equations of motion. The question is then to understand how do these particles manage to organize themselves in such a way as to form a coherent structure on a large scale. The ``structure'' will be described by few variables (temperature, pressure \ldots) governed by autonomous equations (Euler's equations, Navier-Stokes's equation, heat equation \ldots). The microscopic specificities of the system will appear on this scale only through the thermodynamics (equation of state) and through the transport coefficients. Unfortunately, we are very far from understanding how to derive such macroscopic equations for physical relevant interactions.
One of the main ingredients that we need to obtain the macroscopic laws is that the particles, which evolve deterministically, have a behavior that one can consider almost as being random. The reason for this is that the dynamical system considered is expected to have a very sensitive dependence on the initial conditions and therefore is chaotic. This `` deterministic chaos'' is a poorly understood subject for systems with many degrees of freedom and even a precise consensual formulation is missing.
A first simplification to attack these problems consists in replacing the deterministic evolution of particles {\textit{ab initio}} by purely stochastic evolutions. Despite this simplification we notice that the derivation of the macroscopic evolution laws is far from being trivial. For example, we do not have any derivation of a system of hyperbolic conservation laws from a stochastic microscopic system after shocks. Nevertheless, since the pioneering work of Guo, Papanicolaou,Varadhan (\cite{GPV1}) and Yau (\cite{yau1}), important progresses have been performed in several well understood situations by the development of robust probabilistic and analytical methods (see \cite{KL} and \cite{Sp} for reviews).
In this course we will be mainly (but not only) interested in hybrid models for which the time evolution is governed by a combination of deterministic and stochastic dynamics. These systems have the advantage to be mathematically tractable and conserve some aspects of the underlying deterministic evolution. The stochastic noise has to be chosen in order to not destroy the main features of the Hamiltonian system that we perturb.
The central macroscopic equation of these lecture notes is the heat equation:
\begin{equation*}
\begin{cases}
\partial_t u = \partial_{x} ( D(u) \partial_x u), \quad x \in {\mathring U}, \quad t > 0,\\
\quad u(0, x) = u_{0} (x), \quad x \in U,\\
\quad u(t,x)=b(x), \quad x \in \partial U, \quad t>0.
\end{cases}
\end{equation*}
Here $u(t,x)$ is a function of the time $ t \ge 0$ and the space $x \in U \subset \RR^d$, $d \ge 1$, starting from the initial condition $u_0$ and subject to boundary conditions prescribed by the function $b$. The advantage of the heat equation with respect to other macroscopic equations such as the Euler or Navier-Stokes equations is that the notion of solution is very well understood. The dream would be to start from a system of $N \gg 1$ particles whose interactions are prescribed by Newton's laws and to show that in the large $N$ limit, the empirical energy converges in the diffusive time scale $t=\tau N^2$ to $u$ ($\tau$ is the microscopic time and $t$ the macroscopic time). In fact, this picture is expected to be valid only under suitable conditions and to fail for some low dimensional systems. In the case where the heat equation (or its variants) holds we say that the system has a normal behavior. Otherwise anomalous behavior occurs and the challenging question (even heuristically) is to know by what we shall replace the heat equation and what is the time scale over which we have to observe the system in order to see this macroscopic behavior ( \cite{BLRB}, \cite{Dhar},\cite{LLP1} for reviews).
The course is organized as follows. In Chapter \ref{ch:models} we introduce the models studied. Chapter \ref{ch:normal} is concerned with models which have a normal diffusive behavior. In Chapter \ref{ch:anomalous} we are interested in systems producing an anomalous diffusion. An important issue not discussed here is the effect of disorder on diffusion problems. In order to deal with lecture notes of a reasonable size, many of the proofs have been suppressed or only roughly presented.
\section[Models]{Chains of oscillators}
\label{ch:models}
\subsection{Chains of oscillators with bulk noise}
Chains of coupled oscillators are usual microscopic models of heat conduction in solids. Consider a finite box $\Lambda_N =\{ 1 ,\ldots, N\}^d \subset \ZZ^d$, $d\ge 1$, whose boundary $\partial \Lambda_N$ is defined as $\partial \Lambda_N = \{ x \notin \Lambda_N \, ; \, \exists y \in \Lambda_N, \; |x-y|=1\}$. Here $|\cdot|$ denotes the Euclidian norm in $\RR^d$ and $`` \cdot"$ the corresponding scalar product. Let us fix a nonnegative pair interaction potential $V$ and a pinning potential $W$ on $\RR$. The atoms are labeled by $x \in \Lambda_N$. The momentum of atom $x$ is $p_x \in {\RR}$ and its displacement from its equilibrium position {\footnote{We restrict us to the case where $q_x \in \RR^n$ with $n=1$ because the relevant dimension of the system is the dimension $d$ of the lattice. Most of the results stated in this manuscript can be generalized to the case $n \ge 1$.}} is ${q_x} \in \RR$. The energy ${\mc E}_x$ of the atom $x$ is the sum of the kinetic energy, the pinning energy and the interaction energy:
\begin{equation}
\label{eq:1-energy-site-x}
{\mc E}_x= \frac{|p_x|^2}2 + W(q_x) + \frac{1}{2} \sum_{\substack{|y-x|=1,\\ y\in \Lambda_N}} V(q_x - q_y).
\end{equation}
The Hamiltonian is given by
\begin{equation}
\label{eq:1-hamilt}
\mathcal H_N = \sum_{x \in \Lambda_N} {\mc E}_x + \partial \, {\mc H}_N
\end{equation}
where ${\partial} \, {\mc H}_N$ is the part of the Hamiltonian corresponding to the boundary conditions which are imposed.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=1]
\tikzfading[name=fade out, inner color=transparent!45,
outer color=transparent!100]
\fill [blue!20, path fading=fade out] (-1,-4) rectangle (12,1);
\tikzstyle{atome}=[draw,circle,ball color=white];
\foreach \y in {0,...,9}{
\node[atome] (\y) at (\y,0) {};
\node[atome] (\y+1) at (\y+1,0) {};
\node[draw,rectangle, minimum width=0.5cm, minimum height=0.5cm,fill=black!40] (p\y) at (\y,-1.5){};
}
\foreach \y in {0,...,9}{
\draw [decorate,decoration={snake,amplitude=2pt,segment length=5pt}] (\y) -- (\y+1);
\draw [decorate,decoration={snake,amplitude=2pt,segment length=5pt}] (\y) -- (p\y);
}
\node[draw,rectangle, minimum width=0.5cm, minimum height=0.5cm,fill=black!40] (plast) at (9+1,-1.5){};
\draw [decorate,decoration={snake,amplitude=2pt,segment length=5pt}] (9+1) -- (plast);
\end{tikzpicture}
\caption{A one-dimensional chain of pinned oscillators with free boundary conditions}
\end{figure}
We will consider the following cases:
\begin{itemize}
\item Periodic boundary conditions: we identify the site $1$ to the site $N$ and denote the corresponding box by ${\TT}_N$, the discrete torus of length $N$ (then $\partial {\mc H}_N =0$).
\item Free boundary conditions: this corresponds to the absence of boundary conditions, i.e. to $\partial {\mc H}_N =0$.
\item Fixed boundary conditions: introduce the positions $q_y=0$, $y\in \partial \Lambda_N$, of some fictive walls. We add to the Hamiltonian ${\mc H}_N$ a boundary term $\partial {\mc H}_N = \partial^{\rm{f}}\, {\mc H}_N$ given by
\begin{equation*}
\partial^{\rm{f}}\, {\mc H}_N = \sum_{\substack{|y-x|=1,\\ x \in \Lambda_N, y\in \partial\Lambda_N}} V(q_x - q_y)= \sum_{\substack{|y-x|=1,\\ x \in \Lambda_N, y\in \partial\Lambda_N}} V(q_x).
\end{equation*}
\item Forced boundary conditions: site ${\bf 1}=(1,\ldots,1)$ is in contact with a wall at position $q_0=0$ and each site $y \in \partial \Lambda_N \backslash \{ 0\}$ is driven by a constant force $\tau_y$. This results in a boundary term $\partial {\mc H}_N = \partial^{\rm{\tau}} \, {\mc H}_N$ given by
\begin{equation}
\partial^{\rm{\tau}} \, {\mc H}_N = \sum_{\substack{|y-x|=1,\\ x \in \Lambda_N, y\in \partial\Lambda_N}} V(q_x - q_y) - \sum_{ y\in \partial\Lambda_N \backslash \{ 0 \} } \tau_y q_y.
\end{equation}
\end{itemize}
The equations of motion of the atoms are
\begin{equation}
\label{eq:1-Newton}
{\dot {q_x}}= \partial_{p_x} {\mc H}_N, \quad {\dot p_x} = - \partial_{q_x} {\mc H}_N
\end{equation}
and the generator ${\mc A}_N$ of the system is given by the Liouville operator
\begin{equation*}
{\mc A}_N = \sum_{x \in {\Lambda}_N} \left\{ \partial_{p_x} {\mc H}_N \, \partial_{q_x} \, - \, \partial_{q_x} {\mc H}_N \, \partial_{q_x} \right\}.
\end{equation*}
It will be also useful to consider the chain of oscillators in infinite volume, i.e. replacing $\Lambda_N$ by $\ZZ^d$, $d\ge 1$, in the definitions above. The formal generator ${\mc A}_N$ is then denoted by ${\mc A}$. The dynamics can be defined for a large set of initial conditions if $V$ and $W$ do not behave too badly (\cite{LLL}, \cite{MPP}, \cite{BO-livre1}). We define the set $\Omega$ as the subset of ${\RR}^{\ZZ^d}$ given by
\begin{equation}
\Omega= \bigcap_{\alpha >0} \left\{ \xi \in \RR^{\ZZ^d} \, ; \, \sum_{x \in \ZZ^d} e^{-\alpha|x|} |\xi_x|^2 < +\infty \right\}
\end{equation}
and ${\tilde \Omega}= \Omega \times \Omega$. We equip $\Omega$ with its natural product topology
and its Borel $\sigma$-field and $\tilde \Omega$ by the corresponding product topology. For $X=\Omega$ or $X={\tilde \Omega}$, the set of Borel probability measures on $X$ is
denoted by ${\mathcal P} (X)$. A function $f: X \to \RR$ is said to be \emph{local}
if it depends of $\xi$ only through the coordinates $\{ \xi_x\, ; \, x \in \Lambda_f\}$,
$\Lambda_f$ being a finite box of $\ZZ$. We also introduce the sets $C_0^k(X)$, $k \ge 1$ (resp. $k=0$),
of bounded local functions on $X$ which are differentiable up to order $k$ with bounded partial derivatives (resp. continuous and bounded).
In the rest of the manuscript, apart from specific cases, we will assume that one of the following conditions hold:
\begin{itemize}
\item The potentials $V$ and $W$ have bounded second derivatives. Then the infinite dynamics $(\omega(t))_{t \ge 0}$ can be defined for any initial condition $\omega^0=(\bq^0, \bp^0) \in {\tilde \Omega}$. Moreover ${\tilde \Omega}$ is invariant by the dynamics. This defines a semigroup $(P_t)_{t \ge 0}$ on $C_0^0 ({\tilde \Omega})$ and the Chapman-Kolmogorov equations
\begin{equation}
(P_t f)(\omega) - f(\omega) = \int_0^t (P_s {\mc A} f ) (\omega) \, ds \; = \; \int_0^t ({\mc A} P_s f ) (\omega) \, ds
\end{equation}
are valid for any $f \in C_0^1 ({\tilde \Omega})$.
\item The potential $W=0$ and the interaction potential $V$ has a second derivative uniformly bounded from above and below. It is more convenient to go over the deformation field $\eta_{(x,y)}= q_y -q_x$, $|x-y|=1$, which by construction is constrained to have zero curl. In $d=1$ we will denote $\eta_{(x-1,x)}= q_{x}- q_{x-1}$ by $r_x$. The dynamics (\ref{eq:1-Newton}) can be read as a dynamics for the deformation field and the momenta. Given say $q_0$, the scalar field $\bq=\{ q_x\}_{x \in \ZZ^d}$ can be reconstructed from $\eta$. In the sequel, when $W=0$, we will use these coordinates without further mention. The dynamics for the coordinates $\omega=(\eta, \bp)= (\eta_{(x,x+\be)}, p_x)_{|\be|=1, x \in \ZZ}$ can be defined if the initial condition satisfies $\omega^0 \in {\tilde \Omega}$. Moreover the set ${\tilde \Omega}$ is invariant by the dynamics. This defines a semigroup $(P_t)_{t \ge 0}$ on $C_0^0 ({\tilde \Omega})$ and the Chapman-Kolmogorov equations
\begin{equation}
\label{eq:CK1}
(P_t f)(\omega) - f(\omega) = \int_0^t (P_s {\mc A} f ) (\omega) \, ds \; = \; \int_0^t ({\mc A} P_s f ) (\omega) \, ds
\end{equation}
are valid for any $f \in C_0^1 ({\tilde \Omega})$.{\footnote{The generator ${\mc A}$ has to be written in terms of the deformation field.}}
\end{itemize}
Let us first consider the problem related to the characterization of equilibrium states. For simplicity we take the finite volume dynamics with periodic boundary conditions. Then it is easy to see that the system conserves one or two physical quantities depending on whether the chain is pinned or not. The total energy ${\mc H}_N$ is always conserved. If $W=0$ the system is translation invariant and the total momentum $\sum_{x} p_x$ is also conserved. Notice that because of the periodic boundary conditions the sum of the deformation field $\sum_{x} \eta_{(x,x+e_i)}$ is automatically fixed equal to $0$ for any $i=1, \ldots, d$.
Liouville's Theorem implies that the uniform measure $\lambda^N$ on the manifold $\Sigma^N$ composed of the configurations with a fixed total energy (and possibly a fixed total momentum) is invariant for the dynamics. The micro canonical ensemble is defined as the probability measure $\lambda^N$. The dynamics restricted to $\Sigma^N$ is not necessarily ergodic. Two examples for which one can show it is not the case are the harmonic lattice ($V$ and $W$ quadratic) and the Toda lattice ($d=1$, $W=0$, $V ( r )=e^{-r} -1 +r $) which is a completely integrable system (\cite{Toda}). In fact what is really needed for our purpose is not the ergodicity of the finite dynamics but of the infinite dynamics. We expect that even if the finite dynamics are never ergodic the fraction of $\Sigma^N$ corresponding to non ergodic behavior decreases as $N$ increases, and probably disappears as $N=\infty$ (apart from very peculiar cases). Therefore a good notion of ergodicity has to be stated for infinite dynamics. The definition of a conserved quantity is not straightforward in infinite volume (the total energy of the infinite chain is usually equal to $+ \infty$). To give a precise definition we will use the notion of space-time invariant probability measures for the infinite dynamics defined above.
The infinite volume Gibbs grand canonical ensembles are such probability measures. They form a set of probability measures indexed by one (pinned chains) or $d+2$ (unpinned chains) parameters and are defined by the so-called Dobrushin-Landford-Ruelle's equations. To avoid a long discussion we just give a formal definition (see e.g. \cite{Giac1} for a detailed study).
\begin{itemize}
\item Pinned chains ($W \ne 0$): the infinite volume Gibbs grand canonical ensemble $\mu_{\beta}$ with inverse temperature $\beta>0$ is the probability measure on ${\tilde \Omega}$ whose density with respect to the Lebesgue measure is
\begin{equation*}
Z^{-1} (\beta) \exp \left( - \beta \sum_{x \in \ZZ^d} {\mc E}_x \right).
\end{equation*}
\item Unpinned chains ($W=0$): the infinite volume Gibbs grand canonical ensemble {\footnote{They are defined with respect to the gradient fields $\eta_{(x,y)}$. It would be more coherent to call them {\textit{gradient Gibbs measures}}. }} $\mu_{\beta, {\bar p}, \tau}$ with inverse temperature $\beta>0$, average momentum ${\bar p} \in \RR$ and tension $\tau=\beta^{-1} \lambda \in \RR^d$ is the probability measure on ${\tilde \Omega}$ whose density with respect to the Lebesgue measure is
\begin{equation}
\label{eq.Gibbsmeasures}
Z^{-1} (\beta, {\bar p}, \tau) \; \exp \left( - \beta \, \sum_{x \in \ZZ^d} \{ {\mc E}_x - {\bar p}\, p_x - \sum_{i=1}^d \tau_i \, \eta _{(x, x+e_i)}\} \right).
\end{equation}
\end{itemize}
Observe that in the one dimensional unpinned case we have simply product measures and that the tension $\tau$ is equal to the average of $V'(r_x)$.
Fix an arbitrary Gibbs grand canonical ensemble $\mu$. A probability measure $\nu$ is said to be $\mu$-regular if for any finite box $\Lambda \subset \ZZ^d$ whose cardinal is denoted by $|\Lambda|$, the relative entropy of $\nu |_{\Lambda}$ w.r.t. $\mu |_{\Lambda}$ is bounded above by $C | \Lambda|$ for a constant $C$ independent of $\Lambda$. We recall that the relative entropy $H(\nu|\mu)$ of $\nu \in {\mc P} (X)$ with respect to
$\mu \in {\mc P} (X)$, $X$ being a probability space, is defined as
\begin{equation}
\label{eq:ent009}
H(\nu | \mu) = \sup_{\phi} \left\{ \int \phi \, d\nu -
\log \left( \int e^{\phi} \, d\mu \right) \right\},
\end{equation}
where the supremum is carried over all bounded measurable functions $\phi$ on $X$.
For any arbitrary Gibbs grand canonical ensembles $\mu$ and $\mu'$, $\mu$ is $\mu'$-regular and $\mu'$ is $\mu$-regular. Therefore $\nu$ is $\mu$-regular is equivalent to $\nu$ is $\mu'$-regular and we simply say that $\nu$ is regular.
A notion of ergodicity for infinite dynamics which is suitable to derive rigorously large scale limits of interacting particle systems is the following.
\begin{definition}[Macro-Ergodicity] {\footnote{The name has been proposed by S. Goldstein.}} We say that the dynamics generated by ${\mc A}$ is {\textit{macro-ergodic}} if and only if the only space-time invariant {\footnote{Observe that a probability measure $\nu$ is time invariant for the infinite dynamics if and only if $\int {\mc A} f \, d\nu =0$ for any $f \in C_0^1 ({\tilde \Omega})$. This is a consequence of the Chapman-Kolmogorov equations (\ref{eq:CK1}).}} regular measures $\nu$ for ${\mc A}$ are mixtures (i.e. generalized convex combinations) of Gibbs grand canonical ensembles.
\end{definition}
If the microscopic dynamics is macro-ergodic, then, by using the relative entropy method developed in \cite{OVY}, we can derive the hydrodynamic equations {\footnote{The notion of hydrodynamic limits is detailed in Section \ref{subsec:hl-vf} and Section \ref{sec:hydroLimEuler}.}} in the Euler time scale of the chain before the appearance of the shocks, at least in $d=1$ (\cite{BO-livre1}). These limits form a triplet of compressible Euler equations (for energy ${\mf e}$, momentum ${\mf p}$ and deformation ${\mf r}$) of the form
\begin{equation}
\label{eq:Euler-general0}
\begin{cases}
\partial_t {\mf r} = \partial_q {\mf p}\\
\partial_t {\mf p} = \partial_q {\mf \tau}\\
\partial_t {\mf e} = \partial_q ({\mf p} \tau)
\end{cases}
\end{equation}
where the pressure $\tau:= \tau (\mf r, {\mf e} -\tfrac{{\mf p}^2}{2})$ is a suitable thermodynamic function depending on the potential $V$. A highly challenging open question is to extend these results after the shocks. The proof can be adapted to take into account the presence of mechanical boundary conditions (\cite{Even-Olla}).
We do not claim that the macro-ergodicity is a necessary condition to get Euler equations for purely Hamiltonian systems. We could imagine that weaker or different conditions are sufficient but in the actual state of the art the macro-ergodicity is a clear and simple mathematical statement of what we could require from deterministic systems in order to derive Euler equations rigorously. We refer the interested reader to \cite{Bricmont-Chance} and \cite{Szasz-ergodic} for interesting discussions about the role of ergodicity in statistical mechanics.
\subsubsection{Conserving noises}
In \cite{FFL}, Fritz, Funaki and Lebowitz prove a weak form of macro-ergodicity for a chain of anharmonic oscillators under generic assumptions on the potentials $V$ and $W$ that we do not specify here (see \cite{FFL}).
\begin{theorem}[\cite{FFL}] {\footnote{The proof given in \cite{FFL} assumes $W \ne 0$ but it can be adapted to the unpinned one dimensional case (see {\cite{BO-livre1}}). It would be interesting to extend this theorem to the general unpinned case.}} Consider the pinned chain $W \ne 0$ generated by ${\mc A}$ or an unpinned chain $W=0$ in $d=1$. The only regular time and space invariant measures for ${\mc A}$ which are such that conditionally to the positions configuration $\bq:=\{q_x \, ;\, x \in \ZZ^d\}$ the law of the momenta $\bp:=\{ p_x \, ;\, x \in \ZZ^d\}$ is exchangeable are given by mixtures of Gibbs grand canonical ensembles.
\end{theorem}
They also proposed to perturb the dynamics by a stochastic noise that consists in exchanging at random exponential times, independently for each pair of nearest neighbors site $x,y \in \ZZ^d$, $|x-y|=1$, the momenta $p_x$ and $p_y$. The formal generator ${\mc L}$ of this dynamics, that we will call the {\textit{stochastic energy-momentum conserving dynamics}}, is given by ${\mc L} ={\mc A} +\gamma {\mc S}$, $\gamma >0$, where ${\mc A}$ is the Liouville operator and ${\mc S}$ is defined for any local function $f: {\tilde \Omega} \to \RR$ by
\begin{equation}
\label{eq:cons-noise-exch-mom}
({\mc S} f)(\bq,\bp) = \sum_{\substack{x,y \in \ZZ^d\\ |x-y|=1}} \left[ f(\bq,\bp^{x,y}) -f(\bq, \bp) \right].
\end{equation}
Here the momenta configuration $\bp^{x,y}$ is the configuration obtained from $\bp$ by exchanging $p_x$ with $p_y$. The previous discussion about existence of the dynamics on ${\tilde \Omega}$ for the deterministic case and its relation with its formal generator is also valid for this dynamics and the other dynamics defined in this section.
With some non-trivial entropy estimates we get the following result.
\begin{theorem}[\cite{FFL}]
\label{th:1:ergexc}
Consider the pinned ($W\ne 0$) or the one-dimensional unpinned ($W \ne 0$) stochastic energy-momentum conserving dynamics. The only regular time and space invariant measures for these dynamics are given by mixtures of Gibbs grand canonical ensembles, i.e. the stochastic energy-momentum conserving dynamics is macro-ergodic.
\end{theorem}
Consequently the stochastic energy-momentum conserving dynamics is macro-ergodic. By using the relative entropy method developed in \cite{OVY}, one can show it has in the Euler time scale and before the appearance of the shocks the same hydrodynamics (\ref{eq:Euler-general0}) as the deterministic model. This is because the noise has some macroscopic effects only in the diffusive time scale (\cite{BO-livre1}).
We consider now a different stochastic perturbation. Let us define the flipping operator $\sigma_x: \bp \in \Omega \to {\bp}^x \in \Omega$ where ${\bp}^x$ is the configuration such that $({\bp}^x)_z =p_z$ for $z \ne x$ and $({\bp}^x)_x =-p_x$. In \cite{FFL} is also proved that the only time-space regular stationary measures for the Liouville operator ${\mc A}$ such that conditionally to the positions the momenta distribution is invariant by any flipping operator $\sigma_{x}$ are mixtures of Gibbs grand canonical ensembles with zero momentum average. Then we consider the dynamics on ${\tilde \Omega}$ generated by ${\mc L} ={\mc A} +\gamma {\mc S}$, $\gamma >0$, with ${\mc S}$ the noise defined by
\begin{equation}
\label{eq:cons-noise-flip}
({\mc S} f)(\bq, \bp)= \cfrac{1}{2} \sum_{x \in \ZZ^d} \left[ f(\bq, \bp^x) -f(\bq,\bp)\right]
\end{equation}
for any local function $f: {\tilde \Omega} \to \RR$. This dynamics conserves the energy and the deformation of the lattice but destroys all the other conserved quantities. We call this system the \textit{velocity-flip model} (sometimes the \textit{stochastic energy conserving model}).
\begin{theorem}[\cite{FFL}]
\label{th:1:ergflip}
Consider the pinned $d$-dimensional velocity-flip model or the one-dimensional unpinned velocity-flip model. The only regular time and space invariant measures are given by mixtures of Gibbs grand canonical ensembles. In other words the velocity-flip model is macro-ergodic.
\end{theorem}
Since the velocity flip-model does not conserve the momentum its Gibbs invariant measures are given by (\ref{eq.Gibbsmeasures}) with ${\bar p}=0$. In particular the average currents with respect to theses measures is zero. Therefore assuming propagation of local equilibrium in the Euler time scale we get that it has trivial hydrodynamics in this time scale: initial profile of energy does not evolve. This is only in the diffusive scale that an evolution should take place.
\subsubsection{NESS of chains of oscillators perturbed by an energy conserving noise}
The models defined in the previous sections can also be considered in a non-equilibrium stationary state (NESS) by letting them in contact with thermal baths at different temperatures and imposing various mechanical boundary conditions. Let us only give some details for the NESS of the one-dimensional velocity-flip model.
Consider a chain of $N$ unpinned oscillators where the particle $1$ (resp. $N$) is subject to a constant force $\tau_\ell$ (resp. $\tau_r$). Moreover we assume that the particle $1$ (reps. $N$) is in contact with a Langevin thermal bath at temperature $T_{\ell}$ (resp. $T_r$). The generator ${\mc L}_N$ of the dynamics on the phase space $\Omega_N=\RR^{N-1} \times \RR^N$ is given by
\begin{equation}
\label{eq:1:ness-gen}
{\mc L}_N= {\mc A}^{\tau_\ell, \tau_r}_N +\gamma {\mc S}_N + \gamma_\ell {\mc B}_{1,T_\ell} + \gamma_r {\mc B}_{N,T_r}, \quad \gamma>0,
\end{equation}
where ${\mc A}^{\tau_\ell, \tau_r}_{N}$ is the Liouville operator,
${\mc B}_{j,T}$ the generator of the Langevin bath at temperature $T$
acting on the $j$--th particle and ${\mc S}_N$ the generator of the
noise. The strength of noise and thermostats are regulated by
$\gamma$, $\gamma_{\ell}$ and $\gamma_r$ respectively. The Liouville operator is
defined by
\begin{equation}
\label{eq:ss51}
\begin{split}
{\mathcal A }^{\tau_\ell, \tau_r}_{N}=
\sum_{x=2}^{N} \left(p_{x} - p_{x-1}\right) \partial_{r_x} +
\sum_{x=2}^{N-1}\left(V'(r_{x+1}) - V'(r_{x})\right)
\partial_{p_{x}}\\
- \left(\tau_\ell - V'(r_2)\right) \partial_{p_1}
+ \left(\tau_r - V'(r_N)\right) \partial_{p_N}.
\end{split}
\end{equation}
The generators of the thermostats are given by
\begin{equation}
\label{eq:ss49}
{\mc B}_{j,T} = T \partial_{p_j}^2 - p_j \partial_{p_j}.
\end{equation}
The noise corresponds to independent velocity change of sign, i.e.
\begin{equation}
\label{eq:ss41}
({\mc S}_N f ) (\br,\bp) = \frac{1}{2} \sum_{x=2}^{N-1} \left(f(\br,\bp^{x}) - f(\br,\bp)\right), \quad f: \Omega_N \to \RR .
\end{equation}
We will also consider the case where the chain has fixed boundary conditions.
\begin{proposition}[\cite{BO1}, \cite{BO2},\cite{Car}]
Consider a finite chain of pinned or unpinned oscillators with fix, free or forced boundary conditions in contact with two thermal baths at different temperatures and perturbed by one of the energy conserving noises defined above. Then, there exists a unique non-equilibrium stationary state for this dynamics which is absolutely continuous w.r.t. Lebesgue measure.
\end{proposition}
\begin{proof}
The proof of the existence of the invariant state can be obtained from the knowledge of a suitable Liapounov function. To prove the uniqueness of the invariant measure it is sufficient to prove that the dynamics is irreducible and has the strong-Feller property. Some hypoellypticity, control theory and conditioning arguments are used to achieve this goal.
\cqfd
\end{proof}
\subsection{Simplified perturbed Hamiltonian systems}
\label{sec:intro-phd}
Introducing a noise into the deterministic dynamics help us to solve some ergodicity problems. Nevertheless, as we will see, several challenging problems remain open for chains of oscillators perturbed by a conservative noise. In \cite{BS} we proposed to simplify still these models and the main message addressed there is that the models introduced in \cite{BS} have qualitatively the same behaviors as the unpinned chains. For simplicity we define only the dynamics in infinite volume.
Let $U$ and $V$ be two potentials on $\RR$ and consider the Hamiltonian system $(\omega (t) )_{t \ge 0} = ( \, {\br} (t) , {\bp} (t) \,)_{t \ge 0}$ described by the equations of motion
\begin{equation}
\label{eq:1:generaldynamics}
\frac{dp_x}{dt} = V'(r_{x+1}) -V'(r_x),
\qquad \frac{dr_x}{dt} = U' (p_x) -U' (p_{x-1}),
\qquad x \in \ZZ,
\end{equation}
where $p_x$ is the momentum of particle $x$, $q_x$ its position and $r_x=q_{x} -q_{x-1}$ the ``deformation''. Standard chains of oscillators are recovered for a quadratic kinetic energy $U(p)=p^2 /2$. The dynamics conserves (at least) three physical quantities: the total momentum $\sum_{x} p_{x}$, the total deformation $ \sum_{x} r_{x}$ and the total energy $\sum_x {\mc E_x}$ with ${\mc E}_x= V(r_x) + U(p_x)$. Consequently, every Gibbs grand canonical ensemble ${\nu}_{\beta,\lambda, \lambda'}$ defined by
\begin{equation}
\label{eq:invmeas007}
d{\nu}_{\beta,\lambda,\lambda'} (\eta) = \prod_{x \in \ZZ} {\mc Z} (\beta,\lambda,\lambda')^{-1}
\exp\left\{ -\beta {\mc E}_x -\lambda p_{x} -\lambda' r_{x} \right\} \, d r_x \, d p_x
\end{equation}
is invariant under the evolution. To simplify we assume that the potentials $U$ and $V$ are smooth potentials with second derivatives bounded by below and from above.
To overcome our ignorance about macro-ergodicity of the dynamics, as before, we add a stochastic conserving perturbation. In the general case $U \ne V$, the Hamiltonian dynamics can be perturbed by the energy-momentum conserving noise acting on the velocities (as proposed in \cite{FFL}) but conserving the three physical invariants mentioned above. Then the infinite volume dynamics can be defined on the state space ${\tilde \Omega}$. Its generator ${\mc L}$ is given by ${\mc L} = {\mc A} +\gamma {\mc S}$, $\gamma >0$, where
\begin{equation}
\label{eq:1:ASdynum}
\begin{split}
&({\mc A} f) (\br ,\bp) =\sum_{x \in \ZZ} \left\{\, (V'(r_{x+1}) -V'(r_x)) \partial_{p_x} f\, + \, (U' (p_x) -U' (p_{x-1})) \partial_{r_x} f \right\}\, (\br,\bp)\\
& ({\mc S} f) = \sum_{x \in \ZZ} \left[ f(\br, \bp^{x,x+1}) - f (\br, \bp) \right]
\end{split}
\end{equation}
for any $f \in C_0^1 ({\tilde \Omega})$.
\begin{theorem}[\cite{BS}]
Assume that the potentials $U$ and $V$ are smooth potentials with second derivatives bounded by below and from above. The dynamics generated by ${\mc L}={\mc A} +\gamma {\mc S}$ with $\gamma>0$ and ${\mc A}, {\mc S}$ given by (\ref{eq:1:ASdynum}) is macro-ergodic. Consequently, before the appearance of the shocks, in the Euler time scale, the hydrodynamic limits are given by a triplet of compressible Euler equations.
\end{theorem}
Our motivation being to simplify as much as possible the dynamics considered in~\cite{BBO1,BBO2} without destroying the anomalous behavior of the energy diffusion, we mainly focus on the symmetric case $U=V$. Then the $\bp$'s and $\br$'s play a symmetric role so there is no reason that momentum conservation is more important than deformation conservation. We propose thus to add a noise conserving only the energy and $\sum_{x} [ r_x + p_x]$. It is more convenient to use the variables $\{ \eta_x \, ; \, x \in \ZZ \} \in \RR^{\ZZ}$ defined by $\eta_{2x}= p_x$ and $\eta_{2x -1} =r_x$ so that (\ref{eq:1:generaldynamics}) becomes
\begin{equation}
\label{eq:1:dyneq}
d\eta_x = \left[ V' (\eta_{x+1}) -V' (\eta_{x-1}) \right] dt, \quad x \in \ZZ.
\end{equation}
We might also interpret the dynamics for the $\eta$'s as the dynamics of an interface whose height (resp. energy) at site $x$ is $\eta_x$ (resp. $V(\eta_x)$). It is then quite natural to call the quantity $\sum_x \eta_x$ the ``volume".
Hence, we introduce the so-called {\textit{stochastic energy-volume conserving dynamics}}, which is still described by (\ref{eq:1:dyneq}) between random exponential times where two nearest neighbors heights $\eta_x$ and $\eta_{x+1}$ are exchanged. Observe that in the momenta-deformation picture this noise is less degenerate than the momenta exchange noise since exchange between momenta and positions is now allowed. The generator ${\mc L}$ of the infinite volume dynamics, well defined on the state space $\Omega$, is given by ${\mc L} ={\mc A} +\gamma {\mc S}$, $\gamma>0$, where for any $f \in C_0^1 (\Omega)$,
\begin{equation}
\label{eq:1:ASdynum2}
\begin{split}
& ({\mc A}f)(\eta) = \sum_{x \in \ZZ} \left[ V' (\eta_{x+1}) -V' (\eta_{x-1}) \right] (\partial_{\eta_x} f )(\eta),\\
& ({\mc S}f)(\eta) = \sum_{x \in \ZZ} \left[ f(\eta^{x,x+1}) -f(\eta) \right].
\end{split}
\end{equation}
The noise still conserves the total energy and the total volume but destroys the conservation of momentum and deformation. Therefore, only two quantities are conserved and the invariant Gibbs grand canonical measures of the stochastic dynamics correspond to the choice $\lambda=\lambda'$ in~(\ref{eq:invmeas007}). We denote ${\nu}_{\beta,\lambda,\lambda}$ (resp. ${\mc Z}(\beta,\lambda,\lambda)$) by $\mu_{\beta, \lambda}$ (resp. $Z (\beta,\lambda)$).
\section{Normal diffusion}
\label{ch:normal}
Normal diffusion of energy in purely deterministic homogeneous chains of oscillators is expected to hold in high dimension ($d\ge 3$) or if momentum is not conserved, i.e. in the presence of a pinning potential. The problem of anomalous diffusion will be discussed in the next chapter. In this chapter we consider the case of normal diffusion.
The first step to show such normal behavior is to prove that the transport coefficient, the thermal conductivity, is well defined. Once it has been achieved, the following non-equilibrium problems can be considered:
\begin{itemize}
\item Hydrodynamic limits in the diffusive time scale $t \ve^{-2}$, $\ve$ being the scaling parameter: if the system has trivial hydrodynamics in the time scale $t \ve^{-1}$, i.e. if momentum is not conserved, we would like to show that in the diffusive time scale, the macroscopic energy profile evolves according to a diffusion equation. If the system has non-trivial hydrodynamics given by the Euler equations in the hyperbolic scaling (i.e. if momentum is conserved), in the diffusive time scale, we would like to derive the incompressible Navier-Stokes equations. These would be obtained by starting with some initial momentum macroscopic profile of order ${\mc O}(\ve)$ but an energy profile of order ${\mc O} (1)$.
\item Validity of Fourier's law: we consider the NESS of the system in contact at the boundaries with thermal baths at different temperatures. Fourier's law expresses that the average of the energy current in the NESS is proportional to the gradient of the local temperature. The proportionality coefficient is called the thermal conductivity.
\end{itemize}
Assume for simplicity that $d=1$ and that the energy is the only conserved quantity. The corresponding microscopic current, denoted by $j^{e}_{x,x+1}$, is defined by the local energy conservation law
\begin{equation}
{\mc L} {\mc E}_x = - \nabla j^{e}_{x-1,x}
\end{equation}
where ${\mc L}$ is the generator of the infinite dynamics under investigation and $\nabla$ is the discrete gradient defined for any $(u_x)_x \in \RR^{\ZZ}$ by $\nabla u_x = u_{x+1} -u_x$. In the current state of the art, in all the problems mentioned above, the usual approach consists to prove that there exist functions $\varphi_x =\theta_x \varphi_0$ and $h_x =\theta_x h_0$ (actually only approximate solutions are needed) such that the following decomposition
\begin{equation}
\label{eq:2:fluc-diss}
j^e_{x,x+1} = \nabla \varphi_x +{\mc L} h_x
\end{equation}
holds. Here $\theta_x$ denotes the shift by $x \in \ZZ^d$. Equation (\ref{eq:2:fluc-diss}) is called a {\textit{microscopic fluctuation-dissipation equation}}. Then, taking arbitrary large integer $\ell \ge 1$, by using a multi-scale analysis we replace the block averaged function $\tfrac{1}{2\ell +1} \sum_{|y-x| \le \ell} \nabla \varphi_y$ by $D({\mc E}^{\ell}_x) \nabla {\mc E}^\ell_x$ where the function $D$ is identified to a diffusion coefficient which depends on the empirical energy ${\mc E}_x^\ell = \tfrac{1}{2\ell +1} \sum_{|y-x| \le \ell} {\mc E}_y$ in the mesoscopic box of length $(2\ell +1)$ centered around $x$. Intuitively, ${\mc L} h_x$ represents rapid fluctuation (integrated in time, it is a martingale) and the term $\nabla \varphi_x$ represents the dissipation. Gradient models are systems for which the current is equal to the gradient of a function ($h_x=0$ with the previous notations).
There are at least two reasons for which the problems listed above are difficult:
\begin{itemize}
\item The existence of a microscopic fluctuation-dissipation equation has been given for the first time for reversible systems. It has been extended to asymmetric systems satisfying a {\textit{sector condition}}. Roughly speaking this last condition means that the antisymmetric part of the generator is a bounded perturbation of the symmetric part of the generator {\footnote{The antisymmetric (resp. symmetric) part of the generator ${\mc L}$ is given by $\tfrac{{\mc L}-{\mc L}^*}{2}$ (resp. $\tfrac{{\mc L}+{\mc L}^*}{2}$) where ${\mc L}^*$ is the adjoint of ${\mc L}$ in $\bb L^2 (\mu)$, $\mu$ being any Gibbs grand canonical measure. For the models considered in this course, the antisymmetric part is ${\mc A}$ and due to the deterministic dynamics, the symmetric part is ${\mc S}$ and due to the noise. }}. Later this condition has been relaxed into the so-called {\textit{graded sector condition}}: there exists a gradation of the space where the generator is defined and the asymmetric part is bounded by the symmetric part on each graded part (see \cite{KLO-book}, \cite{HIToth} and references therein). The Hamiltonian systems perturbed by a noise are non-reversible and since the noise (the symmetric part of the generator) is very degenerate, none of these conditions hold.
\item The system evolves in a non compact space and one needs to show that energy cannot concentrate on a site. This technical problem turns out to be difficult since no general techniques are available. For deterministic nonlinear chains the bounds on the average energy moments are usually polynomial in the size $N$ of the system. Typically we need bounds of order one with respect to $N$.
\end{itemize}
\subsection{Anharmonic chain with velocity-flip noise}
\subsubsection{Linear response theory: Green-Kubo formula}
\label{subsec:lrGK}
The Green-Kubo formula is one of the most important formulas of non-equilibrium statistical mechanics. In the two problems mentioned in the introduction of the chapter (hydrodynamic limits and Fourier's law) the limiting objects are defined via some macroscopic coefficients which can be expressed by a Green-Kubo formula. The latter is a formal expression and showing that it is indeed well defined is a difficult problem. It is usually introduced in the context of the linear response theory that we describe below.
Consider a one dimensional unpinned chain of $N$ harmonic oscillators with forced boundary conditions and perturbed by the velocity-flip noise. The two external constant forces are denoted by $\tau_\ell$ and $\tau_r$. Furthermore on the boundary particles $1$ and $N$, Langevin thermostats are acting at different temperature $T_\ell=\beta_{\ell}^{-1}$ and $T_r=\beta_r^{-1}$. The generator ${\mc L}_N$ of the dynamics is given by (\ref{eq:1:ness-gen}) and we denote the unique non-equilibrium stationary state by $\mu_{\rm{ss}}$. The expectation w.r.t. $\mu_{\rm{ss}}$ is denoted by $\langle \cdot \rangle_{ss}$.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=0.7]
\filldraw[gray!50,left color=red!40, right color=blue!40,rounded corners=10pt]
(0,-3)--(13,-3)--(13,3)--(-0,3)--(-3,3)--(-3,-3)--(0,-3) ;
\tikzstyle{atome}=[draw,circle,ball color=white];
\foreach \y in {0,...,9}{
\node[atome] (\y) at (\y,0) {};
\node[atome] (\y+1) at (\y+1,0) {};
}
\foreach \y in {0,...,9}{
\draw [decorate,decoration={snake,amplitude=2pt,segment length=5pt}] (\y) -- (\y+1);
}
\node [draw,rectangle, minimum width=0.5cm, minimum height=0.5cm,fill=red!40] (bath-left) at (0,-2){$T_\ell$};
\node [draw,rectangle, minimum width=0.5cm, minimum height=0.5cm,fill=blue!40] (bath-right) at (10,-2){$T_r$};
\draw [decorate,decoration={snake,amplitude=0.5pt,segment length=5pt}] (bath-left) -- (0);
\draw [decorate,decoration={snake,amplitude=0.5pt,segment length=5pt}] (bath-right) -- (9+1);
\node (force-right) at (12,0){$\tau_r$};
\draw [->,ultra thick] (9+1)--(force-right);
\node (force-left) at (-2,0){$\tau_\ell$};
\draw [->,ultra thick] (0)--(force-left);
\end{tikzpicture}
\caption{The unpinned chain with boundary thermal reservoirs and forced boundary conditions.}
\end{figure}
The energy {\footnote{The definition of the energy is slightly modified w.r.t. (\ref{eq:1-energy-site-x}). It is more convenient since the energies are then independent random variables in the Gibbs grand canonical ensemble.}} of atom $x$ is defined by
\begin{equation*}
{\mc E}_1 =\cfrac{p_1^2}{2}, \quad {\mc E}_x =\cfrac{p_x^2}{2} + V(r_x), \quad x=2, \ldots,N.
\end{equation*}
The local conservation of
energy is expressed by the microscopic continuity equation
\begin{equation*}
{\mc L}_N ({\mc E}_x) = -\nabla j^e_{x-1,x}, \quad x=1, \ldots, N,
\end{equation*}
where the energy current $j^e_{x,x+1}$ from site $x$ to site $x+1$ is
given by
\begin{equation}
\begin{split}
j^e_{0,1} &= -\tau_{\ell} p_1 + \gamma_\ell (T_{\ell} -p_1^2), \\
j^e_{N,N+1} &=-\tau_r p_N -\gamma_r (T_r -p_N^2), \\
j^e_{x,x+1} &= - p_x V' (r_{x+1}), \, x=1,\ldots,N-1.
\end{split}
\end{equation}
The energy current $j^e_{0,1}$ (and similarly for $j^e_{N,N+1}$) is
composed of two terms: the term $-\tau_\ell p_1$ corresponds to the
work done on the first particle by the linear force and the term $
\gamma_\ell (T_{\ell} -p_1^2)$ is the heat current due to the left
reservoir.
Let $P_s$ be the velocity of the center
of mass of the system and $J_{s}$ be the average energy current, which are defined by
$$P_s=\langle p_x \rangle_{ss} \quad \text{and}\qquad
J_s=\langle j^e_{x,x+1} \rangle_{ss} .
$$
We have the simple relation between these two quantities
\begin{equation}
\label{eq:JV}
J_s = -\tau_\ell P_s + \gamma_\ell (T_\ell -\langle p_1^2 \rangle_{ss} ), \quad J_s = -\tau_r P_s - \gamma_r (T_r -\langle p_N^2 \rangle_{ss} ).
\end{equation}
The value of $P_s$ can be determined exactly and is independent of the nonlinearities present in the system. By writing that $\langle {\mc L}_N (p_x) \rangle_{ss} =0$ for any $x=1, \ldots,N$ we get that the tension profile, defined by $\tau_x = \langle V'(r_x) \rangle_{ss}$, satisfies
\begin{equation*}
\begin{split}
\tau_2 -\tau_\ell =\gamma_\ell P_s, \quad \tau_r -\tau_N = \gamma_r P_s,\\
\tau_{x+1} -\tau_x = \gamma P_s, \quad x=2, \ldots, N-1.
\end{split}
\end{equation*}
We have then:
\begin{lemma}[\cite{BO2}]
\label{lem:velo}
The velocity $P_s$ of the center of mass is given by
\begin{equation}
\label{eq:V}
P_s = \cfrac{\tau_r -\tau_\ell}{\gamma (N-2) + \gamma_\ell + \gamma_r}
\end{equation}
and the tension profile is linear:
\begin{equation}
\label{eq:1}
\tau_x = \cfrac{\gamma (x-2) + \gamma_\ell}{\gamma (N-2) + \gamma_\ell +
\gamma_r} (\tau_r - \tau_\ell) + \tau_\ell.
\end{equation}
Consequently
\begin{equation}
\lim_{n \to \infty} \tau_{[Nu]} =
\tau_\ell + (\tau_r -\tau_\ell) u, \quad u \in [0,1].
\end{equation}
\end{lemma}
For purely deterministic chain ($\gamma=0$), the
velocity $P_{s}$ is of order $1$, while the tension profile is flat
at the value $\left({\gamma_\ell + \gamma_r}\right)^{-1}\left[ {\gamma_\ell \tau_r + \gamma_r \tau_\ell }\right]$. The first effect of the
noise is to make $P_s$ of order $N^{-1}$ and to give a nontrivial macroscopic tension profile.
It is expected that there exists a positive constant $C$ independent of the size $N$ such that $\langle {\mc E}_x \rangle_{ss} \le C$ for any $x=1, \ldots,N$. Apart from the harmonic case we do not know how to prove such a bound.
We shall denote by ${\tilde f_{ss}}$ the derivative of the stationary state $\mu_{ss}$ with respect to the
local Gibbs equilibrium state $\mu_{lg}$ defined by $\mu_{lg} (d\br, d\bp) = g(\br,\bp) d \br d \bp$ with
\begin{equation}
\label{eq:GLES-ness}
g (\br,\bp) = \prod_{x=1}^N \frac{e^{-\beta_x (\mathcal E_x - \tau_x
r_x)}}{Z(\tau_x\beta_x, \beta_x)},
\end{equation}
where $\beta_x=\beta_{\ell} + \frac{x}{N} (\beta_r -\beta_{\ell})$ and $\tau_x=\tau_{\ell} + \frac{x}{N} (\tau_r -\tau_{\ell})$. In the formula above we have introduced $r_1=0$ to avoid annoying notations.
The function ${\tilde f}_{ss}$ is solution, in the sense of distributions, of the equation
\begin{equation}
\label{eq:ss48}
{\tilde{\mc L}}_N^{*} \, \tilde f_{ss} = 0
\end{equation}
where ${\tilde{\mc L}}_N^*$ is the adjoint of ${\mc L}_N$ in ${\mathbb L}^2 (\mu_{lg})$. We assume that $T_r=T +\delta T, T_\ell= T$ and $\tau_r = \tau -\delta \tau, \tau_\ell= \tau$ with $\delta T, \delta \tau$ small. At first order in $\delta T$ and $\delta \tau$, we have
\begin{equation*}
\begin{split}
{ \tilde {\mc L}}_N^{*}
&= {\mc L}_{N, {\rm{eq.}}}^{*} + \gamma_r \delta T \partial_{p_N}^2 -\delta \tau \partial_{p_N} -\frac {\delta T}{T^2 N} \sum_{x=1}^{N-1} \left(j^e_{x,x+1} + \tau p_x \right) -\frac{\delta \tau}{NT} \sum_{x=1}^{N-1} p_x+ o(\delta T, \delta \tau)
\end{split}
\end{equation*}
where $ {\mc L}_{ N, {\rm{eq.}}}^{*}=-{ \mathcal A}^{\tau,\tau}_{N} + \gamma {\mc S}_N + \gamma_\ell {\mc B}_{1, T} + \gamma_r {\mc B}_{N, T} $ is the adjoint in ${\mathbb L}^{2} (\mu^N_{\tau,T})$ of
\begin{equation}
\label{eq:ss43}
{\mc L}_{N,{\rm{eq.}}} ={ \mathcal A}_N^{\tau,\tau} + \gamma {\mc S}_N + \gamma_\ell {\mc B}_{1, T} + \gamma_r {\mc B}_{N, T}
\end{equation}
and $\mu_{\tau,T}^N$ is the finite volume Gibbs grand canonical ensemble with tension $\tau$ and temperature $T$. We now expand $\tilde f_{ss}$ at the linear order in $\delta T$ and $\delta \tau$:
\begin{equation}
\label{eq:ss55tau}
\tilde f_{ss} = 1 + \tilde u \, \delta T +\tilde v \, \delta \tau +
o(\delta T, \delta \tau)
\end{equation}
and we get that $\tilde u$ and $\tilde v$ are solution of
\begin{equation}
\label{eq:ss51}
\begin{split}
\mc L_{N,{\rm{eq.}}}^{*} \tilde u &= \frac 1{T^2 N} \sum_{x=1}^{N-1}
\left(j^e_{x,x+1} + \tau p_x \right)
\\
\mc L_{N, {\rm{eq.}}}^{*} \tilde v &= \frac{1}{NT}
\sum_{x=1}^{N-1} p_x .
\end{split}
\end{equation}
It is clear that the function $h_x$ appearing in the microscopic fluctuation-dissipation equation (\ref{eq:2:fluc-diss}) is closely related (up to a time reversal) to the functions ${\tilde u}$, ${\tilde v}$, i.e. to the {\textit{first order correction to local equilibrium}}.
We can now compute the average energy current at the first order in $\delta T$ and $\delta \tau$ as $N \to \infty$ but we need to introduce some notation. We recall that the generator of the infinite dynamics is given by ${\mc L} ={\mc A} +\gamma {\mc S}$ where, for any $f \in C_0^1 ({\tilde \Omega})$,
\begin{equation*}
\begin{split}
&({\mc A} f)(\br, \bp) = \sum_{x \in \ZZ} \left[ \left(p_x - p_{x-1}\right) \partial_{r_x} f + \left(V'(r_{x+1}) - V'(r_{x})\right) \partial_{p_{x}} f \right] (\br,\bp),\\
&({\mc S} f) (\br, \bp) = \frac{1}{2}\sum_{x \in \ZZ} \left[ f(\br, \bp^x) -f (\br, \bp) \right].
\end{split}
\end{equation*}
Let ${\bb H}:={\bb H}_{\tau,T}$ be the completion of the vector space of bounded local functions w.r.t. the semi-inner product $\ll \cdot, \cdot \gg$ defined for bounded local functions $f,g : {\tilde \Omega} \to \RR$, by
\begin{equation}
\label{eq:scalarproduct>>}
\ll f, g \gg = \sum_{x \in \ZZ} \left\{ \mu_{\tau,T} (f \theta_x g) - \mu_{\tau,T} (f) \mu_{\tau,T} (g)\right\}.
\end{equation}
Observe that in ${\bb H}$ every constant $c\in \RR$ and discrete gradient $\psi=\theta_1 f - f$ is equal to zero since for any local bounded function $h$ we have $\ll c, h \gg =0$ and $\ll \psi, h \gg =0$.
Assuming they exist let ${\tilde J}_s$ and ${\hat P_s}$ be the limiting average energy current and velocity:
\begin{equation}
\label{eq:PJJJ}
{\tilde J}_s = \lim_{N \to \infty} N \langle j^e_{0,1} \rangle_{ss}, \quad {\hat P}_{s} = \lim_{N \to \infty} N \langle p_0 \rangle_{ss},
\end{equation}
and define ${\hat J}_{s} = {\tilde J}_s + \tau {\hat P}_s$. We expect that as $N$ goes to infinity and, at first order in $\delta T$ and $\delta \tau$,
\begin{equation*}
\left(
\begin{array}{c}
{\hat J}_s\\
{\hat P}_s
\end{array}
\right)
= - \,
\kappa (T,\tau) \,
\left(
\begin{array}{c}
\delta T\\
\delta \tau
\end{array}
\right)
\end{equation*}
with
\begin{equation}
\label{eq:1:cond-mat}
\kappa (T,\tau)=
\left(
\begin{array}{cc}
\kappa^e& \kappa^{e,r}\\
\kappa^{r,e}&\kappa^r
\end{array}
\right)
\end{equation}
the {\textit{ thermal conductivity}} matrix. Assume for simplicity that $N=2k$ is even. By (\ref{eq:ss55tau}) and (\ref{eq:ss51}), we get that
\begin{equation*}
\begin{split}
N \langle p_0 \rangle_{ss}=N \langle p_k \rangle_{ss}& = N \int p_k \, {\tilde f}_{ss} \; d\mu_{lg}\\
&= N\, \delta T \, \int p_k \, {\tilde u} \; d\mu_{lg} \; + \; N \, \delta \tau \, \int p_k \, {\tilde v} \; d\mu_{lg} \;+ \; o(\delta T, \delta \tau)\\
&= -\cfrac{\delta T}{T^2} \, \int p_k \; (-{\mc L}_{N, {\rm{eq.}}}^*)^{-1} \Big( \sum_{x=1}^{N-1} (j_{x,x+1}^e + \tau p_x) \Big) \; d\mu_{lg} \\
&\; - \; \cfrac{\delta \tau}{T} \, \int p_k \; (-{\mc L}_{N, {\rm{eq.}}}^*)^{-1} \Big( \sum_{x=1}^{N-1} p_x \Big) \; d\mu_{lg} \;+ \; o(\delta T, \delta \tau).
\end{split}
\end{equation*}
Since $\tfrac{d\mu_{lg}}{d\mu_{\tau, T}^N}$ is equal to $1+ O(\delta T, \delta \tau)$, we can replace $\mu_{\lg}$ by $\mu_{\tau,T}^N$ in the last terms of the previous expression. Using that ${\mc L}_{N, {\rm{eq.}}}^*$ is the adjoint of ${\mc L}_{N, {\rm{eq.}}}$ in ${\bb L}^2 (\mu_{\tau,T}^N)$ and denoting by $\langle \cdot, \rangle_{\tau,T}$ the scalar product in ${\bb L}^2 (\mu_{\tau,T}^N)$, we obtain that
\begin{equation*}
\begin{split}
N \langle p_0 \rangle_{ss}&= -\cfrac{\delta T}{T^2} \, \left\langle (-{\mc L}_{N, {\rm{eq.}}})^{-1} p_k \; , \; \sum_{x=1}^{N-1} (j_{x,x+1}^e + \tau p_x) \; \right\rangle_{\tau,T} \\
&\; - \; \cfrac{\delta \tau}{T} \, \left\langle (-{\mc L}_{N, {\rm{eq.}}})^{-1} p_k \; , \; \sum_{x=1}^{N-1} p_x \; \right\rangle_{\tau,T} \;+ \; o(\delta T, \delta \tau)\\
&= -\cfrac{\delta T}{T^2} \, \left\langle (-{\mc L}_{2k, {\rm{eq.}}})^{-1} p_k \; , \; \sum_{y=-k+1}^{k-1} (j_{y+k,y+k+1}^e + \tau p_{y+k}) \; \right\rangle_{\tau,T} \\
&\; - \; \cfrac{\delta \tau}{T} \, \left\langle (-{\mc L}_{2k, {\rm{eq.}}})^{-1} p_k \; , \; \sum_{y=-k+1}^{k-1} p_{y+k} \; \right\rangle_{\tau,T} \;+ \; o(\delta T, \delta \tau)
\end{split}
\end{equation*}
In the first order terms of the previous expression we can recenter everything around $k$ by a translation of $-k$ and we get
\begin{equation*}
\begin{split}
N \langle p_0 \rangle_{ss}&= -\cfrac{\delta T}{T^2} \, \left\langle (-{\mc L}_{\Lambda_k, {\rm{eq.}}})^{-1} p_0 \; , \; \sum_{y=-k+1}^{k-1} (j_{y,y+1}^e + \tau p_{y}) \; \right\rangle _{\tau,T} \\
&\; - \; \cfrac{\delta \tau}{T} \, \left\langle (-{\mc L}_{\Lambda_k, {\rm{eq.}}})^{-1} p_0 \; , \; \sum_{y=-k+1}^{k-1} p_{y} \; \right\rangle_{\tau,T} \;+ \; o(\delta T, \delta \tau)
\end{split}
\end{equation*}
where $\Lambda_k= \{-k+1, \ldots,k\}$ and
\begin{equation*}
{\mc L}_{\Lambda_k, {\rm{eq.}}}={\mc A}_{\Lambda_k}^{\tau,\tau} +\gamma {\mc S}_{\Lambda_k} + \gamma_\ell {\mc B}_{-k,T} +{\gamma_r} {\mc B}_{k, T}
\end{equation*}
with
\begin{equation*}
\begin{split}
{\mathcal A }^{\tau, \tau}_{\Lambda_k} =
\sum_{x=-k+2}^{k} \left(p_{x} - p_{x-1}\right) \partial_{r_x} +
\sum_{x=-k+2}^{k-1}\left(V'(r_{x+1}) - V'(r_{x})\right)
\partial_{p_{x}}\\
- \left(\tau- V'(r_{-k+2})\right) \partial_{p_{-k+1}}
+ \left(\tau - V'(r_{k})\right) \partial_{p_{k}}
\end{split}
\end{equation*}
and
\begin{equation*}
({\mc S}_{\Lambda_k} f ) (\br,\bp) = \frac{1}{2} \sum_{x=-k+2}^{k-1} \left(f(\br,\bp^{x}) - f(\br,\bp)\right).
\end{equation*}
A similar formula can be obtained for $N\langle j_{0,1}^e \rangle_{ss}$. As $k \to \infty$, the finite volume Gibbs measure converges to the infinite volume Gibbs measure. Moreover, we expect that since $k \to \infty$ the effect of the boundary operators ${\mc B}_{\pm k,T}$ around the site $0$ disappears so that $(-{\mc L}_{\Lambda_k, {\rm{eq.}}})^{-1} p_0$ converges to $(-{\mc L})^{-1} p_0$. Therefore, in the thermodynamic limit $N \to \infty$ ({\textit{i.e.}} $k \to \infty$), the transport coefficients are given by the Green-Kubo formulas
\begin{equation}
\label{eq:1:ke}
\begin{split}
\kappa^e = T^{-2} \ll j^e_{0,1} +\tau p_0\,,\, (-{\mc L})^{-1}\, (j^e_{0,1} +\tau p_0) \gg,\\
\kappa^{e,r} =T^{-1} \ll p_0\,,\, (-{\mc L})^{-1} \,( j^e_{0,1}+\tau p_0) \gg,
\end{split}
\end{equation}
and
\begin{equation}
\label{eq:1:kr}
\begin{split}
\kappa^r =T^{-1} \ll p_{0} \,,\,(- {\mc L})^{-1}\, (p_0) \gg,\\
\kappa^{r,e} =T^{-2} \ll j^e_{0,1} + \tau p_0\,,\, (-{\mc L})^{-1} \,( p_0) \gg.
\end{split}
\end{equation}
The argument above is formal. In fact even proving the existence of the transport coefficients defined by (\ref{eq:1:ke}), (\ref{eq:1:kr}) is a non-trivial task. The existence of ${\hat P}_s$ defined by the second limit in (\ref{eq:PJJJ}) can be made rigorous since we have the exact expression of $P_s$. From Lemma \ref{lem:velo}, we have, even for $\delta \tau ,\delta T$ that are not small,
$${\hat P}_s = -\cfrac{\delta \tau}{\gamma}.$$
On the other hand we show in Theorem \ref{th:GK} that the quantities $\kappa^r, \kappa^{r,e}$, formally given by (\ref{eq:1:kr}), can be defined in a slightly different but rigorous way, and are then equal to
\begin{equation}
\label{eq:krr}
{\kappa^r} = \gamma^{-1}, \quad {\kappa}^{r,e} =0.
\end{equation}
Thus we can rigorously establish the validity of the linear response theory for the velocity ${\hat P}_s$.
\subsubsection{Existence of the Green-Kubo formula}
\label{subsec:GK0}
One of the main results of \cite{BO2} is the existence of the Green-Kubo formula for the conductivity matrix. Let ${\bb H}^a$ (resp. ${\bb H}^s$) be the set of functions $f:{\tilde \Omega} \to \RR$ antisymmetric (resp. symmetric) in $\bp$, i.e. $f(\br,\bp)=-f(\br,-\bp)$ (resp. $f(\br,\bp)=f(\br,-\bp)$) for every configuration $(\br,\bp) \in {\tilde \Omega}$. For example, the functions $j^e_{0,1}$, $p_0$ and every linear combination of them are antisymmetric in $\bp$.
\begin{theorem}[\cite{BO2}, \cite{BHLLO}]
\label{th:GK}
Let $f,g \in {\bb H}^a$. The limit
\begin{equation*}
\sigma(f, g) =\lim_{\substack{z \to 0\\ z>0}} \ll f\, ,\, (z - {\mc L})^{-1} \, g \gg
\end{equation*}
exists and $\sigma(f,g)=\sigma(g,f)$. Therefore, the conductivity matrix $\kappa (T,\tau)$ is well defined in the following sense: the limits
\begin{equation}
\begin{split}
\kappa^e =\lim_{\substack{z \to 0\\ z>0}} T^{-2} \ll j^e_{0,1} +\tau p_0\,,\, (z-{\mc L})^{-1}\, (j^e_{0,1} +\tau p_0) \gg,\\
\kappa^{e,r} =\lim_{\substack{z \to 0\\ z>0}} T^{-1} \ll p_0\,,\, (z-{\mc L})^{-1} \,( j^e_{0,1}+\tau p_0) \gg,\\
\kappa^r =\lim_{\substack{z \to 0\\ z>0}} T^{-1} \ll p_{0} \,,\,(z- {\mc L})^{-1}\, (p_0) \gg=\gamma^{-1},\\
\kappa^{r,e} =\lim_{\substack{z \to 0\\ z>0}} T^{-2} \ll j^e_{0,1} + \tau p_0\,,\, (z-{\mc L})^{-1} \,( p_0) \gg
\end{split}
\end{equation}
exist and are finite. Moreover Onsager's relation $\kappa^{e,r}=\kappa^{r,e} (=0) $ holds.
\end{theorem}
We have a nice thermodynamical consequence of the previous result. If $\delta T$ and $\delta \tau$ are small and of the same order, the system cannot be used as a refrigerator or a
boiler: at the first order, a gradient of tension does
not contribute to the heat current ${\hat J}_s$.
The argument above says nothing about the possibility to
realize a heater or a refrigerator if $\delta \tau $ is not of the
same order as $\delta T$. For the harmonic chain, we will see that it is possible to get a heater if $\delta
\tau $ is of order $\sqrt {\delta T}$.
\begin{remark}
\begin{enumerate}
\item The existence of the Green-Kubo formula is also valid for a pinned or unpinned chain in any dimension.
\item Observe that with respect to the establishment of a microscopic fluctuation-dissipation equation (\ref{eq:2:fluc-diss}) the computation of the Green-Kubo formula is less demanding since only the knowledge of $\sum_x h_x$ is necessary.
\end{enumerate}
\end{remark}
The proof of Theorem \ref{th:GK} is based on functional analysis arguments. The first main observation is that there exists a spectral gap for the operator ${\mc S}$ restricted to the space ${\bb H}^a$.
\begin{lemma}
\label{lem:sg}
The noise operator ${\mc S}$ lets ${\bb H}^a$ and ${\bb H}^s$ invariant. For any local function $f \in {\bb H}^a$ we have that
\begin{equation}
\label{eq:1:sg1}
\ll f, f \gg \; \le \; \ll f, -{\mc S} f \gg.
\end{equation}
Moreover, for any local function $f \in {\bb H}^a$, there exists a local function $h \in {\bb H}^a$ such that
$${\mc S}h =f.$$
\end{lemma}
\begin{proof}
Since the Gibbs states are Gaussian states in the $p_x$'s it is convenient to decompose the operator ${\mc S}$ (which acts only on the $p_x$'s) in the orthogonal basis of Hermite polynomials. The the lemma follows easily.
\cqfd\end{proof}
\begin{proof}[Theorem \ref{th:GK}]
We observe first that ${\bb H}^a$ and ${\bb H}^s$ are orthogonal Hilbert spaces such that ${\bb H} ={\bb H}^a \oplus {\bb H}^s$. It is also convenient to define the following semi-inner product
\[
\ll u,w\gg_1=\ll u, (-\mc S) w\gg.
\]
Let $\bb H^1$ be the associated Hilbert space. We also define the Hilbert space $\bb H^{-1}$ via the duality
given by the ${\bb H}$ norm, that is
\[
\|u\|_{-1}^2=\sup_{w} \{ \, 2 \ll u,w\gg -\ll w,w\gg_1 \, \}
\]
where the supremum is taken over local bounded functions $w$. By Lemma \ref{lem:sg} we have that ${\bb H}^a\subset {\bb H}^{-1}$. Thus $g \in {\bb H}^{-1}$.
Let $w_z$ be the solution of the resolvent equation $(z- {\mc L}) w_z = g$. We have to show that $\ll f, w_z \gg$ converges as $z$ goes to $0$. We decompose $w_z$ into $w_z = w^-_z +w^+_z$, $w^-_z \in {\bb H}^a$ and $w^+_z \in {\bb H}^s$. Since ${\bb H}^a$ is orthogonal to ${\bb H}^s$ and $f \in {\bb H}^a$ we have $\ll f, w_z \gg = \ll f, w_z^- \gg$. It is thus sufficient to prove that $(w_z^-)_{z>0}$ converges weakly in ${\bb H}$ as $z \to 0$.
Since ${\mc A}$ inverts the parity and $\mc S$ preserves it and ${\bb H}^a \oplus {\bb H}^s =\bb H$ and $g \in {\bb H}^a$, we have, for any $\mu,\nu>0$,
\begin{equation}
\label{eq:cfgst}
\begin{split}
&\nu w_{\nu}^+ - {\mc A} w_{\nu}^- -\gamma {\mc S} w_{\nu}^+ =0,\\
&\mu w_{\mu}^{-} - {\mc A} w_\mu^+ - \gamma {\mc S} w_{\mu}^- =g.
\end{split}
\end{equation}
Taking the scalar product with $w_\mu^+$ (resp. $w_{\nu}^-$) on both sides of the first (resp. second) equation of (\ref{eq:cfgst}), we get
\begin{equation}
\label{eq:oldsplit}
\begin{split}
&\nu\ll w_{\mu}^+,w_{\nu}^+\gg -\ll w_{\mu}^+, {\mc A} w_{\nu}^-\gg+\gamma\ll w_{\mu}^+,w_{\nu}^+\gg_1=0,\\
&\mu\ll w_{\nu}^-,w_{\mu}^-\gg -\ll w_{\nu}^-, {\mc A} w_{\mu}^+\gg+\gamma \ll w_{\mu}^-,w_{\nu}^-\gg_1=\ll w_{\nu},g \gg.
\end{split}
\end{equation}
Summing the above equations we have
\begin{equation}
\label{eq:baseone}
\nu\ll w_{\mu}^+,w_{\nu}^+\gg+\mu\ll w_{\nu}^-,w_{\mu}^-\gg+\gamma\ll w_{\mu},w_{\nu}\gg_1=\ll w_{\nu},g\gg
\end{equation}
Putting $\mu=\nu$ we get
\[
\nu\ll w_{\nu} , w_\nu \gg +\gamma \ll w_{\nu},w_{\nu}\gg_1\leq \| w_{\nu}\|_1\| g\|_{-1}.
\]
Hence $(w_\nu)_{\nu >0}$ is uniformly bounded in $\bb H^1$ and by the spectral gap property so is $( w_\nu^- )_{\nu >0}$ in $\bb H$. Moreover, $( \nu w_{\nu} )_{\nu>0}$ converges strongly to $0$ in ${\bb H}$ as $\nu \to 0$. We can then extract weakly convergent subsequences. Taking first the limit, in \eqref{eq:baseone}, $\nu\to 0$ and then $\mu\to 0$ along one such subsequence (converging to $w_*$) we have
\[
\gamma \ll w_{*},w_{*}\gg_1=\ll w_{*},g\gg.
\]
Next, taking the limit along different weakly convergent subsequences (let $w^*$ be the other limit) we have
\[
\gamma \ll w_{*},w^{*}\gg_1=\ll w^{*},g\gg
\]
and, exchanging the role of the two sequences
\[
2\gamma \ll w_{*},w^{*}\gg_1=\ll w_{*},g\gg +\ll w^{*},g\gg =\gamma \ll w_{*},w_{*}\gg_1+\gamma\ll w^{*},w^{*}\gg_1
\]
which implies $w_*=w^*$, that is all the subsequences have the same limit. Thus $(w_\nu)_{\nu>0}$ converges weakly in ${\bb H}^1$ as well as $(w_{\nu}^-)_{\nu>0}$ in $\bb H$ by Lemma \ref{lem:sg}.
\cqfd\end{proof}
In the harmonic case, $V ( r )=r^2/2$, much more is known. Indeed one easily checks that the exact microscopic fluctuation-dissipation equation (\ref{eq:2:fluc-diss}) holds with
\begin{equation}
\label{eq:fluct-diss-eq-harm}
h_x =\frac{1}{2\gamma} r_{x+1} (p_x +p_{x+1}) - \frac{r_{x+1}^2}{4}, \quad \varphi_x =- \frac{1}{2\gamma} (r_x r_{x+1} + p_x^2).
\end{equation}
It follows that we can compute explicitly $(z-{\mc L})^{-1}j^{e}_{0,1}$ and obtain that the value of the conductivity matrix:
\begin{equation*}
\kappa (\tau, T)=
\left(
\begin{array}{cc}
\frac{1}{2 \gamma} & 0\\
0& \frac{1}{\gamma}
\end{array}
\right).
\end{equation*}
This value will be recovered by considering the hydrodynamic limits of the system (Theorem \ref{thm:Sim1}) and also by establishing the validity of Fourier's law (see Theorem \ref{thm:BO1}).
\subsubsection{Expansion of the Green-Kubo formula in the weak coupling limit}
In the previous subsection we proved the existence of the Green-Kubo formula showing that the transport coefficient is well defined if some noise is added to the deterministic dynamics. We are now interested in the behavior of the Green-Kubo formula as the noise vanishes. We investigate this question in the weak coupling limit, i.e. assuming that the interaction potential is of the form $\ve V$ where $\ve \ll 1$ is the (small) coupling parameter. For notational simplicity we consider the one dimensional infinite pinned system but the arguments given below are easily generalized to the (pinned or unpinned) $d \ge 1$-dimensional case {{\footnote{If $W=0$ the variables $q_x$ have to take values in a compact manifold.}}}. The expansion presented in this section is formal but we will precise at the end of the section what has been rigorously proved. In order to emphasize the dependence of $\kappa^e$ (denoted in the sequel by $\kappa$) in the coupling parameter $\ve$ and the noise intensity $\gamma$, we denote $\kappa$ by $\kappa(\ve, \gamma)$. Here we propose a formal expansion of the conductivity $\kappa$ in the form
\begin{equation}
\label{eq:expK}
\kappa (\ve,\gamma)= \sum_{n \ge 2} \kappa_n (\gamma) \ve^n.
\end{equation}
Then we study rigorously the first term of this expansion $\kappa_2 (\gamma)$. It is intuitively clear that the expansion starts from $\ve^2$ since the Green-Kubo formula is a quadratic function of the energy current and that the latter is of order $\ve$ (see (\ref{current: general formula})).
When the system is uncoupled ($\ve=0$), the dynamics is given by the generator ${\mc L}_0={\mc A}_0 +\gamma S$ with ${\mc S}$ the flip noise defined by (\ref{eq:cons-noise-flip}) and
\begin{equation*}
{\mc A}_0 =\sum_{x \in \ZZ} p_x \partial_{q_x} - W' (q_x) \partial_{p_x}.
\end{equation*}
When $\ve >0$, the generator of the coupled dynamics is denoted by
\begin{equation}
\label{eq:geneve}
{\mc L}_{\ve} = {\mc L}_0 + \ve {\mc G}
\end{equation}
where
$$ {\mc G} =\sum_{x \in \ZZ} \, V'(q_x -q_{x-1})(\partial_{p_{x-1}} - \partial_{p_x}).$$
The energy of each cell, which is the sum of the internal energy and
of the interaction energy, is defined by
\begin{equation}
\label{eq:11}
{\mc E}_x^\ve = {\mc E}_x + \frac \ve 2 \left(V(q_{x+1} - q_{x}) + V(q_{x} -
q_{x-1})\right), \quad {\mc E}_x = \frac{p_x^2}{2} + W(q_x).
\end{equation}
Observe that ${\mc E}_x ={\mc E}_x^0$ is the energy of the
isolated system $x$. The dynamics generated by ${\mc L}_0$ preserves all the
individual energies ${\mc E}_x$. The dynamics generated by ${\mc L}_\ve$ conserves the total energy. The
corresponding energy currents $\ve j_{x,x+1}$,
defined by the local conservation law
$$
{\mc L}_{\ve} {\mc E}_x^\ve = \ve \left(j_{x-1,x}-j_{x,x+1}\right)
$$
are given by
\begin{equation}
\label{current: general formula}
\ve\, j_{x,x+1} = -\frac {\ve}{2}\, (p_x + p_{x+1}) \cdot V'(q_{x+1} -q_{x}).
\end{equation}
Let us denote by $\mu_{\beta,\ve}=\langle \cdot \rangle_{\beta,\ve}$
the canonical Gibbs measure at temperature $\beta^{-1}>0$ defined by the Dobrushin-Lanford-Ruelle equations, which of course depends on the interaction $\ve V$. We shall assume in all the cases considered
that $\mu_{\beta,\ve}$
is analytical in $\ve$ for sufficiently small $\ve$ (when applied to
local functions). In particular we assume that the potentials $V$ and
$W$ are such that the Gibbs state is unique and has spatial exponential decay of correlations (this holds under great general conditions on $V$ and $W$, see \cite{GEO}).
In order to emphasize the dependence in $\ve$ we reintroduce some notation. For any given local functions $f,g$, define the semi-inner product
\begin{equation}
\label{eq:sipt}
\ll f,g \gg_{\beta, \ve} \ =\ \sum_{x\in \ZZ} [\langle\theta_x f, g\rangle_{\beta,\ve} -
\langle f\rangle_{\beta, \ve} \langle g\rangle_{\beta,\ve}].
\end{equation}
We recall that $\theta_x$ is the shift operator by $x$. The sum is finite in the case $\ve =0$, and converges for $\varepsilon>0$ thanks to the exponential decay of the spatial correlations. Denote by $\bb H_{\ve} = {\bb L}^2( \ll\cdot ,\cdot \gg_{\beta,\varepsilon} )$ the corresponding closure. We define the subspace of antisymmetric functions in the velocities
\begin{equation}
\label{eq:antisymm}
\bb H_\ve^a = \left\{ f \in \bb H_\ve: f({\bf q},-{\bf p}) = -f({\bf q},{\bf p})\right\}.
\end{equation}
Similarly we define the subspace of symmetric functions in $\bf p$ as
$\bb H_\ve^s$. On local functions this decomposition of a function into symmetric and antisymmetric parts is
independent of $\ve$. Let us denote by $\mc P_{\ve}^a$ and $\mc P_{\ve}^s$
the corresponding orthogonal projections, whose definition in fact
does not depend on $\ve$. Therefore we sometimes omit the index $\ve$ in the notation. Finally, for any function $f\in {\bb L}^2(\mu_{\beta,\ve})$, define
$$
(\Pi_\ve f) ({\mathbb {\mc E}}) = \mu_{\beta,\ve} (f|{\mathbb{\mc E}}),\quad Q_\ve =\rm{Id}-\Pi_\ve
$$
where ${\mathbb{\mc E}}:=\{ {\mc E}_x \, ; \, x \in \ZZ \}$.
According to Theorem \ref{th:GK} the conductivity is defined by
\begin{equation}
\label{eq:ghj1}
\kappa (\ve, \gamma)= \ve^2 \lim_{\nu \to 0} \ll j_{0,1} \, , \, (\nu -{\mc L}_{\ve})^{-1} j_{0,1} \gg_{\beta,\ve}.
\end{equation}
It turns out that, for calculating the terms in the expansion (\ref{eq:expK}), it is convenient to choose $\nu = \ve^2\lambda$ in (\ref{eq:ghj1}), for a $\lambda >0$, and solve the resolvent equation
\begin{equation}\label{eq:poisson}
(\lambda\ve^2-\mc L_\ve) u_{\lambda,\ve} = \ve j_{0,1}
\end{equation}
for the unknown function $u_{\lambda,\ve}$. The factor $\ve^2$ is the natural scaling in view of
the subsequent computations.
We assume that a solution of \eqref{eq:poisson} is in the form
\begin{equation}
\label{eq:6}
u_{\lambda,\ve} =\sum_{n \ge 0} U_{\lambda,n} \ve^n= \sum_{n \ge 0} (v_{\lambda,n} + w_{\lambda,n}) \ve^n,
\end{equation}
where $\Pi v_{\lambda,n} =Q w_{\lambda,n}=0$, i.e. $w_{\lambda,n} = \Pi U_{\lambda,n}$ and $v_{\lambda,n} =Q U_{\lambda,n}$. Here $\Pi=\Pi_0$ and $Q=
Q_0$ refer to the uncoupled measure $\mu_{\beta,0}${\footnote{The reason to use the orthogonal decomposition of $U_{\lambda,n}= v_{\lambda,n} + w_{\lambda,n}$ is that at some point we will have to consider, for a given function $f$, the solution $h$ to the Poisson equation ${\mc L}_0 h= f $ . The minimal requirement for the existence of $h$ is that $\Pi f =0$.}}. Given such an expression we can, in principle, use it in \eqref{eq:ghj1} to write
\begin{equation}
\label{eq:kcoef}
\begin{split}
\kappa(\ve, \gamma)&=\lim_{\lambda\to 0}\sum_{n\geq0}\ve^{n+1}\ll j_{0,1},v_{\lambda,n}+w_{\lambda,n}\gg_{\beta,\ve}\\
&=\sum_{n\geq 1}\lim_{\lambda\to 0}\ve^{n}\ll j_{0,1},v_{\lambda,n-1}\gg_{\beta,\ve}
\end{split}
\end{equation}
where we have used the fact that that $\ll j_{0,1}, w_{\lambda,\ve} \gg_{\beta,\ve}=0$ and we have,
arbitrarily, exchanged the limit with the sum. Note that this is not yet
of the type \eqref{eq:expK} since the terms in the expansion depend
themselves on $\ve$. To identify the coefficients $\kappa_n$ we
would need to expand in $\ve$ also the expectations. This is not
obvious since the functions $v_{\lambda,n}$ are non local.
Let us consider the operator $\mf L = \Pi \mc G \mc P^a (-\mc L_0)^{-1} \mc G\Pi.$ We show below that the operator $\mf L$ is a generator of a Markov process so that
$(\lambda- {\mf L})^{-1}$ is well defined for $\lambda>0$. Pluging (\ref{eq:6}) in (\ref{eq:poisson}) we obtain the following hierarchy
\begin{equation}
\label{eq:w0}
\begin{split}
&v_{\lambda,0}=0,\\
&w_{\lambda,0}=(\lambda-\mf L)^{-1}\Pi \mc G \mc P^a (-\mc L_0)^{-1}j_{0,1},\\
&v_{\lambda,1}= (-\mc L_0)^{-1}\left[j_{0,1}+\mc G w_{\lambda,0}\right],\\
&w_{\lambda,n}=(\lambda-\mf L)^{-1}\Pi \mc G {\mc P}^a (-\mc L_0)^{-1}\left[-\lambda
v_{\lambda, n-1} + Q \mc G v_{\lambda,n}\right], \qquad n\ge 1\\
&v_{\lambda,n+1}= (-\mc L_0)^{-1}\left[-\lambda v_{\lambda, n-1} + \mc G
w_{\lambda,n} + Q\mc G v_{\lambda,n}\right], \qquad n\ge 1.
\end{split}
\end{equation}
Observe that in the previous equations the (formal) operator $(-{\mc L}_0)^{-1}$ is always applied to functions $f$ such that $\Pi f =0$ (this is the minimal requirement to have consistent equations). This is however not sufficient to make sense of the functions $v_{\lambda, n}$ and $w_{\lambda,n}$. Nevertheless, by using an argument similar to the one given in Theorem \ref{th:GK}, we have that the local operator ${\mc T}_0$ on ${\bb H}_0^a$ defined by
$${\mc T}_0 f = \lim_{\nu \to 0} {\mc P}^a (\nu -{\mc L}_0)^{-1} f, \quad f \in \bb H_{0}^a,$$ is well defined. Therefore, it is possible to make sense, as a distribution, of
\begin{equation}
\label{eq:alpha01}
\alpha_{01}= \Pi \mc G \mc P^a (-\mc L_0)^{-1}j_{0,1}:=\Pi \mc G {\mc T}_0 j_{0,1}.
\end{equation}
Nevertheless, the function $w_{\lambda,0}$ is still not well defined since we are not sure that $ {\mc T}_0 j_{0,1}$ is in the domain of ${\mc G}$.
Even if the previous computations are formal a remarkable fact is that the operator ${\mf L}$, when applied to functions of the internal energies, coincides with the Markov generator ${\mf L}_{\rm{GL}}$ of a reversible Ginzburg-Landau dynamics on the internal energies. Let us denote by $\rho_{\beta}$ the distribution of the internal energies ${\mathbf{\mc E}}=\{ {\mc E}_x \, ; \, x \in \ZZ \}$ under the Gibbs measure $\mu_{\beta,0}$. It can be written in the form
$$
d\rho_{\beta} ({\mathbf{\mc E}})= \prod_{x \in \ZZ} Z_\beta^{-1} \exp( - \beta {\mc E}_x -U({\mc E}_x)) d{\mc E}_x
$$
for a suitable function $U$. We denote the formal sum $\sum_x U({\mc E}_x)$
by ${\mc U}:={\mc U} ({\mathbf{\mc E}})$. We denote also, for a given value of the
internal energy ${\tilde {\mc E}}_x$ in the cell $x$, by $\nu_{{\tilde {\mc E}}_x}^{x}$ the microcanonical probability measure in the cell $x$. i.e. the uniform probability measure on the manifold
$$
\Sigma_{{\tilde {\mc E}}_x}:=\{ (q_x,p_x) \in \Omega \, ; \, {\mc E}_x (q_x,p_x)
\, =\, {\tilde {\mc E}}_x \}.
$$
Then, the generator ${\mf L}_{\rm{GL}}$ is given by
\begin{equation}
\label{eq:12}
{\mf L}_{\rm{GL}} = \sum_x e^{\mc U} (\partial_{{\mc E}_{x+1}} - \partial_{{\mc E}_x} )
\left[e^{-\mc U} \gamma^2 ({\mc E}_x, {\mc E}_{x+1})
(\partial_{{\mc E}_{x+1}} - \partial_{{\mc E}_x} ) \right],
\end{equation}
where
\begin{equation}
\label{gamma square}
\gamma^2 ({\tilde {\mc E}}_0, {\tilde {\mc E} }_{1})= \int_{\Sigma_{{\tilde {\mc E}}_0} \times
\Sigma_{{\tilde {\mc E}}_{1}}} \big( j_{0,1}\; {\mc T}_0\, j_{0,1} \big) \; d\nu_{{\tilde {\mc E}}_0}^0 d\nu_{{\tilde {\mc E}}_{1}}^1.
\end{equation}
The operator ${\mf L}_{\rm{GL}} $ is well defined only if the function $\gamma^2$ has some regularity properties, that are actually proven in specific examples \cite{LO-0, DL}. We can show that the Dirichlet forms {\footnote{They are well defined even if $\gamma^2$ is not regular.}} associated to $\mf L$ and ${\mf L}_{{\rm{GL}}}$ coincide. Then in the cases where $\gamma^2$ is proven to be smooth \eqref{eq:12} is well defined and $\mf L={\mf L}_{{\rm{GL}}}$.
\begin{proposition}[\cite{BHLLO}]
\label{prop:Lweak}
For each local smooth functions $f,g$ of the internal energies only we have
\begin{equation}\label{dir-form}
\ll g ,(-\mf L) f \gg_{\beta,0} = \ll g, (-{\mf L}_{\rm{GL}}) f \gg_{\beta,0}.
\end{equation}
\end{proposition}
The operator ${\mf L}_{\rm{GL}}$ is the generator of a Ginzburg-Landau dynamics
which is reversible with respect to $\rho_{\beta}$, for any $\beta>0$. It is conservative
in the energy $\sum_x {\mc E}_x$ and the corresponding currents are given by $\theta_x \alpha_{0,1}$ where $\alpha_{0,1}$ has been defined in (\ref{eq:alpha01}). The corresponding finite size
dynamics appears in \cite{LO-0,DolgoL} as the weak coupling limit of a
finite number $N$ (fixed) of cells weakly coupled by a potential
$\varepsilon V$ in the limit $\varepsilon \to 0$ when time $t$ is rescaled
as $t \ve^{-2}$. Moreover, the hydrodynamic limit of the Ginzburg-Landau dynamics is then given (in the diffusive time scale $tN^2$, $N \to + \infty$), by a heat equation with diffusion coefficient which coincides with $\kappa_2$ as given by \eqref{eq:gke} below (\cite{Var0}). This is summarized in Figure \ref{fig:wclimit000}.
\begin{center}
\begin{figure}
\begin{tikzpicture}[node distance = 4cm, auto, scale=0.75]
\node[draw, rectangle,fill=blue!25] (MS) at (-5,5) {$N$ cells coupled by $\ve V$};
\node[draw,rectangle,fill=blue!25] (HE) at (5,5) {$\partial_t T = \nabla ({\kappa} (\ve,\gamma) \nabla T)$};
\draw[->,>=latex,dashed] (MS) -- (HE) node[sloped,midway]{$tN^2$, $N \to \infty$, $\varepsilon \sim 1$};
\node[draw,rectangle,fill=blue!25] (GL) at (-5,0) {Ginzburg-Landau dynamics for $N$ particles};
\draw[->,>=latex] (MS) -- (GL) node[midway]{${\varepsilon}^{-2} t$, $\varepsilon \to 0$};
\node[draw,rectangle,fill=blue!25] (HE2) at (5,4) {$\partial_t T = \nabla ({\kappa}_2 (\gamma) \nabla T)$};
\draw[->,>=latex] (GL) -- (HE2) node[sloped, pos=0.7]{$tN^2$, $N \to \infty$};
\end{tikzpicture}
\caption{The relation between the hydrodynamic limit, the weak coupling limit and the Green-Kubo expansion. The dotted arrow (hydrodynamic limits in the diffusive time scale) has not been proved. The weak coupling limit (vertical arrow) has been proved in \cite{LO-0} (see also \cite{DolgoL}) and the diagonal arrow (hydrodynamic limits for a Ginzburg-Landau dynamics) has been obtained in \cite{Var0} in some cases which however do not cover our cases. In \cite{BHLLO} it is argued that $\kappa (\ve, \gamma) \sim \ve^2 \kappa_2 (\gamma)$ as $\ve \to 0$.}
\label{fig:wclimit000}
\end{figure}
\end{center}
According to the previous expansion it makes sense to define $\kappa_2 (\gamma)$ by
\begin{equation}
\label{eq:k2345}
\kappa_2 (\gamma)= \lim_{\ve \to 0} \lim_{\lambda \to 0} \left\{ \ll j_{0,1}, {\mc T}_0 j_{0,1} \gg_{\beta,\ve} + \ll j_{0,1}, \mc T_0 {\mc G} w_{\lambda,0} \gg_{\beta,\ve} \right\}
\end{equation}
if the limits exist. In fact, a priori, it is not even clear that the term ${\mc T}_0 {\mc G} w_{\lambda,0}$ makes sense since $w_{\lambda,0}$ is not well defined. In \cite{BHLLO} we argue that
\begin{equation}
\label{eq:gke}
\begin{split}
\kappa_2 (\gamma) = & \left\langle \gamma^2_{0,1}
\right\rangle_{\beta} - \ll \alpha_{0,1} \, , \, (-{\mf L}_{GL})^{-1} \alpha_{0,1} \gg_{\beta}.
\end{split}
\end{equation}
Here $\langle \cdot \rangle_{\beta}$ and $\ll \cdot \gg_{\beta}$ refer to the scalar products w.r.t. $\rho_{\beta}$. In the special case $W=0$ {\footnote{If $W=0$ the variables $q_x$ have to take values in a compact manifold.}}, we prove rigorously in \cite{BHLLO} that we can make sense for any $\lambda,\ve$ of the term in the righthandside of (\ref{eq:k2345}) and that (\ref{eq:gke}) is valid, supporting the conjecture that (\ref{eq:gke}) is valid in more general situations. Observe that (\ref{eq:gke}) is the Green-Kubo formula for the diffusion coefficient of the Ginzburg-Landau dynamics.
In specific examples, it is possible to study the behavior of $\kappa_2 (\gamma)$ defined by (\ref{eq:gke}) in the vanishing noise limit $\gamma \to 0$:
\begin{enumerate}
\item Harmonic chain: it is known that the conductivity of the (deterministic) harmonic chain is $\kappa(\ve,0)=\infty$. If $\gamma>0$, $\kappa (\ve,\gamma)= c \gamma^{-1} \ve^{-2}$, $c>0$ a constant, and we get thus that $\lim_{\gamma \to 0} \kappa_2 (\gamma)=\infty$.
\item Disordered pinned harmonic chain: $V$ is quadratic and the one-site potential $W$ is site-dependent given by $W_x (q)=\nu_{x} q^2$ where $\{\nu_x \, ; \, x \in \ZZ\}$ is a sequence of independent identically distributed positive bounded random variables {\footnote{Even if this model does not belong {\textit{stricto sensu}} to the class of models discussed above it is easy to generalize to this case, at least formally, the previous results.}}. It is known (\cite{BH}) that $\kappa (\ve, 0)=0$ so that $\kappa_2 (\ve, 0) =0$. It can be proved that $\kappa_2 (\gamma)$ vanishes as $\gamma$ goes to $0$.
\item Harmonic chain with quartic pinning potential: $V$ is quadratic and $W(q)=q^4$. Then it can be shown that $\limsup_{\gamma \to 0} \kappa_2 (\gamma) < \infty$. This upper bound does not prevent the possibility that $\lim_{\gamma \to 0} \kappa_2 (\gamma)=0$.
\end{enumerate}
To prove these results we use the upper bound $\kappa_2 (\gamma) \le \left\langle \gamma^2_{0,1}
\right\rangle_{\beta}$. Recalling (\ref{gamma square}) we see that if we are able to compute ${\mc T}_0 j_{0,1}$ then we can estimate $\left\langle \gamma^2 ({\mc E}_{0}, {\mc E}_1)
\right\rangle_{\beta}$. It is exactly what is done in \cite{BHLLO} for the specific cases above.
It would be highly interesting to have a rigorous derivation of the formal expansion above. Bypassing this problem, another relevant issue is to decide if genuinely $\lim_{\gamma \to 0} \kappa_2 (\gamma) $ is zero or not. Some authors (see \cite{dRH0} and references therein) conjecture that, in some cases, the conductivity of the deterministic chain $\kappa(\ve,0)$ has a trivial weak coupling expansion ($\kappa(\ve,0) ={\mc O} (\ve^n)$ for any $n \ge 2$). Showing that $\kappa_2 (\gamma) \to 0$ as $\gamma \to 0$ would support this conjecture.
\subsection{Harmonic chain with velocity-flip noise}
In this section we assume that $V ( r )=r^2/2$.
\subsubsection{Hydrodynamic limits}
\label{subsec:hl-vf}
As explained in the beginning of this chapter an interesting problem consists to derive a diffusion equation for a chain of oscillators perturbed by an energy conserving noise. Consider a one dimensional unpinned chain of $N$ harmonic oscillators with periodic boundary conditions perturbed by the velocity flip noise in the diffusive scale. In other words let $\omega (t) =(\br (t), \bp(t))_{t \ge 0}$ be the process with generator $N^2 {\mc L}_N=N^2 \left[ {\mc A}_N + \gamma {\mc S}_N \right]$ where ${\mc S}_N$ is given by (\ref{eq:cons-noise-flip}), $\ZZ^d$ being replaced by $\TT_N$, the discrete torus of length $N$, and ${\mc A}_N$ is the Liouville operator of a chain of unpinned harmonic oscillators with periodic boundary conditions. The system conserves two quantities: the total energy $\sum_{x \in \TT_N} {\mc E}_x$, ${\mc E}_x =\frac{p_x^2}{2} +\frac{r_x^2}{2}$, and the total deformation of the lattice $\sum_{x \in \TT_N} r_x$. Consequently, the Gibbs equilibrium measures $\nu_{\beta, \tau}$ are indexed by two parameters $\beta>0$, the inverse temperature, and $\tau \in \RR$, the pressure. They take the form
\begin{equation*}
d\nu_{\beta, \tau} (d\br, d \bp) = \prod_{x \in \TT_N}{\mc Z}^{-1} (\beta,\tau) \, \exp \{ -\beta ({\mc E}_x -\tau r_x)\} dr_x dp_x
\end{equation*}
where
$${\mc Z} (\beta, \tau) = \cfrac{2\pi}{\beta} \exp(\beta \tau^2 /2).$$
Observe the following thermodynamic relations
$$\int {\mc E}_x \, d\nu_{\beta, \tau} =\beta^{-1} + \tfrac{\tau^2}{2}, \quad \int {r}_x \, d\nu_{\beta, \tau} = \tau$$
or equivalently
\begin{equation*}
\tau =\int {r}_x \, d\nu_{\beta, \tau}, \quad \beta = \left\{ \int {\mc E}_x \, d\nu_{\beta, \tau} - \tfrac{ \Big( \int {r}_x \, d\nu_{\beta, \tau} \Big)^2}{2}\right\}^{-1}.
\end{equation*}
\begin{definition}
Let $\TT=[0,1)$ be the continuous torus. Let ${\mf e}_0: \TT \to \RR$ and ${\mf r}_0:\TT \to \RR$ be two continuous macroscopic profiles such that ${\mf e}_0 > \tfrac{{\mf r}_0^2}{2}$. A sequence of probability measures $(\mu^N)_{N \ge 1}$ on $(\RR \times \RR)^{\TT_N}$ is said to be a sequence of Gibbs local equilibrium states associated to the energy profile ${\mf e}_0$ and the deformation profile ${\mf r}_0$ if
\begin{equation*}
d\mu^N (d\br, d\bp) = \prod_{x \in \TT_N} {\mc Z}^{-1} (\beta_0 (\tfrac{x}{N}) ,\tau_0 (\tfrac{x}{N}))\, \exp \{ -\beta_0 (x/N) ({\mc E}_x -\tau_0 (x/N) r_x)\} dr_x dp_x
\end{equation*}
where the functions $\beta_0$ and $\tau_0$ are defined by
\begin{equation*}
\tau_0 = {\mf r}_0, \quad \beta_0 = \{ {\mf e}_0 - \tfrac{{\mf r}_0^2}{2}\}^{-1}.
\end{equation*}
\end{definition}
Once we have the microscopic fluctuation-dissipation equation (see (\ref{eq:fluct-diss-eq-harm})) and assuming the propagation of local equilibrium in the diffusive time scale it is easy to guess the hydrodynamic equations followed by the system. In \cite{Sim} the following theorem is proved.
\begin{theorem}[\cite{Sim}]
\label{thm:Sim1}
Consider the unpinned velocity-flip model with periodic boundary conditions. Let $(\mu^N)_N$ be a sequence of Gibbs local equilibrium states {\footnote{One can consider more general initial states, see \cite{Sim}.}} associated to a bounded energy profile ${\mf e}_0$ and a deformation profile ${\mf r}_0$. For every $t \ge 0$, and any test continuous functions $G,H: \TT \to \RR$, the random variables
\begin{equation}
\label{eq:vf-empiricaldensities}
\Big(\frac1N\sum_{x \in\TT_N} {\displaystyle{ G(\tfrac{x}{N}) r_x (tN^2), \frac1N \sum_{x \in\TT_N} H(\tfrac{x}{N}) {\mc E}_x (tN^2)}} \Big)
\end{equation}
converge in probability as $N \to \infty$ to $$\Big(\int_{\TT} G(y) {\mf r} (t,y) dy, \int_{\TT} H(y) {\mf e} (t,y)dy \Big)$$ where $\mf r$ and $\mf e$ are the (smooth) solutions to the hydrodynamical equations
\begin{equation}
\label{eq:2:hl-re1}
\begin{cases}
\partial_t {\mf r} =\frac 1{\gamma}\, \partial_y^2 \, {\mf r},\\
\partial_t {\mf e}= \frac1{2 \gamma} \, \partial_y^2 \, \left[ {\mf e} \, + \, \frac{ {\mf r}^2}{2} \right]
\end{cases}
,\quad y \in \TT,
\end{equation}
with initial conditions ${\mf r} (0,y) ={\mf r}_0 (y)$, ${\mf e}(0,y)={\mf e}_0 (y)$.
\end{theorem}
The proof of this theorem is based on Yau's relative entropy method (\cite{yau1}, \cite{OVY}). The general strategy is simple. Let $\mu_t^N$ be the law of the process at time $tN^2$ starting from $\mu^N$ and let ${\tilde \mu}^N_t$ be a sequence of Gibbs local equilibrium state corresponding to the deformation profile ${\mf r}_t (\cdot):={\mf r} (t,\cdot)$ and energy profile ${\mf e}_t (\cdot):={\mf e}(t,\cdot)$ solution of (\ref{eq:2:hl-re1}). We expect that since $\mf e$ and ${\mf r}$ are the hydrodynamic profiles, the probability measure of the process $\mu_t^N$ is close, in some sense, to the local Gibss state ${\tilde \mu}_t^N$. Yau's relative entropy method consists to show that the entropic distance {\footnote{There is some abuse of language here since the relative entropy is not a distance between probability measures.}}
\begin{equation}
\label{eq:Hyauest}
H_N (t):=H(\mu_t^N | {\tilde \mu}_t^N) =o(N)
\end{equation}
between the two states is relatively small. Assuming (\ref{eq:Hyauest}), in order to prove for example the convergence of the empirical energy, we use the entropy inequality {\footnote{It is a trivial consequence of the definition (\ref{eq:ent009}).}} which states that for any $\alpha>0$ and test function $\phi$
\begin{equation}
\label{eq:ent-inequ}
\int \phi d\mu_t^N \le \tfrac{H(\mu_t^N | {\tilde \mu}_t^N) }{\alpha} +\cfrac{1}{\alpha} \log \left( \int e^{\alpha \phi} d{\tilde \mu}_t^N \right).
\end{equation}
We take then $\alpha=\delta N$, $\delta>0$, and
$$\phi=\left| \frac1N \sum_{x \in\TT_N} H({\tfrac{x}{N}}) {\mc E}_x - \int_{\TT} H(y) {\mf e} (t,y)dy \right|.$$
Since ${\tilde \mu}_t^N$ is fully explicit and even product, by using large deviations estimates, it is possible to show that
\begin{equation}
\limsup_{N \to \infty} \cfrac{1}{\delta N} \log \left( \int e^{\delta N \phi} d{\tilde \mu}_t^N \right) =I(\delta)
\end{equation}
where $I(\delta ) \to 0$ as $\delta \to 0$. By using (\ref{eq:Hyauest}), we are done. It remains then to prove (\ref{eq:Hyauest}) and for this we rely on a Gronwall inequality for the entropy production ($C>0$ is a constant)
\begin{equation}
\label{eq:entprodbound}
\partial_t H_N \le C H_N (t) + o(N).
\end{equation}
The proof of (\ref{eq:entprodbound}) is quite evolved and we refer the interested reader to \cite{Sim}, \cite{BO-livre1} (see also \cite{KL} for some overview on the subject). It is in this step that the macro-ergodicity of the dynamics is used in order to derive the so-called one-block estimate.
For non-gradient systems, i.e. systems such that the microscopic currents of the conserved quantities are not given by discrete gradients {\footnote{Observe that if a system is gradient then a microscopic fluctuation-dissipation equation (\ref{eq:2:fluc-diss}) holds with a zero fluctuating term.}}, the previous strategy has to be modified. Indeed, in order to have (\ref{eq:Hyauest}) it is necessary to replace the local equilibrium Gibbs state ${\tilde \mu}_t^N$ by a local equilibrium state with a first order correction term of the form
\begin{equation}
\label{eq:firstoderGibbs}
\begin{split}
&d{\hat \mu}_t^{N} (d\br, d\bp) \\
&= Z_{t,N}^{-1} \, \prod_{x \in \TT_N} \exp \left\{ -\beta_t (x/N) ({\mc E}_x -\tau_t (x/N) r_x) +\tfrac{1}{N} F(t, x/N) (\theta_x g) (\br, \bp) \right \} dr_x dp_x
\end{split}
\end{equation}
where $Z_{t,N}$ is a normalization constant,
\begin{equation*}
\tau_t = {\mf r}_t, \quad \beta_t = \{ {\mf e}_t - \tfrac{{\mf r}_t^2}{2}\}^{-1}
\end{equation*}
and the functions $F$ and $g$ are judiciously chosen. The choice is guided by the fluctuation-dissipation relation (\ref{eq:fluct-diss-eq-harm}) and done in order to obtain the first order "Taylor expansion" (\ref{tay}) below.
Let $\Omega^N =(\RR \times \RR)^{\TT_N}$ be the configurations space and denote
\begin{equation}
{\hat H}_N(t):=H\left(\mu_t^N \vert {\hat \mu}_t^N \right)=\int_{\Omega^N} f_t^N(\omega) \log \frac{f_t^N(\omega)}{\phi_t^N(\omega)} d\nu_* (\omega)\ ,
\label{eq:relentr2}
\end{equation}
where $f_t^N$ is the density of $\mu_t^N$ with respect to the Gibbs reference measure $\nu_*:=\nu_{1,0}$. In the same way, $\phi_t^N$ is the density of ${\hat \mu}_t^N$ with respect to $\nu_*$ (which is fully explicit). The goal is to get (\ref{eq:entprodbound}) with $H_N$ replaced by ${\hat H}_N$ .
We begin with the following entropy production bound. Let us denote by ${\mc L}_N^*=-\mc A_N+\gamma {\mc S}_N$ the adjoint of $\mc L_N$ in $\mathbb{L}^2(\nu_*)$.
\begin{lemma}
\label{entropy}
\begin{equation*}
\partial_t {\hat H} _N(t) \leq \int \frac{1}{\phi_t^N}\left(N^2\mc L_N^*\phi_t^N-\partial_t\phi_t^N\right) f_t^N d\nu_{*}= \int \left[ \frac{1}{\phi_t^N}\left(N^2\mc L_N^*\phi_t^N-\partial_t\phi_t^N\right) \right] d\mu_t^N \ .
\end{equation*}
\end{lemma}
\begin{proof}
We have that $f_t^N$ solves the Fokker-Plack equation $\partial_t f_t^N = N^2 {\mc L}_N^* f_t^N$. Assuming it is smooth to simplify, we have
\begin{equation*}
\begin{split}
\partial_t {\hat H} _N(t)& = \int \partial_t f^N_t [ 1+ \log f_t^N] d\nu_* - \int \partial_t f_t^N \log \phi_t^N d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*\\
&=N^2 \int {\mc L}_N^* f_t^N [\log f_t^N -\log \phi_t^N] d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*\\
&= N^2 \int f_t^N {\mc L}_N [ \log \tfrac{f_t^N}{\phi_t^N}] d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*\\
&= N^2 \int \tfrac{f_t^N}{\phi_t^N} {\mc L}_N [ \log \tfrac{f_t^N}{\phi_t^N}] \phi_t^N d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*\\
& \le N^2 \int {\mc L}_N [ \tfrac{f_t^N}{\phi_t^N}] \phi_t^N d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*\\
&=N^2 \int \cfrac{f_t^N}{\phi_t^N} \, {\mc L}_N^*\phi_t^N d\nu_* - \int \partial_t \phi_t^N \frac{f_t^N}{\phi_t^N} d\nu_*
\end{split}
\end{equation*}
where we used that for any positive function $h$, $h {\mc L}_N (\log h) \le {\mc L}_N h$ (this is a consequence of Jensen's inequality).
\cqfd\end{proof}
We define $\xi_x:=({\mc E}_x,r_x)$ and $\pi(t,q):=({\mf e}(t,q),{\mf r}(t,q)).$ If $f$ is a vectorial function, we denote its differential by $Df$.
\begin{proposition}[\cite{Sim}]
\label{prop:M1}
Let $(\lambda, \beta)$ be defined by $\beta= ({\mf e}- \tfrac{{\mf r}^2}{2})^{-1}$ and $\lambda= -\beta {\mf r}$. The term $ (\phi_t^N)^{-1}\left(N^2\mc L_N^*\phi_t^N-\partial_t\phi_t^N\right)$ can be expanded as
\begin{multline}
(\phi_t^N)^{-1}\left(N^2\mc L_N^*\phi_t^N-\partial_t\phi_t^N\right) \\
=\sum_{k=1}^5 \sum_{x \in \T_N} v_k\left(t,\frac{x}{N}\right) \left[J_x^k-H_k\left({\pi}\left(t,\frac{x}{N}\right)\right)-(DH_k)\left({\pi}\left(t,\frac{x}{N}\right)\right)\cdot \left({\xi}_x-{\pi}\left(t,\frac{x}{N}\right)\right)\right] \\ +o(N)
\label{tay}
\end{multline}
where
\begin{equation*}
\begin{array}{| c | c | c | c |}
\hline k & J_x^k & H_k(\mf{e},\mf{r}) & v_k(t,q) \\
\hline \hline 1 & p_x^2+r_xr_{x-1}+2\gamma p_x r_{x-1} & \mf{e}+{\mf r}^2/2 & -(2\gamma)^{-1} \partial^2_q \beta(t,q) \\
2 & r_x+\gamma p_x & {\mf r} & -{\gamma}^{-1} \partial^2_q\lambda(t,q) \\
3 & p_x^2\ (r_x+r_{x-1})^2 & (2{\mf e}-{\mf r}^2) \left({\mf e}+3{\mf r}^2/2\right) & (4\gamma)^{-1} [\partial_q \beta(t,q)]^2 \\
4 & p_x^2 \ (r_x+r_{x-1}) & {\mf r} \ (2{\mf e}-{\mf r}^2)& {\gamma}^{-1} \partial_q\beta(t,q) \ \partial_q\lambda(t,q) \\
5 & p_x^2 & {\mf e}-{\mf r}^2/2 & {\gamma}^{-1} [\partial_q \lambda(t,q)]^2 \\
\hline
\end{array}
\end{equation*}
\end{proposition}
Observe that $H_k (e,r)$ is equal to $\int J_x^k d\nu_{\beta,\tau}$ where $\beta,\tau$ are related to $e,r$ by the thermodynamic relations. Thus, the terms appearing in the righthand side of (\ref{tay}) can be seen as first order ``Taylor expansion". The form of the first order correction in (\ref{eq:firstoderGibbs}) plays a crucial role in order to get such expansions.
A priori the first term on the right-hand side of (\ref{tay}) is of order $N$, but we want to take advantage of these microscopic Taylor expansions to show it is in fact of order $o(N)$.
First, we need to cut-off large energies in order to work with bounded variables only. To simplify, we assume they are bounded ab initio.
Let $\ell$ be some integer (dividing $N$). We introduce some averaging over microscopic blocks of size $\ell$ and we will let $\ell \to \infty$ after $N\to \infty$. We decompose $\TT_N$ in a disjoint union of $p=N/\ell$ boxes $\Lambda_\ell (x_j)$ of length $\ell$ centered at $x_j$, $j \in \{ 1, \ldots,p\}$. The microscopic averaged profiles in a box of size $\ell$ around $y \in \TT_N$ are defined by
\begin{equation*}
{\tilde \xi}_{\ell} (y) = \cfrac{1}{\ell} \sum_{x \in \Lambda_\ell (y)} \xi_x.
\end{equation*}
Similarly we define
$${\tilde J}^k_\ell (y) =\cfrac{1}{\ell} \sum_{x \in \Lambda_\ell (y)} J_x^k.$$
In (\ref{tay}) we rewrite the sum $\sum_{x \in {\TT}_N}$ as $\sum_{j=1}^p \sum_{x \in \Lambda_{\ell} (x_j)}$ and, by using the smoothness of the function $v_k$, $H_k$, it is easy to replace the term
$$\cfrac{1}{N}\sum_{x \in \T_N} v_k\left(t,\frac{x}{N}\right) \left[J_x^k-H_k\left({\pi}\left(t,\frac{x}{N}\right)\right)-(DH_k)\left({\pi}\left(t,\frac{x}{N}\right)\right)\cdot \left({\xi}_x-{\pi}\left(t,\frac{x}{N}\right)\right)\right] $$
by
$$\cfrac{1}{p} \sum_{j=1}^p v_k\left(t,\frac{x_j}{N}\right) \left[{\tilde J}^k_\ell (x_j) \, - \, H_k\left({\pi}\left(t,\frac{x_j }{N}\right)\right)-(DH_k)\left({\pi}\left(t,\frac{x_j}{N}\right)\right)\cdot \left({\tilde \xi}_\ell (x_j) -{\pi}\left(t,\frac{x_j}{N}\right)\right)\right]$$
in the limit $N, \ell \to \infty$ with some error term of order $o(1)$.
Then, the strategy consists in proving the following crucial estimate, often called the {\textit{one-block estimate}}: we replace the empirical average current ${\tilde J}^k_\ell (x_j)$ which is averaged over a box centered at $x_j$ by its mean with respect to a Gibbs measure with the parameters corresponding to the microscopic averaged profiles ${\tilde \xi}_\ell (x_j)$, i.e. $H_k ({\tilde \xi}_\ell (x_j) )$. This non-trivial step is achieved thanks to some compactness argument and the macro-ergodicity of the dynamics.
Consequently we have to deal with terms in the form
\begin{equation}
\label{eq:phild}
\begin{split}
\cfrac{1}{p} \sum_{j=1}^p v_k\left(t,\frac{x_j}{N}\right) & \left[ H_k \left( {\tilde \xi}_\ell (x_j) \right) \, - \, H_k\left({\pi}\left(t,\frac{x_j }{N}\right)\right) \right.\\
& \left. -(DH_k)\left({\pi}\left(t,\frac{x_j}{N}\right)\right)\cdot \left({\tilde \xi}_\ell (x_j) -{\pi}\left(t,\frac{x_j}{N}\right)\right)\right]
\end{split}
\end{equation}
The final step consists then in applying the entropy inequality (\ref{eq:ent-inequ}) with respect to ${\hat \mu}_t^N$ with $\phi:=\phi_{\ell, N}$ given by (\ref{eq:phild}) and $\alpha=\delta N$, $\delta>0$ fixed but small. This will produce some term of order ${\hat H}_N (t) /N$ plus the term
\begin{equation*}
\limsup_{\ell \to \infty} \limsup_{N \to \infty} \cfrac{1}{\delta N} \log \left( \int e^{\delta N \phi} d{\hat \mu}_t^N \right) =I(\delta).
\end{equation*}
By using some large deviations estimates (observe that ${\hat \mu}_t^N$ is explicit and product at first order in $N$) one can show that $I(\delta)$ is nonpositive for $\delta$ sufficiently small. Thus we get the desired Gronwall inequality.
There is some additional difficulty that we hid under the carpet in the sketch of the proof. Since the state space is non compact, a control of high energies is required for the initial cut-off. This is a highly non trivial problem {\footnote{A similar problem appears in \cite{OVY} where the authors derived Euler equations for a gas perturbed by some ergodic noise. There, to overcome this difficulty, the authors replace ab initio the kinetic energy by the relativistic kinetic energy.}}. In the harmonic case considered here this control is obtained thanks to the following remark: the set of mixtures of Gaussian probability measures {\footnote{A Gibbs local equilibrium state is a Gaussian state in the harmonic case.}} is preserved by the (harmonic) velocity-flip model. Since for Gaussian measures all the moments are expressed in terms of the covariance matrix, required bounds can be obtained by a suitable control of the covariance matrices appearing in the mixture.
The extension of this result in the anharmonic case is a challenging open problem (see however \cite{Olla-Sasada} where equilibrium fluctuations are considered for an anharmonic chain perturbed by a conservative noise acting on the momenta and positions).
\subsubsection{Fourier's law}
Since in the harmonic case an exact fluctuation-dissipation equation is available Fourier's law can be obtained without too much work {\footnote{The {\textit{a posteriori}} simple but fundamental remark that an exact fluctuation-dissipation equation exists for the harmonic model (see (\ref{eq:fluct-diss-eq-harm})) is the real contribution of \cite{BO1}. }}.
\begin{theorem}[\cite{BO1,BO2}]
\label{thm:BO1}
Consider the one-dimensional harmonic chain in contact with two heat baths and with forced boundary conditions as in Section \ref{subsec:lrGK}. Then Fourier's law holds:
\begin{equation}
\label{eq:2:flwertyui}
{\tilde J}_s:=\lim_{N \to \infty} N \langle j_{0,1}^e \rangle_{ss} = \cfrac{1}{2\gamma} \left\{ (T_\ell -T_r) +(\tau_\ell^2 -\tau_r^2)\right\}
\end{equation}
and we have
\begin{equation}
\label{eq:jthermh}
\begin{split}
{\hat J}_\ell=\lim_{N \to \infty} N(\langle p_1^2 \rangle_{ss} -T_\ell) =\cfrac{1}{2\gamma \gamma_{\ell}} \left[ (T_r -T_\ell) +(\tau_{\ell} -\tau_r)^2 \right],\\
{\hat J}_r=\lim_{N \to \infty} N(T_r -\langle p_N^2 \rangle_{ss} )= \cfrac{1}{2\gamma \gamma_r} \left[ (T_r -T_\ell) - (\tau_{\ell} -\tau_{r})^2 \right].
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
We divide the proof in two steps:
\begin{itemize}
\item We first prove that there exists a constant $C$ independent of $N$ such that $|\langle j_{0,1}^e \rangle_{ss} | \le C/N$. This is obtained by using the fluctuation-dissipation equation and the fact that $\langle j^e_{x,x+1}\rangle_{ss}$ is independent of $x$:
\begin{eqnarray}
\langle j^e_{0,1} \rangle_{ss} &=& \cfrac{1}{N-3} \sum_{x=2}^{N-2} \langle j^e_{x,x+1} \rangle_{ss}\\ \nonumber
&=& -\cfrac{1}{2\gamma} \cfrac{1}{N-3} \sum_{x=2}^{N-2} \left\langle\nabla \left[ p_{x}^2 + r_{x} r_{x+1}\right]\right\rangle_{ss}\\ \nonumber
&=& \cfrac{1}{2\gamma}{\frac{1}{N-3}} \left\{ (\langle p_2^2 \rangle_{ss} +\langle r_2 r_3\rangle_{ss} )-(\langle p_{N-1}^2 \rangle_{ss} +\langle r_{N-1} r_{N}\rangle_{ss} )\right\}.
\end{eqnarray}
By using simple computations, one can show that $(\langle p_2^2 \rangle_{ss} +\langle r_2 r_3\rangle_{ss} )-(\langle p_{N-1}^2 \rangle_{ss} +\langle r_{N-1} r_{N}\rangle_{ss} )$ is uniformly bounded in $N$ by a positive constant.
\item Now we have only to evaluate the limit of each term appearing in $(\langle p_2^2 \rangle_{ss} +\langle r_2 r_3\rangle_{ss} )-(\langle p_{N-1}^2 \rangle_{ss} +\langle r_{N-1} r_{N}\rangle_{ss} )$. Notice that assuming local equilibrium we easily get the result. The first step implies that $\langle j_{0,1}^e \rangle_{ss}$ and $ \langle j_{N,N+1}^{e} \rangle_{ss}$ vanish as $N \to +\infty$. Since $V_s$ goes to $0$ by Lemma \ref{lem:velo}, one has that $\langle p_1^2 \rangle_{ss}$ and $\langle p_{N}^2 \rangle_{ss}$ converge respectively to $T_\ell$ and $T_r$. By using some ``entropy production bound'' one can propagate this local equilibrium information to the particles close to the boundaries and show (\ref{eq:2:flwertyui}).
\end{itemize}
\cqfd\end{proof}
It follows from this Theorem that the system can be used as a heater but not as a refrigerator. Assume for example that $T_r > T_{\ell}$. The term ${\hat J}_\ell$ (resp. ${\hat J}_r$) is the macroscopic heat current from the left reservoir to the system (resp. from the system to the right reservoir). Whatever the values of $\tau_\ell, \tau_r$ are, ${\hat J}_{\ell} >0$ and we can not realize a refrigerator. But if $(T_r -T_{\ell}) < (\tau_r -\tau_{\ell})^2$ then ${\hat J}_r <0$ and we realized a heater.\\
The proof of the validity of Fourier's law for anharmonic chains perturbed by an energy conserving noise is still open.
\subsubsection[Macroscopic Fluctuation Theory] {Macroscopic Fluctuation Theory for the energy conserving harmonic chain}
The macroscopic fluctuation theory (\cite{BdSGJLL1}) is a general approach developed by Bertini, De Sole, Gabrielli, Jona-Lasinio and Landim to calculate the large deviation functional of the empirical profiles of the conserved quantities of Markov processes in a NESS. Its main interest is that it can be applied to a large class of boundary driven diffusive systems and does not require the explicit form of the NESS but only the knowledge of two thermodynamic macroscopic parameters of the system, the diffusion coefficient $D(\rho)$ and the mobility $\chi(\rho)$. This theory can be seen as an infinite dimensional generalization of the Freidlin-Wentzel theory \cite{Freidlin-Wentzel} and is based on the large deviation principle for the hydrodynamics of the system.
In order to explain (roughly) the theory we consider for simplicity a Markovian system $\{ \eta (t) :=\{ \eta_x (t) \in \RR \, ; \, x \in \{1, \ldots N\} \}_{t \ge 0}$ with only one conserved quantity, say the density $\rho$, in contact with two reservoirs at each extremity. Here $N$ is the size of the system which will be sent to infinity. We denote by $\mu^N_{ss}$ the nonequilibrium stationary state of $\{\eta(t)\}_{t \ge 0}$. For any microscopic configuration $\eta:=\{\eta_x \, ;\, x \in \{1, \ldots,N\} \}$ let
\begin{equation*}
\pi^N (\eta, \cdot)=\sum_{x=1}^{N-1} \eta_x {\bf 1}_{\big[\tfrac{x}{N}, \tfrac{x+1}{N}\big)} (\cdot)
\end{equation*}
be the empirical density profile. In the diffusive time scale, we assume that $\pi^N (\eta (tN^2), \cdot)$ converges as $N$ goes to infinity to $\rho_t (\cdot):=\rho(t,\cdot)$ solution of
\begin{equation*}
\begin{cases}
\partial_t \rho = \partial_y (D(\rho) \partial_y \rho), \quad y \in [0,1], \quad t\ge 0,\\
\rho(t,0) =\rho_\ell, \quad \rho(t,1)=\rho_r,\quad t \ge 0, \\
\rho(0, \cdot) =\rho_0 (\cdot)
\end{cases}
\end{equation*}
where $\rho_0 (\cdot)$ is the initial density profile, $D(\rho)>0$ is the diffusion coefficient and $\rho_\ell$, $\rho_r$ the densities fixed by the reservoirs. As $t \to \infty$ the solution $\rho_t$ of the hydrodynamic equation converges to a stationary profile $\bar \rho :[0,1] \to \RR$ solution of $D(\bar \rho) \partial_y {\bar \rho} =J=const.$ with ${\bar \rho}(0)=\rho_\ell$, ${\bar \rho} (1)=\rho_r$. We assume that under $\mu_{ss}^N$, the empirical density profile $\pi^N (\eta, \cdot)$ converges to ${\bar \rho}$. This assumption is nothing but a law of large numbers for the random variables $\pi^N$.
We are here interested in the corresponding large deviation principle. Thus, we want to estimate the probability that in the NESS $\mu_{ss}^N$ the empirical density profile $\pi^N$ is close to an atypical macroscopic profile $\rho(\cdot) \neq {\bar \rho}$. This probability typically is of order $e^{-N {\bb V} (\rho)}$ where $\bb V$ is the rate function:
\begin{equation*}
\mu_{ss}^N ( \pi^N (\eta, \cdot) \approx \rho(\cdot) ) \approx e^{-N {\bb V} (\rho)}.
\end{equation*}
The goal of the macroscopic fluctuation theory is to obtain information about this functional.
The condition to be fulfilled by the system to apply the theory of Bertini et al. is that it satisfies a {\textit{dynamical large deviation principle}} with a rate function which takes a quadratic form {\footnote{Such property has been proved to be valid for a large class of stochastic dynamics (\cite{KOV89}, \cite{KL}).}} like (\ref{eq:formI}).
Let us first explain what we mean by dynamical large deviation principle. Imagine we start the system from a Gibbs local equilibrium state corresponding to the macroscopic profile $\rho_0$. We want to estimate the probability that the empirical density $\pi^N (\eta (tN^2), \cdot)$ is close during the macroscopic time interval $[0,T]$, $T$ fixed, to a smooth macroscopic profile $\gamma (t, y)$ supposed to satisfy \footnote{This assumption avoids taking into account the cost to produce the initial profile, cost which is irrelevant for us.} $\gamma(0,\cdot)=\rho_0$. This probability is exponentially small in $N$ with a rate $I_{[0,T]} ( \gamma\, | \, {\rho_0})$
\begin{equation}
\label{eq:IIII}
{\PP} \left[ \pi^N ( \eta (tN^2), y) \approx \gamma (t,y), \; (t,y) \in [0,T] \times [0,1] \right] \sim e^{- N I_{[0,T]} (\gamma | \rho_0)}.
\end{equation}
The rate function is assumed to be of the form
\begin{equation}
\label{eq:formI}
\begin{split}
I_{[0,T]} (\gamma \, | \, \rho_0) & = \frac{1}{2} \int_0^T dt \int_0^1 dy \; \chi (\rho (t,y)) \left[ (\partial_y H) (t,y) \right]^2
\end{split}
\end{equation}
where $ \partial_y H$ is the extra gradient external field needed to produce the fluctuation $\gamma$, namely such that
\begin{equation}
\label{eq:Hgamma}
\partial_t \gamma = \partial_y \left [ D(\gamma) \partial_y \gamma \, - \, \chi(\gamma) \partial_y H \right].
\end{equation}
Thus, $I_{[0,T]} (\gamma | \rho_0)$ is the work done by the external field $\partial_y H$ to produce the fluctuation $\gamma$ in the time interval $[0,T]$. The function $\chi$ appearing in (\ref{eq:Hgamma}) is the second thermodynamic parameter (with the diffusion coefficient $D$) mentioned in the beginning of this section. The two parameters $D$ and $\chi$ are in fact related together by the Einstein relation so that knowing one of them and the Gibbs states of the microscopic model is sufficient to obtain the second.
To show this result the strategy is the following. We perturb the Markov process $\{\eta (t)\}_{t \ge 0}$ thanks to the function $H:=H(\gamma)$, which is solution of the Poisson equation (\ref{eq:Hgamma}), by adding locally a small space inhomogeneous drift provided by $\partial_y H$. In doing so we obtain a new Markov process $\{ \eta^H (t)\}_{t \ge 0}$ such that in the diffusive time scale $\pi^N (\eta^H (tN^2), \cdot)$ converges to $\gamma (\cdot)$. Let ${\bb P}^H$ (resp. $\PP^0$) be the probability measure on the empirical density paths space induced by $\{\eta^H (tN^2)\}_{t \in[0,T]}$ (resp. $\{\eta(tN^2)\}_{t \in [0,T]}$) . Then, by using hydrodynamic limits techniques similar to the ones explained in Section \ref{subsec:hl-vf} we show that in the large $N$ limit, under ${{\mathbb P}^H}$, the Radon-Nikodym derivative is well approximated by {\footnote{We use Girsanov transform to express the Radon-Nikodym derivative. A priori it is not a functional of the empirical density and we need to establish some {\textit{replacement lemma}} (see \cite{KL}). }}
\begin{equation*}
\cfrac{{\rm d}\PP^0}{{\rm d}{{\PP^H}}} (\pi ) \approx \exp\left\{ -N I_{[0,T]}( \pi | \rho_0)\right\}.
\end{equation*}
Here $\pi:=\{\pi (t,y)\, ; \, t \in[0,T], y \in [0,1]\}$ is any space-time density profile. Thus, since
\begin{equation*}
\PP^0 \left[\pi^N (\eta (tN^2), \cdot) \sim \gamma(t, \cdot),\; t\in [0,T]\right]
={\EE^H} \left[ \cfrac{{\rm d}\PP^0}{{\rm d}{{\PP}^H}} (\pi) \, {\bf 1}_{\left\{ \pi (t, \cdot) \sim \gamma (t, \cdot), t \in [0,T] \right\} }\right]
\end{equation*}
we obtain (\ref{eq:IIII}).
The macroscopic fluctuation theory claims that the large deviations functional ${\bb V} (\rho)$ of the empirical density in the NESS coincides with the quasi-potential ${\bb W} (\rho)$ defined by
\begin{equation*}
{\bb W} (\rho) = \inf_{\substack{\gamma: \gamma (-\infty) = {\bar \rho}\\ \gamma (0) =\rho}} I_{[-\infty, 0]} (\gamma | {\bar \rho}).
\end{equation*}
Here $I_{[-\infty, 0]}$ is obtained from $I_{[0,T]}$ by a shift in time by $-T$, $T$ being sent to $+\infty$ afterwards. In words, the quasi potential determines the cost to produce a fluctuation equal to $\gamma$ at $t=0$ when the system is macroscopically in the stationary profile ${\bar \rho}$ at $t=-\infty$.
Thus, the problem is reduced to computing $W$. It can be shown that $W$ solves (at least formally) the infinite-dimensional Hamilton-Jacobi equation
\begin{equation}
\frac{1}{2} \left \langle \partial_y \left[ \frac{\delta {\bb W}}{ \delta \rho} \right] \, , \chi( \rho) \, \partial_y \left[ \frac{\delta {\bb W}}{ \delta \rho} \right] \right\rangle + \left\langle \frac{\delta {\bb W}}{ \delta \rho}\, , \, \partial_y \left[ D(\rho) \partial_y \rho \right] \right\rangle =0
\end{equation}
where $\langle \cdot, \cdot \rangle$ denotes the usual scalar product in ${\mathbb L}^2 ([0,1])$. Note that there is no uniqueness of solutions (${\bb W}=0$ is a solution) and up to now a general theory of infinite dimensional Hamilton-Jacobi equations is still missing. This implies that we have in fact to solve by hand the variational problem and the solution is only known for few systems. This is an important limitation of the macroscopic fluctuation theory. Even getting interesting qualitative properties on ${\bb W}$ is difficult.
The rigorous implementation of this long program has only been carried for the boundary driven Symmetric Simple Exclusion Process and extended with less rigor to a few other systems (see \cite{BDSGLL-SSEP}, \cite{Bod-Giac}, \cite{Farfan} for rigorous results).
Let us now try to apply this theory for the harmonic chain with velocity-flip noise. Since we have a fully explicit microscopic fluctuation-dissipation equation (even when some harmonic pinning is added) we can easily guess what is the form of the hydrodynamic equations under various boundary conditions by assuming that the propagation of local equilibrium in the diffusive time scale holds. Nevertheless, let us observe that a rigorous derivation is missing, the obstacle being a sufficiently good control of the high energies {\footnote{This control is only available in the case of periodic boundary conditions (\cite{Sim}).}}. The boundary conditions we impose to the system are the following. At the left (resp. right) end we put the chain in contact with a Langevin bath at temperature $T_\ell$ (resp. $T_r$) and consider the system with fixed boundary conditions or with forced boundary conditions with the same force $\tau$ at the two boundaries.
Then, for the unpinned chain, the equations (\ref{eq:2:hl-re1}) are still valid but they are supplemented with the boundary conditions (\cite{BKLL1})
\begin{equation}
\left[{\mf e} -\cfrac{{\mf r}^2}{2} \right] (t,0)= T_\ell, \quad \left[{\mf e} -\cfrac{{\mf r}^2}{2} \right] (t,1)= T_r,
\end{equation}
since the Langevin baths fix the temperatures at the boundaries and
\begin{equation}
\partial_y {\mf r} (t, 0) = \partial_y {\mf r} (t, 1)=0
\end{equation}
for fixed boundary conditions (the total length of the chain is constant {\footnote{Indeed, by (\ref{eq:2:hl-re1}), we have $\partial_t (\int_{0}^1 {\mf r} (t,y) dy) = \gamma^{-1} \int_{0}^1 {\partial}_y^2 {\mf r} (t,y) dy= \gamma^{-1} [ \partial_y {\mf r} (t, 1) -\partial_y {\mf r} (t,0) ] =0$.}}) and
\begin{equation}
{\mf r} (t, 0) = {\mf r} (t, 1)= \tau
\end{equation}
for forced boundary conditions.
If the chain is pinned by the harmonic potential $W(q)=\nu q^{2} /2$ then only the energy is conserved and the macroscopic diffusion equation takes the form
\begin{equation}
\begin{cases}
\partial_t {\mf e} = \partial_{y} (\kappa \partial_y {\mf e}),\\
{\mf e}(0,y) = {\mf e}_0 (y),\\
{\mf e} (t,0)= T_\ell, \; {\mf e} (t,1)=T_r,
\end{cases}
\quad y \in (0,1)
\end{equation}
where the conductivity $\kappa$ is equal to (\cite{BKLL1})
\begin{equation}
\label{eq:conductivityuiop}
\kappa= \cfrac{1/\gamma}{2 + \nu^2 + \sqrt{\nu (\nu +4)}}.
\end{equation}
Assuming a good control of high energies, it is possible to derive the dynamical large deviations function of the empirical conserved quantities. The goal would be to compute the large deviation functional of the NESS which according to the macroscopic fluctuation theory coincides with the quasi potential. We recall that the quasi potential is defined by a variational problem and that it depends only on two thermodynamic quantities, the diffusion coefficient and the mobility (the latter are matrices if several conserved quantities are involved).
Let us first consider the pinned velocity flip model where the energy is the only conserved quantity. It turns out that the mobility is a quadratic function. Consequently, the methods exposed in Theorem 6.5 of \cite{BdSGJLL2} apply and the variational formula can be computed. The quasi potential ${\bb V}(\cdot)$ is given by (\cite{BKLL1})
\begin{equation}
\label{eq:2:VKMP}
{\bb V} (e)=\int_0^1 dq \left[ \cfrac{e(q)}{F(q)} -1 -\log\left( \cfrac{e(q)}{F(q)}\right) - \log \left( \cfrac{F' (q)}{T_r -T_\ell }\right) \right]\, ,
\end{equation}
where $F$ is the unique non decreasing solution of
\begin{equation}
\begin{cases}
{F''} = \cfrac{F-e}{F^2} (F')^2\, ,\\
F(0)=T_\ell , \; F(1) =T_r\, .
\end{cases}
\end{equation}
Surprisingly, the function ${\bb V}$ is independent of the pinning value $\nu$ and of the intensity of the noise $\gamma$. It is thus natural to conjecture that in the NESS of the unpinned velocity flip model the large deviation function of the empirical energy profile coincides with ${\bb V}$ but we did not succeed to prove it. Observe that at equilibrium ($T_\ell=T_r$), $F(q)=T_\ell=T_r$ and the last term in (\ref{eq:2:VKMP}) disappears so that the quasi potential is local. On the other hand, if $T_\ell \ne T_r$, this is no longer the case and this reflects the presence of long-range correlations in the NESS. In particular, an approximation of the NESS by a Gibbs local equilibrium state in the form (\ref{eq:GLES-ness}) would not give the correct value of the quasi potential.
For the unpinned chain we have two conserved quantities. Solving the variational problem of the quasi potential for these two conserved quantities is a very difficult open problem{\footnote{Here we do not have any exactly solvable model like the Symmetric Simple Exclusion Process which could give us some hints for the form of the quasi-potential.}} (see \cite{B2} for a partial result for some other stochastic perturbation of the harmonic chain).
\section{Anomalous diffusion}
\label{ch:anomalous}
An anomalous large conductivity is observed experimentally in carbon nanotubes and numerically in chains of oscillators without pinning, where numerical evidence shows a conductivity diverging with the size $N$ of the system like $N^{\alpha}$, with $\alpha<1$ in dimension $d=1$, and like $\log N$ in dimension $d=2$. If some nonlinearity is present in the interaction, finite conductivity is observed numerically in all pinned case or in dimension $d \ge 3$ (\cite{Dhar},\cite{LLP1}). Consequently it has been suggested that conservation of momentum is an important ingredient for the anomalous conductivity in low dimensions (see however \cite{ZZWZ}).
In Chapter \ref{ch:normal} we considered chains of oscillators perturbed by a noise conserving only energy and destroying the possible momentum conservation. In the harmonic case we obtained Fourier's law and in the anharmonic case we proved existence of the Green-Kubo formula for the thermal conductivity.
In this chapter the added perturbation conserves both energy and momentum (energy and volume for the Hamiltonian systems considered in Section \ref{sec:intro-phd}). These systems qualitatively have the same behavior as Hamiltonian chains of oscillators (without any noise), i.e. anomalous transport for unpinned chain in dimension $d=1,2$ and normal transport otherwise. We could even be more optimistic and hope that they share with the deterministic systems common limits for the energy fluctuation fields, two point correlation functions ${\ldots}$ This is because one expects that the microscopic details of the dynamics are irrelevant. Therefore some {\textit{universality}} should hold. Recently H. Spohn (\cite{Sp-NFH}), by following ideas of \cite{VB}, used the nonlinear fluctuating hydrodynamics theory to classify very precisely the different expected universality classes. The nonlinear fluctuating hydrodynamics theory is based on the assumption that the microscopic dynamics evolve in the Euler time scale according to a system of conservation laws. The theory is macroscopic in the sense that all the predictions are done starting from this system of conservation laws without further references to the microscopic dynamics. Since we have seen that the presence of the energy-momentum conserving noise does not change the form of the hydrodynamic equations, the theory claims in fact that the limit of the fluctuations fields of the conserved quantities for purely deterministic chains of oscillators and for noisy energy-momentum conserving chains are exactly the same.
\subsection{Harmonic chains with momentum exchange noise}
Getting some information on the behavior of the energy fluctuation field in the large scale limit remains challenging. So far, satisfactory but not complete results have only been obtained in the harmonic case. The anharmonic case is much more difficult.
In \cite{BBO1}, \cite{BBO2} we explicitly compute the time correlation current for a system of harmonic oscillators perturbed by an energy-momentum conserving noise {\footnote{It is straightforward to adapt the proofs given in \cite{BBO1}, \cite{BBO2} to the case of the momenta exchange noise.}} and we find that it behaves, for large times, like $t^{-d/2}$ in the unpinned cases, and like $t^{-d/2-1}$ when on-site harmonic potential is present.
These results are given in the Green-Kubo formalism. Their counterpart in the NESS formalism have been considered in \cite{LMMP} but a rigorous proof is still missing. Several variations of the Green-Kubo formula can be found in the literature: one can start with the infinite system in the canonical ensemble, as we did in Subsection~\ref{subsec:GK0}, or with a finite system, in the canonical or micro-canonical ensembles, sending the size of the system to infinity. It is widely believed that all these definitions coincide (also in the case of infinite conductivity). As shown in \cite{BBO2}, this is essentially true for the energy-momentum conserving harmonic chain. Here we consider the simplest possible definition avoiding to discuss the rigorous definition of the canonical ensemble in infinite volume and the problem of equivalence of ensembles.
The set-up is the following. We consider a chain perturbed by the energy-momentum conserving noise (see (\ref{eq:cons-noise-exch-mom})) with periodic boundary conditions. Its Hamiltonian is given by ${\mc H}_N=\sum_{x \in \TT_N^d} {\mc E}_x$ where the energy ${\mc E}_x$ of atom $x$ is
\begin{equation}
\label{eq:4-energy-site-x}
{\mc E}_x= \frac{|p_x|^2}2 + W(q_x) + \frac{1}{2} \sum_{\substack{|y-x|=1}} V(q_x - q_y).
\end{equation}
The system is considered at equilibrium under the Gibbs grand-canonical measure
\begin{equation*}
d\mu_{N,T} =\frac{e^{-{\mc H}_N /T}}{Z_{N,T}} d\bq d \bp
\end{equation*}
where $Z_{N,T}$ is the renormalization constant.
The Green-Kubo formula for the thermal conductivity in the direction $e_k$, $1 \le k \le d$, is {\footnote{By symmetry arguments this is independent of $k$.}} the limiting variance of the energy current $J^{e,\gamma}_{x,x+e_k} ([0,t])$ up to time $t$ in the direction $e_k$ in a space-time box of size $N \times t$:
\begin{equation}
\label{eq:5;GKtyu}
\kappa (T) = \frac{1}{2T^2} \lim_{t \to +\infty} \lim_{N \to \infty}{\mathbb E}_{\mu_{N,T}} \left[ \left(\frac{1}{\sqrt{N^{d} t}}\sum_{x \in \TT_N^d} J^{e,\gamma}_{x,x+e_1} ([0,t]) \right)^2 \right].
\end{equation}
The energy currents $\{ J^{e,\gamma}_{x,x+e_k} ([0,t])\, ; \, k=1, \ldots, d\}$ are defined by the energy conservation law
\begin{equation*}
{\mc E}_x (t) -{\mc E}_x (0) = \sum_{k=1}^d \left( J^{e,\gamma}_{x-e_k,x} ([0,t]) -J^{e, \gamma}_{x,x+e_k} ([0,t]) \right).
\end{equation*}
The energy current up to time $t$ can be written as
\begin{equation}
J^{e, \gamma}_{x,x+e_k} ([0,t]) = \int_0^t {j}_{x,x+e_k}^{e,\gamma} (s) ds + M_{x,x+e_k} (t)
\end{equation}
where $M_{x,x+e_k} (t)$ is a martingale and ${j}^{e,\gamma}_{x,x+e_k}$ is the instantaneous current which has the form
\begin{equation}
\label{eq:curebtjean}
{j}_{x,x+e_k}^{e,\gamma} ={\tilde j}^e_{x,x+e_k}+ \gamma \left[ p_{x+e_k}^2 -p_x^2 \right], \quad {\tilde j}^e_{x,x+e_k}= -\frac{1}{2} V' (q_{x+e_k} -q_x) (p_{x+e_k} + p_x).
\end{equation}
The term ${\tilde j}_{x,x+e_k}^e$ is the Hamiltonian contribution while the gradient term is due to the noise.
We now expand the square in (\ref{eq:5;GKtyu}). Notice first that since we have periodic boundary conditions the gradient term appearing in (\ref{eq:curebtjean}) does not contribute. By a time reversal argument one can show that the cross term between the martingale and the time integral of the instantaneous current vanishes. Moreover a simple computation shows that the square of the martingale term gives a contribution equal to $\gamma$ (see \cite{BBO2} for details). Thus we obtain
\begin{equation}
\label{eq:5;GKtyu6}
\begin{split}
\kappa (T) &= T^{-2} \lim_{t \to +\infty} \lim_{N \to \infty} \frac{1}{2N^{d} t}{\mathbb E}_{\mu_{N,T}} \left[ \left( \sum_{x \in \TT_N^d} \int_{0}^t {\tilde j}_{x,x+e_k}^e (s) ds \right)^2 \right] \; + \; \gamma \\
&=T^{-2} \lim_{t \to +\infty} \lim_{N \to \infty} \sum_{x \in \TT_N^d} \int_0^{+\infty} ds \left(1- \frac{s}{t} \right)^{+} {\mathbb E}_{\mu_{N,T}} \left[ {\tilde j}^e_{0,e_k} (0) \,{\tilde j}_{x,x+e_k}^e (s) \right] \, ds \; + \: \gamma
\end{split}
\end{equation}
where the last line is obtained by time and space stationarity of the Gibbs measure and $u^+$ denotes $\max (u,0)${\footnote{Observe that replacing $(1-\tfrac{s}{t})^+$ by $e^{-s/t}$ and $\lim_{N \to \infty} \sum_{x \in \TT_N^d}$ by $\sum_{x \in \ZZ^d}$ we formally get an expression similar to the Green-Kubo formula of Theorem \ref{th:GK}.}}. It is then clear that the divergence of the Green-Kubo formula, {\textit{i.e.}} anomalous transport, is due to a slow decay of the time correlation function $C(t)$ defined by
\begin{equation}
\label{eq:4:CTTT}
C(t)= \lim_{N \to \infty} \sum_{x \in \TT_N^d} {\mathbb E}_{\mu_{N,T}} \left[ {\tilde j}^e_{0,e_k} (0) \, {\tilde j}_{x,x+e_k}^e (t) \right].
\end{equation}
\begin{theorem}[\cite{BBO2}]
\label{th:GK-harm-exch}
Consider the harmonic case: $V (r )=\alpha r^2$, $W(q)= \nu q^2$ where $\alpha>0$ and $\nu \ge 0$. \\
Then the limit defining $C(t)$ in (\ref{eq:4:CTTT}) exists and can be computed explicitly. In particular, we have that $C(t) \sim t^{-d/2}$ if $\nu=0$ and $C(t) \sim t^{-d/2 -1}$ if $\nu>0$. \\
Consequently, the limit (\ref{eq:5;GKtyu6}) exists in $(0,+\infty]$ and is finite if and only if $d \ge 3$ or $\nu>0$. When finite, $\kappa (T)$ is independent of $T$ and can be computed explicitly.
\end{theorem}
\begin{proof}
We compute the Laplace transform $L_N (z)=\int_0^{+\infty} e^{-zt} C_N (t) dt$, $z>0$, of $C_N (t)=\sum_{x \in \TT_N^d} {\mathbb E}_{\mu_{N,T}} \left[ {\tilde j}_{0,e_k} (0) \, {\tilde j}_{x,x+e_k}^e (t) \right]$. Since we have
\begin{equation*}
L_N (z) = N^{-1} \mu_{N,T} \left[ \left(\sum_{x \in \TT_N^d} {\tilde j}_{x,x+,e_k}^e \right) \, (z-{\mc L}_N)^{-1} \left( \sum_{x \in \TT_N^d} {\tilde j}_{x,x+,e_k}^e \right) \right]
\end{equation*}
it is equivalent to solve the resolvent equation $(z-{\mc L}_N) h_N = \sum_{x \in \TT_N^d} {\tilde j}_{x,x+e_k}^e$. Notice that ${\mc L}_N$ maps polynomial functions of degree $2$ into polynomial functions of degree $2$ and that $\sum_{x} {\tilde j}^e_{x,x+e_k}$ is a polynomial function of degree $2$. Thus, the function $h_N$ is a polynomial function of degree $2$. Moreover it has to be space translation invariant since $\sum_{x} {\tilde j}^e_{x,x+e_k}$ is. Therefore we can look for a function $h_N$ of the form
\begin{equation*}
h_N = \sum_{x,y} a (y-x) p_x p_y + \sum_{x,y} b (y-x) p_x q_y + \sum_{x,y} c (y-x) q_x q_y
\end{equation*}
where $a,b$ and $c$ are functions from $\TT_N^d$ into $\RR$. We compute explicitly $a,b$ and $c$ and we get $a=c=0$ while $b$ is the solution to
$$(z+2\nu -\gamma \Delta) b = - \alpha (\delta_{e_k} -\delta_{-e_k})$$
where $\Delta$ is the discrete Laplacian. Then we deduce $L_N (z)$, hence $C_N (t)$ by inverse Laplace transform. The limit $C(t) = \lim_{N \to + \infty} C_N (t)$ follows.
\cqfd\end{proof}
Consequently in the unpinned harmonic cases in dimension $d = 1$ and $2$, the conductivity
of our model diverges as $N$ goes to infinity. Otherwise it converges as $N \to \infty$. In the anharmonic case we obtained some upper bounds showing that the divergence cannot be worse than in the harmonic case. These upper bounds also show that the conductivity cannot be infinite if $d\ge 3$ (see \cite{BBO2} for details and precise statements).
\subsection{A class of perturbed Hamiltonian systems}
In \cite{BS} is proposed a class of models for which anomalous diffusion is observed. These models have been introduced in Section \ref{sec:intro-phd} of Chapter \ref{ch:models}. The goal of \cite{BS} was to show that these systems have a behavior very similar to that of the standard one-dimensional chains of oscillators conserving momentum {\footnote{They could be defined in any dimension.}}.
\subsubsection{Definition of thermodynamic variables}
\label{sec:def_thermo}
Let us fix a potential $V$ and consider the stochastic energy-volume conserving model defined by the generator ${\mc L}={\mc A} + \gamma {\mc S}$, $\gamma \ge 0$, where ${\mc A}$ and ${\mc S}$ are given by (\ref{eq:1:ASdynum2}). Recall that the Gibbs grand-canonical probability measures $\mu_{\beta, \lambda}$, $\beta >0$, $\lambda \in \RR$, defined on $\Omega$ by
\begin{equation*}
d\mu_{\beta,\lambda} (\eta) =\prod_{x \in \ZZ} Z(\beta,\lambda)^{-1} \exp \left\{ -\beta V(\eta_x) -\lambda \eta_x \right\} d\eta_x
\end{equation*}
form a family of invariant probability measures for the infinite dynamics. We assume that
the partition function $Z$ is well defined on $(0,+\infty) \times {\mathbb R}$. The following thermodynamic relations relate the chemical potentials $\beta, \lambda$ to the mean volume $v$ and the mean energy $e$ under $\mu_{\beta,\lambda}$:
\begin{equation}
\label{eq:tr}
\begin{split}
&v(\beta,\lambda) =\mu_{\beta,\lambda} (\eta_x)= -\partial_{\lambda} \Big(\log Z(\beta,\lambda)\Big), \\
&e(\beta,\lambda)= \mu_{\beta,\lambda} (V(\eta_x))= -\partial_{\beta}\Big(\log Z(\beta,\lambda)\Big).
\end{split}
\end{equation}
These relations can be inverted by a Legendre transform to express $\beta$ and~$\lambda$
as a function of~$e$ and~$v$.
Define the thermodynamic entropy $S \, : \, (0,+\infty) \times \RR \to [-\infty,+\infty)$ as
\begin{equation*}
S(e,v)= \inf_{ \lambda \in \RR, \beta >0} \Big\{ \beta e + \lambda v +
\log Z (\beta,\lambda) \Big\}.
\end{equation*}
Let ${\mc U}$ be the convex domain of $(0,+\infty) \times \RR$ where
$S(e,v) >- \infty$ and $\mathring{\mc U}$ its interior.
Then, for any $(e,v):=(e(\beta, \lambda),v(\beta,\lambda))
\in {\mathring{\mc U}}$, the parameters $\beta,\lambda$ can be obtained as
\begin{equation}
\label{eq:14}
\beta = (\partial_e S ) (e,v),
\qquad
\lambda= (\partial_v S) (e,v).
\end{equation}
We also introduce the tension
$\tau(\beta,\lambda)= \mu_{\beta,\lambda} (V' (\eta_0))=-\lambda / \beta$.
The microscopic energy current $j^{e,\gamma}_{x,x+1}$ and volume current $j^{v,\gamma}_{x,x+1}$ are given by
\begin{equation}
\label{eq:current_gamma}
\begin{split}
&j^{e,\gamma}_{x,x+1}= - V'(\eta_x) V' (\eta_{x+1}) - \gamma \nabla [ V(\eta_x)],\\
&j^{v,\gamma}_{x,x+1}= -[ V' (\eta_x) +V' (\eta_{x+1})] - \gamma \nabla [ \eta_x].
\end{split}
\end{equation}
With these notations we have
\begin{equation}
\label{eq:averages_of_currents}
\mu_{\beta,\lambda} (j_{x,x+1}^{e, \gamma})=-\tau^2,
\qquad
\mu_{\beta,\lambda} (j_{x,x+1}^{v,\gamma})=-2\tau.
\end{equation}
In the sequel, with a slight abuse of notation,
we also write $\tau$ for $\tau(\beta(e,v), \lambda(e,v))$
where $\beta(e,v)$ and $\lambda(e,v)$ are defined by relations~(\ref{eq:14}).
\subsubsection{Hydrodynamic limits}
\label{sec:hydroLimEuler}
Consider the finite \emph{closed} stochastic energy-volume dynamics with periodic boundary conditions,
that is the dynamics generated by ${\mc L}_{N,{\rm{per}}} ={\mc A}_{N,{\rm{per}}} + \gamma {\mc S}_{N,{\rm{per}}}$ where
\begin{equation}
\label{eq:A_per}
\Big({\mathcal A}_{N,{\rm per}} f \Big)(\eta) =
\sum_{x \in \TT_N} \left[V'(\eta_{x+1})-V'(\eta_{x-1}) \right] \partial_{\eta_x}f(\eta),
\end{equation}
and
\begin{equation*}
\Big({\mc S}_{N,{\rm per}} f\Big)(\eta) =\sum_{x \in \TT_N} \left[ f(\eta^{x,x+1}) -f(\eta) \right].
\end{equation*}
We choose to consider the dynamics on $\TT_N$ rather than on $\ZZ$ to avoid (nontrivial) technicalities. We are interested in the macroscopic behavior of the two conserved quantities on a macroscopic time-scale $Nt$ as $N \to \infty$.
\begin{remark} The results of this section shall be compared to the results of Section \ref{subsec:hl-vf}. For the velocity-flip model, the hydrodynamic limits where trivial in the Euler time scale. It was only in the diffusive time scale that some evolution of the profiles was observed and the hydrodynamic limits were given by parabolic equations (see (\ref{eq:2:hl-re1}). Here, the evolution is not trivial in the Euler time scale and the hydrodynamic limits are given by hyperbolic equations (see below (\ref{eq:syslim}).
\end{remark}
We assume that the system is initially distributed according to a local Gibbs equilibrium
state corresponding to a given energy-volume profile ${X_0} : \TT \to {\mathring{\mc U}}$:
\begin{equation*}
X_0=\left(
\begin{array}{c}
{\mf e}_0\\
{\mf v}_0
\end{array}
\right),
\end{equation*}
in the sense that, for a given system size~$N$,
the initial state of the system is described by the following product probability measure:
\begin{equation}
\label{eq:Ges}
d\mu_{{\mf e}_0, {\mf v}_0}^N(\eta) =
\prod_{x \in \TT_N} \frac{ \exp\left\{ - {\mf \beta}_0 (x/N) V(\eta_x)
- {\mf \lambda}_0 (x/N) \eta_x \right\}}{Z ( {\mf \beta}_0 (x/N), {\mf \lambda}_0 (x/N))}\,
d\eta_x,
\end{equation}
where $(\beta_0 (x/N), \lambda_0 (x/N))$ is actually a function of
$({\mf e}_0 (x/N), {\mf v}_0 (x/N))$ through relations~(\ref{eq:14}).
Starting from such a state, we expect the state of the system at time $N t$ to be
close, in a suitable sense, to a local Gibbs equilibrium measure
corresponding to an energy-volume profile
\begin{equation*}
X(t,\cdot)=\left(
\begin{array}{c}
{\mf e}(t, \cdot)\\
{\mf v} (t,\cdot)
\end{array}
\right),
\end{equation*}
satisfying a suitable partial differential equation with initial condition~$X_0$
at time $t = 0$. In view of~(\ref{eq:averages_of_currents}), and assuming propagation of local
equilibrium, it is not difficult
to show that the expected partial differential equation
is the following system of two conservation laws:
\begin{equation}
\label{eq:syslim}
\begin{cases}
&\partial_t {\mf e} -\partial_q \tau^2 =0,\\
&\partial_t {\mf v} - 2 \partial_{q} \tau=0,
\end{cases}
\end{equation}
with initial conditions ${\mf e}(0,\cdot) = {\mf e}_0(\cdot), {\mf v}(0,\cdot) = {\mf v}_0(\cdot)$.
We write~(\ref{eq:syslim}) more compactly as
\begin{equation*}
\partial_t X + \partial_q {\mf J}(X) =0, \qquad X (0,\cdot)=X_0 (\cdot),
\end{equation*}
with
\begin{equation}
\label{eq:JJJ}
{\mf J} (X)= \left(
\begin{array}{c}
-\tau^2 ( {\mf e}, {\mf v})\\
-2\tau ({\mf e}, {\mf v})
\end{array}
\right).
\end{equation}
The system of conservation laws (\ref{eq:syslim}) has other
nontrivial conservation laws. In particular, the thermodynamic entropy~$S$
is conserved along a smooth solution of~(\ref{eq:syslim}):
\begin{equation}
\label{eq:consentropy}
\partial_t S ({\mf e},{\mf v})= 0.
\end{equation}
Since the thermodynamic entropy is a strictly concave function
on $\mathring{\mc U}$, the system~(\ref{eq:syslim}) is strictly
hyperbolic on $\mathring{\mc U}$ (see~\cite{Serre}). The two real
eigenvalues of $(D{\mf J})({\bar \xi})$ are $0$
and $-\left[ \partial_{e} (\tau^2) + 2 \partial_v (\tau)\right]$,
corresponding respectively to the two eigenvectors
\begin{equation}
\left(
\begin{array}{c}
-\partial_v \tau\\
\partial_e \tau
\end{array}
\right), \qquad
\left(
\begin{array}{c}
\tau\\
1
\end{array}
\right).
\end{equation}
It is well known that classical solutions to systems of $n \ge 1$ conservation laws in general develop shocks in finite times, even when starting from smooth initial conditions. If we consider weak solutions rather than classical solutions, then a criterion is needed to select a unique, relevant solution among the weak ones. For scalar conservation laws ($n=1$), this criterion is furnished by the so-called entropy inequality and existence and uniqueness of solutions is fully understood. If $n\ge2$, only partial results exist (see~\cite{Serre}). This motivates the fact that we restrict our analysis to smooth solutions before the appearance of shocks.
We assume that the potential $V$ satisfies the following
\begin{assumption}
\label{ass:V}
The potential $V$ is a smooth, non-negative function such that the partition function $Z(\beta, \lambda) = \int_{-\infty}^{\infty} \exp\left( -\beta V(r) -\lambda r \right)\, dr$ is well defined for $\beta>0$ and $\lambda \in \RR$ and there exists a positive constant $C$ such that
\begin{equation}
\label{eq:ass-pot1}
0 < V'' (r) \le C,
\end{equation}
and
\begin{equation}
\label{eq:ass-pot2}
\limsup_{| r| \to + \infty} \frac{r V' (r)}{V(r)} \in(0, +\infty),
\end{equation}
\begin{equation}
\label{eq:ass-pot3}
\limsup_{| r| \to + \infty} \frac{[V' (r)]^2}{V(r)} < +\infty.
\end{equation}
\end{assumption}
\emph{Provided} we can prove that the infinite volume
dynamics is macro-ergodic, then we can rigorously prove (even if $\gamma=0$),
using the relative entropy method of Yau (\cite{yau1}), that~(\ref{eq:syslim}) is
indeed the hydrodynamic limit in the smooth regime, \textit{i.e.}
for times~$t$ up to the appearance of the first shock (see for
example~\cite{KL,TV}). Observe that the expected hydrodynamic limits do not depend on $\gamma$. We need to assume $\gamma>0$ to ensure the macro-ergodicity of the dynamics.
\begin{remark}
As argued in~\cite{TV}, it turns out that the conservation of thermodynamic entropy (\ref{eq:consentropy}) is fundamental for Yau's method where, in the expansion of the time derivative of relative entropy, the cancelation of the linear terms is a consequence of the preservation of the thermodynamic entropy.
\end{remark}
Averages with respect to the empirical energy-volume measure are defined, for continuous functions $G,H : \TT \to \RR$, as (similarly to (\ref{eq:vf-empiricaldensities}))
\[
\left(\begin{array}{c}
{\mc E}_N (t,G) \cr
{\mc V}_N (t,H)
\end{array} \right)
= \left( \begin{array}{c}
\displaystyle \frac1N \sum_{x \in \TT_N} G\left(\frac{x}{N}\right) \, V(\eta_x (t) ) \cr
\displaystyle \frac{1}{N} \sum_{x \in \TT_N} H\left(\frac{x}{N}\right) \, \eta_x (t)
\end{array} \right).
\]
We can then state the following result.
\begin{theorem}[\cite{BS}]
\label{th:4:hl}
Fix some $\gamma > 0$ and consider
the dynamics on the torus ${\bb T}_N$ generated by ${\mc L}_{N,{\rm per}}$ where the potential $V$ satisfies Assumption~\ref{ass:V}. Assume that the system is initially distributed according to a local Gibbs state~(\ref{eq:Ges}) with smooth energy profile ${\mf e}_0$ and volume profile ${\mf v}_0$. Consider a positive time $t$ such that the solution $({\mf e}, {\mf v})$ to (\ref{eq:syslim}) belongs to~${\mathring{\mc U}}$ and is smooth on the time interval~$[0,t]$. Then, for any continuous test functions $G,H:\TT \to \RR$, the following convergence in
probability holds as $N \to +\infty$:
\[
\Big({\mc E}_N (tN, G), {\mc V}_N (tN,H)\Big) \longrightarrow
\left(\int_{\TT} G(q) {\mf e} (t,q) dq, \int_{\TT} H(q) {\mf v} (t,q) dq\right).
\]
\end{theorem}
The derivation of the hydrodynamic limits beyond the shocks for systems of $n \geq 2$ conservation laws is very difficult and is one of the most challenging problems in the field of hydrodynamic limits. The first difficulty is of course our poor understanding of the solutions to such systems. Recently, J. Fritz proposed in~\cite{F0} to derive hydrodynamic limits for hyperbolic systems (in the case $n=2$) by some extension of the compensated-compactness approach~\cite{Tartar} to stochastic microscopic models. This program has been achieved in~\cite{FT} (see also the recent paper \cite{MR2807137}), where the authors derive the classical $n=2$ Leroux system of conservation laws. In fact, to be exact, only the convergence to the set of entropy solutions is proved, the question of uniqueness being left open. It nonetheless remains the best result available at this time. The proof is based on a strict control of entropy pairs at the microscopic level by the use of logarithmic Sobolev inequality estimates. It would be very interesting to extend these methods to systems such as the ones considered here.
\subsubsection{Anomalous diffusion}
We investigate now the problem of anomalous diffusion of energy for these models.
If $V ( r )=r^2$ then Theorem \ref{th:GK-harm-exch} is mutatis mutandis valid and we get the same conclusions: the time-space correlations for the current behave for large time $t$ like $t^{-1/2}$ . Thus the system is super-diffusive (see \cite{BS} for the details).
For generic anharmonic potentials, we can only provide numerical evidence of the super-diffusivity.
However, it is difficult to estimate numerically the time autocorrelation functions of the
currents because of their expected long-time tails, and because statistical errors are
very large (in relative value) when $t$ is large.
Also, for finite systems (the only ones we can simulate on a computer),
the autocorrelation is generically exponentially decreasing for anharmonic potentials,
and, to obtain meaningful results, the thermodynamic limit $N \to \infty$ should be taken before
the long-time limit.
A more tenable approach consists in studying a nonequilibrium system in its steady-state.
We consider a finite system of length $2N+1$ in contact with two
thermostats which fix the value of the energy at the boundaries.
The generator of the dynamics is given by
\begin{equation}
\label{eq:generator_open}
{\mc L}_{N} = {\mc A}_{N} + \gamma {\mc S}_N + \lambda_\ell {\mc B}_{-N,T_\ell} +
\lambda_{r} {\mc B}_{N,T_r},
\end{equation}
where ${\mc A}_N$ and ${\mc S}_N$ are defined by
\begin{equation*}
\begin{split}
&({\mathcal A}_{N}f)(\eta) =\sum_{x =-(N-1)}^{N-1}
\Big(V'(\eta_{x+1})-V'(\eta_{x-1})\Big)(\partial_{\eta_x} f)(\eta) \\
&{\phantom{({\mathcal A}_{N}f)(\eta) =}}-V' (\eta_{N -1}) \, (\partial_{\eta_N} f)(\eta) + V' (\eta_{-N+1}) \, (\partial_{\eta_{-N}} f)(\eta),\\
&({\mc S}_N f)(\eta) =\sum_{x=-N}^{N-1} \left[ f(\eta^{x,x+1}) -f(\eta) \right],
\end{split}
\end{equation*}
and
${\mc B}_{x,T} = T \partial_{\eta_x}^2 -V' (\eta_x) \partial_{\eta_x}$.
The positive parameters $\lambda_\ell$ and $\lambda_r$ are the intensities of the thermostats and $T_\ell, T_r$ the ``temperatures'' of the thermostats.
The generator ${\mc B}_{x,T}$ is a thermostatting mechanism. In order to fix the energy at site $-N$ (resp. $N$) to the value $e_\ell$ (resp. $e_r$), we have to choose $\beta_\ell= T_{\ell}^{-1}$ (resp. $\beta_r = T_r^{-1}$) such that $e(\beta_\ell,0)=e_\ell$ (resp. $e(\beta_r, 0)=e_r$). We denote by $\langle \cdot \rangle_{\rm ss}$ the unique stationary state for the dynamics generated by ${\mc L}_{N}$.
The energy currents $j^{e,\gamma}_{x,x+1}$, which are such that
${\mc L}_{N,{\rm open}} (V (\eta_x)) = -\nabla j_{x-1,x}^{e,\gamma}$ (for $x=-N, \ldots,N-1$),
are given by the first line of (\ref{eq:current_gamma})
for $x = -N+1,\dots,N-1$ while
\begin{equation*}
\begin{split}
&j^{e,\gamma}_{-N-1,-N} = \lambda_\ell \left[ T_\ell V'' (\eta_{-N}) -(V' (\eta_{-N}))^2 \right],\\
&j^{e,\gamma}_{N,N+1} = -\lambda_r \left[ T_r V'' (\eta_{N}) -(V' (\eta_{N}))^2 \right].
\end{split}
\end{equation*}
Since $\langle {\mc L}_{N,{\rm open}} (V (\eta_x)) \rangle_{\rm ss} = 0$,
it follows that, for any $x = -N,\ldots,N+1$,
$\langle j_{x,x+1}^{e,\gamma} \rangle_{\rm ss}$ is equal to a
constant $J_N^\gamma(T_\ell,T_r)$ independent of $x$. In fact,
\begin{equation}
\label{eq:total_current_NESS}
J^{\gamma}_N(T_\ell,T_r) = \left\langle \mathcal{J}_N^\gamma
\right\rangle_{\rm ss},
\qquad
\mathcal{J}_N^\gamma = \frac{1}{2N}\sum_{x=-N-1}^{N} j_{x,x+1}^{e,\gamma}.
\end{equation}
The latter equation is interesting from a numerical viewpoint since it allows to
perform some spatial averaging, hence reducing the statistical error of the results. We estimate the
exponent $\delta \ge 0$ such that
\begin{equation}
\label{eq:def_delta}
\kappa (N) :=N J_N^{\gamma} \sim N^{\delta}
\end{equation}
using numerical simulations. If $\delta=0$, the system is a normal conductor of energy.
If on the other hand $\delta > 0$, it is a superconductor.
The numerical simulations giving the value of $\delta$ are summarized in Table 1.
They have been performed for the harmonic chain $V ( r ) =r^2 /2$, the quartic potential $V ( r ) =r^2/2 + r^4/4$ and the exponential potential $V ( r ) = e^{-r}+r-1$. In Section~\ref{sec:exp.case} we will motivate our interest in the exponential potential.
\begin{table}
\captionsetup{justification=centering}
\label{tabtabexponents}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
$\gamma$ & $V ( r ) =r^2 /2$ & $V ( r ) =r^2/2 + r^4/4$ & $V ( r ) = e^{-r}+r-1$ \\
\hline
$0$ & $1$ & $0.13$ & $1$ \\
$0.01$\phantom{0} & -- & 0.14 & 0.12 \\
$0.1$\phantom{00} & 0.50 & 0.27 & 0.25 \\
$1$\phantom{0.00} & 0.50 & 0.43 & 0.33 \\
\hline
\end{tabular}
\end{center}
\caption{Conductivity exponents}
\end{table}
Exponents in the harmonic case agree with their expected values. For nonlinear potentials, except for the singular value $\delta = 1$ when $\gamma = 0$ and $V ( r ) = e^{-r}+r-1$, the exponents seem to be monotonically increasing with $\gamma$. A similar behavior of the exponents is observed for Toda chains~\cite{ILOS} with a momentum conserving noise. This strange behavior casts some doubts on the convergence of conductivity exponents $\delta$ with respect to system size $N$ (see the comment after Theorem 3 in \cite{BG}). A detailed study, including the nonlinear fluctuating hydrodynamics predictions, is available in \cite{Spohn-Stoltz}.
Note also that the value found for $\gamma = 0$ with the anharmonic FPU potential $V(r) = r^2/2 + r^4/4$ is smaller than the corresponding value for standard oscillators chains, which is around 0.33 (see~\cite{MDN07}). We performed also numerical simulations for a ``rotor'' model, $V ( r ) = 1- \cos ( r )$, and we found $\delta \approx 0.02$, i.e. a normal conductivity. A similar picture is observed for the usual rotor {\footnote{The variable $r$ has to be interpreted as an angle and belongs to the torus $2 \pi \TT$.}} model which is composed of a chain of unpinned oscillators with interaction potential $V ( r )=1-\cos ( r )$. The normal behavior is conjectured to be due to the absence of long waves (that carry energy ballistically) because some rotors turning fast in between will break them (\cite{ILOS-11}). See \cite{Spohn2014-rotators} and references therein for a recent study of the rotors model.
\subsection[Harmonic interactions]{Fractional superdiffusion for a harmonic chain with bulk noise}
In this section we consider the energy-volume conserving model with quadratic potential. Fix $\lambda \in \bb R$ and $\beta>0$, and consider the process $\{\eta (t) ; t \geq 0\}$ generated by (\ref{eq:1:ASdynum2}) with $V(\eta)= \eta^2 /2$ and with initial distribution $\mu_{\beta, \lambda}$. Notice that the distribution of the process $\{\eta (t) +\rho; t \geq 0\}$ with initial measure $\mu_{\beta, \rho+\lambda}$ is the same for all values of $\lambda \in \RR$. Therefore, we can assume, without loss of generality, that $\lambda = 0$. We write $\mu_\beta= \mu_{\beta,0}$ to simplify notation, and denote by $\bb P$ the law of $\{\eta (t) ; t \geq 0\}$ and by $\bb E$ the expectation with respect to $\bb P$. The {\em energy correlation function} $\{S_t (x); x \in \bb Z, t \geq 0\}$ is defined as
\begin{equation}
S_t(x) = \tfrac{\beta^2}{2} \; \bb E\big[\big(\eta_0(0)^2-\tfrac{1}{\beta} \big) \big( \eta_x (t)^2 -\tfrac{1}{\beta} \big)\big]
\end{equation}
for any $x \in \bb Z$ and any $t \geq 0$. The constant $\frac{\beta^2}{2}$ is just the inverse of the variance of $\eta_x^2-\frac{1}{\beta}$ under $\mu_\beta$. By translation invariance of the dynamics and the initial distribution $\mu_\beta$, we see that
\begin{equation}
\tfrac{\beta^2}{2}\, \bb E\big[ \big(\eta_x(0)^2 -\tfrac{1}{\beta} \big) \big(\eta_y (t)^2 -\tfrac{1}{\beta}\big)\big]=S_t(y-x)
\end{equation}
for any $x, y \in \bb Z$.
\begin{theorem} [\cite{BGJ}]
\label{t1}
Let $f,g: \bb R \to \bb R$ be smooth functions of compact support. Then,
\begin{equation}
\label{ec1.12}
\lim_{n \to \infty} \tfrac{1}{n}\sum_{x, y \in \bb Z} f\big( \tfrac{x}{n} \big) g\big( \tfrac{y}{n}\big) S_{tn^{3/2}}(x-y) = \iint f(x)g(y) P_t(x-y) dx dy,
\end{equation}
where $\{P_t(x); x \in \bb R, t \geq 0\}$ is the fundamental solution {\footnote{Since the skew fractional heat equation is linear, it can be solved explicitly by Fourier transform.}} of the skew fractional heat equation on $\RR$
\begin{equation}
\label{ec1.13}
\partial_t u = -\tfrac{1}{\sqrt{2}}\big\{ (-\Delta)^{3/4} - \nabla (-\Delta)^{1/4}\big\} u.
\end{equation}
\end{theorem}
A fundamental step in the proof of this theorem is the analysis of the correlation function $\{{\bb S}_t(x,y); x \neq y \in \bb Z, t \geq 0\}$ given by
\begin{equation}
{\bb S}_t(x,y) = \tfrac{\beta^2}{2} \bb E \big[ \big(\eta_0 (0)^2-\tfrac{1}{\beta} \big) \eta_x(t) \eta_y(t)\big]
\end{equation}
for any $t \geq 0$ and any $x \neq y \in \bb Z$. Notice that this definition makes perfect sense for $x =y$ and, in fact, we have ${\bb S}_t(x,x) = S_t(x)$. For notational convenience we define ${\bb S}_t(x,x)$ as equal to $S_t(x)$. However, these quantities are of different nature, since $S_t(x)$ is related to {\em energy fluctuations} and ${\bb S}_t(x,y)$ is related to {\em volume fluctuations} (for $x \neq y$).
\begin{remark}
It is not difficult to see that with a bit of technical work the techniques actually show that the distribution valued process $\{ {\mc E}_t^n (\cdot) \, ; \, t\ge 0\}$ defined for any test function $f$ by
$${\mc E}_{t}^n (f) = \cfrac{1}{\sqrt n} \sum_{x \in \ZZ} f \big (\tfrac{x}{n} \big) \{ \eta_x (tn^{3/2})^2 - \tfrac{1}{\beta}\}$$
converges as $n$ goes to infinity to an infinite dimensional $3/4$-fractional Ornstein-Uhlenbeck process, i.e. the centered Gaussian process with covariance prescribed by the right hand side of (\ref{ec1.12}).
\end{remark}
\begin{remark}
It is interesting to notice that $P_t$ is the maximally asymmetric $3/2$-Levy distribution.
It has power law as $|x|^{-5/2}$ towards the diffusive peak and stretched exponential as $\exp[-|x|^3]$ towards the exterior of the sound cone (\cite[Chapter 4]{UZ}). As mentioned to us by H. Spohn, this reflects the expected physical property that no propagation beyond the sound cone occurs.
\end{remark}
\begin{remark}
With a bit of technical work the proof of this theorem can be adapted to obtain a similar statement for a harmonic chain perturbed by the momentum exchanging noise (see \cite{JKO2} where such statement is proved for the Wigner function). In this case the skew fractional heat equation is replaced by the (symmetric) fractional heat equation.
\end{remark}
\begin{proof}
Denote by $\mc C_c^\infty(\bb R^d)$ the space of infinitely differentiable functions $f: \bb R^d \to \bb R$ of compact support. Then, $\| f \|_{2,n}$ denotes the weighted $\ell^2(\bb Z^d)$-norm
\begin{equation}
\|f\|_{2,n} = \sqrt{ \vphantom{H^H_H}\smash{\tfrac{1}{n^d} \sum_{x \in \bb Z^d} f\big( \tfrac{x}{n} \big)^2}}.
\end{equation}
Let $g \in \mc C_c^\infty(\bb R)$ be a fixed function.
For each $n \in \bb N$ we define the field $\{\mc S_t^n; t \geq 0\}$ as
\begin{equation}
\mc S_t^n(f) = \tfrac{1}{n}\!\! \sum_{x, y \in \bb Z} g\big(\tfrac{\vphantom{y}x}{n} \big) f\big( \tfrac{y}{n}\big) S_{tn^{3/2}}(y-x)
\end{equation}
for any $t \geq 0$ and any $f \in \mc C_c^\infty(\bb R)$.
\noindent
By the Cauchy-Schwarz inequality we have the {\em a priori} bound
\begin{equation}
\label{ec3.4}
\big| \mc S_t^n(f) \big| \leq \|g\|_{2,n} \| f\|_{2,n}
\end{equation}
for any $t \geq 0$, any $n \in \bb N$ and any $f,g \in \mc C_c^\infty(\bb R)$. For a function $h \in \mc C_c^\infty(\bb R^2)$ we define $\{ Q_t^n(h); t \geq 0\}$ as
\begin{equation}
Q_t^n(h) = \tfrac{1}{n^{3/2}} \sum_{x \in \ZZ} \; \sum_{y\ne z \in \bb Z} g\big(\tfrac{\vphantom{y}x}{n} \big) h \big(\tfrac{y}{n}, \tfrac{\vphantom{y} z}{n}\big) {\bb S}_{t n^{3/2}}(y-x,z-x).
\end{equation}
Notice that $Q_t^n(h)$ depends only on the symmetric part of the function $h$. Therefore, we will always assume, without loss of generality, that $h(x,y) = h(y,x)$ for any $x,y \in \bb Z$. We point out that $Q_t^n(h)$ does not depend on the values of $h$ at the diagonal $\{x=y\}$. We have the {\em a priori} bound
\begin{equation}
\label{ec3.6}
\big| Q_t^n(h) \big| \leq 2 \|g\|_{2,n} \|{\tilde h} \|_{2,n},
\end{equation}
where $\tilde h$ is defined by ${\tilde h} \big(\tfrac{\vphantom{y}x}{n} , \tfrac{y}{n}\big) = h \big(\tfrac{\vphantom{y}x}{n} , \tfrac{y}{n}\big) \, {\bf 1}_{x \ne y}.$
For a function $f \in \mc C_c^\infty(\bb R)$, we define $\Delta_n f: \bb R \to \bb R$ as
\begin{equation}
\Delta_n f \big( \tfrac{x}{n}\big) = n^2\Big( f\big(\tfrac{x\!+\! 1}{n} \big) + f\big( \tfrac{x\!-\! 1}{n} \big) - 2 f\big(\tfrac{x}{n}\big) \Big).
\end{equation}
In other words, $\Delta_n f$ is a discrete approximation of the second derivative of $f$. We also define $\nabla_n f \otimes \delta : \sfrac{1}{n} \bb Z^2 \to \bb R$ as
\begin{equation}
\label{eq:3.9}
\big(\nabla_n f \otimes \delta\big) \big( \tfrac{x}{n}, \tfrac{y}{n}\big) =
\begin{cases}
\frac{n^2}{2}\big(f\big(\tfrac{x+1}{n}\big) - f\big(\tfrac{x}{n}\big)\big); & y =x\!+\! 1\\
\frac{n^2}{2}\big(f\big(\tfrac{x}{n}\big) - f\big(\tfrac{x-1}{n}\big)\big); & y =x\!-\! 1\\
0; & \text{ otherwise.}
\end{cases}
\end{equation}
Less evident than the interpretation of $\Delta_n f$, $\nabla_n f \otimes \delta$ turns out to be a discrete approximation of the (two dimensional) distribution $f'\!(x) \otimes \delta(x=y)$, where $\delta(x=y)$ is the $\delta$ of Dirac at the line $x=y$. We have that
\begin{equation}
\label{ec3.10}
\tfrac{d}{dt} \mc S_t^n(f) = -2Q_t^n(\nabla_n f \otimes \delta) + \mc S_t^n(\tfrac{1}{\sqrt n}\Delta_n f).
\end{equation}
In this equation we interpret the term $Q_t^n(\nabla_n f \otimes \delta)$ in the obvious way.
By the {\em a priori} bound \eqref{ec3.4}, the term $ \mc S_t^n(\frac{1}{\sqrt n} \Delta_n f)$ is negligible, as $n \to \infty$.We describe now the equation satisfied by $Q_t^n(h)$. For this we need some extra definitions. For $h \in \mc C_c^\infty(\bb R^2)$ we define $\Delta_n h : \bb R^2 \to \bb R$ as
\begin{equation}
\Delta_n h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y}{n}\big) = n^2\Big( h\big( \tfrac{\vphantom{y}x+1}{n}, \tfrac{y}{n}\big)+h\big( \tfrac{\vphantom{y}x-1}{n}, \tfrac{y}{n}\big)+h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y+1}{n}\big)+ h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y-1}{n}\big) - 4 h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y}{n}\big)\Big).
\end{equation}
In words, $\Delta_n h$ is a discrete approximation of the $2d$ Laplacian of $h$. We also define $\mc A_n h: \bb R \to \bb R$ as
\begin{equation}
\mc A_n h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y}{n}\big) = n\Big(h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y-1}{n}\big)+ h\big( \tfrac{\vphantom{y}x-1}{n}, \tfrac{y}{n}\big)- h\big( \tfrac{\vphantom{y}x}{n}, \tfrac{y+1}{n}\big)-h\big( \tfrac{\vphantom{y}x+1}{n}, \tfrac{y}{n}\big)\Big).
\end{equation}
The function $\mc A_n h$ is a discrete approximation of the directional derivative $(-2,-2) \cdot \nabla h$. Let us define $\mc D_n h : \sfrac{1}{n} \bb Z \to \bb R$ as
\begin{equation}
\mc D_n h\big( \tfrac{x}{n} \big) = n \Big( h \big(\tfrac{x}{n}, \tfrac{x+1}{n}\big) - h \big( \tfrac{x-1}{n}, \tfrac{x}{n} \big) \Big)
\end{equation}
and $\widetilde {\mc D}_n h :\sfrac{1}{n} \bb Z^2 \to \bb R$ as
\begin{equation}
\widetilde{\mc D}_n h (\tfrac{x}{n},\tfrac{y}{n}) =
\begin{cases}
n^2 \big(h\big(\tfrac{x}{n}, \tfrac{x+1}{n}\big)-h\big(\tfrac{x}{n}, \tfrac{x}{n}\big)\big); & y =x+1\\
n^2 \big(h\big(\tfrac{x-1}{n}, \tfrac{x}{n}\big)-h\big(\tfrac{x-1}{n}, \tfrac{x-1}{n}\big)\big); & y=x-1\\
0; & \text{ otherwise.}
\end{cases}
\end{equation}
The function $\mc D_n h$ is a discrete approximation of the directional derivative of $h$ along the diagonal $x=y$, while $\widetilde{\mc D}_n h$ is a discrete approximation of the distribution $\partial_y h(x,x) \otimes \delta(x=y)$.
Finally we can write down the equation satisfied by the field $Q_t^n(h)$:
\begin{equation}
\label{ec3.15}
\tfrac{d}{dt} Q_t^n(h) = Q_t^n\big(n^{-1/2}\Delta_n h+ n^{1/2}\mc A_n h\big) -2 \mc S_t^n\big( \mc D_n h\big) + 2 Q_t^n\big(n^{-1/2}\widetilde{\mc D}_n h\big).
\end{equation}
Given $f \in \mc C_c^\infty(\bb R)$, if we choose $h:=h_n (f)$ such that
$$n^{-1/2}\Delta_n h+ n^{1/2}\mc A_n h = 2 \nabla_n f \otimes \delta$$
then summing \eqref{ec3.10} and \eqref{ec3.15} we get
\begin{equation*}
\tfrac{d}{dt} \mc S_t^n(f) = - \tfrac{d}{dt} Q_t^n(h) + \mc S_t^n(\tfrac{1}{\sqrt n}\Delta_n f) -2 \mc S_t^n\big( \mc D_n h\big) + 2 Q_t^n\big(n^{-1/2}\widetilde{\mc D}_n h\big).
\end{equation*}
We integrate in time the previous expression. By the a priori bounds, the term $\int_0^t \mc S_s^n(\tfrac{1}{\sqrt n}\Delta_n f) ds $ is small as well as $\int_0^t \tfrac{d}{ds} Q_s^n (h) ds =Q_t^n(h) - Q_0^n (h)$. The term $$\int_0^t Q_s^n( n^{-1/2} \widetilde{\mc D}_n h)\, ds$$ is quite singular since it involves an approximation of a distribution but it turns out to be negligible, although this does not follow directly from the a priori bounds (see \cite{BGJ}). By using Fourier transform one can see that $-2 \mc D_n h$ converges to $ -\tfrac{1}{\sqrt{2}}\big\{ (-\Delta)^{3/4} - \nabla (-\Delta)^{1/4}\big\} f$ and we are done.
\cqfd\end{proof}
\subsection[Exponential interactions]{Anomalous diffusion for a perturbed Hamiltonian system with exponential interactions}
\label{sec:exp.case}
We investigate here in more details the exponential case $V_{\rm{exp}} ( r ) = e^{-r} -1 +r$. The deterministic system with generator (\ref{eq:A_per}) and with the exponential potential above is well known in the integrable systems literature {\footnote{It seems that although the Hamiltonian structure of the Kac-van-Moerbecke system was known, the interpretation of the latter as a chain of oscillators with exponential kinetic energy and exponential interaction was not observed before \cite{BS}.}}. It has been introduced in \cite{KVM} by Kac and van Moerbecke and was shown to be completely integrable. Consequently, using Mazur's inequality, it is easy to show that the energy transport is ballistic (\cite{BS}).
As we will see the situation dramatically changes when the momentum exchange noise is added: the energy transport is no more ballistic but superdiffusive. Thus the situation is similar to the harmonic case. Nevertheless we expect the time autocorrelation of the current to decay like $t^{-2/3}$. We are not able to show this but we proved in \cite{BG} lower bounds sufficient to imply superdiffusivity.
The results are stated in infinite volume: we consider the stochastic energy-volume conserving dynamics $\{\eta (t)\}_{t \ge 0}$ with potential $V:=V_{\rm{exp}}$. Its generator is given by ${\mc L}={\mc A} + \gamma {\mc S}$ where ${\mc A}$ and ${\mc S}$ are defined by (\ref{eq:1:ASdynum2}). Since the exponential potential grows very fast as $r\to -\infty$, some care has to be taken to show that the infinite dynamics is well defined (see \cite{BG}). We recall that grand canonical Gibbs measures are denoted by $\mu_{\beta,\lambda}$ and take the form
\begin{equation*}
d\mu_{\beta,\lambda} (\eta) = \prod_{x \in \ZZ} \frac{{\rm e}^{- \beta V(\eta_x) -\lambda \eta_x} }{Z(\beta,\lambda)} d\eta_x, \quad \beta>0, \; \lambda + \beta <0.
\end{equation*}
In this section, $\beta$ and $\lambda$ are fixed and we denote by $e$ (resp. $v$) the average energy (resp. volume) w.r.t. $\mu_{\beta,\lambda}$ (see (\ref{eq:tr})).
The microscopic energy current ${j}^{e,\gamma}_{x,x+1}$ and volume current ${j}^{v,\gamma}_{x,x+1}$ are given by
\begin{equation*}
{j}^{e,\gamma}_{x,x+1}(\eta)=-e^{-(\eta_x + \eta_{x+1})}+(e^{- \eta_x} +e^{- \eta_{x+1}})- \gamma \nabla (V (\eta_{x}))
\end{equation*}
and
\begin{equation*}
{j}^{v,\gamma}_{x,x+1}(\eta)=e^{-\eta_x} +e^{-\eta_{x+1}} -\gamma \nabla \eta_x .
\end{equation*}
We will use the compact notations
\begin{equation*}
\omega_x= \left( \begin{array}{c} V(\eta_x) \\ \eta_x \end{array} \right), \quad J_{x,x+1}= \left( \begin{array}{c} j^{e,\gamma}_{x,x+1} \\ j^{v,\gamma}_{x,x+1} \end{array} \right).
\end{equation*}
In the hyperbolic scaling, the hydrodynamical equations for the energy profile ${\mf e}$ and the volume profile ${\mf v}$ take the form
\begin{equation}
\label{eq:hl1euler}
\begin{cases}
\partial_t {\mf e} -\, \partial_q ( ({\mf e} - {\mf v})^2) =0\\
\partial_{t} {\mf v} + 2 \, \partial_q ({\mf e -{\mf v}}) =0.
\end{cases}
\end{equation}
They can be written in the compact form $\partial_t {{X}} + \partial_q {{ \mf J}} ({{X}}) =0$ with
\begin{equation}\label{eq:hl-ss00}
{X}=
\left(
\begin{array}{c}
{\mf e}\\
{\mf v}
\end{array}
\right)
, \quad
{ \mf J }({ {X}})=
\left(
\begin{array}{c}
- ( {\mf e} - {\mf v})^2\\
2 ( {\mf e} - {\mf v})
\end{array}
\right).
\end{equation}
The differential matrix of ${ \mf J}$ is given by
\begin{equation*}
(\nabla{ \mf J})({X})=
2 \left(
\begin{array}{cc}
- ( {\mf e} - {\mf v})& {\mf e} - {\mf v} \\
1 & -1
\end{array}
\right).
\end{equation*}
For given $({e}, {v})$ we denote by $({ T}^{+}_t )_{t \ge 0}$ (resp. $({T}^{-}_t)_{t \ge 0}$) the semigroup on $S(\RR) \times S(\RR)$ generated by the linearized system
\begin{equation}\label{linearized system 1}
\partial_t \varepsilon + {M}^T\, \partial_q \varepsilon =0, \quad ({\text{resp.}} \; \partial_t \varepsilon - {M}^T \, \partial_q \varepsilon =0 ),
\end{equation}
where
\begin{equation*}
{M} := {M} ({e}, {v})= [\nabla { \mf J}] (\omega), \quad \omega= \left(
\begin{array}{c}
{ e}\\
{v}
\end{array}
\right).
\end{equation*}
We omit the dependence of these semigroups on $(e,v)$ for lightness of the notations. Above $S(\RR)$ denotes the Schwartz space of smooth rapidly decreasing functions.
The first result of \cite{BG} gives a lower bound on the time-scale for which a non-trivial evolution of the energy-volume fluctuation field can be observed.
We take the infinite system at equilibrium under the Gibbs measure $\mu_{\bar \beta,\bar \lambda}$ corresponding to a mean energy $\bar e$ and a mean volume $\bar v$. Our goal is to study
the energy-volume fluctuation field in the time-scale $tn^{1+\alpha}$, $\alpha \ge0$:
\begin{equation}
\label{eq:YY}
\mathcal{Y}^{n,\alpha}_t (\bG) =\frac{1}{\sqrt{n}} \sum_{x\in \ZZ}
\bG\left(x/n\right) \cdot \left({\omega}_x (tn^{1+\alpha}) - {\bar{\omega}}\right),
\end{equation}
where for $q\in{\mathbb{R}}$, $x \in \ZZ$,
\[
\bG(q) = \left(
\begin{array}{c}
G_{1} (q)\\
G_2 (q)
\end{array}
\right), \quad
{\omega}_x= \left(
\begin{array}{c}
V (\eta_x) \\
\eta_x
\end{array}
\right)
\]
and $G_1, G_2$ are test functions belonging to $S(\RR)$.
We need to introduce some notation. For each $z \ge 0$, let $H_z (x) = (-1)^z e^{x^2} \tfrac{d^z}{dx^z} e^{-x^2}$ be the Hermite polynomial of order $z$ and $h_z (x) =(z! {\sqrt{2\pi}})^{-1} H_{z} (x) e^{-x^2}$ the Hermite function. The set $\{ h_z, z\ge 0\}$ is an orthonormal basis of ${\bb L}^2 (\RR)$. Consider in ${\bb L}^2(\RR)$ the operator $K_0 = x^2-\Delta$, $\Delta$ being the Laplacian on $\RR$. For an integer $k \ge 0$, denote by ${\bb H}_k$ the Hilbert space obtained by taking the completion of $S (\RR)$ under the norm induced by the scalar product $\langle \cdot,\cdot\rangle_{k}$ defined by $\langle f, g \rangle_k= \langle f, K_0^k g \rangle_0$, where $\langle \cdot,\cdot\rangle_0$ denotes the inner product of $\LL^2 (\RR)$ and denote by ${\bb H}_{-k}$ the dual of ${\bb H}_k$, relatively to this inner product. Let $\langle\cdot\rangle$ represent the average with respect to the Lebesgue measure.
If $E$ is a Polish space then $D(\RR^+, E)$ (resp, $C(\RR^+, E)$) denotes the space of $E$-valued functions, right continuous with left limits (resp. continuous), endowed with the Skorohod (resp. uniform) topology. Let $Q^{n,\alpha}$ be the probability measure on ${D}(\RR^+,{\bb H}_{-k} \times {\bb H}_{-k})$ induced by the fluctuation field ${\mc Y}^{n,\alpha}_t$ and $\mu_{\beta,\lambda}$. Let $\mathbb{P}_{\mu_{ \beta,\lambda}}$ denote the probability measure on ${ D}(\RR^+, \RR^{\ZZ})$ induced by $(\eta(t))_{t\geq{0}}$ and $\mu_{\beta, \lambda}$. Let $\mathbb{E}_{\mu_{\beta,\lambda}}$ denote the expectation with respect to $\mathbb{P}_{\mu_{\beta,\lambda}}$.
\begin{theorem}[\cite{BG}]
\label{th:fluct-hs}
Fix an integer $k>2$. Denote by $Q$ the probability measure on $C(\RR^+, {\bb H}_{-k} \times {\bb H}_{-k})$ corresponding to a stationary Gaussian process with mean $0$ and covariance given by
\begin{equation*}
{\mathbb E}_{Q} \left[ \mathcal{Y}_t (\mb H) \, \mathcal{ Y}_s (\mb G) \right] = \langle \,{T}_t^{-} \mb H\; \cdot \; \chi \; {T}_s^{-}\mb G \, \rangle
\end{equation*}
for every $0 \le s \le t$ and $\mb H, \mb G$ in ${\bb H}_k \times {\bb H}_k$. Here ${\chi}:={\chi} ({ \beta}, {\lambda})$ is the equilibrium covariance matrix of ${ \omega}_0$. Then, the sequence $(Q^{n,0})_{n \ge 1}$ converges weakly to the probability measure $Q$.
\end{theorem}
The theorem above means that in the hyperbolic scaling the fluctuations are trivial: the initial fluctuations are transported by the linearized system of (\ref{eq:hl1euler}). To see a nontrivial behavior we have to study, in the transport frame, the fluctuations at a longer time scale $t n^{1+\alpha}$, with $\alpha>0$. Thus, we consider the fluctuation field ${\widehat {\mc Y}}_{\cdot}^{n,\alpha}$, $\alpha>0$, defined, for any $\bG \in S(\RR) \times S(\RR)$, by
\begin{equation}\label{longer density field}
{\widehat {\mc Y}}_t^{n,\alpha} (\bG)= {\mc Y}_t^{n, \alpha} \left( {T}^+_{t n^{\alpha}} \bG \right).
\end{equation}
Our second main theorem shows that the correct scaling exponent $\alpha$ is greater than $1/3$:
\begin{theorem}[\cite{BG}]
\label{th:fluct-ds}
Fix an integer $k>1$ and $\alpha<1/3$. Denote by $Q$ the probability measure on $C(\RR^+, {\bb H}_{-k} \times {\bb H}_{-k})$ corresponding to a stationary Gaussian process with mean $0$ and covariance given by
\begin{equation*}
{\mathbb E}_{Q} \left[ \mathcal{Y}_t (\mb H) \, \mathcal{ Y}_s (\mb G) \right] = \langle \, \mb H \;\cdot \; {\chi} \; \mb G \rangle
\end{equation*}
for every $0 \le s \le t$ and $\mb H, \mb G$ in ${\bb H}_k \times {\bb H}_k$. Then, the sequence $(Q^{n,\alpha})_{n \ge 1}$ converges weakly to the probability measure $Q$.
\end{theorem}
The proofs of these theorems can be reduced to the proof of a so-called {\textit{equilibrium Boltzmann-Gibbs principle}}. Let us explain what it means. Observables can be divided into two classes: non-hydrodynamical and hydrodynamical. The first ones are the non conserved quantities and they fluctuate on a much faster scale than the conserved ones. Hence, they should average out and only their projection on the hydrodynamical variables should persist in the scaling limit.
For any local function $g:= g(\eta)$, the projection ${\mc P}_{e, {v}} \, g$ of $g$ on the fields of the conserved quantities is defined by
\begin{equation*}
({\mc P}_{{ e}, { v}} g)(\eta)=g (\eta) -{\tilde g} ( e, v) -(\nabla {\tilde g})(e, v) \cdot ({\omega_0}-{\omega})
\end{equation*}
where ${\tilde g} (e, v) = \langle g \rangle_{\mu_{\beta,\lambda}}$ and $\nabla {\tilde g}$ is the gradient of the function ${\tilde g}$.
As explained above we expect that in the Euler time scale, for any test function $\bH \in S(\RR) \times S(\RR)$, the space-time variance
\begin{equation}
\label{eq:4:BGvar}
\lim_{n \to \infty} {\mathbb E}_{\mu_{\beta, \lambda}} \left[ \left( \int_0^{t} \, \cfrac{1}{\sqrt{n}} \sum_{x \in \ZZ} \mb H\left( x/n \right)\cdot \left[\theta_x {\mc P}_{e, v}\, g \, (\eta (sn) )\right] \, ds\right)^2\right] =0
\end{equation}
vanishes as $n$ goes to infinity. In fact it suffices to show (\ref{eq:4:BGvar}) for the function $g=J_{0,1}${ \footnote{For Theorem \ref{th:fluct-ds}, the Boltzmann-Gibbs principle has to be proved in the longer time scale $t n^{1+ \alpha}$ and in the transport frame.}}. Thus let us first define the {\textit{normalized}} currents by
\begin{equation}
\label{norm. currents}
\begin{split}
{\hat J}_{x,x+1}& = \left[\theta_x {\mc P}_{\bar e, \bar v}\, J_{0,1} \right]= \left(
\begin{array}{c}
{j}^{e,\gamma}_{x,x+1}(\eta)\\
{j}^{v,\gamma}_{x,x+1}(\eta)
\end{array}
\right)
- {{\mf J}} ({ \omega}) - (\nabla {{\mf J}}) ({\omega})
\left(
\begin{array}{c}
V_{\rm{exp}} (\eta_x) -{e}\\
\eta_x -{v}
\end{array}
\right).
\end{split}
\end{equation}
To estimate the space-time variance involved we use the following inequality (see \cite{KLO-book}):
\begin{equation}
\label{eq:4:inequS}
\begin{split}
{\mathbb E}_{\mu_{\beta,\lambda}} \left[ \left( \int_0^t f (\eta (s n^{1+\alpha})) \, ds \right)^2\right] \, &\le \, \cfrac{C t} {n^{1+\alpha}} \left\langle f\;, \left( \cfrac{1}{t n^{1+\alpha}} - \gamma {\mc S} \right)^{-1} f \right\rangle_{\mu_{\beta,\lambda}}
\end{split}
\end{equation}
where $f= \cfrac{1}{\sqrt{n}} \sum_{x \in \ZZ} \mb H\left( x/n \right)\cdot {\hat J}_{x,x+1}$. Due to the very simple form of the operator ${\mc S}$ the RHS of (\ref{eq:4:inequS}) can be estimated and shown to vanish as $n$ goes to infinity. Nevertheless it has to be done with some care since ${\mc S}$ is very degenerate so that without the term $\frac{1}{t n^{1+\alpha}}$ the RHS of (\ref{eq:4:inequS}) blows up.
Theorem \ref{th:fluct-ds} does not exclude the possibility of normal fluctuations, i.e. the convergence in law of the fluctuation field of the two conserved quantities to an infinite dimensional Ornstein-Uhlenbeck process in the diffusive time scale ($\alpha=1$). To see that it is not the case we will show that the diffusion coefficient ${\mc D}:={\mc D} ({e}, { v})$ appearing in this hypothetical limiting process would be infinite, excluding thus this possibility. Up to a constant matrix coming from a martingale term (due to the noise) and thus irrelevant for us (see \cite{BBO2}, \cite{BS}), the matrix coefficient ${\mc D}$ is defined by the Green-Kubo formula
\begin{equation}
\label{eq:GK00}
{\mc D} = \int_{0}^{\infty} {\mathbb E}_{\mu_{\beta, \lambda}} \left[ \sum_{x \in \ZZ} {\hat J}_{x,x+1} (t) \left[ {\hat J}_{0,1} (0) \right]^T \right] \, dt.
\end{equation}
The signature of the superdiffusive behavior of the system is seen in the divergence of ${\mc D}$, i.e. in a slow decay of the current-current correlation function. To study the latter we introduce its Laplace transform
\begin{equation*}
\begin{split}
{\mc F} (\gamma, z) = \int_{0}^{\infty} e^{-z t} \, {\mathbb E}_{\mu_{\beta, \lambda}} \left[ \sum_{x \in \ZZ} {\hat J}_{x,x+1} (t) \left[ {\hat J}_{0,1} (0)\right]^T \right] \, dt
\end{split}
\end{equation*}
which is well defined for any $z>0$. This can be rewritten as
\begin{equation*}
\begin{split}
{\mc F} (\gamma, z) = \ll {\hat J}_{0,1}, (z-{\mc L})^{-1} {\hat J}_{0,1} \gg_{\beta,\lambda}
\end{split}
\end{equation*}
where $\ll \cdot, \cdot \gg_{\beta,\lambda}$ is the semi-inner product defined with respect to $\mu_{\beta,\lambda}$ in the same way as in (\ref{eq:scalarproduct>>}).
Our third theorem is the following lower bound on ${\mc F} (\gamma, z)$. Observe that ${\mc F} (\gamma, z)$ is a square matrix of size $2$ whose entry $(i,j)$ is denoted by ${\mc F}_{i,j}$.
\begin{theorem}[\cite{BG}]
\label{th:diffusivity}
Fix $\gamma>0$. There exists a positive constant $c:=c(\gamma)>0$ such that
\begin{equation*}
{\mc F}_{1,1} (\gamma, z) \ge c z^{-1/4}
\end{equation*}
and
\begin{equation*}
{\mc F}_{i,j} (\gamma, z)=0, \quad (i,j) \ne (1,1).
\end{equation*}
Moreover, there exists a positive constant $C:=C(\gamma)$ such that for any $z>0$,
\begin{equation}
\label{eq:F11c}
C^{-1} {\mc F}_{1,1} (1,z/\gamma) \le {\mc F}_{1,1} (\gamma, z) \le C {\mc F}_{1,1} (1,z/\gamma).
\end{equation}
\end{theorem}
The last part of the theorem follows easily by a scaling argument and is in fact also valid for general potentials $V$ and for generic ``standard" anharmonic chains of oscillators. In \cite{BDLLOP, ILOS, BS}, numerical simulations indicate a strange dependence w.r.t. the noise intensity $\gamma>0$ of the exponent $\delta$ in the energy transport coefficient $\kappa (N) \sim N^{\delta}$ ($N$ is the system size, see (\ref{eq:def_delta}) for the definition of $\kappa (N)$): $\delta:=\delta(\gamma)>0$ is increasing with the noise intensity $\gamma$. This is very surprising since the more stochasticity in the model is introduced, the more the system is superdiffusive! The inequality (\ref{eq:F11c}) shows that the time decay of the current autocorrelation function is independent of $\gamma$ (up to possible slowly varying functions corrections, i.e. in a Tauberian sense). It is common folklore that there should be a simple relationship between the slow long-time tail decay of the autocorrelation of the current in the Green-Kubo formula (described by some power law decay) and the divergence of the thermal conductivity of open systems in their steady states. The argument is that the autocorrelation should be integrated over times of order~$N$. If we believe in this argument it means that the numerical simulations of \cite{BDLLOP, ILOS, BS} are not converged. There is however no clear mathematical result backing up this belief.\\
The proof of the first part of Theorem \ref{th:diffusivity} is based on the three following arguments.
\begin{itemize}
\item The first idea consists in performing the microscopic change of variables $\xi_x = e^{-\eta_x}$, $x \in \ZZ$, that defines a new Markov process $\{\xi (t)\}_{t \ge 0}= \{ \xi_x (t) \, ; \, x \in \ZZ\}_{t \ge 0}$ with state space $(0,+\infty)^{\ZZ}$ and conserving $\sum_x \xi_x$ and $\sum_{x} \log \xi_x$. Its generator is given by ${\tilde {\mc L}} ={\tilde {\mc A}} +\gamma {\tilde{ \mc S}}$ where for any local differentiable function $f$,
\begin{equation*}
\begin{split}
&({\tilde {\mc A}} f)(\xi) = \sum_{x \in \ZZ} \xi_x (\xi_{x+1} -\xi_{x-1} ) (\partial_{\xi_x} f )(\xi),\\
&({\tilde {\mc S}} f)(\xi) = \sum_{x \in \ZZ} \left[ f(\xi^{x,x+1}) -f(\xi)\right].
\end{split}
\end{equation*}
The invariant measures for $(\xi (t))_{t \ge 0} $ are obtained from the Gibbs measures $\mu_{\beta, \lambda}$ by the change of variables above. They form a family $\{ \nu_{\rho, \theta} \}_{\rho, \theta}$ of translation invariant product measures indexed by two parameters $\rho$ and $\theta$ which satisfy
$$\rho =\nu_{\rho, \theta} (\xi_x), \quad \theta= \nu_{\rho, \theta} (\log \xi_x). $$
In fact the marginal of $\nu_{\rho, \theta}$ is a Gamma distribution. The parameters $(\rho, \theta)$ are in a one-to-one explicit correspondence with the parameters $(e,v)$.
Rewriting ${\hat J}_{x,x+1}$ with these new variables we see that it is sufficient to prove a similar statement for the process $(\xi (t) )_{t \ge 0}$ under the equilibrium probability measure $\nu_{\rho,\theta}$. Introducing the inner product $\ll \cdot, \cdot \gg$ defined, for any local functions $f,g$ on $(0,+\infty)^{\ZZ}$ by
\begin{equation*}
\ll f, g \gg =\sum_{x \in \ZZ} \left\{ \nu_{\rho, \theta} (f \, \theta_x g) - \nu_{\rho,\theta} (f) \nu_{\rho, \theta} (g) \right\}
\end{equation*}
we can show that the proof of the first claim of Theorem \ref{th:diffusivity} reduces to showing that there exists a positive constant $c$ such that for any $z>0$,
\begin{equation}
\label{eq:4:w01}
\ll W_{0,1} , (z -{\tilde {\mc L}})^{-1} W_{0,1} \gg \ge c z^{-1/4}
\end{equation}
where $W_{0,1} (\xi) =(\xi_0 -\rho) (\xi_1 -\rho)$.
\item The second step consists in using a variational formula to express the LHS of (\ref{eq:4:w01}). Indeed we have
\begin{equation*}
\begin{split}
\ll W_{0,1} , (z -{\tilde {\mc L}})^{-1} W_{0,1} \gg &=\sup_{g} \left\{ 2 \ll W_{0,1}, g \gg - \ll g, (z -\gamma {\tilde{\mc S}}) g \gg \right.\\
& \quad \quad \left. - \ll {\tilde{\mc A}} g, (z-\gamma {\tilde{\mc S}})^{-1} {\tilde{\mc A}}g \gg \right\}
\end{split}
\end{equation*}
where the supremum is taken over local compactly supported smooth functions~$g$. To get a lower bound it is sufficient to find a function $g$ for which one can show that
$$2 \ll W_{0,1}, g \gg - \ll g, (z -\gamma {\tilde{\mc S}}) g \gg - \ll {\tilde{\mc A}} g, (z-\gamma {\tilde{\mc S}})^{-1} {\tilde{\mc A}}g \gg \; \ge \; c z^{-1/4}.$$
\item Let ${\bb H}$ be the Hilbert space obtained by completion of the set of local functions w.r.t. the inner product $\ll \cdot, \cdot \gg$. Since $\nu_{\rho, \theta}$ is a product of Gamma distributions, the set of multivariate Laguerre polynomials form an orthogonal basis of ${\bb H}$. It is then possible to decompose $\bb H$ as an orthogonal sum $\oplus_{n \in \N} {\bb H}_n$ of subspaces ${\bb H}_n$ such that
\begin{equation*}
{\tilde{\mc S}}: {\bb H}_n \to {\bb H}_n, \quad {\tilde{\mc A}}:{\bb H}_n \to {\bb H}_{n-1} \oplus {\bb H}_n \oplus {\bb H}_{n+1}.
\end{equation*}
The function $W_{0,1}$ belongs to ${\bb H}_2$. Then we restrict the variational formula to functions $g \in {\bb H}_2$ and we estimate the corresponding new variational problem which is still infinite dimensional but involves only functions belonging to ${\bb H}_1 \oplus {\bb H}_2 \oplus {\bb H}_3$. To solve this variational problem we adapt ideas developed first in the context of Asymmetric Simple Exclusion Process (\cite{B1}, \cite{LQSY}) and exploited later for other models. One of the difficulties comes again from the fact that the noise is degenerate.
\end{itemize}
The extension of Theorem \ref{th:diffusivity} to other interacting potentials is a challenging problem. The general strategy presented here could be carried out but the orthogonal basis (formed by Laguerre polynomials in the exponential case) is no longer explicit and only defined by some recurrence relations.
\begin{acknowledgement}
This work has been supported by the Brazilian-French Network in Mathematics and the French Ministry of Education through the ANR grant EDNHS. The referees deserve thanks for careful reading and many useful comments.
\end{acknowledgement}
|
\section{Derivation of probability $p_Q^N$ for a singlet state $\psi_N^{(-)}$}
\subsection*{Lossless detection}
We aim at computing the following probability distribution for an arbitrary singlet state $\psi_N^{(-)}$
\begin{align*}
p_Q^N(n,m \mid\theta)=\big\lvert\langle \psi^{(-)}_{N}\vert n, N-n, (N-m)_{H+\theta }, m_{V+\theta}\rangle\big\rvert^2.
\end{align*}
At first, we will express the state $\Ket{(N-m)_{H+\theta }, m_{V+\theta}}$ in terms of the $(H,V)$ basis. The two bases, i.e. $(H,V)$ and $(H+\theta,V+\theta)$ are linked by the following rotation
\begin{equation}\left(
\begin{array}{c}b_{H+\theta }^{\dagger}\\
b_{V+\theta }^{\dagger}
\end{array}\right) =\left(
\begin{array}{cc}\cos{\theta }&\sin{\theta }\\
\- -\sin{\theta }&\cos{\theta }
\end{array}\right) \left( \begin{array}{c}b_{H}^{\dagger}\\b_{V }^{\dagger}
\end{array}\right).
\label{eq:11}
\end{equation}
Using Eq.~(\ref{eq:11}) we obtain
\begin{displaymath}
\begin{split}
&\Ket{(N-m)_{H+\theta }, m_{V+\theta}}=\frac{(b_{H+\theta }^{\dagger})^{N-m}(b_{V+\theta}^{\dagger})^{m}}{\sqrt{(N-m)!m!}}\Ket{0} \\&{}=\frac{1}{\sqrt{(N-m)!m!}}\sum_{p=0}^{N-m}\sum_{q=0}^{m}\binom{N-m}{p}\binom{m}{q}(-1)^{q}\\&\qquad(\sin{\theta })^{N-m-p+q}(\cos{\theta })^{m-q+p}\\&\qquad(b_{H }^{\dagger})^{p+q}(b_{V }^{\dagger})^{N-(p+q)}\Ket{0}
\end{split}
\end{displaymath}
\begin{displaymath}
\begin{split}
&{}=\frac{1}{\sqrt{(N-m)!m!}}\sum_{p=0}^{N-m}\sum_{q=0}^{m}\binom{N-m}{p}\binom{m}{q}(-1)^{q}\\&\qquad(\sin{\theta })^{N-m-p+q}(\cos{\theta })^{m-q+p}\\&\qquad\sqrt{(p+q)!(N-(p+q))!}\Ket{p+q,N-(p+q)}.
\end{split}
\end{displaymath}
We are now ready to compute $p_Q^N(n,m \mid\theta )$. The corresponding probability amplitude is given by
\begin{displaymath}
\begin{split}
&\langle \psi^{(-)}_{N}\vert n, N-n, (N-m)_{H+\theta }, m_{V+\theta}\rangle\\&={}\frac{1}{\sqrt{(N+1)(N-m)!m!}}\sum_{k=0}^{N}(-1)^{k}\sum_{p=0}^{N-m}\sum_{q=0}^{m}\binom{N-m}{p}\binom{m}{q}\\&\qquad(-1)^{q}(\sin{\theta })^{N-m-p+q}(\cos{\theta })^{m-q+p}\\&\qquad\sqrt{(p+q)!(N-(p+q))!}\\&\qquad\langle k, N-k, N-k, k\vert p+q,N-(p+q)\rangle \\&{}=\frac{1}{\sqrt{(N+1)(N-m)!m!}}\sum_{k=0}^{N}(-1)^{k}\sum_{p=0}^{N-m}\sum_{q=0}^{m}\binom{N-m}{p}\binom{m}{q}\\&\qquad(-1)^{q}(\sin{\theta })^{N-m-p+q}(\cos{\theta })^{m-q+p}\\&\qquad\sqrt{(p+q)!(N-(p+q))!}\\&\qquad\delta _{nk}\,\delta _{p+q, N-k}\\
&={}\frac{(-1)^{n}}{\sqrt{(N+1)(N-m)!m!}}\sum_{p=0}^{N-m}\sum_{q=0}^{m}\binom{N-m}{p}\binom{m}{q}\\&\qquad(-1)^{q}(\sin{\theta })^{N-m-p+q}(\cos{\theta })^{m-q+p}\\&\qquad\sqrt{(p+q)!(N-(p+q))!}\,\delta _{p+q, N-n}.
\end{split}
\end{displaymath}
In order to get rid of the Kronecker delta $\delta _{p+q, N-n}$, we notice that $p\in(0, N-m)$ and $0\leq q=N-n-p\geq m-n$. This implies the summation over $q$ from $q=i=\max\{0,m-n\}$ to $j=\min\{N-n,m\}$. Hence, the above probability amplitude simplifies to
\begin{equation}
\begin{split}
&\langle \psi^{(-)}_{N}\vert n, N-n, (N-m)_{H+\theta }, m_{V+\theta}\rangle \\&{}=(-1)^{n}\sqrt{\xi_{nm}^{(N)}}\sum_{q=i}^{q=j}\binom{N-m}{N-n-q}\binom{m}{q}\\&\qquad(-1)^{q}(\sin{\theta })^{2q+n-m}(\cos{\theta })^{N-(2q+n-m)}\\&{}=(-1)^{n}\sqrt{\xi_{nm}^{(N)}}(\cos{\theta })^{N}\sum_{q=i}^{q=j}\binom{N-m}{N-n-q}\binom{m}{q}\\&\qquad(-1)^{q}(\tan{\theta })^{2q+n-m},
\label{amp}
\end{split}
\end{equation}
where $\xi_{nm}^{(N)}=\frac{(N-n)!n!}{(N+1)(N-m)!m!}$. Square of the absolute value of (\ref{amp}) gives the probability $p_Q^N(n,m \mid\theta )$
\begin{align*}
&p_Q^N(n,m \mid\theta )=\xi _{nm}^{(N)}(\cos{\theta })^{2N}\\&\qquad\cdot\Bigg(\sum_{q=i}^{q=j}\binom{N-m}{N-n-q}\binom{m}{q} (-1)^{q}(\tan{\theta })^{2q+n-m}\Bigg)^{2}.
\end{align*}
\subsection*{Imperfect detection}
We assume that the probability distribution describing losses in photon counting detectors is given by the Bernoulli binomial distribution with the success probability (efficiency of detectors) $\eta $. We also assume that losses in each polarization mode are independent but, for simplicity, equal.
Imperfect detectors will measure $x$ and $y$ photons in modes $a_{H}$ and $b_{V + \theta }$, instead of $n\ge x$ and $m\ge y$, respectively. Therefore, for a $2N$-photon singlet state $\lvert\psi^{(-)}_{N}\rangle$, probability $p_Q^N(x, y\mid n, m, \theta )$ of measuring $x$ and $y$ photons, assuming that before losses there were $n$ and $m$ photons in modes $a_{H}$ and $b_{V + \theta }$, is given by
\begin{align}
p_Q^N&{}(x, y\mid n, m, \theta, \eta)= \binom{n}{x}\binom{m}{y}\eta ^{x+y}\\
&{}(1-\eta )^{n+m-x-y}p_Q^N(n,m \mid \theta ), \nonumber
\end{align}
where $n+m\leq 2N$. Since the numbers $n$ and $m$ are not known, the probability $p_Q^N(x, y\mid \theta )$ of detection $x$ and $y$ photons in modes $a_{H}$ and $b_{V + \theta }$ equals
\begin{equation}
P_Q^N(x, y\mid\theta,\eta)= \sum_{n=x}^{N}\sum_{m=y}^{N}p_Q^N(x, y\mid n, m, \theta,\eta)
\label{eq:prob_loss}.
\end{equation}
\section{Violation of chained inequality (8) by a singlet state $\psi_N^{(-)}$}
\subsection*{Lossless detection}
We rewrite inequality (8) as follows
\begin{align}
\label{eq:13}
&
\sum_{n,m=0}^{N}\lvert m - n\rvert \Lambda (n, m\mid N, \theta )\\&\qquad \geq
\sum_{n,m=0}^{N}\lvert m - n\rvert p_Q^N(n,m \mid\theta'),\nonumber
\end{align}
where we denote
\begin{displaymath}
\Lambda (n,m\mid N,\theta )= (2L-1)\,p_Q^N(n,m \mid\theta ),
\end{displaymath}
and notice that for $L\to\infty$, $\theta = \tfrac{\pi}{4L} \to0$ and $\theta'= \tfrac{(2L - 1)\pi}{4L}\to\tfrac{\pi }{2}$.
From now on, the notation $LHS$ and $RHS$ will stand for the left-hand side and the right-hand side of inequality~(\ref{eq:13}). At first, we will consider the RHS. Please notice that
\begin{align}
\label{eq:15}
&\sum_{n,m=0}^{N}\lvert m - n\rvert p_Q^N(n,m \mid\theta' )\\
&{}=\kern-.5em\sum_{n+m<N}^{N}\kern-.5em\lvert m - n\rvert p_Q^N(n,m \mid\theta' )+\kern-.5em\sum_{n+m> N}^{N}\kern-.5em\lvert m - n\rvert p_Q^N(n,m \mid\theta' )\nonumber\\&\qquad{}+\sum_{n+m=N}^{N}\lvert m - n\rvert p_Q^N(n,m \mid\theta' ).\nonumber
\end{align}
In Eq.~(\ref{amp}) we have that $i=q_{min}=\max\{0,m-n\}$ and $j=q_{max}=\min\{N-n,m\}$. For simplicity, we will denote $K(q)=2q+n-m$. Now, we have to consider the following three cases. $1^{\circ }$ If $n+m<N$ then $q_{max}=m$ and $K(q_{max})=n+m<N$. Hence, $K(q)<N$ and $(\cos{\theta'})^{N}(\tan{\theta'})^{K(q)}\rightarrow 0$ if $L\rightarrow \infty$ $(\theta' \rightarrow 0 )$. $2^{\circ }$ Similarly, if $n+m<N$ then $q_{max}=N-n$ and $K(q_{max})=2N-(n+m)<N$. Thus, again $K(q)<N$ and $(\cos{\theta'})^{N}(\tan{\theta'})^{K(q)}\rightarrow 0$. From $1^{\circ }$ and $2^{\circ }$ we conclude that the two first sums in (\ref{eq:15}) vanish for $L\rightarrow \infty$. $3^{\circ }$ If $n+m=N$, we obtain $q_{max}=m=N-n$ and $K(q_{max})=n+m=N$. Since $(\cos{\theta'})^{N}(\tan{\theta'})^{N} \to 1$ and $(\cos{\theta'})^{N}(\tan{\theta'})^{K(q)} \to 0$ for $K(q)<K(q_{max})$, therefore only the last component of the third sum in (\ref{eq:15}) contributes. Since for $n+m=N$ and $q=m$, $\xi_{nm}^{(N)}=\frac{1}{N+1}$ and $\binom{N-m}{N-n-q} \binom{m}{q}=1$, we obtain $p_Q^N(n,m \mid\theta' )\rightarrow \xi_{nm}^{(N)}=\frac{1}{N+1}$, which is grater than $0$. Hence, the expression (\ref{eq:15}) for $L\rightarrow \infty$ takes the following form
\begin{align}
\label{eq:16}
&\sum_{n,m=0}^{N}\lvert m - n\rvert p_Q^N(n,m \mid\theta')\\&\qquad{}=\frac{1}{N+1}\sum_{\substack{n,m=0\\ n+m=N}}^{N}\lvert m - n\rvert\nonumber\\&\qquad{}=\begin{cases} \frac{2}{N+1}\sum_{n=0}^{\frac{N-1}{2}}(N-2n) &\text{for $N$ -- odd,} \\ \frac{2}{N+1}\sum_{n=0}^{\frac{N-2}{2}}(N-2n)
&\text{for $N$ -- even,} \end{cases}\nonumber\\&\qquad{}=\begin{cases} \frac{\frac{1}{2}N^{2}+N+\frac{1}{2}}{(N+1)} &\text{for $N$ -- odd,}\\ \frac{\frac{1}{2}N^{2}+N}{(N+1)} &\text{for $N$ -- even.} \end{cases}\nonumber
\end{align}
Thus, we conclude that for large population
\begin{align}
&RHS \propto N, \quad \textrm{for} \quad L \to \infty.
\end{align}
\begin{figure}
\centering
\includegraphics[width=6.5cm]{Rosolek_SM1.eps}
\caption{Bell parameter $B_{Q}^N$ as a function of number of settings $L$ computed for different photon numbers $2N$: $2N=2$ (red line), $2N=4$ (blue line), $2N=6$ (black line), $2N=8$ (green line), $2N=10$ (orange line), $2N=12$ (yellow line).}
\label{Limit}
\end{figure}
Now, we will show that the function $\Lambda (n,m\mid N,\theta )$ vanishes for $\theta \to 0$ ($L \to \infty$). Inserting $p_Q^N(n,m \mid\theta)$, $\Lambda$ can be rewritten as follows
\begin{displaymath}
\begin{split}
&\Lambda (n,m\mid N, \theta )= \xi_{nm}^{(N)} \Bigg(\sqrt{2L-1}\cos^N{\theta}\\&\qquad{}\cdot\sum_{q=i}^{q=j}\binom{N-m}{N-n-q}
\binom{m}{q}(-1)^{q}(\tan{\theta})^{2q+n-m}\Bigg)^{2}.
\end{split}
\end{displaymath}
Please notice that for $m\neq n$ (if $m=n$, $\lvert m - n\rvert=0$) and $q \geq \max\{0,m-n\}$ we have that $K(q)=2q+n-m>0$. Using de L'Hospital rule, we check that $\sqrt{2L-1}\tan{\theta}\to 0$ for $L \to \infty$. From this fact and knowing that $(\cos{\theta})^{N}\to 1$ and $(\tan{\theta})^{l}\to 0$ ($l$ is arbitrary positive number) for $L \to \infty$, we show that $\sqrt{2L-1}(\cos{\theta})^{N}(\tan{\theta})^{K(q)}\to 0$ for $L \to \infty$. Hence, $\Lambda (n,m\mid N, \theta )\to 0$. From the above considerations we conclude that for an arbitrary $N$
\begin{align}
& LHS\rightarrow 0, \quad \textrm{for} \quad L\to \infty.
\end{align}
We introduce a Bell parameter in the form of $B_{Q}^N={}\mathrm{LHS} - \mathrm{RHS}$, which according to LHV theories is always positive. Fig.~\ref{Limit} shows $B_{Q}^{N}$ as a function of the number of settings $L$ for various photon numbers $2N$. Independently of $N$, for sufficiently large $L$, $B_{Q}^{N}$ tends to a negative value.
\begin{figure*}[th]
\centering
\raisebox{3.0cm}{(a)}\hskip-0.3cm
\includegraphics[width=5.7cm]{Rosolek_SM2a.eps}
\raisebox{3.0cm}{(b)}\hskip-0.3cm
\includegraphics[width=5.7cm]{Rosolek_SM2b.eps}
\raisebox{3.0cm}{(c)}\hskip-0.3cm
\includegraphics[width=5.7cm]{Rosolek_SM2c.eps}
\caption{Bell parameter $B_{Q}^N$ as a function of efficiency of detectors $\eta $ evaluated for different number of settings $L$: $L=2$ (red line), $L=4$ (blue line), $L=6$ (green line), $L=8$ (orange line), $L=10$ (yellow line) for (a) $2N=2$, (b) $2N=8$ (c) $2N=12$.}
\label{fig:L2N2}
\end{figure*}
\subsection*{Imperfect detection}
We have investigated how fragile is the violation of $B_Q^N$ in case of a non-unit detection efficiency $\eta$. In order to compute dependence of the Bell parameter on $\eta$, we replace the probability $p_Q^N(n,m \mid\theta)$ with
$P_Q^N(x,y \mid\theta,\eta)$ given in (\ref{eq:prob_loss}). Inequality (8) is modified as follows
\begin{align}
&
\sum_{x,y=0}^{N}\lvert x-y\rvert \Lambda (x,y \mid N, \theta, \eta )\\&\qquad \geq
\sum_{x,y=0}^{N}\lvert x-y\rvert P_Q^N(x,y \mid\theta', \eta),\nonumber
\end{align}
where we denote
\begin{displaymath}
\Lambda (x,y\mid N,\theta, \eta )= (2L-1)\,P_Q^N(x,y \mid\theta, \eta ),
\end{displaymath}
Fig.~\ref{fig:L2N2}a displays numerical results obtained for the simplest case $2N=2$ and for various number of settings. As expected, violation of the local bound is possible only above the usual minimal value of efficiency of detectors $\eta >\frac{2}{1+\sqrt{2}}\approx 83\%$. The violation takes its maximal value $-0.4$ in the limit of perfect detection, $\eta = 1$. The violation persists for all $L$'s but is most pronounced for the highest (considered) number of settings $L=10$ and reaches $-0.8$. For larger photon numbers, in order to observe any violation, the minimal efficiency as well as the minimal number of settings increase, see Fig.~\ref{fig:L2N2}bc. For $2N=8$ one must have $L\ge 4$ to have a violation. For $2N=12$ the threshold is $L \ge 6$. Interestingly, the minimal efficiency required for violation can be smaller for larger number of settings, compare the green and yellow curves in Fig.~\ref{fig:L2N2}c.
\section{Violation of chained inequality (8) by a squeezed vacuum state (SV) in case of lossless detection}
We rewrite inequality (8) for a squeezed vacuum state (BSV) (1) in a similar way as we did in the previous paragraph
\begin{align}
\label{eq:13BSV}
&\sum_{N=0}^{\infty} \lambda_N^2
\sum_{n,m=0}^{N}\lvert m - n\rvert \Lambda (n, m\mid N, \theta )\\&\qquad \geq
\sum_{N=0}^{\infty} \lambda_N^2
\sum_{n,m=0}^{N}\lvert m - n\rvert p_Q^N(n,m \mid\theta').\nonumber
\end{align}
Using Eq.~(\ref{eq:16}) and the fact that $\lambda _{N}^{2}=\cosh^{-4}(g)(N+1)\tanh^{2N}(\Gamma )$, we compute the $RHS$ in the limit of $L \to \infty$ to be
\begin{displaymath}
RHS = \frac{\sinh^{3}(2\Gamma )}{\sinh(4\Gamma )}>0.
\end{displaymath}
Since the $LHS$ for $L \to \infty$ tends to zero independently of $N$ (it was shown in the previous paragraph), we conclude that the Bell parameter defined as $B_Q=LHS-RHS$ of (\ref{eq:13BSV}) takes negative values $B_Q \to -e^{2\Gamma}$.
\section{Violations of chained inequalities (8) by a squeezed vacuum state (SV) in case of imperfect detection}
In the case of imperfect detection, $\eta<1$, probability of detecting $x$ and $y$ photons by Alice and Bob, respectively, reads
\begin{equation}
P_Q(x, y\mid\theta, \eta) = \sum_{N=0}^{\infty}\lvert\lambda _{N}\rvert^{2}P_Q^N(x, y\mid\theta, \eta),
\end{equation}
where $\lvert\lambda _{N}\rvert^{2}$ is the probabilistic weight of $\lvert\psi^{(-)}_{N}\rangle$ in the squeezed vacuum state, and $P_Q^N(x, y\mid\theta, \eta)$ were discussed in the previous sections, see (\ref{eq:prob_loss}) .
However, as we are not able to sum above expression to infinity, we introduce a cut-off parameter $N_{max}$, and define
\begin{equation}
P^{approx}_Q(x, y\mid\theta, \eta, N_{max}) = \sum_{N=0}^{N_{max}}\lvert\lambda _{N}\rvert^{2}P_Q^N(x, y \mid\theta, \eta), \label{PR}
\end{equation}
where
$N_{max}$ is
such that for a given amplification gain $\Gamma$ we nearly fulfill the normalization condition
\begin{equation}
\sum_{N=0}^{N_{max}}\lvert\lambda _{N}\rvert^{2}\geq 0.99.
\end{equation}
We have carefully checked whether we took large enough values of $N_{max}$ for our numerical computations, so that the results presented in the main text do not change significantly for larger values.
We take the probabilities given in the previous section (\ref{PR}) and directly insert it into inequality (8)
\begin{align}
\label{INEQQ}
{}&\sum_{x,y=0}^{N_{max}}\lvert x - y\rvert (2L-1) P^{approx}_Q(x, y\mid \theta, \eta, N_{max})\\ \geq&{}\sum_{x,y=0}^{N_{max}}\lvert x - y\rvert P^{approx}_Q(x, y\mid \theta', \eta, N_{max}).\nonumber
\end{align}
For the gains $\Gamma<1$, we have checked that $N_{max}=10$ gives a good approximation for the infinite sum.
The most important numerical results for our discussion, concerning violation of the inequality (\ref{INEQQ}), are shown in Fig.\ 5 of the main text.
\end{document}
|
\section{Introduction}
Proposed in 1984, quantum key distribution (QKD) allows two users to exchange provably secure keys via a potentially insecure quantum channel~\cite{bb84}.~Since then, QKD has attracted much attention and significant progress has been made in both theory and practice~\cite{Gisin2002, Scarani2009}.~On the application front, however, the operating distance of practical fibre-based QKD systems is limited to about 150 km~\cite{lo2014}, which is mainly due to the high background noise produced by commonly used semiconductor single-photon detectors~\cite{Hadfield2009, eisaman2011} (SPDs) and the stringent demand on the minimum classical-post-processing (CPP) block size~\cite{Scarani2008, tomamichel2012, lim2014}.~Here, we present, for the first time, a compact and autonomous QKD system that is capable of distributing provably-secure cryptographic key over 307\,km of ultra-low-loss optical fibre (51.9\,dB loss). The system is based on a recently developed standard semiconductor (inGaAs) SPDs~\cite{korzh2014a} with record low background noise and a novel efficient finite-key security analysis for QKD. This demonstrates the feasibility of practical long-distance QKD based on standard fibre-optic telecom components.
In order to achieve long-distance QKD, existing systems usually resort to using superconducting nanowire single-photon detectors (SNSPD) which can achieve a low dark count rate (DCR), but require cryogenic temperatures (\textless3~K). In addition, a weaker security framework is usually taken, i.e., by assuming individual attacks instead of coherent attacks. Crucially, all record-distance demonstrations to date have ignored corrections due to finite-length keys~\cite{Scarani2008, tomamichel2012, lim2014}; in particular, it has been shown that corrections due to finite-length keys are non-negligible for realistic CPP block sizes. This means that previous QKD demonstrations might be overly optimistic in the achievable distance.~Table \ref{tab1} summarises recent fibre based QKD demonstrations. For distances beyond 160\,km, the systems required the use of SNSPDs and thus are incompatible with compact implementations. To the best of our knowledge, only a handful of QKD implementations take finite-length key corrections into account for their security analyses. For example, see the demonstration by Lucamarini {\it et al}~\cite{lucamarini2013}, which distributed provably-secure keys using a fibre of about 90\,km. Although we have restricted our comparison to fibre based prepare-and-measure discrete-variable experiments, we note that significant progress has also been made in continuous variable \cite{jouguet2013} and free-space QKD \cite{nauerth2013, wang2013}.
\begin{table*}
\caption{\label{tab1} Summary of notable QKD demonstrations, outlining the maximum achievable distance, detector type used, and the level of security. Where the finite-key analysis was not carried out the CPP block size, $n_\tn{cpp}$, was left blank and the security parameter, $\epsilon_\text{qkd}$ could not be defined.}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
&\multicolumn{2}{c}{Quantum channel}&\multicolumn{2}{c}{Detector}&&\multicolumn{3}{c}{Security}\\ \hline
& Length&Attenuation& Type & Temperature & Protocol & Attacks & Finite-key size & $\epsilon_\text{qkd}$ & $r_\text{sec}$ \\
& (km) & (dB) & & (K) & & & & & (bps) \\ \hline
This work & 307 & 51.9 & InGaAs~\footnotemark[3] & 153 & COW & Collective & 6.6$\times 10^5$ &$4 \times 10^{-9}$ & 3.18 \\
Wang 2012 \cite{wang2012}& 260 & 52.9 & SNSPD & 1.7 & DPS & Individual & - & - & 1.85 \\
Stucki 2009 \cite{stucki2009}& 250 & 42.9 & SNSPD & 2.5 & COW & Collective & - & - & 15 \\
Takesue 2007 \cite{takesue2007}& 200 & 42.1 & SNSPD & 3 & DPS & Individual & - & - & 12.1 \\
Liu 2010 \cite{liu2010}&200& - & SNSPD & 2.4 & BB84 & Collective & - & - & 15 \\
Rosenberg 2009 \cite{rosenberg2009} & 135 & 27.8 & SNSPD & 3 & BB84 & Collective & - & - & 0.2 \\
Namekata 2011 \cite{namekata2011}& 160 & 33.6 & InGaAs~\footnotemark[4] & 193 & DPS & Individual & - & - & 490 \\
Yuan 2009 \cite{yuan2009} & 100 & 20 & InGaAs~\footnotemark[4] & 243 & BB84 & Collective & - & - & $1.01\times10^4$ \\
Shimizu 2014 \cite{shimizu2014}& 90~\footnotemark[1] & 30 & SNSPD & 2.5 & DPS & Individual & - & - & 1100\\
Lucamarini 2013 \cite{lucamarini2013} & 80~\footnotemark[2] & 16 & InGaAs~\footnotemark[4] & 243 & BB84 & Collective &$\sim 10^{9}$ & $\sim 10^{-10}$ & $1.20 \times 10^{5}$\\
Walenta 2014 \cite{walenta2014}& 25~\footnotemark[2] & 5.3 & InGaAs~\footnotemark[4] & 293 & COW & Collective & $10^{6}$ & $4 \times 10^{-9}$ & $2.25 \times 10^4$ \\
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{Installed fibre link}
\footnotetext[2]{Wavelength multiplexing of classical and quantum channels over a single fibre}
\footnotetext[3]{Free-running operation}
\footnotetext[4]{Gated operation}
\end{table*}
\section{Results}
Our system is based on the coherent one-way (COW)~\cite{stucki2005} QKD protocol, where the bit string is encoded in the time of arrival of weak coherent laser pulses (WCPs) and the channel disturbance is monitored by measuring the visibility of the interference between neighbouring pulses. That is, bit~0~and~1~are sent using $\ket{\alpha_0}:=\ket{0}\ket{\alpha}$ and $\ket{\alpha_1}:=\ket{\alpha}\ket{0}$, respectively. On Bob's side, he simply recovers the bit value by measuring the arrival time of the laser pulse, e.g., bit 1 is detected if there is a detection in the later time. To detect attacks on $\ket{\alpha_0}$ and $\ket{\alpha_1}$,~Alice randomly sends an additional test state,~$\ket{\alpha_t}:=\ket{\alpha}\ket{\alpha}$,~to check for phase coherence between any two successive laser pulses. Therefore, phase coherence can be checked in any of these sequences,~$\ket{\alpha_0}\ket{\alpha_1}$, $\ket{\alpha_0}\ket{\alpha_t}$, $\ket{\alpha_t}\ket{\alpha_1}$, $\ket{\alpha_t}$, $\ket{\alpha_t}\T$,~by using an imbalanced interferometer with a pulse delay on Bob's side. The physical implementation is outlined in figure~\ref{fig:setup}.
The security of our QKD system is based on the universally composable security framework~\cite{RennerThesis2005, Quade2009}.~In particular, we say that our QKD is $\epsilon_{\tn{qkd}}$-{secure} whenever it is $\eps_{\tn{sec}}$-{secret} and $\eps_{\tn{cor}}$-{correct}. Here, $\eps_{\tn{sec}}$-secret means that the output secret key is distinguishable from an ideal secret key with probability at most $\eps_{\tn{sec}}$, and $\eps_{\tn{cor}}$-correct means that probability of Alice and Bob having identical secret keys is at least $1-\eps_{\tn{cor}}$. Using this security framework and under the assumption of collective attacks~\cite{Branciard2008}, we are able to derive a bound on the maximum value of the extractable secret key length, $\ell$, in terms of a fixed security parameter $\epsilon_{\tn{qkd}}$ and the observed statistics. \todo{We note that in Ref~\cite{Moroder2012}, it has been shown that there is only a small advantage for the adversary if she uses general attacks instead of collective attacks. }Accordingly, this allows us to select the appropriate family of universal hash functions for privacy amplification, which extracts a secret key of size $\ell$ from a weakly random string of size $n_\tn{cpp}$. Specifically, for some parameter $\beta \in (0,\epsilon_{\tn{qkd}}/4)$, laser pulse intensity $\mu$, CPP block size, $n_{\tn{cpp}}$, and the number of bits revealed during error correction, $m_{\rm IR}$, we can extract a $\epsilon_{\tn{qkd}}$-secure key of length (see Supplementary Information (SI) for the full security analysis)
\begin{multline}\label{Main_Eq}
\ell \leq \max_{\beta}\Bigg\lfloor n_\tn{cpp}\left[1-\hat{Q}-(1-\hat{Q})h\left(\frac{1+\xi(\mu,\hat{V}}{2}\right)\right] \\
-7\sqrt{n_\tn{cpp}\log_2 \frac{1}{\beta}}-m_{\rm IR}-\log_2\frac{2}{4\eps_{\tn{cor}}\beta^2} \Bigg \rfloor, \end{multline}
where $h(x)$ is the binary entropy function and $\xi(a,b):=(2\hat{V}-1)\exp(-\mu)-2\sqrt{(1-\exp(-2\mu))\hat{V}(1-\hat{V})}$. Here, $\hat{Q}$ and $\hat{V}$ are the measured quantum bit error rate (QBER) in the raw string and the visibility, respectively.
\begin{figure*}
\includegraphics[width=17cm]{Fig1}
\caption{\label{fig:setup} Experimental setup of the COW QKD system. Alice's part (left) consists of a continuous wave DFB laser at 1550\,nm which is modulated using an electro-optic intensity modulator, before passing through a set of attenuators to regulate the photon number per pulse. The state preparation frequency is set to 625\,MHz. Bob's system consists of an asymmetric beam splitter which provides a passive choice of measurement basis, following which the photons either travel directly to the $\text{SPD}_\text{D}$ (Data detector) or pass through an imbalanced interferometer. The phase of the interferometer is maintained such that $\text{SPD}_\text{M}$ is on the destructive port, and the visibility can be calculated by registering the detections due to the interfering (two neighbouring pulses) and non-interfering events (single pulse). The two InGaAs/InP NFAD SPDs are both cooled by a single stirling cooler. The quantum channel consists of ULL fibre whilst a clock signal is distributed from Alice to Bob via the synchronisation channel, both of which are equal in length. All of the processes from the state preparation through to the sifting are carried with the use of a field-programable gate array (FPGA) on each side, whilst the subsequent CPP (error correction, etc) is completed in on-board software.~An ethernet link is used as the service channel and all of the passed messages are authenticated.}
\end{figure*}
\begin{figure}
\includegraphics[width=9cm]{skr_ull}
\caption{\label{fig:skr}~(\textbf{a}) Numerical optimisation of the SKRs versus distance for different CPP block sizes $n_\tn{cpp}=10^s$ with $s = 4, 5, 6, 7$ (left to right). The results are calculated using the old bound~\cite{walenta2014} (dashed blue lines) and the bound presented in this work (solid red lines). The asymptotic limit is represented by the black dash-dot line. Experimental parameters used for the calculation were taken as for the 307~km experimental measurement and the visibility was set to 98.0\%.~(\textbf{b}) Experimental final secret key rate versus distance. Theoretical plots for different temperature limits are plotted, which show the advantage of increasing the temperature at shorter fibre lengths, which reduces the detector saturation.~(\textbf{c}) QBER (black triangles) and visibility (blue squares) measured for each fibre length.}
\end{figure}
In our system, the QBER is directly measured by counting how many bits are flipped during the error corruption step, meaning that, provided the verification step passes, the uncertainty in this measurement is zero. For the estimation of visibility, we need to consider the problem of random sampling without replacement, i.e., based on the observed visibility $V_{\tn{obs}}$ in the monitoring line, we want to estimate the visibility $V_{\tn{key}}$ of the quantum signals used to generate the secret key. More formally, for some positive deviation term $t$, we like to show that $V_{\tn{key}}\leq V_{\tn{obs}}-t$ is highly unlikely; note that $\hat{V}=V_{\tn{obs}}-t$. In the literature, finite-key security analyses commonly use the Chernoff-Hoeffding tail inequality or Serfling tail inequality to solve this problem. However, these inequalities are only optimal for sufficiently large sample sizes, and thus are not suitable for long-distance QKD where sample sizes are small. To resolve this issue, we derive a new tail inequality that is specifically tailored for random sampling without replacement. In particular, we exploit the fact that the distribution of errors in a random sample is described by the hypergeometric distribution, and then use tight bounds for binomial coefficients to derive an upper bound on $\Pr\left[ V_{\tn{key}}\leq V_{\tn{obs}}-t \right]$. More precisely, we are able to show that the following relation,
\begin{equation}\label{Main_Eq2}
V_{\tn{key}} \overset{}{\geq} V_{\tn{obs}}-t (n_{\tn{cpp}},n_\tn{vis}, V_{\tn{obs}},\beta),
\end{equation}holds with probability at least $1-\beta$, and where $n_{\tn{vis}}$ is the number of events used to calculate $V_{\tn{obs}}$. The explicit expression of $t (n_{\tn{cpp}},n_\tn{vis}, V_{\tn{obs}},\beta)$ is deferred to the Methods section. In comparison to existing tail inequalities, the new inequality takes into account the measured error rate of the random sample and thus provides a much sharper bound on the tail event, i.e.,$V_{\tn{key}}\leq V_{\tn{obs}}-t $. To illustrate the tightness of this new bound, figure~\ref{fig:skr}a shows the numerically optimised secret key rates (SKRs) for different CPP block sizes based on the previous bound~\cite{walenta2014} and the new tail inequality. The performance enhancement with small $n_{\tn{cpp}}$ is significant, allowing key distribution to around 310\,km even with $n_{\tn{cpp}}=10^5$, which would not have been possible with the old bound. At 300\,km the SKR reaches 84.5\% of the asymptotic limit with $n_{\tn{cpp}}=10^7$, which is an order of magnitude smaller than previously thought required~\cite{lucamarini2013}.
Another enabling factor of this work has been our recent demonstration that detectors based on InGaAs/InP negative feedback avalanche diodes (NFADs)~\cite{Itzler2011} can operate with extremely low noise in the free-running regime~\cite{korzh2014a}. This marked an improvement of over two orders of magnitude compared to previous InGaAs detectors. Due to a separate avalanche and multiplication~\cite{Itzler2011} structure typically used in such SPDs, where InGaAs is used for the absorption region, and InP for the avalanche region, dark carriers are generated by two different mechanisms. In fact, at temperatures above 200\,K, thermal dark carrier generation in the absorption region is dominant: below this temperature, trap-assisted tunnelling (TAT) in the multiplication region becomes the dominant effect. Generally, TAT has a small temperature dependence, however, the breakdown voltage ($V_{\text{BR}}$) of InP has a linear dependance on temperature. This means that reducing the temperature below 200\,K still leads to a reduction of the DCR, because the operating voltage is reduced, lowering the TAT contribution which is a field-dependent effect~\cite{korzh2014b}. The NFADs used in this report have a ($V_{\text{BR}}$) temperature coefficient such that below 200\,K the DCR drops by approximately a factor of two with every 10\,K. We have observed this trend down to temperatures of less than 150\,K, where a DCR of a few counts per second can be achieved at detection efficiencies of more than 20\%, which is comparable to many SNSPDs~\cite{eisaman2011}.
\begin{figure}
\includegraphics[width=9cm]{stability_200km}
\caption{\label{fig:stability} QKD system stability over 64 hours with 200\,km of fibre showing~(\textbf{a}) the SKR and (\textbf{b}) the corresponding QBER and visibility as a function of time. The detector temperature was set to 183 K. one of the advantages of the COW protocol is that the bit string is encoded in the time basis and the phase basis is only used for the monitoring of the channel disturbance. Hence, even if the visibility is low, the errors in the raw bit string are not necessarily affected, meaning only the eavesdroppers mutual information is altered but not the $m_{\rm IR}$ in Eq.~\ref{Main_Eq}. This is contrary to the DPS protocol \cite{shimizu2014}, where the bit string is encoded in the phase, such that both of the terms would be affected. As seen in (\textbf{a}), the QBER fluctuations are significantly smaller than those of the visibility, highlighting the importance of this.}
\end{figure}
In practice it is not always beneficial to operate the detector at the lowest temperature since, for a given dead-time, the after-pulse probability increases exponentially with reducing temperatures~\cite{korzh2014b}.~This is not a problem because as the channel length is increased (i.e., increasing channel loss), the detection rates drop exponentially, allowing a longer dead-time to be applied without effecting the detection rates. Therefore, for each distance there exists an optimum detector temperature, which should be set such that the DCR is close to being the dominant source of the QBER.
The QKD system was tested over fibre lengths of 100-307\,km and the operating temperature was varied between 223-153\,K, whilst the dead time was varied between 8-115~$\mu s$\,(see SI for details of the experimental settings).~The SKRs achieved at each distance are shown in figure~\ref{fig:skr}b and the corresponding error rates are plotted in figure~\ref{fig:skr}c. Theoretical curves are plotted for the minimum and maximum detector temperatures to demonstrate the adaptability of the system at different distances.\,A SKR of 12.7\,kbps was generated at 104\,km, while at 307~km the SKR was 3.18~bps. Since the DCR increases faster with temperature above 200\,K~\cite{korzh2014b}, it was not feasible to increase the temperature to more than 223\,K, where a minimum achievable dead time was around 8\,$\mu$s. For this reason, the NFAD SPDs are optimal for distances $\textgreater$100\,km.\,For shorter distances, it is better to use rapid gating detectors operating as high as room temperature, which do not require dead time \cite{walenta2012}.~At the longest distance, the visibility dropped to $97.0\%$ (from $\textgreater 98\%$) due to the increased difficulty of its stabilisation (see Methods), due to the large integration times required. We adapted the $n_\tn{cpp}$ for each fibre length to maintain reasonable collection times (see SI for details) and could use $n_\tn{cpp}=6.6\times10^5$ even at the longest distance. Figure~\ref{fig:stability} shows the system performance over a continuous period of 70\,hours at a distance of 200\,km, showing that stable operation could be maintained, with automatic tracking of the temporal alignment and the visibility. The average QBER and visibility were $1.55\%$ and $97.7\%$ respectively whilst the SKR was around 900\,bps.
\section{Conclusion}
To summarise, we have demonstrated a robust and autonomous QKD system, based on practical and compact InGaAs SPDs, which achieves secure key distribution over 307~km.~Moreover, we have sharpened the finite-key security analysis, which improves the performance of the QKD protocol when using small CPP block sizes. This has enabled us to provide a quantifiable security parameter for the complete protocol ($\epsilon_{\tn{qkd}} = 4\times10^{-9}$), which has not been possible before for QKD systems operating over 150\,km.~This work demonstrates that practical, robust and autonomous QKD is feasible over very long distances even with standard telecom components in a rack mounted architecture.\newline
\section{\label{sec:methods}Additional Information}
\textbf{Experimental details} The secret key post-processing flow is illustrated in figure~\ref{fig:setup}.~The information reconciliation (IR) is carried out using the CASCADE algorithm~\cite{Brassard1994}, which was chosen over an LDPC solution~\cite{walenta2014} due to its increased efficiency of processing small blocks and the fact that high throughput was not required. The IR processing block size was typically 2-3 orders of magnitude smaller than $n_{\tn{cpp}}$, meaning that a frequent measure of the QBER could be obtained, which is useful for active stabilisations of parameters such as the intensity modulator bias voltage. Once the error corrected string reaches the size of the CPP block, error verification is carried out, followed by computation of equation~(\ref{Main_Eq}) to set the compression ratio of the subsequent privacy amplification (PA) step. We operate in a trusted detector scenario, meaning that the dark count contribution to $\hat{V}$ and $\hat{Q}$ in equation~(\ref{Main_Eq}) are subtracted. To facilitate this, only the detector DCR has to be determined accurately, however, it is straightforward to characterise, even at random times. The security level in this work was chosen such that $\beta = 10^{-9}$,~whilst the failure probabilities of the error verification and service channel authentication were $\sim10^{-11}$ and $\sim10^{-15}$.~This gives us an upper bound on the total security parameter for our system of $\epsilon_{\tn{qkd}} = 4 \times 10^{-9}$,~for all fibre lengths tested in this work.
To achieve long term operational stability, the fibre length mismatch between the synchronisation and quantum channels is tracked by oversampling the input detector signal at 2.5\,GHz and using the subsequent statistics to shift the detection window even before significant errors are induced. The visibility specifically, was stabilised by adjusting the wavelength of the DFB laser through the adjustment of the drive current.
The SPD temperatures were achieved using a stirling cycle cooler, which are significantly more efficient than electrical Peltier coolers and are capable of achieving significantly lower temperatures. Such cooling technology will enable the complete QKD system to be integrated into telecom standard rack mountable package.
\textbf{Sketch of security analysis.~}
Here, we briefly sketch the proof for the bound on the secret key length, i.e., equation~(\ref{Main_Eq}), and the complete security analysis is deferred to the SI. The technical part of our security analysis lies in finding a bound on the smooth min-entropy of ${X}$ given ${E}$, i.e., $H_{\min}^{\epsilon}({X}|{E})$, where $\epsilon$ is the smoothing parameter, ${X}$ is the random variable describing the raw key and ${E}$ represents the overall knowledge of the adversary. To arrive at equation~(\ref{Main_Eq}), we first use the quantum leftover hash lemma~\cite{RennerThesis2005}, which says that the extractable secret key length $\ell$ is approximately equal to the $H_{\min}^{\epsilon}({X}|{E})$.
Second, by using certain chain rules for smooth entropies and assuming collective attacks, we are able to put a lower bound on the smooth min-entropy of ${X}$ given ${E}$ in terms of the von Neumann entropy of ${X}$ given ${E}$. This essentially allows us to use the main results from Branciard {\it et al}~\cite{Branciard2008}., where they found a lower bound on the von Neumann entropy of ${X}$ given ${E}$ for the COW protocol in terms of the expected bit error rate and visibility. In order to account for finite-size correction of the visibility statistics, we derive a tight tail inequality that allows us to relate the expected visibility $V_\tn{key}$ to the observed $V_{\tn{obs}}$. More specifically, let $\lambda=1/2-V_{\tn{obs}}/2$, then the deviation term in Eq.~(\ref{Main_Eq2}) (for simplicity we let $t(n_{\tn{cpp}},n_{\tn{vis}},\lambda,\epsilon)=t$) is given as
\begin{equation}
t=\sqrt{\frac{8(n_\tn{cpp}+n_\tn{vis})\lambda(1-\lambda)}{n_\tn{vis}n_\tn{cpp}}\log\frac{\sqrt{n_\tn{cpp}+n_\tn{vis}}C}{\sqrt{2\pi n_\tn{cpp}n_\tn{vis}\lambda(1-\lambda)}\epsilon}}, \end{equation}
where
\begin{equation}
C=e^{\left(\frac{1}{8(n_\tn{cpp}+n_\tn{vis})}+\frac{1}{12n_\tn{vis}}-\frac{1}{12n_\tn{vis}\lambda+1}-\frac{1}{12n_\tn{vis}(1-\lambda)+1}\right)}.
\end{equation} In other words, $V_{\tn{key}} \geq V_{\tn{obs}}-t=:\hat{V}$ is true except with probability $\epsilon$. Using this tail inequality together with the above arguments, we thus arrive at an upper bound for $\ell$ which is essentially dependent on the security parameter $\epsilon_\tn{qkd}$, observed statistics, the CPP block size $n_\tn{cpp}$ and the random sample size $n_\tn{vis}$.
\textbf{Ultra-low-loss fibre.~} Beyond improving the performance of the QKD system, significant gains can be achieved by reducing the loss of the quantum channel. Due to this, developments in optical fibre technology hold at important role for QKD performance. The ultra-low-loss fibres based on silica were achieved by taking careful considerations to the doping levels and the manufacturing process (refer to SI for full discussion). The fibers used in the current work have an average attenuation of 0.160\,dB/km without splices and connectors.~It is expected that it could be possible to create fibers with an attenuation of less than 0.1\,dB/km in the future \cite{tsujikawa2007}, which would mean that the distance of QKD could increase to over 500\,km.
\section*{Acknowledgements}
We would like to acknowledge Nino Walenta for useful discussions and Olivier Guinnard for technical assistance. We thank ID Quantique for providing the error correction software used in this work as well as Mathilde Soucarros for technical support.~This work was supported by the Swiss NCCR QSIT project.
|
\section{Introduction}
The astrophysical site that produces elements heavier than iron by rapid neutron capture reactions (r-process) remains poorly understood. This is particularly true for the third peak in the abundance of r-process elements near atomic mass $\approx195$. The need for a high flux of free neutrons for rapid neutron capture reactions requires an astrophysical source with an abundance of free neutrons \citep{BBFH}. Given this, the two astrophysical sites most commonly considered for the origin of r-process elements are core-collapse supernovae and binary neutron star (NS) (or neutron star--black hole) mergers.
The neutrino-driven wind produced after a successful core-collapse supernova explosion was originally viewed as a promising site for the production of r-process nuclei \citep[e.g.][]{Meyer1992}. However, increasingly sophisticated treatments of the dynamics and neutrino transport in neutrino-driven winds have systematically failed to produce the thermodynamic conditions necessary for producing third-peak r-process nuclides \citep[e.g.][]{Qian1996,Thompson2001}. This is even more true of many modern core-collapse simulations, which find that the neutrino driven wind has a proton-rich rather than neutron-rich composition \citep[e.g.][]{Hudepohl2010}. Despite this general failure of core-collapse supernova models to produce r-process nuclei, there are sufficient remaining uncertainties that this conclusion is not yet definitive. In addition to uncertainties in neutrino physics and the core-collapse explosion mechanism, strong magnetic fields and rapid rotation change the thermodynamic conditions in neutrino-driven winds, making the conditions more favourable for the production of r-process nuclei \citep[e.g.][]{Metzger2007}.
In contrast to core-collapse supernovae, NS mergers can robustly produce third peak r-process nuclei. In particular, the decompression of neutron-rich matter from nuclear densities produces rapid neutron capture nucleosynthesis and a distribution of elements that is reasonably consistent with observations of the Sun and other stars \citep[e.g.][]{Lattimer1977,Freiburghaus1999}. Unbound, neutron-rich matter is produced dynamically during NS mergers, in e.g. tidal tails, and in winds from the rotationally-supported accretion disk left after the merger \citep[e.g.][]{Lattimer1974,Metzger2009}.
NS mergers are very likely to be the first source of gravitational waves directly detected by the ground-based interferometers Advanced LIGO and VIRGO \citep{Abbott2009, Accadia2012}. In this context, there is renewed motivation for assessing NS mergers as the source of r-process nuclei. In particular, the decay of heavy elements produced during r-process nucleosynthesis powers isotropic electromagnetic emission that is predicted to be one of the most promising sources of electromagnetic emission coincident with a gravitational-wave detection \citep[e.g.][]{Metzger2010,Barnes2013}. Constraints on r-process production during NS mergers thus directly constrains the properties of electromagnetic sources powered by r-process nucleosynthesis and, at least in principle, the rate of NS mergers themselves.
Given that NS mergers, unlike supernovae, robustly produce third peak r-process nuclei, a natural hypothesis is that the majority of the r-process elements in nature are produced by NS mergers. The observed abundances of r-process elements in galactic stars provide some of the most direct constraints on this hypothesis. Such observations show that even very metal poor stars have a solar r-process abundance pattern and that there is only modest scatter in the r-process to iron abundances of stars \citep[e.g.][]{Sneden2008}. In simple models of galactic chemical enrichment, these observations are difficult to reproduce \citep[e.g.][]{Qian2000, Argast2004, Matteucci2014}. This is because NS mergers are relatively rare compared to supernovae, which increases the scatter in the abundance of r-process elements and decreases the likelihood that early generations of stars (i.e.\ low-metallicity stars) are enriched by r-process nucleosynthesis.
However, existing quantitative models of galactic r-process enrichment rely on simplified chemical evolution calculations that do not capture the full complexity of how galaxies form. In particular, stellar feedback during galaxy formation drives powerful winds which redistribute heavy elements far from their birth locations \citep{Veilleux2005}. This mixing is not present in idealized chemical evolution calculations, which thus likely overestimate the star-to-star scatter in heavy-element abundances.
Motivated by these potential deficiencies of existing models of galactic r-process enrichment, this paper studies the abundances of r-process elements using state-of-the-art galaxy formation simulations, which include galactic winds and hydrodynamic flows associated with structure formation (thus alleviating some of the problems that plague simple semi-analytical chemical evolution models). We explore a large range of parameters values in a simple model for r-process enrichment due to NS mergers, finding stellar abundance ratios in broad agreement with observations.
In Section~\ref{sec:sim} we describe the simulation we used; the details of the different r-process element enrichment models can be found in Section~\ref{sec:models}. In Section~\ref{sec:results} we show the abundances for these models and how they correlate with stellar age and galactocentric distance. We discuss and summarize our results and their convergence in Section~\ref{sec:concl}.
\section{Method} \label{sec:sim}
We use a newly developed version of TreeSPH (P-SPH) which adopts the Lagrangian `pressure-entropy' formulation of the smoothed particle hydrodynamics (SPH) equations \citep{Hopkins2013PSPH}. The gravity solver is a heavily modified version of \textsc{gadget}-2 \citep[last described in][]{Springel2005}. P-SPH also includes substantial improvements in the artificial viscosity, entropy diffusion, adaptive timestepping, smoothing kernel, and gravitational softening algorithm.
This work is part of the Feedback in Realistic Environments (FIRE) project\footnote{http://fire.northwestern.edu/} \citep{Hopkins2014FIRE}, which consists of several cosmological `zoom-in' simulations. Here we make use of galaxy \textbf{m12i}, a galaxy with mass similar to the Milky Way at $z=0$. Its initial conditions are taken from the AGORA project \citep{Kim2014}. The simulation is described in \citet{Hopkins2014FIRE} and references therein and we will only summarize its main properties here.
This cosmological simulation assumes a $\Lambda$CDM cosmology with parameters consistent with the 9-yr Wilkinson Microwave Anisotropy Probe (WMAP) results \citep{Hinshaw2013}, $\Omega_\mathrm{m}= 1-\Omega_\Lambda = 0.272$, $\Omega_\mathrm{b}= 0.0455$, $h = 0.702$, $\sigma_8 = 0.807$ and $n = 0.961$.
\begin{table*}
\begin{center}
\caption{\label{tab:sims} \small Simulation parameters: simulation resolution identifier, initial mass of gas particles ($m_\mathrm{baryon}$), mass of dark matter particles ($m_\mathrm{DM}$), minimum baryonic force softening ($\epsilon_\mathrm{baryon}$), minimum dark matter force softening ($\epsilon_\mathrm{DM}$), final simulation redshift ($z_\mathrm{final}$), number of star particles within 50~proper~kpc of the galaxy centre at $z_\mathrm{final}$ ($N_\mathrm{star}$).}
\begin{tabular}[t]{llllllr}
\hline
\hline \\[-3mm]
resolution & $m_\mathrm{baryon}$ (M$_\odot$) & $m_\mathrm{DM}$ (M$_\odot$) & $\epsilon_\mathrm{baryon}$ ($h^{-1}$pc) & $\epsilon_\mathrm{DM}$ ($h^{-1}$pc) & $z_\mathrm{final}$ & $N_\mathrm{star}$ \\
\hline \\[-4mm]
low & $4.5\times10^5$ & $2.3\times10^6$ & 30 & 250 & 0 & 117,901 \\
fiducial & $5.7\times10^4$ & $2.8\times10^5$ & 14 & 100 & 0 & 611,995 \\
high & $7.1\times10^3$ & $3.5\times10^4$ & 14 & 100 & 2.4 & 410,995 \\
\hline
\end{tabular}
\end{center}
\end{table*}
Because of the expansion of the Universe and the collapse of structure, the volume of the zoom-in region depends on redshift, but is always much larger than the virial radius. It is about a Mpc in radius at $z=0$. The (initial) particle masses for baryons and dark matter are $5.7\times10^4$~M$_\odot$ and $2.8\times10^5$~M$_\odot$, respectively, for our fiducial resolution. We also consider simulations with eight (two) times lower and higher mass (spatial) resolution. Particle masses and minimum physical softening lengths for all the simulations used in this work are listed in Table~\ref{tab:sims}, as well as the final redshift of each simulation and the number of star particles included in our analysis, i.e.\ all star particles within 50~kpc of the centre of the galaxy. The minimum physical baryonic force softening length for our fiducial resolution is 14~$h^{-1}$pc. We adopt a quintic spline kernel with an adaptive size, keeping the mass within the kernel approximately equal. The average number of neighbours in the smoothing kernel is 62.
Star formation takes place in molecular, self-gravitating gas above a hydrogen number density of $n_\mathrm{H}^\star>10$~cm$^{-3}$, where the molecular fraction is calculated following \citet{Krumholz2011} and the self-gravitating criterion following \citet{Hopkins2013SelfGrav}. Stars are formed at the rate $\dot\rho_\star=\rho_\mathrm{molecular}/t_\mathrm{ff}$, where $t_\mathrm{ff}$ is the free-fall time. The star formation parameters are kept fixed when we vary the resolution. Star particles inherit their metal abundances from their progenitor gas particle.
We assume an initial stellar mass function (IMF) from \citet{Kroupa2002}. Radiative cooling and heating are computed in the presence of the cosmic microwave background (CMB) radiation and the ultraviolet (UV)/X-ray background from \citet{Faucher2009}. Self-shielding is accounted for with a local Sobolev/Jeans length approximation. A temperature floor of 10~K or the CMB temperature is imposed.
The primordial abundances are $X = 0.76$ and $Y = 0.24$, where $X$ and $Y$ are the mass fractions of hydrogen and helium, respectively. The simulation has a metallicity floor at $Z=10^{-4}$~Z$_\odot$, because yields are very uncertain at lower metallicities. The abundances of 11 elements (H, He, C, N, O, Ne, Mg, Si, S, Ca and Fe) produced by massive and intermediate-mass stars (through Type Ia supernovae, Type II supernovae, and stellar winds) are computed following \citet{Iwamoto1999}, \citet{Woosley1995}, and \citet{Izzard2004}. There is \emph{no} sub-resolution metal diffusion in our simulation.
Supernovae and stellar winds are modelled by having a fraction of the mass ejected by a star particle transferred to its neighbouring gas particle $j$ as follows:
\begin{equation}
f_j = \dfrac{\frac{m_j}{\rho_j} W(r_j,h_\mathrm{sml})}{\Sigma_i \frac{m_i}{\rho_i} W(r_i,h_\mathrm{sml})},
\end{equation}
where $h_\mathrm{sml}$ is the smoothing length of the star particle (determined in the same manner as for gas particles), $r_i$ is the distance from the star particle to neighbour $i$, $W$ is the quintic SPH kernel, and the summation is over all SPH neighbours of the star particle, i.e.\ 62 on average.
The FIRE simulations include an implementation of stellar feedback by supernovae, radiation pressure, stellar winds, and photo-ionization and photo-electric heating (see \citealt{Hopkins2014FIRE} and references therein for details). For the purposes of the present paper, we emphasize that these simulations produce galaxies with stellar masses reasonably consistent with observations over a wide range of dark matter halo masses. This is a consequence of galactic winds driven by stellar feedback. The same winds also redistribute the heavy elements created in stars, both into the surrounding circum-galactic medium (CGM) and elsewhere in the galactic disk via a `galactic fountain.' This feedback-induced mixing is important for the chemical evolution of the galaxy.
The galaxy in our fiducial simulation is an isolated, star-forming, bulge-dominated galaxy. It builds up a disc component at low redshift ($z<0.5$), which kinematically comprises of about a third of the stars in the galaxy. Since these stars are young, however, their contribution to the total stellar light (or light-weighted kinematics) is larger. Measured within the central 20~kpc at $z=0$, the star formation rate is about 3~M$_\odot$yr$^{-1}$, the stellar mass is $3\times10^{10}$~M$_\odot$, the total gas mass is $9\times10^{9}$~M$_\odot$, and the mass of the star-forming gas is $9\times10^{8}$~M$_\odot$. We have repeated our analysis for more disc-dominated as well as for more bulge-dominated galaxies within the FIRE sample \citep{Hopkins2014FIRE} and found our results for the r-process abundances to be very similar.
\subsection{r-process models} \label{sec:models}
R-process elements are implemented as tracers and have no effect on the simulation. As a result, we are able to implement multiple models of r-process enrichment in the same hydrodynamic simulation, saving considerable computational time.
We assume that NS mergers are responsible for all of the r-process enrichment. However, the rates of NS mergers are quite uncertain so we consider a variety of models to bracket this uncertainty. Using the known population of binary neutron stars in the Milky Way, \citet{Abadie2010} estimate that the merger rate is $\sim 10^{-4}$ yr$^{-1}$, but that the rate could range from $10^{-6}-10^{-3}$ yr$^{-1}$ under extreme assumptions. For the purposes of understanding the enrichment of low-metallicity stars, it is also important to know the delay time distribution that characterizes the merger rate as a function of time after a burst of star formation. Population synthesis calculations suggest a rate $\propto t^{-1}$ once lower mass core collapse supernovae have exploded, i.e., after a delay of $\sim 10^7$ yrs (e.g., \citealt{Belczynski2006}). Given these constraints, we take the
rate of NS mergers ($R_\mathrm{NS}$) to be
\begin{equation} \label{eqn:rate}
R_\mathrm{NS} = \frac{A M_\mathrm{star}}{t} \ \mathrm{for} \ t > t_\mathrm{min}
\end{equation}
and $R_\mathrm{NS}=0$ otherwise, where $A$ is the number of NS mergers per unit of stellar mass, $M_\mathrm{star}$ is the stellar mass under consideration, $t$ is time since the formation of the star particle, and $t_\mathrm{min}$ is the minimum time needed for a NS merger to take place. In the simulations the rate of NS mergers is implemented stochastically, since we do not resolve individual stars. We take fiducial values of $A=10^{-5}$~M$_\odot^{-1}$ and $t_\mathrm{min}=3\times10^7$~yr. The stellar mass of our Milky Way model at $z = 0$ is about $3\times10^{10}$~M$_\odot$, with half of the stars having formed by $z\approx1$. Thus our fiducial choice of $A$ corresponds to a present day merger rate of $\sim10^{-4}$~yr$^{-1}$, similar to that suggested by observations.
Given the significant uncertainties in the NS merger rate, however, we explore a range of models with $t_\mathrm{min}$ ranging from $3\times10^6$ to $3\times10^8$~yr and $A$ ranging from $3\times10^{-6}$ to $3\times10^{-5}$~M$_\odot^{-1}$ (see Table~\ref{tab:models}). Although we focus on model 0, with $A=10^{-5}$~M$_\odot^{-1}$ and $t_\mathrm{min}=3\times10^7$~yr, we stress that all of the models that we consider are plausible and so provide an estimate of the systematic uncertainty in the chemical evolution of r-process elements implied by the uncertainty in the NS merger rate and delay time distribution.
Our calculations do not include NS kicks, though it would clearly be of interest to do so in future work. The median offset of short-duration gamma-ray bursts is modest, about 4.5~kpc, though $\sim25$ per cent of events have offsets greater than 10~kpc \citep{Fong2013}. We discuss this uncertainty further in Section~\ref{sec:concl}.
The amount of r-process elements ejected per NS merger is uncertain. However, an ejecta mass $\sim10^{-2} M_\odot$ per merger together with a merger rate $\sim 10^{-4}$ yr$^{-1}$ in the Milky Way is sufficient to produce the total mass of r-process nuclei in the Galaxy, which is $\sim10^4 M_\odot$ assuming all stars have roughly solar abundances.
Since in our simulation the r-process elements act only as tracer particles, we choose to simply normalize our six models to solar metallicity, i.e. to set [r-process/Fe]=0 when [Fe/H]=0. This should be a reasonable approximation unless there is a significant systematic variation in the r-process yields per NS merger with redshift or stellar population age.
Another important parameter characterizing r-process enrichment is the mass of the ambient interstellar medium (ISM) that is initially enriched with r-process elements. Treating the ejecta from NS mergers as an expanding remnant analogous to a supernova remnant, the final momentum of the remnant during the momentum conserving phase is \citep{Cioffi1988}
\begin{equation}\label{eqn:momentum}
M v|_\mathrm{final} \approx 6 \times 10^4 E_{50}^{13/14} n_0^{-1/7}~\mathrm{M_\odot\,km\,s^{-1}},
\end{equation}
where $n_0$ is the ambient density in cm$^{-3}$ and we have scaled the initial energy of the remnant, $E_{50}$, to correspond to an ejecta mass of $0.01$~M$_\odot$ moving at $v = 0.1$~c, i.e. $10^{50}$~ergs \citep{Piran2013}. Equation~\ref{eqn:momentum} is for solar metallicity, but the final momentum is only about twice as large for a primordial composition. The total swept-up mass when the neutron-rich ejecta comes into approximate pressure equilibrium with the ambient medium is set by evaluating Equation~\ref{eqn:momentum} given a final velocity comparable to the sound speed or turbulent velocity of the ambient medium. This suggests that the neutron-rich ejecta is initially incorporated into $M_{\rm swept} \sim 10^{3.5-4}$~M$_\odot$ of the ambient ISM, depending on the exact ambient density and turbulent velocity.
In simple chemical evolution calculations, the neutron-rich ejecta are typically taken to mix with \emph{only} a mass $M_{\rm swept}$ \citep[e.g.][]{Qian2000, Argast2004}. It is unlikely, however, that this is correct, given the vigorous turbulent mixing induced by stellar feedback, galactic winds, galactic fountains, galaxy mergers, instabilities in the differentially rotating galactic disc, etc. (as also argued by \citealt{Piran2014}).
To address this, one would ideally like to have a hydrodynamic simulation with a mass resolution comparable to $M_{\rm swept}$ since the simulation would (at least in principle) self-consistently resolve additional mixing produced by galaxy-scale turbulence and winds. This is somewhat prohibitive, however. Our fiducial simulation has a baryonic particle mass of $5.7 \times 10^4 M_\odot$ while our highest resolution simulation discussed in Section~\ref{sec:res} has a baryonic particle mass of $7 \times 10^3$~M$_\odot$. However, the mass in the kernel -- which is the mass to which the metal enrichment is applied -- is roughly 62 times larger. In Section~\ref{sec:concl} we show and discuss the dependence of our results on resolution. It is important to bear in mind that in all of our simulations the r-process enrichment initially occurs in a mass somewhat larger than the initial mass swept-up by a NS merger remnant. As we show below, the uncertainties in r-process enrichment associated with uncertainties in the NS merger rate are comparable to those due to the finite numerical resolution of our simulations.
\begin{table}
\begin{center}
\caption{\label{tab:models} \small Parameters of r-process enrichment models: model number, number of NS mergers per $M_\odot$ of stars (see Eq.~\ref{eqn:rate}), minimum time needed for the first NS merger (see Eq.~\ref{eqn:rate}), NS merger rate at $z=0$. For each model we vary only one parameter, indicated in bold.}
\begin{tabular}[t]{crrc}
\hline
\hline \\[-3mm]
model & $A$ [M$_\odot^{-1}$] & $t_\mathrm{min}$ [yr] & $R_\mathrm{NS}(z=0)$ [yr$^{-1}$]\\
\hline \\[-4mm]
0 & $10^{-5}$ & $3\times10^7$ & $1.5\times10^{-4}$\\
1 & $10^{-5}$ & $\mathbf{3\times10^6}$ & $2.1\times10^{-4}$\\
2 & $10^{-5}$ & $\mathbf{10^7}$ & $ 1.8\times10^{-4}$\\
3 & $10^{-5}$ & $\mathbf{10^8}$ & $1.1\times10^{-4}$ \\
4 & $\mathbf{3\times10^{-6}}$ & $3\times10^7$ & $4.4\times10^{-5}$\\
5 & $\mathbf{3\times10^{-5}}$ & $3\times10^7$ & $4.4\times10^{-4}$\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Results} \label{sec:results}
Throughout the paper we express abundance ratios of a star compared to those of the Sun as
\begin{equation}
[A/B]=\mathrm{log}_{10}\left(\frac{N_A}{N_B}\right)_\mathrm{star}-\mathrm{log}_{10}\left(\frac{N_A}{N_B}\right)_\odot,
\end{equation}
where $A$ and $B$ are different elements, $N_{\mathrm{A}}$ and $N_{\mathrm{B}}$ are number densities. Given that our simulations use a metallicity floor at $Z=10^{-4}$~Z$_\odot$ for all metal species, except tracer species, [Fe/H]$\ge-4$ and [Mg/H]$\ge-4$. We therefore have to be careful interpreting our results at low metallicity. Results are shown for [Fe/H]$\ge-3.5$, a factor three above the metallicity floor, where our results are not strongly affected by the choice of metallicity floor. The initial conditions, however, do not include a metallicity floor for r-process elements, so particles can have $\mathrm{[\rp/Fe]}=-\infty$. We take those particles into account as well when we calculate the median and percentiles below.
\begin{figure}
\center
\includegraphics[scale=.46]{figures/star_abund_rp0_z0.eps}
\caption {\label{fig:rpFe} [Mg/Fe] (top panel) and [r-process/Fe] (bottom panel) as a function of [Fe/H] at $z=0$ for our fiducial resolution simulation. The colour coding indicates the logarithm of the stellar mass per pixel and both images are 200 by 200 pixels. Black curves show the median (solid curves) and 16th and 84th percentiles (dashed curves). [r-process/Fe] has been normalized so that its median is zero at $\mathrm{[Fe/H]}=0$. [Mg/Fe] is not normalized but is set by the supernova yields. Black plusses and downward arrows are observed [Eu/Fe] detections and upper limits, respectively \citep{Suda2008}. The median [Mg/Fe] and [r-process/Fe] are both fairly constant with metallicity, although they decrease slightly at $\mathrm{[Fe/H]}>0$. [r-process/Fe] decreases strongly at $\mathrm{[Fe/H]}<-3$. The scatter in [Mg/Fe] is small at all metallicities, but the scatter in [r-process/Fe] increases significantly towards lower metallicity.}
\end{figure}
Figure~\ref{fig:rpFe} shows the $z=0$ abundance ratios of Mg (top panel) and r-process elements (bottom panel) to Fe as a function of metallicity for our fiducial resolution simulation. We include all star particles at radii $R<50$~kpc from the centre of the galaxy, but our results do not depend on this choice. Black curves show the median (solid curves) and 16th and 84th percentiles (dashed curves). As described in Section~\ref{sec:models}, [r-process/Fe] has been normalized so that its median is zero at $\mathrm{[Fe/H]}=0$. [Mg/Fe] is computed directly from the heavy element yields used in the simulation and is not normalized. The bottom panel of Figure~\ref{fig:rpFe} also shows observational data points for [Eu/Fe], where plusses show detections and downward arrows show upper limits \citep{Suda2008}.\footnote{We exclude carbon-enhanced metal-poor stars, since their origin is not well understood.} Eu is an ideal tracer of r-process enrichment, because it is almost solely produced in r-process events \citep{Burris2000}. It is not possible to compare our models and the observations in detail, because of unknown selection effects in the heterogeneous observational sample. Global trends are likely robust.
Mg is the $\alpha$ element most easily observed and therefore most useful for comparison with observations. It is thought to be primarily produced in Type II supernovae. Consistent with observations of the Milky Way \citep[e.g.][]{Cayrel2004, Arnone2005}, we find very little scatter in [Mg/Fe] for both metal-poor and metal-rich stars, little variation in the median at $\mathrm{[Fe/H]}<-1$ and a small decrease at high metallicity, due to the increasing importance of Type Ia supernovae, which mainly produce iron-peak elements. The median [Mg/Fe] in our simulation is, however, lower by about 0.3~dex as compared to observations. This discrepancy is well-known and is due to the \citet{Woosley1995} yields for Mg being too low by a factor of a few \citep[e.g.][]{Timmes1995, Thomas1998, Portinari1998, Lia2002}. Because Mg is not a dominant coolant in our simulation, this discrepancy in Mg abundance has no effect on the evolution of the galaxy and could be straightforwardly fixed by adopting better yields in future simulations or by increasing the Mg abundance in post-processing. We choose to leave the Mg abundances as is in Figure~\ref{fig:rpFe} to accurately reflect the actual results in our simulation. Moreover, the trends in Mg abundance with metallicity and the scatter in Figure~\ref{fig:rpFe} are robust, since Mg is almost purely produced in Type II supernovae. In future work, we plan to compare the abundance ratios of the various metal species, the mass-metallicity relation, and metallicity gradients in our different galaxies to observations. We include the Mg results in the present paper primarily to highlight the different scatter relative to the r-process elements.
As is the case for Mg, the decrease of [r-process/Fe] at $\mathrm{[Fe/H]}>0$ in Figure~\ref{fig:rpFe} is due to Fe production by Type Ia supernovae, which become relatively more important at late times. This is also consistent with observations, although there the decrease sets in at somewhat lower metallicities. This could be due to selection biases in the observations or to the absense of certain mixing processes in our simulation. See Section \ref{sec:concl} for more discussion on the latter possibility. The median [r-process/Fe] is approximately flat at $\mathrm{[Fe/H]}<0$ in the simulation at our fiducial resolution, whereas the data show a small increase to $\mathrm{[\rp/Fe]}\approx0.5$. In both the observations and simulation the scatter increases substantially towards low metallicity. Note that many of the low-metallicity observations are upper limits (shown as downward arrows). Overall, our simulation is in reasonable agreement with observations, especially taking into account the uncertainties in the metal yields as well as the unknown selection effects in the data.
\begin{figure}
\center
\includegraphics[scale=.5]{figures/star_rp_median_scatter.eps}
\caption {\label{fig:models} Median [r-process/Fe] (top panel) and scatter ($\sigma$) in [r-process/Fe] (bottom panel) as a function of [Fe/H] at $z=0$ for the six different models listed in Table~\ref{tab:models} and indicated in the legend. The black, solid curve is for model 0, also shown in Figure~\ref{fig:rpFe}. The thin, black, dot-dashed curve is model 0 for a low resolution simulation, described in Section~\ref{sec:res}. Dotted curves show variations of $t_\mathrm{min}$. Dashed curves show variations of $A$. As $t_\mathrm{min}$ increases or $A$ decreases, [r-process/Fe] decreases and the scatter increases, especially at lower metallicities.}
\end{figure}
As mentioned before, the NS merger rate is highly uncertain and we therefore explore different NS merger models. Figure~\ref{fig:models} shows the median [r-process/Fe] (top panel) and scatter in [r-process/Fe] (bottom panel) as a function of [Fe/H] for the six different models listed in Table~\ref{tab:models}. The solid, black curve is identical to the one in Figure~\ref{fig:rpFe}. The scatter is calculated by subtracting the 16th percentile from the 84th percentile (dashed curves in Figure~\ref{fig:rpFe}) and dividing by two. Dotted curves show variations of $t_\mathrm{min}$. Dashed curves show variations of $A$, as indicated in the legend. We reiterate that given the uncertainties in the NS merger rate, all of these models are plausible.
The different models span a range of median [r-process/Fe] values of only $0.4$~dex at $\mathrm{[Fe/H]}=-2$, but differ significantly at lower metallicity. With higher $t_\mathrm{min}$, i.e.\ a larger delay time before the first NS merger takes place, or lower $A$, i.e.\ fewer NS mergers per unit of stellar mass, the median [r-process/Fe] decreases at low metallicity, the drop in r-process element abundances moves to higher metallicity, and the scatter in [r-process/Fe] increases. This is due to NS mergers becoming more rare, which increases the stochasticity. Vice versa, when we decrease $t_\mathrm{min}$ or increase $A$, [r-process/Fe] increases, the low-metallicity drop moves to lower metallicities, and the scatter decreases. For all our models, the median and scatter are large enough that an appreciable fraction of low-metallicity stars has high r-process abundance ratios. The 84th percentile of the distribution has $[\rp/\mathrm{Fe}]>0$ for all models.
\begin{figure}
\center
\includegraphics[scale=.5]{figures/star_rpFe_median.eps}
\caption {\label{fig:zradius} Median galactocentric radius at $z=0$ (top panel) and median redshift at which the stars were formed (bottom panel) as a function of [r-process/Fe] for different [Fe/H] bins as indicated in the legend. As the metallicity increases, the distance from the galaxy centre and the stellar age decrease. At fixed metallicity, distance and age are lowest for stars with $\mathrm{[\rp/Fe]}\approx0-0.5$ and increase as the absolute value of [r-process/Fe] increases. This shows that the stars responsible for the large scatter in [r-process/Fe] are, on average, old and currently in the halo of the galaxy.}
\end{figure}
Lower-metallicity stars and stars with [r-process/Fe] more different from solar are, on average, older and located at larger distances from the centre of the galaxy. This is quantified in Figure~\ref{fig:zradius}, where the median galactocentric radius ($\langle R\rangle$, top panel) and the median formation redshift ($\langle z_\mathrm{form}\rangle$, bottom panel) of stars are shown as a function of [r-process/Fe]. The iron abundance increases for the different curves from top to bottom as indicated by the legend. The star particles have been selected in the same way as for Figs.~\ref{fig:rpFe} and~\ref{fig:models}, i.e.\ within 50~kpc of the centre of the galaxy.
Stars with supersolar metallicities are found close to the centre and have been formed very recently, with $\langle R\rangle=0.7$~kpc and $\langle z_\mathrm{form}\rangle=0.1$. Radius and formation redshift increase with decreasing stellar metallicity. Stars with $-3.5<\mathrm{[Fe/H]}<-3$ have $\langle R\rangle=16$~kpc and $\langle z_\mathrm{form}\rangle=2.9$. This is quantitatively consistent with observations of the Milky Way, where halo stars are found to be older and more metal poor than disc stars. Note that most of our low-metallicity stars were not accreted, but formed in situ at high redshift and moved to larger radii through dynamical effects.
At high metallicity, stars with $\mathrm{[\rp/Fe]}\approx0$ have the smallest $R$ and $z_\mathrm{form}$. This minimum moves to $\mathrm{[\rp/Fe]}\approx0.6$ for $-3<\mathrm{[Fe/H]}<-2$. Outliers in [r-process/Fe] are found at larger radii and are older, on average. For example, for $\mathrm{[Fe/H]}\approx-1$, $\langle R\rangle=2$~kpc and $\langle z_\mathrm{form}\rangle=0.3$ at $\mathrm{[\rp/Fe]}\approx0$, but at $\mathrm{[\rp/Fe]}\approx-1$, $\langle R\rangle=8$~kpc and $\langle z_\mathrm{form}\rangle=2.1$. This indicates that the scatter is driven by inhomogeneous chemical evolution at high redshift. Stars formed at high redshift are moved to larger radii through dynamical effects. The dependence of $\langle R\rangle$ and $\langle z_\mathrm{form}\rangle$ on [r-process/Fe] disappears for stars with $\mathrm{[Fe/H]}<-3$.
\subsection{Resolution test} \label{sec:res}
As discussed in Section~\ref{sec:models}, the mass of the ISM initially enriched by r-process nuclei is one of the key parameters setting the chemical evolution of neutron-rich nuclei. Typically it is assumed that the r-process elements mix only with the initial swept-up mass in the NS merger remnant \citep[e.g.][]{Qian2000, Argast2004}. Given the many sources of turbulent mixing in galaxies, however, it is by no means clear that this physical picture is correct. It is, however, the case that due to limited resolution, the r-process nuclei in our simulations initially enrich an ISM mass larger than the ejecta mass swept up prior to the NS merger remnant reaching pressure equilibrium with its surroundings (see Eq.~\ref{eqn:momentum} and associated discussion). In particular, we reiterate that in our SPH formulation all gas particles inside the kernel of a given star particle are enriched, i.e.\ 62 on average, weighted according to their SPH weights. Therefore, the enriched mass, i.e.\ about $3.5\times10^6$~M$_\odot$, is approximately $62$ times larger than the particle mass for our fiducial resolution, although it is not enriched uniformly. To test the sensitivity of our results to the mass resolution, we ran a simulation with 8 (2) times \emph{lower} mass (spatial) resolution down to $z=0$ and repeated our analysis. We checked that the galoctocentric distances and formation redshifts of stars shown in Figure~\ref{fig:zradius} are similar to the ones at low resolution. The Mg abundances and scatter are also very similar in both simulations. We conclude that these properties are converged.
The lower resolution results for model 0 are shown as thin, dot-dashed curves in Figure~\ref{fig:models}. Comparing to the fiducial resolution, we see that our r-process abundances are not completely converged. At low resolution, [r-process/Fe] decreases slightly towards $\mathrm{[Fe/H]}=0$ and there is no drop at low metallicity.\footnote{We do see a drop at this resolution for model 4, at $-3.5<\mathrm{[Fe/H]}<-3$, i.e.\ at lower metallicity than for our fiducial resolution simulation.} The scatter is lower at $\mathrm{[Fe/H]}<-1$, but still appreciable. This is not surprising, because lower resolution simulations mix the r-process elements into a larger initial mass, i.e.\ the smoothing kernel, reducing the dispersion in [r-process/Fe]. We note, however, that the lower resolution results are similar to some of our physically plausible variants at the fiducial resolution, in particular the models with more NS mergers per M$_\odot$ of stars (model 5) or shorter delay times (models 1 and 2).
\begin{figure}
\center
\includegraphics[scale=.5]{figures/star_rp_median_scatter_z2p4.eps}
\caption {\label{fig:models_z2p4} Same as Figure~\ref{fig:models}, but for a high resolution simulation at $z=2.4$. The black, dot-dashed curve is model 0 for our fiducial resolution at $z=2.4$. We normalized [r-process/Fe] at $\mathrm{[Fe/H]}=-1$ due to the low number of stars at higher metallicities. At higher resolution, the drop in median [r-process/Fe] moves to higher metallicity and becomes less pronounced; the scatter in [r-process/Fe] also increases. Our different models for the NS merger rate show significant variations for the median [r-process/Fe] at low metallicity, as well as for scatter in [r-process/Fe] at $\mathrm{[Fe/H]}>-3$. Models 1, 2, and 5 at high resolution are reasonably similar to the fiducial resolution model 0.
}
\end{figure}
We also ran a high resolution simulation with 8 (2) times \emph{higher} mass (spatial) resolution, but only down to $z=2.4$, since it is more computationally expensive. Figure~\ref{fig:zradius} shows that the low-metallicity stars are formed at high redshift and thus the abundance ratios at the low-metallicity end are already in place at $z=2.4$. Figure~\ref{fig:models_z2p4} shows the median [r-process/Fe] and scatter for our high-resolution simulation at $z=2.4$. We normalized [r-process/Fe] at $\mathrm{[Fe/H]}=-1$ due to the low number of stars at higher metallicities. The lack of sufficiently many stars also causes the results to be somewhat noisy at high metallicity. The black, dot-dashed curve in Figure~\ref{fig:models_z2p4} is model 0 for our fiducial resolution at $z=2.4$. Comparing it to the solid line in Figure~\ref{fig:models} shows that there is indeed not much evolution in the r-process element abundances at low metallicity from $z = 2.4$ to $z = 0$.
Figure~\ref{fig:models_z2p4} shows that at higher resolution, the drop in median [r-process/Fe] moves to higher metallicity ($\mathrm{[Fe/H]}\approx-2$) and becomes less abrupt. The scatter increases at $\mathrm{[Fe/H]}>-3$. However, the 84th percentile of the distribution remains supersolar at all metallicities. Moreover, models 1, 2, and 5 at high resolution are reasonably similar to the fiducial resolution model 0, with the median [r-process/Fe] only decreasing at $\mathrm{[Fe/H]}<-2.5$, perhaps more consistent with the observations shown in Figure~\ref{fig:rpFe}.
The resolution tests shown here demonstrate that simulated r-process element abundance ratios are sensitive to the resolution of the simulation at $\mathrm{[Fe/H]}\lesssim-2$ (for the resolutions we were able to explore). The qualitative change in our results with resolution is similar to what one would expect on physical grounds, namely that higher mass resolution generally leads to somewhat larger scatter and a lower median r-process abundance at low metallicity. Comparing to the observational data shown in Figure~\ref{fig:rpFe}, Figure~\ref{fig:models_z2p4} shows that our higher resolution simulations seem to match the observed median slope of [r-process/Fe] as a function of metallicity less well. This could be an indication of extra mixing processes these simulations are not capturing; see Section 4 for a discussion. Again, the observations are very heterogeneous with unknown selection effects, so a direct comparison is non-trivial. This sensitivity to resolution is, however, specific to NS merger models and is not present in our [Mg/Fe] or [Fe/H] abundance ratios, which are produced by supernovae. The difference, of course, is that NS mergers are much rarer. Another key result of our resolution tests is that the uncertainty in the NS merger rate and the delay time distribution produces an uncertainty in the r-process enrichment that is comparable to the differences between our fiducial and high-resolution simulations.
\section{Discussion and Conclusions} \label{sec:concl}
We have quantified the abundance pattern of neutron-rich r-process nuclei in cosmological zoom-in simulations of a Milky Way-mass galaxy enriched by simplified implementations of binary neutron star (NS) mergers. We have compared the r-process nucleosynthesis to that of Mg and Fe, which are predominantly produced by Type II and Ia supernovae, respectively. We have explored a range of models for the rate and delay time of NS mergers, intended to roughly bracket the wide range of models consistent with the known binary NS population and population synthesis calculations of merger rates (see \S~\ref{sec:models}). The r-process elements are passive tracers that do not affect the simulation dynamics in any way. Our primary conclusions are:
\begin{itemize}
\item Neutron star (NS) mergers can produce [r-process/Fe] abundance ratios and scatter that are broadly consistent with observations for stars with $-2\lesssim \mathrm{[Fe/H]}\lesssim 0$.
\item The uncertain rate and delay time distribution of NS mergers results in a large uncertainty in the r-process abundances. Some of our assumed NS merger models resemble available observations better than others.
\item The results for the r-process abundance ratios are not fully converged, particularly at low metallicity, i.e.\ at $\mathrm{[Fe/H]}\lesssim-2$. This is physical and is due to the low rate of NS mergers and the difficulty resolving mixing on scales of the ISM of galaxies even in our highest resolution cosmological zoom-in simulations.
\item At low metallicity, the scatter in [r-process/Fe] is large enough to explain the r-process abundances of observed low-metallicity stars. However, we also predict a large population of stars with significantly subsolar [r-process/Fe], which have not yet been observed. The existence of this population depends sensitively on the rate of turbulent mixing in the ISM. It is very likely that our simulations do not capture all of the key mixing processes and so may overestimate the number of low-metallicity stars with subsolar [r-process/Fe].
\item The low-metallicity stars in our simulations are, on average, formed at high redshift ($z=2-3$) and large galactocentric radius ($\langle R\rangle=10-20$~kpc).
\item The scatter in [r-process/Fe] at fixed metallicity is driven by old stars at large distances from the centre of the galaxy.
\end{itemize}
The motivation for studying chemical enrichment with simulations of galaxy formation is that the simulations include a wide variety of mixing processes not present in phenomenological chemical evolution models. In particular, our galaxy formation simulation includes physically motivated (but still subgrid on scales smaller than giant molecular clouds) treatments of stellar feedback that lead to stellar masses of galaxies comparable to those observed over a wide range of dark matter halo masses \citep{Hopkins2014FIRE}. This consistency is a consequence of galactic winds efficiently redistributing gas (and with it heavy elements) from the scale of the galaxy to much larger radii in the halo. Our simulations are also based on a formulation of SPH, called P-SPH, that resolves the historical problems of SPH in capturing hydrodynamic instabilities that are important for mixing, in particular the Kelvin-Helmholtz and Rayleigh-Taylor instabilities \citep{Hopkins2013PSPH}.
We suspect that the two most important mixing processes in our simulations for the purposes of stellar abundance patterns are galactic winds and subsequent re-accretion of previously enriched gas from the surrounding halo and galactic fountains redistributing mass throughout the galaxy. Instabilities within the differentially rotating galactic disc, galaxy mergers, and gas accretion can also contribute to the mixing of the ISM.
In addition, there is turbulent mixing within the galaxy itself, although this is not well-resolved in fully cosmological simulations like those we have carried out here. As we discuss below, our results differ from previous analyses of r-process enrichment (e.g., \citealt{Argast2004}), suggesting that these additional mixing processes are indeed important. Despite the enhanced mixing introduced by stellar feedback, there is an important sense in which our calculations \emph{underestimate} the mixing of heavy elements: metals in our simulation are stuck to gas particles and do not diffuse to neighbouring gas particles.
Material unbound during NS mergers is the only known astrophysical site with conditions that robustly produce r-process nuclei. In particular, current models of nucleosynthesis in core-collapse supernovae fail to produce the requisite conditions, although the uncertainties remain large \citep[e.g.][]{Qian1996,Thompson2001, Hudepohl2010}.
The primary objections to NS mergers as the source of r-process nuclides have been on nucleosynthetic grounds. First, observations of stars with a range of metallicities find that heavy r-process nuclei are also accompanied by elements with atomic masses of ~90-120 \citep{Sneden2008, Qian2007}. This does not occur in the very low electron fraction conditions typically considered in NS merger ejecta, which do not produce nuclei with atomic mass ~90-120. However, recent work has shown that accretion disc outflows produced from the disc left after a NS merger have higher electron fractions than the tidal tail unbound during the dynamical phase of the merger \citep[e.g.][]{Metzger2008, Just2014}. These two sources of nucleosynthesis during NS mergers can thus satisfy the observational requirements on the source of neutron-rich heavy elements.
The second nucleosynthetic objection to NS mergers as the origin of the r-process elements is that the rarity of NS mergers would lead to too much scatter in the star-to-star r-process abundances and the absence of low-metallicity stars with solar r-process abundances (e.g., \citealt{Argast2004}). Our calculations significantly alleviate this objection, although with some remaining caveats discussed below.
One of our primary conclusions is that NS mergers can produce [r-process/Fe] abundance ratios and scatter that appear reasonably consistent with observations for stars with $-2\lesssim \mathrm{[Fe/H]}\lesssim 0$ (see Fig.~\ref{fig:rpFe}). This suggests that NS mergers may indeed account for most of the r-process nuclei in the Universe. In addition, the scatter in [r-process/Fe] increases with decreasing [Fe/H] in our calculations, as is also observed. In many of our models, there is a decrease in the median stellar [r-process/Fe] ratio at $\mathrm{[Fe/H]}\lesssim-2$ (see Fig.~\ref{fig:models}). Even so, there are still many stars with high abundance ratios of r-process elements at all subsolar metallicities we can probe. In particular, for all of our models the 84th percentile in the [r-process/Fe] distribution is supersolar at all metallicities. This is in contrast to the simplified chemical evolution models of \citet{Argast2004}, which find a very sharp drop in [r-process/Fe] at $\mathrm{[Fe/H]}\approx-2$ and no stars at $\mathrm{[Fe/H]}\lesssim-3$ that are enriched with r-process elements.
A more quantitative comparison to observations will require understanding the observational selection effects that enter into existing samples of r-process element abundance measurements. For example, we find that stars of metallicity $-2<\mathrm{[Fe/H]}<0$ show a clear trend in increasing galactocentric radius with decreasing [r-process/Fe] abundance (see Fig.~\ref{fig:zradius}), so that any kinematic selection effects for halo stars would bias the inferred median [r-process/Fe] ratios. In the future, quantifying these effects using a direct comparison between our simulations and observations is likely to be a fruitful way of testing the NS merger origin of r-process nuclei. A specific test of our models would be additional observations of low-metallicity stars ($\mathrm{[Fe/H]}\lesssim-2$). All of our calculations show that there should be some low-metallicity stars with $[\rp/\mathrm{Fe}]<-1$ (see e.g.\ Fig.~\ref{fig:rpFe}). If such stars are in fact completely lacking observationally, this would rule out the models described in this paper and point to either a different origin for r-process elements in the lowest metallicity stars or to physical processes not included in our simulation, such as additional turbulent diffusion. Turbulent diffusion results in fewer low-metallicity stars and a reduction of the scatter in both [Mg/Fe] and [r-process/Fe].
The primary uncertainty in our conclusions is that our results for the stellar r-process abundance patterns are not fully converged, particularly at low metallicity, i.e.\ at $\mathrm{[Fe/H]}\lesssim-2$. This is not surprising and is a consequence of the rarity of NS mergers as a source of heavy elements. By contrast, the abundance ratios of Mg to Fe are numerically converged. At our highest resolution, the gas particle mass is $7 \times 10^3 M_\odot$ so that the total mass of gas (within the SPH kernel) that is enriched per `merger' is on average about $4 \times 10^5$~M$_\odot$. This is significantly larger than the mass of the ISM swept-up by the ejecta created during a NS merger (see Sec.~\ref{sec:models}). It is not at all clear, however, that the latter is a reasonable model for the mass of the ISM that the r-process material is mixed into, given the many additional sources of mixing in the ISM \citep[e.g.][]{Yang2012} and the fact that mixing on scales smaller than the scale height of the galactic disk is not well resolved in our simulations. Thus, it may in fact be the case that our simulations describe a reasonable `initial' mass of the ISM enriched by r-process nuclei. Ultimately, however, it will be necessary to carry out even higher-resolution simulations, particularly for understanding the r-process abundances at low metallicity. Moreover, either grid-based calculations or SPH simulations with explicit metal diffusion at the scale of the kernel would be useful for determining the extent to which small-scale mixing modifies the convergence of the abundances at low metallicity. In the absence of such mixing the results cannot converge at the lowest metallicity in an SPH simulation in which metals are locked into gas particles. In addition, high resolution simulations of individual neutron star merger remnants mixing into a turbulent ISM would also be valuable. Larger turbulent velocities \emph{decrease} the mass of the ISM swept up prior to the NS merger remnant reaching pressure equilibrium with its surroundings (Eq.~\ref{eqn:momentum}) but \emph{increase} the resulting turbulent mixing of the r-process enriched material with the surrounding ISM.
We find that the variations in [r-process/Fe] abundances introduced by considering a range of plausible NS merger rates and delay times is comparable to or larger than the changes introduced by numerical resolution. Specifically, if the delay time for a significant fraction of NS mergers is $\lesssim 10^7$ years or if the NS merger rate is at the higher end of the allowed values, the decrease in median [r-process/Fe] is relatively modest until [Fe/H] $\lesssim -2.5$ in our highest resolution simulation (see Fig.~\ref{fig:models_z2p4}). Moreover, these constraints need only apply at $z \gtrsim 3$ when most of the low-metallicity stars form (see Fig.~\ref{fig:zradius}).
A second uncertainty in our analysis is that we have not implemented NS kicks, which may also contribute to redistributing r-process nuclei relative to heavy elements produced in supernovae. In this respect, our calculations may also \emph{underestimate} the mixing of r-process nuclei. Based on the modest offsets of most short-duration GRBs relative to their host galaxies \citep{Fong2013} we suspect that NS kicks are not likely to be important for the majority of r-process nuclei produced in NS mergers. However, they may be important for enriching low metallicity halo gas. This clearly needs to be explored in detail in future work.
In summary, we have explored a range of values for the NS merger rate per unit stellar mass and the delay time for the first NS mergers to take place, all reasonably consistent with current observational constraints. Our results for the r-process abundance ratios are broadly consistent with available observations. The abundance ratios of r-process elements are not fully converged with the resolution of the simulation, especially at low metallicity. However, variations with resolution are of similar magnitude as variations between different allowed models for the NS merger rate. Considering this as well as the additional uncertainties discussed above, we conclude that NS mergers could well be the source of the majority of the r-process elements in nature.
\section*{Acknowledgements}
We would like to thank Dan Kasen, Patrick Fitzpatrick, and Brian Metzger for useful conversations and the anonymous referee for helpful comments. This work was supported in part by NASA grant NNX10AJ96G, NSF grant AST-1206097, the David and Lucile Packard Foundation, and by a Simons Investigator Award from the Simons Foundation to EQ. DK is supported in part by Hellman Fellowship at UCSD and NSF grant AST-1412153. CAFG is supported by NASA through Einstein Postdoctoral Fellowship Award number PF3-140106 and by NSF through grant number AST-1412836. The simulations here used computational resources granted by the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number OCI-1053575; specifically allocations TG-AST120025 (PI Kere\v{s}), TG-AST130039 (PI Hopkins).
\bibliographystyle{mn2e}
|
\section{Motivation}
Recent simulations of galaxy mergers are able to produce remnants that
contain two or more supermassive black holes
\citep[BHs;][]{hof07,ama10,kul12}. When such BHs accrete, they can
appear as two or more active galactic nuclei (AGNs) on sub-galactic
scales \citep{van12,ble13}. Systematic surveys for such multiple AGNs
can impose observational constraints on AGN activation and tidally
enhanced star formation \citep[e.g.,][]{liu12,kos12}. Such surveys
can also constrain the BH merger rate, a key quantity for predicting
the signals expected for pulsar timing arrays and gravity-wave
detectors \citep{hob10,dot12,sha13}. These topics are of fundamental
importance in astrophysics, so all reports of multiple AGN candidates
warrant close vetting
\citep[e.g.,][]{wan09,liu10,smi10,fu11a,fu12,com12,kos12}.
This Letter focuses on one such case, SDSS\,J150243.09+111557.3
(J1502+1115 hereafter), originally identified as a dual AGN candidate
at a redshift of $z = 0.39$ with double-peaked profiles of [OIII]
\citep{smi10} and later shown to be a merging system with an
unobscured, primary AGN, J1502P, offset by 1.4\arcsec\, (7.4 kpc)
\footnote{$H_0$ = 70 km s$^{-1}$ Mpc$^{-1}$, $\Omega_M$ = 0.3,
$\Omega_\Lambda$ = 0.7} from a dust-obscured, secondary AGN J1502S
\citep[][F11 hereafter]{fu11b}. \citet[][D14 hereafter]{dea14}
recently advanced J1502S as {\em itself\/} hosting two AGNs and, thus,
two supermassive BHs. Adding in the AGN J1502P then suggests that the
merging system J1502+1115 hosts three supermassive BHs on sub-galactic
scales.
Specifically, D14 reported evidence from the European VLBI Network
(EVN) that J1502S exhibits two flat-spectrum radio sources,
J1502SE/SW, offset by 26 mas (140 pc), with each source being
energized by its own supermassive BH. D14 reached this intriguing
interpretation of J1502SE/SW as a close binary BH after ruling out a
double-hotspot scenario, wherein both hotspots are energized by a
single, central BH in one configuration that defines the well-studied
Compact Symmetric Objects \citep[CSOs;][]{pec00}. When observed with
sufficient sensitivity and resolution, an object with double hotspots
should have an edge-brightened structure. This Letter reports
evidence from the Very Long Baseline Array \citep[VLBA;][]{nap94} for
just such structure in an image of J1502S with higher sensitivity and
resolution than the EVN images. \S~\ref{imaging} presents the new
VLBA imaging and \S~\ref{implications} explores its implications. A
summary and conclusions appear in \S~\ref{summary}.
\section{VLBA Imaging}\label{imaging}
J1502+1115 was observed during a 6-hour session with 9 VLBA antennas
on 2013 June 29 (UT) under proposal BW102 = 13A-241. The tenth VLBA
antenna, at Fort Davis, TX, was unavailable due to equipment
malfunction. J1504+1029 was used as a phase and relative amplitude
calibrator (Table 1). The switching time between it and J1502+1115
was 180~s, with about a third of the session devoted to J1504+1029.
About once per hour, OQ\,208 and J1507+1236 were observed. OQ\,208 is
a compact and slowly-varying source \citep[][and references
therein]{wu13} observed to check the absolute flux density
calibration. J1507+1236 (Table 1) is a VLBA calibrator with a well
known position that was observed to check the quality of the phase
referencing.
\begin{deluxetable}{llccc}
\tabletypesize{\scriptsize}
\tablecolumns{5}
\tablewidth{0pc}
\tablecaption{VLBA Astrometry and Photometry at 5 GHz}\label{tab1}
\tablehead{
\colhead{Source} & \colhead{Peak} & \colhead{Position} &
\colhead{Switching} & \colhead{Integrated}\\
\colhead{ } & \colhead{R.A./Decl.} & \colhead{Error} &
\colhead{Angle} & \colhead{Flux Density\tablenotemark{a}}\\
\colhead{ } & \colhead{(J2000)} & \colhead{(mas)} &
\colhead{(\arcdeg)} & \colhead{($\mu$Jy)}\\
\colhead{(1)}&\colhead{(2)}&\colhead{(3)}&\colhead{(4)}&\colhead{(5)}}
\startdata
J1504+1029\tablenotemark{b} &
15 04 24.979782 & 0.1 & \nodata & \nodata \\
& 10 29 39.19840 & 0.1 & & \\
J1507+1236\tablenotemark{c,d} &
15 07 21.758063 & 0.3 & 2.23 & \nodata \\
& 12 36 29.07573 & 0.4 & & \\
J1502SE\tablenotemark{c} &
15 02 43.180261 & 0.2 & 0.87 & 891$\pm$81 \\
& 11 15 57.06831 & 0.3 & & \\
J1502SW\tablenotemark{c} &
15 02 43.178473 & 0.2 & 0.87 & 1010$\pm$92 \\
& 11 15 57.06508 & 0.3 & & \\
J1502P\tablenotemark{c} &
15 02 43.088667 & 0.2 & 0.87 & 254$\pm$45 \\
& 11 15 57.40016 & 0.3 & & \\
\enddata
\tablecomments{Col.~(1): Source. Col.~(2): Position at peak. Units
of right ascension are hours, minutes, and seconds, and units of
declination are degrees, arcminutes, and arcseconds. Col.~(3):
Position error. Col.~(4): Switching angle to the phase calibrator.
Col.~(5): Integrated flux density. Error is the quadratic sum of
(i) the image rms times the square root of the number of synthesized
beam areas integrated over, (ii) a 3\% uncertainty due to
phase-calibration errors and (iii) a 5\% uncertainty in the
amplitide scale.}
\tablenotetext{a}{Measured from the images made from self-calibrated
data.}
\tablenotetext{b}{Phase calibrator. The tabulated position and errors
are assumed and adopted from the Petrov catalog rfc\_2012b. The
latest Petrov catalog is at at http://astrogeo.org/petrov/.}
\tablenotetext{c}{Positions are measured from the images made from
phase-referenced data that have not been self-calibrated. The
conservative error estimates include measurement errors, systematic
phase-referencing errors and the errors in the position of the phase
reference calibrator.}
\tablenotetext{d}{Phase-referencing check. The astrometric catalog
position from Petrov rfc\_2014a is R.A.=15 07 21.758075 ($\pm$ 0.13
mas) and Decl. = 12 36 29.07581 ($\pm$ 0.21 mas). It differs from
our measured position by 0.18 mas in R.A. and 0.08 mas in Decl.}
\end{deluxetable}
A total of 256~MHz per circular polarization centered on $\nu$ =
4.980~GHz (5~GHz hereafter) were recorded using the new RDBE/MARK5C
wide-band system at 2 Gbps. Every 1~s the VLBA DiFX correlator
\citep{del11} produced 512 contiguous 0.5-MHz channels. About 3.4
hours were accrued on J1502+1115, during which the VLA position for
J1502S (F11) was used for pointing and correlation. Correlation
parameters were chosen to ensure distortion-free imaging at the
position of J1502P, offset by 1.4\arcsec\, from J1502S (F11). VLBA
system temperatures and gains were recorded for amplitude calibration.
Observations of OQ\,208 on 2013 July 1 UT were retrieved from the
archive of the Jansky Very Large Array (VLA) \citep[VLA;][]{per11}.
Release 4.1 of the Common Astronomy Software Applications (CASA)
package \citep{mcm07} was employed to calibrate and edit the data in
an automated fashion
\footnote{science.nrao.edu/facilities/vla/data-processing/pipeline}.
The final edits and imaging were done in the Astronomical Image
Processing System \citep[AIPS;][]{gre03}. To best match the VLBA
data, the final VLA image of OQ\,208 utilized a total bandwidth of
128~MHz per circular polarization centered at $\nu$ = 5~GHz. Direct
comparison with the primary flux calibrator 3C\,286 yielded a VLA flux
density of $2.33\pm0.07$~Jy, with the error dominated by an estimated
3\% uncertainty in the amplitude scale \citep{per13}.
AIPS was used for VLBA calibration and imaging. The calibration
strategies documented in Appendix C of the AIPS Cookbook were
followed, including doing a bandpass calibration and making
corrections for the ionosphere and for updated Earth orientation
parameters. The phase calibrator J1504+1029 was imaged using standard
methods based on multiple iterations of self-calibration and imaging.
Its final image was then used to determine the phase and amplitude
corrections to apply to all other data. For OQ\,208, only the
amplitudes were of interest. For J1507+1236, J1502S, and J1502P, the
phases enabled phase-referenced imaging, and those images were used to
measure the positions in Table~1. After exploring various imaging
schemes, a robustness of 1 as implemented in AIPS was adopted.
The OQ\,208 data were self-calibrated and imaged, yielding an
integrated VLBA flux density of $1.78$~Jy, 31\% lower than the VLA
value. This finding corroborated a recent report of low calibrated
flux densities from another VLBA user group. The cause is under
investigation and seems to affect mainly the Polyphase Filter Bank
personality of the RDBE system when the bandpass is calibrated in what
is believed to be the proper manner. Meanwhile, for this paper, we
have scaled the VLBA amplitudes upward by a factor of 1.31 to align
the VLA and VLBA photometry for OQ\,208, thereby calibrating the VLBA
amplitude scale to an accuracy of about 5\%.
The J1507+1236 data were self-calibrated and imaged, after which the
peak flux density increased by 78\% while the integrated flux density
did not change significantly, as expected when the emission is
concentrated by the improved phases. Compared to 1502+1115, many
fewer observations were made of J1507+1236, and those observations
involved a switching angle about 2.5 times larger (Table~1). Thus the
quality of the J1502+1115 data is expected to be much higher. From
the self-calibrated data for J1504+1029 and J1507+1236, the integrated
flux densities are $990\pm50$~mJy and $208\pm10$~mJy, respectively,
with errors dominated by the 5\% uncertainty in the amplitude scale.
The J1502+1115 data contains information on J1502P and J1502S. D14
did not detect J1502P, whereas our phase-referenced image of it
detects one radio source. Our phase-referenced image of J1502S shows
that it consists of two radio sources, as reported by D14, with a
summed flux density of less than 2 mJy. This image was used to
self-calibrate the phases of the J1502+1115 data. Because J1502S is
so faint, the phase-only self-calibration was based on data coherently
averaged for 35~m over all frequencies and both polarizations. This
long coherent average was possible because the initial phase
referencing had already removed the short-term phase fluctuations.
This phase self-calibration was used to reduce residual systematic
calibration offsets. For J1502S, the improved calibration recovered
an additional $\sim$ 10\% in the peak flux densities and $\sim$ 5\% in
the integrated flux densities. Self-calibration at low
signal-to-noise (S/N) can adjust the flux density upwards excessively
by gathering noise, or downward by failing to calibrate the residual
phase fluctuations. To account for these effects, our experiments
with several coherent averages lead us to include an estimate of a 3\%
uncertainty in the error budget for the integrated flux densities
(term (ii) in Table~1).
Figure~1 shows the self-calibrated VLBA image of J1502S. The achieved
rms noise is in line with the estimated thermal noise. Figure~1
reveals that J1502S consists of two radio sources, labeled J1502SW and
J1502SE. This image was used for the photometry in Table~1. With
adequate S/N, structures as large as about 20 mas could be represented
in Figure~1.
\begin{figure}
\plotone{f1.ps}
\caption{VLBA image of Stokes $I\/$ emission from the dust-obscured
AGN J1502S at $\nu$ = 5 GHz after self calibration. The rms noise
is $\sigma$ = 21~$\mu$Jy~beam$^{-1}$ and the synthesized beam
dimensions at FWHM are 3.5 mas $\times$ 1.5 mas with an elongation
position angle PA = $-$4\arcdeg\, (boxed hatched ellipse). Allowed
contours are at $-$6, $-$4, $-$2, 2, 4, 6, 8, 10, ... and 30 times 1
$\sigma$. Negative contours are dashed and positive ones are solid.
The coordinate origin is at the VLA position for J1502S (F11).
Clear edge-brightening is observed, suggesting that the two sources
are hotspots energized by a single, centrally-located BH. The
western extension of J1502SE has a peak of 5.7
$\sigma$.}\label{fig1}
\end{figure}
J1502P was imaged twice from the J1502+1115 data, once in its
phase-referenced form and once in its self-calibrated form. The data
used for imaging retained the 0.5-MHz channels from the correlator to
prevent bandwidth smearing, and had its phase center shifted by
$-$1.329\arcsec\, in right ascension and 0.330\arcsec\, in
declination. J1502P was detected as a slightly resolved source in
both images, with the self-calibration providing a 9\% enhancement in
the peak flux density and no change in the integrated flux density.
The signal from J1502P did not enter into the self-calibration, so
this is an independent check of the quality of that calibration. An
elliptical Gaussian fit in the image plane was consistent with J1502P
being unresolved with major and minor axes less than 3.5 mas and 1.5
mas, respectively. Since J1502P is so point-like, no image is
presented.
\section{Implications}\label{implications}
\subsection{J1502SE/SW}\label{j1502s}
From EVN images at 1.7 and 5~GHz, D14 reported evidence that the
dust-obscured AGN exhibits two flat-spectrum radio sources,
J1502SE/SW, offset by 26 mas (140 pc), with each source being
energized by its own supermassive BH. While intriguing, this
interpretation of a close binary BH hole was adopted only after D14
had ruled out a more prosaic double-hotspot scenario. In a
double-hotspot scenario both hotspots are energized by a single,
central BH. Such a configuration occurs in a well-studied class of
radio-selected AGN, the CSOs \citep[][and references therein]{an12}.
CSOs have radio extents of less than 1 kpc; slow-moving emission on
both sides of the central engine; and weak, if any, radio variability
\citep[e.g.,][]{fas01}. When observed with sufficient sensitivity and
resolution, the overall structure of a CSO can appear edge-brightened,
that is, the emission is brightest at the CSO's extremities, called
hotspots, and becomes fainter closer to its central BH. The radio
sources J1502SE/SW were only slightly resolved in the EVN images,
making it difficult to look for edge-brightening. However, the VLBA
image in Figure 1 clearly shows the expected edge-brightened structure
for J1502SE/SW. For this reason, we conclude that the double-hotspot
scenario should be reconsidered as a viable interpretation for
J1502SE/SW.
When compared to CSOs in the compilation of \citet{an12}, D14 noted
that J1502SE/SW has several unusual properties, including an
atypically low spectral power of $P_{\rm{1.7~GHz}} = 7\times10^{23}$ W
Hz$^{-1}$ and an atypically flat spectral index of
$\alpha^{\rm{1.7~GHz}}_{\rm{5~GHz}} \sim -0.1\pm0.1$, especially given
its projected linear size. These properties provide important clues
as to the nature of J1502SW/SE. Below, we discuss these properties
within an alternate, and yet related, context that stems from two
findings for AGNs with radio powers similar to J1502SE/SW.
First, the radio source J1148+5924 has a spectral power of
$P_{\rm{1.4~GHz}} \sim 2\times10^{23}$ W Hz$^{-1}$, a factor of a few
below that of J1502S \citep{tay98}. J1148+5924 is dominated by radio
emission with an extent of less than 1 kpc, shows a slow apparent
separation speed of 0.3c on parsec scales, and displays slow radio
variability \citep{tay98,pec00,fas01}. Such traits are shared by the
powerful CSOs \citep{an12}, suggesting that J1148+5924 is a low-power
CSO. Moreover, J1502S shows evidence for an overall rotational
symmetry on scales 0.1-10 kpc (D14). J1148+5924 also shows evidence
for an overall rotational symmetry on similar scales, possibly caused
by interactions between the low-power radio outflow and the
circumnuclear gas and dust \citep{wro84,tay98,per01}. Redshifted HI
is seen in absorption against the two-sided structures in J1148+5924
\citep{pec98}, and this infalling gas could fuel and/or distort the
slow-moving radio outflow.
Second, the obscured AGN J1502S was originally [OIII]-selected
\citep{smi10} and has a luminosity of $L({\rm [OIII]}) = 3.6 \times
10^{42}$ erg s$^{-1}$. In their analysis of the radio properties of
obscured AGN that have $L({\rm [OIII]}) \sim 10^{42}$ erg s$^{-1}$,
\citet{lal10} find a high incidence of spectra that are flat or rising
between 1.4 and 8.4 GHz. Those authors speculate that such radio
spectra could be caused by free-free absorption in the ionized gas
traced in [OIII]. J1502S has a similar [OIII] luminosity. This hints
that free-free absorption could contribute to the flat spectra D14
measure for J1502SE/SW, although achieving flatness between 1.7-15.7
GHz could be a challenge for such an absorption model. Still, this
suggestion is testable with VLBA imaging spanning a wide frequency
range \citep[e.g.,][]{mar14}, which could also help reveal emission
from a centrally located origin of activity.
D14 cite a 10\% uncertainty in the amplitude scale of the EVN at 5
GHz. Factoring this into their reported photometry leads to flux
densities of 857$\pm$99 $\mu$Jy and 872$\pm$100 $\mu$Jy for J1502SE
and J1502SW, respectively, in 2011 April 12. Comparison with the VLBA
photometry (Table 1) obtained more than two years later implies no
time variability, another trait consistent with CSO-like behavior.
From the EVN images, the apparent separation between J1502SE and
J1502SW increases by less than 5 pc yr$^{-1}$. For comparson, for
J1148+5924 the apparent separation increases by 0.1 pc yr$^{-1}$. If
J1502SE/SW have a similarly-slow apparent separation, it would be
advantageous to monitor that separation with the VLBA's higher
sensitivity and higher resolution.
More broadly, establishing a double-hotspot scenario for the
dust-obscured AGN J1502S could have important implications for
feedback in obscured AGNs, whether those AGN are discovered in surveys
for X-ray or infrared continuum or for narrow optical emission lines
\citep[][and references therein]{lal10} or in increasingly deep VLBA
surveys \citep[][and references therein]{mid13,del14}. Conversely,
our study underscores the importance of culling double-hotspot sources
from VLBA surveys seeking candidate binary BHs on parsec scales
\citep[e.g.,][]{bur11,tin11} or strong gravitational lenses on mas
scales \citep{wil01}.
\subsection{J1502P}\label{j1502p}
From Table~1 the VLBA detection of J1502P at $\nu$ = 5 GHz has a
rest-frame brightness temperature of 4.8 million K, or more if truly
unresolved. Such levels are not achieved by even the most compact
startbursts \citep{con92}. Also, the VLBA image is too shallow to
detect even the most luminous radio supernovae beyong $z \sim 0.1$
\cite[e.g.,][]{mid13}. Thus the VLBA detection indicates that J1502P
must be AGN driven, confirming, independent of F11, that J1502P hosts
an AGN. At $\nu$ = 5 GHz, the VLA photometry (F11) localizes the
emission to a diameter of 300 mas (1.6 kpc), whereas the EVN
photometry (D14), obtained only about two months earlier, recovers
less than 20\% of the VLA signal. Both this fact and the overall
steep-spectrum nature of J1502P (F11) suggest that source-resolution
effects are at play, a motivation for deeper VLBA imaging.
\section{Summary and Conclusions}\label{summary}
We used the VLBA at 5~GHz to image J1502+1115, a merging system at $z
= 0.39$ containing a dust-obscured AGN, J1502S, offset by 1.4\arcsec\,
(7.4 kpc) from an unobscured AGN, J1502P (F11).
Regarding J1502S, D14 advocate it as hosting two AGNs and, thus, two
supermassive BHs, based on EVN images at 1.7 and 5~GHz showing two
slightly-resolved, flat-spectrum radio sources, J1502SE/SW, offset by
26 mas (140 pc). D14 reached their intriguing interpretation of a
140-pc binary BH after discounting a double-hotspot scenario, wherein
both hotspots are energized by a single, central BH, a configuration
occuring amongst radio-selected CSOs. When observed with sufficient
sensitivity and resolution, an object with double hotspots should have
an edge-brightened appearance. We find clear evidence for such
edge-brightening in our VLBA image of J1502S that has higher
sensitivity and resolution than the EVN images. We thus conclude that
the double-hotspot scenario should be reconsidered as a viable
interpretation for J1502SE/SW. We also suggest that free-free
absorption by J1502S's [OIII] emitting gas could help flatten the
radio spectra of J1502SE/SW. Future VLBA imaging can further test the
double-hotspot scenario for J1502SE/SW, as well as the potential role
of free-free absorption. A double-hotspot scenario could have
important consequences for feedback modes in obscured AGNs, a key
population for understanding evolutionary linkages between galaxies
and the BHs they host. Future VLBA imaging of samples of low-power
CSOs is also needed to investigate trends between spectral indices and
projected linear sizes, to help distinguish low-power CSOs from close
binary BHs. And because the detection of polarized intensity is
relatively rare among powerful CSOs \citep{hel07}, such VLBA imaging
should include polarimetry as a possible discriminant between
low-power CSOs and close binary BHs.
Regarding J1502P, as it is detected in our VLBA image, it must be AGN
driven. This confirms, independent of F11, that J1502P hosts an AGN.
The VLBA detection is faint, only 254 $\mu$Jy. Future, deeper VLBA
imaging is needed to characterize it further.
\acknowledgements We thank the anonymous referee and the commentator,
Roger Deane, for prompt and helpful feedback. This work made use of
the Swinburne University of Technology software correlator, developed
as part of the Australian Major National Research Facilities Programme
and operated under licence. We acknowledge using Ned Wright's
Cosmology Calculator \citep{wri06}. We are grateful to Drew Medlin
for providing only the calibrator scans from the VLA data archived on
2013 July 1.
{\it Facilities:} \facility{VLA}, \facility{VLBA}.
|
\section{Formalism and ingredients}
The effective Lagrangian method is an important theoretical tool in
describing the various processes around the resonance region. In
this section, we introduce the theoretical formalism and ingredients
to study the $\bar{p} p \to \phi \phi$ reaction by using the
effective Lagrangian method.
The basic tree level Feynman diagrams for the $\bar{p} p \to \phi
\phi$ reaction are depicted in Fig.~\ref{pbarpdiagram}. In addition
to the ``background" diagrams, such as the $t$-channel
[Fig.~\ref{pbarpdiagram} (b)] and $u$-channel
[Fig.~\ref{pbarpdiagram} (c)] $N^*(1535)$ resonance exchange which
have been considered in the previous calculation~\cite{Shi:2010un},
we include the $s$-channel diagram [Fig.~\ref{pbarpdiagram} (a)]
through either a scalar meson ($f_0$) or a tensor meson ($f_2$) in
our present calculation.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=0.58]{pbarpfmd.eps}
\caption{Feynman diagrams for $\bar{p} p \to \phi \phi$ reaction.
The contributions from $t$- and $u$-channel $N^*(1535)$ resonance
exchange, and $s$-channel $f_0$ or $f_2$ resonance are considered.}
\label{pbarpdiagram}
\end{center}
\end{figure}
The invariant scattering amplitudes that enter our model for the
calculation of the total and differential cross sections for the
reaction
\begin{eqnarray}
\bar{p}(p_1,s_1) p(p_2,s_2) \to \phi(p_3,\lambda_1)
\phi(p_4,\lambda_2)
\end{eqnarray}
are defined as
\begin{eqnarray}
-iT_i = \bar{\nu}(p_1,s_1)A_i^{\mu \nu}u(p_2,s_2)
\epsilon^*_{\mu}(p_3,\lambda_1) \epsilon^*_{\nu}(p_4,\lambda_2),
\label{ti}
\end{eqnarray}
where $\nu(p_1,s_1)$ and $u(p_2,s_2)$ are Dirac spinors for
anti-proton and proton, respectively, while
$\epsilon_{\mu}(p_3,\lambda_1)$ and $\epsilon_{\nu}(p_4,\lambda_2)$
are polarization vectors for the $\phi$ mesons. The subscript $i$
stands for the $s$-channel $f_0$ or $f_2$ process, the $t$- and
$u$-channel $N^*(1535)$ resonance exchange.
The explicit expressions for the reduced $A_{N^*(1535)}^{\mu \nu}$
can be found in Ref.~\cite{Shi:2010un}. Here, we only give details
about the $s$-channel $f_0$ and $f_2$ amplitudes, $A^{\mu
\nu}_{f_0}$ and $A^{\mu \nu}_{f_2}$, associated to the diagram of
Fig.~\ref{pbarpdiagram} (a). They are obtained from the following effective
interaction Lagrangian~\cite{Renner:1971mu,Kochelev:1999zf,Oh:2003aw}:
\begin{eqnarray}
{\mathcal L}_{f_0 \bar{p} p} &=& g_{f_0 \bar{p} p}
\bar{\Psi}_{\bar{p}} f_0 \Psi_{p} \,+{\rm h.c.}, \label{f0nn} \\
{\mathcal L}_{f_0 \phi \phi} &=& g_{f_0 \phi \phi} m_{\phi} \phi_{\mu} \phi^{\mu} f_0 \, ,\label{f0phiphi} \\
{\mathcal L}_{f_2 \bar{p} p} &=& -i \frac{g_{f_2 \bar{p} p}}{m_{N}} \bar{\Psi}_{\bar{p}}
(\gamma_{\mu} \partial_{\nu} + \gamma_{\nu} \partial_{\mu}) \Psi_{p} f^{\mu\nu}_2 \, + {\rm h.c.}, \label{f2nn} \\
{\mathcal L}_{f_2 \phi \phi} &=& g_{f_2 \phi \phi} m_{\phi}
\phi_{\mu} \phi_{\nu} f^{\mu \nu}_2 \, . \label{eq:eqknstar}
\end{eqnarray}
With the above Lagrangians, the reduced
$A_i^{\mu}$ amplitudes in Eq.~(\ref{ti}) can be easily obtained,
\begin{eqnarray}
A_{f_0}^{\mu \nu} &=& -g_{f_0 \bar{p}p} g_{f_0 \phi\phi} m_{\phi} G_{f_0}(q_s) g^{\mu \nu} \, f_s, \label{eq:af0} \\
A_{f_2}^{\mu \nu} &=& i \frac{g_{f_2\bar{p}p} g_{f_2 \phi \phi} m_{\phi}}{m_{N}}
\Big [ \gamma_{\rho} (p_1 - p_2)_{\sigma} + \gamma_{\sigma} (p_1-p_2)_{\rho} \Big ] \nonumber \\
&& \times G^{\rho\sigma \mu\nu}_{f_2}(q_s) \, f_s, \,
\label{eq:af2}
\end{eqnarray}
where the propagators for the scalar meson $f_0$ and the tensor mesor $f_2$
are, respectively,
\begin{eqnarray}
G_{f_0}(q_s) &=& \frac{i }{s-M^2_{f_0} + i M_{f_0}\Gamma_{f_0}}, \\
G^{\mu\nu\rho\sigma}_{f_2}(q_s) &=& \frac{i}{s-M^2_{f_2}+i
M_{f_2}\Gamma_{f_2}} P^{\mu\nu \rho\sigma} (q_s),
\end{eqnarray}
and
\begin{eqnarray}
P^{\mu \nu \rho\sigma}(q_s) &=& \frac{1}{2}(\bar{g}^{\mu\rho}\bar{g}^{\nu\sigma} +
\bar{g}^{\mu\sigma}\bar{g}^{\nu\rho}) - \frac{1}{3}\bar{g}^{\mu\nu}\bar{g}^{\rho\sigma} \, , \\
\bar{g}^{\mu\nu} &=& -g^{\mu \nu} + \frac{q^{\mu}_s q^{\nu}_s}{s},
\end{eqnarray}
with $q_s = p_1 + p_2$ the momentum of $f_0$ or $f_2$ and
$s=q^2_s$ the invariant mass square of the $\bar{p} p$ system.
As can be seen from Eqs.~(\ref{eq:af0}) and (\ref{eq:af2}), in the
tree-level approximation, only the products, $g_{f_0 \bar{p}p}
g_{f_0 \phi\phi}$ and $g_{f_2 \bar{p}p} g_{f_2 \phi\phi}$ enter the
invariant amplitudes. $M_{f_0}$ ($M_{f_2}$) and $\Gamma_{f_0}$
($\Gamma_{f_2}$) are the mass and the total decay width of the $f_0$
($f_2$) meson. We take them as free parameters and determine them by
fitting to the total cross section of the $\bar{p} p \to \phi \phi$
reaction~\cite{Evangelista:1998zg} using MINUIT.
In Eqs.~(\ref{eq:af0}) and (\ref{eq:af2}), we have also included
the relevant off shell form factors~\footnote{We take the following
form factor for $t$- and $u$-channel $N^*(1535)$ ($\equiv N^*$)
resonance exchange as in Ref.~\cite{Shi:2010un}:
\begin{eqnarray}
F_{N^*(1535)} &=& \frac{\Lambda^2_{N^*} -M^2_{N^*}}{\Lambda^2_{N^*}
- q^2_{N^*}}, \nonumber
\end{eqnarray}
with $q^2_{N^*}$ the 4-momentum of the exchanged $N^*(1535)$
resonance. In general, the cutoff parameter $\Lambda_{N^*}$ for
$N^*(1535)$ resonance should be at least a few hundred MeV larger
than the $N^*(1535)$ mass, and thus in the range of $2$ to $4$
GeV.} for $f_0$ and $f_2$ mesons. We adopt here the common scheme
used in many previous works,
\begin{eqnarray}
&& f_s =\frac{\Lambda^4_i}{\Lambda^4_i+(s - M_i^2)^2}, \quad i= f_0,
f_2 \, . \label{sff}
\end{eqnarray}
The cutoff parameters, $\Lambda_{f_0}$ and $\Lambda_{f_2}$, are
constrained between $0.6$ and $1.2$ GeV. This way, we can reduce the
number of free parameters.
\section{Numerical results and discussion}
The differential cross section for $\bar{p} p \to \phi \phi$
reaction at the center of mass ($\rm c.m.$) frame can be expressed as
\begin{equation}
{d\sigma \over d{\rm cos}\theta}={1\over 64\pi s}{
|\vec{p}_3^{\text{~c.m.}}| \over |\vec{p}_1^{\text{~c.m.}}|} \left (
{1\over 4}\sum_{s_1,s_2,\lambda_1,\lambda_2}|T|^2 \right ),
\label{eq:pipdcs}
\end{equation}
where $\theta$ denotes the angle of the outgoing $\phi$ meson
relative to the beam direction in the $\rm c.m.$ frame, while
$\vec{p}_1^{\rm{~c.m.}}$ and $\vec{p}_3^{\text{~c.m.}}$ are the
3-momentum of the initial $\bar{p}$ and final $\phi$ meson.
First, by including the contributions from the $s$-channel scalar
meson $f_0$~\footnote{In general, we should study the role of the
scalar meson and tensor meson together. However, because of the
limitation of the experimental measurements and scarcity of the
information about the relevant mesons, hence, we separately study
them in this work.} and $t$- and $u$-channel $N^*(1535)$ resonance
(corresponding to $T= T_{f_0} + T_{N^*(1535)}$), with fixed cutoff
parameters $\Lambda_{f_0}$ and $\Lambda_{N^*(1535)}$, we perform a
$\chi^2$ fit (Fit I) to the total cross section data for $\bar{p} p
\to \phi \phi$~\cite{Evangelista:1998zg}. There are a total of $20$
data points.
By constraining the value of the cutoff parameter $\Lambda_{f_0}$
between $0.6$ and $1.2$ GeV and $\Lambda_{N^*(1535)}$ around $3.0$ GeV
based on the results of Ref.~\cite{Shi:2010un}, we obtain a minimal
$\chi^2/\mathrm{d.o.f.} = 2.1$ with $\Lambda_{f_0} = 0.6$ GeV and
$\Lambda_{N^*(1535)} = 3.05$ GeV. The fitted parameters are:
$g_{f_0 \bar{p}p} g_{f_0 \phi \phi} = 0.45 \pm 0.08$, $M_{f_0} =
2174 \pm 3$ MeV, and $\Gamma_{f_0} = 167 \pm 27$ MeV.
Second, instead of a scalar meson, we study the case of a tensor
meson $f_2$ in the $s$-channel and $t$- and $u$-channel $N^*(1535)$
resonance (corresponding to $T= T_{f_2} + T_{N^*(1535)}$), and we
perform a second $\chi^2$ fit (Fit II). In this case, we get a
minimal $\chi^2/\mathrm{d.o.f.} = 1.4$ with $\Lambda_{f_2} = 0.65$
GeV and $\Lambda_{N^*(1535)} = 3.05$ GeV. The fitted parameters are:
$g_{f_2 \bar{p}p} g_{f_2 \phi \phi} = -0.12 \pm 0.02$, $M_{f_2} =
2192 \pm 4$ MeV, and $\Gamma_{f_2} = 177 \pm 30$ MeV.
Based on the value of the $\chi^2/\mathrm{d.o.f.}$, Fit II is
preferred to Fit I. It seems to indicate that the $\bar{p} p \to
\phi \phi$ reaction is dominated by the exchange of a strange tensor
meson with quantum number $J^{PC} = 2^{++}$ in the $s$-channel, in
agreement with the study of Ref.~\cite{Evangelista:1998zg}. In
addition, a partial-wave analysis of the $\pi^- p \to \phi \phi n$
reaction shows that the $\phi \phi$ system is dominant by two
$J^{PC} = 2^{++}$ states~\cite{Etkin:1982bw}, one an $S$ wave and
the other a $D$ wave. The mass of the $S$ wave state is $M = 2160
\pm 50$ MeV, with a decay width $\Gamma = 310 \pm 70$ MeV. The mass
is in agreement with our fitted result for the tensor meson.
Next, we show the corresponding fitted results for the total cross
sections in Fig.~\ref{pbarptcs}, in comparison with the experimental
data from Ref.~\cite{Evangelista:1998zg}. In Fig.~\ref{pbarptcs},
the dashed curve stands for the contributions from the $t$- and
$u$-channel $N^*(1535)$ resonance, and the dash-dotted and dotted
lines stand for the contributions from the $s$-channel scalar meson
$f_0$ and tensor meson $f_2$, respectively, while the total results
of Fit I and Fit II are shown by dash-dot-dotted and solid curves.
From Fig.~\ref{pbarptcs}, one can see that the experimental total
cross section can be described fairly well by including the
contributions from both the $N^*(1535)$ resonance and the scalar
meson $f_0$ or tensor meson $f_2$. The contributions from
$N^*(1535)$ resonance dominates above $W = 2.25$ GeV, while the bump
structure around $W = 2.2$ GeV can be well reproduced by considering
the contributions from the strange mesons $f_0$ and $f_2$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=0.45]{pbarptcs.eps}
\caption{Total cross sections for $\bar{p} p \to \phi \phi$
reaction. The experimental data are taken from
Ref.~\cite{Evangelista:1998zg}. The curves are the contributions
from $s$-channel $f_0$ (dash-dotted) and $f_2$ (dotted), $t$- and
$u$-channel $N^*(1535)$ resonance (dashed), and the total results of
Fit I (dash-dot-dotted) and Fit II (solid).} \label{pbarptcs}
\end{center}
\end{figure}
With the above fitted parameters, the corresponding differential
cross sections for $\bar{p} p \to \phi \phi$ reaction at the energy
around the fitted masses of $f_0$ and $f_2$, $W = 2.15$ GeV, $W =
2.20$ GeV, and $W = 2.25$ GeV, are shown in Fig.~\ref{pbarpdcs}(a),
Fig.~\ref{pbarpdcs}(b), and Fig.~\ref{pbarpdcs}(c), respectively.
From Fig.~\ref{pbarpdcs}, we see that the shapes of the angular
distributions are similar, mainly because both the scalar meson and
the tensor meson decay to $\phi \phi$ in the $s$-wave. But, there
are still a little bit difference in the two cases, especially for
the energies of $W = 2.20$ GeV and $W = 2.25$ GeV, because the
production of a scalar meson $f_0$ from $\bar{p} p$ is in $s$-wave,
while the $\bar{p}p$ to the tensor meson $f_2$ is in the $D$-wave.
These predictions can be checked by future experiments.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=0.42]{dcs215.eps}
\includegraphics[scale=0.42]{dcs220.eps}
\includegraphics[scale=0.42]{dcs225.eps}
\caption{Differential cross sections for $\bar{p} p \to \phi \phi$
reaction. The curves are the contributions from $s$-channel scalar
meson $f_0$ (dash-dotted) and tensor meson $f_2$ (dotted), $t$- and
$u$-channel $N^*(1535)$ resonance (dashed), and the total results of
Fit I (dash-dot-dotted) and Fit II (solid).} \label{pbarpdcs}
\end{center}
\end{figure}
\section{Summary}
In this paper, we have phenomenologically reanalyzed the $\bar{p} p
\to \phi \phi$ reaction within an effective Lagrangian approach and
the isobar model. In addition to the ``background" contributions
from $t$- and $u$-channel $N^*(1535)$ resonance, we studied the role
of scalar meson ($f_0$) and tensor meson ($f_2$) in the $s$-channel.
Unfortunately, the information about the $f_0$ and $f_2$ meson with
mass around $2.2$ GeV is scarce~\cite{pdg2012}. Thus, in the present
work, we have taken the masses, the total decay widths, and the
coupling constants, $g_{f_0\bar{p}p}g_{f_0 \phi \phi}$ and
$g_{f_2\bar{p}p}g_{f_2\phi\phi}$ as free parameters, and we fitted
them to the experimental data on the $\bar{p} p \to \phi \phi$
reaction in Ref.~\cite{Evangelista:1998zg}. The fitted results are:
$M_{f_0} = 2174 \pm 3$ MeV, $\Gamma_{f_0} = 167 \pm 27$ MeV,
$M_{f_2} = 2192 \pm 4$ MeV, and $\Gamma_{f_2} = 177 \pm 30$ MeV. The
fitted results are shown that the $\bar{p} p \to \phi \phi$ reaction
is dominated by the exchange of a strange tensor meson with quantum
number $J^{PC} = 2^{++}$ in the $s$-channel, which is in agreement
with the previous analysis~\cite{Evangelista:1998zg,Etkin:1982bw}.
In this respect, we have shown how the experimental measurements for
the $\bar{p} p \to \phi \phi$ reaction could lead to valuable
information on scalar and tensor mesons with masses around $2.2$
GeV.
Finally, we would like to stress that due to the important role
played by the resonant contribution in the $\bar{p} p \to \phi \phi$
reaction, the bump structure around $W = 2.2$ GeV in the total cross
section can be well reproduced, and more accurate data on this
reaction can be used to improve our knowledge on the strange mesons
$f_0$ and $f_2$, which is at present poorly known. This work
constitutes a first step in this direction.
\section*{Acknowledgments}
We would like to thank Jun Shi and Xu Cao for useful discussions.
This work is partly supported by the Ministry of Science and
Technology of China (2014CB845406), the National Natural Science
Foundation of China under grants: 11105126, 11375024 and 11175220.
We acknowledge the one Hundred Person Project of Chinese Academy of
Science (Y101020BR0).
|
\section{Introduction}
The control of quantum mechanical systems has offered further motivation for the study of control systems on Lie groups, and in particular on $SU(n)$ and its Lie subgroups, as the evolution of a closed quantum system can be often modeled as a right invariant system varying on such Lie groups (see, e.g., \cite{IAAATACSI}, \cite{Mikobook} and references therein). Among these models, systems on $SU(2)$ arguably represent the simplest non trivial case, still a very rich one from a mathematical point of view. These {\it two-level quantum systems} are of fundamental interests in quantum physics and in quantum information, since they are the basic building block in the circuit based implementation of quantum information processing (see, e.g., \cite{NielsenandChuang}). A natural requirement in these implementations is to perform quantum operations (evolutions) in minimum time, both to shorten the overall time of computation and to avoid the effects of the interaction with the environment (de-coherence). For these reasons these systems have been studied in many aspects and their (time) optimal control has been the subject of many papers (see, e.g., \cite{Ugo1}, \cite{Ugo2}, \cite{newpaperonTimeOptimal}, \cite{newpaperbis}, \cite{QSL}, and references therein.). Here we add to this literature providing an explicit description of all optimal trajectories. This is done for a system with two orthogonal controls $u_x$ and $u_y$ (cf. model (\ref{basicmodel}) below) with have to satisfy a bound $u_x^2+u_y^2 \leq \gamma^2$ at every time, with positive $\gamma$ and $\gamma \leq 1$.
In particular, the model we consider is given by
\be{basicmodel}
\dot X=\sigma_z X + u_x \sigma_x X+u_y \sigma_y X, \qquad X(0)={\bf 1},
\ee
where $X \in SU(2)$ and $\sigma_{x,y,z}$ are the Pauli matrices, which form a basis of the Lie algebra $su(2)$. They are defined as
\be{Paulimat}
\sigma_x:=\frac{1}{2}\pmatrix{ 0 & i \cr i & 0}, \qquad
\sigma_y:=\frac{1}{2} \pmatrix{ 0 & -1 \cr 1 & 0},
\qquad \sigma_z:=\frac{1}{2}\pmatrix{ i & 0 \cr 0 & -i}.
\ee
The Lie algebra $su(2)$ is equipped with an inner product between matrices, $\langle \cdot, \cdot \rangle$, defined as $\langle A, B \rangle:=Tr(A B^\dagger)$, so that the associated norm is $\|A\|:=\sqrt{\langle A, A \rangle}$. With these definitions the norm of the Pauli matrices is $\frac{1}{\sqrt{2}}$.
We want to find, for every final condition $X_f$, the controls $u_x,u_y$, that steer the state of system (\ref{basicmodel}) from the identity to $X_f$ in minimum time, with the requirement that $u_x^2+u_y^2 \leq \gamma^2$, $\gamma^2 \leq 1$.
\begin{remark}{\label{primo}}
Requiring a small bound on the norm of the control as compared to the size of the drift in (\ref{basicmodel}) is quite natural in NMR experiments where the control is usually a perturbation.
\end{remark}
\begin{remark}{\label{primobis}}
The more general time optimal control problem for the system
\be{basicmodelgeneral}
\dot U=\pm \omega_0 \sigma_zU+v_x \sigma_x U+v_y \sigma_y U, \qquad U(0)={\bf 1},
\ee
with $\omega_0>0$ with $v_x^2+ v_y^2 \leq \omega_0^2 \gamma^2$ can be reduced to the problem for system (\ref{basicmodel}). Define
$X(t):=U(\frac{t}{\omega})$, and new controls $u_{x,y}(t):=\frac{1}{\omega_0}v_{x,y}\left(\frac{t}{\omega_0}\right)$ we have that once the minimum time problem for
\be{piuomenomodel}
\dot X=\pm \sigma_zX+u_x \sigma_x X+u_y \sigma_y X, \qquad U(0)={\bf 1},
\ee
is solved with controls $u_x$ and $u_y$ and minimum time $T$, and $u_x^2+u_y^2 \leq \gamma^2$ the original optimal control for (\ref{basicmodelgeneral}) is solved with $v_{x,y}(t)=\omega_0 u_{x,y}(\omega_0 t)$, in time $\frac{T}{\omega_0}$ to drive to the same final condition. The optimal control problem for system (\ref{piuomenomodel}) is the same as the one we have stated in the case $+$. In the case $-$ it can be reduced to it. Assume we have solved the minimum time problem for system (\ref{basicmodel}) for the final condition $X_f^{-1}$ and with controls $u_x$ and $u_y$ over an interval $[0,T]$. Then it is easily verified that the control $-u_x$, $-u_y$ over the same interval $[0,T]$ solves the problem of driving system (\ref{piuomenomodel}) with the $-$ from the identity to $X_f$, in minimum time.
\end{remark}
\vs
The paper is organized as follows. In section \ref{para}, we will select one of the methods to parametrize elements in $SU(2)$, and prove a simple property of the control system (\ref{basicmodel}) which will allow us to consider only two parameters rather than three when studying time optimal trajectories. In view of these facts, we will be able to perform the whole geometric analysis in the unit disk in the complex plane. We also recall how to apply the maximum principle of optimal control in this case and the form of the extremal controls and trajectories. In section \ref{Threecases} we solve the time optimal control problem for {\it diagonal operators}. As a limit of these trajectories, we identify a particular optimal trajectory which is a circle and plays a fundamental role for the whole analysis. All optimal trajectories leading to diagonal operators are outside this circle while all others are inside. Therefore we call this curve the {\it separatrix}.\footnote{Note this terminology is used with a slightly different meaning usually in mathematics, where a separatrix is a curve separating different behaviors of solutions of a differential equation. Here our curves are projections of solutions of differential equations obtained for different values of parameters rather than initial conditions.}
For the special case $\gamma=1$, the separatrix curve coincides with the trajectory corresponding to the SWAP operator. The optimal trajectories for points outside the separatrix are the same ones that lead to diagonal operators. The optimal trajectories for points inside the separatrix are described in section \ref{inside}. Here we give the general picture as a conjecture which is supported by theoretical results and simulations. In order to complete the proof though, we use the additional assumption $\gamma \geq \frac{1}{\sqrt{3}}$. In section \ref{Geo}, we provide a discussion of the results and show how these lead to a simple method to find the optimal control once the final condition is chosen.
In this section we also compare our results with other work on the control of systems on $SU(2)$ and two level quantum systems and in particular \cite{newpaperonTimeOptimal} and \cite{newpaperbis}.
\section{Parametrization of $SU(2)$ and general properties of the model}\label{para}
\subsection{Parametrization of the final conditions in the optimal control problem}
It is well known that the Lie group $SU(2)$ is diffeomorphic to the sphere $S^3 \subseteq \RR^4$ and it is Lie-homeomorphic to the Lie group of unit quaternions, $SH$, $x + y \vec i + c \vec j+ d \vec k$, with $x^2+y^2+c^2+d^2=1$, the homeomorphism being given by
\be{homeo}
SH \ni x + y \vec i + c \vec j+ d \vec k \Leftrightarrow \pmatrix{x+yi & -(c+id) \cr (c-id) & x-iy} \in SU(2). \ee
By writing $-(c+id)=e^{i\phi}M$ and $x+iy=e^{i\psi}\sqrt{1-M^2}$, with $0 \leq M \leq 1$, $\psi, \phi \in [0,2\pi)$, we can write any matrix $X_f\in SU(2)$ using the three parameters $\psi,\phi,$ and $M$, as
\be{parametriz}
X_f:=\pmatrix{e^{i\psi} \sqrt{1-M^2} & e^{i \phi} M \cr
- e^{-i\phi}M & e^{-i \psi} \sqrt{1-M^2}}.
\ee
We shall some times normalize the parameter $\psi$ and use the parameter $x_{\psi}$ instead, defined as $x_{\psi}:=\frac{\psi -\pi}{\pi}$, with $x_\psi \in [-1,1)$.
The parameter $\phi$ of the final condition $X_f$ in (\ref{parametriz}) does not affect the time optimal control problem, in the sense that matrices that differ only by the parameter $\phi$ can be reached in the same minimum time. This is a consequence of the following proposition.
\bp{eliminaz} The minimum time to reach $X_f \in SU(2)$, is the same as the minimum time to reach
$e^{\sigma_z \alpha} X_f e^{-\sigma_z \alpha}$, for any $\alpha \in \RR$.
\ep
\bpr
Let $u_x$ and $u_y$ optimal controls steering the state $X$ of (\ref{basicmodel}) from the identity to $X_f$, in time $T_{opt}$ and let $X^o:=X^o(t)$ the corresponding trajectory. Define for $j=x,y$ the constants $\beta_{jk}$ such that
\be{rotation}
e^{\sigma_z \alpha}\sigma_{j}e^{-\sigma_z \alpha}=\sum_{k=x,y}\beta_{j,k} \sigma_k.
\ee
Define new controls $v_{x}, v_{y}$, for $k=x,y$, as $v_{k}:=\sum_{j=x,y} \beta_{j,k} u_j$. Moreover notice that $v_x^2+v_y^2=u_x^2+u_y^2$ so that, if $u_x,u_y$ is an admissible control so is $v_x,v_y$. With the control $v_x,v_y$, the trajectory solution of (\ref{basicmodel}) is $U(t)=e^{\sigma_z \alpha} X^o(t) e^{-\sigma_z \alpha}.$ In fact, differentiating $U(t)$ and using (\ref{basicmodel}) for $X^o$ and (\ref{rotation}), we obtain
\[
\dot U:=e^{\sigma_z \alpha} \dot X e^{-\sigma_z \alpha}=\sigma_z U+(\sum_{j=x,y} u_{j}(\sum_{k=x,y} \beta_{j,k} \sigma_k)) U=
\]
\[\sigma_z U+(\sum_{k=x,y}(\sum_{j=x,y} \beta_{j,k}u_j) \sigma_k)U=
\sigma_z U+(\sum_{k=x,y}v_k \sigma_k)U.
\]
This shows that the optimal time to reach $e^{\sigma_z \alpha} X_f e^{-\sigma_z \alpha}$ is not greater than the one to reach $X_f$. By exchanging the roles of $X_f$ and $e^{\sigma_z \alpha} X_f e^{-\sigma_z \alpha}$, the opposite is seen to be true. Therefore the minimum time is the same in the two cases as stated.
\epr
\br{generalization}
The proof can be generalized with only formal modifications to more general systems on (Lie subgroups of) $SU(n)$, and more general systems of the form $\dot X=AX+\sum_{j=1}^m u_j B_j X$. We can replace the element of the form $e^{\sigma_z \alpha}$ with any element $K$ of (the Lie subgroup of) $SU(n)$, which commutes with $A$ and it such that $\texttt{span}\{ K B_1 K^\dagger,\ldots, K B_m K^\dagger \}=\texttt{span}\{ B_1,\ldots, B_m\}$.
\er
\vs
In view of Proposition \ref{eliminaz} the only element that is relevant to determine the minimum time to reach $X_f$ in (\ref{parametriz}) is the element (1,1) in the matrix $X_f$. This will be parametrized by phase $\psi$ (or $x_\psi$) and magnitude $M$ or, more often, by its real and imaginary parts, i.e., as a point $x+iy$ in the unit disk in the complex plane. To every (optimal) trajectory in $SU(2)$ there corresponds a curve starting from (1,0) in the unit disk. Points in the unit disk correspond to classes of matrices in $SU(2)$ which can reached in the same minimum time.
\vs
\subsection{The Pontryagin maximum principle and the expression of optimal candidates}
\vs
Consider the problem of driving the state $X$ of (\ref{basicmodel}) from the identity to a final condition $X_f$, with bound $u_x^2+u_y^2 \leq \gamma^2$, in minimum time. The {\it Pontryagin Maximum Principle} states that, if $u_x,u_y$ is optimal, and $X_o$ is the optimal trajectory, then there exists a nonzero matrix $\tilde M \in {su(2)}$, such that, for almost every $t$, $u_x(t),$ $u_y(t)$, are the values of $v_x$ and $v_y$, that maximize the Hamiltonian function
\be{PMPHam}
H(\tilde M,X_o,v_x,v_y):=\langle \tilde M, X_o^\dagger \sigma_z X_o\rangle+v_x \langle \tilde M, X_o^\dagger \sigma_x X_o\rangle+v_y \langle \tilde M, X_o^\dagger \sigma_y X_o\rangle.
\ee
Furthermore $H(\tilde M,X_o(t),u_x(t),u_y(t))$ is constant for almost
every $t$.
Define,
$b_{x,y,z}:= \langle \tilde M, X_o^\dagger \sigma_{x,y,z} X_o\rangle$. The maximization condition, implies that
\be{afterminimiz}
u_{x,y}=\gamma \frac{b_{x,y}}{\sqrt{b_x^2+ b_y^2}}
\ee unless $b_x$ and $b_y$ are both zero, in which case the corresponding arc is called {\it singular}.
Differentiating $b_{x,y,z}$ with respect to time, using (\ref{basicmodel}), and the standard commutation relations for the Pauli matrices\footnote{$[\sigma_x, \sigma_y]=\sigma_z,$ $[\sigma_y, \sigma_z]=\sigma_x,$ $[\sigma_z, \sigma_x]=\sigma_y.$}, we arrive at the following system of differential equations for $b_x,$ $b_y$ and $b_z$.
\be{bx}
\dot b_x=b_z u_y - b_y,
\ee
\be{by}
\dot b_y= b_x -b_z u_x,
\ee
\be{bz}
\dot b_z=b_y u_x-b_x u_y.
\ee
On a non singular arc, given the expression of the controls $u_x$ and
$u_y$ in (\ref{afterminimiz}) we have that $b_z$ is constant. This together with the fact that the Hamiltonian (\ref{PMPHam}), which takes the form $H=b_z+\gamma \sqrt{b_x^2+b_y^2}$, is also constant, implies that the controls $u_x$ and $u_y$ (for nonsingular extremals) can be written as (cf., the solutions of (\ref{bx}), (\ref{by}))
\be{controlli}
u_x=\gamma \sin(\omega t + \tilde \phi), \qquad u_y=-\gamma \cos(\omega t + \tilde \phi),
\ee
for some frequency $\omega \in \RR$ and phase $\tilde \phi \in \RR$. For singular arcs where $b_x \equiv b_y \equiv 0$, from (\ref{bz}) $b_z=const \not= 0$ which\footnote{If it was equal to zero it would imply $\tilde M=0$ which is excluded from the maximum principle.} therefore gives from (\ref{bx}), (\ref{by}), $u_x\equiv 0$, $u_y \equiv 0$. Therefore singular arcs starting from a point $X_1$ have the form $e^{\sigma_z t} X_1$, for $t\in [0,t_1]$ for some $t_1 >0$. We shall see in Theorem \ref{Diagsumma} and its proof that these arcs are never optimal.\footnote{General conditions to discard singular arcs are discussed in \cite{UgoKP} and the references therein.} Therefore in the optimal control problem we can restrict ourselves to nonsingular arcs.
\vs
Using the controls (\ref{controlli}) in (\ref{basicmodel}), the resulting differential equation {\it can be explicitly integrated} (see, e.g., \cite{CT} p.446). Direct verification shows that the solution is given by
\be{soluzexpli}
X(t,\omega,\tilde \phi):=\pmatrix{ e^{i \omega \tau}(\cos(a \tau)+ i \frac{b}{a} \sin(a \tau)) & e^{i (\omega \tau + \tilde \phi)} \frac{\gamma}{a} \sin(a \tau) \cr - e^{-i(\omega \tau + \tilde \phi)} \frac{\gamma}{a} \sin(a \tau) & e^{-i \omega \tau}(\cos(a \tau)-i \frac{b}{a} \sin(a \tau))},
\ee
for $\tau=\frac{t}{2}$, $b:=1-\omega$, $a:=\sqrt{\gamma^2 + b^2}$. For given $\omega$ and $\tilde \phi$, the time $T$ is the minimum time to reach if $X_f:=X(T,\omega,\tilde \phi)$ if there is no smaller $T_1$ and pair $\omega_1$ and $\tilde \phi_1$ such that $X_f:=X(T_1,\omega_1,\tilde \phi_1)$ .
\vs
In the expression (\ref{soluzexpli}), the phase of the element $(1,2)$ does not affect the (minimum) time to reach a given target, in the sense that we can always tune $\tilde \phi$ to give an arbitrary phase to the (1,2) element of the final condition, which provides an alternative way to prove Proposition \ref{eliminaz}.
\vs
{\bf Notation:} In the following we shall replace the notation $\tau$ with $t$, with the understanding that the new $`t'$ is half the $`t'$ we have mentioned so far.
\subsection{Properties of extremal curves}\label{prope}
Any candidate optimal is represented by a parametric curve in the complex plane, and in particular inside the unit disk, which starts from the point $(1,0)$ and represents the $(1,1)$ element of the trajectory of (\ref{basicmodel}). These curves can be parametrized by the frequency $\omega$ of the optimal control candidates while the phase does not play any role. They are explicitly given by (cf. (\ref{soluzexpli})
\be{curvex}
x(t):=x_{\omega}(t)=\cos(\omega t)\cos(at)-\frac{b}{a}\sin(\omega t)\sin(at),
\ee
\be{curvey}
y(t):=y_\omega(t)= \sin(\omega t)\cos(at)+\frac{b}{a}\cos(\omega t)\sin(at),
\ee
with $b:=1-\omega,$ $a=\sqrt{b^2+\gamma^2}.$
We also have (cf. (\ref{soluzexpli})) for the distance of the point from the origin,
\be{rquadro}
1-M^2(t):=r^2(t):=x^2(t)+y^2(t)=1-\frac{\gamma^2}{a^2}\sin^2(a t).
\ee
The phase $\psi(t)$ is given (cf. (\ref{soluzexpli})) for $0 \leq t \leq \frac{\pi}{2a}$ by
\be{fase1}
\psi(t)=\omega t + \arctan\left( \frac{b}{a} \tan(a t) \right),
\ee
and for $\frac{\pi}{2a} < t \leq \frac{\pi}{a}$,
\be{fase2}
\psi(t)=\omega t+ \pi + \arctan\left( \frac{b}{a} \tan(a t) \right).
\ee
\vs
In the following there will be some values of the frequency $\omega$ which play an important role. We define them at the outset. In particular we define $\omega^*:\frac{1+\gamma^2}{2}$, $\omega_c:=2 \omega^*=1+\gamma^2$. Correspondingly, we define $b^*:=1-\omega^*$, $b_c:=1-\omega_c$,
$a^*:\sqrt{\gamma^2+(b^*)^2}$ and $a_c:=\sqrt{\gamma^2+(b_c)^2}$.
We record few properties of the extremal trajectories.
\vs
{\bf Fact 1} From equation (\ref{rquadro}), we have:
\[
\frac{d r^2}{dt}= \frac{-2\sin(at)\cos(at)}{a},
\]
which implies, that $r(t)$ is decreasing for $t\in[0,\frac{\pi}{2a}]$, and it is increasing for $t\in [\frac{\pi}{2a},\frac{pi}{a}]$. At the time $t=\frac{\pi}{a}$ the trajectory reaches the boundary of the unit disk.
Moreover, since $\frac{d(r^2)}{dt}|_{t=0}=\frac{d^2(r^2)}{dt^2}|_{t=0}=\frac{d^3(r^2)}{dt^3}|_{t=0}=0$ and $\frac{d^4(r^2)}{dt^4}|_{t=0} =8a^2$, we have that given $\omega_1$ and $\omega_2$
and letting $a_{1,2}$ and $r_{1,2}(t)$ the corresponding value for the constant $a$ and $r(t)$,
if $a_1>a_2$, for $t$ in a neighborhood of $0$, we have $r_1(t)>r_2(t)$.
\vs
{\bf Fact 2} Calculating $\frac{d \psi}{d t}$ from (\ref{fase1}), (\ref{fase2}), we obtain
\be{dpsidtau}
\frac{d \psi}{d t}=\frac{-\gamma^2 \sin^2(a t)\omega + a^2}{a^2 \cos^2( a t)+b^2 \sin^2(a t)}.
\ee
Equation (\ref{dpsidtau}) implies that for $\omega \leq 1$ the phase is always increasing.
Moreover when $\omega>0$, we have:
\[
-\gamma^2 \sin^2(a \tau)\omega + a^2= \omega^2-(2+\gamma^2\sin^2(at))\omega+ (\gamma^2+1)
\geq \omega^2-(2-\gamma^2)\omega+(\gamma^2+1).
\]
Since the last polynomial is positive when $\omega>\omega_c:=1+\gamma^2$, we derive that the phase is increasing
for $\omega\leq 1$ and for $\omega\geq\omega_c$.
\vs
{\bf Fact 3}: Because of the {\it existence} of the optimal control, every point in the unit disk is reached by at least one curve and among those that reach the point at least one is optimal.
\vs
{\bf Fact 4}: The {\it singular} curve corresponds to the {\it boundary} of the unit disk. Therefore every point in the interior of the unit disk must be reached by an optimal trajectory which contains a nonsingular arc. We shall in fact see in Theorem \ref{Diagsumma} (under the assumption $\gamma \leq 1$) that even for the points on the boundary the optimal trajectories are nonsingular, and this implies that all the optimal trajectories do not contains singular arcs.
\vs
{\bf Fact 5}: (Principle of Optimality) If a curve reaching a point $P$ is optimal, then that curve is optimal for every point on that curve before $P$.
\vs
{\bf Fact 6}: When two curves intersect at a point $P$ they cannot be both optimal at the point $P$. In fact, if they reach $P$ at point at different times, then, obviously, the one that reaches at greater time is not optimal. If the reach $P$ at the same time, then we could possibly switch from one value of $\omega$ to the other in the control and still have an optimal trajectory. This contradicts the fact that all the nonsingular extremals have the form (\ref{controlli}) (cfr. Fact 4).
If a curve is optimal for every point before a point $P$ and not optimal after $P$ we say that a curve {\it looses optimality at $P$}.
\section{Optimal control problem for diagonal final conditions and the separatrix curve}\label{Threecases}
\subsection{Diagonal operators}\label{operatoridiagonali}
Assume the final condition $X_f:=\pmatrix{e^{i\psi_f} & 0 \cr 0 & e^{-i \psi_f}}$, that is, we want to drive in minimum time to a point on the boundary of the unit disk. According to formula (\ref{rquadro}) extremal trajectories reach the boundary of the unit disk at times $T=\frac{k\pi}{a}$. If $T$ is the final time in (\ref{soluzexpli}), we have the two equations
\be{equa1}
T=\frac{k \pi}{a}, \qquad k \geq 0,
\ee
\be{equa2}
\omega T+aT=\psi_f +2m\pi, \qquad m \in \ZZ,
\ee
which give respectively the condition on the norm of the off diagonal term and on the phase of the diagonal term.\footnote{In the condition (\ref{equa1}) we have used the fact that the time has to be nonnegative (in fact positive if $\psi \not=0$).} Plugging (\ref{equa1}) into (\ref{equa2}), we have
\be{equa3}
k \pi (1+ \frac{\omega}{a})=\psi_f + 2 m \pi.
\ee
A study of the function $f(\omega):=\frac{\omega}{a}$ for $\omega \in (- \infty, \infty)$ reveals that this function is bounded below by $-1$, so that, when $\psi_f \in (0,2 \pi)$, (\ref{equa3}) can only be verified for $m \geq 0$. Let use denote by $T_{k,m}$ the time $T$ which is given by equation (\ref{equa1}) with $k$ and verifying the constraint (\ref{equa2}). Notice that not all pairs $k>0,m\geq 0$ are feasible (the function $\frac{\omega}{a}$ is bounded). We shall show that no matter what $\psi_f \in (0,2\pi)$ is, the minimum of these times is $T_{1,0}$ which is feasible.\footnote{This means that there exists an $\omega$ satisfying (\ref{equa3}) with $k=1$ and $m=0$.}.
The proof can be achieved in two steps given by the following two lemmas. The result for the diagonal case is summarized in Theorem \ref{Diagsumma}. Proofs of Lemmas \ref{Lemma1} and \ref{Lemma2} are given in Appendix A.
\bl{Lemma1} For every $k>0$ and $m>0$,
\be{TKMTK0}
T_{k,m} \geq T_{k,0}.
\ee
\el
\bl{Lemma2} For every $k >0$,
\be{TK0T10}
T_{k,0} \geq T_{1,0}.
\ee
\el
\bt{Diagsumma}
Assume $\gamma \leq 1$. Then the minimum time to reach a diagonal operator $X_f :=\pmatrix{e^{i \psi_f} & 0 \cr
0 & e^{-i \psi_f}}$, $\psi_f \in (0, 2 \pi)$ is
\be{Tmin}
T_{min}=T_{1,0}(\psi_f):=\frac{\psi_f (2 \pi - \psi_f)}{\pi - \psi_f +\sqrt{\pi^2 + \gamma^2 \psi_f (2\pi - \psi_f)}},
\ee
which is obtained with the controls (\ref{controlli}), with $\tilde \phi$ arbitrary and $\omega$ given by\footnote{Recall that $x_\psi:=\frac{\psi -\pi}{\pi}$}
\be{omegaopt}
\omega= \frac{x_{\psi_f}}{1- x^2_{\psi_f}} (- x_{\psi_f}+ \sqrt{1+ \gamma^2(1-x^2_{\psi_f})}).
\ee
\et
\bpr
The theorem summarizes the previous two Lemmas. The expression of the optimal
frequency $\omega$ is obtained from (\ref{equa3}), (\ref{aa}), with $k=1$ and $m=0$.
To make sure that this time is optimal we need to compare it with the one obtained with the singular trajectory which is $T_{sing}(\psi_f)=\psi_f$. In fact we have $T_{min}< T_{sing}$. This follows from
\be{singula00}
\frac{T_{min}}{\psi_f} = \frac{2\pi - \psi_f}{\pi - \psi_f + \sqrt{\pi^2 + \gamma^2 \psi_f (2 \pi - \psi_f)}}< 1=\frac{T_{sing}}{\psi_f}.
\ee
\epr
A consequence of this theorem is also that no optimal trajectory can be contain a singular arc, because the singular arc can be followed in smaller time.
\vs
\subsection{The separatrix curve}
Reconsider formula (\ref{omegaopt}). There is a one to one correspondence between values of $x_{\psi_f} \in (-1,1)$ (alternatively values of $\psi_f \in (0,2\pi)$) and $\omega$. In fact $-\infty <\omega < \frac{1+\gamma^2}{2}:=\omega^*$ and $\lim_{\psi_f \rightarrow 0} \omega =-\infty$ and $\lim_{\psi_f \rightarrow 2 \pi} \omega =\frac{1+\gamma^2}{2}=\omega^*$.
Consider now the trajectory corresponding exactly to $\omega= \omega^*=\frac{1+\gamma^2}{2}$. In this case
$a=a^*=\omega=\omega^*$ and the parametric equations of (\ref{curvex}) and (\ref{curvey}) become
\be{curvexsepa}
x(t)=\frac{2}{1+\gamma^2} \cos(\omega^*t)-\frac{1-\gamma^2}{1+\gamma^2},
\ee
\be{curveysepa}
y(t)=\frac{2}{1+\gamma^2} \cos(\omega^*t)\sin(\omega^*t).
\ee
This represents a circle with center in
\be{centro}
P=\left(\frac{\gamma^2}{1+\gamma^2}, 0 \right),
\ee
and radius $\frac{1}{1+\gamma^2}$. We shall call this circle the `{\it separatrix}'. The following lemma justifies this name.
\bl{stannofuori}
All the optimal trajectories corresponding to diagonal operators (described in subsection \ref{operatoridiagonali}) intersect the separatrix curve only in the point $(1,0)$.
\el
The proof is in Appendix A.
Figures \ref{Traie12} and \ref{Traie1} give some plots of the trajectories outside the separatrix, leading to diagonal operators for the cases $\gamma=\frac{1}{2}$ and $\gamma=1$ respectively. The separatrix is the red circle in both cases. The cases $\omega=-3$, $\omega=0$ and $\omega=\frac{1}{2}$ are displayed explicitly for $\gamma=\frac{1}{2}$ and the same values of $\omega$'s and $\omega=\frac{8}{9}$ are displayed for $\gamma=1$. As $\omega \rightarrow \omega^*$, the trajectories tend to the separatrix.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{Traiettoriefuoregamma12}
\caption{Optimal trajectories to reach the boundary of the unit disk (in blue) for various values of $\omega$ and $\gamma=\frac{1}{2}$. The outermost trajectory is the one corresponding to $\omega=-3$, the next one (reaching the point (-1,0)) corresponds to $\omega=0$. The innermost trajectory is the one corresponding to $\omega=\frac{1}{2}$. The separatrix is the red circle centered at the point $(\frac{1}{5}, 0)$.}
\label{Traie12}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{Traiettoriefuoregamma1}
\caption{Optimal trajectories to reach the boundary of the unit disk (in blue) for various values of $\omega$ and $\gamma=1$. The outermost trajectory is the one corresponding to $\omega=-3$, the next one (reaching the point (-1,0)) corresponds to $\omega=0$. The next on is the one corresponding to $\omega=\frac{1}{2}$. The innermost trajectory is the one corresponding to $\omega=\frac{8}{9}$. The separatrix is the red circle centered at the point $(\frac{1}{2}, 0)$.}
\label{Traie1}
\end{figure}
The following proposition states two important properties of the optimal trajectories outside the separatrix.
\bp{Filling}
\begin{enumerate}
\item The trajectories corresponding to $\omega \in (-\infty, \omega^*)$ loose optimality after reaching the boundary of the unit disk.
\item Every point outside the separatrix is reached by an optimal trajectory (before reaching the boundary) corresponding to a single value of $\omega$, with $\omega \in (-\infty, \omega^*)$
\end{enumerate}
\ep
\bpr
To prove 1., recall from Fact 2 of subsection \ref{prope} that the phase $\psi$ is always increasing, since $\omega^*\leq 1$. This means that any of the trajectories corresponding to $\omega \in (-\infty, \omega^*)$ after hitting the boundary will necessarily intersect another trajectory corresponding to a larger value of (final) $\psi_f$ which is optimal. Therefore such a trajectory looses optimality at the boundary.
To prove 2., Consider a point $P$ outside the separatrix and assume by contradiction that none of the curves reaching the boundary and corresponding to $\omega \in (-\infty, \omega^*)$ contains such a point. In particular, denote by $C_{\psi_f}$ any such curve corresponding to the phase $\psi_f \in (0, 2\pi)$. By the existence of the optimal control for $P$ there exists an optimal trajectory ending in $P$, which we denote by $C_P$, defined in $[0,t_P]$, with $t_P <\frac{\pi}{a}$ (see (\ref{rquadro}). All the trajectories $C_{\psi_f}$ and $C_P$ never intersect (except for the point $(1,0)$).
Express the trajectory $C_{\psi_f}$ and $C_P$ as polar equations $r=r(\psi)$ with $\psi$ the (variable) phase. In particular we write $r=r_f( \psi)$ for $C_{\psi_f}$ and $r=r_P( \psi)$ for $C_P$. With this notation, we say that $C_{\psi_f}$ is {\it above} $C_P$ if $r_f( \psi)$ is greater than $r_P(\psi)$ for one (and therefore all since they cannot intersect) $\psi \not=0$ which are in the common domain of the function $r_f$ and $r_P$. Analogously we say that $C_{\psi_f}$ is {\it below} $C_P$ if $r_f(\psi)$ is smaller than $r_P(\psi)$. Consider the set $A_P$ ($B_P$) of all $\psi_f \in (0,2\pi)$ which are such that $C_{\psi_f}$ is above (below) $C_P$. It is important to notice that both $A_P$ and $B_P$ are not empty. $A_P$ is not empty because it definitely contains all $\psi_f$'s smaller than the phase of $P$ since the phase is always increasing from formula (\ref{dpsidtau}). $B_P$ is not empty because it is enough to take a curve $C_{\psi_f}$ sufficiently close to the separatrix to leave $P$ on the right. Moreover $A_P \bigcup B_P =(0,2\pi)$. By continuity (again using the fact that $C_{\psi_f}$ and $C_P$ never intersect) $A_P$ and $B_P$ are both open set. Since they are not empty this contradicts $A_P \bigcup B_P =(0,2\pi)$ because of the connectedness of $(0,2\pi)$. \epr
\subsubsection{SWAP operator}{\label{swap}}
The SWAP operator, is the operator that in quantum information theory corresponds to a logic operation $NOT$. It inverts the state of a two level quantum system. It is given in the computational basis by
\be{SWAP}
X_{SWAP}:=\pmatrix{0 & 1 \cr -1 & 0},
\ee
which corresponds to the origin of the unit disk. In formula (\ref{soluzexpli}), we need $a=\gamma$, $b=0$ and $\omega=1$ (resonance condition \cite{UgoKP}) and, minimum time $T_{min}(X_{SWAP})=\frac{\pi}{2\gamma}$. The optimal trajectory is
\be{OptSWAP}
X(t)=\pmatrix{e^{i t} \cos(\gamma t) & \sin (\gamma t) \cr - \sin(\gamma t) & e^{-i t} \cos(\gamma t)}.
\ee
Figure \ref{figurewithvariousgamma} displays the various trajectories for the values of $\gamma=\frac{2}{7},\frac{1}{2},\frac{2}{3},1$, until the trajectories self intersect and therefore are no longer optimal.\footnote{The full trajectories are closed in the case where $\gamma$ is a rational number.} The trajectory corresponding to $\gamma=1$ is a circle of radius $\frac{1}{2}$ centered at $(\frac{1}{2},0)$. Which coincides with the separatrix in this case.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{Figure1Variousgamma}
\caption{Optimal trajectories for the SWAP operator for various values of $\gamma$. Trajectories are no longer optimal after they self-intersect. The trajectories correspond to the following values of $\gamma$: $\gamma=\frac{2}{7}$, red; $\gamma=\frac{1}{2}$, blue; $\gamma=\frac{2}{3}$, green; $\gamma=1$, black.}
\label{figurewithvariousgamma}
\end{figure}
\section{The optimal control problem inside the separatrix}\label{inside}
From now on we denote by $\mathcal{S}$ the closed region inside the separatrix.
For points in $\mathcal{S}$, the frequency $\omega$ of the optimal control must be greater than or equal to $\omega^*=\frac{1+ \gamma^2}{2}$. In fact, as we have seen in the provious section,
the trajectories corresponding to $\omega < \omega^*$ do not intersect the separatrix before touching the boundary (Lemma \ref{stannofuori}) and, after touching the boundary, they are not optimal anymore (Proposition \ref{Filling}). Therefore, for all points in $\mathcal{S}$, the optimal trajectories are with omega, $\omega \geq \omega^*$.
In order to study the behavior of these trajectories with respect to the separatrix we consider a trajectory $(x_\omega(t), y_\omega(t))$ in (\ref{curvex}), (\ref{curvey}) and the function
\be{funzionedistanza}
\Delta_\omega(t):=\left(x_\omega(t)-\frac{\gamma^2}{\gamma^2+1}\right)^2+ y_\omega^2(t)-\frac{1}{(1+\gamma^2)^2},
\ee
which gives the difference between the square of the distance of the trajectory (as a function of $t$) from the center of the separatrix and the square of the radius of the separatrix. $\Delta_\omega(t)$ is identically zero for $\omega=\omega^*$, i.e., on the separatrix. Using (\ref{rquadro}) and (\ref{curvex}), we obtain
\be{iii}
\frac{\Delta_\omega(t)}{\gamma^2}=\frac{2}{1+\gamma^2}-\frac{1}{a^2}\sin^2(a t)-\frac{2}{1+\gamma^2}(\cos(\omega t) \cos(a t)- \frac{b}{a}\sin(\omega t) \sin(a t)).
\ee
\bl{lemmino}
Assume $\omega \in [\omega_*, 3 \omega^*)$. Then there exists an
$\epsilon=\epsilon_\omega > 0$ such that $(x_\omega(t), y_{\omega}(t))$ is in $\mathcal{S}$ for every $t\in [0, \epsilon)$. Assume $\omega > 3 \omega^*$. Then there exists an $\epsilon=\epsilon_\omega$ such that $(x_\omega(t), y_{\omega}(t))$ is outside $\mathcal{S}$ for every $t\in (0, \epsilon)$.
\el
\bpr
We calculate the derivatives of $\frac{\Delta_\omega(t)}{\gamma^2}$ at $t=0$. The first three derivatives give zero while the fourth one is greater than or equal to zero for $\omega \leq \omega^*$ and $\omega \geq \omega^*$ otherwise it is smaller than zero. The case $\omega \leq \omega^*$ corresponds to the trajectories of subsection \ref{operatoridiagonali} and the separatrix itself. The case $\omega > \omega^* $ also corresponds to trajectories that starts outside of the separatrix. Trajectories corresponding to $\omega \in (\omega_*, 3 \omega^*)$ start inside the separatrix.
\epr
\bc{exclusione}
Trajectories corresponding to $\omega > 3 \omega^*$ are not optimal.
\ec
\bpr
Using Proposition \ref{Filling}, these trajectories are not optimal since they intersect the optimal ones going to the boundary of the unit disk.
\epr
From the above two results, all points in $\mathcal{S}$ will have optimal trajectories corresponding to values of $\omega$ in the interval $[\omega^*, 3 \omega^*]$.
\vs
In the interval $[\omega^*,3 \omega^*]$ a particularly important role is played by the curve corresponding to $\omega_c:=2 \omega^*=\gamma^2+1$. This curve presents a cuspid point, i.e., a point where both $\dot x$ and $\dot y$ are zero. In particular having defined $a_c$ as the value of $a$ corresponding to $\omega_c$, i.e., $a_c:=\gamma \sqrt{1+\gamma^2}$, from (\ref{curvex}) and (\ref{curvey}), we obtain
\be{derivatx}
\dot x_{\omega_c}(t)=-\sin(\omega_c t) \cos(a_c t),
\ee
\be{derivaty}
\dot y_{\omega_c}(t)=\cos(\omega_c t) \cos(a_c t),
\ee
and both derivatives are zero when $t =\frac{\pi}{2 a_c}$. We shall call the trajectory corresponding to $\omega=\omega_c$ until the point corresponding to $t =\frac{\pi}{2 a_c}$, the {\it critical trajectory}. Its final point is
\be{finalcritic}
x_{\omega_c}(\frac{\pi}{2 a_c})=\frac{\gamma}{\sqrt{1+\gamma^2}}\sin\left(\pi \frac{\sqrt{1+\gamma^2}}{2\gamma}\right), \qquad y_{\omega_c}(\frac{\pi}{2 a_c})=- \frac{\gamma}{\sqrt{1+\gamma^2}}\cos\left(\pi \frac{\sqrt{1+\gamma^2}}{2\gamma}\right).
\ee
It is in particular a point on the circle centered at the origin with radius $\frac{\gamma}{\sqrt{1+\gamma^2}}$. Such a circle centered at the origin will play an important role in our proof below. We call it the {\it critical circle}.
\vs
The general picture of the optimal synthesis for points inside the separatrix is summarized in Theorem \ref{Insideseparatrix}.
\bt{Insideseparatrix}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{Insideseparagamma1}
\caption{Optimal trajectories inside the separatrix (in red) for $\gamma=1$. The critical
trajectory is in black, while the trajectories for $\omega=1.1 \omega^*$,
$\omega=1.2 \omega^*$, $\omega=1.5 \omega^*$, $\omega=1.8 \omega^*$, are in blue (starting closer to the separatrix when $\omega \rightarrow \omega^*=1$ and starting closer to the critical trajectory when $\omega \rightarrow \omega^*=2$).}
\label{Inssepa1}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{Insiderseparagamma12gooder}
\caption{Optimal trajectories inside the separatrix (in red) for $\gamma=\frac{1}{2}$. The critical
trajectory is in black, while the trajectories for $\omega=1.1 \omega^*$,
$\omega=1.2 \omega^*$, $\omega=1.5 \omega^*$, $\omega=1.8 \omega^*$, are in blue (starting closer to the separatrix when $\omega \rightarrow \omega^*=\frac{1+\gamma^2}{2}=
\frac{1+\frac{1}{4}}{2}$ and starting closer to the critical trajectory when $\omega \rightarrow 2\omega^*=1+\gamma^2=\frac{5}{4}$).}
\label{Inssepa2}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{ILLUSTRAZIONEPROOF2}
\caption{Geometric objects used in the proof of Theorem \ref{Insideseparatrix}. The critical trajectory and the critical circle are in red ($\gamma=\frac{1}{sqrt{2}}$). We deform the critical trajectory by adding a $\-epsilon \lambda$ to the phase. The corresponding deformed curve is in green. Trajectories for $\omega\in (2 \omega^*, 3 \omega^*]$ never reach the region $R$ in the figure.}
\label{Perlaprova}
\end{figure}
Assume $\gamma \in [\frac{1}{\sqrt{3}}, 1]$. The only optimal trajectories for points in $\mathcal{S}$ correspond to $\omega \in [\omega^*, 2 \omega^*]$. The trajectory corresponding to $\omega^*$ is the optimal for points of the separatrix. The trajectory corresponding to $\omega_c=2 \omega^*$ until the point (\ref{finalcritic}) is optimal for point on the critical trajectory. For any other points inside the separatrix, there exists a unique value of $\omega \in (\omega^*, 2\omega^*)$ and a corresponding optimal trajectory leading to that point.
\et
We believe this theorem holds for general values of $\gamma \leq 1$ but we were able to completely prove it only for $\gamma \in [\frac{1}{\sqrt{3}}, 1]$. The situation is described in Figure \ref{Inssepa1} for the case $\gamma=1$ and Figure \ref{Inssepa2} for the case $\gamma=\frac{1}{2}$, respectively. In both figures, the red circle is the separatrix and the black trajectory inside the separatrix is the critical trajectory. Optimal trajectories depicted in blue start from the point $(1,0)$ and end, loosing optimality, on the critical trajectory.
The proof of Theorem \ref{Insideseparatrix} is presented in Appendix B. We give here the main ideas and discuss where the assumption $\gamma\geq \frac{1}{\sqrt{3}}$ is used. We consider the critical curve, i.e., with $\omega=\omega_c$ and $a=a_c$. Starting from the point $(1,0)$ the distance from the origin decreases monotonically according to formula (\ref{rquadro}) and the last point is on the critical circle. Under the assumption $\gamma \geq \frac{1}{\sqrt{3}}$ the whole critical trajectory is in the first quadrant.\footnote{This renders the geometry of the problem easier to visualize. If this is not verified, the critical trajectory looks like a spiral winding around the origin more and more times as $\gamma \rightarrow 0$.} We introduce a parameter $\lambda:=\sin(a_c t)$, with $t \in [0, \frac{\pi}{2 a_c}]$, i.e., $\lambda \in [0,1]$, to parametrize the critical trajectory. Because of (\ref{rquadro}), $\lambda$ indicates the distance of the point on the critical trajectory from the origin, which goes from $1$
to $\sqrt{1-\frac{\gamma^2}{a_c^2}}$. For a given value of $\lambda$, i.e., for points on the same circle, we compare the phase of any trajectory (corresponding to a given value of $\omega$ and $a$)
with the phase for the critical trajectory. In doing that, we assume $0 \leq t \leq \frac{\pi}{2a}$ and we use formula (\ref{fase1}) for the phase. We find that the phase for the generic trajectory is always bigger than the one for the critical trajectory, Lemma \ref{LemmaB1}. This has several consequences: 1) Every trajectory corresponding to $\omega \in (\omega^*,3\omega^*]$, $\omega \not= \omega_c$ that intersects the critical trajectory has to do so at a time $t > \frac{\pi}{2a}$ (Corollary \ref{CorollarioB1}). 2) All trajectories corresponding $\omega \in (2 \omega^*, 3 \omega^*]$ which under the assumption $\gamma \geq \frac{1}{\sqrt{3}}$ are also in the first and second quadrant until $\frac{\pi}{2a}$, do not reach any of the points below the critical curve and outside the critical circle (i.e no points in the region R of Figure \ref{Perlaprova}), before hitting the boundary of the unit disk at time $t:=\frac{\pi}{a}$ (Corollary \ref{CorollarioB2}).
We then consider a curve obtained from the critical trajectory by slightly modifying it lowering the phase by a small quantity $\epsilon \lambda$, for $\lambda\in [0,1]$. This curve is in the region below the critical curve and outside the critical circle. Since the trajectories with $\omega \in (2\omega^*, 3\omega^*]$ cannot reach them optimally (they touch the boundary of the unit disk first (Corollary \ref{CorollarioB2})), the only trajectories left are the ones in $[\omega^*,2 \omega^*)$. Because of the existence of the optimal control, there is at least one value $\omega \in [\omega^*,2 \omega^*)$ and the corresponding trajectory which reaches the point corresponding to $\lambda$. Equating the two phases
up to a multiple of $2k\pi$, we find that, for a given $\lambda$, there exists a unique $\omega$ such that the two curve intersect at that $\lambda$ (Lemma \ref{Correspondence}). Here the assumption $\gamma \geq \frac{1}{\sqrt{3}}$ is used to show that only the case $k=-1$ has to be used in the multiple $2k\pi$. In fact Lemma \ref{Correspondence} establishes a on to one correspondence $\omega=\omega(\lambda)$ between the values $\lambda \in [0,1]$ and optimal values $\omega \in [\omega^*, 2\omega^*]$
As a consequence of the principle of optimality all points in $\mathcal{S}$ are reached by an optimal trajectory with $\omega \in [\omega^*, 2 \omega^*]$. This excludes however the points between the critical curve and the $\epsilon$-deformed curve. However these points are recovered at the limit as $\epsilon \rightarrow 0$. This is in the conclusion of the proof of Theorem \ref{Insideseparatrix} in Appendix B.
\section{Discussion}\label{Geo}
The above analysis provides a description of the optimal trajectories for every element in $SU(2)$. It also gives a very simple method to find the optimal control for an element $X_f \in SU(2)$.
Given such element one first singles out the $(1,1)$ element and the point $P_f$ in the unit disk and checks whether $P_f$ is inside or outside the separatrix. If $P_f$ is outside one has to use the trajectories described in section \ref{operatoridiagonali}, i.e., with $\omega \in (-\infty, \omega^*)$. The choice of $\omega$ can be made by successive approximations (for example using a simple bisection algorithm) by examining the plots for trajectories which leave $P_f$ on the right or on the left and getting closer and closer to the trajectory which actually contain $P_f$. If $P_f$ is inside the separatrix, the same procedure can be performed with the trajectories described in Section \ref{inside}. Once $\omega$ is found, one finds the corresponding $t$, either by tracing the plot or by solving an optimization problem minimizing (in $t$) the distance of the point on the trajectory from $P_f$. The last step is to adjust the phase $\tilde \phi$ in (\ref{soluzexpli}) (with the values found for $\omega$ and $t$) so that the element $(1,2)$ in (\ref{soluzexpli}) also coincides with the corresponding element in $X_f$. This completely determines the optimal controls in (\ref{controlli}).
Figure \ref{FigSUmmary} describes the work we have done to find the optimal control for the Hadamard gate $X_f:=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
1 & 1 \\ -1 & 1\end{array}\right)$ and $\gamma=\frac{1}{\sqrt{2}}$. The point $P_f$ is the point $(\frac{1}{\sqrt{2}}, 0)$ which is inside the separatrix curve. We have drawn a small circle around this point. The two curves in blue in the figure correspond to $\omega=1.2\omega^*$, $\omega=1.4\omega^*$ and $\omega^*=\frac{3}{4}$ in this case. The optimal curve is found for $\omega \approx 1.28 \omega^*$ and is the curve in red crossing the small circle in the figure. The optimal time is found to be approximately $t_{opt}\approx \pi+0.2$. The total phase of the $(1,2)$ element in (\ref{soluzexpli}) must be zero, therefore, we choose $\tilde \phi=-\omega t_{opt}=1.28 \omega^*(\pi +0.2)$. These values have to be replaced in (\ref{controlli}) to give the optimal controls.\footnote{Recall that we have replaced the notation $\tau$ with $t$ therefore we should have $2t$ in (\ref{controlli}) instead of $t$.}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\textwidth]{FIGURAFINALCOMPREHENSIVE}
\caption{Search for the optimal control for the Hadamard gate
in the case $\gamma=\frac{1}{\sqrt{2}}$. The separatrix is in black and the critical trajectory and critical circle are in red. Two trial trajectories corresponding to $\omega=1.2*\omega^*$ and $\omega=1.4*\omega^*$ are in blue and the (approximate) optimal trajectory is in red. In green it is depicted a typical trajectory for $\omega>2\omega^*$. In this case it is $\omega=2.4\omega^*$. As predicted by the theoretical analysis, this trajectory follows the critical trajectory with higher phase getting close to the critical circle and then goes further away from the origin until reaching the boundary of the unit disk.}
\label{FigSUmmary}
\end{figure}
\vs
While we were completing this work, other authors \cite{newpaperbis} submitted a paper on the same topic, building upon their previous work in \cite{newpaperonTimeOptimal} on the case $\omega_0=0$ (cf. Remark \ref{primobis}) and previous work in \cite{Boozer}, \cite{Kirillova}. In the paper \cite{newpaperbis}, the authors parametrize elements in $SU(2)$ with the so-called Hopf parameters\footnote{ These are defined as $\theta_{1,2,3}$ when writing $x=\cos(\theta_1) \cos(\theta_2)$, $y=\cos(\theta_1)\sin(\theta_2)$, $c=\sin(\theta_1)\cos(\theta_3)$, $d=\sin(\theta_1)\sin(\theta_3)$ (\ref{homeo}).} and the Euler parameters of the elements of $SU(2)$. They derive the dynamical equations in terms of these parameters and consider the optimal control problem in this setting. They prove properties of the optimal trajectories and give an algorithm to find the optimal controls. Our geometric analysis of the optimal trajectories in the unit disk provides an alternative approach which, beside giving a very straightforward method to find the time optimal control, as we have seen above, highlights the general picture of the optimal trajectories. Main features of this picture are the existence of a closed curve which separates two classes of optimal trajectories (the separatrix) and of a special (non-smooth) trajectory inside this curve (the critical trajectory) which is some sort of limit of all other trajectories and it is where these trajectories loose optimality. It will be interesting in the future to investigate if, how and in what cases these features can be found in higher dimensional time optimal control systems on Lie groups.
\vs
\section*{Acknowledgement} D. D'Alessandro's research was supported by ARO MURI under Grant
W911NF-11-1-0268. D. D'Alessandro also acknowledges the kind hospitality of the Department
of Mathematics at the University of Padova, where part of this work was
performed. Graphs were drawn using `fooplot' at
www.fooplot.com.
|
\section{Results and motivation}
One constructs the Jacobi metric $ds^2_{\rm JM}$ of classical mechanics by f\/ixing the total energy~$E$ of the system and
multiplying the kinetic energy metric $ds^2_{\rm K}$ by the conformal factor
\begin{gather*}
f =2(E-V),
\qquad
ds^2_{\rm JM} = f ds^2_{\rm K},
\end{gather*}
where~$V$ is the potential energy.
It is well-known that the geodesics for this Jacobi--Maupertuis metric (henceforth JM metric for short) are, up to
reparameterization, exactly the solutions to Newton's equations having energy~$E$.
(See Proposition~\ref{JMtoNewton} below for a~careful statement.
See~\cite[Theorem 3.7.7]{AbMar} for another discussion and a~nice proof.) The domain of the Jacobi metric is the domain
in conf\/iguration space where this conformal factor is non-negative and is called the Hill region:
\begin{gather*}
{\mathcal H} = \{q: f (q)\ge 0\}.
\end{gather*}
The Hill region includes the Hill boundary (sometimes called the zero velocity surface) where the conformal factor, and
hence the metric, vanishes:
\begin{gather*}
\partial {\mathcal H}= \{q:f(q) = 0\}.
\end{gather*}
A~``regular point'' $q_0$ of the Hill boundary is one for which $df(q_0) \ne 0$.
Here is our main result.
\begin{theorem}\label{Theorem1}
Any JM geodesic which comes sufficiently close to a~regular point $q_0$ of the Hill boundary contains a~pair of
conjugate points close to $q_0$ which are conjugate along a~short arc close to $q_0$.
In particular, such a~geodesic fails to minimize JM length.
\end{theorem}
This theorem is a~direct consequence of a~structure theorem, Theorem~\ref{Structure} below, regarding the conjugate
locus of near-boundary points, and results from Seifert's seminal paper~\cite{Seifert} which we recall in the next
section.
{\bf Motivations.} Two questions motivated this paper.
1.~Can the calculus of variations, applied to the JM metric reformulation of mechanics, uncover new results regarding the
classical three-body problem? The direct method of the calculus of variations breaks down at the Hill boundary since
curves lying in the boundary have zero JM length.
A~deeper understanding of the behaviour of near-boundary JM geodesics seems necessary to the further development of JM
variational methods in case where the Hill boundary is not empty.
For some results in celestial mechanics based on JM variational methods in instances where the Hill boundary is not
empty see~\cite{Moeckel} and~\cite{Soave_Terr} whin this direction
2.~Does the fact that JM curvatures tend to positive inf\/inity imply there are conjugate points near the boundary? Let~$q$
be a~point near a~regular point of the Hill boundary and let~$y$ denote its Riemannian distance from the boundary.
The sectional curvatures~$K$ of two-plane through~$q$ which contains the normal direction to the boundary tends to
positive inf\/inity like $1/ y^3$ as $y \to 0$.
The classical Bonnet--Meyer's estimate says that if the curvatures~$K$ along a~geodesic through~$q$ are greater than or
equal to a~positive constant $K_0$ then there must be a~point conjugate to~$q$ along the geodesic and lying within
$\pi/\sqrt{K_0} $ from~$q$.
This suggests the existence of conjugate points within $y^{3/2}$ from our point~$q$.
However, the {\it JM distance} of~$q$ to the boundary is also of order $y^{3/2}$ for small~$y$.
The two distances are of the same order.
These naive estimates do not tell us if Bonnet--Meyers ``wins'' to beat out the closeness of the boundary by creating
a~conjugate point before we have ``ref\/lected'' of\/f the boundary and left the region in which the Bonnet--Meyers curvature
estimate holds.
Theorem~\ref{Theorem1} asserts that, indeed, Bonnet--Meyers wins.
3.~The recent work~\cite{Ramis} claims that the harmonic oscillator, when it is reformulated in terms of JM geodesics, has
positive Lyapunov exponents.
This surprise, and trying to better understand it, was the seed that planted this paper.
\section{Mechanics and Seifert's coordinates}
By Newton's equations on a~manifold~$M$ we mean a~system of second-order dif\/ferential equations of the form
\begin{gather}
\nabla_{\dot \gamma} \dot \gamma = - \nabla V (\gamma).
\label{N}
\end{gather}
Here $\nabla$ is the Levi-Civita connection associated with a~f\/ixed Riemannian metric $ds^2_{\rm K}$ on~$M$.
(The subscript `K' is for `kinetic'.)
$V$ is a~chosen smooth function on~$M$ called the ``potential''.
The total energy
\begin{gather*}
H = \frac{1}{2} \langle \dot \gamma, \dot \gamma \rangle_{\gamma} + V(\gamma)
\end{gather*}
is constant along any solution to~\eqref{N}.
The inner product is the one def\/ined by the metric $ds^2_{\rm K}$.
Fix a~value $H=E$ for this energy and form the conformal factor
\begin{gather*}
f = 2(E-V)
\end{gather*}
and the resultant Jacobi--Maupertuis metric
\begin{gather}
ds^2_{\rm JM} = f ds^2_{\rm K}.
\label{JM}
\end{gather}
The following well-known proposition connects solutions to~\eqref{N} with Jacobi geodesics.
\begin{proposition}\label{JMtoNewton}
Solutions to~\eqref{N} with energy~$E$ are, after reparameterization, geodesics for the metric~\eqref{JM}
which
lie inside the Hill region $ f \ge 0$ and touch the Hill boundary $f=0$ in at most two points.
Conversely, any geodesic for the Jacobi metric lying inside the Hill region and touching the boundary in no more than
two points is a~reparameterization of a~solution to~\eqref{N}.
\end{proposition}
For a~proof see~\cite[Theorem 3.7.7]{AbMar}.
Special care must be taken with geodesics at the Hill boundary.
We have $f = 2(E -V(\gamma(t)) = \| \dot \gamma \|^2)$
along solutions $\gamma(t)$ to~\eqref{N}.
It follows that such a~solution hits the boundary at a~time $t_0$ if and only if $\dot \gamma (t_0) = 0$.
We call such a~solution a~``brake orbit''.
The point $q_0 = \gamma (t_0)$ where the solution hits the boundary is called the ``brake point'' since it has
instantaneously stopped.
Uniqueness of solutions to~\eqref{N} shows that a~brake orbit retraces its own path when we pass the brake instant:
$\gamma(t_0 + t) = \gamma(t_0 -t)$.
When we speak of Jacobi geodesics which hit the Hill boundary we mean exactly these brake orbits, up to
reparameterization.
If a~brake orbit hits the Hill boundary at two distinct points then it is periodic, shuttling back and forth forever
between these two brake points, with Newtonian period twice the Newtonian time it takes to get from one point to the
other.
Conversely, any periodic orbit having one brake point must have another distinct brake point.
Seifert's primary aim in~\cite{Seifert} was to establish the existence of such periodic brake solutions.
Suppose that the brake point $q_0$ is a~regular point of the boundary: $df(q_0) \ne 0$, i.e.~$\nabla V(q_0) \ne 0$.
Then Seifert proved that for small~$\epsilon$ the sub-arc $\gamma([t_0, t_0 + \epsilon])$ of~$\gamma$ is a~minimizing JM
geodesic which realize the JM distance from $\gamma(t_0 + \epsilon)$ to the boundary.
A~Taylor expansion yields $\gamma(t_0 + h) = q_0 - \frac{1}{2} h^2 \nabla V(q_0) + O(h^4)$ showing that this brake
orbit, as a~non-parameterized curve, is smooth and intersects the boundary orthogonally at the brake point.
Seifert solved Newton's equations with initial conditions $(\gamma, \dot \gamma) = (q, 0)$ on the Hill boundary to form
a~system of coordinates $(x, y)$ with $x = (x_1, \ldots, x_{n-1})$, $x_i, y \in \mathbb{R}$ and $n=\dim(M)$ for which
the~$y$-curves $x_i=\const$, $y=t$ are reparameterized brake orbits with brake instant $t=0$.
In these coordinates the Hill boundary is given by $y=0$ and the~$x_i$ coordinatize points on the Hill boundary.
We center the coordinates at a~regular point $q_0$ of the boundary, meaning that~$q_0$ has coordinates $(0,0)$.
We will call such coordinates ``cylinder coordinates'' or ``Seifert coordinates''.
When the coordinate domain has the form $W \times [0, \epsilon]$ we call the resulting sets in the manifold ``cylinder
neighborhoods'' of height~$\epsilon$.
The ``roof'' of the cylinder is the locus $y = \epsilon$.
The ``vertical lines'' are the images of $\{q_* \} \times [0, \epsilon]$ and are brake orbits.
We will say that the direction $-\dd{}{y}$ is `straight down'.
It is the tangent direction f\/ield to the brake orbits headed to the boundary.
{\bf Properties of Seifert coordinates.}
We recall some properties of Seifert's cylinder coordinates.
Within the f\/ive bullet points
\begin{itemize}\itemsep=0pt
\item[(1)] The JM distance of a~point to the Hill boundary is $y^{3/2}$
(\cite{Seifert}, the displayed equation just preceding his equation (46)).
\item[(2)] The JM metric is given by the coordinate expression
\begin{gather}
ds^2_{\rm JM} = ydy^2 + yf(x,y)\Big(\sum dx_i^2 + \sum h_{ij} (x,y) dx_i dx_j\Big),
\label{SeifertModel}
\end{gather}
where $f(0,0) = 1$ and $h_{ij}(x,y) = O(x^2+y^2)$ \cite[equation~(46)]{Seifert}.
\item[(3)]
For any suf\/f\/iciently small cylinder neighborhood~$A$ of $q_0$ and any~$\delta$ smaller than $45$ degrees
there is a~smaller cylinder neighborhood~$B$ of $q_0$ such that every geodesic which enters into~$B$ must exit and
leave~$A$ through the roof of~$A$, making an angle of less than~$\delta$ with the vertical line as it enters and leaves
(see Fig.~\ref{parabola}).
\begin{figure}[t]\centering
\includegraphics[width=5.0cm]{parabola3}
\caption{A geodesic which enters into $B$ must enter and leave
through the roof of $A$~at a~steep upward angle.}
\label{parabola}
\end{figure}
\item[(4)]
Along any of the geodesics $\gamma(t) = (x(t), y(t))$ described in the previous item, the height function
$y(t)$ is strictly convex relative to Newtonian time~$t$, with a~unique local minimum~\cite[Fig.~3, also Theorem~\ref{Theorem1}]{Seifert}.
\item[(5)] If a~geodesic enters into a~suf\/f\/iciently small cylinder neighborhood of height~$h$ then it leaves that
neighborhood within a~short Euclidean time of at most $C \sqrt{h}$ where~$C$ is any constant greater than $2 \sqrt{2}/\|\nabla V (q_0)\|$ \cite[Fig.~3 and Theorem~\ref{Theorem1}]{Seifert}.
\end{itemize}
Items (3) and (4) are not proved exactly as stated in Seifert.
We give proofs in Section~\ref{Section5} below.
For~$q$ a~point in a~cylinder set~$A$ let $q_* \in \partial {\mathcal H} \cap A$ denote the brake point along the brake
orbit connecting~$q$ to the boundary.
In terms of cylinder coordinates, if $q = (x,y)$ then $q_* = (x,0)$.
Let $C(q) \subset \bar {\mathcal H} \cap A$ denote the f\/irst conjugate locus to~$q$ for the restriction of the JM metric
to~$A$.
The points of $C(q)$ are the points conjugate to~$q$ along geodesic arcs {\it lying in}~$A$.
\begin{theorem}[structure theorem]\label{Structure}
Let~$A$ be a~cylinder set whose height~$\epsilon$ is sufficiently small.
Then the conjugate locus $C(q) \subset A$ of any point $q \in A$ has the following properties.
$C(q)$ is a~smooth hypersurface which intersects the Hill boundary tangentially at $q_*$ and in no other point.
As a~singularity of the exponential map, $C(q)$ represents the fold singularity.
Every geodesic arc through~$q$ in~$A$ lies entirely on the side of $C(q)$ closest to~$q$.
With the single exception of the brake orbit $[q, q_*]$, if such a~geodesic arc touches $C(q)$ then it touches it
tangentially.
Every geodesic through~$q$ whose initial tangent vector~$v$ is sufficiently close to the ``straight down'' direction
touches $C(q)$ $($see Fig.~{\rm \ref{envelope})}.
\end{theorem}
\begin{remark}
Compare this theorem with part (C) of Theorem 3.3 of~\cite{Warner} where Warner shows fold-type
conjugate loci occur stably and generically for Riemannian metrics.
\end{remark}
\begin{remark} If we take a~geodesic which starts at~$q$ and touches $C(q)$ and extend it slightly beyond $C(q)$ then
it will fail to minimize.
The extent to which it fails to minimize is measured by the {\it index} of Morse theory.
This index is 1 for all the geodesics of the structure theorem, this being the dimension of the kernel described towards
the end of Appendix~\ref{A2} in the paragraph {\bf Fold}.
\end{remark}
Let us continue with the notation of Theorem~\ref{Structure}.
If~$v$ is a~tangent vector to $q \in A$ then we say that~$v$ ``points downward'' if $dy(v) < 0$.
Consider the cone of downward-pointed velocities $v \in T_q {\mathcal H}$ with the additional property that the geodesic
with initial condition $(q,v)$ touches the conjugate locus $C(q)$ to~$q$ at a~point $c \in C(q)$ below~$q$: $y(c) <
y(q)$.
Call this set of vectors~$v$ the ``downward conjugate cone'' at~$q$ and denote it by ${\rm DC}(q)$.
\begin{theorem}\label{Theorem3}
The downward conjugate cone ${\rm DC}(q)$ is an open cone containing the brake direction.
As $q \to q_*$ along the brake segment $[q, q_*]$, the cone ${\rm DC}(q)$ limits to the open downward pointed cone consisting
of all vectors~$v$ whose angle with the straight down direction is less than $45$~degrees.
\end{theorem}
\section{Throwing balls: the model example}
The idea of our proof is to reduce the study of geodesics near the Hill boundary to that of a~model example for which
the geodesics can be found exactly.
The model is
\begin{gather}
\label{model}
ds^2_{\text{falling}} = y\big(dx^2 + dy^2\big),
\qquad
y\ge 0.
\end{gather}
In the model $f = y$, $ds^2_{\rm K} = dx^2 + dy^2$, $V(x,y) = -\frac{1}{2}y$ and $E =0$.
The corresponding Newton's equations are
\begin{gather}
\ddot x = 0,
\qquad
\ddot y = 1/2
\label{modelEqs}
\end{gather}
with energy
\begin{gather*}
H(x,y,\dot x, \dot y) = \frac{1}{2}\big(\dot x^2 + \dot y^2\big) -\frac{1}{2}y.
\end{gather*}
{\bf Freshman physics.} The af\/f\/ine change of variables $z = h_0 - 2gy$, $x =x$
turns these Newton's equations into the equation $\ddot z = -g$, $\ddot x = 0$
which governs the height~$z$ of a~ball thrown under the inf\/luence of the earth's
constant gravitational f\/ield of strength~$g$, pointed down.
This is the well-studied problem of ballistics from the 1st week or so of most beginning physics courses.
We are throwing balls or shooting cannons from a~f\/ixed point~$q$ with $z(q) < h_0$.
The Hill region $f \ge 0$ is $z \le h_0$.
The speed of our throws at a~f\/ixed point are all equal and are such that the maximum possible height we can reach, the
height reached if we hurl our ball straight up, is the height~$h_0$.
Turn Fig.~\ref{envelope}
upside down to see a~familiar picture of many balls being thrown at the same speed from
a~f\/ixed point to form a~sprinkler pattern, or if you prefer, the arcs of light seen in a~f\/ireworks display.
\begin{figure}[h]\centering \includegraphics[width=65mm]{envelope3}
\caption{Geodesics leaving a~point and headed toward the
boundary in the model example.
Turn the f\/igure upside down to see the trace of thrown balls, a~sprinkler, or f\/ireworks.}
\label{envelope}
\end{figure}
The general solution to our model Newton's equations~\eqref{modelEqs} is the family of parabolas:
\begin{gather}
x = x_0 + v_1 t,
\qquad
y = y_0 + v_2 t - \frac{1}{4} t^2.
\label{Solutions}
\end{gather}
The parameters $(x_0,y_0, v_1, v_2)$ are the initial conditions at time $t=0$ for our dif\/ferential
equations~\eqref{modelEqs}.
The energy along any one member of this family of solutions is $H(x_0,y_0, v_1, v_2)= \frac{1}{2} (v_1^2 + v_2^2)
-\frac{1}{2}y_0$.
We want this energy to be zero which means that $v_1^2 + v_2^2 = y_0$ so that the allowable velocities $(v_1, v_2)$
through $P_0$ vary over a~circle.
For each velocity in this circle we get a~parabola through $P_0$.
The {\it envelope} of this one-parameter family of parabolas is the conjugate locus.
\begin{lemma}\label{lemma1}
The envelope of the geodesics through $P_0 = (x_0, y_0)$ is the conjugate point locus to $P_0$ for the model metric
$($equation~\eqref{model}$)$ and is the parabola $y = \frac{1}{4 y_0} (x-x_0)^2$ tangent to the boundary $y=0$ at $(x_0, 0)$.
As a~singularity, the envelope realizes the simplest of the stable singularities of maps $\mathbb{R}^2 \to
\mathbb{R}^2$, the fold singularity, whose normal form near $(0,0)$ is $(u,v) \mapsto (u, v^2)$.
$($See~{\rm \cite[\textit{Theorems}~4.4 \textit{and}~4.5]{GG}} for results on the fold singularity.$)$
\end{lemma}
We prove the lemma in Appendices~\ref{A1} and~\ref{A2}.
In Appendix~\ref{A1} we review the def\/inition of `envelope' and show that the envelope is indeed the conjugate locus.
In Appendix~\ref{A2} we compute our specif\/ic envelope and show that the map for which it is a~singularity is a~simple fold.
More important than the exact formula for the envelope given in the lemma is the fact that it represents a~stable
singularity.
We also review the def\/inition of the fold and of a~stable singularity in Appendix~\ref{A2}.
We urge the reader to see the discussion in~\cite{Levi}, especially Fig.~5.6
for another good picture and a~discussion of this model example.
\subsection{Higher dimensions}
To place the model example (equation~\eqref{model}) in higher dimensions, take $x = (x_1, \ldots, x_{n-1}) \in
\mathbb{R}^{n-1}$ and work in the upper half space $y\ge 0$ of $\mathbb{R}^n = \mathbb{R}^{n-1} \times \mathbb{R}$ with
coordinates $(x,y)$.
Understand~$dx^2$ to mean the Euclidean metric~$\sum dx_i^2$.
Euclidean rotations about the vertical axes $x = x_0$ are isometries for the model metric.
The conjugate locus is obtained by taking the envelope just worked out in the lemma above for the planar case and
rotating it about the vertical axis through~$P_0$ to obtain a~hypersurface of revolution.
The lemma above holds as is.
In the equation for the conjugate locus we interpret $(x-x_0)^2$ to mean $\|x - x_0\|^2: = \sum (x_i - x_{i 0})^2$.
The singularity is again a~fold.
The normal form for the fold map from $\mathbb{R}^n \to \mathbb{R}^n$ remains the same, remembering to write $(x,y) \in
\mathbb{R}^{n-1} \times \mathbb{R} = \mathbb{R}^{n}$.
\section{Reduction to the model example}
\begin{proof}[Proof of Theorem~\ref{Structure}, the structure theorem]
Scale Seifert's cylinder coordinates by
$(x,y) $ $\to (\epsilon x, \epsilon y)$.
Here $x \in \mathbb{R}^{n-1}$, $y \in \mathbb{R}$, $y \ge 0$ and $\epsilon > 0$ is suf\/f\/iciently small.
We ask the reader to take a~glance at Seifert's metric normal form (equation~\eqref{SeifertModel})
and the function~$f$ there.
Taylor expand~$f$ about the origin: $f(x,y)= 1 + ax + by +O(x^2 + y^2)$ and write $f_1 (x,y) = ax + by$ for the linear
term.
In the rescaled coordinates Seifert's metric normal form (equation~\eqref{SeifertModel}) becomes:
\begin{gather*}
ds^2_{\rm JM} = \epsilon^3 y\Big(dy^2 + \Big(1 + \epsilon f_1 + O\big(\epsilon^2\big)\Big(\sum dx_i^2 + \sum \epsilon^2 h_{ij} (x,y) dx_i dx_j\Big)\Big)\Big)
\\
\phantom{ds^2_{\rm JM}}
= \epsilon^3 y \big(dy^2 + dx^2 + \epsilon f_1 dx^2 + O\big(\epsilon^2\big)\big)
\end{gather*}
where the $O(\epsilon^2)$ term only contains $dx_i$, $dx_j$ terms (no $dy$'s).
Dividing a~metric by a~positive constant does not change its geodesics.
Divide our metric by $\epsilon^3$ to get the metric:
\begin{gather*}
ds^2 = y\big(dy^2 + dx^2 + \epsilon f_1 dx^2 + O\big(\epsilon^2\big)\big),
\end{gather*}
whose conjugate locus is identical (after rescaling) to that of the original Seifert form.
Now view this expression for $ds^2$ as an instance of the Jacobi--Maupertuis principle.
In other words take the energy~$E$ to be $0$ so that the overall conformal factor~$y$ corresponds to the same potential
$V= -({1/2})y$ as in our model example.
View the term in parenthesis as the ``underlying metric''.
Now play the JM game in reverse, to write out Newton's equations, in Hamiltonian form, based on the structure of this
metric.
The kinetic energy metric part of our model has changed from $dy^2 + dx^2$ to $dy^2 + (1 + \epsilon f_1) dx^2 +
O(\epsilon^2)$ where the $O(\epsilon^2)$ error term does not involve $dy$ but only $dx_i dx_j$ terms.
Set $y = x_0$ momentarily so that we can write the metric tensor of this metric in the uniform manner $\sum g_{ab}
dx_a dx_b$ with $a$, $b$ now running from $0$ to $n-1$.
The Hamiltonian whose Hamilton's equations are Newton's equations is $\frac{1}{2} (\sum g^{ab} p_a p_b - y)$ where
$g^{ab}$ is the inverse to the matrix of metric coef\/f\/icients $g_{ab}$.
We see that
\begin{gather*}
(g)_{ab} = \left(
\begin{matrix}
1 & 0
\\
0 & I + \epsilon f_1 I + O\big(\epsilon^2\big)
\end{matrix}
\right)
\end{gather*}
from which it follows that the inverse matrix is
\begin{gather*}
g^{ab} = \left(
\begin{matrix}
1 & 0
\\
0 & I - \epsilon f_1 I + O\big(\epsilon^2\big)
\end{matrix}
\right)
\end{gather*}
yielding the Hamiltonian
\begin{gather*}
H_{\epsilon} = \frac{1}{2} \big(p_x^2 + p_y^2 - \epsilon f_1 p_x^2 + O\big(\epsilon^2\big)\big) - \frac{1}{2}y,
\end{gather*}
where $p_x^2$ means $\sum p_i^2$.
This Hamiltonian is a~small order~$\epsilon$ perturbation of the Hamiltonian
\begin{gather*}
H_0 = \frac{1}{2} \big(p_x^2 + p_y^2\big) - \frac{1}{2}y
\end{gather*}
for our model problem, solved in the last section.
Write $(x,y, p) \mapsto \Phi_t^{\epsilon} (x,y, p)$ for the Hamiltonian f\/low of our perturbed Hamiltonian $H_{\epsilon}$
with the unperturbed model f\/low being $\Phi_t^0$.
Since $H_{\epsilon}$ is within $C \epsilon$ of $H^0$ in the $C^k$ topology over compact sets (any~$k$ up to the
smoothness of the original problem), we have that their associated f\/lows are also close, provided we restrict to compact
subsets.
In other words, if $(x,y, p; t)$ are conf\/ined to vary over a~compact subset of $K \subset \mathbb{R}^n \times
\mathbb{R}^n \times \mathbb{R}$, then the restrictions of the associated Hamiltonian f\/lows $\Phi_t^{\epsilon}$ and
$\Phi_{t}^0$ are $O(\epsilon)$ close in the $C^{k}$ topology.
(Yes,~$k$, not $k-1$.
We get $k \to k -1$ when we dif\/ferentiate~$H$ to get the Hamiltonian vector f\/ield.
But we add $1$ back when we integrate the vector f\/ield to get the f\/low.) Now we worked out the details of the
unperturbed f\/low in in the last section (equation~\eqref{Solutions}).
Why can we restrict to compact sets? The geodesics for the model problem all leave a~given cylinder set in a~bounded
time $T_0$, and hence the same is true of the perturbed problem, with a~perhaps somewhat bigger $T_0$, say $2T_0$ to be
safe.
Fix the point $q = (x_0, y_0)$ (in rescaled variables).
The perturbed geodesics through $q = (x_0, y_0)$ are obtained by solving Hamilton's equations for $p =(p_x, p_y)$ lying
in the sphere $H_{\epsilon} (x_0, y_0, p) = 0$.
Thus, in computing the geodesics and conjugate locus we need only vary~$p$ and~$t$ over a~compact set of the form
$S^{n-1} \times [0, 2T_0]$.
We want to compare the singular loci of the map $(p, t) \mapsto \pi(\Phi_t^{\epsilon} (x_0,y_0, p))$ to that of the
unperturbed map $(p,t) \mapsto \pi(\Phi_t^{0} (x_0,y_0, p))$.
Here $\pi(x,y,p) = (x,y)$ is the projection onto conf\/iguration space.
The unperturbed map $(p,t) \mapsto \pi(\Phi_t^{0} (x_0,y_0, p))$ is the subject of Lemma~\ref{lemma1} of the previous section and
is structurally stable.
Hence there is an $\epsilon_0 >0 $ suf\/f\/iciently small so that for all $\epsilon < \epsilon_0$ the singularities of $(p,
t) \mapsto \pi(\Phi_t^{\epsilon} (x,y, p))$ are all folds (and are close to those of $ \pi \circ \Phi_t^0$).
The f\/lows being within order~$\epsilon$ we know that the maps $(p, t) \mapsto \Phi_t^{\epsilon} (x,y, p)$ and $(p,t)
\mapsto \Phi_t^0 (x,y, p)$ are $C^2$ close for~$\epsilon$ suf\/f\/iciently small.
Take~$\epsilon$ small enough that structural stability holds: the singular locus of the perturbed map is a~fold.
This locus is our conjugate locus.
The conjugate locus must touch the Hill boundary at the brake orbit through~$q$ as before.
Since the only geodesic through~$q$ touching the Hill boundary is the brake orbit, this is the only point where $C(q)$
intersects the boundary.
Since the conjugate locus and the boundary are both smooth hypersurfaces, and $C(q)$ lies entirely on one side of the
boundary, it must touch it tangentially.
\end{proof}
\section{Proof of the Seifert properties (3)--(5) and Theorem~\ref{Theorem3}}\label{Section5}
We begin with the model problem, taking $z = h_0 - gy$ so we can use the ball-throwing analogy.
The steeper the angle of the throw, the closer we get to the Hill boundary $z =h_0$ and only the straight-up throw
touches the boundary.
The cut-of\/f angle of 45 degrees is angle of maximal horizontal throw: at a~f\/ixed speed this is the upward angle to throw
a~ball so as to achieve the maximum horizontal distance before the ball hits the ground again at $z = z_0$.
Any higher angle and the ball drops short of the 45 degree throw, and the corresponding arc hits the conjugate locus
{\it before the ball hits the ground}, i.e.\ closer to the Hill boundary then when we started.
(Any lower angle and the ball hits the ground before it hits the conjugate locus, and hits the ground short of the 45
degree throw.) Thus if the angle of throw with the vertical is less than 45 degrees then the point where the geodesic
hits the conjugate locus $C(q)$ is closer to the Hill boundary than the starting point~$q$.
\begin{proof}[Proof of Seifert properties (3), (4), and (5).] Properties (3), (4) and (5) above regarding the Seifert coordinates are
easily verif\/ied for the model problem.
The reader can work out precise algebraic relations relating angles of steepness to heights.
The conditions involved in the three properties are open conditions in the $C^2$-topology on curves.
The real problem is an order~$\epsilon$ perturbation of the model problem as measured in the $C^1$-topology on the space
of vector f\/ields.
Consequently the geodesics for the real problem lie within an~$\epsilon$-$C^2$ neighborhood of those of the model
problem.
Consequently these properties continue to hold for the real problem, provided we take~$\epsilon$ small enough.
The precise constants involved will need to be relaxed a~bit.
The smaller we take~$\epsilon$, the closer we are to being able to use the same algebraic relations which the reader may
have worked out in the model problem.
{\it Property $(3)$.} Let us see the details of this argument for property~(3).
In the unperturbed model example, all the geodesics are parabolas.
A~bit of algebra shows that they can be written $y-y_m = \frac{1}{y_m} (x-x_m)^2$ where the vertex of the parabola,
which is the minimum value of~$y$ lies at $(x_m, y_m)$.
One then computes that $|dy/dx| \ge 1$ provided $|x-x_m| \ge \frac{1}{2}y_m$ which is to say $y \ge \frac{5}{4} y_m$.
In other words, once we reach a~height of $\lambda y_m$ or greater, $\lambda > 5/4$ along the parabola, the tangent line
to this parabola is less than 45~degrees
from the vertical.
It follows that if we take any constant $\lambda > 5/4$ then there is an~$\epsilon$ suf\/f\/iciently small, such that any
geodesic which enters into the cylinder neighborhood of height $\epsilon_1 < \epsilon$ will be leaving through the roof
of the cylinder of height $\lambda \epsilon$ and with a~tangent direction to the vertical of angle $45$ degrees or less.
The angle of the tangent with the vertical in the model parabolas decreases monotonically with their height, so the same
is true of the perturbed example, for~$\epsilon$ suf\/f\/iciently small.
We can increase~$\lambda$ so as to guarantee that this angle is, say, $42$ degrees, for example.
The constant in property (5) is verif\/ied by rewriting the model problem with a~constant~$g$: $\ddot y = g$ and observing
that~$g$ corresponds to the length of the force, or gradient of~$V$, and then doing a~bit of algebra and scaling.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Theorem3}]
To prove Theorem~\ref{Theorem3}, recall, as described a~few paragraphs up, that in the model problem the
downward pointed cone ${\rm DC}(q)$ is the cone of vectors making an angle of 45~degrees with respect to the vertical,
regardless of the initial point $q= (x_0, y_0)$.
The real case is a~small perturbation of the model case with the size of the perturbation tending to zero as we tend to
the boundary.
\end{proof}
\section{Proof of main theorem}
\begin{proof}
Consider the regular point $q_0$ of the boundary together with cylindrical neighborhoods centered on $q_0$ for which the
properties of Seifert hold.
Now by property (3) any geodesic which enters into the cylinder of height $\epsilon_1$ must leave through the roof of
a~cylinder of height $\lambda \epsilon_1$ at an angle closer to 42~degrees to the vertical.
Here~$\lambda$ a~f\/ixed constant, somewhat bigger than~$5/4$.
(We could have taken any degree less than~45 in place of~42~degrees.)
By Theorem~\ref{Theorem3}, for $\epsilon_1$ suf\/f\/iciently small,
these geodesic arcs all have conjugate pairs $q_1$, $q_2$ with $q_1$ being at height $y= \lambda \epsilon_1$ and with
$q_2$ being at a~lower height $ y < \lambda \epsilon_1$.
We refer the reader again to Fig.~\ref{parabola}.
We now simply insist that our geodesics enter the cylinder of height $\epsilon_1$ about $q_0$.
Any such geodesic is of the type described in the previous paragraph.
We are guaranteed our conjugate pair along this geodesic.
\end{proof}
|
\section{Introduction}\label{sec-intro}
Consider the one-dimensional reaction-diffusion equation
\begin{equation}\label{general-eqn}
u_{t}=u_{xx}+f(t,x,u), \quad x\in\R,\,\,t\in\R,
\end{equation}
where $f(t,x,u)$ is of ignition type, that is, there exists $\theta\in(0,1)$ such that for all $t\in\R$ and $x\in\R$, $f(t,x,u)=0$ for $u\in[0,\theta]\cup\{1\}$ and $f(t,x,u)>0$ for $u\in(\theta,1)$. Such an equation arises in the combustion theory (see e.g. \cite{BeLaLi90,BeNiSc85}). The number $\theta$ is usually referred to as the ignition temperature. The front propagation concerning this equation was first investigated by Kanel (see \cite{Ka60,Ka61,Ka62,Ka64}) in the space-time homogeneous media, i.e., $f(t,x,u)=f(u)$; he proved that all solutions, with initial data in some subclass of continuous functions with compact support and values in $[0,1]$, propagate at the same speed $c_{*}>0$, which is the speed of the unique traveling wave solution $\psi(x-c_{*}t)$, where $\psi$ satisfies
\begin{equation*}
\begin{split}
&\psi_{xx}+c_{*}\psi_{x}+f(\psi)=0,\quad\lim_{x\ra-\infty}\psi(x)=1\quad\text{and}\quad\lim_{x\ra\infty}\psi(x)=0.
\end{split}
\end{equation*}
Also see \cite{ArWe75,ArWe78,FiMc77,FiMc80} and references therein for the treatment of traveling wave solutions of \eqref{general-eqn} in space-time homogeneous media.
Recently, equation \eqref{general-eqn} in the space heterogeneous media, i.e., $f(t,x,u)=f(x,u)$, has attracted a lot of attention. In terms of space periodic media, that is, $f(x,u)$ is periodic in $x$, Berestycki and Hamel proved in \cite{BeHa02} the existence of pulsating fronts or periodic traveling waves of the form $\psi(x-c_{*}t,x)$, where $\psi(s,x)$ is periodic in $x$ and satisfies a degenerate elliptic equation with boundary conditions $\lim_{s\ra-\infty}\psi(s,x)=1$ and $\lim_{s\ra\infty}\psi(s,x)=0$ uniformly in $x$. In the work of Weinberger (see \cite{We02}), he proved from the dynamical system viewpoint that solutions with general non-negative compactly supported initial data spread with the speed $c_{*}$. We also refer to \cite{Xin91,Xin92,Xin93} for related works.
In the general space heterogeneous media, wavefront with a profile is no longer appropriate, and we are looking for more general wavefronts such as transition fronts in the sense of Berestycki and Hamel (see \cite{BeHa07,BeHa12}), that is,
\begin{defn}\label{def-transition-wave}
A global-in-time solution $u(t,x)$, $x\in\R$, $t\in\R$ of \eqref{general-eqn} is called a {\rm transition front} if there is a function $\xi:\R\ra\R$, called, {\rm interface location function}, such that
\begin{equation*}
\begin{split}
&u(t,x)\ra1\,\,\text{uniformly in}\,\,t\,\,\text{and}\,\,x\leq\xi(t)\,\,\text{as}\,\,x-\xi(t)\ra-\infty,\,\,and\\
&u(t,x)\ra0\,\,\text{uniformly in}\,\,t\,\,\text{and}\,\,x\geq\xi(t)\,\,\text{as}\,\,x-\xi(t)\ra\infty.
\end{split}
\end{equation*}
A transition front $u(t,x)$ is called {\rm critical} if for any transition front $\tilde{u}(t,x)$ there exists a function $\zeta:\R\ra\R$ such that
\begin{equation*}
\begin{split}
u(t,x)&\geq\tilde{u}(t,x),\quad x\leq\zeta(t),\\
u(t,x)&\leq\tilde{u}(t,x),\quad x\geq\zeta(t)
\end{split}
\end{equation*}
for all $t\in\R$.
\end{defn}
Transition fronts are proper generalizations of traveling waves in homogeneous media and periodic traveling waves (or pulsating fronts) in periodic media. It is easily seen that the interface location function $\xi(t)$ in Definition \ref{def-transition-wave} is unique up to addition by bounded functions. Transition fronts in the above sense are also called {\it generalized traveling waves} in some literature, especially in the time heterogeneous media, i.e., $f(t,x,u)=f(t,u)$ (see \cite{Sh11-1}). Roughly speaking, critical transition fronts are the steepest ones among all transition fronts. It is known that the existence of a transition front implies the existence of a critical transition front and critical transition fronts (if exist) are unique up to phase shift (see Lemmas \ref{lm-critical-traveling-wave1} and Lemma \ref{lm-critical-traveling-wave2}).
In the work of Nolen and Ryzhik (see \cite{NoRy09}), and Mellet, Roquejoffre and Sire (see \cite{MeRoSi10}), transition fronts with additional properties, such as, time monotonicity, finite speed, exponential decay ahead of the interface, etc., are proven to exist in the general space heterogeneous media
of ignition type (the work \cite{NoRy09} also deals with transition fronts in space random media of ignition type). Later, stability and uniqueness of such transition fronts are also established in \cite{MNRR09}. These results are then generalized by Zlato\v{s} (see \cite{Zl13}) to the equations in space heterogeneous cylindrical domains of ignition type.
However, there is little study of transition fronts in general time heterogeneous and random media of ignition type.
In the current paper, we first study front propagation phenomenon in the reaction-diffusion equation \eqref{general-eqn} in general time heterogeneous media, that is,
\begin{equation}\label{main-eqn}
u_{t}=u_{xx}+f(t,u), \quad x\in\R,\,\,t\in\R.
\end{equation}
Here are the assumptions on $f(t,u)$:
\begin{itemize}
\item[\rm(H1)] There is a $\theta\in(0,1)$, called the ignition temperature, such that for all $t\in\R$,
\begin{equation*}
\begin{split}
f(t,u)&=0,\quad u\in(-\infty,\theta]\cup\{1\},\\
f(t,u)&>0,\quad u\in(\theta,1),\\
f(t,u)&<0,\quad u>1.
\end{split}
\end{equation*}
The family of functions $\{f(t,u), u\in\R\}$ is locally uniformly H\"{o}lder continuous. The family of functions $\{f(t,u), t\in\R\}$ is locally uniformly Lipschitz continuous. For any $t\in\R$, $f(t,u)$ is continuously differentiable for $u\geq\theta$.
\item[\rm(H2)] There are Lipschitz continuous functions $f_{\inf}$, $f_{\sup}$ satisfying
\begin{equation*}
\begin{split}
&f_{\inf}, f_{\sup}\in C^1([\theta,\infty),\R)\\
&f_{\inf}(u)=0=f_{\sup}(u)\,\,\text{for}\,\,u\in[0,\theta]\cup\{1\},\\
&0>(f_{\inf})_{u}(1)\geq(f_{\sup})_{u}(1),\\
&0<f_{\inf}(u)<f_{\sup}(u)\,\,\text{for}\,\,u\in(\theta,1)
\end{split}
\end{equation*}
such that $f_{\inf}(u)\leq f(t,u)\leq f_{\sup}(u)$ for $u\in[\theta,1]$ and $t\in\R$.
\end{itemize}
We prove
\begin{thm}\label{thm-transition-wave}
Suppose $\rm(H1)$ and $\rm(H2)$.
\begin{itemize}
\item[\rm(1)] (Existence of transition front) Equation \eqref{main-eqn} admits a transition front $u(t,x)$, $x\in\R$, $t\in\R$ in the sense of Definition \ref{def-transition-wave}, where the function $\xi:\R\ra\R$ is continuously differentiable and satisfies $u(t,\xi(t))=\theta$ for all $t\in\R$. Moreover, the following properties hold:
\begin{itemize}
\item[\rm(i)] (Monotonicity of the transition front) $u_{x}(t,x)<0$ for $x\in\R$ and $t\in\R$;
\item[\rm(ii)] (Uniform steepness) $\sup_{t\in\R}u_{x}(t,\xi(t))<0$;
\item[\rm(iii)] (Finite speed) $\sup_{t\in\R}|\xi'(t)|<\infty$;
\item[\rm(iv)] (Uniform decaying estimates) there exists a continuous and strictly decreasing function $v:\R\ra(0,1)$ satisfying $v(x)\geq 1-c_{1}e^{c_{2}x}$, $x\leq-c_{3}$ for some $c_{1},c_{2},c_{3}>0$ and $v(x)=\theta e^{-cx}$, $x\geq0$ for some $c>0$ such that
\begin{equation*}
\begin{split}
u(t,x+\xi(t))&\geq v(x),\quad x\leq0;\\
u(t,x+\xi(t))&\leq v(x),\quad x\geq0.
\end{split}
\end{equation*}
\end{itemize}
\item[\rm(2)] (Periodicity) If $f(t+T,u)=f(t,u)$, then \eqref{main-eqn} admits a periodic traveling wave $u(t,x)$, that is, there are a constant $c\in\R$ and a function $\psi:\R\times\R\ra(0,1)$ satisfying
\begin{equation*}\label{eqn-periodic-tw}
\begin{cases}
\psi_{t}=\psi_{xx}+c\psi_{x}+f(t,\psi),\cr
\lim_{x\ra-\infty}\psi(t,x)=1,\,\,\lim_{x\ra\infty}\psi(t,x)=0\,\,\text{uniformly in}\,\,t\in\R,\cr
\psi(t,\cdot)=\psi(t+T,\cdot)\,\,\text{for all}\,\,t\in\R
\end{cases}
\end{equation*}
such that $u(t,x)=\psi(t,x-ct)$ for $x\in\R$ and $t\in\R$.
\end{itemize}
\end{thm}
Clearly, due to the space homogeneity of \eqref{main-eqn}, if $u(t,x)$ is a transition front of \eqref{main-eqn}, then any space translation of $u(t,x)$ is also a transition front. All these consists of a family of transition fronts propagating to the right. By space reflection, we obtain another family propagating to the left. We see that the transition front constructed in Theorem \ref{thm-transition-wave} has a time-dependent profile given by $\psi(t,x)=u(t,x+\xi(t))$, which is a solution of
\begin{equation}\label{eqn-time-profile}
\begin{cases}
\psi_{t}=\psi_{xx}+\xi'(t)\psi_{x}+f(t,\psi),\cr
\lim_{x\ra-\infty}\psi(t,x)=1,\,\,\lim_{x\ra\infty}\psi(t,x)=0\,\,\text{uniformly in}\,\,t\in\R.
\end{cases}
\end{equation}
We then study front propagation phenomena in reaction-diffusion equations in random media, that is,
\begin{equation}
\label{random-eq}
u_t=u_{xx}+f(\sigma_{t}\omega,u),\quad x\in \R, \,\, t\in\R,
\end{equation}
where $\omega\in\Omega$, $((\Omega,\mathcal{F},P),\{\sigma_{t}\}_{t\in\R})$ is a metric dynamical system
(i.e. $(\Omega,\mathcal{F},P)$ is a probability space, the mapping $(t,\omega)\mapsto \sigma_t(\omega):\R\times\Omega\ra\R$ is measurable,
$\sigma_t\circ\sigma_s=\sigma_{t+s}$ for any $s,t\in\R$, and $P(\sigma_t F)=P(F)$ for any $t\in\R$ and $F\in\mathcal{F}$) and $f:\Omega\times\R\to\R$ satisfies
\begin{itemize}
\item[\rm(H3)] $f:\Omega\times\R\to\R$ is measurable, and for each $\omega\in\Omega$, $f^\omega(t,u)=f(\theta_{t}\omega,u)$ satisfies $\rm(H1)$ and $\rm(H2)$.
\end{itemize}
We look for random traveling wave solutions of \eqref{random-eq} in the following sense (see \cite{Sh04}).
\begin{defn}
\label{random-wave-def} A family $\{u(t,x;\omega)\}_{\omega\in\Omega}$ of global-in-time solutions of \eqref{random-eq} is called a
{\rm random traveling wave} if there are measurable functions $\Psi:\R\times\Om\to\R$ and $\xi:\R\times\Om\to \R$ such that
$$
u(t,x;\omega)=\Psi(x-\xi(t;\omega),\sigma_t \omega),
$$
and for each fixed $\omega\in\Omega$, $u(t,x;\omega)$ is a transition front of \eqref{random-eq}, that is,
$$
\lim_{x\to -\infty}\Psi(x,\sigma_t \omega)=1,\quad \lim_{x\to\infty}\Psi(x,\sigma_t\omega)=0
$$
uniformly in $t\in\R$.
\end{defn}
We prove
\begin{thm}
\label{thm-random-wave} Assume $\rm(H3)$.
\begin{itemize}
\item[\rm(1)] Equation \eqref{random-eq} admits a random traveling wave $u(t,x;\omega)=\Psi(x-\xi(t;\omega),\sigma_t \omega)$, where the function $\xi:\R\times \Omega\ra\R$ is continuously differentiable in $t\in\R$ and satisfies $u(t,\xi(t,\omega);\omega)=\theta$ for all $t\in\R$ and $\omega\in\Omega$.
\item[\rm(2)] If $((\Omega,\mathcal{F},P),\{\sigma_t\}_{t\in\R})$ is an ergodic metric dynamical system, then there are $c^*\in\R$ and $\Psi^*(\cdot)\in C_{\rm unif}^b (\R,\R)$ such that for a.e.
$\omega\in\Omega$,
$$
\lim_{t\to\infty}\frac{\xi(t;\omega)}{t}=c^{*},
$$
$$
\lim_{t\to\infty}\frac{1}{t}\int_0^ t\Psi(x,\sigma_s\omega)ds=\Psi^*(x),\quad \forall \,x\in\R,
$$
and
$$
\lim_{x\to -\infty}\Psi^*(x)=1,\quad \lim_{x\to\infty} \Psi^*(x)=0.
$$
\item[\rm(3)] If $((\Om,\mathcal{F},P),\{\sigma_t\}_{t\in\R})$ is a compact flow, then there is $\Om_0\subset\Om$ with $\si_t(\Om_0)=\Om_0$ such that $\Om_0$ is a residual subset of $\Om$ (i.e., $\Om_{0}$ is the intersection of countably many open dense subsets of $\Om$) and the map $\om\in\Om\mapsto \Psi(\cdot,\om)\in C_{\rm unif}^b(\R,\R)$ is continuous at $\om\in\Om_0$.
\end{itemize}
\end{thm}
We remark that if $\Omega=\{f(\cdot+\tau,\cdot)|\tau\in\R\}$ equipped with open compact topology and $\sigma_t g(\cdot,\cdot)=g(\cdot+t,\cdot)$ for $g\in\Omega$ and $t\in\R$, where $f(t,u)$ satisfies (H1), (H2) and $f(t+T,u)=f(t,u)$,
then Theorem \ref{thm-random-wave} (3) implies that \eqref{main-eqn} admits a periodic traveling wave solution, which recovers Theorem \ref{thm-transition-wave} (2). In the case that $\Omega={\rm hull}(f):={\rm cl}\{f(\cdot+\tau,\cdot)|\tau\in\R\}$ with open compact topology and $\sigma_t g(\cdot,\cdot)=g(\cdot+t,\cdot)$ for $g\in \Omega$ and $t\in\R$, where $f(t,u)$ satisfies $\rm(H1)$ and $\rm(H2)$ and is almost periodic in $t$ uniformly with respect to $u$, whether \eqref{main-eqn} admits almost periodic traveling wave solutions remains open. This issue together with the uniqueness and stability of transition fronts of \eqref{main-eqn} are studied in \cite{SS-1}.
We also remark that time-periodic traveling waves were first investigated by Alikakos, Bates and Chen (see \cite{AlBaCh99}) in time periodic bistable media. For time heterogeneous bistable equations, transition fronts with a time-dependent profile satisfying \eqref{eqn-time-profile} and their uniqueness and stability have been investigated by Shen (see e.g. \cite{Sh99-1,Sh99-2,Sh04,Sh06}). There are similar results for time heterogeneous KPP equations (see e.g. \cite{NaRo12,Sh11,RR14}). Transition fronts have also been proven to exist in space heterogeneous Fisher-KPP type equations (see \cite{NRRZ12,Zl12}). But it is far from being clear in space heterogeneous media of bistable type due to the wave blocking phenomenon (see \cite{LeKe00}) except the one established in \cite{NoRy09} under additional assumptions.
To this end, we comment on the differences between the analysis in the present paper and that in \cite{Sh06} and \cite{NoRy09}. In \cite{Sh06}, transition fronts of the following equation
\begin{equation}\label{eqn-time-hetero-bistable}
u_{t}=u_{xx}+f_{B}(t,u),\quad (t,x)\in\R\times\R
\end{equation}
in time heterogeneous bistable case were studied. In particular, $u\equiv0$ and $u\equiv1$ are two stable solutions and $\theta(t)$, $t\in\R$ is the unstable solution between $0$ and $1$. The method used in \cite{Sh06} has a strong dynamical system favor. More precisely, instead of focusing on \eqref{eqn-time-hetero-bistable}, the following family
$$
u_{t}=u_{xx}+f_{B}(t+s,u),\quad (t,x)\in\R\times\R,\quad s\in\R
$$
were treated as a whole. Since the analysis in \cite{Sh06} heavily relies on the uniform instability of the solution $\theta(t)$, the method can not be applied in our case.
In \cite{NoRy09}, transition fronts in space-heterogeneous ignition equations were treated by studying the following equation
\begin{equation}\label{eqn-time-hetero-ignition}
u_{t}=u_{xx}+f_{I}(x,u),\quad (t,x)\in\R\times\R
\end{equation}
with the neutral stable solution $u\equiv0$, the stable solution $u\equiv1$ and the ignition temperature $\theta\in(0,1)$. The proof of the existence of transition fronts is constructive via the construction of approximating solution $u^{n}(t,x)$, where $u^{n}(t,x)$ is the unique solution of \eqref{eqn-time-hetero-ignition} with well-constructed initial data at initial time $t=-n$. An important property of $u^{n}(t,x)$ is the time monotonicity, i.e., $u^{n}_{t}(t,x)>0$, which implies $\dot{\xi}^{n}(t)>0$, where $\xi^{n}(t)$ is the interface location defined by
\begin{equation*}
\xi^{n}(t)=\sup\{x\in\R|u^{n}(t,x)=\theta\}.
\end{equation*}
The fact that $\xi^{n}(t)$ is increasing plays a very important role in the analysis done in \cite{NoRy09}.
In our case, we first construct approximating solutions $u^{n}(t,x)$ as in \cite{NoRy09}, but our approximating solutions satisfy space monotonicity, i.e., $u_{x}^{n}(t,x)<0$, instead of time monotonicity. We then look at the interface location $\xi^{n}(t)$ defined to be the unique point such that $u^{n}(t,\xi^{n}(t))=\theta$. However, due to the time-dependence of the nonlinearity $f(t,u)$, $\xi^{n}(t)$ oscillates, and therefore, the analysis in \cite{NoRy09} does not apply. A major part of the present paper is devoted to the analysis of the propagation of $\xi^{n}(t)$ with oscillations.
The rest of the paper is organized as follows. In Section \ref{sec-entire-sol}, we construct a global-in-time solution of \eqref{main-eqn} as the limit of approximating solution sequence. Section \ref{sec-bd-interface-width} is devoted to the boundedness of interface width. In Section \ref{Non-Flat Estimate Near Ignition Temperature}, we prove that the derivative of the approximating solution sequence near the ignition temperature is uniformly negative. In Section \ref{uniform-estimate}, we establish uniform estimates behind and ahead of the interface for the approximating solution sequence. In Section \ref{sec-transition-wave-const}, we prove the existence of transition fronts of \eqref{main-eqn} and finish the proof Theorem \ref{thm-transition-wave}. In Section \ref{sec-random-wave}, we investigate random traveling wave solutions of \eqref{random-eq} and prove Theorem \ref{thm-random-wave}.
\section{Construction of Global-in-Time Solutions}\label{sec-entire-sol}
In this section, we construct a global-in-time solution of \eqref{main-eqn}. Throughout this section, we assume (H1) and (H2).
First, we consider the space-time homogeneous equation
\begin{equation}\label{eqn-ig-inf}
u_{t}=u_{xx}+f_{\inf}(u).
\end{equation}
By $\rm(H2)$, $f_{\inf}$ is of standard ignition type. Classical results (see e.g. \cite{ArWe75,ArWe78,FiMc77}) ensure the existence of a unique constant $c_{\inf}>0$ and a twice continuously differentiable function $\phi$ satisfying
\begin{equation}\label{traveling-homo}
\begin{cases}
\phi_{xx}+c_{\inf}\phi_{x}+f_{\inf}(\phi)=0,\cr
\phi_{x}<0,\quad\lim_{x\ra-\infty}\phi(x)=1\,\,\text{and}\,\,\lim_{x\ra\infty}\phi(x)=0
\end{cases}
\end{equation}
such that $\phi(x-c_{\inf}t)$, $x\in\R$, $t\in\R$ and its translations are traveling wave solutions of \eqref{eqn-ig-inf}. Thus, we may assume, without loss of generality, that $\phi(0)=\theta$. Since $f_{\inf}(u)=0$ for $u\in[0,\theta]$, a direct computation gives
\begin{equation}\label{explicit-sol}
\phi(x)=\theta e^{-c_{\inf}x},\quad x\geq0.
\end{equation}
\begin{lem}\label{lem-entire-sol}
For any $s<0$, there is a unique $x_{s}\in\R$ such that the solution $u(t,x;s)$, $t\geq s$ of \eqref{main-eqn} with $u(s,x;s)=\phi(x-x_{s})$ satisfies $u(0,0;s)=\theta$. Moreover, $x_{s}\ra-\infty$ as $s\ra-\infty$.
\end{lem}
\begin{proof}
Fix any $s<0$. Let $u^{y}(t,x;s)$ be the solution of \eqref{main-eqn} with $u^{y}(s,x;s)=\phi(x-y)$. By comparison principle, $u^{y}(t,x;s)\geq\phi(x-y-c_{\inf}(t-s))$ for $t\geq s$. In particular, $u^{y}(0,0;s)\geq\phi(c_{\inf}s-y)$. Note that if $y>c_{\inf}s$, then $\phi(c_{\inf}s-y)>\phi(0)=\theta$ by monotonicity. Thus, $u^{y}(0,0;s)>\theta$ if $y>c_{\inf}s$.
On the other hand, let us fix some constant $M>0$ such that $f(t,u)\leq Mu$ for all $u\geq0$ and $t\in\R$. Such an $M$ exists by $\rm(H1)$ and $\rm(H2)$. Now, set $v^{y}(t,x;s)=e^{-c_{\inf}(x-y-y_{0}-c(t-s))}$ for some $y_{0}\in\R$ and $c>0$ to be chosen. By \eqref{explicit-sol}, we can easily find an $y_{0}\in\R$ such that $\phi(x-y)\leq v^{y}(s,x;s)$ for all $y\in\R$. We fix such an $y_{0}$. We compute
\begin{equation*}
(v^{y})_{t}-(v^{y})_{xx}-f(t,v^{y})=c_{\inf}(c-c_{\inf})v^{y}-f(t,v^{y}).
\end{equation*}
Thus, if we choose $c>0$ such that $c_{\inf}(c-c_{\inf})\geq M$, then $v^{y}$ is a sup-solution of \eqref{main-eqn}, which leads to $u^{y}(t,x;s)\leq v^{y}(t,x;s)=e^{-c_{\inf}(x-y-y_{0}-c(t-s))}$ by comparison principle. In particular, $u^{y}(0,0;s)\leq e^{-c_{\inf}(cs-y-y_{0})}$. Thus, $u^{y}(0,0;s)<\theta$ for $y\ll-1$. Continuity of the solution with respect to $y$ then ensures the existence of some $x_{s}$ as in the statement of the lemma.
The uniqueness follows from comparison principle. In fact, if there are $x_{s}$ and $x_{s}^{*}$ with $x_{s}\neq x_{s}^{*}$, then we have either $u^{x_{s}}(0,x;s)<u^{x_{s}^{*}}(0,x;s)$ or $u^{x_{s}^{*}}(0,x;s)<u^{x_{s}}(0,x;s)$ for all $x\in\R$ by comparison principle, since either $\phi(x-x_{s})<\phi(x-x_{s}^{*})$ or $\phi(x-x_{s}^{*})<\phi(x-x_{s})$ holds for all $x\in\R$. Hence, for different $x_{s}$ and $x_{s}^{*}$, we can not have both $u^{x_{s}}(0,0;s)=\theta$ and $u^{x_{s}^{*}}(0,0;s)=\theta$.
The ``moreover" part is a simple consequence of the estimate
\begin{equation}\label{estimate-lower-bound}
u^{x_{s}}(t,x;s)\geq\phi(x-x_{s}-c_{\inf}(t-s)).
\end{equation}
In fact, if $\inf_{s<0}x_{s}>-\infty$, then for all $s\ll0$, $u^{x_{s}}(0,x;s)\geq\phi(0-x_{s}-c_{\inf}(0-s))>\theta$. It is a contradiction.
\end{proof}
From the above lemma, we can construct a global-in-time solution.
\begin{thm}\label{thm-entire-sol}
There exists a sequence $\{s_{n}\}_{n\in\N}\subset(-\infty,0)$ with $s_{n}\ra-\infty$ as $n\ra\infty$ and a function $u(t,x)$, $x\in\R$, $t\in\R$ continuously differentiable in $t$ and twice continuously differentiable in $x$ such that for any compact $K\subset\R\times\R$, the following limits
\begin{equation*}
\begin{split}
\lim_{n\ra\infty}u(t,x;s_{n})&=u(t,x),\quad\lim_{n\ra\infty}u_{t}(t,x;s_{n})=u_{t}(t,x),\\
\lim_{n\ra\infty}u_{x}(t,x;s_{n})&=u_{x}(t,x),\quad\lim_{n\ra\infty}u_{xx}(t,x;s_{n})=u_{xx}(t,x)
\end{split}
\end{equation*}
exist and are uniform in $(t,x)\in K$. In particular, $u(t,x)$, $x\in\R$, $t\in\R$ is a global-in-time solution of \eqref{main-eqn}.
\end{thm}
\begin{proof}
It is a consequence of Lemma \ref{lem-entire-sol}, a priori estimates for parabolic equations (see e.g. \cite{Fri64}), Arzela-Ascoli theorem and the diagonal argument.
\end{proof}
The global-in-time solution $u(t,x)$ constructed in Theorem \ref{thm-entire-sol} is a candidate for the expected transition front. All we need is to show that this solution satisfies certain non-degenerate and uniform decaying estimates. This, however, can be deduced from the boundedness of interface width, the steepness estimate, and the uniform decaying estimates of the approximating solutions $u(t,x;s)$, which are the objectives of Section \ref{sec-bd-interface-width}, Section \ref{Non-Flat Estimate Near Ignition Temperature} and Section \ref{uniform-estimate}, respectively. In Section \ref{sec-transition-wave-const}, we finish the construction of transition fronts.
In the rest of this section, we derive some fundamental properties of $u(t,x;s)$.
\begin{lem}\label{lem-basic-prop}
For any $s<0$ and $t\geq s$, there hold the following properties:
\begin{equation*}
\lim_{x\ra-\infty}u(t,x;s)=1,\quad\lim_{x\ra\infty}u(t,x;s)=0\quad\text{and}\quad u_{x}(t,x;s)<0.
\end{equation*}
\end{lem}
\begin{proof}
The limit at $+\infty$ follows from the following two-sided estimates
\begin{equation}\label{aprior-estimate-two-side1}
\phi(x-x_{s}-c_{\inf}(t-s))\leq u(t,x;s)\leq e^{-c_{\inf}(x-x_{s}-y_{0}-c(t-s))},
\end{equation}
where the lower bound and the upper bound are constructed in Lemma \ref{lem-entire-sol}. The limit at $-\infty$ follows from the following two-sided estimates
\begin{equation*}\label{aprior-estimate-two-side11}
\phi(x-x_{s}-c_{\inf}(t-s))\leq u(t,x;s)\leq 1,
\end{equation*}
where the lower bound is constructed in Lemma \ref{lem-entire-sol} and the upper bound is due to the fact that $u\equiv 1$ is a solution of \eqref{main-eqn}
and $u(s,x;s)<1$ for all $x\in\R$.
We now show $u_{x}(t,x;s)<0$. Clearly, it is the case if $t=s$. So we assume $t>s$. Since $\phi(x-x_{s})$ is strictly decreasing, we apply maximum principle to $u(t,x+y;s)-u(t,x;s)$ for any $y>0$ to conclude that $u(t,x+y;s)<u(t,x;s)$. That is, $u(t,x;s)$ is strictly decreasing. For contradiction, suppose $u_{x}(t_{0},x_{0};s)=0$ for some $t_{0}>s$ and $x_{0}\in\R$. Let $u^{o}(t;t_{0},a)$ be the solution of the ODE $u_{t}=f(t,u)$ with $u^{o}(t_{0};t_{0},a)=a=u(t_{0},x_{0};s)$. Note $u^{o}(t;t_{0},a)$ extends naturally for $t<t_{0}$. Let $v(t,x;s)=u(t,x;s)-u^{o}(t;t_{0},a)$. It satisfies $v(t_{0},x_{0};s)=0$, $v_{x}(t_{0},x_{0};s)=0$ and the linear equation
\begin{equation}\label{eqn-linear-zero}
v_{t}=v_{xx}+q(t,x)v,
\end{equation}
where
\begin{equation*}
\begin{split}
q(t,x)=\left\{\begin{aligned}
\frac{f(t,u(t,x;s))-f(t,u^{o}(t;t_{0},a))}{u(t,x;s)-u^{o}(t;t_{0},a)},&\quad u(t,x;s)\neq u^{o}(t;t_{0},a),\\
0,&\quad u(t,x;s)=u^{o}(t;t_{0},a)
\end{aligned} \right.
\end{split}
\end{equation*}
is bounded. Applying Angenent's result (see e.g. \cite[Theorem B]{Ang88}) to \eqref{eqn-linear-zero}, there exist $\ep>0$ and $\de>0$ such that $v(t-\de,x;s)$ has at least two zeros in the interval $[x_{0}-\ep,x_{0}+\ep]$. However, due to the monotonicity of $u(t-\de,x;s)$ in $x$, $v(t-\de,x;s)$ has exactly one zero. This is a contradiction. Hence, $u_{x}(t,x;s)<0$.
\end{proof}
By Lemma \ref{lem-basic-prop}, for any $\la\in(0,1)$, $s<0$ and $t\geq s$, there is a unique $\xi_{\la}(t;s)\in\R$ such that
\begin{equation*}
u(t,\xi_{\la}(t;s);s)=\la.
\end{equation*}
The case $\la=\theta$ is of particular interest and it does play an important role in our later arguments. Notice $\xi_{\theta}(s;s)=x_{s}$ for all $s<0$. As usual, we refer to the point $(\xi_{\la}(t;s),\la)$ on the solution curve as the interface and $\xi_{\la}(t;s)$ as the interface location. The following lemma shows the continuous differentiability of $\xi_{\la}(t,s)$ in $t$.
\begin{lem}\label{lem-negativity}
Let $\la\in(0,1)$. For any $s<0$, the interface location $\xi_{\la}(t,s)$ is continuously differentiable in $t$ for $t>s$. Moreover, there holds
\begin{equation*}
\frac{d\xi_{\la}(t;s)}{dt}=-\frac{u_{t}(t,\xi_{\la}(t;s);s)}{u_{x}(t,\xi_{\la}(t;s);s)}.
\end{equation*}
\end{lem}
\begin{proof}
The continuity follows from the continuity of $u(t,x;s)$ and its monotonicity in $x$ by Lemma \ref{lem-basic-prop}. We show the continuous differentiability. Since $u(t,\xi_{\la}(t;s);s)=\la$ for $t\geq s$, we have $u(t+\ep,\xi_{\la}(t+\ep;s);s)-u(t,\xi_{\la}(t;s);s)=0$. Thus,
\begin{equation*}
\begin{split}
&\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t,\xi_{\la}(t;s);s)}{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}\times\frac{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}{\ep}\\
&\quad\quad=\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t+\ep,\xi_{\la}(t+\ep;s);s)}{\ep}
\end{split}
\end{equation*}
Since $\xi_{\la}(t;s)$ is continuous in $t$, we have
\begin{equation*}
\lim_{\ep\ra0}\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t,\xi_{\la}(t;s);s)}{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}=u_{x}(t,\xi_{\la}(t;s);s)<0
\end{equation*}
by Lemma \ref{lem-basic-prop}. In particular, $\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t,\xi_{\la}(t;s);s)}{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}\neq0$ for all small $\ep$. Thus,
\begin{equation*}
\begin{split}
&\frac{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}{\ep}\\
&\quad\quad=\frac{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}{u(t,\xi_{\la}(t+\ep;s);s)-u(t,\xi_{\la}(t;s);s)}\times\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t+\ep,\xi_{\la}(t+\ep;s);s)}{\ep}
\end{split}
\end{equation*}
Passing to the limit $\ep\ra0$ in the above equality, we conclude from the limit
\begin{equation*}
\lim_{\ep\ra0}\frac{u(t,\xi_{\la}(t+\ep;s);s)-u(t+\ep,\xi_{\la}(t+\ep;s);s)}{\ep}=-u_{t}(t,\xi_{\la}(t;s);s)
\end{equation*}
for $t>s$
that $\frac{d\xi_{\la}(t;s)}{dt}=\lim_{\ep\ra0}\frac{\xi_{\la}(t+\ep;s)-\xi_{\la}(t;s)}{\ep}$ exists and
\begin{equation*}
\frac{d\xi_{\la}(t;s)}{dt}=-\frac{u_{t}(t,\xi_{\la}(t;s);s)}{u_{x}(t,\xi_{\la}(t;s);s)}
\end{equation*}
for $t>s$,
which also implies the continuity of $\frac{d\xi_{\la}(t;s)}{dt}$ in $t$ for $t>s$. Hence, $\xi_{\la}(t;s)$ is continuously differentiable in
$t$ for $t>s$.
\end{proof}
We remark that due to the time-dependence of the nonlinear term $f(t,u)$, the time derivative $u_{t}(t,\xi_{\la}(t;s);s)$ does not have a fixed sign in general, and hence, $\frac{d\xi_{\la}(t;s)}{dt}$ does not have a fixed sign, which means $\xi_{\la}(t;s)$ oscillates and it is an unpleasant fact and does cause a lot of troubles (we point out that in the space heterogeneous case, the interface always propagates in one direction due to the time monotonicity, see \cite{MeRoSi10,NoRy09}). But, the estimate \eqref{estimate-lower-bound} forces $\xi_{\la}(t;s)$ to approach $+\infty$ as time $t$ elapses. However, the estimate \eqref{estimate-lower-bound} does not tell much information about how does $\xi_{\la}(t;s)$ approach $+\infty$. Later, in Lemma \ref{lem-rightward-prop-above-temp} and Lemma \ref{lem-propogation}, we characterize the rightward propagation of $\xi_{\la}(t;s)$, which plays the crucial role in deriving the boundedness of interface width and the exponential decay of the transition front ahead of the interface. We also note that $u_{t}(t,\xi_{\la}(t;s);s)$ is uniformly bounded in $t\ge s+\delta_0$ for any $\delta_0>0$, but temporarily we are not sure if $u_{x}(t,\xi_{\la}(t;s);s)$ is uniformly away from $0$. But it is the case, see Theorem \ref{thm-non-flat}. Hence, $\frac{d\xi_{\la}(t;s)}{dt}$ is uniformly bounded in $t\ge s+\delta_0$ for any $\delta_0>0$, that is, the interfaces cannot propagate faster than certain speed.
\section{Bounded Interface Width}\label{sec-bd-interface-width}
In this section, we show that the distance between their interface locations of any two interfaces remains bounded as time elapses. Throughout this section, we consider \eqref{main-eqn} and assume $\rm(H1)$ and $\rm(H2)$. The main result of this section is given by
\begin{thm}\label{thm-bd-interface-width}
For any $\la_{1},\la_{2}\in(0,1)$, there exists $C=C(\la_{1},\la_{2})>0$ such that
\begin{equation*}
|\xi_{\la_{1}}(t;s)-\xi_{\la_{2}}(t;s)|\leq C(\la_{1},\la_{2})
\end{equation*}
for all $s<0$, $t\geq s$.
\end{thm}
To prove the above theorem, we first prove some lemmas and propositions. First of all, we characterize the rightward propagation of interfaces above the ignition temperature. Let $f_{B}$ be a continuously differentiable function satisfying
\begin{equation}\label{bistable-nonlinearity}
\begin{split}
&f_{B}(0)=0,\,\,f_{B}(u)<0\,\,\text{for}\,\,u\in(0,\theta),\\
&f_{B}(u)=f_{\inf}(u)\,\,\text{for}\,\,u\in[\theta,1]\,\,\text{and}\,\,\int_{0}^{1}f_{B}(u)du>0.
\end{split}
\end{equation}
Since $f_{\inf}(u)>0$ for $u\in(\theta,1)$, such an $f_{B}$ exists. Clearly, $f_{B}$ is of standard bistable type and $f_{B}(u)\leq f(t,u)$ for all $u\in[0,1]$ and $t\in\R$. Hence, there exist (see e.g.\cite{ArWe75,ArWe78,FiMc77}) a unique constant $c_{B}>0$ and a wave profile $\phi_{B}$ satisfying $(\phi_{B})_{x}<0$, $\phi_{B}(-\infty)=1$ and $\phi_{B}(\infty)=0$ such that $\phi_{B}(x-c_{B}t)$ and its translations are traveling wave solutions of
\begin{equation}\label{eqn-bi-1}
u_{t}=u_{xx}+f_{B}(u).
\end{equation}
\begin{lem}\label{lem-rightward-prop-above-temp}
Let $\la\in(\theta,1)$. For any $\ep>0$, there is $t_{\ep,\la}>0$ such that
\begin{equation*}
\xi_{\la}(t;s)-\xi_{\la}(t_{0};s)\geq (c_{B}-\ep)(t-t_{0}-t_{\ep,\la})
\end{equation*}
for $s<0$, $t\geq t_{0}\geq s$.
\end{lem}
\begin{proof}
Let us fix a $\la\in(\theta,1)$. We first define
\begin{equation*}
\begin{split}
\psi_{*}(x)=\left\{\begin{aligned}
\la,&\quad x\leq0,\\
\max\{-C^{*}x+\la,0\},&\quad x\geq0,
\end{aligned} \right.
\end{split}
\end{equation*}
where $C^{*}>0$ is such that $\inf_{s<0,t\geq s}\inf_{x\in\R}u_{x}(t,x;s)\geq-C^{*}$. Such an $C^{*}$ exists by a priori estimates
for parabolic equations. Clearly, for any $s<0$ and $t_{0}\geq s$, we have
\begin{equation}\label{special-function-1}
\psi_{*}(x)\leq u(t_{0},x+\xi_{\la}(t_{0};s);s),\quad x\in\R.
\end{equation}
Next, for $t_{0}\geq s$, let $u_{B}(t,x;t_{0})$, $t\geq t_{0}$ be the solution of \eqref{eqn-bi-1} with initial data $u_{B}(t_{0},x;t_{0})=\psi_{*}(x)(\leq u(t_{0},x+\xi_{\la}(t_{0};s);s)$ by \eqref{special-function-1}). Thus, time homogeneity and comparison principle ensure
\begin{equation*}
u_{B}(t-t_{0},x;0)=u_{B}(t,x;t_{0})\leq u(t,x+\xi_{\la}(t_{0};s);s),\quad x\in\R,\,\,t\geq t_{0}.
\end{equation*}
By the stability of traveling wave solutions of \eqref{eqn-bi-1} (see \cite[Theorem 3.1]{FiMc77}) and the conditions satisfied by $\psi_{*}$, there exist $z_{0}=z_{0}(\la)\in\R$, $K=K(\la)>0$ and $\om=\om(\la)>0$ such that
\begin{equation*}
\sup_{x\in\R}|u_{B}(t-t_{0},x;0)-\phi_{B}(x-c_{B}(t-t_{0})-z_{0})|\leq Ke^{-\om(t-t_{0})}.\quad t\geq t_{0}.
\end{equation*}
In particular, for $t\geq t_{0}$ and $x\in\R$
\begin{equation}\label{estimate-l-bd}
u(t,x+\xi_{\la}(t_{0};s);s)\geq u_{B}(t-t_{0},x;0)\geq\phi_{B}(x-c_{B}(t-t_{0})-z_{0})-Ke^{-\om(t-t_{0})}.
\end{equation}
Let $T_{0}=T_{0}(\la)>0$ be such that $Ke^{-\om T_{0}}=\frac{1-\la}{2}$ and denote by $\xi_{B}(\frac{1+\la}{2})$ the unique point such that $\phi_{B}(\xi_{B}(\frac{1+\la}{2}))=\frac{1+\la}{2}$. Setting $x=c_{B}(t-t_{0})+z_{0}+\xi_{B}(\frac{1+\la}{2})$ in \eqref{estimate-l-bd}, we find for any $t\geq t_{0}+T_{0}$
\begin{equation*}
u(t,c_{B}(t-t_{0})+z_{0}+\xi_{B}(\frac{1+\la}{2})+\xi_{\la}(t_{0};s);s)\geq\phi_{B}(\xi_{B}(\frac{1+\la}{2}))-Ke^{-\om T_{0}}=\la.
\end{equation*}
Monotonicity then yields
\begin{equation}\label{propagation-result-1}
\xi_{\la}(t;s)-\xi_{\la}(t_{0};s)\geq c_{B}(t-t_{0})+z_{0}+\xi_{B}(\frac{1+\la}{2}), \quad t\geq t_{0}+T_{0}.
\end{equation}
Finally, we consider $\xi_{\la}(t;s)$ for $t\in[t_{0},t_{0}+T_{0}]$. To do so, let $\vp(x)=\max\{\hat{\vp}(x),0\}$ for $x\in\R$, where $\hat{\vp}$ is the unique solution of the following problem
\begin{equation*}
-\hat{\vp}_{xx}=f_{\inf}(\hat{\vp}),\quad \hat{\vp}(0)=\la,\quad \hat{\vp}_{x}(0)=0.
\end{equation*}
The function $\hat{\vp}$ satisfies the following properties:
\begin{itemize}
\item it is even and strictly decreasing for $x\geq0$;
\item it is strictly concave down for $x\in(-z_{1},z_{1})$, where $z_{1}>0$ is such that $\hat{\vp}(z_{1})=\theta$;
\item it is linear for $x\geq z_{1}$ with a negative slope.
\end{itemize}
Then, we can easily find a shift $z_{*}<0$ such that $\vp(x-z_{*})\leq\psi_{*}(x)$ for $x\in\R$. Denote by $u_{I}(t,x;t_{0})$ the solution of $u_{t}=u_{xx}+f_{\inf}(u)$ with $u_{I}(t_{0},x;t_{0})=\vp(x-z_{*})$. Since $-\vp_{xx}\leq f_{\inf}(\vp)$, we obtain from the maximum principle that $u_{I}(t,x;t_{0})\geq u_{I}(t_{0},x;t_{0})=\vp(x-z_{*})$ for all $t>t_{0}$. In particular, $u_{I}(t,z_{*};t_{0})\geq\la$ for all $t\geq t_{0}$.
Since $\vp(x-z_{*})\leq\psi_{*}(x)\leq u(t_{0},x+\xi_{\la}(t_{0};s);s)$, comparison principle implies that $u_{I}(t,x;t_{0})\leq u(t,x+\xi_{\la}(t_{0};s);s)$ for $t\geq t_{0}$. Setting $x=z_{*}$, we in particular have $u(t,z_{*}+\xi_{\la}(t_{0};s);s)\geq u_{I}(t,z_{*};t_{0})\geq\la$ for $t\geq t_{0}$. Monotonicity then yields
\begin{equation}\label{propagation-result-2}
\xi_{\la}(t;s)\geq z_{*}+\xi_{\la}(t_{0};s),\quad t\geq t_{0}.
\end{equation}
The result then follows from \eqref{propagation-result-1} and \eqref{propagation-result-2}.
\end{proof}
As seen in the proof of Lemma \ref{lem-rightward-prop-above-temp}, the bistable traveling waves $\phi_{B}(x-c_{B}t)$ push the approximation solutions $u(t,x;s)$ move rightward in some average sense. This property can also be derived if we use ignition traveling waves of $u_{t}=u_{xx}+f_{\inf}(u)$. The reason for using bistable traveling waves is that bistable traveling waves attract a larger class of initial data than ignition traveling waves do, and therefore, it is more flexible and convenient to use bistable traveling waves.
Next, for $\ka>0$, set $c^*_{\ka}=2\sqrt{\ka}$ and $\la_{\ka}=\sqrt{\ka}$. Clearly, $\la_{\ka}^{2}-c^*_{\ka}\la_{\ka}+\ka=0$, and hence, $e^{-\la_{\ka}x}$ is a solution of $\psi''+c^*_{\ka}\psi'+\ka\psi=0$. It is well-known that $c^*_{\ka}=\min_{\la>0}\frac{\ka+\la^{2}}{\la}=\frac{\ka+\la_{\ka}^{2}}{\la_{\ka}}$ is the minimal speed of a KPP traveling wave (see e.g. \cite{KPP37}).
For $\ka>0$, $s<0$ and $t\geq s$, define
\begin{equation}\label{new-interface}
\xi(t;s)=\inf\Big\{y\in\R\Big|u(t,x;s)\leq e^{-\la_{\ka}(x-y)},\quad x\in\R\Big\}.
\end{equation}
Due to the second estimate in \eqref{aprior-estimate-two-side1}, $\xi(t;s)$ is well-defined if $\la_{\ka}\leq c_{\inf}$, that is, $\ka\in(0,c_{\inf}^{2}]$. Here, we use the $\ka$-independent notation for $\xi(t;s)$, but this should not cause any trouble, since later in Lemma \ref{lem-bd-width}, we only need one small $\ka$. The following result controls the rightward propagation of $\xi(t;s)$.
\begin{lem}\label{lem-new-interface-bd}
Let $\ka\in(0,c_{\inf}^{2}]$. Set $\ka_{0}=\sup_{u\in(0,1)}\frac{f_{\sup}(u)}{u}$ and $c_{\ka_{0}}=\frac{\ka_{0}}{\la_{\ka}}+\la_{\ka}$. Then,
\begin{equation*}
\xi(t;s)-\xi(t_{0};s)\leq c_{\ka_{0}}(t-t_{0})
\end{equation*}
for all $s<0$, $t\geq t_{0}\geq s$.
\end{lem}
\begin{proof}
For $s<0$, $t\geq t_{0}\geq s$, define
\begin{equation*}
v(t,x;t_{0})=e^{-\la_{\ka}(x-\xi(t_{0};s)-c_{\ka_{0}}(t-t_{0}))}.
\end{equation*}
Since $c_{\ka_{0}}=\frac{\ka_{0}}{\la_{\ka}}+\la_{\ka}$, i.e., $\la_{\ka}^{2}-c_{\ka_{0}}\la_{\ka}+\ka_{0}=0$, we readily check that $v_{t}=v_{xx}+\ka_{0}v$. By the definition of $\ka_{0}$, we have $\ka_{0}v\geq f_{\sup}(v)$ for all $v\geq0$. It then follows from $v(t_{0},x;t_{0})=e^{-\la_{\ka}(x-\xi(t_{0};s))}\geq u(t_{0},x;s)$ by \eqref{new-interface} and the comparison principle that $v(t,x;t_{0})\geq u(t,x;s)$ for $t\geq t_{0}$, which leads to the result.
\end{proof}
Note the definition of $\xi(t;s)$ in \eqref{new-interface} and Lemma \ref{lem-new-interface-bd} does not guarantee any continuity of $\xi(t;s)$ in $t$. But, if we know $\xi(t;s)$ is increasing from $\xi(t_{0};s)$ for $t>t_{0}$, then it is controlled continuously by Lemma \ref{lem-new-interface-bd}. This observation is important in the next technical lemma, which is crucial in proving Theorem \ref{thm-bd-interface-width}.
\begin{lem}\label{lem-bd-width}
There exists $\la_{*}\in(\theta,1)$ such that for any $\la\in(\theta,\la_{*}]$, there is $C=C(\la)>0$ such that
\begin{equation*}
|\xi_{\la}(t;s)-\xi(t;s)|\leq C
\end{equation*}
for all $s<0$, $t\geq s$.
\end{lem}
\begin{proof}
We follow \cite[Lemma 2.5]{Zl13}.
Recall that for given $\ka>0$, $c^*_{\ka}=2\sqrt{\ka}$, and that $c_{B}>0$ is the unique speed of traveling wave solutions of \eqref{eqn-bi-1}. We fix some $\ka\in(0,c_{\inf}^{2}]$ such that $c^*_{\ka}<c_{B}$, and set $\ep=\frac{c_{B}-c^*_{\ka}}{2}$ in Lemma \ref{lem-rightward-prop-above-temp}.
Let
\begin{equation*}
\la_{*}=\min\big\{u>0\big|f_{\sup}(u)=\ka u\big\}.
\end{equation*}
As a consequence, we have $f(t,u)\leq f_{\sup}(u)\leq \ka u$ for all $u\in[0,\la_{*}]$.
Fix an $\la\in(\theta,\la_{*}]$. Let $C_{0}=\max\{\xi(s;s)-\xi_{\la}(s;s),1\}$ (note $C_{0}$ is independent of $s$) and
$C_{1}=C_{0}+c_{B}t_{\ep,\la}$, where $t_{\ep,\la}$ is as in Lemma \ref{lem-rightward-prop-above-temp}. Notice the estimate $\xi_{\la}(t;s)-\xi(t;s)\leq C$ for some large $C>0$ is trivial. We show $\xi(t;s)-\xi_{\la}(t;s)\leq C_1$. Suppose this is not the case, then we can find some $t_{1}\geq s_{1}$ such $\xi(t_{1};s_{1})-\xi_{\la}(t_{1};s_{1})>C_1$. Let
\begin{equation*}
t_{0}=\sup\big\{t\in[s_{1},t_{1}]\big|\xi(t;s_{1})-\xi_{\la}(t;s_{1})\leq C_{0}\big\}.
\end{equation*}
We claim $\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})\leq C_{0}$. It is trivial if there are only finitely many $t\in[s_{1},t_{1}]$ such that $\xi(t;s_{1})-\xi_{\la}(t;s_{1})\leq C_{0}$. So we assume there are infinitely many such $t$ and the claim is false. Then, there exists a sequence $\{\tilde{t}_{n}\}_{n\in\N}\subset[s_{1},t_{0})$ such that $\xi(\tilde{t}_{n};s_{1})-\xi_{\la}(\tilde{t}_{n};s_{1})\leq C_{0}$ for $n\in\N$ and $\tilde{t}_{n}\ra t_{0}$ as $n\ra\infty$. Moreover, $\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})=\tilde C_{1}>C_{0}$. It then follows that for all $n\in\N$
\begin{equation*}
\xi(\tilde{t}_{n};s_{1})-\xi_{\la}(\tilde{t}_{n};s_{1})\leq C_{0}=C_{0}-\tilde C_{1}+\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1}),
\end{equation*}
that is,
\begin{equation*}
\tilde C_{1}-C_{0}+\xi_{\la}(t_{0};s_{1})-\xi_{\la}(\tilde{t}_{n};s_{1})\leq \xi(t_{0};s_{1})-\xi(\tilde{t}_{n};s_{1})\leq c_{\ka_{0}}(t_{0}-\tilde{t}_{n}),
\end{equation*}
where the second inequality is due to Lemma \ref{lem-new-interface-bd}.
Passing $n\ra\infty$, we conclude from the continuity of $\xi_{\la}(t;s_{1})$ in $t$ (see Lemma \ref{lem-negativity}) that $\tilde C_{1}-C_{0}\leq0$. It is a contradiction. Hence, the claim is true, that is, $\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})\leq C_{0}$. It follows that $t_{0}<t_{1}$.
Instead of $\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})\leq C_{0}$, there must hold
\begin{equation}\label{equality-technical}
\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})=C_{0}.
\end{equation}
Suppose \eqref{equality-technical} is not true, then we can find some $\ep_{0}>0$ such that $\xi(t_{0};s_{1})-\xi_{\la}(t_{0};s_{1})=C_{0}-\ep_{0}$. Since $\xi(t;s_{1})-\xi_{\la}(t;s_{1})>C_{0}$ for $t\in(t_{0},t_{1}]$ by the definition of $t_{0}$, we deduce from Lemma \ref{lem-new-interface-bd} that for $t\in(t_{0},t_{1}]$
\begin{equation*}
\begin{split}
C_{0}<\xi(t;s_{1})-\xi_{\la}(t;s_{1})&\leq\xi(t_{0};s_{1})+c_{\ka_{0}}(t-t_{0})-\xi_{\la}(t_{0};s_{1})+\xi_{\la}(t_{0};s_{1})-\xi_{\la}(t;s_{1})\\
&=C_{0}-\ep_{0}+c_{\ka_{0}}(t-t_{0})+\xi_{\la}(t_{0};s_{1})-\xi_{\la}(t;s_{1}).
\end{split}
\end{equation*}
Since $\xi_{\la}(t;s_{1})$ is continuous in $t$, we fix some $t>t_{0}$ but close to $t_{0}$ such that $|c_{\ka_{0}}(t-t_{0})+\xi_{\la}(t_{0};s_{1})-\xi_{\la}(t;s_{1})|\leq\frac{\ep_{0}}{2}$, which then leads to $C_{0}<C_{0}-\frac{\ep_{0}}{2}$. It is a contradiction. Hence, \eqref{equality-technical} holds.
Next, we look at the time interval $[t_{0},t_{1}]$ and set $\tilde{\xi}(t;s_{1})=\xi(t_{0};s_{1})+c^*_{\ka}(t-t_{0})$ for $t\in[t_{0},t_{1}]$. Note both $\xi_{\la}(t;s_{1})$ and $\tilde{\xi}(t;s_{1})$ are continuous on $[t_{0},t_{1}]$ and $\xi_{\la}(t_{0};s_{1})<\tilde{\xi}(t_{0};s_{1})$ by \eqref{equality-technical}. We claim that $\xi_{\la}(t;s_{1})<\tilde{\xi}(t;s_{1})$ for all $t\in[t_{0},t_{1}]$. Suppose this is not the case and let $t_{2}=\min\{t\in[t_{0},t_{1}]|\xi_{\la}(t;s_{1})=\tilde{\xi}(t;s_{1})\}$. Clearly, $t_{2}\in(t_{0},t_{1}]$. Define
\begin{equation*}
v(t,x;t_{0})=e^{-\la_{\ka}(x-\tilde{\xi}(t;s_{1}))},\quad x\in\R,\,\, t\in[t_{0},t_{2}].
\end{equation*}
Since $\la_{\ka}^{2}-c^*_{\ka}\la_{\ka}+\ka=0$, we easily check $v_{t}=v_{xx}+\ka v$. Now, consider the parabolic domain
\begin{equation*}
D=\big\{(t,x)\in[t_{0},t_{2}]\times\R\big|x\geq\tilde{\xi}(t;s_{1})\big\}.
\end{equation*}
We see that for $(t,x)\in D$, $x\geq\tilde{\xi}(t;s_{1})\geq\xi_{\la}(t;s_{1})$, which leads to $u=u(t,x;s_{1})\leq\la\in(0,\la_{*}]$ by monotonicity, and then, $f(t,u)\leq \ka u$ as noted in the beginning of the proof. Also, at the initial moment $t_{0}$, we have $u(t_{0},x;s_{1})\leq e^{-\la_{\ka}(x-\xi(t_{0};s_{1}))}=v(t_{0},x;t_{0})$, and at the boundary point $x=\tilde{\xi}(t;s_{1})$, we trivially have $u(t,\tilde{\xi}(t;s_{1});s_{1})<1=v(t,\tilde{\xi}(t;s_{1});t_{0})$. Thus, comparison principle yields $u(t,x;s_{1})\leq v(t,x;t_{0})$ on $D$, which leads to
\begin{equation*}
u(t,x;s_{1})\leq v(t,x;t_{0})=e^{-\la_{\ka}(x-\tilde{\xi}(t;s_{1}))},\quad x\in\R,\,\, t\in[t_{0},t_{2}].
\end{equation*}
It follows that $\xi(t;s_{1})\leq\tilde{\xi}(t;s_{1})$ for $t\in[t_{0},t_{2}]$ by definition in \eqref{new-interface}. In particular, $\xi(t_{2};s_{1})\leq\tilde{\xi}(t_{2};s_{1})=\xi_{\la}(t_{2};s_{1})$. Since $t_{2}\in(t_{0},t_{1}]$, we have $\xi(t_{2};s_{1})-\xi_{\la}(t_{2};s_{1})>C_{0}$ by the definition of $t_{0}$. It is a contradiction. Thus, the claim follows, that is, $\xi_{\la}(t;s_{1})<\tilde{\xi}(t;s_{1})$ for all $t\in[t_{0},t_{1}]$, and repeating the above arguments, we see
\begin{equation}\label{propagation-estimate-refined}
\xi(t;s_{1})\leq\tilde{\xi}(t;s_{1})=\xi(t_{0};s_{1})+c^*_{\ka}(t-t_{0}),\quad t\in[t_{0},t_{1}].
\end{equation}
It follows from \eqref{propagation-estimate-refined} and Lemma \ref{lem-rightward-prop-above-temp} that for any $t\in[t_{0},t_{1}]$
\begin{equation*}
\begin{split}
\xi(t;s_{1})-\xi_{\la}(t;s_{1})&\leq\xi(t_{0};s_{1})+c^*_{\ka}(t-t_{0})-[\xi_{\la}(t_{0};s_{1})+(c_{B}-\ep)(t-t_{0}-t_{\ep,\la})]\\
&\leq C_{0}+(c_{B}-\ep)t_{\ep,\la}-(c_{B}-c^*_{\ka}-\ep)(t-t_{0})\\
&\leq C_{0}+c_{B}t_{\ep,\la}\\
&=C_1.
\end{split}
\end{equation*}
Thus, we in particular have $\xi(t_{1};s_{1})-\xi_{\la}(t_{1};s_{1})\leq C_{1}$, which is a contradiction. Consequently, $\xi(t;s)-\xi_{\la}(t;s)\leq C_{1}$ for all $s<0$, $t\geq s$. This completes the proof.
\end{proof}
The following proposition is in fact Theorem \ref{thm-bd-interface-width} restricted to the case $\la_{1},\la_{2}\in(0,\la_{*}]$.
\begin{prop}\label{prop-bd-interface-width}
For any $\la_{1},\la_{2}\in(0,\la_{*}]$ there exists $C=C(\la_{1},\la_{2})>0$ such that
\begin{equation*}
|\xi_{\la_{1}}(t;s)-\xi_{\la_{2}}(t;s)|\leq C
\end{equation*}
for all $s<0$ and $t\geq s$, where $\la_{*}\in(\theta,1)$ is as in Lemma \ref{lem-bd-width}.
\end{prop}
\begin{proof}
Fix some $\ka\in(0,c_{\inf}^{2}]$ such that $c^*_{\ka}<c_{B}$ as in the proof of Lemma \ref{lem-bd-width}. Let $\la_{*}\in(\theta,1)$ be as in Lemma \ref{lem-bd-width} and $\la_{1},\la_{2}\in(0,\la_{*}]$. We may assume, without loss of generality, that $\la_{1}<\la_{2}$. Thus, $\xi_{\la_{1}}(t;s)\geq\xi_{\la_{2}}(t;s)$, and
\begin{equation*}
\xi_{\la_{1}}(t;s)-\xi_{\la_{2}}(t;s)\leq\eta_{\la_{1}}(t;s)-\xi_{\la_{*}}(t;s),
\end{equation*}
where $\eta_{\la_{1}}(t;s)$ is the unique point such that $e^{-\la_{\ka}(\eta_{\la_{1}}(t;s)-\xi(t;s))}=\la_{1}$. Since $\eta_{\la_{1}}(t;s)-\xi(t;s)\equiv \hat{C}(\la_{1})$ for some $\hat{C}(\la_{1})>0$, we deduce from Lemma \ref{lem-bd-width} that $\xi_{\la_{1}}(t;s)-\xi_{\la_{2}}(t;s)\leq \hat{C}(\la_{1})+C(\la_{*})$.
\end{proof}
Note that in the presence of Proposition \ref{prop-bd-interface-width}, to finish the proof of Theorem \ref{thm-bd-interface-width}, we only need to bound $\xi_{\theta}(t;s)-\xi_{\la}(t;s)$ for all $\la\in(\theta,1)$ close to $1$. To do so, we need to study the propagation of $\xi_{\theta}(t;s)$.
Let $u^{o}(t;t_{0},a)$ be the solution of the ODE $u_{t}=f(t,u)$ with initial data $u^{o}(t_{0};t_{0},a)=a$. Let $\de\in(0,1-\theta)$. For $t_{0}\in\R$, $t\geq t_{0}$ and $x\in\R$, define
\begin{equation*}
\begin{split}
\om_{+}(t,x;t_{0})&=(\theta-\de)\big[1-\phi(x-x_{s}-C(t-t_{0}))\big]+u^{o}(t;t_{0},1+\de)\phi(x-x_{s}-C(t-t_{0})),\\
\om_{-}(t,x;t_{0})&=-\de\big[1-\phi(x+x_{s}+C(t-t_{0}))\big]+u^{o}(t;t_{0},\theta+\de)\phi(x+x_{s}+C(t-t_{0})),
\end{split}
\end{equation*}
where $C>0$ is some constant. Note that $u^{o}(t;t_{0},1+\de)$ and $u^{o}(t;t_{0},\theta+\de)$ are decreasing and increasing in $t$, respectively, and
$$
\lim_{t\ra\infty}u^{o}(t;t_{0},1+\de)=1=\lim_{t\ra\infty}u^{o}(t;t_{0},\theta+\de).
$$
\begin{lem}\label{lem-sub-super-sol}
For sufficiently large $C>0$, $\om_{+}(t,x;t_{0})$ and $\om_{-}(t,x;t_{0})$ are sup-solution and sub-solution of \eqref{main-eqn}, respectively.
\end{lem}
\begin{proof}
We only prove that $\om_{+}(t,x;t_{0})$ is a super-solution for sufficiently large $C$; $\om_{-}(t,x;t_{0})$ being a sub-solution for sufficiently large $C$ can be proven similarly. We compute
\begin{equation*}
\begin{split}
&(\om_{+})_{t}-(\om_{+})_{xx}-f(t,\om_{+})\\
&\quad\quad=(\theta-\de-u^{o}(t;t_{0},1+\de))\big[C\phi'(x-x_{s}-C(t-t_{0}))+\phi''(x-x_{s}-C(t-t_{0}))\big]\\
&\quad\quad\quad+f(t,u^{o}(t;t_{0},1+\de))\phi(x-x_{s}-C(t-t_{0}))-f(t,\om_{+})\\
&\quad\quad=(C-c_{\inf})(\theta-\de-u^{o}(t;t_{0},1+\de))\phi'(x-x_{s}-C(t-t_{0}))\\
&\quad\quad\quad+(u^{o}(t;t_{0},1+\de)-\theta+\de)f_{\inf}(\phi(x-x_{s}-C(t-t_{0})))\\
&\quad\quad\quad+f(t,u^{o}(t;t_{0},1+\de))\phi(x-x_{s}-C(t-t_{0}))-f(t,\om_{+})\\
&\quad\quad\geq(C-c_{\inf})(\theta-\de-u^{o}(t;t_{0},1+\de))\phi'(x-x_{s}-C(t-t_{0}))\\
&\quad\quad\quad+f(t,u^{o}(t;t_{0},1+\de))\phi(x-x_{s}-C(t-t_{0}))-f(t,\om_{+}),\\
\end{split}
\end{equation*}
where we used the equation in \eqref{traveling-homo} in the second equality.
There are two cases. If $\om_{+}\leq\theta$, then $f(t,\om_{+})=0$ and $u^{o}(t;t_{0},1+\de)\phi(x-x_{s}-C(t-t_{0}))\leq\theta$, which forces $\phi(x-x_{s}-C(t-t_{0}))\leq\theta$ and hence, $x-x_{s}-C(t-t_{0})\geq0$ by monotonicity. We then conclude from \eqref{explicit-sol} or the way $(\phi(z),\phi'(z))$ approaches $(0,0)$ as $z\ra\infty$ that $\phi(x-x_{s}-C(t-t_{0}))$ and $\phi'(x-x_{s}-C(t-t_{0}))$ are comparable, which leads to
\begin{equation}\label{sup-sol-propagation}
(\om_{+})_{t}-(\om_{+})_{xx}-f(t,\om_{+})\geq0
\end{equation}
for sufficiently large $C>0$.
If $\om_{+}>\theta$, then by Taylor expansion,
\begin{equation*}
\begin{split}
&f(t,u^{o}(t;t_{0},1+\de))\phi(x-x_{s}-C(t-t_{0}))-f(t,\om_{+})\\
&\quad\quad=[f(t,u^{o}(t;t_{0},1+\de))-f(t,\om_{+})\big]\phi(x-x_{s}-C(t-t_{0}))\\
&\quad\quad\quad+f(t,\om_{+})\big[1-\phi(x-x_{s}-C(t-t_{0}))\big]\\
&\quad\quad=f_{u}(t,u_{*})(u^{o}(t;t_{0},1+\de)-\theta+\de)\big[1-\phi(x-x_{s}-C(t-t_{0}))\big]\phi(x-x_{s}-C(t-t_{0}))\\
&\quad\quad\quad+f(t,\om_{+})\big[1-\phi(x-x_{s}-C(t-t_{0}))\big],
\end{split}
\end{equation*}
where $u_{*}\in[\om_{+},u^{o}(t;t_{0},1+\de)]$. Note that the condition $\om_{+}>\theta$ forces $x-x_{s}-C(t-t_{0})\leq x_{*}$ for some universal constant $x_{*}>0$. We then conclude from the way $(\phi(z),\phi'(z))$ approaches $(1,0)$ as $z\ra-\infty$ that $1-\phi(x-x_{s}-C(t-t_{0}))$ and $\phi'(x-x_{s}-C(t-t_{0}))$ are comparable, and hence, \eqref{sup-sol-propagation} holds as well for sufficiently large $C>0$.
\end{proof}
The next result concerns the propagation of $\xi_{\theta}(t;s)$.
\begin{prop}\label{prop-bd-propagation}
For any $T_{0}>0$, there exists $h_{0}=h_{0}(T_{0})>0$ such that
\begin{equation*}
|\xi_{\theta}(t+T_{0};s)-\xi_{\theta}(t;s)|\leq h_{0}
\end{equation*}
for all $s<0$, $t\geq s$. Moreover, $h_{0}(T_{0})$ is increasing in $T_{0}$.
\end{prop}
\begin{proof}
Let $\de_{*}=\la_{*}-\theta$, where $\la_{*}$ is as in Lemma \ref{lem-bd-width}. First, since
\begin{equation*}
\begin{split}
\om_{+}(t_{0},x;t_{0})&=(\theta-\de_{*})\big[1-\phi(x-x_{s})\big]+(1+\de_{*})\phi(x-x_{s}),\\
\om_{-}(t_{0},x;t_{0})&=-\de_{*}\big[1-\phi(x+x_{s})\big]+(\theta+\de_{*})\phi(x+x_{s}),
\end{split}
\end{equation*}
we can find some $x_{*}<0$ such that $\om_{+}(t_{0},x_{*}+x_{s};t_{0})\geq1$ and $\om_{-}(t_{0},-x_{*}-x_{s};t_{0})\leq0$,
which yields
\begin{equation*}
\om_{-}(t_{0},x-\xi_{\theta+\de_{*}}(t_{0};s)-x_{*}-x_{s};t_{0})\leq u(t_{0},x;s)\leq\om_{+}(t_{0},x-\xi_{\theta-\de_{*}}(t_{0};s)+x_{*}+x_{s};t_{0})
\end{equation*}
for all $x\in\R$. It then follows from Lemma \ref{lem-sub-super-sol} and comparison principle that
\begin{equation}\label{estimate-lower-upper-bd}
\om_{-}(t,x-\xi_{\theta+\de_{*}}(t_{0};s)-x_{*}-x_{s};t_{0})\leq u(t,x;s)\leq\om_{+}(t,x-\xi_{\theta-\de_{*}}(t_{0};s)+x_{*}+x_{s};t_{0})
\end{equation}
for all $x\in\R$ and $t\geq t_{0}$.
We now fix any $T_{0}>0$. Since
\begin{equation*}
\begin{split}
\om_{+}(t_{0}+T_{0},x+CT_{0};t_{0})&=(\theta-\de_{*})\big[1-\phi(x-x_{s})\big]+u^{o}(t_{0}+T_{0};t_{0},1+\de_{*})\phi(x-x_{s}),\\
&\leq(\theta-\de_{*})\big[1-\phi(x-x_{s})\big]+(1+\de_{*})\phi(x-x_{s}),\\
\om_{-}(t_{0}+T_{0},x-CT_{0};t_{0})&=-\de_{*}\big[1-\phi(x+x_{s})\big]+u^{o}(t_{0}+T_{0};t_{0},\theta+\de_{*})\phi(x+x_{s}),\\
&\geq-\de_{*}\big[1-\phi(x+x_{s})\big]+(\theta+\de_{*})\phi(x+x_{s}),
\end{split}
\end{equation*}
we can find some $x_{**}>0$ such that
\begin{equation*}
\begin{split}
\om_{+}(t_{0}+T_{0},x_{**}+x_{s}+CT_{0};t_{0})&\leq\theta-\frac{\de_{*}}{2},\\
\om_{-}(t_{0}+T_{0},-x_{**}-x_{s}-CT_{0};t_{0})&\geq\theta+\frac{\de_{*}}{2}.
\end{split}
\end{equation*}
This together with \eqref{estimate-lower-upper-bd} gives
\begin{equation*}
\begin{split}
u(t_{0}+T_{0},\xi_{\theta-\de_{*}}(t_{0};s)-x_{*}+x_{**}+CT_{0};s)&\leq\theta-\frac{\de_{*}}{2},\\
u(t_{0}+T_{0},\xi_{\theta+\de_{*}}(t_{0};s)+x_{*}-x_{**}-CT_{0};s)&\geq\theta+\frac{\de_{*}}{2}.
\end{split}
\end{equation*}
By monotonicity, we find
\begin{equation}\label{estimate-propagation-aux}
\begin{split}
\xi_{\theta-\frac{\de_{*}}{2}}(t_{0}+T_{0};s)&\leq\xi_{\theta-\de_{*}}(t_{0};s)-x_{*}+x_{**}+CT_{0},\\
\xi_{\theta+\frac{\de_{*}}{2}}(t_{0}+T_{0};s)&\geq\xi_{\theta+\de_{*}}(t_{0};s)+x_{*}-x_{**}-CT_{0}.
\end{split}
\end{equation}
Finally, to finish the proof, we set
$$
h_{0}=h_{0}(T_{0})=-x_{*}+x_{**}+CT_{0}+C(\theta+\de_{*},\theta-\de_{*}),
$$
where $C(\theta+\de_{*},\theta-\de_{*})>0$ is as in Proposition \ref{prop-bd-interface-width}. Then,
\begin{equation*}
\begin{split}
&\xi_{\theta}(t_{0}+T_{0};s)-\xi_{\theta}(t_{0};s)\\
&\quad\quad\leq\xi_{\theta-\frac{\de_{*}}{2}}(t_{0}+T_{0};s)-\xi_{\theta+\de_{*}}(t_{0};s)\\
&\quad\quad=\xi_{\theta-\frac{\de_{*}}{2}}(t_{0}+T_{0};s)-\xi_{\theta-\de_{*}}(t_{0};s)+\xi_{\theta-\de_{*}}(t_{0};s)-\xi_{\theta+\de_{*}}(t_{0};s)\\
&\quad\quad\leq-x_{*}+x_{**}+CT_{0}+C(\theta+\de_{*},\theta-\de_{*})=h_{0},
\end{split}
\end{equation*}
where we used the first estimate in \eqref{estimate-propagation-aux} and Proposition \ref{prop-bd-interface-width}. Similarly, by the second estimate in \eqref{estimate-propagation-aux} and Proposition \ref{prop-bd-interface-width}, we deduce $\xi_{\theta}(t_{0}+T_{0};s)-\xi_{\theta}(t_{0};s)\geq-h_{0}$. This completes the proof.
\end{proof}
Finally, we prove Theorem \ref{thm-bd-interface-width}.
\begin{proof}[Proof of Theorem \ref{thm-bd-interface-width}]
Note that in the presence of Proposition \ref{prop-bd-interface-width}, we only need to bound $\xi_{\theta}(t;s)-\xi_{\la}(t;s)$ for all $\la\in(\theta,1)$ close to $1$. To do so, let $\de_{*}=\la_{*}-\theta$ be as in the proof of Proposition \ref{prop-bd-propagation}. Recall that $u^{o}(t;t_{0},\theta+\de_{*})$ is increasing in $t$ and $\lim_{t\ra\infty}u^{o}(t;t_{0},\theta+\de_{*})=1$. From which, we can find some $T_{-}>0$ and $x_{-}<0$ such that
\begin{equation*}
\om_{-}(t,x_{-}-x_{s}-C(t-t_{0});t_{0})\geq1-\de_{*},\quad t\geq t_{0}+T_{-}.
\end{equation*}
Using the first inequality in \eqref{estimate-lower-upper-bd}, we find
\begin{equation*}
u(t,\xi_{\theta+\de_{*}}(t_{0};s)+x_{*}+x_{-}-C(t-t_{0});s)\geq1-\de_{*},\quad t\geq t_{0}+T_{-}.
\end{equation*}
By monotonicity,
\begin{equation*}
\xi_{1-\de_{*}}(t;s)\geq\xi_{\theta+\de_{*}}(t_{0};s)+x_{*}+x_{-}-C(t-t_{0}),\quad t\geq t_{0}+T_{-}.
\end{equation*}
Setting $t=t_{0}+T_{-}$ in the above estimate, we find
\begin{equation*}
\xi_{1-\de_{*}}(t_{0}+T_{-};s)\geq\xi_{\theta+\de_{*}}(t_{0};s)+x_{*}+x_{-}-CT_{-}.
\end{equation*}
Since $\xi_{\theta}(t_{0}+T_{-};s)\leq\xi_{\theta}(t_{0};s)+h_{0}(T_{-})$ by Proposition \ref{prop-bd-propagation}, we find
\begin{equation*}
\begin{split}
&\xi_{\theta}(t_{0}+T_{-};s)-\xi_{1-\de_{*}}(t_{0}+T_{-};s)\\
&\quad\quad\leq\xi_{\theta}(t_{0};s)-\xi_{\theta+\de_{*}}(t_{0};s)+h_{0}(T_{-})-x_{*}-x_{-}+CT_{-}\\
&\quad\quad\leq\ep_{*}+h_{0}(T_{-})-x_{*}-x_{-}+CT_{-}
\end{split}
\end{equation*}
by Proposition \ref{prop-bd-interface-width}, where $\ep_{*}=C(\theta,\theta+\de_{*})$. Since $t_{0}\geq s$ is arbitrary, we arrive at
\begin{equation*}
\xi_{\theta}(t;s)-\xi_{1-\de_{*}}(t;s)\leq\ep_{*}+h_{0}(T_{-})-x_{*}-x_{-}+CT_{-},\quad t\geq s+T_{-}.
\end{equation*}
For the time interval $[s,s+T_{-}]$, we consider space-time homogeneous equations
\begin{equation}\label{eqn-space-time-homo}
\begin{split}
u_{t}=u_{xx}+f_{\inf}(u),\quad u_{t}=u_{xx}+f_{\sup}(u).
\end{split}
\end{equation}
Let $u_{\inf}(t,x;s)$ and $u_{\sup}(t,x;s)$ be solutions of the first and the second equation in \eqref{eqn-space-time-homo}, respectively, with $u_{\inf}(s,x;s)=\phi(x-x_{s})=u_{\sup}(s,x;s)$. By comparison principle and homogeneity, we find
\begin{equation*}
u_{\inf}(t-s,x;0)\leq u(t,x;s)\leq u_{\sup}(t-s,x;0),\quad x\in\R,\,\,t\geq s.
\end{equation*}
Denote by $\xi_{1-\de_{*}}^{\inf}(t-s)$ be the unique point such that $u_{\inf}(t-s,\xi_{1-\de_{*}}^{\inf}(t-s);0)=1-\de_{*}$ and by $\xi_{\theta}^{\sup}(t-s)$ be the unique point such that $u_{\sup}(t-s,\xi_{\theta}^{\sup}(t-s);0)=\theta$. Then, for $t\in[s,s+T_{-}]$ we have
\begin{equation*}
-\infty<\inf_{t\in[s,s+T_{-}]}\xi_{1-\de_{*}}^{\inf}(t-s)\leq\xi_{1-\de_{*}}(t;s)<\xi_{\theta}(t;s)\leq\sup_{t\in[s,s+T_{-}]}\xi_{\theta}^{\sup}(t-s)<\infty.
\end{equation*}
Setting
\begin{equation*}
\ep_{**}=\sup_{t\in[s,s+T_{-}]}\xi_{\theta}^{\sup}(t-s)-\inf_{t\in[s,s+T_{-}]}\xi_{1-\de_{*}}^{\inf}(t-s),
\end{equation*}
we find $\xi_{\theta}(t;s)-\xi_{1-\de_{*}}(t;s)\leq\ep_{**}$ for $t\in[s,s+T_{-}]$. Thus, setting
\begin{equation*}
\ep_{***}=\max\big\{\ep_{**},\ep_{*}+h_{0}(T_{-})-x_{*}-x_{-}+CT_{-}\big\},
\end{equation*}
we have
\begin{equation}
\label{width-bd-eq}
\xi_\theta(t;s)-\xi_{1-\delta_{*}}(t;s)\le\epsilon_{***},\quad s<0,\,\, t\geq s.
\end{equation}
The theorem then follows from Proposition \ref{prop-bd-interface-width} and \eqref{width-bd-eq}.
\end{proof}
\section{Uniform Steepness Estimate}\label{Non-Flat Estimate Near Ignition Temperature}
This section is devoted to the uniform steepness of $u(t,x;s)$ near $\xi_{\theta}(t;s)$. Through this section, we assume $\rm(H1)$ and $\rm(H2)$. The main result is the following
\begin{thm}\label{thm-non-flat}
There exist a constant $T_{D}>0$ and a continuous nonincreasing function $\al:[0,\infty)\ra(0,\infty)$ such that for any $M\geq0$ there holds
\begin{equation*}
u_{x}(t,x;s)\leq-\al(M),\quad x\in[\xi_{\theta}(t;s)-M,\xi_{\theta}(t;s)+M]
\end{equation*}
for all $s<0$, $t\geq s+T_{D}$. In particular, the following statements hold.
\begin{itemize}
\item[\rm(i)] For any $\la\in(0,1)$, there is $\al_{\la}>0$ such that
\begin{equation*}
u_{x}(t,\xi_{\la}(t;s);s)\leq-\al_{\la}
\end{equation*}
for all $s<0$ and $t\geq s+T_{D}$. Moreover, the function $\la\mapsto\al_{\la}:(0,1)\ra(0,\infty)$ is continuous and bounded.
\item[\rm(ii)] For any $\la\in(0,1)$, there exists $C_{\la}>0$ such that
\begin{equation*}
\sup_{s<0,t\geq s+T_{D}}\bigg|\frac{d\xi_{\la}(t;s)}{dt}\bigg|\leq C_{\la}.
\end{equation*}
\end{itemize}
\end{thm}
The notation $T_{D}$ stands for the time delay. We understand it as the time that the approximation solutions take to adjust their shapes. The proof of Theorem \ref{thm-non-flat} depends on the boundedness of interface width as in Theorem \ref{thm-bd-interface-width} and the propagation of the interface location $\xi_{\theta}(t;s)$ as in Proposition \ref{prop-bd-propagation}. To prove Theorem \ref{thm-non-flat}, we first prove a lemma.
\begin{lem}[\cite{Sh99-1}]\label{derivative-integral-estimate}
For any $h>0$, $t\geq t_{0}\geq s$, there holds
\begin{equation*}
u_{x}(t,x;s)\leq J(t-t_{0},|x-z|+h)\int_{z-h}^{z+h}u_{x}(t_{0},y;s)dy,
\end{equation*}
where
\begin{equation*}
J(t-t_{0},|x-z|+h)=e^{-\tilde{M}(t-t_{0})}\frac{1}{\sqrt{4\pi(t-t_{0})}}e^{-\frac{(|x-z|+h)^{2}}{4(t-t_{0})}}
\end{equation*}
for some $\tilde{M}>0$.
\begin{proof}
Let $\ep>0$. Set $v_{1}(t,x;s)=u(t,x+\ep;s)$ and $v_{2}(t,x;s)=u(t,x;s)$. By monotonicity, $v_{1}(t,x;s)<v_{2}(t,x;s)$. Clearly, $v(t,x;s)=v_{1}(t,x;s)-v_{2}(t,x;s)$ satisfies
\begin{equation*}
v_{t}=v_{xx}+f(t,v_{1})-f(t,v_{2}).
\end{equation*}
By (H1), there exists $\tilde{M}>0$ such that $f(t,v_{1})-f(t,v_{2})\leq-\tilde{M}(v_{1}-v_{2})$, and hence
\begin{equation*}
v_{t}\leq v_{xx}-\tilde{M}v.
\end{equation*}
By comparison principle, we obtain for $t\geq t_{0}\geq s$
\begin{equation*}
\begin{split}
&u(t,x+\ep;s)-u(t,x;s)\\
&\quad\quad=v(t,x;s)\\
&\quad\quad\leq e^{-\tilde{M}(t-t_{0})}\int_{\R}\frac{1}{\sqrt{4\pi(t-t_{0})}}e^{-\frac{(x-y)^{2}}{4(t-t_{0})}}[u(t_{0},y+\ep;s)-u(t_{0},y;s)]dy\\
&\quad\quad\leq e^{-\tilde{M}(t-t_{0})}\int_{z-h}^{z+h}\frac{1}{\sqrt{4\pi(t-t_{0})}}e^{-\frac{(x-y)^{2}}{4(t-t_{0})}}[u(t_{0},y+\ep;s)-u(t_{0},y;s)]dy\\
&\quad\quad\leq e^{-\tilde{M}(t-t_{0})}\frac{1}{\sqrt{4\pi(t-t_{0})}}e^{-\frac{(|x-z|+h)^{2}}{4(t-t_{0})}}\int_{z-h}^{z+h}[u(t_{0},y+\ep;s)-u(t_{0},y;s)]dy,
\end{split}
\end{equation*}
which leads to the result.
\end{proof}
\end{lem}
Observe that $J(t-t_{0},|x-z|+h)\ra0$ as $t-t_{0}\ra0$, that is, the estimate given in Lemma \ref{derivative-integral-estimate} is degenerate when $t$ approaches $t_{0}$. This is the technical reason why we introduce the time delay $T_{D}$ in the statement of Theorem \ref{thm-non-flat}.
We now prove Theorem \ref{thm-non-flat}.
\begin{proof}[Proof of Theorem \ref{thm-non-flat}]
Set $h_{\theta}=\max\{C(\theta,\frac{1+\theta}{2}),C(\theta,\frac{\theta}{2})\}$, where $C(\theta,\frac{1+\theta}{2})$ and $C(\theta,\frac{\theta}{2})$ are as in Theorem \ref{thm-bd-interface-width}. Theorem \ref{thm-bd-interface-width} then ensures that for all $t_{0}\geq s$
\begin{equation}\label{an-inequality-aux-1234}
\xi_{\theta}(t_{0};s)+h_{\theta}\geq\xi_{\frac{\theta}{2}}(t_{0};s),\quad \xi_{\theta}(t_{0};s)-h_{\theta}\leq\xi_{\frac{1+\theta}{2}}(t_{0};s).
\end{equation}
Now, for any $\tau\geq0$ and $t_{0}\geq s$, applying Lemma \ref{derivative-integral-estimate} with $z=\xi_{\theta}(t_{0};s)$ and $h=h_{\theta}$, we obtain that if $|x-\xi_{\theta}(t_{0};s)|\leq M$, then
\begin{equation}\label{an-estimate-aux-1234}
\begin{split}
u_x(\tau+t_{0},x;s)&\leq J(\tau, M+h_{\theta})\int_{\xi_{\theta}(t_{0};s)-h_{\theta}}^{\xi_{\theta}(t_{0};s)+h_{\theta}}u_{x}(t_{0},y;s)dy\\
&=J(\tau, M+h_{\theta})\big[u(t_{0},\xi_{\theta}(t_{0};s)+h_{\theta};s)-u(t_{0},\xi_{\theta}(t_{0};s)-h_{\theta};s)\big]\\
&\leq J(\tau, M+h_{\theta})\big[u(t_{0},\xi_{\frac{\theta}{2}}(t_{0};s);s)-u(t_{0},\xi_{\frac{1+\theta}{2}}(t_{0};s);s)\big]\\
&=-\frac{1}{2}J(\tau,M+h_{\theta}),
\end{split}
\end{equation}
where we used \eqref{an-inequality-aux-1234} and monotonicity of $u(t_{0},x;s)$ in $x$ in the second inequality.
Finally, fix some $T_{0}$, where $T_{0}$ is as in Proposition \ref{prop-bd-propagation}. Setting $\tau=T_{0}$ in \eqref{an-estimate-aux-1234}, we find that if $|x-\xi_{\theta}(t_{0}+T_{0};s)|\leq M$, then
\begin{equation*}
|x-\xi_{\theta}(t_{0};s)|\leq|x-\xi_{\theta}(t_{0}+T_{0};s)|+|\xi_{\theta}(t_{0}+T_{0};s)-\xi_{\theta}(t_{0};s)|\leq M+h_{0}(T_{0},\de_{*})
\end{equation*}
by Proposition \ref{prop-bd-propagation}, and hence,
$$
u_x(t_{0}+T_{0},x;s)\leq-\frac{1}{2}J(T_{0},M+h_{0}(T_{0},\de_{*})+h_{\theta}).
$$
This completes the proof of the main result. For the ``in particular" part, we argue as follows.
$\rm(i)$ It is a simple consequence of the just-proven result and Theorem \ref{thm-bd-interface-width}.
$\rm(ii)$ It follows from Lemma \ref{lem-negativity}, the uniform boundedness of $u_{t}(t,\xi_{\theta}(t;s);s)$ in $t\ge s+\delta_0$ for any
$\delta_0>0$ and $\rm(i)$.
\end{proof}
\section{Uniform Decaying Estimates}\label{uniform-estimate}
In this section, we investigate the uniform-in-time estimates of $u(t,x+\xi_\theta(t;s),s)$ for $x\le 0$ (referred to as behind the interface) and $x\ge 0$ (referred to as ahead of the interface). Throughout this section, we assume $\rm(H1)$ and $\rm(H2)$.
\subsection{Uniform Decaying Estimates Behind Interface}
In this subsection, we control $u(t,x;s)$ behind the interface. The main results of this subsection are stated in the following theorem.
\begin{thm}\label{thm-uniform-estimate-hehind}
\begin{itemize}
\item[\rm(i)] There is a strictly decreasing function $v:(-\infty,0]\ra[\theta,1)$ satisfying $v(-\infty)=1$ and $v(0)=\theta$ such that
\begin{equation*}
u(t,x+\xi_{\theta}(t;s);s)\geq v(x),\quad x\leq0
\end{equation*}
for all $s<0$, $t\geq s+T_{D}$, where $T_{D}$ is given in Theorem \ref{thm-non-flat}.
\item[\rm(ii)] There exist $\la_{0}\in(\theta,1)$, $r>0$ and $\beta_{0}>0$ such that
\begin{equation*}
u(t,x+\xi_{\theta}(t;s);s)\geq1-(1-\la_{0})\Big[e^{-\be_{0}(t-s)}+e^{r(x+C(\theta,\la_{0}))}\Big],\quad x\leq-C(\theta,\la_{0})
\end{equation*}
for all $s<0$, $t\geq s+T_{D}$, where $C(\theta,\la_{0})$ is as in Theorem \ref{thm-bd-interface-width}.
\end{itemize}
\end{thm}
The first part of the theorem gives an uniform control of $u(t,x;s)$ behind the interface. The second part gives an exponential property of $1-u(t,x;s)$ behind the interface, which leads to the exponential decay behind the interface of the limiting function $1-u(t,x;s)$ as $s\ra-\infty$.
To prove Theorem \ref{thm-uniform-estimate-hehind}, we first prove a lemma giving the exponential property of $u(t,x;s)$ behind some special interface.
\begin{lem}\label{lem-hehind-key}
There exists $\la_{0}\in(\theta,1)$, $r>0$ and $\beta_{0}>0$ such that
\begin{equation*}
u(t,x+\xi_{\la_{0}}(t;s);s)\geq1-(1-\la_{0})\Big[e^{-\be_{0}(t-s)}+e^{rx}\Big],\quad x\leq0
\end{equation*}
for all $s<0$, $t\geq s+T_{D}$
\end{lem}
\begin{proof}
By $\rm(H2)$, there exist $\la_{0}\in(\theta,1)$ and $\be_{0}>0$ such that
\begin{equation}\label{condition-lower-estimate}
f(t,u)\geq\be_{0}(1-u),\quad u\in[\la_{0},1].
\end{equation}
Let $v(t,x;s)=u(t,x+\xi_{\la_{0}}(t;s);s)$. It solves
\begin{equation*}
\begin{split}
\left\{\begin{aligned}
&v_{t}=v_{xx}+\xi_{\la_{0}}'v_{x}+f(t,v),\,\,x\leq0,\,\,t\geq s+T_{D}\\
&v(t,-\infty;s)=1,\,\,v(t,0;s)=\la_{0},\,\,t\geq s+T_{D}\\
&v(s+T_{D},x;s)=u(s+T_{D},x+\xi_{\la_{0}}(s+T_{D};s);s),\,\,x\leq0
\end{aligned} \right.
\end{split}
\end{equation*}
where $\xi_{\la_{0}}'=\frac{d}{dt}\xi_{\la_{0}}(t;s)$. Since $v(t,x;s)\in[\la_{0},1)$ for $x\leq0$ and $t\geq s+T_{D}$, we conclude from \eqref{condition-lower-estimate} that
\begin{equation}\label{estimate-hehind-1}
v(t,x;s)\geq\hat{v}(t,x;s),\quad x\leq0,\,\,t\geq s+T_{D},
\end{equation}
where $\hat{v}(t,x;s)$ is the solution of
\begin{equation*}
\begin{split}
\left\{\begin{aligned}
&\hat{v}_{t}=\hat{v}_{xx}+\xi_{\la_{0}}'\hat{v}_{x}+\beta_{0}(1-\hat{v}),\,\,x\leq0,\,\,t\geq s+T_{D}\\
&\hat{v}(t,-\infty;s)=1,\,\,\hat{v}(t,0;s)=\la_{0},\,\,t\geq s+T_{D}\\
&\hat{v}(s+T_{D},x;s)=u(s+T_{D},x+\xi_{\la_{0}}(s+T_{D};s);s),\,\,x\leq0.
\end{aligned} \right.
\end{split}
\end{equation*}
Let $C_{\la_{0}}$ be as in Theorem \ref{thm-non-flat} $\rm(ii)$ and $\tilde{v}(x)$, $x\leq0$ be the solution of
\begin{equation*}
\begin{split}
\left\{\begin{aligned}
&\tilde{v}_{xx}+C_{\la_{0}}\tilde{v}_{x}+\beta_{0}(1-\tilde{v})=0,\,\,x\leq0\\
&\tilde{v}(-\infty)=1,\,\,\tilde{v}(0)=\la_{0}.
\end{aligned} \right.
\end{split}
\end{equation*}
The above problem is explicitly solvable, and we readily compute
\begin{equation}\label{estimate-hehind-2}
\tilde{v}(x)=1-(1-\la_{0})e^{rx},\quad x\leq0,
\end{equation}
where $r=\frac{-C_{\la_{0}}+\sqrt{C_{\la_{0}}^{2}+4\beta_{0}}}{2}>0$. Setting
\begin{equation}\label{estimate-hehind-3}
\bar{v}(t,x;s)=\hat{v}(t,x;s)-\tilde{v}(x),\,\,x\leq0,\,\,t\geq s+T_{D}
\end{equation}
we easily check that $\bar{v}(t,x;s)$ satisfies
\begin{equation*}
\begin{split}
\left\{\begin{aligned}
&\bar{v}_{t}\geq\bar{v}_{xx}+\xi_{\la_{0}}'\bar{v}_{x}-\beta_{0}\bar{v},\,\,x\leq0,\,\,t\geq s+T_{D}\\
&\bar{v}(t,0;s)=0,\,\,\bar{v}(t,-\infty;s)=0,\,\,t\geq s+T_{D}\\
&\bar{v}(s+T_{D},x;s)=u(s+T_{D},x+\xi_{\la_{0}}(s+T_{D};s);s)-\tilde{v}(x),\,\,x\leq0.
\end{aligned} \right.
\end{split}
\end{equation*}
Since clearly $\bar{v}(s,x;s)\geq\la_{0}-1$, we obtain that
\begin{equation}\label{estimate-hehind-4}
\bar{v}(t,x;s)\geq(\la_{0}-1)e^{-\be_{0}(t-s)},\quad x\leq0,\,\,t\geq s+T_{D}
\end{equation}
where $(\la_{0}-1)e^{-\be_{0}(t-s)}$ is a space-independent solution of $\bar{v}_{t}=\bar{v}_{xx}+\xi_{\la_{0}}'\bar{v}_{x}-\beta_{0}\bar{v}$. The result then follows from \eqref{estimate-hehind-1}, \eqref{estimate-hehind-2}, \eqref{estimate-hehind-3} and \eqref{estimate-hehind-4}.
\end{proof}
We now prove Theorem \ref{thm-uniform-estimate-hehind}.
\begin{proof}[Proof of Theorem \ref{thm-uniform-estimate-hehind}]
$\rm(i)$ For $\la\in[\theta,1)$, we define
\begin{equation*}
M_{\la}=\sup_{s<0,t\geq s+T_{D}}\big[\xi_{\theta}(t;s)-\xi_{\la}(t;s)\big].
\end{equation*}
Clearly, $M_{\la}\leq C(\theta,\la)$ by Theorem \ref{thm-bd-interface-width} with the understanding $C(\theta,\theta)=0$, $M_{\la}\ra\infty$ as $\la\ra1$ by just looking at $u(t_{0},x;s_{0})$ for some $s_{0}<0$ and $t_{0}\geq s_{0}+T_{D}$, and the map $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is nondecreasing. We show that $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is strictly increasing and continuous.
We show that $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is strictly increasing. Fix any $\la_{0}\in[\theta,1)$ and let $\{s_{n}+T_{D}\leq t_{n}\}_{n\in\N}$ be such that $\lim_{n\ra\infty}\xi_{\theta}(t_{n};s_{n})-\xi_{\la_{0}}(t_{n};s_{n})=M_{\la_{0}}$. Since $\inf_{s<0,t\geq s, x\in\R}u_{x}(t,x;s)\geq-C^{*}$ for some $C^{*}>0$ by a priori estimates parabolic equations, we find
\begin{equation}\label{estimate-upper-bd-aux}
\begin{split}
u(t_{n},x+\xi_{\theta}(t_{n};s_{n});s_{n})\leq\min\Big\{-C^{*}\big[x-(\xi_{\la_{0}}(t_{n};s_{n})-\xi_{\theta}(t_{n};s_{n}))\big]+\la_{0},1\Big\}
\end{split}
\end{equation}
for all $x\in(-\infty,\xi_{\la_{0}}(t_{n};s_{n})-\xi_{\theta}(t_{n};s_{n})]$ and all $n\in\N$.
Now, let $\la_{1}\in(\la_{0},1)$. Using \eqref{estimate-upper-bd-aux} and $\lim_{n\ra\infty}\xi_{\theta}(t_{n};s_{n})-\xi_{\la_{0}}(t_{n};s_{n})=M_{\la_{0}}$, we can find an $N$ sufficiently large such that
$\xi_{\la_{1}}(t_{N};s_{N})-\xi_{\theta}(t_{N};s_{N})\leq x_{N}$, where $x_{N}<-M_{\la_{0}}$ is such that $-(C^{*}+1)(x_{N}+M_{\la_{0}})+\la_{0}=\la_{1}$. It then follows
\begin{equation*}
M_{\la_{1}}\geq\xi_{\theta}(t_{N};s_{N})-\xi_{\la_{1}}(t_{N};s_{N})\geq-x_{N}>M_{\la_{0}}.
\end{equation*}
We show that $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is continuous. Fix any $\la_{0}\in[\theta,1)$ and let $\ep_{0}>0$ be small. By Theorem \ref{thm-non-flat}, there is $\al_{0}>0$ such that
\begin{equation*}
\sup_{s<0,t\geq s+T_{D}\atop x\in[\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)-\ep_{0},\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)+\ep_{0}]}u_{x}(t,x+\xi_{\theta}(t;s);s)\leq-\al_{0}.
\end{equation*}
It follows that for all $s<0$, $t\geq s+T_{D}$
\begin{equation}\label{continuity-right}
u(t,x+\xi_{\theta}(t;s);s)\geq-\al_{0}\big[x-(\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s))\big]+\la_{0}
\end{equation}
for $x\in[\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)-\ep_{0},\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)]$, and
\begin{equation}\label{continuity-left}
u(t,x+\xi_{\theta}(t;s);s)\leq-\al_{0}\big[x-(\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s))\big]+\la_{0}
\end{equation}
for $x\in[\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s),\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)+\ep_{0}]$.
Then, comparing \eqref{continuity-right} with the segment $-\al_{0}(x+M_{\la_{0}})+\la_{0}$ for $x\in[-M_{\la_{0}}-\ep_{0},-M_{\la_{0}}]$, we obtain for any $\la\in(\la_{0},\al_{0}\ep_{0}+\la_{0}]$, $\xi_{\la}(t;s)-\xi_{\theta}(t;s)\geq-M_{\la_{0}}-\frac{\la-\la_{0}}{\al_{0}}$ for all $s<0$, $t\geq s+T_{D}$, which together with the fact that $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is strictly increasing implies
\begin{equation*}
M_{\la_{0}}<M_{\la}\leq M_{\la_{0}}+\frac{\la-\la_{0}}{\al_{0}}\ra M_{\la_{0}}\,\,\text{as}\,\,\la\ra\la_{0}^{+}.
\end{equation*}
This show the right continuity at $\la_{0}$.
For the left continuity, for any $\la\in[\la_{0}-\frac{1}{2}\al_{0}\ep_{0},\la_{0})$, we pick a sequence $\{s_{n}+T_{D}\leq t_{n}\}_{n\in\N}$ such that
$$
\lim_{n\ra\infty}\xi_{\theta}(t_{n};s_{n})-\xi_{\la_{0}}(t_{n};s_{n})=M_{\la_{0}}.
$$
Then, comparing \eqref{continuity-left} with the segment $-\frac{1}{2}\al_{0}(x+M_{\la_{0}})+\la_{0}$ for $x\in[-M_{\la_{0}},-M_{\la_{0}}+\ep_{0}]$, we can find an $N$ sufficiently large such that
\begin{equation*}
\xi_{\la}(t_{N};s_{N})-\xi_{\theta}(t_{N};s_{N})\leq-M_{\la_{0}}-\frac{2(\la-\la_{0})}{\al_{0}},
\end{equation*}
which implies $M_{\la}\geq M_{\la_{0}}+\frac{2(\la-\la_{0})}{\al_{0}}$, and then the left continuity at $\la_{0}$.
So far, we have shown that $\la\mapsto M_{\la}:[\theta,1)\ra[0,\infty)$ is strictly increasing, continuous, and satisfies $M_{\theta}=0$ and $M_{\la}\ra+\infty$ as $\la\ra1$. We now define $v:(-\infty,0]\ra[\theta,1)$ as the inverse function of $\la\mapsto-M_{\la}$. It is easily verified that this $v$ satisfies all required properties as in the statement.
$\rm(ii)$ By Lemma \ref{lem-hehind-key}$\rm(ii)$, we have
\begin{equation*}
u(t,x+\xi_{\la_{0}}(t;s);s)\geq1-(1-\la_{0})\Big[e^{-\be_{0}(t-s)}+e^{r[x-(\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s))]}\Big]
\end{equation*}
for $x\leq\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)$. Since $\xi_{\la_{0}}(t;s)-\xi_{\theta}(t;s)\geq- C(\theta,\la_{0})$ by Theorem \ref{thm-bd-interface-width}, we arrive at the result.
\end{proof}
\subsection{Uniform Decaying Estimates Ahead of Interface}\label{subsec-estimate-ahead-interface}
In this subsection, we control $u(t,x;s)$ ahead of the interface. The main result of this subsection is stated in the following theorem.
\begin{thm}\label{thm-decaying-2}
There exist $\hat{T}_{D}>0$ and $c>0$ such that
\begin{equation*}
u(t,x+\xi_{\theta}(t;s);s)\leq\theta e^{-cx},\quad x\geq0
\end{equation*}
for all $s<0$ and $t\geq s+\hat{T}_{D}$.
\end{thm}
This theorem says that $u(t,x;s)$ decays from the interface with a uniform decaying rate. It actually contains much more information than it looks like. For example, since $u(t,\xi_{\theta}(t;s);s)=\theta$, Theorem \ref{thm-decaying-2} then implies $u_{x}(t,\xi_{\theta}(t;s);s)\leq-c\theta$, although we have obtained this information in Theorem \ref{thm-non-flat}.
To prove Theorem \ref{thm-decaying-2}, we first prove several lemmas. The first one concerns the rightward propagation of $\xi_{\theta}(t;s)$.
\begin{lem}\label{lem-propogation}
There exist $T_{*}>0$ and $h_{*}>0$ such that
\begin{equation*}
\xi_{\theta}(t+T_{*};s)-\xi_{\theta}(t;s)\geq h_{*}
\end{equation*}
for all $s<0$, $t\geq s+T_{D}$.
\end{lem}
\begin{proof}
The lemma follows from Theorem \ref{thm-bd-interface-width} and Lemma \ref{lem-rightward-prop-above-temp}.
\end{proof}
The next lemma is the driving force for the so-called sliding method (see \cite{BeNi91}).
\begin{lem}\label{lem-sliding}
Let $c\in(0,\min\{c_{\inf},\frac{h_{*}}{T_{*}}\})$ and $s<0$, where $T_{*}$ and $h_{*}$ are as in Lemma \ref{lem-propogation}. Suppose there is $t_{*}\geq s+T_{D}$ such that $u(t_{*},x+\xi_{\theta}(t_{*};s);s)\leq\theta e^{-cx}$ for $x\geq0$. Then, there exists $T(t_{*})\in(t_{*},\infty)$ such that
\begin{equation*}
u(T(t_{*}),x+\xi_{\theta}(T(t_{*});s);s)\leq\theta e^{-cx},\quad x\geq0.
\end{equation*}
Moreover, there are constants $0<c_{0}<C_{0}$ (independent of $s$ and $t_{*}$) such that
\begin{equation*}
c_{0}\leq T(t_{*})-t_{*}\leq C_{0}.
\end{equation*}
\end{lem}
\begin{proof}
Fix some $\theta_{*}\in(\theta,1)$. For $t\geq t_{*}$, define
\begin{equation}\label{aux-fun-sliding}
v(t,x;t_{*})=\theta_{*}e^{-c(x-\xi_{\theta}(t_{*};s)-c(t-t_{*}))}.
\end{equation}
Clearly, $u(t_{*},x;s)<v(t_{*},x;t_{*})$ for $x\geq\xi_{\theta}(t_{*};s)$ by assumption. By comparison principle, we have $u(t,x;s)<v(t,x;t_{*})$ for $x\geq\xi_{\theta}(t;s)$ for all $t>t_{*}$ with $t-t_{*}$ sufficiently small. In fact, since $u(t_{*},\xi_{\theta}(t_{*};s);s)<v(t_{*},\xi_{\theta}(t_{*};s);t_{*})$, continuity ensures the existence of some $t_{1}>t_{*}$ with $t_{1}-t_{*}$ small such that $u(t,\xi_{\theta}(t;s);s)<v(t,\xi_{\theta}(t;s);t_{*})$ for all $t\in[t_{*},t_{1}]$. Since $v(t,x;t_{*})$ solves $v_{t}=v_{xx}$ and $f(t,u(t,x;s))=0$ for $x\geq\xi_{\theta}(t;s)$, we conclude from the comparison principle that $u(t,x;s)<v(t,x;t_{*})$ for $x\geq\xi_{\theta}(t;s)$ for all $t\in[t_{*},t_{1}]$.
Now, we define
\begin{equation*}
T(t_{*})=\sup\big\{t\geq t_{*}\big|u(\tau,x;s)<v(\tau,x;t_{*}),\,\,x\geq\xi_{\theta}(\tau;s)\,\,\text{holds for all}\,\,\tau\in[t_{*},t)\big\}.
\end{equation*}
Clearly, $T(t_{*})>t_{*}$. Since $\phi(x-x_{s}-c_{\inf}(t-s))\leq u(t,x;s)$ and $c<c_{\inf}$, we conclude that $T(t_{*})<\infty$.
Again, since $v(t,x;t_{*})$ solves $v_{t}=v_{xx}$ and $f(t,u(t,x;s))=0$ for $x\geq\xi_{\theta}(t;s)$, we conclude from the comparison principle that, at time $T(t_{*})$, we must have
\begin{equation*}
\begin{split}
&u(T(t_{*}),x;s)\leq v(T(t_{*}),x;s),\quad x\geq\xi_{\theta}(T(t_{*});s),\\
&u(T(t_{*}),\xi_{\theta}(T(t_{*});s);s)=\theta=v(T(t_{*}),\xi_{\theta}(T(t_{*});s);t_{*}).
\end{split}
\end{equation*}
Using \eqref{aux-fun-sliding}, we readily check $u(T(t_{*}),x;s)\leq\theta e^{-c(x-\xi_{\theta}(T(t_{*});s))}$ for $x\geq\xi_{\theta}(T(t_{*});s)$.
For the ``moreover" part, let $\eta_{\theta}(t;t_{*})$ be the unique point such that $v(t,\eta_{\theta}(t;t_{*});t_{*})=\theta$. Then, $T(t_{*})$ is the first time that $\xi_{\theta}(t;s)$ hits $\eta_{\theta}(t;t_{*})$. Note that $\eta_{\theta}(t;t_{*})$ moves rightward at a constant speed $c$, that is,
\begin{equation*}
\eta_{\theta}(t;t_{*})=\eta_{\theta}(t_{*};t_{*})+c(t-t_{*})=\xi_{\theta}(t_{*};s)+\frac{1}{c}\ln\frac{\theta_{*}}{\theta}+c(t-t_{*}).
\end{equation*}
By Lemma \ref{lem-propogation}, for any $n\in\N$, $\xi_{\theta}(t_{*}+nT_{*};s)\geq \xi_{\theta}(t_{*};s)+nh_{*}$. Since $c<\frac{h_{*}}{T_{*}}$, we can find some $n_{0}$ such that
$\xi_{\theta}(t_{*}+n_{0}T_{*};s)\geq\eta_{\theta}(t_{*}+n_{0}T_{*};t_{*})$, which leads to $T(t_{*})-t_{*}\leq n_{0}T_{*}$. This establishes the upper bound.
For the lower bound, we use Theorem \ref{thm-non-flat}(ii), saying that $\xi_{\theta}(t;s)$ propagates not faster than the speed $C_{*}:=C_{\theta}\geq c_{\inf}$. Therefore, it takes, at least, $\frac{\eta_{\theta}(t_{*};t_{*})-\xi_{\theta}(t_{*};s)}{C_{*}-c}=\frac{1}{c(C_{*}-c)}\ln\frac{\theta_{*}}{\theta}$, for $\xi_{\theta}(t;s)$ to hit $\eta_{\theta}(t;t_{*})$. Thus, $T(t_{*})-t_{*}\geq\frac{1}{c(C_{*}-c)}\ln\frac{\theta_{*}}{\theta}$.
\end{proof}
We remark that the constant $c_{0}$ in the statement of Lemma \ref{lem-sliding} does depend on the choice of $c$ as in the statement of the lemma and $\theta_{*}$ as in the proof. But this will not cause any trouble, because we only need some $c\in(0,\min\{c_{\inf},\frac{h_{*}}{T_{*}}\})$ and some $\theta_{*}\in(\theta,1)$.
Lemma \ref{lem-sliding} lays the foundation for an iteration argument. To run such an argument, we need the exponential decay condition as in the lemma to hold at some initial time greater than $s+T_{D}$. This is given by
\begin{lem}\label{lem-sliding-initial}
Let $c\in(0,\min\{c_{\inf},\frac{h_{*}}{T_{*}}\})$ be small. For any $s<0$ there exists $T_{s}>0$ such that
\begin{equation*}
u(s+T_{s},x+\xi_{\theta}(s+T_{s};s);s)\leq\theta e^{-cx},\quad x\geq0.
\end{equation*}
Moreover, $T_{D}\leq T_{s}\leq\hat{C}_{0}$ for some $\hat{C}_{0}>0$.
\end{lem}
\begin{proof}
Fix some $\theta_{*}\in(\theta,1)$. By Proposition \ref{prop-bd-propagation} and \eqref{estimate-lower-bound}, there exists $h_{D}>0$ such that
\begin{equation}\label{estimate-possible-position}
\xi_{\theta}(t;s)\in[x_{s},x_{s}+h_{D}]\,\,\text{for}\,\,t\in[s,s+T_{D}].
\end{equation}
Now, for $t\geq s$, we define
\begin{equation}\label{aux-fun-sliding-intial}
v(t,x;s)=\theta_{*}e^{-c(x-x_{s}-c(t-s))}.
\end{equation}
Note that for small $c$, we can guarantee that the unique solution of the algebraic equation $\theta_{*}e^{-c(x-x_{s})}=\theta$ is greater than $x_{s}+h_{D}$. Let us denote this solution by $x_{s}+h_{D}+x_{D}$ for some $x_{D}>0$. As in Lemma \ref{lem-sliding}, let $\eta_{\theta}(t;s)$ be the unique point such that $v(t,\eta_{\theta}(t;s);s)=\theta$ and $s+T_{s}$ be the first time that $\xi_{\theta}(t;s)$ hits $\eta_{\theta}(t;s)$.
Since $\eta_{\theta}(s;s)=x_{s}+h_{D}+x_{D}$ and $\eta_{\theta}(t;s)$ moves rightward at a constant speed $c$, \eqref{estimate-possible-position} ensures $T_{s}\geq T_{D}$. On the other hand, by \eqref{explicit-sol} and \eqref{estimate-lower-bound}, we have $\xi_{\theta}(t;s)\geq x_{s}+c_{\inf}(t-s)$, which implies that it will take, at most, $\frac{h_{D}+x_{D}}{c_{\inf}-c}$, for $\xi_{\theta}(t;s)$ to hit $\eta_{\theta}(t;s)$. Thus, $T_{s}\leq\frac{h_{D}+x_{D}}{c_{\inf}-c}$.
Finally, at the first hitting time $s+T_{s}$, we have the estimate
\begin{equation*}
u(s+T_{s},x+\xi_{\theta}(s+T_{s};s);s)\leq\theta e^{-cx},\quad x\geq0.
\end{equation*}
as in the proof of Lemma \ref{lem-sliding}.
\end{proof}
Finally, we prove Theorem \ref{thm-decaying-2}.
\begin{proof}[Proof of Theorem \ref{thm-decaying-2}]
Let $c\in(0,\min\{c_{\inf},\frac{h_{*}}{T_{*}}\})$ be small such that both Lemma \ref{lem-sliding} and Lemma \ref{lem-sliding-initial} hold. By Lemma \ref{lem-sliding-initial}, we have $u(s+T_{s},x+\xi_{\theta}(s+T_{s};s);s)\leq\theta e^{-cx}$ for $x\geq0$. Since $T_{s}\geq T_{D}$ by Lemma \ref{lem-sliding-initial}, we can apply Lemma \ref{lem-sliding} to obtain that at each moment $T^{n}(s+T_{s})=\underbrace{T\circ T\circ\cdots\circ T}_{n\,\,\text{times}}(s+T_{s})$, there holds
\begin{equation*}
u(T^{n}(s+T_{s}),x+\xi_{\theta}(T^{n}(s+T_{s});s);s)\leq\theta e^{-cx},\quad x\geq0
\end{equation*}
for all $n\in\N$. Again, by Lemma \ref{lem-sliding}, $c_{0}\leq T^{n}(s+T_{s})-T^{n-1}(s+T_{s})\leq C_{0}$ for all $n\in\N$. In particular, $T^{n}(s+T_{s})\ra\infty$ as $n\ra\infty$, and $[s+T_{s},\infty)=\cup_{n\in\N}[T^{n-1}(s+T_{s}),T^{n}(s+T_{s})]$.
Next, we claim that there is $\hat{\theta}>0$ such that
\begin{equation}\label{estimate-decaying}
u(t,x+\xi_{\theta}(t;s);s)\leq\hat{\theta}e^{-cx},\quad x\geq0
\end{equation}
for all $s<0$, $t\geq s+T_{s}$. Fix any $n\in\N$ and consider the interval $[T^{n-1}(s+T_{s}),T^{n}(s+T_{s})]$. By the proof of Lemma \ref{lem-sliding}, we have for any $t\in[T^{n-1}(s+T_{s}),T^{n}(s+T_{s})]$ and $x\geq\xi_{\theta}(t;s)$,
\begin{equation*}
\begin{split}
u(t,x;s)&\leq v(t,x;T^{n-1}(s+T_{s}))\\
&=\theta_{*}e^{-c(x-\xi_{\theta}(T^{n-1}(s+T_{s});s)-c(t-T^{n-1}(s+T_{s})))}\\
&=\theta_{*}e^{-c(x-\xi_{\theta}(t;s))}e^{-c(\xi_{\theta}(t;s)-\xi_{\theta}(T^{n-1}(s+T_{s});s))}e^{c^{2}(t-T^{n-1}(s+T_{s}))}\\
&\leq\theta_{*}e^{cC_{*}(t-T^{n-1}(s))}e^{c^{2}(t-T^{n-1}(s+T_{s}))}e^{-c(x-\xi_{\theta}(t;s))}\\
&\leq\theta_{*}e^{cC_{*}C_{0}}e^{c^{2}C_{0}}e^{-c(x-\xi_{\theta}(t;s))}.
\end{split}
\end{equation*}
The claim follows with $\hat{\theta}=\theta_{*}e^{cC_{*}C_{0}}e^{c^{2}C_{0}}$, where $C_{*}=C_{\theta}$ is given in Theorem \ref{thm-non-flat}(ii).
To finish the proof, we fix some $M_{0}>0$ and set $\al_{M_{0}}=\min_{M\in[0,M_{0}]}\al(M)$, where the function $\al(\cdot)$ is given by Theorem \ref{thm-non-flat}. It then follows from the fact that $u(t,\xi_{\theta}(t;s);s)=\theta$ and Theorem \ref{thm-non-flat} that
\begin{equation*}
u(t,x+\xi_{\theta}(t;s);s)\leq-\al_{M_{0}}x+\theta,\quad x\in[0,M_{0}]
\end{equation*}
for $s<0$, $t\geq s+T_{s}$. By monotonicity, we obtain $u(t,x+\xi_{\theta}(t;s);s)\leq-\al_{M_{0}}M_{0}+\theta$ for $x\geq M_{0}$. Note that by enlarging $\hat{\theta}$ if necessary, we may assume without loss of generality that $\hat{\theta}e^{-cM_{0}}>\theta$, which implies $-\al_{M_{0}}x+\theta<\hat{\theta}e^{-cx}$ for all $x\in[0,M_{0}]$. Now, let $x_{*}>M_{0}$ be the smallest point such that $\hat{\theta}e^{-cx_{*}}=-\al_{M_{0}}x+\theta$. All these together, we obtain for $x\geq0$
\begin{equation*}
\begin{split}
u(t,x+\xi_{\theta}(t;s);s)\leq\psi_{*}(x)=\left\{\begin{aligned}
-\al_{M_{0}}x+\theta,&\quad x\in[0,M_{0}],\\
-\al_{M_{0}}M_{0}+\theta,&\quad x\in[M_{0},x_{*}],\\
\hat{\theta}e^{-cx},&\quad x\geq x_{*}.
\end{aligned} \right.
\end{split}
\end{equation*}
Set $c_{*}=\frac{1}{x_{*}}\ln\frac{\theta}{\theta-\al_{M_{0}}M_{0}}$, that is, $\theta e^{-c_{*}x_{*}}=-\al_{M_{0}}M_{0}+\theta$. By further enlarging $\hat{\theta}$ if necessary, we can make $x_{*}$ sufficiently large so that $c_{*}\leq\al_{M_{0}}$, which ensures $\phi_{*}(x)\leq\theta e^{-c_{*}x}$ for $x\geq0$. Hence, $u(t,x+\xi_{\theta}(t;s);s)\leq\theta e^{-c_{*}x}$ for $x\geq0$. The theorem then follows with $\hat{T}_{D}=\sup_{s<0}T_{s}$.
\end{proof}
\section{Transition Fronts in Time Heterogeneous Media}\label{sec-transition-wave-const}
In this section, we investigate front propagation phenomena in \eqref{main-eqn} and prove Theorem \ref{thm-transition-wave}. Throughout this section, we assume $\rm(H1)$ and $\rm(H2)$. We first present two lemmas about critical transition fronts (see Definition \ref{def-transition-wave}).
\begin{lem}[Uniqueness of critical transition fronts]
\label{lm-critical-traveling-wave1}
If $u(t,x)$ and $\tilde{u}(t,x)$ are critical transition fronts of \eqref{main-eqn}, then there is a space shift $\zeta_{0}\in\R$ such that $u(t,x+\zeta_{0})=\tilde{u}(t,x)$ for all $x\in\R$ and $t\in\R$.
\end{lem}
\begin{proof}
It follows from the arguments of \cite[Theorem A]{Sh04}.
\end{proof}
\begin{lem}[Existence of critical transition fronts]
\label{lm-critical-traveling-wave2}
If \eqref{main-eqn} admits a transition front $u(t,x)$, then it admits a critical transition front $u^{c}(t,x)$.
\end{lem}
\begin{proof}
It follows from the arguments of \cite[Theorem A]{Sh04}.
\end{proof}
We now prove Theorem \ref{thm-transition-wave}.
\begin{proof}[Proof of Theorem \ref{thm-transition-wave}]
$\rm(1)$ Let $u(t,x)$, $x\in\R$, $t\in\R$ be the global-in-time solution of \eqref{main-eqn} given in Theorem \ref{thm-entire-sol}.
We first show that there is a continuously differentiable function $\xi:\R\ra\R$ such that $u(t,\xi(t))=\theta$ for all $t\in\R$. Since
\begin{equation*}
\sup_{t<s,t\geq s+T_{D}}\bigg|\frac{d}{dt}\xi_{\theta}(t;s)\bigg|\leq C_{*}
\end{equation*}
by Theorem \ref{thm-non-flat}, we conclude from Arzel\`{a}-Ascoli theorem and the diagonal argument that $\xi_{\theta}(t;s)$ converges to $\xi(t)$ uniformly on any compact set as $s\ra-\infty$ along some subsequence. Since $u(t,\xi_{\theta}(t;s);s)=\theta$ for all $s<0$, $t\geq s$, we find $u(t,\xi(t))=\theta$ for all $t\in\R$. By Theorem \ref{thm-bd-interface-width}, $u(t,x)$ is a transition front of \eqref{main-eqn}.
$\rm(1)(i)$ Since $\lim_{x\ra-\infty}u(t,x)=1$ and $\lim_{x\ra\infty}u(t,x)=0$, $u(t,x)$ is strictly decreasing on some open set. We now fix some $t_{0}$ as an initial moment and consider the solution $u(t,x)$ for $t\geq t_{0}$. Let $y>0$. Since $u(t_{0},x+y)-u(t_{0},x)\leq0$ for all $x$ and $u(t_{0},x+y)-u(t_{0},x)<0$ on some open set, we apply maximum principle to $u(t,x+y)-u(t,x)$ to conclude that $u(t,x+y)-u(t,x)<0$ for all $x\in\R$ and $t>t_{0}$. Since $u(t,x)$ is a global-in-time solution, $u(t,x)$ is strictly decreasing in $x$ for all $t\in\R$. We then conclude $u_{x}(t,x)<0$ from Angenent's result (see \cite[Theorem B]{Ang88}) as in the proof of Lemma \ref{lem-basic-prop}.
$\rm(1)(ii)$ For continuous differentiability, we first use the limit $u_{x}(t,\xi_{\theta}(t;s);s)\ra u_{x}(t,\xi(t))$ as $s\ra\infty$ along some subsequence and Theorem \ref{thm-non-flat} to conclude that $\sup_{t\in\R}u_{x}(t,\xi(t))<0$.
The result then follows from the arguments as in the proof of Lemma \ref{lem-negativity}. In particular, we have
\begin{equation*}
\xi'(t)=-\frac{u_{t}(t,\xi(t))}{u_{x}(t,\xi(t))},\quad t\in\R.
\end{equation*}
As a byproduct, we also have $\sup_{t\in\R}|\xi'(t)|<\infty$.
$\rm(1)(iii)$ Since $u(t,x+\xi_{\theta}(t;s);s)\geq v(x)$ for all $x\leq0$, $s<0$ and $t\geq s+T_{D}$ by Theorem \ref{thm-uniform-estimate-hehind}$\rm(i)$, we have $u(t,x+\xi(t))\geq v(x)$ for $x\leq0$ and $t\in\R$. Moreover, setting $s\ra-\infty$ along some subsequence in the estimate
\begin{equation*}
u(t,x+\xi_{\theta}(t;s);s)\geq1-(1-\la_{0})\Big[e^{-\be_{0}(t-s)}+e^{r(x+C(\theta,\la_{0}))}\Big],\quad x\leq-C(\theta,\la_{0})
\end{equation*}
for $s<0$, $t\geq s+T_{D}$ given by Theorem \ref{thm-uniform-estimate-hehind}$\rm(ii)$, we conclude that
\begin{equation*}
u(t,x+\xi(t))\geq1-(1-\la_{0})e^{r(x+C(\theta,\la_{0}))},\quad x\leq-C(\theta,\la_{0}).
\end{equation*}
That is, $1-u(t,x+\xi(t))$ decays exponentially as $x\ra-\infty$ and the decay is uniform in $t\in\R$.
By Theorem \ref{thm-decaying-2}, we clearly have $u(t,\xi(t))\leq\theta e^{-cx}$ for $x\geq0$ and $t\in\R$. Thus, by setting
\begin{equation*}
\begin{split}
\hat{v}(x)=\left\{\begin{aligned}
\max\Big\{v(x),1-(1-\la_{0})e^{r(x+C(\theta,\la_{0}))}\Big\},\quad x\leq0,\\
\theta e^{-cx},\quad x\geq0,\\
\end{aligned} \right.
\end{split}
\end{equation*}
we find the function satisfying all required properties.
$\rm(2)$ By (1) and Lemma \ref{lm-critical-traveling-wave2}, \eqref{main-eqn} has a critical transition front $u^{c}(t,x)$. We prove that it must be a periodic traveling wave. Clearly, $u^{c}(\cdot+T,\cdot)$ is also a transition front. We show its criticality. Let $u$ be an arbitrary transition front. Then $u(\cdot-T,\cdot)$ is a transition front as well. Thus, for any $t\in\R$, there is a $\zeta(t)\in\R$ such that $u^{c}(t,x)\geq u(t-T,x)$ if $x\leq\zeta(t)$ and $u^{c}(t,x)\leq u(t-T,x)$ if $x\geq\zeta(t)$. Replacing $t$ by $t+T$, we find for any $t\in\R$, $u^{c}(t+T,x)\geq u(t,x)$ if $x\leq\zeta(t+T)$ and $u^{c}(t+T,x)\leq u(t,x)$ if $x\geq\zeta(t+T)$. Hence, $u^{c}(\cdot+T,\cdot)$ is critical.
By Lemma \ref{lm-critical-traveling-wave1}, there exists some $\zeta_{0}\in\R$ such that
\begin{equation}\label{equality-periodic}
u^{c}(t,x+\zeta_{0})=u^{c}(t+T,x),\quad x\in\R,\,\,t\in\R.
\end{equation}
For $t\in\R$, let $\xi_{c}(t)$ be the unique point such that $u^{c}(t,\xi_{c}(t))=\theta$. Assume, without loss of generality, that $\xi_{c}(0)=0$. Setting $t=0$ and $x=\xi_{c}(T)$ in \eqref{equality-periodic}, we find $u^{c}(0,\xi_{c}(T)+\zeta_{0})=u^{c}(T,\xi(T))=\theta$. It follows $\xi_{c}(T)+\zeta_{0}=\xi_{c}(0)=0$, and hence, $\zeta_{0}=-\xi_{c}(T)$. Thus,
\begin{equation}\label{equality-periodic-1}
u^{c}(t,x-\xi_{c}(T))=u^{c}(t+T,x),\quad x\in\R,\,\,t\in\R.
\end{equation}
Let $c_{T}=\frac{\xi_{c}(T)}{T}$. Define
\begin{equation*}
\psi(t,x)=u^{c}(t,x+c_{T}t),\quad x\in\R,\,\,t\in\R.
\end{equation*}
We check $\psi$ satisfies the properties required by a profile of a periodic traveling wave. Using \eqref{equality-periodic-1}, we have for $x\in\R$,
\begin{equation*}
\begin{split}
\psi(t+T,x)&=u^{c}(t+T,x+c_{T}(t+T))\\
&=u^{c}(t+T,x+c_{T}t+\xi_{c}(T))\\
&=u^{c}(t,x+c_{T}t)\\
&=\psi(t,x),
\end{split}
\end{equation*}
that is, $\psi(\cdot+T,\cdot)=\psi$. The uniform-in-time limit at $\pm\infty$ then follows. Since $u^{c}$ solves \eqref{main-eqn}, we readily check $\psi_{t}=\psi_{xx}+c_{T}\psi_{x}+f(t,\psi)$. In conclusion, $u^{c}$ is a periodic traveling wave.
\end{proof}
\section{Transition Fronts in Random Media}\label{sec-random-wave}
In this section, we explore front propagation phenomena in \eqref{random-eq} and prove Theorem \ref{thm-random-wave}.
\begin{proof} [Proof of Theorem \ref{thm-random-wave}]
$\rm(1)$ First of all, for any fixed $\omega\in\Omega$, by Theorem \ref{thm-transition-wave}, \eqref{random-eq} admits a transition front $u^\omega(t,x)$. Clearly, $u^\omega(t,x)$ is a wave-like solution of \eqref{random-eq} in the sense of \cite[Definition 2.3]{Sh04}. Then by \cite[Theorem A (1)]{Sh04},
\eqref{random-eq} admits a random traveling wave solution $u(t,x;\omega)$.
$\rm(2)$ By \cite[Theorem A (2)]{Sh04}, there are $\Psi^*(\cdot)\in C_{\rm unif}^b(\R,\R)$ and $c^*\in\R$ such that for a.e. $\omega\in\Omega$,
$$
\lim_{t\to\infty}\frac{\xi(t;\omega)}{t}=c^*,
$$
$$
\lim_{t\to\infty}\frac{1}{t}\int_0^ t\Psi(x,\sigma_s\omega)ds=\Psi^*(x)\quad \forall \,x\in\R,
$$
and
$$
\lim_{x\to -\infty}\Psi^*(x)=1,\quad \lim_{x\to\infty} \Psi^*(x)=0.
$$
$\rm(3)$ It follows from \cite[Theorem B (2)]{Sh04}.
\end{proof}
|
\section{Introduction}
The Quicksort algorithm invented by Hoare \cite{hoare} sorts $n$ keys
by randomly choosing a key called pivot and rearranging the array by
comparing every key to the pivot, so that
all keys less than or equal to the pivot are on its left
and all keys greater than or equal to the pivot are on its right.
The algorithm is then recursively applied to each of these two smaller arrays (which either might be
empty) till we get trivial arrays of
length $1$ or $0$. The term ``key'' can be a number, word and more generally can be an element of
a finite set, equipped with a transitive relation.
Throughout this note, we assume
that the input array is a random permutation of the positive integers $\{1, \ldots, n\}$ with all the
$n!$ permutations equally likely to be the input.
A generalisation of the algorithm is to randomly choose
$k$ pivots $i_{1}, i_{2}, \ldots, i_{k}$, where $k = 1, 2, \ldots $ and
partition the array to $(k+1)$ subarrays.
The algorithm is recursively applied to each
of the segments that contains at least $(k+1)$ keys and arrays with less than
$(k+1)$ keys are sorted by another algorithm, as insertion sort.
We point out at once that this multipivot Quicksort is a
special case of Hennequin's `generalised Quicksort' \cite{Hen}, where a
random sample of $k(t+1)-1$ keys is chosen from the array
to be sorted and the $(t+1)$-st, $2(t+1)$-th, \ldots,
$(k-1)(t+1)$-th smallest keys are used as pivots.
Obviously, for $t=0$, the array is partitioned to $k$ subarrays,
according to $k-1$ pivots. For
$k=2$, we have the `median of $(2t+1)$' Quicksort. For various multipivot variants, we
also refer the reader to the Ph.D. theses of Sedgewick \cite{sedg} and Tan \cite{tan}.
In this note, we consider the average case
analysis of multipivot Quicksort and compute the constants of integration by Vandermonde
matrices. Let $f(n, k)$ denote the expected cost when a randomly permuted array of $n$ keys is to
be sorted by the application of Quicksort on $k$ pivots.
We deliberately allow some flexibility in the form of the cost,
but a typical example might be the number of
comparisons made. We obtain the following recursive relation:
\begin{align*}
f(n, k) & = T(n, k) \\
& \quad {} + \dfrac{1}{\dbinom{n}{k}} \underbrace{\sum_{i'_{1}} \sum_{i'_{2}} \ldots
\sum_{i'_{k}}}_{i'_{1} < i'_{2} < \ldots < i'_{k}}
\Bigl ( f(i'_{1}-1, k) + f(i'_{2} - i'_{1} -1, k)
+ \ldots + f(n - i'_{k}, k ) \Bigr),
\end{align*}
where $\mathbb{E}(\tau(n, k)) = T(n, k) = \overline{a}(k)n + \overline{b}(k)$ is the average value
of a ``toll function'' $\tau(n, k)$ during the first partitioning stage.
We assume that this is a linear function of $n$.
The recursion may look a complex $k$-index summation,
but can be simplified by noting the ranges of the indices;
\begin{align*}
f(n, k)
& = T(n, k) + \dfrac{1}{\dbinom{n}{k}} \sum_{i'_{1}=1}^{n - k+1} \sum_{i'_{2}=
i'_{1}+1}^{n-k+2} \ldots \sum_{i'_{k} = i'_{k-1} + 1}^{n} \Bigl ( f(i'_{1}-1, k) + \ldots + f(n - i'_{k}, k ) \Bigr ) \\
& = T(n, k) + \dfrac{(k+1)!}{n(n-1) \ldots (n-k+1)}
\sum_{i'_{1} = 1}^{n-k+1} \dbinom{n - i'_{1}}{k-1}f(i'_{1}-1, k),
\end{align*}
since the partitioning of the array according to $k$ pivots yields $(k+1)$
segments and using the fact that the expectations of the average costs in
each segment are equal owing to the uniform distribution of the permutation.
\section{Solution of the Cauchy-Euler differential equation}
With the view of applying generating functions for the solution of our recurrence,
let $f(n, k) = a_{n}$ and consider $h(x)= \displaystyle \sum_{n=0}^{\infty}a_{n}x^{n}$;
\begin{eqnarray*}
\sum_{n=0}^{\infty}\dbinom{n}{k}a_{n}x^{n} =
\sum_{n=0}^{\infty} \dbinom{n}{k}T(n, k)x^{n} + (k+1)\sum_{n=0}^{\infty}
\left (\sum_{i'_{1} = 1}^{n} \dbinom{n - i'_{1}}{k-1}a_{i'_{1}-1} \right )x^{n}.
\end{eqnarray*}
Interchanging the order of summation and multiplying both sides
by $\left (\dfrac{x}{1-x} \right )^{-k}$,
this becomes a $k$-th order differential equation
\begin{eqnarray*}
(1-x)^{k}h^{(k)}(x) = \dfrac{k!\bigl (\overline{a}(k)(x+k)+
\overline{b}(k)(1-x) \bigr )}{(1-x)^{2}} + h(x)(k+1)!,
\end{eqnarray*}
which is a {\em Cauchy--Euler} differential equation. This type of differential
equations is inherent to the analysis of Quicksort and its variants:
we refer the reader to \cite{hc}, \cite{dur}, \cite{Hen, Hennequin} and \cite{sedg}.
Substituting $z=1-x$, we have $h(x)=g(1-x)$ and
\begin{eqnarray*}
(-1)^{k}z^{k}g^{(k)}(z)- g(z)(k+1)! =
\dfrac{k!\left (\overline{a}(k)(1-z+k)+ \overline{b}(k)z \right )}{z^{2}}.
\end{eqnarray*}
Following the analysis of Hennequin \cite{Hen, Hennequin} and Sedgewick \cite{sedg}, we
use the differential operator $\Theta$, with $\Theta g(z):=zg^{\prime}(z)$ for the
solution of the differential equation. It is easily verifiable by
induction that $\binom{\Theta}{k}g(z) =
\frac{z^{k}g^{(k)}(z)}{k!}$ and we have
\begin{eqnarray*}
\bigl ((-1)^{k}\Theta(\Theta -1) \ldots (\Theta -k + 1)-(k+1)! \bigr )g(z) =
\dfrac{k!\left (\overline{a}(k)(1-z+k)+ \overline{b}(k)z \right )}{z^{2}}.
\end{eqnarray*}
The indicial polynomial $\mathcal{P}_{k}(\Theta)$ is equal to
\begin{equation*}
\mathcal{P}_{k}(\Theta)=(-1)^{k}\Theta^{\underline{k}}-(k+1)!,
\end{equation*}
where using the notation from \cite{Gra},
$\Theta^{\underline{k}}:=\Theta(\Theta-1)\cdot \ldots \cdot (\Theta-k+1)$
denotes the falling factorial.
It can be easily proved that the polynomial has $k$ simple roots, with real parts in $[-2, k+1]$.
The solution of the differential equation is
\begin{align}
g(z) & = \dfrac{\overline{a}(k)(k+1)!}{(-2-r_{1})(-2-r_{2})
\ldots (-2-r_{k-1})}\dfrac{\ln(z)}{z^{2}} \notag \\
& \qquad {} + \dfrac{k!}{(-1-r_{1})(-1-r_{2}) \ldots 1}\dfrac{\bigl
(\overline{b}(k)-\overline{a}(k) \bigr )}{z}
+ \sum_{i=1}^{k}s_{i}z^{r_{i}}.
\label{1}
\end{align}
In order to evaluate $\mathcal{S}_{k-1}(-2)=(-2-r_{1})(-2-r_{2})
\ldots (-2-r_{k-1})$, note that
\begin{equation*}
\mathcal{S}_{k-1}(-2) = \mathcal{P}_{k}^{\prime}(-2),
\end{equation*}
thus
\begin{equation*}
\mathcal{S}_{k-1}(-2) = -(k+1)!(H_{k+1} - 1).
\end{equation*}
Moreover,
\begin{equation*}
\mathcal{P}_{k}(-1)= -kk!
\end{equation*}
and in terms of series, we have
\begin{align}
h(x) & = \dfrac{\overline{a}(k)}{H_{k+1} - 1} \sum_{n=0}^{\infty}
\bigl ( (n+1)H_{n}-n \bigr ) x^{n} +
\sum_{n=0}^{\infty} \sum_{i=1}^{k}s_{i}(-1)^{n}\dbinom{r_{i}}{n}x^{n} \notag \\
& \qquad {} +
\dfrac{\overline{a}(k)-\overline{b}(k)}{k} \sum_{n=0}^{\infty}x^{n}.
\label{2}
\end{align}
Extracting the coefficients and noting that $-2$ is the unique root with
the least real part, the expected cost of Quicksort on $k$ uniformly at random
chosen pivots is
\begin{align*}
a_{n} & = \dfrac{\overline{a}(k)}{H_{k+1} - 1} \bigl ((n+1)H_{n}-n \bigr )
+ s_{k}(n+1)+o(n).
\end{align*}
\section{Computation of the integration constants using Vandermonde matrices}
In this section, we compute the constants of integration $s_{i}$ using Vandermonde
determinants. We remark that this approach is employed in \cite{dur},
where the nine integration constants involved in the expected number
of comparisons of `remedian of $3^{2}$' Quicksort
are computed using Vandermonde matrices.
In \cite{Hen, Hennequin}, the constant
corresponding to the root $-2$ is computed by the application of generating
functions and the differential
operator (see Proposition {\bf III.8} in \cite[page 50]{Hen}). Also,
Vandermonde determinants appear in the analysis of multiple Quickselect \cite{panh}.
Our system of equations is
\begin{eqnarray*}
g(1)=g^{\prime}(1)= \ldots = g^{(k-1)}(1) = 0.
\end{eqnarray*}
Differentiating $m$ times Eq. \eqref{1} and setting $z=1$,
\begin{eqnarray}
\sum_{i=1}^{k}s_{i}r_{i}^{\underline{m}}=(-1)^{m+1}m! \biggl
(\dfrac{\overline{a}(k)\bigl(m+1)H_{m}-m)
\bigr)}{(H_{k+1}-1)}+ \dfrac{\overline{a}(k)-\overline{b}(k)}
{k} \biggr ),
\label{3}
\end{eqnarray}
for $m=0, 1, \ldots, k-1.$
In matrix form, Eq. \eqref{3} is
\begin{eqnarray*}
\centering
&& \begin{bmatrix}
1 & 1 & \ldots & 1 \\
r_{1} & r_{2} & \ldots & -2 \\
\vdots & \vdots & \ddots & \vdots \\
r^{\underline{k-1}}_{1} & r^{\underline{k-1}}_{2} & \ldots & (-2)^{\underline{k-1}} \\
\end{bmatrix}
\begin{bmatrix}
s_{1} \\
s_{2} \\
\vdots \\
s_{k}
\end{bmatrix} = \\
\centering
&& \begin{bmatrix}
& -\dfrac{1}{k}\bigl (\overline{a}(k)-\overline{b}(k) \bigr ) \\
& \dfrac{\overline{a}(k)}{H_{k+1} - 1} + \dfrac{1}{k}\bigl (\overline{a}(k)-\overline{b}(k) \bigr ) \\
& \vdots \\
& (-1)^{k}(k-1)!\biggl
(\dfrac{\overline{a}(k)\bigl (kH_{k-1}-(k-1)
\bigr)}{H_{k+1}-1} + \dfrac{\overline{a}(k)-\overline{b}(k)}
{k} \biggr )
\end{bmatrix}
\end{eqnarray*}
It is easy to see that the coefficient matrix is non-singular.
Using the generating function $x^{k-1}=\sum_{j=1}^{\infty} {k-1 \brace j-1}x^{\underline{j-1}}$ \cite{Abr},
where ${n \brace k}$ are the Stirling numbers of the second kind, with ${n \brace k}=0$ for $n<k$,
we can write each power
of $r_{i}$ as a sum of integer multiples of (earlier)
rows of the matrix of coefficients we get naturally (and that of course does not change the determinant).
Hence the determinant of this matrix is the same as the determinant of the Vandermonde matrix,
which is well known to be $\prod_{1 \leq i < j \leq k} (r_{j}-r_{i}) \neq 0$, as the roots are all simple.
Transforming the matrix into a Vandermonde one, we have
\begin{eqnarray*}
\sum_{j=1}^{\infty}(-1)^{j}(j-1)! {k-1 \brace j-1}\biggl (
\dfrac{\overline{a}(k)\bigl (jH_{j}-j
)}{H_{k+1}-1} + \dfrac{\overline{a}(k)-\overline{b}(k)}{k} \biggr ).
\end{eqnarray*}
Note that \cite{Abr},
\begin{equation*}
\sum_{j=1}^{k}(-1)^{j}(j-1)! {k-1 \brace j-1}=(-1)^{k},
\end{equation*}
since $(-1)^{j-1}(j-1)!=(-1)^{\underline{j-1}}$. Also,
\begin{equation*}
(-1)^{\underline{j}} = (-1)(-2)^{\underline{j-1}},
\end{equation*}
hence
\begin{equation*}
\sum_{j=1}^{k}(-1)^{j}j! {k-1 \brace j-1}=(-1)^{k}2^{k-1}.
\end{equation*}
Differentiating the generating function, we have
\begin{equation*}
(k-1)x^{k-2}=\sum_{j=2}^{k} {k-1\brace j-1}x^{\underline{j-1}}\left
(\sum_{i=0}^{j-2}\dfrac{1}{x-i} \right ),
\end{equation*}
therefore
\begin{align}
(-1)^{k}(k-1)2^{k-2}& = \sum_{j=2}^{k}(-1)^{j}j! {k-1 \brace j-1}\left
(\sum_{i=0}^{j-2}\dfrac{1}{i+2} \right)\notag \\
& = \sum_{j=2}^{k}(-1)^{j}j!{k-1 \brace j-1}(H_{j}-1).
\label{4}
\end{align}
Note that Eq. \eqref{4} is a special case of Eq. (36) from \cite{panh}, for $x=1$.
Our linear system now becomes:
\begin{eqnarray*}
\centering
&& \begin{bmatrix}
1 & 1 & \ldots & 1 \\
r_{1} & r_{2} & \ldots & -2 \\
\vdots & \vdots & \ddots & \vdots \\
r^{k-1}_{1} & r^{k-1}_{2} & \ldots & (-2)^{k-1} \\
\end{bmatrix}
\begin{bmatrix}
s_{1} \\
s_{2} \\
\vdots \\
s_{k}
\end{bmatrix} = \\
\centering
&& \begin{bmatrix}
& -\dfrac{1}{k}\bigl (\overline{a}(k)-\overline{b}(k) \bigr ) \\
& \dfrac{\overline{a}(k)}{H_{k+1} - 1} + \dfrac{1}{k}\bigl (\overline{a}(k)-\overline{b}(k) \bigr ) \\
& \vdots \\
& (-1)^{k} \left ((k-1)2^{k-2}
\dfrac{\overline{a}(k)}{H_{k+1}-1} + \dfrac{\overline{a}(k)-\overline{b}(k)}
{k} \right )
\end{bmatrix}
\end{eqnarray*}
The inverse matrix can be factored into a product of an upper and lower
triangular matrices, \cite{Hou}, \cite{Turner}.
In \cite{Hou} an algorithm is presented, where the entries
of the triangular matrices are recursively computed.
Letting ${\bf A}^{-1}$ denote the inverse, we have
\begin{eqnarray*}
{\bf A}^{-1} =
\begin{bmatrix}
1 & \frac{1}{r_{1}-r_{2}} & \frac{1}{(r_{1}-r_{2})(r_{1}-r_{3})} & \ldots \\
0 & \frac{1}{r_{2}-r_{1}} & \frac{1}{(r_{2}-r_{1})(r_{2}-r_{3})} & \ldots \\
0 & 0 & \frac{1}{(r_{3}-r_{1})(r_{3}-r_{2})} & \ldots \\
0 & 0 & 0 & \ldots \\
\vdots & \vdots & \vdots & \ldots
\end{bmatrix}
\begin{bmatrix}
1 & 0 & 0 & \ldots \\
-r_{1} & 1 & 0 & \ldots \\
r_{1}r_{2} & -(r_{1}+r_{2}) & 1 & \ldots \\
-r_{1}r_{2}r_{3} & r_{1}r_{2}+r_{1}r_{3}+r_{2}r_{3} & -(r_{1}+r_{2}+r_{3}) & \ldots \\
\vdots & \vdots & \vdots & \ldots
\end{bmatrix}
\end{eqnarray*}
The constants of integration are given by:
\begin{eqnarray*}
s_{i} =(-1)^{k}\left (\dfrac{\displaystyle \prod_{j \neq i}^{k}
(r_{j}+1)}{\displaystyle \prod_{j \neq i}^{k}(r_{i} - r_{j})}
\biggl( \dfrac{\overline{a}(k)-\overline{b}(k)}{k} \biggr)+ \dfrac{\overline{a}(k)}{H_{k+1}-1}
\dfrac{
\displaystyle \sum_{j=1}^{k-1}\biggl(\prod_{l \neq j}^{k-1}(r_{l}+2)\biggr)}
{\displaystyle \prod_{j \neq i}^{k}(r_{i} - r_{j})}
\right ).
\end{eqnarray*}
The products of pairwise differences of roots $r_{i}$ and $r_{j}$ that naturally arise in {\bf LU} triangular decomposition of
the inverse of Vandermonde matrix form alternating polynomial functions. Putting $i=k$ to the
previous equation,
\begin{equation*}
\dfrac{\displaystyle \prod_{j=1}^{k-1}(r_{j}+1)}{\displaystyle \prod_{j=1}^{k-1}(-2-r_{j})}
=(-1)^{k-1}\dfrac{\mathcal{P}_{k}(-1)}{\mathcal{S}_{k-1}(-2)}=(-1)^{k-1}\dfrac{k}{(k+1)(H_{k+1}-1)}
\end{equation*}
and the sum of the products is
\begin{equation*}
\displaystyle \sum_{j=1}^{k-1}\biggl(\prod_{l \neq j}^{k-1}(r_{l}+2)\biggr)=\mathcal{S}^{\prime}_{k-1}(-2).
\end{equation*}
Differentiating $\mathcal{P}_{k}(\Theta)$ twice and setting $\Theta=-2$,
\begin{equation*}
\mathcal{S}^{\prime}_{k-1}(-2)=\dfrac{\mathcal{P}_{k}^{\prime \prime}(-2)}{2}=\dfrac{(k+1)!(H^{2}_{k+1}-2H_{k+1}
-H^{(2)}_{k+1}+2)}{2},
\end{equation*}
where $H^{(2)}_{k+1}:=\displaystyle \sum_{j=1}^{k+1}\frac{1}{j^{2}}$ denotes the second--order
harmonic number.
The main result of this paper is the following Theorem:
\begin{Theorem}
The expected cost of multipivot Quicksort on $k$ uniformly at random selected pivots
for partitioning an array consisting of $n > k$ distinct keys to
subarrays that each one contains at most $k$ keys is
\begin{align*}
& \dfrac{\overline{a}(k)}{H_{k+1} - 1} \bigl ((n+1)H_{n}-n \bigr )
- \Biggl (\dfrac{\overline{a}(k)}{H_{k+1}-1} \biggl
(\dfrac{H^{2}_{k+1}-2H_{k+1}-H^{(2)}_{k+1}+2}{2(H_{k+1}-1)} \biggr) \\
& \qquad{}+ \dfrac{\overline{a}(k)-\overline{b}(k)}{(k+1)(H_{k+1}-1)} \Biggr )(n+1)
+ o(n),
\end{align*}
where the ``toll function'' has the average value $\overline{a}(k)n+\overline{b}(k)$.
\end{Theorem}
\section*{Acknowledgements}
I thank Dr. David B. Penman for his advice and helpful suggestions regarding this paper.
|
\section{Introduction}
\label{Sec:Introduction}
The $N$-vector model with $O(N)$ group symmetry plays an important role in the understanding of crucial aspects of renormalization group (RG) flows: In four dimensions, it exhibits a Landau pole and corresponds to a trivial theory \cite{Aizenman:1981du,Frohlich:1982tw,Luscher:1987ay,*Luscher:1987ek,*Luscher:1988uq,Rosten:2008ts}. In other words, as an interacting model it is only valid over a finite range of scales, thus constituting an effective low-energy theory. This could affect the possible range of validity of the standard model of particle physics \cite{Dashen:1983ts,Callaway:1983zd,Kuti:1987nr,Luscher:1988gc,Kogut:1988sf,Gockeler:1997dn} and could also play a role in cosmology as, e.g., many inflationary models probably share this feature. On the other hand, in three dimensions the theory exhibits an important example of an interacting RG fixed point \cite{LeGuillou:1977ju,Guida:1998bx}. Such fixed points are crucial in the understanding of scaling and universality in critical phenomena \cite{Wilson:1973jj,Wegner:1976bk,Pelissetto:2000ek} and, more recently, they have been of considerable interest, e.g., in the problem of the ultraviolet (UV) completion of gravity \cite{Weinberg:1980gg, Reuter:2012id}. On a more technical level, well-known examples such as, e.g., the infrared (IR) attractive Wilson-Fisher fixed point (FP) in the $O(N)$ model may provide an important benchmark test for nonperturbative methods, which one may then apply to other problems of interest (see, e.g., Ref.\ \cite{Delamotte:2004zg}).
Extending the $O(N_1)$ vector model by a coupling to another $O(N_2)$ symmetric vector field leads to complex dynamics that has been discussed extensively in the context of multicritical phenomena and systems with competing order parameters \cite{Fisher:1974zz,Kosterlitz:1976zza,Aharony:2002,Aharony:2002zz,Calabrese:2002bm,Folk:2008mi}. Such a theory is characterized by an $O(N_{1})\oplus O(N_{2})$ symmetry which admits a number of interacting (IR attractive) FPs. These travel through the coupling space of the model as the numbers of field components $N_{1}$ and $N_{2}$ are varied. At particular values of $N_1$ and $N_2$, two of these FPs can collide and exchange their stability properties, cf.\ Fig.\ \ref{Fig:collision}. In this context, an IR stable FP is defined as featuring only two positive critical exponents, as this corresponds to the number of relevant couplings that need to be tuned in order to approach the FP. When two FPs collide, a FP with three positive critical exponents trades one of them for a negative exponent, while the second FP picks up the additional relevant direction and becomes unstable. As a consequence, it turns out that for every combination of $N_1$ and $N_2$ there is exactly one stable FP. Of course, this statement assumes that one considers renormalization group trajectories within a single domain of attraction. In general, the parameter/coupling space of the model will allow for separate domains, where different FPs might exist, and may or may not be stable.
\begin{figure}[!h]
\includegraphics[width=0.7\linewidth]{collisionsketch.pdf}
\caption{\label{Fig:collision}We show a sketch of the value of the third largest critical exponent, $\theta_3$, of three different FPs (solid, dotted, and dashed lines) in the $O(N_1)\oplus O(N_2)$ model at fixed $N_1$ as a function of $N_2$. The regime where a FP is stable is indicated by the labels FP1, FP2 and FP3. As the coordinates of two different FPs coincide at $N_2^{\ast}$ and at $N_2^{\ast\ast}$, these FPs exchange their stability properties, and $\theta_{3}$ changes its sign if evaluated beyond the stable regime. An explicit calculation showing this situation can be found in Ref.\ \cite{Eichhorn:2013zza}.}
\end{figure}
In this work, we will for the first time provide a comprehensive analysis of a model with a coupling to an additional field, with a resulting $O(N_1) \oplus O(N_2) \oplus O(N_3)$ symmetry. Note that the $O(N_1) \oplus O(N_2)$ symmetry acts trivially on the third vector, and similarly the first two fields transform as singlets under the $O(N_3)$ symmetry. At first, one might expect that this model will exhibit very similar behavior to that already encountered in the case of the two-field models, and will feature a single stable FP with three relevant directions at every value of $(N_1, N_2, N_3)$. In this study, we present evidence for a rather different behavior, where FPs exhibit a large number of relevant eigendirections in a given range of values for the number of field components. This leads to the absence of stable FP solutions in a part of the parameter space. In the following, we will argue that this is a generic feature of multifield models and is due to a significantly increased number of possible mixed interactions compared to the single-field or two-field models. This behavior is akin to the absence of FPs in the low-energy effective models for phases of strongly-interacting matter \cite{Basile:2005hw,Vicari:2007ma}, or frustrated spin systems \cite{Pelissetto:2000ne,Tissier:2002zz}. In both cases, one observes the absence of stable FP solutions beyond some critical number of field components, indicative of a first order phase transition (see, e.g., Ref.\ \cite{Bak:1976}).
The presence of competing orders can change the nature of a continuous transition or even drive it to be first order. Although there are numerous examples in the literature, this is probably best illustrated using the example of the two-field model \cite{Fisher:1974zz,Kosterlitz:1976zza,Aharony:2002,Aharony:2002zz,Calabrese:2002bm,Folk:2008mi}. Already at the mean-field level one observes a suppression of the coexistence region in the presence of a strong repulsive interaction between the two competing fields \cite{She:2010}. If the coupling assumes a critical value, the coexistence region vanishes and the second-order lines merge to a first order transition. This dramatic change in the phase diagram marks a change in the universality class of the adjacent multicritical point. Fluctuations will favor either one of these scenarios (corresponding to a tetracritical or bicritical point) as long as the associated fixed point is within its domain of attraction.\footnote{This is not the case if the interactions are sufficiently strong and the total number of field components $N = N_{1} + N_{2}$ is larger than some critical value. Then, instead of a multicritical point one observes a genuine first order transition in the phase diagram of the model.} Multifield models feature a different possibility: fluctuation-mediated interactions might not only affect the universality class of the multicritical point, but they might even render it unstable, thus allowing no IR convergent RG trajectory. The difference between two- and multifield models lies in the distinct RG flow topologies and mechanisms of stability trading between different fixed points, which leads to the absence of a stable fixed point. We will investigate these properties in detail in the following sections, as well as the general behavior of systems with a large number of interacting sectors.
The main motivation for this study is to gain an improved understanding of generalized Wilson-Fisher universality classes (multicritical points) and to understand how these might show up in the phase diagram of systems with multiple order parameters. Previous work, in the context of systems with two competing order parameters, has led to the understanding that the effect of fluctuations plays an important role in addressing the stability of a particular continuous transition \cite{Fisher:1974zz,Kosterlitz:1976zza,Aharony:2002,Aharony:2002zz,Calabrese:2002bm,Folk:2008mi}. The $SO(5)$ theory of high-$T_c$ superconductivity \cite{Calabrese:2002bm,Demler:2004} provides one example, where fluctuations of the order parameters are seen to alter the stability properties of the associated fixed point and rule out such a theory as an effective IR description in the region where both order parameters become critical. On the other hand, it has been pointed out that the interplay of two competing order parameters might explain the presence of first order transitions or spatially inhomogeneous phases that exhibit finite wavevector ordering near quantum criticality \cite{She:2010}. Here, we argue that first order transition might be a generic scenario for systems that feature a large number of competing phases.
The outline of this paper is as follows: In Sec.~\ref{Sec:Model} we present the model under consideration in detail. In Sec.~\ref{Sec:Fixed-point analysis} we then explain the results of our study, discussing numerical results and scaling relations for several different FPs. In the Appendix we present the renormalization group flow equations for these models in $d$ dimensions, both in a local potential approximation (LPA) and including anomalous dimensions. Sections \ref{Sec:Model} and \ref{Sec:Fixed-point analysis} are self-contained, and can be read without referring to the technical details of our study.
\section{Model}
\label{Sec:Model}
We consider a model with three different bosonic fields, $\phi_1,\phi_2$, and $\phi_3$, with $N_1,N_2$, and $N_3$ field components, respectively. We derive the $\beta$ functions from the nonperturbative functional flow equation for the (Euclidean) scale dependent effective action $\Gamma_{k}$ \cite{Wetterich:1992yh}, (see Appendix for details, and reviews, e.g., Ref.\ \cite{Berges:2000ew,Polonyi:2001se,Pawlowski:2005xe,Gies:2006wv,Delamotte:2007pf,Rosten:2010vm,Metzner:2011cw}). This method has been shown to yield results in very good agreement with those obtained from the $\epsilon$-expansion and lattice simulations in the case of the $O(N)$ Wilson-Fisher FP, see, e.g., Ref.\ \cite{Canet:2003qd, Bervillier:2007rc,Litim:2010tt,Codello:2012ec} and the $O(N_1)\oplus O(N_2)$ FPs \cite{Eichhorn:2013zza}. To leading order in the derivative expansion \cite{Morris:1994au, *Morris:1994ie, *Morris:1996kn, *Morris:1996xq} our \textit{ansatz} for $\Gamma_{k}$ reads
\begin{eqnarray}
\Gamma_{k} &=& \int\! d^dx \left( \sum_{I=1}^3 Z_{I} \left( \partial_{\mu} \phi_{I} \right)^{2} + U_k (\phi_{1}, \phi_{2}, \phi_{3})\right) ~,
\label{Eq:EffectiveAction}
\end{eqnarray}
where $\phi_{I}^{a}$, $a = 1, \ldots, N_{I}$, and $\phi_{I}^{2} \equiv \phi_{I}^{a} \phi_{I}^{a}$. Here, we have introduced the scale dependent effective potential
\begin{equation}
\hspace{-5pt} U_k = \sum_{l,m,n} \frac{\bar{\lambda}_{l,m,n}}{l!\, m!\, n!} \left( \bar{\rho}_{I} - \bar{\kappa}_{1} \right)^{l} \left( \bar{\rho}_{2} - \bar{\kappa}_{2} \right)^{m} \left( \bar{\rho}_{3} - \bar{\kappa}_{3} \right)^n ,
\label{Eq:EffectivePotential}
\end{equation}
which we have written in terms of the invariants $\bar{\rho}_{I} = \frac{1}{2} \phi_{I}^{2}$, thereby making the $O(N_1) \oplus O(N_2) \oplus O(N_3)$ symmetry manifest. The parameter $k$ defines an infrared momentum cutoff scale, on which the parameters and couplings depend. For brevity we do not indicate the scale dependence explicitly, i.e., $\bar{\lambda}_{l,m,n} = \bar{\lambda}_{l,m,n} (k)$. Similarly, scale dependent wavefunction renormalization factors are simply denoted by $Z_{I}$. We expand the scale dependent effective potential $U_{k}$ around (possibly) nonvanishing scale dependent minima for the fields, $\bar\kappa_{I}$.
For the identification of scaling solutions, we introduce dimensionless renormalized couplings, given by
\begin{eqnarray}
u_k &=& U_k k^{-d} ~, \nonumber\\
\kappa_{I} &=& Z_I k^{2-d}\bar{\kappa}_{I} ~, \nonumber\\
\rho_{I} &=& {Z_I k^{2-d}} \bar{\rho}_{I} ~,\nonumber\\
\lambda_{l,m,n} &=& \bar{\lambda}_{l,m,n} Z_{1}^{-l}\, Z_{2}^{-m}\, Z_{3}^{-n} k^{-d + (l+m+n) (d-2)} ~.
\end{eqnarray}
We truncate the coupling space to a finite-dimensional subspace of the form Eqs.\ \eqref{Eq:EffectiveAction} and \eqref{Eq:EffectivePotential}, which includes all relevant operators, i.e., those with a positive critical exponent at the FP of interest. Including field monomials up to order $4$, $6$, and $8$, defines the local potential approximation, LPA 4/$4+\eta$, LPA 6/$6+\eta$, and the LPA 8/$8+\eta$, respectively (depending on the inclusion of a scale dependent wavefunction renormalization, $\partial_{t} Z_{I} \neq 0$).
In order to distinguish physically meaningful from spurious FPs arising within a given truncation, we demand that a FP can be continued to higher orders in the truncation, and universal quantities, e.g., critical exponents, show signs of convergence. Further, corrections to canonical scaling should not be too large, as otherwise we would not expect our truncation to be reliable. Moreover, we demand that all eigenvalues of $\left. \left(\frac{\partial^2 u_k}{\partial \rho_I \partial \rho_J} \right)\right|_{\rho_I = \kappa_I}$ are non-negative. If this condition is violated, the expansion point for the effective potential does not correspond to its true minimum, and critical exponents evaluated around this point will show poor convergence properties. The parameter $\Delta \equiv \left. \det \left(\frac{\partial^2 u_k}{\partial \rho_I \partial \rho_J} \right) \right|_{\rho_I = \kappa_I}$ serves to separate the space of couplings into different (not necessarily bounded) domains of attraction. Within such a domain there exists at most \textit{one} IR stable FP, characterized by the strength of correlations \cite{Vicari:2006xr}. In the following, we will be interested specifically in IR scaling solutions in the $\Delta \geq 0$ domain, corresponding to a minimum of the effective potential.
With these preliminaries and definitions we now turn to analyze the fixed-point structure of this model.
\section{Fixed-point analysis}
\label{Sec:Fixed-point analysis}
For generic multifield models with $\bigoplus_{I} O(N_{I})$ symmetry, a number of FPs and their stability properties can be deduced from the existence of the $O(N)$ Wilson-Fisher FP. These FPs are typically characterized by an enhancement of symmetry.
\begin{itemize}[leftmargin=*]
\item[] The \textit{isotropic fixed point (IFP)} shows maximal symmetry enhancement: All couplings at a given order in the fields take the same value, i.e., in the three-field model we have $\left. \lambda_{l,m,n} \right|_{l+m+n = 2} \equiv \lambda_{2}$, and similarly for higher order couplings. It is characterized by $O(N)$ symmetry, where $N \equiv \sum N_{I}$. Accordingly, it features additional massless Goldstone modes, even in the case of an underlying discrete symmetry, e.g., with $Z_{2} \oplus Z_{2} \oplus Z_{2}$ symmetry.
\item[] The \textit{decoupled fixed point (DFP)} is characterized by vanishing couplings between different sectors of the theory. In a model with three fields this implies $\lambda_{l,m,n} = 0$ if $l,m \neq 0$, $l,n \neq 0$, or $m,n \neq 0$. The values of the couplings in each sector approach those of the corresponding $O(N_{I})$ Wilson-Fisher FP. However, while the action at that FP is fully decoupled, critical exponents that relate to mixed couplings are nontrivial.
\item[] The \textit{decoupled isotropic fixed point (DIFP)} occurs for the first time in a model with three fields: It is characterized by a partial enhancement of symmetry, as two fields remain fully coupled and the couplings in those sectors become degenerate. Simultaneously, the third field decouples completely and its couplings approach the corresponding values of the Wilson-Fisher FP. There exist three realizations of this FP, as any of the three sectors can be the one to decouple. For generic multifield models, a set of different DIFPs exists, where any number of the fields decouple, and the couplings in the remaining sectors show a symmetry enhancement.
\item[] We may additionally infer the existence of another class of FPs from the knowledge of the anisotropic scaling solution in the two-field model with $O(N_{1})\oplus O(N_{2})$ symmetry. In general, any FP of the two-field model can be extended to the three-field model as a partially decoupled FP, where the third field decouples from the other two and the fixed-point values of its couplings are given by those of the Wilson-Fisher fixed point. In particular, this applies to the \textit{biconical} fixed-point solution in the two-field model, which we identify as the \textit{decoupled biconical fixed point (DBFP)}.
\end{itemize}
In analogy to the two-field case, we will refer to a FP in the three-field model as stable, if it features three relevant directions. This terminology relates to the requirement that these three parameters need to be tuned to reach the multicritical point.
Our model, cf.\ Eqs.\ \eqref{Eq:EffectiveAction} and \eqref{Eq:EffectivePotential}, contains nine running couplings in the potential when we take into account all operators up to fourth order in the fields, that is, the associated parameters $\kappa_I$ and couplings $\lambda_{l,m,n}$ with $l + m + n = 2$. They give rise to the nine largest critical exponents of the model\footnote{These can be calculated from the stability matrix
\begin{equation}
\mathit{\Theta}_{i,j} = \left. \frac{\partial \beta_{g_i}}{\partial g_j} \right|_{\textrm{FP}} ~.
\end{equation}
Here, the $g_i$ label all the (dimensionless renormalized) running couplings/parameters and $\beta_{g_i}$ define the corresponding beta functions. The critical exponents are then given by:
\begin{equation}
\theta_{i} ~ \in - {\rm spec} (\mathit{\Theta}) ~.
\end{equation}
}.
Going to higher orders in the expansion of the effective potential, the number of running couplings $\lambda_{l,m,n}$ increases. Accordingly, the number of critical exponents will increase, but those subleading critical exponents will be irrelevant. The leading order critical exponents will typically receive corrections from the additional higher-order couplings and will therefore vary with the order of the truncation.
\subsection{Isotropic fixed point}
\label{SubSec:Isotropic fixed point}
To determine the critical exponents at the IFP in multifield models, it is crucial to realize that a subset of those is determined by the $O(N)$ Wilson-Fisher exponents. These correspond to the directions in theory space that respect that full symmetry, i.e., those directions that span the Wilson-Fisher theory space. Additional directions in the full theory space break (a subgroup) of the enhanced $O(N)$ symmetry, and their associated critical exponents are therefore not associated with the Wilson-Fisher critical exponents, see Fig.~\ref{IFP_illustration}.
\begin{figure}[!t]
\includegraphics[width=0.7\linewidth, height=120pt, clip=true, trim= 10cm 8cm 0cm 0cm]{illustration_symmetry.pdf}
\caption{\label{IFP_illustration}This illustration of the two-dimensional subspace of the two-field coupling space shows that the Wilson-Fisher theory space is a one-dimensional subspace, corresponding to the $\lambda_{2,0} = \lambda_{0,2}$ line. The critical exponent along this direction corresponds to the largest Wilson-Fisher critical exponent. Another eigendirection of the stability matrix does not respect the enhanced symmetry of the FP.}
\end{figure}
It accordingly follows, that of the nine largest critical exponents at the IFP, two are determined by the scaling exponents of the $O(N)$ Wilson-Fisher FP. Among the additional critical exponents, one can observe a degeneracy, which can be understood from the following considerations:
\begin{table*}[!t]
\renewcommand{\arraystretch}{1.2}
\renewcommand{\tabcolsep}{6pt}
\begin{tabular}{c | c c c c c}
$N = \sum N_{I}$ & $ \theta_{1} = \theta_{2} = y_{2,2}$ & $\theta_{3} = \frac{1}{\nu}$ & $\theta_{4} = \theta_{5} = \theta_{6} = y_{4,4}$ & $\theta_{7} = \theta_{8} = y_{4,2}$ & $\theta_{9} = -\omega$ \\
\hline\hline
3 & 1.790 & 1.362 & 0.086 & -0.380 & -0.756 \\
4 & 1.818 & 1.292 & 0.196 & -0.324 & -0.775 \\
5 & 1.842 & 1.240 & 0.289 & -0.283 &-0.797 \\
\hline\hline
\end{tabular}
\caption{\label{Tab:IFP}Critical indices for the IFP in LPA 12 including anomalous dimensions. Our notation corresponds to the one introduced in Ref.\ \cite{Eichhorn:2013zza}.}
\end{table*}
Sufficiently close to $d = 4$, i.e., in the vicinity of the noninteracting fixed point, the relevant perturbations at the Wilson-Fisher FP are determined by the spin-$l$ representations of the $O(N)$ symmetry group \cite{Wegner:1972zz} (see also Ref.\ \cite{Calabrese:2002bm}). Here, we will assume that such a classification of perturbations also holds for arbitrary dimensions, and only operators up to quartic order need to be taken into account. Defining the $N$-component field $\Phi=(\phi_1,\phi_2,\phi_3)$, we find the following eigendirections of the stability matrix at the IFP in the three-field model:
\begin{itemize}[leftmargin=*]
\item[] A scalar quadratic perturbation at the IFP $\sim m^{2} \Phi^{2}$, where $P_{2,0} = \Phi^2$, defines the critical exponent $\nu$ related to the divergence of the correlation length, i.e., $[m^{2}] = \frac{1}{\nu}$. This critical exponent is thus always positive, corresponding to one relevant direction.
\item[] From the quadratic perturbation $P^{a b}_{2,2} = \Phi^{a} \Phi^{b} - \frac{1}{N} \delta^{a b} \Phi^{2}$ in the spin-$1$ representation of the $O(N)$ symmetry group, we can construct an $O(N)$ invariant operator by a suitable contraction of indices where, e.g., $\mathcal{P}_{2,2} = \phi_{I}^{2} - \frac{N_{I}}{N} \Phi^{2}$, $I = 1,2,3$. Then, the perturbation $\sim v \mathcal{P}_{2,2} $ defines the critical exponent $y_{2,2} = [v] = d - [\mathcal{P}_{2,2}]$. For the three-field model, two independent operators of that form can be constructed. Thus the corresponding critical exponent shows a two-fold degeneracy in the scaling spectrum. We emphasize that these critical exponents are identical to those evaluated for the $O(N)$ symmetric IFP in the two-field or anisotropic $N$-vector models, see, e.g., Ref.\ \cite{Calabrese:2002bm, Folk:2008mi, 2011PhRvB..84l5136H, Eichhorn:2013zza}. These exponents are always positive, adding two further relevant directions at the FP.
\item[] A scalar quartic perturbation $\sim u \Phi^{4}$, where $P_{4,0} = \Phi^{4}$, which is irrelevant at the IFP and defines the Wegner critical exponent $\omega$, yields a negative critical exponent, i.e., $[u] = y_{4,0} = - \omega$.
\item[] A quartic operator in the spin-$1$ representation of the $O(N)$ symmetry group: $P_{4,2}^{a b} = \Phi^{2} P_{2,2}^{a b}$ can be contracted to define the exponent $y_{4,2} = d - [\mathcal{P}_{4,2}]$, which is also given by the value calculated in the two-field model and shows a two-fold degeneracy in the three-field case.
\item[] A quartic perturbation in the spin-$2$ representation of the $O(N)$ symmetry group is given by
\begin{eqnarray}
\hspace{8pt} P_{4,4}^{a b c d} &=& \Phi^{a} \Phi^{b} \Phi^{c} \Phi^{d} \nonumber - \frac{1}{N + 4} \Phi^{2} \left( \Phi^{a} \Phi^{b} \delta^{c d} + p(a,b,c,d) \right) \nonumber \\ && +\: \frac{1}{(N + 2) (N + 4)} ( \Phi^{2} )^{2} \left( \delta^{a b} \delta^{c d} + p(a,b,c,d) \right) ~. \nonumber\\ &&
\end{eqnarray}
Note, that $p(a,b,c,d)$ denotes inequivalent permutation of the indices on the preceding operator, e.g., $\Phi^{a} \Phi^{b} \delta^{c d} \rightarrow \Phi^{c} \Phi^{d} \delta^{a b} + \ldots$ or $\delta^{a b} \delta^{c d} \rightarrow \delta^{a c} \delta^{b d} + \ldots$ The corresponding perturbation defines the critical exponent $y_{4,4} = d - [\mathcal{P}_{4,4}]$, which becomes negative for $N = 3$, cf.\ Fig.\ \ref{Fig:IFP}. This critical exponent with three-fold degeneracy is again determined by the two-field model, see, e.g., Ref.\ \cite{Calabrese:2002bm, Eichhorn:2013zza}.
\end{itemize}
\begin{figure}[!t]
\includegraphics[width=0.8\linewidth]{IFP_N1.pdf}\\
\includegraphics[width=0.8\linewidth]{IFP_N.pdf}
\caption{\label{Fig:IFP}(Color online) We show the fourth-largest critical exponent at the IFP as a function of $N_1$, with $N_2=N_3=1$ (upper panel) and as a function of $N_{1} = N_{2} = N_{3}$ (lower panel), with the LPA 4 result (green circles), LPA 6 (blue squares) and LPA 8 (red diamonds). The LPA clearly converges rapidly. As this exponent is positive beyond the LPA 4 and shows a three-fold degeneracy, the IFP is characterized by a total of six relevant directions.}
\end{figure}
Our explicit numerical results within the LPA up to 8th order of the three-field model confirm this picture. We may therefore directly exploit the LPA to order 12 including anomalous dimensions within the two-field model to obtain the corresponding exponents. Note that a comparable computation in the full three-field model is quite demanding due to the large number of couplings between different sectors of the model. Using our results for the two-field case, see Ref.\ \cite{Eichhorn:2013zza}, we can accordingly determine the nine largest critical exponents of the model, cf.\ Tab.\ \ref{Tab:IFP}. In general the IFP shows a large number of relevant parameters that require tuning to approach the FP. In fact, the IFP is unstable for any integer combination of field components $(N_1,N_2,N_3)$. It is clear that this pattern will persist to generic multifield models where additional fields are coupled to the system -- for each additional sector the number of relevant directions at the IFP increases (at least) by three.
\subsection{Decoupled fixed point}
\label{SubSec:Decoupled fixed point}
At the DFP, the nonvanishing couplings, i.e., the mass parameters and self-couplings lead to one relevant and one irrelevant direction in each sector. The associated critical exponents are those of the corresponding $O(N_{I})$ Wilson-Fisher FP. While the mixed couplings such as $\lambda_{1,1,0}$ vanish at the FP, the corresponding critical exponents are nontrivial. This follows, as the FP is an interacting FP, and these residual interactions affect scaling dimensions of operators even if the corresponding coupling vanishes. In other words, contributions \mbox{$\sim \lambda_{1,1,0} \lambda_{2,0,0}$} in the $\beta$ functions yield nonvanishing entries in the stability matrix even if $\lambda_{1,1,0} = \lambda_{1,0,1} = \ldots = 0$.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{4pt}
\begin{tabular}{c c c | c c c c c c}
$N_1$& $N_2$&$N_3$ & $\theta_{1}$ & $\theta_{2}$ & $\theta_{3}$ & $\theta_{4}$ & $\theta_{5}$ & $\theta_{6}$ \\
\hline\hline
1& 1& 1& 1.571 & 1.571 &1.571 & 0.142 & 0.142& 0.142\\
2&1&1&1.459 & 1.571 & 1.571 & 0.030 & 0.030 & 0.142\\
3&1&1&1.367 & 1.571 & 1.571 & -0.062 & -0.062 & 0.142\\
4&1&1&1.296 & 1.571 &1.571 & -0.133& -0.133& 0.142\\
\hline
2&2&1 & 1.459 & 1.459 & 1.571 & -0.082 & 0.030& 0.030\\
3&2&1& 1.367 & 1.459 & 1.571 & -0.174 & -0.062& 0.030\\
4&2&1&1.296 & 1.459&1.571& -0.245& -0.133 & 0.030\\
3&3&1&1.367 &1.367&1.571& -0.266& -0.062 & -0.062\\
\hline
2&2&2& 1.459 & 1.459 & 1.459 &-0.082 & -0.082 & -0.082\\
3&2&2& 1.367 & 1.459 & 1.459& -0.174 & -0.174 & -0.082\\
\hline\hline
\end{tabular}
\caption{\label{Tab:DFPa}We list the six largest critical exponents as a function of $N_1, N_2, N_3$ at the DFP, employing results from the LPA 12 including anomalous dimensions from \cite{Eichhorn:2013zza} and using the scaling relations Eqs.\ \eqref{Eq:ScalingRelationsDFPa} -- \eqref{Eq:ScalingRelationsDFPc}.}
\end{table}
At the DFP, the eigendirections corresponding to the six largest critical exponents can be determined using a scaling relation: The quartic couplings $\lambda_{1,1,0}, \lambda_{1,0,1}$, and $\lambda_{0,1,1}$ correspond to eigendirections of the FP with critical exponents
\begin{eqnarray}
\theta_{4} &=& \frac{1}{\nu_1} + \frac{1}{\nu_2}-d ~, \label{Eq:ScalingRelationsDFPa} \\
\theta_{5} &=& \frac{1}{\nu_1} + \frac{1}{\nu_3}-d ~, \label{Eq:ScalingRelationsDFPb} \\
\theta_{6} &=& \frac{1}{\nu_2} + \frac{1}{\nu_3}-d ~, \label{Eq:ScalingRelationsDFPc}
\end{eqnarray}
where $\nu_{I} = \frac{1}{\theta_{I}}$, $I = 1,2,3$. These scaling relations can be motivated as follows \cite{Aharony:1976,Aharony:2002}: At the DFP, the decoupling of the three sectors implies $[\phi_I^{2} \phi_{J}^{2}] = [\phi_{I}^{2}] + [\phi_{J}^{2}]$. Furthermore, the scaling dimensions of $\phi_{I}^{2}$ are -- due to the decoupling -- determined by the Wilson-Fisher critical exponents, such that $[\phi_{I}^{2}] = - \frac{1}{\nu_{I}} + d$. The relations Eqs.\ \eqref{Eq:ScalingRelationsDFPa} -- \eqref{Eq:ScalingRelationsDFPc} follow directly.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{4pt}
\textbf{LPA 8}
\vskip 2pt
\begin{tabular}{c c c | c c c c c c}
$N_1$ & $N_2$ & $N_3$ & $\theta_{1}$ & $\theta_{2}$ & $\theta_{3}$ & $\theta_{4}$ & $\theta_{5}$ & $\theta_{6}$ \\
\hline\hline
1 & 1 & 1 & 1.537 & 1.537 & 1.537 & 0.067 & 0.067 & 0.067\\
2 & 1 & 1 & 1.399 & 1.537 & 1.537 & -0.057 & -0.057 & 0.067\\
3 & 2 & 1 & 1.306 & 1.399 & 1.537 & -0.275 & -0.150& -0.057\\
\hline \hline
& & & \multicolumn{2}{c}{$\Delta \theta_{4}$} & \multicolumn{2}{c}{$\Delta \theta_{5}$} & \multicolumn{2}{c}{$\Delta \theta_{6}$} \\ \hline \hline
1 & 1 & 1 & \multicolumn{2}{c}{0.007} & \multicolumn{2}{c}{0.007} & \multicolumn{2}{c}{0.007}\\
2 & 1 & 1 & \multicolumn{2}{c}{-0.007} & \multicolumn{2}{c}{-0.007} & \multicolumn{2}{c}{0.007}\\
3 & 2 & 1 & \multicolumn{2}{c}{-0.020} & \multicolumn{2}{c}{-0.007} & \multicolumn{2}{c}{-0.007}\\ \hline \hline
\end{tabular}
\vskip 10pt
\textbf{LPA 8}$\mathbf{+\eta}$
\vskip 2pt
\begin{tabular}{c c c | c c c c c c}
$N_1$ & $N_2$ & $N_3$ & $\theta_{1}$ & $\theta_{2}$ & $\theta_{3}$ & $\theta_{4}$ & $\theta_{5}$ & $\theta_{6}$ \\ \hline\hline
1 & 1 & 1 & 1.564 & 1.564 & 1.564 & 0.080 & 0.080 & 0.080\\
2 & 1 & 1 & 1.447 & 1.564 & 1.564 & -0.028 & -0.028 & 0.080\\
3 & 2 & 1 & 1.359 & 1.447 & 1.564 & -0.220 & -0.112 & -0.028\\
\hline \hline
& & & \multicolumn{2}{c}{$\Delta \theta_{4}$} & \multicolumn{2}{c}{$\Delta \theta_{5}$} & \multicolumn{2}{c}{$\Delta \theta_{6}$} \\ \hline \hline
1 & 1 & 1 & \multicolumn{2}{c}{0.047} & \multicolumn{2}{c}{0.047} & \multicolumn{2}{c}{0.047}\\
2 & 1 & 1 & \multicolumn{2}{c}{0.038} & \multicolumn{2}{c}{0.038} & \multicolumn{2}{c}{0.047}\\
3 & 2 & 1 & \multicolumn{2}{c}{0.027} & \multicolumn{2}{c}{0.035} & \multicolumn{2}{c}{0.038}\\
\hline\hline
\end{tabular}
\caption{\label{Tab:DFPb}We list the six largest critical exponents as a function of $N_{1}, N_{2}, N_{3}$ at the DFP, within the LPA 8 and LPA 8 with anomalous dimensions (LPA 8$+ \eta$). We do not exploit the scaling relations, but instead evaluate all critical exponents explicitly from the $\beta$ functions at the DFP. We check the scaling relation explicitly and give the deviation to this order of the truncation. The inclusion of the anomalous dimensions leads to a slightly larger violation of the scaling relation.}
\end{table}
{ We observe that the violation of the scaling relation is slightly larger when anomalous dimensions are taken into account. This is not necessarily surprising, as the main effect of a running wavefunction renormalization is \emph{not} to give a sizable improvement in the value of the critical exponents, but instead to provide a first reasonable estimate of the value of $\eta$ itself. We expect that an enlargement of our truncation, including a field-dependent wavefunction renormalization (or, in other words, momentum-dependent interaction terms), will improve the situation.}
\begin{figure}[!t]
\includegraphics[width=0.8\linewidth]{DFP_N.pdf}
\caption{\label{Fig:DFP}(Color online) We show the fourth-largest critical exponent at the DFP as a function of $N_{1} = N_{2} = N_{3}$, with the LPA 4 result (green circles), LPA 6 (blue squares) and LPA 8 (red diamonds). The LPA clearly converges rapidly.}
\end{figure}
Accordingly the stability of this FP can be determined completely from a knowledge of the Wilson-Fisher FP. Employing a LPA to order 12, including anomalous dimensions, we arrive at the results given in Tab.~\ref{Tab:DFPa}, cf.\ Ref.\ \cite{Eichhorn:2013zza}. Our results obtained within the LPA 8 for the three-field model show reasonable agreement with results deduced from the scaling relations, cf.\ Fig.\ \ref{Fig:DFP}. In fact, we may check the quantitative accuracy of the scaling relations explicitly, without referencing the results from the $O(N)$ model. We simply calculate the deviations $\Delta \theta_{4} = \theta_1 + \theta_2 - d - \theta_4$, $\Delta \theta_{5} = \theta_1 + \theta_3 - d - \theta_5$, and $\Delta \theta_{6} = \theta_2 + \theta_3 - d - \theta_6$, shown in Tab.\ \ref{Tab:DFPb} for the given data sets.
{Our explicit numerical results in LPA 8 deviate from the results inferred from the LPA 12 in the two-field model in some cases. There, we expect that an enlarged truncation in the three-field model will give results in full agreement with those deduced from the two-field case. As the number of couplings grows very substantially with the truncation, such an explicit check is beyond the scope of this work. Where LPA 8 and LPA 12 $+ \eta$ results deviate, the latter are more trustworthy.}
\subsection{Decoupled isotropic fixed point}
\label{SubSec:Decoupled isotropic fixed point}
For three interacting fields, we observe a new FP, where only one of the fields decouples, while the other two sectors show an enhancement of symmetry. For the following discussion, we will assume that it is the $\phi_1$-sector that decouples, while the remaining sectors have a $O(N_2+N_3)$ symmetry. Two other possible DIFPs exist, for which one of the other subsectors decouples, respectively.
Clearly, one positive critical exponent is inherited from the $O(N_1)$ symmetric and another one from the $O(N_2+N_3)$ symmetric Wilson-Fisher scaling spectrum. Two critical exponents that are relevant for the stability properties of this FP follow from the critical exponents of the spin-$1$ and spin-$2$ perturbations of the two-field isotropic $O(N_{2} + N_{3})$ FP, i.e., $y_{2,2}$ and $y_{4,4}$. While $y_{2,2}$ is always positive, $y_{4,4}$ becomes positive for $N_2 + N_3 > 2$. We therefore conclude, that the DIFP can only be stable for $N_2 = N_3 = 1$.
There are two further exponents that we need to consider to establish the stability of the DIFP solution. Both follow from a scaling relation exploiting the decoupling of the $1$-sector: At the FP, the operators $\mathcal{O}_1=\phi_1^2 (\phi_2^2+\phi_3^2)$ and $\mathcal{O}_2 = \phi_1^2 \left(\phi_2^2 - \frac{N_2}{N_2+N_3}(\phi_2^2+\phi_3^2) \right)$ correspond to eigendirections of the stability matrix. As we know from the Wilson-Fisher FP that the scaling dimension $[\phi_I^2] = -\frac{1}{\nu_I}+d$, we deduce that
\begin{equation}
[\mathcal{O}_1] = -\frac{1}{\nu_1} - \frac{1}{\nu_{2+3}} + 2d,
\end{equation}
and accordingly the corresponding critical exponent is given by
\begin{equation}
\theta_{6} = \frac{1}{\nu_{1}} + \frac{1}{\nu_{2+3}} - d ~.
\end{equation}
Similarly, we deduce for the second operator that the corresponding critical exponent is given by
\begin{equation}
\theta_{4} = \frac{1}{\nu_{1}} + y_{2,2} - d ~,
\end{equation}
where $y_{2,2}$ is the scaling dimension of the coupling belonging to $\phi_2^2 - \frac{N_2}{N_2+N_3}(\phi_2^2+\phi_3^2)$ in the two-field case, cf.\ Ref.\ \cite{Eichhorn:2013zza}. The first relation is the one that arises for a two-field DFP, and gives a negative critical exponent (for values of $N_{1} > 1$).
At fixed $N_{2} = N_{3} = 1$, it is the critical exponent $\theta_{4}$ that decides about the stability of the FP. Using results from the two-field case (LPA 12 including an anomalous dimension, cf.\ Ref.\ \cite{Eichhorn:2013zza}) to obtain $\theta_{1} = y_{2,2}$, $\theta_{2} = \frac{1}{\nu_{1}}$, and $\theta_{3} = \frac{1}{\nu_{2+3}}$, we arrive at the results shown in Tab.~\ref{Tab:DIFP}. The DIFP is the stable FP for $N_{1} \geq 6$ and $N_{2} = N_{3} = 1$.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{4pt}
\begin{tabular}{c c c | c c c c c c}
$N_1$ & $N_2$ & $N_3$& $\theta_{1}$ & $\theta_{2}$ & $\theta_{3}$ & $\theta_{4}$ & $\theta_{5}$ & $\theta_{6}$ \\
\hline\hline
1 & 1 & 1 & 1.765 & 1.571 & 1.459 & 0.336 & -0.042 & 0.030 \\
2 & 1 & 1 & 1.765 & 1.459 & 1.459 & 0.224 & -0.042 & -0.092 \\
3 & 1 & 1 & 1.765 & 1.367 & 1.459 & 0.132 & -0.042 & -0.174 \\
4 & 1 & 1 & 1.765 & 1.296 & 1.459 & 0.061 & -0.042 & -0.245 \\
5 & 1 & 1 & 1.765 & 1.242& 1.459 & 0.007 & -0.042 & -0.299 \\
6 & 1 & 1 & 1.765 & 1.203 & 1.459 & -0.032 & -0.042 & -0.338 \\
\hline\hline
\end{tabular}
\caption{\label{Tab:DIFP}Critical exponents at the DIFP, using the LPA 12 including anomalous dimensions, and employing the above scaling relations.}
\end{table}
Note that the derivation of the scaling relations is based on the assumption that the operators corresponding to these couplings are eigenoperators of the stability matrix. This property is, to the best of our knowledge, an assumption in $d = 3$ \cite{Calabrese:2002bm}, and is usually not true within a truncation of the RG flow. Nevertheless, the stability properties are not incompatible with explicit numerical results within the LPA 8, where the transition to stability occurs already at $N_1=5$. { Explicitly, the critical exponents in LPA 8 at $N_1=5$ and $N_2=N_3=1$ read $\theta_1=1.783, \theta_2=1.193, \theta_3=1.399, \theta_4=-0.024,\theta_5=-0.027$ and $\theta_6=-0.395$.}
\vskip 5pt
\subsection{Decoupled biconical fixed point}
\label{SubSec:Decoupled biconical fixed point}
\begin{table}[!h]
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{2.7pt}
\begin{tabular}{c c c | c c c c c c}
$N_1$ & $N_2$ &$N_3$& $\theta_{1}$ & $\theta_{2}$ & $\theta_{3}$ & $\theta_{4}$ & $\theta_{5}$ & $\theta_{6}$ \\
\hline\hline
1 & 1.2 & 1 & {\bf 1.753} & {\bf 1.381} & \textit{1.537} & 0.285 & -0.075 & {\bf -0.005} \\
1 & 1.4 & 1 & {\bf 1.535} & {\bf 1.448} & \textit{1.537} & 0.105 & -0.015 & {\bf -0.010} \\
1 & 1.5 & 1 & {\bf 1.537} & {\bf 1.462} & \textit{1.537} & 0.068 & 0.001 & {\bf -0.001} \\
1 & 1.2 & 2 & {\bf 1.753} & {\bf 1.381} & \textit{1.399} & 0.161 & -0.200 & {\bf -0.005} \\
1 & 1.4 & 2 & {\bf 1.535} & {\bf 1.448} & \textit{1.399} & -0.020 & -0.140 & {\bf -0.010} \\
1 & 1.5 & 2 & {\bf 1.537} & {\bf 1.462} & \textit{1.399} & -0.057 & -0.124 & {\bf -0.001} \\
1.2 & 1 & 5 & {\bf 1.753} & {\bf 1.381} & \textit{1.193} & -0.051 & -0.413 & {\bf -0.005} \\
1.5 & 1 & 5 & {\bf 1.537} & {\bf 1.462} & \textit{1.193} & -0.270 & -0.338 & {\bf -0.001} \\
\hline\hline
\end{tabular}
\caption{\label{Tab:BFP}Critical exponents at the DBFP in the LPA 8. Critical exponents emphasized in \textbf{bold} font correspond to those of the two-field BFP, while those in \textit{italic} arise in the sector with $O(N_3)$ group symmetry. Here, we list the DBFP for values of the field components $N_{I}$, where it is characterized by $\Delta > 0$ (see text).}
\end{table}
Beyond the isotropic and decoupled FP solutions, the two-field models feature another scaling solution, which is the biconical FP \cite{PhysRevLett.33.813, Calabrese:2002bm, Folk:2008mi}. It is stable only in a restricted parameter region of these models, where $\left. \Delta^{\textrm{2-field}} \right|_{\textrm{BFP}} = \lambda_{2,0} \lambda_{0,2} - \lambda_{1,1}^2 > 0$, as it transfers the stability from the IFP to the DFP. Certainly, this FP should similarly manifest itself in the three-field case. While one of the three sectors decouples, the two remaining sectors should feature nondegenerate couplings, and we expect that in a given range of the parameter space such a decoupled BFP will be stable.
{ To obtain as precise results as possible, we should make use of all methods available to us. In fact, results obtained using an $\epsilon$-expansion around $d=4$ in the two-field case, allow us to infer the stability of the decoupled biconical fixed point in the three field case in one important instance: }
From Refs.\ \cite{Folk:2008mi, Calabrese:2002bm}, we know that the biconical FP is stable for $N_1 = 1, N_2 = 2$ in two-field models (and similarly when the sectors are interchanged). Combined with the pattern in Tab.\ \ref{Tab:BFP} for the additional critical exponents in the three-field model, we conjecture that the DBFP is stable for $N_1 = 1, N_2 = 2, N_3\geq 2$ (up to a permutation of the three sectors). To calculate the corresponding critical exponents directly in the three-field model, we expect that an extended truncation will be necessary, taking into account a field-dependent wavefunction renormalization.
\subsection{Search for further stable fixed points}
\label{SubSec:Search for further stable fixed points}
We summarize our results obtained so far in Fig.~\ref{Fig:StabilityPlot}. The figure shows the stable FP solution for the corresponding values of field components, $(N_1, N_2, N_3)$. Apparently, no stable FP exists in the range $N_{1} < 6$, $N_{2} = N_{3}=1$ (up to a permutation of the sectors) that can be derived from the known scaling solutions in the one- and two-field models. This motivates an independent analysis of fully coupled FPs in the three-field model, which we describe in the following sections.
\begin{figure}[!t]
\includegraphics[width=0.8\linewidth]{stability_cube_conjectured_modified.pdf}
\caption{\label{Fig:StabilityPlot}(Color online) We show the stable DIFP (large blue dots), the stable DFP (small orange dots) and points without a stable FP (small gray dots) using the LPA 12$+\eta$ results. We also include points where the DBFP is conjectured to be stable (middle-sized green dots).}
\end{figure}
\subsubsection{Stability trading between stable FPs}
\label{SubSec:Stability trading between stable FPs}
Generally, the $\beta$ functions are non-polynomial functions of a large number of couplings in the multifield models. In the local potential type approximations the number of parameters and couplings increases from 9 to 34 if the order of the truncation is changed from 4 to 8. Thus, finding FPs in the multifield models becomes a highly nontrivial search for zeros of the $\beta$ functions in a high-dimensional parameter space. In the following, we consider strategies to identify new FP solutions using a simple example. Consider the following $\beta$ function which is expanded in terms of the coupling $g$ (assuming that higher than quadratic terms are zero):
\begin{equation}
\beta_g = g ( c + g ),
\end{equation}
where $c$ is a function of the parameters of the model (e.g., dimensionality, number of field components, etc.) and possibly other couplings in a given truncation of the theory. Note that such a form captures the essential properties of typical fixed points, as it allows both for a trivial Gaussian FP and a nontrivial interacting FP, as a function of the parameter $c$. The critical exponent at a fixed point is given by
\begin{equation}
\theta = - \left. \frac{\partial\beta_g}{\partial g} \right|_{\textrm{FP}} = \left. - c - 2 g \right|_{\rm FP}.
\end{equation}
Assuming that it is the exponent $\theta$ that decides about the stability of the FP, we may distinguish the following scenarios: For $c < 0$, the interacting FP at $g = -c$ is infrared stable, whereas the Gaussian FP is unstable. As the parameter $c$ increases towards positive values (as a function of, e.g., $N_I$) the two FPs will approach each other. At $c = 0$, both FPs collide and exchange their stability properties. Moving apart again for $c > 0$, the interacting FP has become the unstable one, whereas the noninteracting FP is stable, cf.\ Fig.\ \ref{Fig:StabilityTrading}. This simple example demonstrates that FPs will typically change their stability when they collide, as was also observed in Ref.\ \cite{Eichhorn:2013zza} in the two-field model with $O(N_{1})\oplus O(N_{2})$ symmetry.
\begin{figure}[!h]
\includegraphics[width=0.8\linewidth]{stability_trading3.pdf}
\caption{\label{Fig:StabilityTrading}To illustrate the mechanism how stability properties are exchanged between FPs, we examine the $\beta$-function $\beta = g(c + g)$ for varying parameter $c$ describing the position of the nontrivial FP. As the FPs pass by one another, the local derivatives $\theta = -\beta'(g)$ exchange their sign indicating an interchange of their IR stability properties. }
\end{figure}
Inspired by the above stability-trading mechanism
we may devise a strategy to identify new FP solutions. Our search for FPs will concentrate on the vicinity of points in coupling space, where a known FP loses its stability.
\subsubsection{Fully coupled fixed points in the three-field model}
\label{SubSec:Fully coupled fixed points in the three-field model}
It turns out that the three-field model works in a different way from the mechanism described above, which applies in the two-field case: Within the LPA 4, we observe that the IFP, DFP, and DIFP have partially overlapping stability regions. No similar behavior occurs in two-field models, where stability regions of different FPs always touch, but never overlap, due to the above stability-trading mechanism. However, this changes dramatically as we include anomalous dimensions: While the IFP inhabits the same points, the DIFP is now only stable for $N_1 \geq 4$, $N_2 = N_3=1$ etc.\ Extensive numerical searches did not reveal a stable FP for $N_1 =3$, $N_2 = N_3=1$, and similarly for the cases where the sectors are interchanged. A similar result holds for higher orders of the LPA: As there are a larger number of independent operators that serve as a basis for the LPA, and thus potentially relevant directions in the three-field model, the stability exchange mechanism may not be captured by the simple model considered above. To elucidate the differences in the three-field model, we will focus on results obtained within the LPA to 8th order.
\begin{figure}[!t]
\includegraphics[width=\linewidth]{BFPDFPcollision.pdf}
\caption{\label{BFPDFPcollision}We plot $\lambda_{1,0,1}, \lambda_{0,1,1}$, and $\lambda_{1,1,0}$ as a function of $N_I$ for the three DBFPs. At $N_I=1.244$ they each pass through the origin of that coordinate system, where the DFP sits. For $N_I>1.244$, the three BFPs move away from each other towards more negative values of the mixed couplings.}
\end{figure}
\begin{figure}[!h]
\includegraphics[width=0.8\linewidth]{BFPDFPthetas.pdf}
\includegraphics[width=0.8\linewidth]{BFPDFPcoords.pdf}
\caption{\label{BFPDFPthetas}We plot the fifth-largest critical exponent at the DBFPs (blue points of increasing value) and the corresponding critical exponent at the DFP (red points of decreasing value) as a function of $N_I$ (upper panel). Below, we show the coordinates of the couplings $\lambda_{2,0,0}$ and $\lambda_{0,2,0}$ at the DBFP (blue diamonds and black squares) and the couplings $\lambda_{2,0,0}= \lambda_{0,2,0} = \lambda_{0,0,2}$ at the DFP (red dots) in the vicinity of the collision point.}
\end{figure}
As a first example, let us consider the point $N_1 = N_2 = N_3 \approx 1.25$. Here, the DFP is stable, but it gains three additional relevant directions around $N_1 = N_2 = N_3 \approx 1.244$. As within the two-field model, this point is marked by a collision with a decoupled biconical FP (DBFP). The main difference is that within a three-field model, three generalizations of the BFP exist, cf.\ Sec.\ \ref{SubSec:Decoupled biconical fixed point}. For $N_I =1.25$, $I = 1,2,3$, all of these FPs feature five relevant directions. Toward smaller $N_I$, all three DBFPs approach the DFP, and simultaneously collide with it at $N_I \approx 1.244$. At this point, the DFP gains one relevant direction from each of the three DBFPs, which subsequently feature only four relevant exponents, cf.\ Fig.\ \ref{BFPDFPcollision} and Fig.~\ref{BFPDFPthetas}. Thus this FP loses stability in a fixed-point collision. The central difference to the two-field case lies in the fact that the symmetry of the model which forces the DFP to collide with three other FPs simultaneously (for $N_1 = N_2 = N_3$ there is an exchange symmetry $N_1 \leftrightarrow N_2 \leftrightarrow N_3$). As each of them starts off with five relevant directions, the collision does not produce a stable FP for $N_I \leq 1.24$, but instead leaves behind three FPs with four relevant directions each.
As a second example of new behavior in three-field models, we consider the point where the DBFP collides with the DIFP in the LPA 8. We fix $N_2 =1$ and $N_3=4$: Then the DBFP has five positive critical exponents at $N_1=1$. Going to larger values of $N_1$, it collides with the DIFP at $N_1 \approx 1.6$. During this collision, the DIFP becomes unstable, and the DBFP gains one negative critical exponent, cf.\ Fig.\ \ref{BFPDIFP}. Similar to the previous scenario, this FP collision is not sufficient to make the DBFP stable. Following the DBFP to even larger values of $N_1$, it undergoes another collision, this time being hit by two other FPs simultaneously. This is a novel feature that is not observed in simple two-field models. Starting in the region $N_1 < 1.2$, these FPs do not seem to exist for real values of the couplings -- at least no sign of them showed up in extensive numerical searches. They can be thought of as being created at the collision. Following this collision, they quickly move away from the collision point for increasing values of $N_{1}$. Each of the newly created FPs features four relevant critical exponents, while the DBFP is stable, cf.\ Fig.\ \ref{BFPcollision}. Both new FPs are anisotropic, and define a new universality class that occurs for the first time if three fields are coupled, cf.\ Tab.\ \ref{Tab:AFP}.
\begin{figure}[!h]
\includegraphics[width=0.9\linewidth]{BFPcoords.pdf}\\
\includegraphics[width=0.9\linewidth]{BFPthetas.pdf}\\
\includegraphics[width=0.9\linewidth]{DIFPtheta.pdf}
\caption{\label{BFPDIFP}(Color online) Here, we show the couplings $\lambda_{2,0,0}$, $\lambda_{0,2,0}$, and $\lambda_{1,1,0}$ at the DBFP (blue circles, purple squares, cyan diamonds) and $\lambda_{2,0,0} = \lambda_{0,2,0} = \lambda_{1,1,0}$ (red triangles) at the DIFP (uppermost panel). We show the fourth- and fifth-largest critical exponent (middle panel) at the DBFP as a function of $N_1$ for $N_2=1, N_3=4$. Around $N_1 \approx 1.16$, the three couplings are clearly degenerate, as expected for a collision with the DIFP. At the same point, the fourth critical exponent crosses zero and becomes negative. Simultaneously, the fifth critical exponent at the DIFP (lower panel) crosses zero and becomes positive.}
\end{figure}
\begin{figure}[!h]
\includegraphics[width=\linewidth]{AFPcollision.pdf}\\
\includegraphics[width=\linewidth]{AFPtheta.pdf}\\
\caption{\label{BFPcollision}Here, we show the couplings $\lambda_{1,0,1}$ and $\lambda_{0,1,1}$ at the two anisotropic FPs. As a function of $N_1$ they converge towards zero, which is where the DBFP is sitting. The collision occurs at $N_1 \approx 1.2$, which is where the fifth critical exponent of these two FPs (which numerically is nearly the same for both) approaches zero.}
\end{figure}
\begin{table}[!t]
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{2.7pt}
\begin{tabular}{ccc|ccc|ccc|c}
$N_1$ & $N_2$ & $N_3$ & $\lambda_{2,0,0}$ & $\lambda_{0,2,0}$ & $\lambda_{0,0,2}$ & $\lambda_{1,1,0}$ & $\lambda_{1,0,1}$ & $\lambda_{0,1,1}$ & $\lambda_{1,1,1}$\\ \hline \hline
2 & 2 & 2 & 6.2 & 6.4& 6.4 & 1.8 & 1.8 & -2.2 & -2.6\\
2 & 1 & 2 & 6.5 & 7.5 & 6.7 & 1.2 & 0.5 & -2.1 & -1.4\\
2 & 1 & 3 & 6.2 & 7.0 & 5.8 & 3.2 & 0.7 & -1.8 & -2.2\\
2 & 1 & 4 & 6.0 & 6.5 & 5.1 & 4.6 & 0.6 & -1.3 & -1.6\\ \hline \hline
\end{tabular}
\vskip 4pt
\renewcommand{\arraystretch}{1.4}
\renewcommand{\tabcolsep}{6.8pt}
\begin{tabular}{ccc|ccc|ccc}
{} & {} & {} & $\theta_1$ & $\theta_2$ & $\theta_3$ & $\theta_4$ & $\theta_5$ & $\theta_6$ \\ \hline \hline
2 & 2 & 2 & 1.70 & 1.33 & 1.29 & 0.20 & -0.17 & -0.24\\
2 & 1 & 2 & 1.62 & 1.40 & 1.35 & 0.08 & -0.05 & -0.23\\
2 & 1 & 3 & 1.71 & 1.34 & 1.28 & 0.15 & -0.10 & -0.30\\
2 & 1 & 4 & 1.77 & 1.32 & 1.23 & 0.16 & -0.07 & -0.35\\ \hline \hline
\end{tabular}
\caption{\label{Tab:AFP}We list selected FP values and critical exponents of the first anisotropic FP in the LPA 8, restricting ourselves to couplings at fourth order in the fields, and additionally giving the value of the coupling associated with the three-field operator $\sim \phi_1^2 \phi_2^2 \phi_3^2$.}
\end{table}
Due to the complicated dynamics of FPs in the parameter space of three-field models, it seems that these models do not necessarily feature a stable FP solution for all values of the model parameters. One might now wonder, why FPs of three-field models show such a disproportionate increase in the number of relevant directions for small values of $N_I$. While the Wilson-Fisher FP has one relevant direction, and two-field models always feature FPs with only two relevant directions, three-field models show FPs with more than three relevant directions. We conjecture that this is related to the fact that the transition from a two-field to a three-field model implies the existence of more than one additional new class of operators: While the transition from one to two fields only adds the mixed interactions $\sim \left( \phi_1^{2} \right)^{m} \left( \phi_2^{2} \right)^{n}$, the transition to three-field models features three different classes of mixed interactions $\sim \left( \phi_I^{2} \right)^{m} \left( \phi_J^{2} \right)^{n}$, $I,J = 1,2,3$, in addition to the new three-field couplings $\sim \left( \phi_1^{2}\right)^{l} \left( \phi_2^{2} \right)^{m} \left( \phi_3^{2} \right)^{n}$. These seem to play a particularly important role for small $N_I$ and imply that the IR scaling properties cannot be accounted for by only three free relevant parameters. Note that this does not imply that the corresponding additional couplings need to become relevant. Since these operators do not necessarily correspond to eigendirections of the RG, they may mix with other operators, and yield corrections to the scaling spectrum.
Applying our results to determine the properties of possible multicritical points in phase diagrams for systems with three competing order parameters, we may conclude that models with small $N_I$ will typically feature a first order, rather than a second order (multicritical) transition. In particular, this applies to the phenomenologically relevant model of three interacting $Z_{2}$-Ising fields, i.e., $Z_{2} \oplus Z_{2} \oplus Z_{2}$ symmetry.
\section{Summary and conclusions}
\label{Sec:Summary and conclusions}
Here, we present a renormalization group study of IR stable FP solutions in three-field models with $O(N_1) \oplus O(N_2) \oplus O(N_3)$ symmetry. Our main results regarding the existence of stable FPs are summarized in Fig.~\ref{Fig:StabilityPlot}. Models in this class exhibit FPs that generalize the Wilson-Fisher FP, falling into three distinct categories, each characterized by the degree of symmetry-enhancement. We find a decoupled FP, a partially isotropic FP solution, and a fully isotropic FP solution. Their scaling spectrum can be deduced partially by considering perturbations around the single- and two-field models with $O(N)$ and $O(N_{1})\oplus O(N_{2})$ symmetry, respectively. We proceed by deriving scaling relations between different critical exponents to discover the stability properties of nontrivial FPs in the three-field case. Apart from the generalized Wilson-Fisher scaling solutions, we identify a decoupled biconical FP whose scaling properties are partly inherited from the BFP in the $O(N_{1})\oplus O(N_{2})$ symmetric model. As a main result of this work, we find that these FPs all show a significantly larger number of relevant critical exponents than in the two-field case, in the region of small $N_1$, $N_2$, and $N_3$. We tentatively connect this result to the existence of a large number of mixed interactions, and further conjecture, that similar results will hold for models with $n>3$ interacting fields.
Summarizing our results, we find no IR stable FP for a small number of fields ($N_{1} < 6$, $N_{2} = N_{3} = 1$, up to permutations of the fields) in $d = 3$ dimensions. This result is certainly unexpected, as there is no evidence for similar behavior in coupled two-field models. While, in principle, we cannot exclude the possibility that stable FPs exist in that region of parameter space, we find no evidence for their existence in extensive numerical searches for FPs of the nonperturbative $\beta$ functions. The identification of FPs in the three-field field models with $O(N_{1})\oplus O(N_{2})\oplus O(N_{3})$ is in general a difficult problem, since the search has to proceed through a high-dimensional coupling space. Nevertheless, the understanding of basic stability transitions between different FPs serves as a guiding principle to single out possible candidates for nontrivial FPs. Quite generally, in coupled-field models, stability seems to be inherited from single mergers or collisions of different FPs. Searches around such stability transition points have not yielded any FP that carries over the stability properties from the IFP (stable at small, noninteger values of $N_{I}<1$) to the decoupled FP. This indicates that the dynamics of FPs in $O(N_{1})\oplus O(N_{2})\oplus O(N_{3})$ symmetric three-field, or general $\bigoplus_{I} O(N_{I})$ symmetric multifield models is very different from that encountered in the simpler $O(N_1)\oplus O(N_2)$-type models.
From these results, we may conclude that models with phenomenological relevance such as, e.g., the $Z_{2} \oplus Z_{2} \oplus Z_{2}$ symmetric model will not feature a multicritical point in its phase diagram. Certainly, it is challenging to find three parameters that are accessible experimentally, and may be tuned to the multicritical point. This would be necessary to quantify the scaling behavior close to the corresponding FP, or to show the absence of such a transition. Nevertheless, it is conceivable that in the context of ultracold atomic systems such a control of the system might be achievable \cite{doi:10.1146/annurev-conmatphys-070909-104059}.
Finally, let us comment on the general applicability of our results to other systems of interest. The renormalization group flow equations are derived for general $d$ Euclidean dimensions and can be applied to $d = 2$, relevant for critical behavior of low-dimensional condensed-matter systems, and $d = 4$, for multifield models of inflation, as well as possible extensions of the Higgs sector of the standard model. We leave this for future work.
\begin{acknowledgements}
We thank P.~Calabrese for discussions.
Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation. This work is supported by DOE grant No.\ DE-FG0201ER41195. M.M.S. is supported by the grant ERC-AdG-290623 and DFG grant FOR 723. A.E. would like to thank the University of Heidelberg and the University of Illinois at Chicago for hospitality during the early stages of this work.
\end{acknowledgements}
\begin{appendix}
\label{Appendix}
\section{Functional renormalization group}
\label{Sec:Functional renormalization group}
The nonperturbative functional renormalization group \cite{Wetterich:1992yh,Berges:2000ew,Polonyi:2001se,Pawlowski:2005xe,Gies:2006wv,Delamotte:2007pf} defines a functional flow for the scale dependent effective action $\Gamma_{k}$, which interpolates between the microscopic action $S$ defined at some ultraviolet cutoff scale $k = \Lambda$ and the full effective action $\Gamma_{k\rightarrow 0} = \Gamma$, when the renormalization group parameter $k$ is removed. It is given by
\begin{equation}
\partial_t \Gamma_{k} = \frac{1}{2} \Tr \int\! \frac{d^{d}q}{(2\pi)^{d}} \left(\Gamma_k^{(2)} (q) + R_k(q) \right)^{-1} \partial_t R_k(q) ~,
\label{Eq:WetterichEquation}
\end{equation}
where the logarithmic scale derivative is written in terms of the parameter $t = \ln (k/\Lambda)$, and the second functional derivative $\Gamma_k^{(2)}(p,q) = \frac{\delta^2 \Gamma_k}{\delta \chi(-p) \delta \chi(q)}$, and $\Gamma^{(2)}(q) (2\pi)^{d} \delta^{(d)}(p-q) \equiv \Gamma^{(2)} (p,q)$. Here, $\chi$ denotes the complete field content of our model and the trace $\Tr$ denotes a summation over internal degrees of freedom, i.e., both fields and field components. The regulator function $R_{k}$ implements a mass-like cutoff and regulates the infrared divergences. We take $R_{I J} (q) = R_{I}(q) \delta_{I J}$, $I, J = 1,2,3$, while the momentum dependence is given by $R_{I}(q) = Z_{I} (k^{2} - q^{2}) \theta(k^{2} - q^{2})$, where the wavefunction renormalization is scale dependent. This choice is referred to as the optimized regulator \cite{Litim:2000ci, Litim:2001up} which allows us to derive fully analytic expressions for the nonperturbative $\beta$ functions, and it is thus a convenient choice to identify possible scaling solutions.
\onecolumngrid
\section{Scale dependent effective potential}
\label{Sec:Scale dependent effective potential}
We derive the renormalization group flow equation for the effective potential by plugging our \textit{ansatz} Eq.\ \eqref{Eq:EffectivePotential} into the flow equation \eqref{Eq:WetterichEquation} and projecting the right-hand-side onto a constant field configuration:
\begin{eqnarray}
\partial_t u_k &=& -d u_k +\sum_I (d-2+ \eta_I) \rho_I \partial_{\rho_{I}} u_{k} + { 2 v_{d} \sum_{I} \Big\{ (N_I-1) l_{0}^{(I)} \left( \partial_{\rho_{I}} u_{k} \right)} \nonumber\\ && { +\: l_{R}^{(I)} \left( \left\{ \partial_{\rho_{J}} u_{k} + 2 \rho_{J} \partial_{\rho_{J}}^{2} u_{k} \right\}, \left\{ 4 \rho_{J} \rho_{K} \partial_{\rho_{J}} \partial_{\rho_{K}} u_{k} \right\} \right)} \Big\} ~.
\label{Eq:FlowEquationPotential}
\end{eqnarray}
Here, $v_{d}^{-1} = 2^{d+1} \pi^{\frac{d}{2}} \Gamma\left(\frac{d}{2}\right)$ arises from the volume integration, and the anomalous dimensions are defined as $\eta_I = - \partial_t \ln Z_I$. In the following the notation follows Sec.\ \ref{Sec:Model}. The threshold functions $l_{0}^{(I)}$ and $l_{R}^{(I)}$ (see, e.g., \cite{Braun:2011pp}) define the diagrammatic contributions to the renormalization group flow of the scale dependent effective potential, where the upper index indicates the corresponding sector. Using an optimized regulator function \cite{Litim:2000ci,*Litim:2001up} the threshold functions take the following form:
\begin{eqnarray}
l_{0}^{(I)}(w_{I}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{I}}{d+2}\right)\frac{1}{1+w_{I}} ~, \\
\hspace{-20pt} l_{R}^{(1)}(\{ w_{I} \}, \{ \delta_{I,J}^{2} \}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{1}}{d+2}\right) \frac{\left( 1 + w_{2} \right) \left( 1 + w_{3} \right) - \delta_{2,3}^{2}}{2 \delta_{1,2} \delta_{1,3} \delta_{2,3}
- \delta_{1,2}^{2} \left( 1 + w_{3} \right) - \delta_{1,3}^{2} \left( 1 + w_{2} \right) - \delta_{2,3}^{2} \left( 1 + w_{1} \right)
+ \prod_{J} \left( 1 + w_{J} \right)} ~, \\
\hspace{-20pt} l_{R}^{(2)}(\{ w_{I} \}, \{ \delta_{I,J}^{2} \}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{2}}{d+2}\right)
\frac{\left( 1 + w_{1} \right) \left( 1 + w_{3} \right) - \delta_{1,3}^{2}}{2 \delta_{1,2} \delta_{1,3} \delta_{2,3}
- \delta_{1,2}^{2} \left( 1 + w_{3} \right) - \delta_{1,3}^{2} \left( 1 + w_{2} \right) - \delta_{2,3}^{2} \left( 1 + w_{1} \right)
+ \prod_{J} \left( 1 + w_{J} \right)} ~, \\
\hspace{-20pt} l_{R}^{(3)}(\{ w_{I} \}, \{ \delta_{I,J}^{2} \}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{3}}{d+2}\right)
\frac{\left( 1 + w_{1} \right) \left( 1 + w_{2} \right) - \delta_{1,2}^{2}}{2 \delta_{1,2} \delta_{1,3} \delta_{2,3}
- \delta_{1,2}^{2} \left( 1 + w_{3} \right) - \delta_{1,3}^{2} \left( 1 + w_{2} \right) - \delta_{2,3}^{2} \left( 1 + w_{1} \right)
+ \prod_{J} \left( 1 + w_{J} \right)} ~.
\end{eqnarray}
Note, that the nature of the interactions is such, that only the radial modes $l_{R}^{(I)}$ are affected by the couplings, e.g., $\delta_{1,2}^{2} = 4 \kappa_{1} \kappa_{2} \lambda_{1,1,0}^{2}$, and equivalent couplings between the remaining sectors. One may easily check that in the limit of vanishing couplings, $\delta_{1,3} \rightarrow 0$ and $\delta_{2,3} \rightarrow 0$, the radial contributions $l_{R}^{(I)}$ in the $(1,2)$ sectors reduce to the threshold functions that were already derived in the two-field model \cite{Eichhorn:2013zza}:
\begin{eqnarray}
\hspace{-20pt} l_{0}^{(I)}( w_{I} ) &=& \frac{2}{d} \left( 1 - \frac{\eta_{I}}{d+2}\right)\frac{1}{1+w_{I}} ~,
\label{Eq:ThresholdFunctionL0} \\
\hspace{-20pt} l_{R}^{(1)}(\{ w_{I} \}, \{ \delta_{I,J}^{2} \}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{1}}{d+2}\right) \frac{1 + w_{2}}{\left( 1 + w_{1} \right) \left( 1 + w_{2} \right)- \delta_{1,2}^{2}} ~, \\
\hspace{-20pt} l_{R}^{(2)}(\{ w_{I} \}, \{ \delta_{I,J}^{2} \}) &=& \frac{2}{d} \left( 1 - \frac{\eta_{2}}{d+2}\right) \frac{1 + w_{1}}{\left( 1 + w_{1} \right) \left( 1 + w_{2} \right)- \delta_{1,2}^{2}} ~.
\label{Eq:ThresholdFunctionLTildeR}
\end{eqnarray}
From \eqref{Eq:FlowEquationPotential} the flow equations for the couplings are derived by the differentiation with respect to the fields and successive projection onto a nonvanishing constant background field configuration $\rho_{I} = \kappa_{I}$, defined by the minimum of the effective potential. For some FPs it might be necessary to employ an expansion point where one or several of the $\kappa_{I}$ are vanishing. In this case $\beta_{\kappa_{3}} \equiv 0$. For a detailed discussion of this issue, we refer to Ref.\ \cite{Eichhorn:2013zza}.
We obtain the $\beta$ functions for the couplings $\lambda_{l,m,n}$, $l + m + n \geq 2$:
\begin{equation}
\beta_{\lambda_{l,m,n}} = \left. \left(\frac{\partial^{l+m+n}}{\partial \rho_1^{l} \partial \rho_2^{m} \partial \rho_3^{n}} \partial_t u_k + \beta_{\kappa_1} \frac{\partial^{(l+1)+m+n} u_k}{\partial \rho_{1}^{l+1} \partial \rho_2^m \partial \rho_3^n} + \beta_{\kappa_2} \frac{\partial^{l+(m+1)+n} u_k}{\partial \rho_{1}^{l} \partial \rho_2^{m+1} \partial \rho_3^n} + \beta_{\kappa_3} \frac{\partial^{l+m+(n+1)} u_k}{\partial \rho_{1}^{l} \partial \rho_2^m \partial \rho_3^{n+1}} \right) \right|_{\rho_I = \kappa_I} ~,
\end{equation}
where $\beta_{\lambda_{l,m,n}} \equiv \partial_{t} \lambda_{l,m,n}$. The $\beta$ functions for the scale dependent dimensionless field expectation values $\kappa_{I}$, $I = 1,2,3$, are given by
\begin{eqnarray}
\beta_{\kappa_1} &=& \frac{1}{\Delta} \left\{
- \Delta_{2,3} \partial_{\rho_1} \partial_t u_k + \left(\lambda_{1,1,0} \lambda_{0,0,2} - \lambda_{1,0,1} \lambda_{0,1,1}\right) \partial_{\rho_2} \partial_t u_k \right. \nonumber \\ && \left.\left. +\: \left(\lambda_{1,0,1} \lambda_{0,2,0} - \lambda_{1,1,0} \lambda_{0,1,1}\right)
\partial_{\rho_3} \partial_t u_k \right\} \right|_{\rho_I = \kappa_I} ~,
\end{eqnarray}
\begin{eqnarray}
\beta_{\kappa_2} &=& \frac{1}{\Delta} \left\{ - \Delta_{1,3} \partial_{\rho_2} \partial_t u_k + \left( \lambda_{2,0,0} \lambda_{0,1,1} - \lambda_{1,1,0} \lambda_{1,0,1} \right) \partial_{\rho_3} \partial_t u_k \right. \nonumber\\
&& \left.\left. +\: \left( \lambda_{0,0,2} \lambda_{1,1,0} - \lambda_{1,0,1} \lambda_{0,1,1} \right) \partial_{\rho_1} \partial_t u_k \right\} \right|_{\rho_I = \kappa_I} ~, \\
\beta_{\kappa_3} &=& \frac{1}{\Delta} \left\{
- \Delta_{1,2} \partial_{\rho_3} \partial_t u_k + \left( \lambda_{2,0,0} \lambda_{0,1,1} - \lambda_{1,1,0} \lambda_{1,0,1} \right) \partial_{\rho_2} \partial_t u_k \right. \nonumber\\ && \left.\left. +\: \left( \lambda_{0,2,0} \lambda_{1,0,1} - \lambda_{1,1,0} \lambda_{0,1,1}\right) \partial_{\rho_1} \partial_t u_k \right\} \right|_{\rho_I = \kappa_I} ~.
\end{eqnarray}
Here, we have defined the coupling parameter
\begin{equation}
\Delta_{1,2} = \lambda_{2,0,0} \lambda_{0,2,0} - \lambda_{1,1,0}^{2} ~,
\end{equation}
and equivalently $\Delta_{1,3}$ and $\Delta_{2,3}$ (defined from the remaining quartic couplings in the three-field model), as well as the parameter
\begin{eqnarray}
\Delta &=& - 2 \left( \lambda_{2,0,0} \lambda_{0,2,0} \lambda_{0,0,2} - \lambda_{1,1,0} \lambda_{1,0,1} \lambda_{0,1,1} \right) + \lambda_{0,0,2} \Delta_{1,2} + \lambda_{0,2,0} \Delta_{1,3} + \lambda_{2,0,0} \Delta_{2,3} ~.
\end{eqnarray}
These parameters quantify the symmetry enhancement properties of the system. In particular, for certain symmetry enhanced FPs, these quantities vanish exactly.
\section{Wavefunction renormalization and anomalous dimensions}
\label{Sec:Wavefunction renormalization and anomalous dimensions}
To determine the scale dependence of the field independent renormalization factor $Z_{I}$ from the functional flow equation \eqref{Eq:WetterichEquation} we perform a projection of the flow onto operators of the type $\sim \left( \partial_{\mu} \phi_{I} \right)^{2} $. This yields the scale dependence of the coefficient for the corresponding operator in the effective action. We have
\begin{equation}
\partial_{t} Z_{I} = \frac{(2\pi)^{d}}{\delta^{(d)}(0)} \lim_{Q\rightarrow 0} \frac{\partial}{\partial Q^{2}} \frac{\delta^{2}}{\delta \varphi_{I}(-Q) \delta \varphi_{I}(Q)} \partial_{t} \Gamma_{k} ~,
\label{Eq:RenormalizationFactorProjection}
\end{equation}
where the functional derivatives are taken with respect to the Nambu-Goldstone (NG) degrees of freedom $\varphi_{I}$ in the $I$-sector. Note, there is no summation implied over the $I$-index. For details of the derivation in the context of the $O(N)$ vector model, we refer to \cite{Tetradis:1993ts,Berges:2000ew}.
The anomalous dimensions are defined via the scaling contribution to the wavefunction renormalization, i.e., $\eta_{I} = - \partial_{t} \ln Z_{I}$, and take the following form in the three-field model:
\begin{eqnarray}
\eta_{1} &=& \frac{16 v_{d}}{d} \Xi^{-1} \left\{ \kappa_{2} \lambda_{1,1,0}^2 + \kappa_{3} \lambda_{1,0,1}^2 + \kappa_{1} \left( \lambda_{2,0,0} + 2 \kappa_{2} \Delta_{1,2} + 2 \kappa_{3} \Delta_{1,3} + 4 \kappa_{2} \kappa_{3} \Delta \right)^2 \right. \nonumber\\ && \left. +\: 4 \kappa_{2} \kappa_{3} \left( \Delta_{2,3} \left( \lambda_{2,0,0} + \kappa_{2} \Delta_{1,2} + \kappa_{3} \Delta_{1,3} \right) - \Delta \left(1 + \kappa_{2} \lambda_{0,2,0} + \kappa_{3} \lambda_{0,0,2} \right) \right) \right\} ~,
\label{Eq:Eta1} \\
\eta_{2} &=& \frac{16 v_{d}}{d} \Xi^{-1} \left\{ \kappa_{1} \lambda_{1,1,0}^{2} + \kappa_{3} \lambda_{0,1,1}^2 + \kappa_{2} \left( \lambda_{0,2,0} + 2 \kappa_{1} \Delta_{1,2} + 2 \kappa_{3} \Delta_{2,3} + 4 \kappa_{1} \kappa_{3} \Delta \right)^2 \right. \nonumber\\
&& \left. +\: 4 \kappa_{1} \kappa_{3} \left( \Delta_{1,3} \left( \lambda_{0,2,0} + \kappa_{1} \Delta_{1,2} + \kappa_{3} \Delta_{2,3} \right) - \Delta \left(1 + \lambda_{2,0,0} \kappa_{1} + \lambda_{0,0,2} \kappa_{3} \right) \right) \right\} ~,
\label{Eq:Eta2} \\
\eta_{3} &=& \frac{16 v_{d}}{d} \Xi^{-1} \left\{ \kappa_{1} \lambda_{1,0,1}^2 + \kappa_{2} \lambda_{0,1,1}^2 + \kappa_{3} \left( \lambda_{0,0,2} + 2 \kappa_{1} \Delta_{1,3} + 2 \kappa_{2} \Delta_{2,3} + 4 \kappa_{1} \kappa_{2} \Delta \right)^2 \right. \nonumber\\
&& \left. +\: 4 \kappa_{1} \kappa_{2} \left( \Delta_{1,2} \left( \lambda_{0,0,2} + \kappa_{1} \Delta_{1,3} + \kappa_{2} \Delta_{2,3} \right) - \Delta \left(1 + \lambda_{2,0,0} \kappa_{1} + \lambda_{0,2,0} \kappa_{2} \right) \right) \right\} ~,
\label{Eq:Eta3}
\end{eqnarray}
where the prefactor in the above expressions is defined as:
\begin{eqnarray}
\Xi &=& \Big(1 + 2 \sum_{I} \lambda_{I} \kappa_{I} + 4 \sum_{I < J} \kappa_{I} \kappa_{J} \Delta_{I,J} + 8 \kappa_{1} \kappa_{2} \kappa_{3} \Delta \Big)^2 ~.
\end{eqnarray}
Note, that Eqs.\ \eqref{Eq:Eta1} -- \eqref{Eq:Eta3} reduce to the two-field results in the limit $\delta_{1,3}^{2} = \delta_{2,3}^{2} = 0$ (cf.\ Appendix \ref{Sec:Scale dependent effective potential})
\begin{eqnarray}
\hspace{-20pt} \eta_{1} &=& \frac{16 v_{d}}{d} \frac{ \kappa_{2} \lambda_{1,1}^2 + \kappa_{1} \left(\lambda_{2,0} + 2 \kappa_{2} \Delta_{1,2} \right)^2}{\left(1 + 2 \kappa_{1} \lambda_{2,0} + 2 \kappa_{2} \lambda_{0,2} + 4 \kappa_{1} \kappa_{2} \Delta_{1,2} \right)^2} ~, \\
\hspace{-20pt} \eta_{2} &=& \frac{16 v_{d}}{d} \frac{ \kappa_{1} \lambda_{1,1}^2 + \kappa_{2} \left(\lambda_{0,2} + 2 \kappa_{1} \Delta_{1,2} \right)^2}{\left(1 + 2 \kappa_{1} \lambda_{2,0} + 2 \kappa_{2} \lambda_{0,2} + 4 \kappa_{1} \kappa_{2} \Delta_{1,2} \right)^2} ~,
\end{eqnarray}
where we have written $\lambda_{1,1} \equiv \lambda_{1,1,0}$, etc., as the $3$-sector effectively decouples in this case. From here, it is easy to identify the decoupling and symmetry enhancement scenarios for the spectrum of anomalous dimensions, by considering the proper limits of the couplings between different sectors. These results for the two-field case are equivalent to the anomalous dimensions given in Refs.\ \cite{Friederich:2010hr,Wetzel:2013}.
\twocolumngrid
\end{appendix}
|
\section{Introduction}
Localization of modes in different physical systems is an interesting and puzzling phenomenon. It can be generated by the underlying geometry or by randomness. In this paper we consider the case in which localization is induced by the presence of a point scatterer on a two-dimensional plate, which is deterministic but requires either renormalization or spectral theory to be properly defined.
The specific model is called the \v{S}eba billiard and was introduced in \cite{Seba} to study quantum chaos. See also \cite{Shigehara1,Shigehara2,rudnick, Ueberschar} for further developments. \v{S}eba considered a limiting case of a standard model of ergodic dynamics, the Sinai billiard \cite{Sinai}, which is a rectangle with a disk removed. In the standard quantization of that model one considers the Laplace operator with zero (Dirichlet) boundary conditions on the boundaries of the rectangle and of the disk. In the \v{S}eba model the disk is shrunk to a point with a suitable renormalization. That renormalization can be interpreted as a choice of a self-adjoint extension \cite{reedsimon2} of the Laplacian on the rectangle with the point removed. The point is then called a {\em point scatterer}. We show that the presence of such a scatterer has a dramatic effect on the localization of low-lying modes. For other two-dimensional structures, a localization for the modes of the Laplacian was studied Sapoval \emph{et al.} \cite{fractal, irregular, review1} in the case of irregular geometry or fractal boundaries. Filoche and Mayboroda \cite{bilaplacian} discovered that localization can be achieved for modes of the bi-Laplacian $ \Delta^2$ on a rectangle with a point removed. For this fourth order operator the natural boundary conditions require the mode and its gradient to vanish at the boundary. Physically, this boundary condition means that the plate is {\em clamped} at the boundary and at the interior point.
In \cite{bilaplacian}, numerical analysis of the modes of the bi-Laplacian showed strong localization on one side of the clamped interior point. Somewhat surprisingly, the same phenomenon occurs for the \v{S}eba billiard, that is, for a model with quantum mechanical origins. As pointed out in \cite{bilaplacian} this phenomenon does occur for limits of eigenfunctions on Sinai billiards with shrinking disks. In our
language that means that localization does not occur without renormalization.
\section{Formalism}
Point scatterers are formally defined by a Schr\"{o}dinger operator \(-\Delta+c\delta_{\mathbf{x}_0}\) where \(c\) is constant and \(\delta_{\mathbf{x}_0}\) is the Dirac delta function located at a specific point \(\mathbf{x}_0\). More precisely, it is a self-adjoint extension of the Laplacian whose domain consists of the functions vanishing at \(\mathbf{x}_0\). A point scatterer in a rectangle with the Dirichlet boundary condition is called the \v{S}eba billiard \cite{Seba}.
Consider a rectangle \(\Omega = [0,a] \times [0,b]\) with \(a, b>0\) and the Dirichlet Laplacian \[-\Delta : H^2(\Omega)\cap H_0^1(\Omega) \rightarrow L^2(\Omega).\] Then we have the eigenvalues \(0<E_1\le E_2\le\cdots\) of \(-\Delta\) with the corresponding \(L^2\)-normalized eigenfunctions \(\phi_1,\phi_2, \cdots\).
On the other hand, we construct a point scatterer at \(\mathbf{x}_0 =(x_0,y_0)\in \Omega\) as follows: First, restrict the domain of the Dirichlet Laplacian \(-\Delta\) to the functions vanishing at \(\mathbf{x}_0\in \Omega\). By the theory of self-adjoint extension developed by von Neumann, such a symmetric operator has a family of self-adjoint extensions \(-\Delta_{\alpha,\mathbf{x}_0}\) with a parameter \(\alpha \in (-\infty,\infty ]\). More precisely, let \(G_z\) be the integral kernel of the resolvent \((-\Delta-z)^{-1}: L^2(\Omega) \rightarrow L^2(\Omega)\), namely,
\[G_z(\mathbf{x},\mathbf{x}') = \sum_{n=1}^\infty \frac{\phi_n(\mathbf{x}) \phi_n(\mathbf{x}')}{E_n-z}\]
so that for \(f \in L^2 (\Omega)\),
\[(-\Delta-z)^{-1} f (\mathbf{x}) = \int_\Omega G_z(\mathbf{x},\mathbf{x}') f(\mathbf{x}') d\mathbf{x}'.\]
Then for \(z\in \rho (-\Delta_{\alpha,\mathbf{x}_0})\), the integral kernel of \((-\Delta_{\alpha,\mathbf{x}_0}-z)^{-1} : L^2(\Omega)\rightarrow L^2(\Omega)\) reads
\begin{multline}\label{dom}
(-\Delta_{\alpha,\mathbf{x}_0}-z)^{-1}(\mathbf{x},\mathbf{x}') \\= G_z(\mathbf{x},\mathbf{x}') + \left[\alpha-F(z) \right]^{-1} G_z(\mathbf{x}_0,\mathbf{x}') G_z(\mathbf{x},\mathbf{x}_0)
\end{multline}
where
\begin{equation}\label{Fz}
F(z)=\sum_{n=1}^\infty \phi_n(\mathbf{x}_0)^2\left( \frac{1}{E_n-z}-\frac{E_n}{E_n^2+1} \right) \end{equation}
(see Fig.~\ref{Fz}).
The coupling constant \(\alpha\in (-\infty,\infty]\) can be considered a parameter related to the strength of the point scatterer. Note that the point scatterer annihilates as \(\alpha \rightarrow\pm \infty\) whereas it acts stronger when \(|\alpha|\ll\infty\).
Now we consider the spectral property of \v{S}eba billiards. Let \(\sigma(P)\) denote the spectrum of an operator \(P\) and let \(\mult(z,P)\) denote the multiplicity of an eigenvalue \(z\in \sigma(P)\).
As the Dirichlet Laplacian \(-\Delta\) has a purely discrete spectrum, so does \(-\Delta_{\alpha,\mathbf{x}_0}\). In addition, some eigenvalues of \(-\Delta_{\alpha,\mathbf{x}_0}\) remain in \(\sigma(-\Delta)\) regardless of the coupling constant \(\alpha\) while the others do not. Hence, for \(\alpha\in\mathbb{R},\) we divide \(\sigma(-\Delta_{\alpha,\mathbf{x}_0})\) into the following two types:
\begin{enumerate}
\item Perturbed eigenvalues: \(\sigma(-\Delta_{\alpha,\mathbf{x}_0})\setminus\sigma(-\Delta) \) and
\item Unperturbed eigenvalues: \(\sigma(-\Delta_{\alpha,\mathbf{x}_0})\cap\sigma(-\Delta)\)
\end{enumerate}
where each of them is obtained by different conditions as follows:
\begin{thm}\label{pert}
For \(\alpha\in \mathbb{R}\), \(z\in\sigma(-\Delta_{\alpha,\mathbf{x}_0})\setminus\sigma(-\Delta)\) if and only if \[\alpha=F(z).\]
Then \(\mult(z,-\Delta_{\alpha,\mathbf{x}_0})=1\) with the corresponding eigenfunctions \[\psi(\mathbf{x})=N^{-1} G_z(\mathbf{x},\mathbf{x}_0)\] where \(N=\|G_z(\bullet,\mathbf{x}_0)\|_{L^2(\Omega)}\) is the normalization constant.
\end{thm}
\begin{figure}
\centering
\includegraphics[width=60mm]{Fz.png}
\caption{A schematic graph of \(F(z)\) defined by Eq.\eqref{Fz}.}
\label{Fz}
\end{figure}
\begin{thm}\label{unpert}
Define \(\mu\) and \(\mu_0\) as
\begin{align}\mu(z)&\equiv\mult(z,-\Delta)=\#\{n\ge1 ~|~ z=E_n\}
\\ \mu_0(z) &\equiv \#\{n\ge1 ~|~ z=E_n, \phi_n(\mathbf{x}_0)=0\}.\end{align}
Then for \(\alpha\in \mathbb{R},~z\in \sigma(-\Delta_{\alpha,\mathbf{x}_0})\cap\sigma(-\Delta) \) if and only if
\[\mu_0 (z) \ge 1 ~\text{ or }~ \mu(z)\ge 2\]
Also,
\[\mult(z,-\Delta_{\alpha,\mathbf{x}_0})=\begin{cases}
\mu(z), & \mbox{ if }\mu_0(z) =\mu(z)\\
\mu(z)-1, & \mbox{ if } \mu_0(z) <\mu(z)
\end{cases}\] with the corresponding eigenspaces
\[\left\{\sum_{z=E_n} c_n\phi_n ~\middle|~ \sum_{z=E_n} c_n\phi_n(\mathbf{x}_0)=0, \quad c_n\in\mathbb{C}\right\}.\]
\end{thm}
Proofs can be found in Chapter 2 of \cite{pseudolaplacian} with generalized statements for a compact Riemannian manifold of dimension two or three. The coupling constant \(\alpha\) in Eq.\eqref{dom} can be obtained by following the notations provided by Albeverio \emph{et al.} \cite{solvable}. Note that \(\alpha\) also corresponds to the inverse of the coupling constant \(v_B\) or \(\overline{v}_\theta\) in Shigehara's setting \cite{Shigehara1,Shigehara2}.
We may interpret Theorem~\ref{unpert} as that the Laplacian eigenfunctions vanishing at \(\mathbf{x}_0\) do not feel the presence of the point scatterer. So not only do they remain as the eigenfunctions of \(-\Delta_{\alpha,\mathbf{x}_0}\), but also the associated eigenvalues stay in \(\sigma(-\Delta_{\alpha,\mathbf{x}_0})\) for any \(\alpha\).
On the other hand, by combining Theorem~\ref{pert} and \ref{unpert} we obtain that the eigenvalues of the point scatterer are interlaced between those of the Dirichlet Laplacian. In other words, for \(\alpha\in (-\infty,\infty]\), let \(z_1(\alpha)\le z_2(\alpha)\le \cdots \) be the eigenvalues of \(-\Delta_{\alpha,\mathbf{x}_0}\). Then we have
\[z_1(\alpha) \le E_1 \le z_2(\alpha) \le E_2 \le z_3(\alpha) \le E_3 \le \cdots.\]
In addition, for \(n \ge 1\),
\begin{align*}\lim_{\alpha\rightarrow\infty}z_n(\alpha) &= E_n \\
\lim_{\alpha\rightarrow -\infty}z_{n+1}(\alpha) &= E_n\\
\lim_{\alpha\rightarrow -\infty}z_{1}(\alpha) &=-\infty\end{align*}
\section{Localization of Eigenfunctions}\label{result}
In this section, we show several examples of perturbed eigenfunctions localized on a plate due to the point scatterer with a suitable coupling constant \(\alpha\in\mathbb{R}\).
Let \(\Omega=[0,a]\times [0,b]\) with \(a=\sqrt{E}\) and \(b=1/\sqrt{E}\) so every plate has unit area for any \(E>0\) which is the eccentricity of the plate. The unperturbed eigenfunctions obtained by Theorem~\ref{unpert} are independent of \(\alpha\) so they have no chance to be localized at all. In order to avoid such cases as much as possible, first we assume the eccentricity \(E\) to be irrational so that all \(E_n\)'s are nondegenerate. In addition, let \(\frac{a}{x_0}\) be irrational to minimize the case in which \(\phi_n\) vanishes at \(\mathbf{x}_0\).
In this paper, we choose a specific value \(\frac{a}{x_0}=2\pi\) (Fig.~\ref{plate}). However, it should be noted that the qualitative property we observe also holds for other values of \(\frac{a}{x_0}\) as long as they are irrational.
\begin{figure}
\centering
\includegraphics[width=60mm]{plate.png}
\caption{(Color online) Geometry of a point scatterer at \(\mathbf{x}_0=(x_0,y_0) \) (marked as \(\times\)) in \(\Omega\). The left part of the plate divided by \(\mathbf{x}_0\) is denoted by \(\Omega_1=[0,x_0]\times[0,b]\). }
\label{plate}
\end{figure}
By Theorem~\ref{pert}, if \(z_n(\alpha)\) is a perturbed eigenvalue of \(-\Delta_{\alpha,\mathbf{x}_0}\) then the corresponding normalized eigenfunction \(\psi_{n,\alpha} \in L^2(\Omega)\) satisfies the following \(L^2\)-identity:
\begin{equation}\label{psina}
\psi_{n,\alpha}(\mathbf{x}) = N_{n,\alpha}^{-1}\sum_{n'=1}^\infty \frac{\phi_{n'}(\mathbf{x}) \phi_{n'}(\mathbf{x}_0)}{E_{n'}-z_n(\alpha)}
\end{equation}
where \(N_{n,\alpha}\) is the \(L^2\)-normalization constant.
We now investigate the localization of the perturbed eigenfunctions given by Eq.\eqref{psina} which depends on the mode number \(n\), the coupling constant \(\alpha\), and the eccentricity \(E\). Among those three variables, we mainly concentrate on \(n\) and \(E\). It should be noted that \(\alpha\) is chosen to maximize the localization property for each situation.
In order to quantify the localization of multiple modes with ease, we introduce two kinds of measurement: First, we define the \(L^2\)-norm ratio \(R_1(n,\alpha)\) as
\begin{equation}\label{R1na}
R_1(n,\alpha) = \left(\int_{\Omega_1} |\psi_{n,\alpha}(\mathbf{x})|^2 d\mathbf{x}\right)^\frac{1}{2},
\end{equation}
where \(\Omega_1=[0,x_0]\times[0,b]\) denotes the left part of the plate divided by the point scatterer.
In addition, let \(A(n,\alpha)\) be the amplitude at \(\mathbf{x}_0\):
\begin{equation}\label{Ana}
A(n,\alpha)=|\psi_{n,\alpha}(\mathbf{x}_0)|.
\end{equation}
For simplicity, let us omit \(E\) in those notations since it is already embedded in every \(E_n\) and \(\phi_n\) of Eq.\eqref{psina}.
Note that we assume that all eigenfunctions are \(L^2\)-normalized. Then \(R_1(n,\alpha)\) measures the ratio of the \(L^2\)-norm localized in \(\Omega_1\). For instance, \(R_1(n,\alpha)=0\) and \(R_1(n,\alpha)=1\) imply that \(\psi_{n,\alpha}\) is completely localized in \(\Omega\setminus\Omega_1\) and \(\Omega_1\), respectively. On the other hand, \(A(n,\alpha)\) measures how much the point scatterer at \(\mathbf{x}_0\) attracts the amplitude of modes.
\begin{figure}
\centering
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode1_unpert.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode2_ecc10pi_a007.png}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode6_unpert.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode7_ecc10pi_a007.png}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode159_unpert.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{mode160_ecc10pi_a007.png}
\end{subfigure}
\caption{(Color online) Several modes on a plate with eccentricity \(E=10\pi\). The figures on the left- and right-hand columns correspond to the Dirichlet Laplacian and a point scatterer at \(\mathbf{x}_0\), respectively. For the point scatterer, the coupling constant \(\alpha\) is chosen to maximize the localization. One can observe several modes localized on the left- or the right-hand side of the point scatterer.}\label{eigftn}
\end{figure}
\subsection{Point scatterer acting as a barrier}
Now we provide numerical results showing that the low-level eigenfunctions with \(n\ge 2\) localize to the left or the right of \(\mathbf{x}_0\) where the point scatterer is located. In Fig.~\ref{eigftn}, we compare some eigenfunctions localized by a point scatterer (right-hand column) to those of the Dirichlet Laplacian (left-hand column) where \(E=10\pi\). These modes are examples in which the point scatterer acts as a barrier confining the amplitude distribution to the left or right of itself.
Instead of presenting the amplitude distribution of every localized eigenfunction on the plate \(\Omega\), let us draw a graph of the \(L^2\)-norm ratio \(R_1(n,\alpha)\) as a function of the mode number \(n\) for each \(E\) fixed. The eigenfunction \(\psi_{n,\alpha}\) is considered to be localized in terms of the \(L^2\)-norm ratio if \(R_1(n,\alpha)<0.1\) or \(R_1(n,\alpha)>0.9\).
Fig.~\ref{PR} compares \(R_1(n,\alpha)\) of the first 500 eigenfunctions of \(-\Delta_{\alpha,\mathbf{x}_0}\) to those of the Dirichlet Laplacian where \(E=\frac{\pi}{3}\) and \(E= 10\pi\). For each \(E\), \(\alpha\) is chosen to maximize the number of localized modes. The blue (black) points and green (gray) points represent the eigenvalues given by Theorem~\ref{pert} and \ref{unpert}, respectively. Note that if the modes are localized completely to the right or the left of \(\mathbf{x}_0\) then all points in the graph will be polarized to either 0 or 1. When eccentricity is small (\(E=\frac{\pi}{3}\)), the point scatterer weakly perturbs the \(L^2\)-norm ratio of modes but it is hard to say these modes are localized enough. On the other hand, when eccentricity is large (\(E=10\pi\)), one can observe a strong localization especially at the low-level modes. Video clips for the continuous transition of Fig.~\ref{PR} from the Dirichlet Laplacian to a point scatterer for \(E=\frac{\pi}{3}, ~E=10\pi\) are given in \url{http://math.berkeley.edu/~lmj0425/seba_PR_pi3.avi} and \url{http://math.berkeley.edu/~lmj0425/seba_PR_10pi.avi}, respectively. Note that the Dirichlet Laplacian is equivalent to the point scatterer with \(\alpha =\infty\).
\begin{figure}
\centering
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{PR_pi3_a100.png}
\end{subfigure} \hfill
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{PR_pi3_a007.png}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{PR_10pi_a100.png}
\end{subfigure} \hfill
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=\textwidth]{PR_10pi_a007.png}
\end{subfigure}
\caption{(Color online) \(L^2\)-norm ratio \(R_1\) of the first 500 modes with a point scatterer at \(\mathbf{x}_0\). \(\alpha\) is chosen to maximize the number of localized modes.
One can observe that the \(L^2\)-norm ratio tends to polarize to either 0 or 1 for a strong (e.g. \(|\alpha|\ll\infty\)) point scatterer. Such a tendency appears remarkably to perturbed lower level modes in (d) as the eccentricity \(E\) increases.}\label{PR}
\end{figure}
Now we discuss how far the localization in terms of the \(L^2\)-norm ratio maintains its influence up to the higher-level eigenfunctions. It has been proved by Keating \emph{et al.}. \cite{local} that the eigenfunctions of \v{S}eba billiards tend to localize around eight points in momentum space as the level of the mode increases. In other words, the localization in position space we observe in this paper is an intermediate phenomenon that tends to diminish as the mode number increases.
However, one can observe that the localization effect extends to higher-level eigenfunctions as the eccentricity \(E\) increases. Fig.~\ref{locmodes_ecc} displays the number of localized modes out of the first 500 modes as a function of eccentricity \(E\). The coupling constant \(\alpha\) is chosen to maximize the number of localized modes for each \(E\). Therefore, we can conclude that the point scatterer induces a strong localization as a barrier confining the amplitude of low-level eigenfunctions to either \(\Omega_1\) or \(\Omega\setminus\Omega_1\).
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{locmodes-ecc.png}
\caption{The number of localized modes (\(R_1 <0.1\) or \(R_1 >0.9\) ) out of the first 500 modes on the plate of eccentricity \(E\). For each \(E\), the coupling constant \(\alpha\) is chosen to maximize the number of localized modes.}
\label{locmodes_ecc}
\end{figure}
\subsection{Point scatterer acting as an attractor}
On the other hand, the eigenfunction corresponding to the lowest eigenvalue \(z_1(\alpha) \in (-\infty,E_1)\) shows a different behavior: It tends to localize around \(\mathbf{x}_0\) so we can say the point scatterer attracts the amplitude of the first mode.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{mode1_eccpi3_a-048.png}
\caption{(Color online) Mode 1 of a point scatterer at \(\mathbf{x}_0\) on a plate with \(E=\frac{\pi}{3}\) and \(\alpha=-0.48\). The associated eigenvalue is \(z_1(\alpha)=-1.29\times 10^4\). The amplitude is highly localized around the point scatterer but not biased to either the left or the right of it.}
\label{mode1}
\end{figure}
A numerical simulation indicates that the amplitude at \(\mathbf{x}_0\) mainly depends on the mode number. In particular, the first mode with the associated eigenvalue \(z_1(\alpha) \in (-\infty, E_1)\) tends to localize around \(\mathbf{x}_0\) as \(z_1(\alpha)\rightarrow -\infty\), or equivalently, as \(\alpha\rightarrow -\infty\). Fig.~\ref{mode1} shows the eigenfunction of \(-\Delta_{\alpha,\mathbf{x}_0}\) corresponding to \(z_1(\alpha)=-1.29\times 10^4\) with \(\alpha=-0.48\). Since the amplitude localizes around the point scatterer evenly, our first criterion using the \(L^2\)-norm ratio cannot detect this type of localization. So we introduce the second measurement \(A(n,\alpha)\), the amplitude of the mode at \(\mathbf{x}_0\), to investigate the behavior described above.
Fig.~\ref{Amp_a} displays how the presence of the point scatterer with the coupling constant \(\alpha\) affects \(A(n,\alpha)\) of the first four modes where \(E=\frac{\pi}{3}\) [green (gray) lines] and \(E=10\pi\) [blue (black) lines]. Regardless of the eccentricity, the amplitude of the first mode at \(\mathbf{x}_0\) blows up as \(\alpha\rightarrow -\infty\) but such localization does not occur in the other modes. This can be justified by the Fourier series representation of the perturbed eigenfunction in Eq.~\eqref{psina} since, for each \(\phi_{n}\), the Fourier coefficients \[\frac{\phi_{n}(\mathbf{x}_0)}{E_{n}-z}\] get relatively uniform as \(z\rightarrow -\infty\). On the other hand, if \(E_j\le z \le E_{j+1}\) for some \(j\ge 1\), then the Fourier coefficients corresponding to \(E_n\)'s near \(z\) prevail in the summation which prevents the amplitude of higher modes from diverging at a certain point.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{Amp_a.png}
\caption{(Color online) Amplitude of the first four modes of at \(\mathbf{x}_0\) for the coupling constant \(\alpha \in [-5,5]\). Green (gray) lines and blue (black) lines correspond to the eccentricities \(E=\frac{\pi}{3}\) and \(E=10\pi\), respectively. The first mode tends to localize around the point scatterer as \(\alpha\rightarrow -\infty\) regardless of the eccentricity while the others maintain low amplitude for all \(\alpha\)'s.}
\label{Amp_a}
\end{figure}
\section{Conclusion}
We have shown that the point scatterer placed on a plate behaves as a barrier confining the low-energy eigenfunctions. Although it has been proved that such a localization property has to diminish as the mode number increases, we can increase the number of localized modes by elongating the plate. Note that the lowest eigenfunction should be excluded from this phenomenon since the point scatterer attracts its amplitude when the corresponding eigenvalue is large and negative regardless of the eccentricity of the plate.
\begin{acknowledgements}
The author is greatly indebted to Maciej Zworski for introducing the topic with inspiring discussions. The author was supported by Samsung Scholarship.
\end{acknowledgements}
\nocite{*}
\bibliographystyle{apsrev4-1}
|
\section{Introduction and statement of results}\label{s1}
A \emph{partition} of a nonnegative integer $n$ is a non-increasing sequence of positive integers whose sum is $n$. The partition function $p(n)$ then counts the number of distinct partitions of $n$. The generating function for $p(n)$ was shown by Euler to be
\[
\sum_{n=0}^\infty p(n)q^n=\prod_{n=1}^\infty\frac1{1-q^n}=1+q+2q^2+3q^3+5q^4+7q^5+\dotsb.
\]
Ramanujan famously observed the congruences $p(5k+4)\equiv 0 \pmod 5$, $p(7k+5)\equiv 0 \pmod 7$, and $p(11k+6)\equiv 0\pmod{11}$. More recently, Ono in \cite{ono-annals} and Folsom, Kent, and Ono in \cite{folsom-kent-ono}
have used Serre's theory of modular forms modulo $p$ to prove general results about the $p$-adic behavior of $p(n)$ for all $p\geq 5$. As a consequence, the behavior of $p(n)$ is well understood modulo $p$ for all $p\geq5$.
The behavior of $p(n)$ modulo 2 and 3 is far less well understood. It is widely believed that $p(n)$ is equidistributed modulo $2$ and $3$, but little is known. Subbarao conjectured in~\cite{subbarao} that for any arithmetic progression $B\pmod{A}$, there are infinitely many integers $N\equiv B\pmod{A}$ for which $p(N)$ is even, and also infinitely many such $N$ for which $p(N)$ is odd. In~\cite{ahlgren-ono}, Ahlgren and Ono conjectured that for any arithmetic progression $B\pmod{A}$, there are infinitely many integers $N\equiv B\pmod{A}$ for which $p(N)\not\equiv0\pmod{3}$.
Ono \cite{ono} established half of Subbarao's conjecture, proving that there are infinitely many $N$ in every arithmetic progression for which $p(n)$ is even. Boylan and Ono \cite{boylan-ono} then used the local nilpotency of the Hecke algebra, as observed by Serre in \cite[p. 115]{serre}, to prove the odd case of Subbarao's conjecture in the case $A=2^s$. Radu proves the full conjecture, along with Ahlgren and Ono's for modulo $3$, in~\cite{radu}, using work \cite{deligne-rapoport} of Deligne and Rapoport that applies the structure of the Tate curve to study the Fourier coefficients of modular forms.
In this note, we adapt the method of Boylan and Ono to provide a simpler, more explicit proof of Ahlgren and Ono's conjecture in the case $A=3^s$. We rely on Bella\"iche and Khare's generalization \cite{bellaiche-khare} of Serre's explicit description \cite{nicolas-serre1, nicolas-serre2} of the Hecke algebra for modulo $2$ reductions of level 1 modular forms. We are then able to show the following.
\begin{theorem}\label{mainthm3}
Let $s$ and $a$ be positive integers. There are infinitely many positive integers $n\equiv a\pmod {3^s}$ such that $p(n)\not\equiv 0\pmod 3$.
\end{theorem}
In Section~\ref{nilpotency}, we discuss Bella\"iche and Khare's work. In Section~\ref{proofmain}, we prove Theorem~\ref{mainthm3}.
\section{Local nilpotency}
\label{nilpotency}
Throughout, let $\Delta=\Delta(z)$ denote the discriminant modular form, and $\tau(n)$ denote the Fourier coefficients of $\Delta$. Let $p$ be a prime number, and let $T_p$ be the $p$th Hecke operator, which by definition acts on a modular form $f(z)=\sum_{n=0}^\infty a(n)q^n$ of weight $k$ by
\[
T_p f(z)=\sum_{n=0}^\infty(a(np)+p^{k-1}a(n/p))q^n
\]
where $a(x)=0$ if $x\notin\mathbb{Z}$. In~\cite{nicolas-serre1}, Nicolas and Serre compute for each $f$ the minimal integer $g=g(f)$ such that for any $g$ primes $p_1,\dotsc,p_g$, we have
\[
T_{p_1}\dotsb T_{p_g} f(z)\equiv0\pmod{2}.
\]
The statement that the algebra generated by the Hecke operators is locally nilpotent modulo $2$ just means that $g(f)$ is finite for every $f$. Additionally, Nicolas and Serre prove in~\cite{nicolas-serre2} that each $T_p$ can be written as a power series in $T_3$ and $T_5$, thereby permitting the enumeration of all sets of primes $p_1,\dotsc,p_{g-1}$ such that $T_{p_1}\dotsb T_{p_{g-1}} f(z)$ is \emph{not} zero modulo $2$. As a result of this description of the Hecke algebra modulo $2$, Boylan and Ono's method in~\cite{boylan-ono} becomes completely explicit.
Since our goal is to recreate Boylan and Ono's work in the modulo $3$ case, we replace Nicolas and Serre's conclusions with the following extension by Bella\"iche and Khare. Write $T_p'=T_p$ if $p\equiv2\pmod3$, and $T_p'=1+T_p$ if $\ell\equiv1\pmod3$. Then the operators $T_p'$ act locally nilpotently on the ring $S(\mathbb{F}_3)$ of level 1 cusp forms with integer coefficients taken modulo 3, in the sense that given such a modular form $f$, there is some minimal integer $g$ such that $T'_{p_1}\cdots T'_{p_g} f=0$ for any sequence of Hecke operators $T_{p_1}',\ldots, T_{p_g}'$. As such, since $\Delta$ and $-\Delta$ are the only cusp forms modulo 3 satisfying $T_p'\vert f=0$ for all $p$, we have some maximal sequence $p_1,\ldots, p_{g-1}$ such that $T_{p_1}'\cdots T_{p_{g-1}}' f=\pm\Delta$. Moreover, we have the following description of the Hecke algebra on modular forms with coefficients in $\mathbb F_3$.
\begin{theorem}[\cite{bellaiche-khare}, Theorem 24]\label{bk}
The algebra of Hecke operators on $S(\mathbb{F}_3)$ is isomorphic to the power series ring $\mathbb{F}_3[[x,y]]$, with an isomorphism given by sending $T_2'=T_2$ to $x$ and $T_7'=1+T_7$ to $y$.
Assuming this identification, we have $T_p'\equiv x\pmod{(x,y)^2}$ if and only if $p$ is congruent to $2\pmod3$ but not $8\pmod9$,
$T_p'\equiv y\pmod{(x,y)^2}$ if and only if $p$ is congruent to $1\pmod3$ and not split in the splitting field of $X^3-3$, and otherwise $T_p'\equiv 0\pmod{(x,y)^2}$.
\end{theorem}
This theorem in particular implies that for any nonzero $f\in S(\mathbb{F}_3)$ there are some positive integers $k$ and $\ell$ such that $(T_{2}')^k (T_7')^\ell f=\pm\Delta$. We may now proceed according to Boylan and Ono's strategy.
\section{Partitions modulo 3} \label{proofmain}
In this section we prove Theorem \ref{mainthm3}. We start by proving a basic lemma, similar to Corollary 1.4 of \cite{radu}.
\begin{lemma}
\label{manyfromone}
Suppose that for every $s$, every arithmetic progression modulo $3^s$ contains at least one $N$ such that $p(N)\not\equiv0\pmod{3}$. Then for every $s$, every arithmetic progression modulo $3^s$ contains infinitely many such $N$.
\end{lemma}
\begin{proof}
We prove the contrapositive. Suppose there exist some $r$ and $s$, with $0\le r<3^s$, for which there are only finitely many $N\equiv r\pmod{3^s}$ with $p(N)\not\equiv0\pmod3$. Then there is some $a_0\ge0$ such that $p(3^sa+r)\equiv0\pmod3$ for all $a\ge a_0$. Let $t$ be such that $3^t>3^sa_0+r$. Then we have
\[
p(3^s\cdot 3^ta+3^sa_0+r)\equiv0\pmod3
\]
for all $a\ge0$, from which we conclude that $p(N)\equiv0\pmod3$ for all $N\equiv 3^sa_0+r\pmod{3^{s+t}}$.
\end{proof}
\begin{comment}
\begin{theorem}
\label{one}
For every $s$, every arithmetic progression modulo $3^s$ contains at least one $N$ such that $p(N)\not\equiv0\pmod{3}$.
\end{theorem}
\end{comment}
Let the integers $a_s(n)$ be defined by the generating function $\sum_{n=0}^{\infty}a_s(n)q^n=\Delta^{\frac{9^s-1}8}$, and let $r_s(j)$ be defined by $1+\sum_{n=1}^\infty r_s(j)q^{3\cdot9^sj}=\prod_{n=1}^\infty(1-q^{3\cdot9^s})$. Our next lemma is similar to Lemma 2.1 of~\cite{boylan-ono}.
\begin{lemma}
\label{recursion}
We have
\[
a_s(n)\equiv p\left(\frac{n-\frac{9^s-1}8}{3}\right)+\sum_{j=1}^\infty r_s(j)p\left(\frac{n-\frac{9^s-1}8}3-9^sj\right)\pmod3.
\]
\end{lemma}
\begin{proof}
We may compute
\begin{align*}
\Delta^{\frac{9^s-1}8} &= \left(q\prod_{n=1}^\infty(1-q^n)^{24}\right)^{\frac{9^s-1}8} \\
&\equiv q^{\frac{9^s-1}8}\prod_{n=1}^\infty(1-q^{3\cdot 9^sn})\prod_{n=1}^\infty\frac1{1-q^{3n}} \pmod 3 \\
&\equiv \left(1+\sum_{j=1}^\infty r_s(j)q^{3\cdot 9^sj}\right)\left(\sum_{k=0}^\infty p(k)q^{3k+\frac{9^s-1}8}\right) \pmod 3.
\end{align*}
The lemma follows by comparing coefficients.
\end{proof}
Finally, we require the following lemma concerning nonzero coefficients of $\Delta^{\frac{9^s-1}8}$.
\begin{lemma}
\label{nonzero}
There are fixed nonnegative integers $k$ and $\ell$, not both zero, such that the following holds. Let $p_1,\dotsc,p_k$ be distinct primes which satisfy $p_i\equiv 2\pmod3$ and $p_i\not\equiv 8\pmod9$ for all $i$ . Let $q_1,\dotsc,q_\ell$ be distinct primes such that for all $j$ we have $q_j\equiv 1\pmod{9^{s+1}}$ and $q_j$ does not split in the splitting field of $X^3-3$. Then for any $n_0$ satisfying $(n_0, p_1\cdots p_k q_1\cdots q_\ell)=1$ such that $\tau(n_0)\not\equiv 0\pmod3$, there is some $d|q_1\dotsb q_\ell$ for which we have $a_s(n_0p_1\dotsb p_kd)\not\equiv0\pmod3$.
\end{lemma}
\begin{remark}
Given values of $k$, $\ell$, and $s$, it is always possible to find corresponding primes $p_1,\ldots,p_k$ and $q_1,\ldots, q_\ell$; indeed by the Chebotarev density theorem, the primes that are valid choices for $p_i$ have density $1/3$ within the primes, and the primes that are valid choices for $q_j$ have density $1/(2\cdot 3^{s+1})$.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{nonzero}.]
Identify the Hecke algebra modulo 3 with the ring $\mathbb{F}_3[[x,y]]$ as described in Theorem \ref{bk} and let $p_1,\dotsc,p_k,q_1,\dotsc,q_\ell$ be such that $T_{p_i}'\equiv x\pmod{(x,y)^2}$ and $T_{q_j}'\equiv y\pmod{(x,y)^2}$, and that
\[
T_{p_1}'\dotsb T_{p_k}'T_{q_1}'\dotsb T_{q_\ell}'\Delta(z)^{\frac{9^s-1}8}\equiv\pm\Delta\pmod{3}.
\]
Note that having fixed $k$ and $l$ so that such $p_i$s and $q_j$s exist, any such sequence of primes satisfying the hypotheses of Lemma~\ref{nonzero} satisfy the same equation. By comparing coefficients, we conclude that for any $n_0$ not divisible by the $p_i$s and $q_j$s with a nonzero corresponding coefficient of $\Delta$, we have
\begin{equation}
\sum_{d|q_1\dotsb q_\ell}a_s(n_0p_1\dotsb p_kq_1\dotsb q_\ell)\not\equiv0\pmod 3.
\end{equation}
Lemma ~\ref{nonzero} follows.
\end{proof}
Finally, combining Lemma~\ref{nonzero} with Lemmas~\ref{manyfromone} and~\ref{recursion}, we prove Theorem~\ref{mainthm3}.
\begin{proof}[Deduction of Theorem~\ref{mainthm3} from Lemma~\ref{nonzero}.]
Choose $k$ and $\ell$ such that $(T_2')^k (T_7')^\ell \Delta^{\frac{9^s-1}{8}}=\pm \Delta$ let $p_1,\ldots,p_k$, $q_1,\ldots, q_\ell$ be as in the statement of Lemma \ref{nonzero}. Then, by Lemma \ref{nonzero}, given any $n_0$ with $\tau(n_0)\not\equiv 0\pmod 3$ and $(n_0,p_1\cdots p_k q_1\cdots q_\ell)=1$ we have some $d\vert q_1\cdots q_\ell$ such that $a_s(n_0p_1\cdots p_k d)\not\equiv 0\pmod 3$. By Lemma~\ref{recursion}, for this $d|q_1\dotsb q_\ell$, we may write
\[
a_s(n_0p_1\dotsb p_kd)\equiv p\left(\frac{n_0p_1\dotsb p_kd-\frac{9^s-1}8}{3}\right)+\sum_{j=1}^\infty r_s(j)p\left(\frac{n_0p_1\dotsb p_kd-\frac{9^s-1}8}3-9^sj\right)\pmod3.
\]
Since $a_s(p_1\dotsb p_kd)\not\equiv0\pmod3$, we conclude that $p\left(\frac{p_1\dotsb p_kd-\frac{9^s-1}8}3-9^sj\right)\not\equiv0\pmod3$ for some $j\geq 0$, so for some $n$ satisfying $n\equiv \frac{p_1\dotsb p_kd-\frac{9^s-1}8}3 \pmod{9^s}$ we have $p(n)\not \equiv 0\pmod 3$. Hence, by Lemma \ref{manyfromone} it suffices to show that as we vary $n_0$ the quantity $\frac{n_0p_1\dotsb p_kd-\frac{9^s-1}8}3$ covers all residue classes modulo $9^s$. Since $d\equiv1\pmod{9^{s+1}}$ regardless of its precise factorization, it may be dropped. To show that $\frac{n_0p_1\dotsb p_k-\frac{9^s-1}8}3$ covers all residue classes as we vary $n_0$, we note first note the standard congruence for the $\tau$ function,
$$
\tau(n)\equiv \sigma_1(n) \pmod 3,
$$
is valid for all $n\not \equiv 0\pmod 3$, and in particular $\tau(p)\equiv 2\pmod 3$ when $p\equiv 1\pmod 3$. In addition, we note that $k$ must be even, since $a_s(n)\not\equiv 0\pmod 3$ implies $n\equiv 1 \pmod 3$, but if $k$ is odd we have $a_s(p_1\cdots p_k D)\not\equiv 0 \pmod 3$ even though $p_1\cdots p_k D \equiv 2 \pmod 3$, a contradiction. So we have that $p_1\cdots p_k\equiv 1 \pmod 3$, so to show that the terms $\frac{n_0p_1\dotsb p_k-\frac{9^s-1}8}3$ cover all residue classes we need only find $n_0$ in all residue classes $n\pmod{ 9^{s+1}}$ satisfying $n\equiv 1\pmod 3$. But to do so we may simply take $n_0$ to be any prime satisfying $n_0\equiv n \pmod {9^{s+1}}$, of which there are infinitely many. Theorem \ref{mainthm3} follows by applying Lemma \ref{manyfromone}.
\end{proof}
\begin{comment}
\begin{lemma}
There are infinitely many primes $q$ satisfying the hypotheses of Lemma~\ref{nonzero}, that is, are $1\pmod{9^{s+1}}$ and do not split in the splitting field of $X^3-3$.
\end{lemma}
\begin{proof}
This follows from applying the Chebotarev Density Theorem to $\mathbb{Q}[\sqrt[3]{3},\zeta]$ where $\zeta$ is a primitive $9^{s+1}$th root of unity, since $p$ splits in $\mathbb{Q}[\zeta]$ precisely when $p\equiv1\pmod{9^{s+1}}$. We find that the proportion of such primes is $\frac{2}{3(9^{s+1}-3^{2s+1})}$. [NOTE: or something like that]
\end{proof}
\end{comment}
\nocite{*}
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:introduction}
In many recent developments involving cosmological models and brane
collisions in higher-dimensional gravity theories
\cite{Lu:1996jk, Behrndt:2003cx, Gibbons:2005rt, Chen:2005jp, Kodama:2005fz,
Kodama:2005cz, Kodama:2006ay, Arroja:2006zz, Koyama:2006ni, Arroja:2007ss,
Binetruy:2007tu, Binetruy:2008ev, Arroja:2008ma, Maeda:2009tq,
Maeda:2009zi,Gibbons:2009dr, Maeda:2009ds, Uzawa:2010zza, Minamitsuji:2010fp,
Maeda:2010ja, Maeda:2010aj, Minamitsuji:2010kb, Nozawa:2010zg,
Minamitsuji:2010uz, Minamitsuji:2011jt, Uzawa:2010zz, Maeda:2011sh,
Maeda:2012xb, Minamitsuji:2012if, Blaback:2012mu,
Uzawa:2013koa, Uzawa:2013msa, Uzawa:2014kka},
the dynamical $p$-branes carrying charges have played
important roles. In the classical solution of a single $p$-brane, the
coupling of the dilaton to the field strength includes the parameter $N$.
Since these brane solutions with $N=4$ are related to well-known
D-branes and M-branes in supergravity theories, they certainly exhibit
many attractive properties in the higher-dimensional spacetime.
Some static solutions with $N\ne 4$ also have supersymmetry after
dimensional reductions to lower-dimensional theory
\cite{Lu:1995cs, Cvetic:2000dm}. The time-dependent
generalizations of these solutions are thus important examples
of higher-dimensional gravity theories.
The dynamical brane solution with the cosmological constant can be obtained
by choosing the coupling constant appropriately
\cite{Minamitsuji:2010fp, Maeda:2010aj, Minamitsuji:2010uz}.
For a single 2-form field strength and a nontrivial dilaton, we have
found that the dynamical single 0-brane solution describes the
Milne universe \cite{Maeda:2010aj, Minamitsuji:2010uz}.
The field equations give an asymptotically de Sitter solution if
the scalar field is trivial \cite{Minamitsuji:2010uz},
which is a generalization of the Kastor-Traschen
solution in the four-dimensional Einstein-Maxwell theory
\cite{Kastor:1992nn, Brill:1993tm}.
The construction of intersecting branes with a cosmological constant
is a natural generalization of the single cosmological brane solutions.
The time-dependent intersecting branes we have mainly discussed
are localized only along the relative or overall transverse directions
in a higher-dimensional background, which are delocalized intersecting brane
systems. However, in the higher-dimensional gravity theories,
one of the branes is localized along the relative transverse directions
but delocalized along the overall transverse directions, which are
partially localized branes solutions. If the background has the
cosmological constant, there is little known about the dynamics of the
intersecting brane system for not only the delocalized case
but also the partially localized one.
In the present paper,
we will explore the possible generalization of these solutions
to the case of the intersecting brane systems
with cosmological constants, although
similar single brane solutions have been
analyzed in Ref.~\cite{Minamitsuji:2010uz}.
We recall these
arguments for constructions of the solution and
modify the ansatz of the fields.
A brane configuration has to satisfy an
intersection rule which is
an algebraic equation that relates the coupling of the dilaton to
the dimensionality of the branes.
The intersection rule implies that only the 0-brane can depend on time
and the dynamical 0-brane commutes with the static $p$-branes.
We will study the dynamical intersecting brane solutions for
not only the delocalized case but also the partially localized one.
The paper is constructed as follows:
In Secs. \ref{sec:pl} and \ref{sec:cn},
we derive the dynamical intersecting brane solutions
with cosmological constants in a $D$-dimensional theory
following the approach developed in Ref.~\cite{Minamitsuji:2010uz}.
We then illustrate how the dynamical solution
of two or $n$ intersecting branes arise
under the condition of $N\ne 4$
in the $D$-dimensional theory.
The spacetime starts with the structure of
the combined 0-branes.
If they do not have the same charges,
a singularity hypersurface appears before they meet
as the time decreases for $D>4$.
We then discuss the dynamics of two 0-branes with static $p$-branes
(or the dynamics of two black holes) in Sec. \ref{sec:c}.
If there exists one uncompactified extra dimension [0$-$8-brane
system or 0$-(D-1)$-brane systems ($p\le 7$)] and two brane systems have
the same brane charges, the solution describes a
collision of two branes (or two black holes),
which is similar to the result in
Refs.~\cite{Gibbons:2005rt, Maeda:2010aj, Maeda:2012xb}.
In Sec.~\ref{sec:sug}, applications of these solutions to five- or
six-dimensional
supergravity models are discussed. We consider in detail the construction
yielding the dynamical 0- or 1-brane in the Nishino-Salam-Sezgin model.
We also provide brief discussions for a time-dependent brane system in
Romans' supergravity model. We describe how our Universe could be
represented in the present formulation via an appropriate compactification
and give the application to cosmology.
We show that there exists no accelerating expansion of our Universe,
although the conventional power-law expansion of the Universe is possible.
We then discuss the dynamics of two 0- or 1-branes with smeared branes.
If two brane systems have the same brane charges
with smearing some dimensions,
the solution describes a collision of two brane backgrounds.
There is a curvature singularity in the dynamical brane background if we
set a particular value for the constant parameters. Then the solution implies
that the presence of the singularities is signaling
possible instabilities, making the solutions sick or unphysical.
We study the classical stability of the solutions in Sec.~\ref{sec:in}.
Our preliminary analysis will present that the energy
of Klein-Gordon scalar fields in the dynamical brane background grows
with time for inertial observers approaching the singularity.
In terms of using the preliminary analysis performed in
Refs.~\cite{oai1, Quevedo:2002tm, oai2, Cornalba:2003ze},
the Klein-Gordon modes will be studied, arriving at the preliminary
conclusion of instability.
Section \ref{sec:cd}
will be devoted to the summary and conclusions.
\section{Dynamical partially localized intersecting brane backgrounds
with cosmological constants}
\label{sec:pl}
In this section,
we will construct the partially localized time-dependent brane systems
in $D$ dimensions with cosmological constants.
We consider a $D$-dimensional theory
composed of the metric $g_{MN}$, the scalar field $\phi$,
cosmological constants $\Lambda_I~(I=r,~s)$,
and two antisymmetric tensor field strengths
of rank $(p_r+2)$ and $(p_s+2)$.
The action in $D$ dimensions is
given by
\Eqr{
S&=&\frac{1}{2\kappa^2}\int \left[\left(R-2\e^{\alpha_r\phi}
\Lambda_r-2\e^{\alpha_s\phi}\Lambda_s\right)\ast{\bf 1}_D
-\frac{1}{2}\ast d\phi \wedge d\phi\right.\nn\\
&&\left.\hspace{-0.5cm}
-\frac{1}{2}\frac{1}{\left(p_r+2\right)!}
\e^{\epsilon_rc_r\phi}\ast F_{(p_r+2)}\,\wedge\,
F_{\left(p_r+2\right)}
-\frac{1}{2}\frac{1}{\left(p_s+2\right)!}
\e^{\epsilon_sc_s\phi}\ast F_{(p_s+2)}\,\wedge\, F_{(p_s+2)}
\right],
\label{pl:action:Eq}
}
where $R$ denotes the Ricci scalar constructed from the $D$-dimensional
metric $g_{MN}$\,, $\alpha_I~(I=r,~s)$ are constants,
$\kappa^2$ denotes the $D$-dimensional gravitational constant,
$\ast$ is the Hodge operator in the $D$-dimensional spacetime, and
$F_{\left(p_r+2\right)}$ and $F_{\left(p_s+2\right)}$
are $\left(p_r+2\right)$ and
$\left(p_s+2\right)$-form field strengths, respectively.
The constant parameters $c_I$
and
$\epsilon_I~(I=r,~s)$ are defined by
\Eqrsubl{pl:parameters:Eq}{
c_I^2&=&N_I-\frac{2(p_I+1)(D-p_I-3)}{D-2},
\label{pl:c:Eq}\\
\epsilon_I&=&\left\{
\begin{array}{cc}
+&{\rm if~the}~p_I-{\rm brane~is~electric}\\
-&~~~{\rm if~the}~p_I-{\rm brane~is~magnetic}\,,
\end{array} \right.
\label{pl:epsilon:Eq}
}
respectively.
Here $N_I$ is constant.
The $(p_r+2)$-form, $(p_s+2)$-form
field strengths $F_{(p_r+2)}$, $F_{(p_s+2)}$
are given by the
$\left(p_r+1\right)$-form, $\left(p_s+1\right)$-form gauge
potentials $A_{\left(p_r+1\right)}$,
$A_{\left(p_s+1\right)}$, respectively:
\Eq{
F_{(p_r+2)}=dA_{(p_r+1)}\,,~~~~~F_{(p_s+2)}=dA_{(p_s+1)}\,.
}
For the $D$-dimensional action (\ref{pl:action:Eq}),
the field equations read
\Eqrsubl{pl:equations:Eq}{
&&\hspace{-0.5cm}R_{MN}=\frac{2}{D-2}\left(\e^{\alpha_r\phi}\Lambda_r
+\e^{\alpha_s\phi}\Lambda_s\right)g_{MN}
+\frac{1}{2}\pd_M\phi \pd_N \phi\nn\\
&&+\frac{1}{2}\frac{\e^{\epsilon_rc_r\phi}}
{\left(p_r+2\right)!}
\left[\left(p_r+2\right)
F_{MA_2\cdots A_{\left(p_r+2\right)}}
{F_N}^{A_2\cdots A_{\left(p_r+2\right)}}
-\frac{p_r+1}{D-2}\,g_{MN}\,F^2_{\left(p_r+2\right)}\right]\nn\\
&&+\frac{1}{2}\frac{\e^{\epsilon_sc_s\phi}}
{\left(p_s+2\right)!}
\left[\left(p_s+2\right)
F_{MA_2\cdots A_{\left(p_s+2\right)}}
{F_N}^{A_2\cdots A_{\left(p_s+2\right)}}
-\frac{p_s+1}{D-2}\,g_{MN}\,F_{\left(p_s+2\right)}^2\right],
\label{pl:Einstein:Eq}\\
&&\hspace{-0.5cm}\lap\phi-\frac{1}{2}\frac{\epsilon_rc_r}
{\left(p_r+2\right)!}
\e^{\epsilon_rc_r\phi}F^2_{\left(p_r+2\right)}
-\frac{1}{2}\frac{\epsilon_sc_s}{\left(p_s+2\right)!}
\e^{\epsilon_sc_s\phi}F^2_{\left(p_s+2\right)}
-2\alpha_r\e^{\alpha_r\phi}\Lambda_r
-2\alpha_s\e^{\alpha_s\phi}\Lambda_s=0\,,
\label{pl:scalar:Eq}\\
&&\hspace{-0.5cm}
d\left[\e^{\epsilon_rc_r\phi}\ast F_{\left(p_r+2\right)}\right]=0\,,
\label{pl:gauge-r:Eq}\\
&&\hspace{-0.5cm}
d\left[\e^{\epsilon_sc_s\phi}\ast F_{\left(p_s+2\right)}\right]=0\,,
\label{pl:gauge-s:Eq}
}
where $\lap$ denotes the Laplace operator with respect to the
$D$-dimensional metric $g_{MN}$\,.
The $D$-dimensional metric involving the intersecting branes with cosmological
constant can be put in the general form
\Eqr{
&&\hspace{-0.7cm}
ds^2=h^{a_r}_r(x, y, z)h_s^{a_s}(x, v, z)q_{\mu\nu}
(\Xsp)dx^{\mu}dx^{\nu}+h^{b_r}_r(x, y, z)h_s^{a_s}(x, v, z)\gamma_{ij}
(\Ysp_1)dy^idy^j\nn\\
&&\hspace{-0.5cm}
+h^{a_r}_r(x, y, z)h_s^{b_s}(x, v, z)w_{mn}(\Ysp_2)dv^{m}dv^{n}
+h^{b_r}_r(x, y, z)h_s^{b_s}(x, v, z)u_{ab}(\Zsp)dz^adz^b,
\label{pl:metric:Eq}
}
where $q_{\mu\nu}$ is the $(p+1)$-dimensional metric depending
only on the $(p+1)$-dimensional coordinates $x^{\mu}$,
$\gamma_{ij}$ is the $(p_s-p)$-dimensional metric depending
only on the $(p_s-p)$-dimensional coordinates $y^i$,
$w_{mn}$ is the $(p_r-p)$-dimensional metric depending
only on the $(p_r-p)$-dimensional coordinates $v^m$ and
$u_{ab}$ is the $(D+p-p_r-p_s-1)$-dimensional metric depending
only on the $(D+p-p_r-p_s-1)$-dimensional coordinates $z^a$.
Here we assume that the parameters $a_I~(I=r,~s)$ and
$b_I~(I=r,~s)$ in the metric (\ref{pl:metric:Eq}) are given by
\Eq{
a_I=-\frac{4\left(D-p_I-3\right)}{N_I\left(D-2\right)},~~~~~~~~~
b_I=\frac{4\left(p_I+1\right)}{N_I\left(D-2\right)}.
\label{pl:parameter:Eq}
}
The brane configuration is illustrated
in Table \ref{table_A}.
\begin{table}[h]
\caption{\baselineskip 14pt
Intersections of $p_r-p_s$-branes in the metric (\ref{pl:metric:Eq}),
where $p'=p_s+p_r-p$.
}
\label{DpDp}
{\scriptsize
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
Case&&0&1& $\cdots$ & $p$ & $p+1$ & $\cdots$ & $p_s$ & $p_s+1$ &
$\cdots$ & $p'$ & $p'+1$ & $\cdots$ & $D-1$
\\
\hline
&$p_r$ & $\circ$ & $\circ$ & $\circ$ & $\circ$ &&&& $\circ$ & $\circ$ &
$\circ$ &&&
\\
\cline{3-15}
$p_r-p_s$ & $p_s$ & $\circ$ & $\circ$ & $\circ$ & $\circ$ & $\circ$ &
$\circ$ & $\circ$ && & & & &
\\
\cline{3-15}
&$x^N$ & $t$ & $x^1$ & $\cdots$ & $x^p$ & $y^1$ & $\cdots$ & $y^{p_s-p}$
& $v^1$ & $\cdots$ & $v^{p_r-p}$ & $z^1$ & $\cdots$ & $z^{D-p'-1}$
\\
\hline
\end{tabular}
\end{center}
}
\label{table_A}
\end{table}
The dynamical brane solutions are characterized by
two warp factors, $h_r$ and $h_s$, depending
on the $(D+p-p_r-p_s-1)$-dimensional coordinates transverse to the
corresponding brane as well as
the $(p+1)$-dimensional world-volume coordinate.
In the case of intersection involving two branes,
the powers of warp factors have to obey the intersection rule,
and then split the coordinates in four parts
\cite{Behrndt:1996pm, Bergshoeff:1996rn, Bergshoeff:1997tt}.
One is coordinates of the world-volume spacetime, $x^{\mu}$,
which are common to the $p_r$-, $p_s$-branes.
The others are coordinates of the overall transverse space,
$z^a$, and coordinates of the relative transverse space, $y^i$ and $v^m$,
which are transverse to only one of the $p_r$-, $p_s$-branes.
In this section, we consider the intersections of a $p_r$- and a $p_s$-brane
with the following conditions in $D$-dimension.
We assume that the functions $h_r$ and $h_s$ depend not only on
overall transverse coordinates but also on the corresponding
relative coordinates and world-volume coordinates.
We therefore may write
$h_r=h_r(x, y, z),\, h_s=h_s(x, v, z)$.
We give the expression for the field strengths $F_{\left(p_r+2\right)}$,
$F_{\left(p_s+2\right)}$ and scalar field $\phi$ of a
$p_r$-brane intersecting with a $p_s$-brane over a $p$-brane configuration:
\Eqrsubl{pl:ansatz:Eq}{
\e^{\phi}&=&h_r^{2\epsilon_rc_r/N_r}\,
h_s^{2\epsilon_sc_s/N_s},
\label{pl:dilaton:Eq}\\
F_{\left(p_r+2\right)}&=&\frac{2}{\sqrt{N_r}}
d\left[h^{-1}_r(x, y, z)\right]\wedge\Omega(\Xsp)\wedge\Omega(\Ysp_2),
\label{pl:strength-r:Eq}\\
F_{\left(p_s+2\right)}&=&\frac{2}{\sqrt{N_s}}
d\left[h^{-1}_s(x, v, z)\right]\wedge\Omega(\Xsp)\wedge\Omega(\Ysp_1),
\label{pl:strength-s:Eq}
}
where $\Omega(\Xsp)$, $\Omega({\Ysp}_1)$, and $\Omega({\Ysp}_2)$
are the volume $\left(p+1\right)$-form, $\left(p_s-p\right)$-form,
and $\left(p_r-p\right)$-form, respectively
\Eqrsubl{pl:volume:Eq}{
\Omega(\Xsp)&=&\sqrt{-q}\,dx^0\,\wedge\,dx^1\wedge\,\cdots\,\wedge\,dx^p,\\
\Omega(\Ysp_1)&=&\sqrt{\gamma}\,dy^1\,\wedge\,dy^2\,\wedge\,\cdots\,\wedge
\,dy^{p_s-p},\\
\Omega(\Ysp_2)&=&\sqrt{w}\,dv^1\,\wedge\,dv^2\,\wedge\,\cdots\,\wedge
\,dv^{p_r-p}.
}
Here, $q$, $\gamma$, and $w$ denote
the determinants of the metrics $q_{\mu\nu}$,
$\gamma_{ij}$, and $w_{mn}$, respectively.
\subsection{Power-law expanding universe}
In this subsection, we consider the field Eqs.~(\ref{pl:ansatz:Eq}) with
$c_I~(I=r,~s)\neq 0$.
The parameters $\alpha_I~(I=r,~s)$ are assumed to be
\Eqr{
\alpha_r=-\epsilon_r c_r\,,~~~~~\alpha_s=-\epsilon_s c_s.
\label{pl:alpha:Eq}
}
Let us first consider the gauge field Eqs.~(\ref{pl:gauge-r:Eq}),
(\ref{pl:gauge-s:Eq}).
Using the assumptions (\ref{pl:metric:Eq})
and (\ref{pl:ansatz:Eq}), we have
\Eqrsubl{pl:gauge2:Eq}{
&&\hspace{-0.3cm}
d\left[h_s^{4(\chi+1)/N_s}\pd_i h_r\left(\ast_{\Ysp_1}dy^i\right)
\wedge\Omega(\Zsp)+h_s^{4\chi/N_s}\pd_a h_r\left(\ast_{\Zsp}dz^a\right)
\wedge\Omega(\Ysp_1)\right]=0,
\label{pl:gauge2-r:Eq}\\
&&\hspace{-0.3cm}
d\left[h_r^{4(\chi+1)/N_r}\pd_m h_s\left(\ast_{\Ysp_2}dv^m\right)
\wedge\Omega(\Zsp)+h_r^{4\chi/N_r}\pd_a h_s\left(\ast_{\Zsp}dz^a\right)
\wedge\Omega(\Ysp_2)\right]=0,
\label{pl:gauge2-s:Eq}
}
where $\ast_{\Ysp_1}$, $\ast_{\Ysp_2}$, and $\ast_{{\rm Z}}$
denote the Hodge operator on $\Ysp_1$, $\Ysp_2$, and Z, respectively,
and $\chi$ is given by
\Eq{
\chi=
p+1-\frac{\left(p_r+1\right)\left(p_s+1\right)}{D-2}
+\frac{1}{2}\epsilon_r\epsilon_sc_rc_s.
\label{pl:chi:Eq}
}
In the following, we discuss the case of $\chi=0$ because the
relation $\chi=0$ is consistent with the intersection rule
which has been found in
\cite{Binetruy:2007tu, Maeda:2009zi, Minamitsuji:2010kb, Minamitsuji:2010uz,
Strominger:1995ac, Townsend:1995af, Douglas:1995bn,
Papadopoulos:1996uq, Tseytlin:1996bh, Gauntlett:1996pb,
Cvetic:1996gq, Papadopoulos:1996ca,
Russo:1996if, Argurio:1997gt, Tseytlin:1997cs, Ohta:1997gw,
Youm:1997hw, Argurio:1998cp, Miao:2004bn,
Chen:2005uw}.
Setting $\chi=0$\,, the Eq.~(\ref{pl:gauge2-r:Eq}) gives
\Eq{
h_s\lap_{\Ysp_1}h_r+\lap_{\Zsp}h_r=0,
~~~\pd_{\mu}\pd_i h_r+\frac{4}{N_s}\pd_{\mu}\ln h_s\pd_ih_r=0,
~~~\pd_{\mu}\pd_a h_r=0,
\label{pl:gauge3:Eq}
}
where $\lap_{\Ysp_1}$, and $\lap_{\Zsp}$ are
the Laplace operators on the space of $\Ysp_1$, and
Z, respectively.
On the other hand, Eq.~(\ref{pl:gauge2-s:Eq}) leads to
\Eq{
h_r\lap_{\Ysp_2}h_s+\lap_{\Zsp}h_s=0,
~~~\pd_{\mu}\pd_m h_s+\frac{4}{N_r}\pd_{\mu}\ln h_r\pd_mh_s=0,
~~~\pd_{\mu}\pd_a h_s=0\,,
\label{pl:gauge4:Eq}
}
where we used Eq.~(\ref{pl:chi:Eq}) and $\triangle_{\Ysp_2}$ is
the Laplace operators on the space of $\Ysp_2$,
Now we consider the Einstein Eq.~(\ref{pl:Einstein:Eq}).
Using the ansatz (\ref{pl:metric:Eq}), (\ref{pl:ansatz:Eq}),
and intersection rule $\chi=0$, the Einstein equations become
\Eqrsubl{pl:cEinstein:Eq}{
&&\hspace{-0.5cm}R_{\mu\nu}(\Xsp)
-\frac{4}{N_r}h_r^{-1}D_{\mu}D_{\nu}h_r
-\frac{4}{N_s}h_s^{-1}D_{\mu}D_{\nu}h_s
+\frac{2}{N_r}\pd_{\mu}\ln h_r\left[\left(1-\frac{4}{N_r}\right)
\pd_{\nu}\ln h_r-\frac{4}{N_s}\pd_{\nu}\ln h_s\right]\nn\\
&&~~~~+\frac{2}{N_s}\pd_{\mu}\ln h_s\left[\left(1-\frac{4}{N_s}\right)
\pd_{\nu}\ln h_s-\frac{4}{N_r}\pd_{\nu}\ln h_r\right]\nn\\
&&~~~~-\frac{2}{D-2}\left(\Lambda_rh_r^{-2+a_rp_r}
h_s^{a_s-2\epsilon_r\epsilon_sc_rc_s/N_s}
+\Lambda_sh_r^{a_r-2\epsilon_r\epsilon_sc_rc_s/N_r}
h_s^{-2+a_sp_s}\right)q_{\mu\nu}\nn\\
&&~~~~-\frac{1}{2}q_{\mu\nu}h_r^{-4/N_r}h_s^{-4/N_s}
\left[a_rh_r^{-1}\left(h_s^{4/N_s}\lap_{\Ysp_1}h_r+\lap_{\Zsp}h_r\right)
+a_sh_s^{-1}\left(h_r^{4/N_r}\lap_{\Ysp_2}h_s+\lap_{\Zsp}h_s\right)\right]\nn\\
&&~~~~-\frac{1}{2}q_{\mu\nu}\left[a_r
h_r^{-1}\lap_{\Xsp}h_r-a_rq^{\rho\sigma}\pd_{\rho}\ln h_r
\left\{\left(1-\frac{4}{N_r}\right)
\pd_{\sigma}\ln h_r-\frac{4}{N_s}
\pd_{\sigma}\ln h_s\right\}\right.\nn\\
&&\left. ~~~~+a_sh_s^{-1}\triangle_{\Xsp}h_s
-a_sq^{\rho\sigma}\pd_{\rho}\ln h_s\left\{\left(1-\frac{4}{N_s}\right)
\pd_{\sigma}\ln h_s-\frac{4}{N_r}\pd_{\sigma}\ln h_r\right\}\right]=0,
\label{pl:cEinstein-mu:Eq}\\
&&\hspace{-0.5cm}\frac{2}{N_r}h_r^{-1}\left(\pd_{\mu}\pd_i h_r
+\frac{4}{N_s}\pd_{\mu}\ln h_s\pd_ih_r\right)=0,~~~~
\frac{2}{N_s}h_s^{-1}\left(\pd_{\mu}\pd_m h_s
+\frac{4}{N_r}\pd_{\mu}\ln h_r\pd_mh_s\right)=0,
\label{pl:cEinstein-mm:Eq}\\
&&\hspace{-0.5cm}\frac{2}{N_r}h_r^{-1}\pd_{\mu}\pd_a h_r
+\frac{2}{N_s}h_s^{-1}\pd_{\mu}\pd_a h_s=0,
\label{pl:cEinstein-ma:Eq}\\
&&\hspace{-0.5cm}
R_{ij}(\Ysp_1)-\frac{1}{2}h_r^{4/N_r}\gamma_{ij}\left[b_r
h_r^{-1}\lap_{\Xsp}h_r-b_rq^{\rho\sigma}\pd_{\rho}\ln h_r
\left\{\left(1-\frac{4}{N_r}\right)
\pd_{\sigma}\ln h_r-\frac{4}{N_s}
\pd_{\sigma}\ln h_s\right\}\right.\nn\\
&&\left. ~~+a_sh_s^{-1}\triangle_{\Xsp}h_s
-a_sq^{\rho\sigma}\pd_{\rho}\ln h_s\left\{\left(1-\frac{4}{N_s}\right)
\pd_{\sigma}\ln h_s-\frac{4}{N_r}\pd_{\sigma}\ln h_r\right\}\right]\nn\\
&&~~-\frac{1}{2}\gamma_{ij}h_s^{-4/N_s}\left\{b_rh_r^{-1}
\left(h_s^{4/N_s}\lap_{\Ysp_1}h_r+\lap_{\Zsp}h_r\right)
+a_sh_s^{-1}\left(h_r^{4/N_r}\lap_{\Ysp_2}h_s+\lap_{\Zsp}h_s\right)
\right\}\nn\\
&&~~-\frac{2}{D-2}\left[\Lambda_rh_r^{-2+a_rp_r+4/N_r}
h_s^{a_s-2\epsilon_r\epsilon_sc_rc_s/N_s}
+\Lambda_sh_r^{a_r-2(\epsilon_r\epsilon_sc_rc_s-2)/N_r}
h_s^{-2+a_sp_s}\right]\gamma_{ij}=0,
\label{pl:cEinstein-ij:Eq}\\
&&\hspace{-0.5cm}
\frac{8}{N_rN_s(D-2)^2}\left[(p_r+1)(p_s+1)-(D-2)(p_r+p_s+2)
\right]\pd_i\ln h_r\pd_m\ln h_s=0\,,
\label{pl:cEinstein-im:Eq}\\
&&\hspace{-0.5cm}R_{mn}(\Ysp_2)-\frac{1}{2}h_s^{4/N_s}w_{mn}\left[a_r
h_r^{-1}\lap_{\Xsp}h_r-a_rq^{\rho\sigma}\pd_{\rho}\ln h_r
\left\{\left(1-\frac{4}{N_r}\right)
\pd_{\sigma}\ln h_r-\frac{4}{N_s}
\pd_{\sigma}\ln h_s\right\}\right.\nn\\
&&\left. ~~+b_sh_s^{-1}\lap_{\Xsp}h_s
-b_sq^{\rho\sigma}\pd_{\rho}\ln h_s\left\{\left(1-\frac{4}{N_s}\right)
\pd_{\sigma}\ln h_s-\frac{4}{N_r}\pd_{\sigma}\ln h_r\right\}\right]\nn\\
&&~~-\frac{1}{2}w_{mn}h_r^{-4/N_r}\left\{a_rh_r^{-1}
\left(h_s^{4/N_s}\lap_{\Ysp_1}h_r+\lap_{\Zsp}h_r\right)
+b_sh_s^{-1}\left(h_r^{4/N_r}\lap_{\Ysp_2}h_s+\lap_{\Zsp}h_s\right)
\right\}\nn\\
&&~~-\frac{2}{D-2}\left[\Lambda_rh_r^{-2+a_rp_r}
h_s^{a_s-2(\epsilon_r\epsilon_sc_rc_s-2)/N_s}
+\Lambda_sh_r^{a_r-2\epsilon_r\epsilon_sc_rc_s/N_r}
h_s^{-2+a_sp_s+4/N_s}\right]w_{mn}=0,
\label{pl:cEinstein-mn:Eq}\\
&&\hspace{-0.5cm}
R_{ab}(\Zsp)-\frac{1}{2}h_r^{4/N_r}h_s^{4/N_s}u_{ab}\left[b_r
h_r^{-1}\lap_{\Xsp}h_r-b_rq^{\rho\sigma}\pd_{\rho}\ln h_r
\left\{\left(1-\frac{4}{N_r}\right)
\pd_{\sigma}\ln h_r-\frac{4}{N_s}\pd_{\sigma}\ln h_s\right\}\right.\nn\\
&&\left. +b_sh_s^{-1}\lap_{\Xsp}h_s
-b_sq^{\rho\sigma}\pd_{\rho}\ln h_s\left\{\left(1-\frac{4}{N_s}\right)
\pd_{\sigma}\ln h_s-\frac{4}{N_r}\pd_{\sigma}\ln h_r\right\}\right]\nn\\
&&-\frac{1}{2}u_{ab}\left[b_rh_r^{-1}
\left(h_s^{4/N_s}\lap_{\Ysp_1}h_r+\lap_{\Zsp}h_r\right)
+b_sh_s^{-1}\left(h_r^{4/N_r}\lap_{\Ysp_2}h_s+\lap_{\Zsp}h_s\right)
\right]\nn\\
&&\hspace{-0.2cm}
-\frac{2u_{ab}}{D-2}\left[\Lambda_rh_r^{-2+a_rp_r+4/N_r}
h_s^{a_s-2(\epsilon_r\epsilon_sc_rc_s-2)/N_s}
+\Lambda_sh_r^{a_r-2(\epsilon_r\epsilon_sc_rc_s-2)/N_r}
h_s^{-2+a_sp_s+4/N_s}\right]=0,
\label{pl:cEinstein-ab:Eq}
}
where
$D_{\mu}$ is the covariant derivative constructed from
the metric $q_{\mu\nu}$, $\triangle_{\Xsp}$,
$\triangle_{\Ysp_1}$,
$\triangle_{\Ysp_2}$,
and $\triangle_{\Zsp}$
are
the Laplace operators on
$\Xsp$, $\Ysp_1$, $\Ysp_2$, and $\Zsp$,
respectively, and
$R_{\mu\nu}(\Xsp)$, $R_{ij}(\Ysp_1)$, $R_{mn}(\Ysp_2)$,
and $R_{ab}(\Zsp)$ are the Ricci tensors
with respect to the metrics $q_{\mu\nu}(\Xsp)$, $\gamma_{ij}(\Ysp_1)$,
$w_{mn}(\Ysp_2)$\,, and $u_{ab}(\Zsp)$, respectively.
From Eqs.
(\ref{pl:cEinstein-mm:Eq}) and (\ref{pl:cEinstein-ma:Eq})\,,
the warp factors $h_r$ and $h_s$ can be expressed as
\Eqrsubl{pl:warp:Eq}{
&&h_r= h_0(x)+h_1(y, z),~~~~h_s=h_s(v, z)\,,~~~~~~{\rm For}~~
\pd_{\mu}h_s=0\,,
\label{pl:warp1:Eq}\\
&&h_r= h_r(y, z),~~~~h_s= k_0(x)+k_1(v, z)\,,~~~~~~{\rm For}~~
\pd_{\mu}h_r=0.
\label{pl:warp2:Eq}
}
If we require that the background satisfies
\Eqr{
\pd_{\mu}h_s=0\,,~~~p=p_r=0\,,~~~\Lambda_s=0\,,
~~~\chi=0,
\label{pl:condition:Eq}
}
the Einstein Equations (\ref{pl:cEinstein:Eq}) reduce to
\Eqrsubl{pl:c2Einstein:Eq}{
&&\hspace{-0.3cm}
-\frac{2}{N_r}\left[2h_r^{-1}\frac{d^2h_0}{dt^2}
-\left(1-\frac{4}{N_r}\right)\left(\pd_t\ln h_r\right)^2\right]
+\frac{2}{D-2}\Lambda_rh_r^{-2}\nn\\
&&+\frac{1}{2}h_r^{-4/N_r}h_s^{-4/N_s}
\left[a_rh_r^{-1}\left(h_s^{4/N_s}\lap_{\Ysp_1}h_1+\lap_{\Zsp}h_1\right)
+a_sh_s^{-1}\lap_{\Zsp}h_s\right]\nn\\
&&-\frac{1}{2}a_r\left[
h_r^{-1}\frac{d^2h_0}{dt^2}-\left(1-\frac{4}{N_r}\right)
\left(\pd_t\ln h_r\right)^2\right]
=0,
\label{pl:c2Einstein-mu:Eq}\\
&&\hspace{-0.3cm}R_{ij}(\Ysp_1)+\frac{1}{2}b_rh_r^{4/N_r}\gamma_{ij}\left[
h_r^{-1}\frac{d^2h_0}{dt^2}-\left(1-\frac{4}{N_r}\right)
\left(\pd_t\ln h_r\right)^2\right]
-\frac{2}{D-2}\Lambda_rh_r^{-2+\frac{4}{N_r}}\gamma_{ij}\nn\\
&&-\frac{1}{2}\gamma_{ij}h_s^{-4/N_s}\left[b_rh_r^{-1}
\left(h_s^{4/N_s}\lap_{\Ysp_1}h_1+\lap_{\Zsp}h_1\right)
+a_sh_s^{-1}\lap_{\Zsp}h_s\right]=0,
\label{pl:c2Einstein-ij:Eq}\\
&&\hspace{-0.3cm}R_{ab}(\Zsp)
+\frac{1}{2}b_rh_r^{4/N_r}h_s^{4/N_s}u_{ab}\left[
h_r^{-1}\frac{d^2h_0}{dt^2}-\left(1-\frac{4}{N_r}\right)
\left(\pd_t\ln h_r\right)^2\right]
-\frac{2\Lambda_r}{D-2}h_r^{-2+\frac{4}{N_r}}h_s^{\frac{4}{N_s}}u_{ab}\nn\\
&&-\frac{1}{2}u_{ab}\left[b_rh_r^{-1}
\left(h_s^{4/N_s}\lap_{\Ysp_1}h_1+\lap_{\Zsp}h_1\right)
+b_sh_s^{-1}\lap_{\Zsp}h_s\right]=0.
\label{pl:c2Einstein-ab:Eq}
}
Note that Eq. (\ref{pl:cEinstein-mn:Eq}) becomes trivial for $p=p_r=0$.
By combining
the above equations and setting $p=p_r=0$,
the Einstein equations for $N_r\ne 4$ lead to
\Eqrsubl{pl:solution1:Eq}{
&&R_{ij}(\Ysp_1)=0,~~~~
R_{ab}(\Zsp)=0,
\label{pl:Ricci:Eq}\\
&&h_r=h_0(t)+h_1(y, z),~~~~h_s=h_s(z)\,,
\label{pl:h:Eq}\\
&& \left(\frac{dh_0}{dt}\right)^2
+N_r\left(1-\frac{4}{N_r}\right)^{-1}\Lambda_r=0
\,,~~~~
h_s^{4/N_s}\lap_{\Ysp_1}h_1+\triangle_{\Zsp}h_1=0,
\label{pl:warp1-1:Eq}\\
&&\triangle_{\Zsp}h_s=0\,.
\label{pl:warp1-2:Eq}
}
Finally, we check the scalar field equation for the case of $p=p_r=0$.
Substituting Eqs.~(\ref{pl:ansatz:Eq}), (\ref{pl:warp:Eq}), and
(\ref{pl:condition:Eq}) and
the intersection rule $\chi=0$ into Eq.~(\ref{pl:scalar:Eq}),
we have
\Eqr{
&&\frac{\epsilon_rc_r}{N_r}h_r^{-b_r4/N_r}h_s^{-b_s4/N_s}\left[
-h_r^{-1}\frac{d^2h_0}{dt^2}+\left(1-\frac{4}{N_r}\right)
\left(\pd_t\ln h_r\right)^2
+N_r\Lambda_rh_r^{-2}\right]\nn\\
&&~~~~~~+\frac{\epsilon_rc_r}{N_r}h_r^{-1}\left(h_s^{4/N_s}\lap_{\Ysp_1}h_1
+\lap_{\Zsp}h_1\right)
+\frac{\epsilon_sc_s}{N_s}h_s^{-1}\lap_{\Zsp}h_s=0.
\label{pl:scalar-e:Eq}
}
Hence the scalar field equation (\ref{pl:scalar-e:Eq}) reads
\Eqrsubl{pl:scalar solution:Eq}{
&&\frac{d^2h_0}{dt^2}=0\,,~~~
\left(\frac{dh_0}{dt}\right)^2
+N_r\left(1-\frac{4}{N_r}\right)^{-1}\Lambda_r=0\,,
~~~h_s^{4/N_s}\lap_{\Ysp_1}h_1+\lap_{\Zsp}h_1=0,
\label{pl:scalar solution1:Eq}\\
&&\lap_{\Zsp}h_s=0.
\label{pl:scalar solution2:Eq}
}
These are consistent with the Einstein equations (\ref{pl:solution1:Eq})\,.
The function $h_r$ can depend on the
coordinate $t$ only if $N_r\ne 4$\,. For $N_r=4$,
the scalar field equation leads to $\Lambda_r=0$\,.
We can find the solution in which the
$p_s$-brane part depends on $x^{\mu}$.
For $p=p_s=0$, $\Lambda_r=0$\,, and
$\pd_th_r=0$, we have
\Eqrsubl{pl:solution3:Eq}{
&&
R_{mn}(\Ysp_2)=0,~~~~R_{ab}(\Zsp)=0,
\label{pl:Ricci3:Eq}\\
&&h_r=h_r(z),~~~~h_s=k_0(t)+k_1(v, z)\,,
\label{pl:h3:Eq}\\
&&\frac{d^2k_0}{dt^2}=0\,,~~~
\left(\frac{dk_0}{dt}\right)^2
+N_s\left(1-\frac{4}{N_s}\right)^{-1}\Lambda_s=0\,,
~~~h_r^{4/N_r}\lap_{\Ysp_2}k_1+\triangle_{\Zsp}k_1=0,
\label{pl:warp3-1:Eq}\\
&&\triangle_{\Zsp}h_r=0.
\label{pl:warp3-2:Eq}
}
It is clear that there is a solution for $k_0(t)$
such as $\partial_th_s\ne 0$ unless $N_s=4$. For $N_r=4$,
the field equations lead to $\Lambda_r=0$\,.
If $F_{\left(p_r+2\right)}=0$ and $F_{\left(p_s+2\right)}=0$,
the warp factors $h_1$ and $k_1$ are trivial functions.
Then the $D$-dimensional spacetime is no longer warped
\cite{Binetruy:2007tu}.
Moreover, Eqs.~(\ref{pl:solution1:Eq}) and (\ref{pl:solution3:Eq})
imply the two cases. First, $p_r$-, $p_s$-branes are delocalized.
These are localized only along the overall transverse directions.
Second, the 0-brane is completely localized on the $p_s$- (or $p_r$-) brane
which is localized only along the overall transverse directions, which is
a partially localized $p_r-0$ (or $0-p_s$) brane system.
As an example, we set
\Eq{
p=p_r=0,~~~~
\gamma_{ij}=\delta_{ij}\,,~~~
u_{ab}=\delta_{ab}\,,~~~
h_s=h_s(z)\,,
\label{pl:flat metric:Eq}
}
where $\delta_{ij}$ and $\delta_{ab}$ are the $p_s$-
and $(D-p_s-1)$-dimensional Euclidean metrics,
respectively. Equation (\ref{pl:warp1-1:Eq}) gives
\Eqr{
h_0(t)=c_0t+c_1\,,~~~~
c_0=\pm\sqrt{N_r\left(\frac{4}{N_r}-1\right)^{-1}\Lambda_r}\,,
\label{pl:h0:Eq}
}
where $c_0$ and $c_1$ are constants.
Hence, solutions exist
for $N_r<4$ if $\Lambda_r>0$
and vice versa.
If the functions $h_1$ and $h_s$ satisfy
the coupled partial differential equations
\Eq{
h_s^{4/N_s}\lap_{\Ysp_1}h_1+\lap_{\Zsp}h_1=0,~~~~\lap_{\Zsp}h_s=0\,,
\label{pl:hrhs:Eq}
}
the harmonic function $h_s$ that satisfies the equation in
\eqref{pl:warp1-2:Eq} takes the form
\Eqr{
h_s(z)=1+\sum_\ell\frac{M_\ell}{|z^a-z^a_\ell|^{D-p_s-3}},
}
where $z_\ell^a$ is the location of the $\ell$ th $p_s$-brane and $M_\ell$
is constant. Since we consider the case in which the $p_s$-branes
coincide at the same location in the overall transverse directions,
the harmonic function $h_s$ can be written by the following form
\cite{Youm:1999zs, Youm:1999ti, Minamitsuji:2011jt}:
\Eq{
h_s(z)=\frac{M}{|z^a- z^a_0|^{D-p_s-3}},
}
where $M$ is constant and the stack of $p_s$-branes is located at
the same points $z^a_0$ along the $z$ directions.
We can find solutions for the harmonic function
$h_1$ in the case where each of the $p_s$-branes does not
coincide at the same location in the overall transverse directions.
If we set $D-p_s\neq 3$ and $2-4N_s^{-1}\left(D-p_s-3\right)\neq 0$
for the overall transverse space, Eq.~\eqref{pl:hrhs:Eq}
can be solved as
\Eqr{
h_1(y, z)=1+\sum_{\ell}\frac{M_\ell}{\left[|y^i-y^i_\ell|^2
+\frac{4M^{4/N_s}}{\{2-4N_s^{-1}\left(D-p_s-3\right)\}^2}|z^a-z^a_0|
^{2-4N_s^{-1}\left(D-p_s-3\right)}\right]^{\zeta_r}}\,,
\label{pl:h_r solution2:Eq}
}
where $M_\ell$ is constant and $\zeta_r$ is given by
\Eq{
\zeta_r=\frac{1}{2}\left[p_s-1+
\frac{(2-4N_s^{-1})\left(D-p_s-3\right)+2}
{2-4N_s^{-1}\left(D-p_s-3\right)}\right]\,.
}
Hence, the functions $h_r$ and $h_s$ can be expressed as
\Eqrsubl{pl:solutions1:Eq}{
\hspace{-1cm}h_r(t, y, z)&=&c_0\,t+c_1\nn\\
&&+\sum_{\ell}\frac{M_\ell}{\left[|y^i-y^i_\ell|^2
+\frac{4M^{4/N_s}}{\{2-4N_s^{-1}\left(D-p_s-3\right)\}^2}|z^a-z^a_0|
^{2-4N_s^{-1}\left(D-p_s-3\right)}\right]^{\zeta_r}},
\label{pl:solution-r:Eq}\\
\hspace{-1cm}h_s(z)&=&\frac{M}{|z^a-z^a_0|^{D-p_s-3}},
\label{pl:solution-s:Eq}
}
where $c_0$, $c_1$, $M_\ell$\,, and $M$ are constant parameters
and $y^i_\ell$ and $z^a_0$ are
constants representing the positions of the branes.
The curvature singularities appear at $h_r=0$
in the $D$-dimensional metric (\ref{pl:metric:Eq}).
Moreover, there is also a singularity at $z^a=z^a_0$
unless the scalar field is trivial.
Upon setting $D-p_s=3$ and $N_s=4$,
the solutions of Eq.~(\ref{pl:hrhs:Eq}) are
given by
\Eqrsubl{pl:solutions1-2:Eq}{
h_r(t, y, z)&=&c_0t+c_1
+\sum_{\ell}\frac{M_\ell}{\left[|y^i-y^i_\ell|^2
+M|z^a-z^a_0|^2\right]^
{\frac{1}{2}(p_s+1)}},
\label{pl:solution-r-2:Eq}\\
h_s(z)&=&M\ln|z^a-z^a_0|\,.
\label{pl:solution-s-2:Eq}
}
In the case of $D-p_s=5$ and $N_s=4$, the functions $h_r$ and $h_s$
can be written by
\Eqrsubl{pl:solutions1-3:Eq}{
h_r(t, y, z)&=&c_0t+c_1
+\sum_{\ell}M_\ell\left[|y^i-y^i_\ell|^2
-p_sM\ln |z^a-z^a_0|\right]\,,
\label{pl:solution-r-3:Eq}\\
h_s(z)&=&\frac{M}{|z^a-z^a_0|^2}\,.
\label{pl:solution-s-3:Eq}
}
The solutions \eqref{pl:solutions1-2:Eq} and
\eqref{pl:solutions1-3:Eq} have
a singular hypersurface at infinity
as well as at $h_r=0$\,, because the $D$-dimensional metric depends on
the logarithmic function of the transverse coordinates.
These solutions also give a singularity at $z^a=z^a_0$
if the dilaton is nontrivial.
It is possible to find the solution for
$\pd_th_r=0$ and $\pd_th_s\ne 0$ if the roles of $\Ysp_1$ and
$\Ysp_2$ are exchanged. The solution of the field equations for
$D-p_r\neq3$ and $D-p_r\neq5$ can be written by
\Eqrsubl{pl:solutions2:Eq}{
\hspace{-1cm}h_s(t, v, z)&=&c_0t+c_1\nn\\
&&+\sum_{\ell}\frac{M_\ell}{\left[|v^m-v^m_\ell|^2
+\frac{4M^{4/N_r}}{\{2-4N_r^{-1}\left(D-p_r-3\right)\}^2}|z^a-z^a_0|
^{2-4N_r^{-1}\left(D-p_r-3\right)}\right]^{\zeta_s}},
\label{pl:solution2-s:Eq}\\
\hspace{-1cm}h_r(z)&=&\frac{M}{|z^a-z^a_0|^{D-p_r-3}}\,,
\label{pl:solution2-r:Eq}
}
where $\zeta_s$ is given by
\Eq{
\zeta_s=\frac{1}{2}\left[p_r-1+
\frac{(2-4N_r^{-1})\left(D-p_r-3\right)+2}
{2-4N_r^{-1}\left(D-p_r-3\right)}\right]\,.
}
If we set $D-p_r=3$, $D-p_r=5$\,, and $N_r=4$, the harmonic functions
$h_r$ and $h_s$ have
logarithmic spatial dependence like \eqref{pl:solutions1-2:Eq} and
\eqref{pl:solutions1-3:Eq}.
Assuming $\Lambda_r>0$ and
introducing a new time coordinate $\tau$ by
\Eq{
\frac{\tau}{\tau_0}=
\left(c_0t+c_1\right)^{\frac{(N_r-2)(D-2)+2}{N_r(D-2)}}\,,~~~~~
\tau_0=\frac{N_r(D-2)}{c_0\left[(N_r-2)(D-2)+2\right]},
}
we find the $D$-dimensional metric (\ref{pl:metric:Eq}) as
\Eqr{
ds^2&=&
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{N_r(D-2)}{(N_r-2)(D-2)+2}}h_1
\right]^{-\frac{4(D-3)}{N_r(D-2)}}h_s^{a_s}
\left[-d\tau^2
\right.\nn\\
&&\left.\hspace{-1cm}
+\left\{1+\left(\frac{\tau}{\tau_0}\right)^
{-\frac{N_r(D-2)}{(N_r-2)(D-2)+2}}h_1\right\}^{\frac{4}{N_r}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{4}{(N_r-2)(D-2)+2}}\left(
\gamma_{ij}dy^idy^j+h_s^{\frac{4}{N_s}}u_{ab}dz^adz^b\right)\right].\nn\\
\label{pl:s-metric:Eq}
}
Since $h_s$ does not approach constant in any region,
the whole spacetime cannot be homogeneous and isotropic.
But on each $z^a={\rm const}$ slice
the spacetime becomes
a homogeneous and isotropic universe.
In the limit $\tau\rightarrow\infty$,
the function $h_1$ can be negligible in the warp factor.
This is guaranteed by a scalar field with
the exponential potential.
The accelerating universe is obtained
on each $z^a={\rm const}$
slice if $N_r<2$,
which corresponds to the case of a positive cosmological constant.
For $\frac{2(D-3)}{D-2}<N_r<2$,
the solution provides a power-law inflationary universe,
and for $N_r>\frac{2(D-3)}{D-2}$,
the scale factor diverges at
$\tau=\tau_{\infty}>0$,
taking the involution $\tau\to \tau_\infty-\tau$.
Finally, for $N_r=\frac{2(D-3)}{D-2}$,
we obtain a de Sitter universe
which will be discussed in the next subsection.
\subsection{de Sitter universe}
Next, we consider the solution with a dilaton which is
the case of $c_I=0~(I=r~{\rm or}~s)$.
In terms of $c_I=0$, Eq.~(\ref{pl:c:Eq}) gives
\Eq{
N_I=\frac{2(D-p_I-3)(p_I+1)}{(D-2)}.
\label{ds:N:Eq}
}
If we assume
\Eqr{
c_r=0\,,~~~c_s\ne 0\,,~~~p=p_r=0\,,~~~
N_r=\frac{2(D-3)}{(D-2)}\,,~~~
\alpha_r=-\frac{N_sa_s}{2\epsilon_sc_s}\,,~~~\Lambda_s=0\,,
}
the field equations reduce to
\Eqrsubl{ds:equations:Eq}{
&&
R_{ij}(\Ysp_1)=0,~~~~~R_{ab}(\Zsp)=0,\\
&&
h_r(t, y, z)=h_0(t)+h_1(y, z),~~~~~
\left(\frac{dh_0}{dt}\right)^2
-\frac{2(D-3)^2}{(D-2)(D-1)}\Lambda_r=0,
\label{ds:warp1:Eq}\\
&&
h_s^{4/N_s}\lap_{\Ysp_1}h_1+\triangle_{\Zsp}h_1=0,~~~~~\triangle_{\Zsp}h_s=0.
\label{ds:warp2:Eq}
}
Then Eq.~(\ref{ds:warp1:Eq}) gives
\Eq{
h_0=c_0t+c_1\,,
}
where $c_1$ is an integration constant and $c_0$ is given by
\Eq{
c_0=\pm(D-3)\sqrt{\frac{2}{(D-2)(D-1)}\Lambda_r}\,.
\label{ds:c0:Eq}
}
Thus there is no solution for $\Lambda_r<0$.
If the metric $u_{ab}({\rm Z})$ is assumed to be
Eq.~(\ref{pl:flat metric:Eq}),
the function ${h}_1$ is given by Eq.~(\ref{pl:h_r solution2:Eq}).
Now we introduce a new time coordinate $\tau$ by
\Eq{
c_0\tau=\ln t,
}
where we have set $c_0>0$ for simplicity.
Then the $D$-dimensional metric
(\ref{pl:metric:Eq}) can be expressed as
\Eqr{
ds^2&=&h_s^{a_s}\left[-\left(1+c_0^{-1}
\e^{-c_0\tau}h_1\right)^{-2}d\tau^2
+\left(1+c_0^{-1}\e^{-c_0\tau}h_1\right)^{2/(D-3)}
\left(c_0\e^{c_0\tau}\right)^{2/(D-3)}
\right.\nn\\
&&\left.\times\left\{\gamma_{ij}(\Ysp_1)dy^idy^j
+h_s^{4/N_s}u_{ab}(\Zsp)dz^adz^b\right\}\right].
\label{ds:metric:Eq}
}
The function $h_s$ does not become constant in any region. Then,
the $D$-dimensional spacetime cannot be de Sitter spacetime.
However, the spacetime gives a homogeneous and isotropic
universe on each $y^i={ \rm const}$, $z^a={ \rm const}$ slice.
If we set $h_s=$const and $h_1=h_1(z)$,
Eq.~(\ref{ds:metric:Eq}) becomes the solution which has been discussed by
Refs.~\cite{Maki:1992tq, Maki:1994wc} (see also \cite{Ivashchuk:1996zv}).
Furthermore, for $D=4$ and by setting $h_s=1$,
the solution is the Kastor-Traschen one \cite{Kastor:1992nn}.
\section{The intersection involving $n$ brane backgrounds}
\label{sec:cn}
The construction that we have analyzed in Sec.~\ref{sec:pl}
is a special case of a more general construction of intersecting branes
with a cosmological constant. In effect, we have been studying the special
case of intersections involving a two-brane. The time-dependent brane
with a cosmological constant property is a 0-brane, represented by a
2-form. To describe more general intersections on a time-dependent
background, one simply incorporates additional branes
in a dynamical background. Without loss of the time dependence, it is
possible to also add $n$ delocalized branes. This also has one
important further refinement. Instead of power-law expansion,
the support of a 0-brane might be accelerated expansion,
where $D$-dimensional geometry is an asymptotically de Sitter spacetime.
The $n$ intersection allows the time dependence of only 0-branes but
not of the $p$-branes ($p\ne 0$).
The reason for this is that the time-dependent brane we have obtained
can be performed in the case of $\chi=0$, where
$\chi$ is defined by \eqref{pl:chi:Eq}.
So the coefficient of the time dependence is simply proportional
to the cosmological constant that we have explored in Sec~\ref{sec:pl}:
the Einstein equations give $p=0$.
In this section, we discuss the intersection of the
delocalized $n$ branes
in the higher-dimensional gravity theory with the cosmological constants.
The general action describing the intersection involving the
$n$ brane system is given by
\Eqr{
S&=&
\frac{1}{2\kappa^2}\int \left[\left(R-2\sum_I\e^{\alpha_I\phi}
\Lambda_I\right)\ast{\bf 1}_D
-\frac{1}{2}\ast d\phi \wedge d\phi \right.\nn\\
&&\left.
-\frac{1}{2}\sum_I \frac{\e^{\epsilon_Ic_I\phi}}{(p_I+2)!}\ast F_{(p_I+2)}
\wedge F_{(p_I+2)}\right],
\label{cn:action:Eq}
}
where $\kappa^2$ denotes the $D$-dimensional gravitational constant,
$R$ is the $D$-dimensional Ricci scalar constructed from
the $D$-dimensional metric $g_{MN}$, $\phi$ is a scalar field,
$F_{(p_I+2)}$ is the antisymmetric tensor fields of rank $(p_I+2)$,
${\ast}$ is the Hodge dual operator in the $D$-dimensional spacetime, and
$c_I$ and $\epsilon_I$ are constants defined by
\Eqrsubl{cn:p:Eq}{
c_I^2&=&N_I-\frac{2(p_I+1)(D-p_I-3)}{D-2},
\label{cn:c:Eq}\\
\epsilon_I&=&\left\{
\begin{array}{cc}
+ &~{\rm for~the~electric~brane}\,,
\\
- &~~~{\rm for~the~magnetic~brane}\,.
\end{array} \right.
\label{cn:epsilon:Eq}
}
Here $I$ denotes the type of the corresponding branes.
The $D$-dimensional action (\ref{cn:action:Eq}) give the field equations
\Eqrsubl{cn:field eq:Eq}{
&&\hspace{-1cm}R_{MN}=\frac{2}{D-2}\sum_I\e^{\alpha_I\phi}\Lambda_I\,g_{MN}
+\frac{1}{2}\pd_M\phi \pd_N \phi\nn\\
&&+\frac{1}{2}\sum_I\frac{1}{(p_I+2)!}\e^{\epsilon_Ic_I\phi}
\left[(p_I+2)F_{MA_2\cdots A_{p_I+2}} {F_N}^{A_2\cdots A_{p_I+2}}
-\frac{p_I+1}{D-2}g_{MN} F^2_{(p_I+2)}\right],
\label{cn:Einstein:Eq}\\
&&\hspace{-1cm}
\lap\phi-2\sum_I\alpha_I\e^{\alpha_I\phi}\Lambda_I
-\frac{1}{2}\sum_I\frac{\epsilon_Ic_I}{(p_I+2)!}
\e^{\epsilon_Ic_I\phi}F^2_{(p_I+2)}=0\,,
\label{cn:scalar:Eq}\\
&&\hspace{-1cm}
d\left[\e^{\epsilon_Ic_I\phi}\ast F_{(p_I+2)}\right]=0\,,
\label{cn:gauge:Eq}
}
where $\lap$ denotes the Laplace operator with respect to the
$D$-dimensional metric $g_{MN}$\,.
We adopt the ansatz that $D$-dimensional metric can be written by
\Eqr{
\hspace{-10mm}
ds^2=-A(t, z)dt^2
+\sum_{\alpha=1}^pB^{(\alpha)}(t, z)(dx^{\alpha})^2
+C(t, z)u_{ab}(\Zsp) dz^a dz^b,
\label{cn:metric:Eq}
}
where $u_{ab}(\Zsp)$ denotes the metric of the
$(D-p-1)$-dimensional $\Zsp$ space which depends only on
the $(D-p-1)$-dimensional coordinates $z^a$.
Concerning the functions $A$, $B^{(\alpha)}$ and $C$,
we assume
\Eq{
A=\prod_I\left[h_I(t,z)\right]^{a_I},~~~
B^{(\alpha)}=\prod_I\left[h_I(t,z)\right]^{\delta^{(\alpha)}_I},~~~
C=\prod_I\left[h_I(t,z)\right]^{b_I}\,,
}
where the constants $a_I$, $b_I$ and $\delta^{(\alpha)}_I$
are given by
\Eq{
a_I=-\frac{4(D-p_I-3)}{N_I(D-2)},~~~~b_I=\frac{4(p_I+1)}{N_I(D-2)},~~~~
\delta^{(\alpha)}_I=\left\{
\begin{array}{cc}
a_I&~{\rm for}~~\alpha\in I\,,\\ b_I &~{\rm for}~~\alpha
\in \hspace{-.8em}/ I\,.
\end{array} \right.
\label{cn:parameter:Eq}
}
The function $h_I(t,z)$ is a straightforward generalization
of the static solution associated with the brane $I$
in an intersecting brane system~\cite{Argurio:1997gt, Argurio:1998cp}
to the dynamical one.
We further require that the dilaton $\phi$ and the form fields
$F_{(p+2)}$ satisfy the following conditions
\Eq{
\e^{\phi}=\prod_Ih_I^{2\epsilon_Ic_I/N_I},~~~
F_{(p_I+2)}=\frac{2}{\sqrt{N_I}}d(h_I^{-1})\wedge\Omega(\Xsp_I)\,,
\label{cn:fields:Eq}
}
where ${\Xsp}_I$ is the space associated with the brane $I$,
and the volume $(p_I+1)$-form $\Omega(\Xsp_I)$ is written by
\Eq{
\Omega(\Xsp_I)=dt\wedge dx^{p_1}\wedge \cdots \wedge
dx^{p_I}\,.
}
\subsection{Power-law expanding universe}
Firstly, we consider the Einstein equations (\ref{cn:Einstein:Eq}) with
$c_I\neq 0~(I=0,\cdots,~n-1)$.
We assume that the parameters $\alpha_I~(I=0,\cdots,~n-1)$ are given by
\Eqr{
\alpha_I=-\epsilon_I c_I.
\label{cn:alpha:Eq}
}
We impose the condition with respect to the components of
$D$-dimensional metric~\cite{Argurio:1998cp}
\Eq{
A^{(D-p-3)}
\prod_{\alpha=1}^p\,B^{(\alpha)} \,C=1 \,,~~~~~~~
A^{-1}\prod_{\alpha \in I}\left(B^{(\alpha)}\right)^{-1}
\e^{\epsilon_Ic_I\phi}=h^2_I\,.
\label{cn:extremal:Eq}
}
The Einstein equations \eqref{cn:Einstein:Eq} become
\Eqrsubl{cn:cEinstein:Eq}{
&&\hspace{-0.5cm}
\sum_{I,I'}\left(\frac{2}{N_I}\delta_{II'}-M_{II'}
\right)\pd_t\ln h_I\pd_t\ln h_{I'}
+\frac{2}{D-2}\sum_I\Lambda_Ih_I^{-2+a_Ip_I}\prod_{I'\ne I}
h_{I'}^{a_{I'}-\frac{2\varepsilon_I\varepsilon_{I'}c_Ic_{I'}}{N_{I'}}}\nn\\
&&~~~~+\frac{1}{2}\sum_{I}b_I\left[\left(1-\frac{4}{N_I}\right)\pd_t\ln h_I
-\sum_{I'\ne I}\frac{4}{N_{I'}}\pd_t\ln h_{I'}\right]\pd_t\ln h_I
\nn\\
&&~~~~-\frac{1}{2}\sum_{I}\left(\frac{4}{N_I}+b_I\right)h_I^{-1}\pd_t^2h_I
+\frac{1}{2}\prod_{I'}h_{I'}^{-4/N_{I'}}\sum_{I}a_{I}
h_{I}^{-1}\lap_{\Zsp}h_{I}=0,
\label{cn:cEinstein-tt:Eq}\\
&&\hspace{-0.5cm}\sum_I\frac{2}{N_{I}}h^{-1}_I\pd_t\pd_ah_I+\sum_{I,I'}
\left(M_{II'}-\frac{2}{N_{I'}}\delta_{II'}
\right)\pd_t\ln h_I\pd_a\ln h_{I'}=0,
\label{cn:cEinstein-ti:Eq}\\
&&\hspace{-0.5cm}\prod_{J'}h_{J'}^{-a_{J'}}
\sum_{\gamma}\prod_{J}h_{J}^{\delta_J^{(\gamma)}}
\sum_{I} \delta^{(\gamma)}_I\left[h_I^{-1}\pd_t^2h_I
-\left\{\left(1-\frac{4}{N_I}\right)\pd_t\ln h_I
\right.\right.\nn\\
&&\left.\left.~~~~-\sum_{I'\ne I}\frac{4}{N_{I'}}
\pd_t\ln h_{I'}\right\}\pd_t\ln h_I\right]
-\prod_{J'}h_{J'}^{-b_{J'}}
\sum_{\gamma}\prod_{J}h_{J}^{\delta_J^{(\gamma)}}\sum_I
\delta^{(\gamma)}_Ih_I^{-1}\lap_{\Zsp}h_I\nn\\
&&~~~~-\frac{4}{D-2}\sum_I\Lambda_I
h_I^{-2+\delta_I^{(\gamma)}p_I}\prod_{I'\ne I}
h_{I'}^{\delta_{I'}^{(\gamma)}
-\frac{2\varepsilon_I\varepsilon_{I'}c_Ic_{I'}}{N_{I'}}}=0,
\label{cn:cEinstein-ab:Eq}\\
&&\hspace{-0.5cm}
R_{ab}(\Zsp)+\frac{1}{2}u_{ab}\prod_Jh_J^{4/N_J}\sum_{I}
b_I\left[ h_I^{-1}\pd_t^2h_I
-\left\{\left(1-\frac{4}{N_I}\right)\pd_t\ln h_I
-\sum_{I'\ne I}\frac{4}{N_{I'}}\pd_t\ln h_{I'}\right\}\pd_t\ln h_I\right]\nn\\
&&~~~~-\frac{1}{2}u_{ab}\sum_Ib_I h_I^{-1}\lap_{\Zsp}h_I
-\sum_{I,I'}\frac{2}{N_I}\left(M_{II'}-\frac{2}{N_{I'}}\delta_{II'}
\right)\pd_a\ln h_I\pd_b\ln h_{I'}\nn\\
&&~~~~-\frac{2}{D-2}\sum_I\Lambda_Ih_I^{-2+a_Ip_I+\frac{4}{N_I}}\prod_{I'\ne I}
h_{I'}^{a_{I'}-\frac{2(\varepsilon_I\varepsilon_{I'}c_Ic_{I'}-2)}{N_{I'}}}
u_{ab}=0,
\label{cn:cEinstein-ab2:Eq}
}
where $R_{ab}(\Zsp)$ is the Ricci tensor with respect to
the metric $u_{ab}(\Zsp)$, and $M_{II'}$ is defined by
\Eqr{
&&M_{II'}\equiv \frac{1}{4}\left[a_Ia_{I'}
+\sum_{\alpha}\delta^{(\alpha)}_I\delta^{(\alpha)}_{I'}
+(D-p-3)b_Ib_{I'}\right]
+\frac{2}{N_IN_{I'}}\epsilon_I\epsilon_{I'}c_Ic_{I'}
\,.
\label{cn:M:Eq}
}
The Eq.~(\ref{cn:cEinstein-ti:Eq}) can be rewritten as
\Eq{
\sum_{I,I'}\left[M_{II'}+\frac{2}{N_{I}} \delta_{II'}
\frac{\pd_t\pd_a \ln h_I}{\pd_t \ln h_I \pd_a \ln h_I} \right]
\pd_t \ln h_I \pd_a \ln h_{I'}=0.
\label{cn:cEinstein-ti2:Eq}
}
One can find that the equation (\ref{cn:cEinstein-ti2:Eq})
is equivalent to satisfying that
\Eq{
\frac{\pd_t\pd_a \ln h_I}{\pd_t \ln h_I \pd_a \ln h_I}=k_I
\,.
\label{cn:cEinstein-ti3:Eq}
}
Then we have
\Eq{
M_{II'}+\frac{2}{N_{I}}k_I \delta_{II'}=0.
\label{cn:cEinstein-ti4:Eq}
}
Eqs.~(\ref{cn:c:Eq}), (\ref{cn:parameter:Eq}) and (\ref{cn:cEinstein-ti2:Eq})
give
\Eqr{
M_{II} &=&\frac{1}{4}\left[(p_I+1)a_I^2 + (p-p_I)b_I^2
+(D-p-3)b_I^2\right] +\frac{2}{N_I^2} c_I^2 \nonumber \\
&=& \frac{2}{N_I}.
\label{cn:M2:Eq}
}
Combining the \eqref{cn:M2:Eq} with (\ref{cn:cEinstein-ti4:Eq})
the constant $k_I$ in~\eqref{cn:cEinstein-ti4:Eq} is $k_I=-1$, which
implies
\Eq{
M_{II'}= \frac{2}{N_{I'}}\delta_{II'}.
\label{cn:M3:Eq}
}
Taking account of these results, the equation (\ref{cn:cEinstein-ti:Eq})
yields
\Eq{
\pd_t \pd_a[h_I(t,z)]=0
\,.
}
Hence we find
\Eq{
h_I(t, z)= K_I(t)+H_I(z)\,.
\label{cn:warp:Eq}
}
For $I\neq I'$, ~\eqref{cn:M3:Eq} provides the intersection rule
on the dimension $\bar{p}$ of the intersection for each pair of branes
$I$ and $I'$ $(\bar{p}\leq p_I, p_{I'})$~\cite{Ohta:1997gw, Youm:1997hw}:
\Eq{
\bar{p}=\frac{(p_I+1)(p_{I'}+1)}{D-2}-1-
\frac{1}{2}\epsilon_Ic_I\epsilon_{I'}c_{I'}.
\label{cn:rule:Eq}
}
Under the assumptions given above, we next reduce the gauge field equations.
In terms of the ansatz~\eqref{cn:fields:Eq}, the Bianchi identity
$dF_{(p_I+2)}=0$ is automatically satisfied;
\Eq{
h_I^{-1}(2\pd_a \ln h_I \pd_b \ln h_I
+ h_I^{-1}\pd_a\pd_b h_I)
dz^a\wedge dz^b\wedge\Omega(\Xsp_I)=0.
}
Utilising \eqref{cn:fields:Eq}, the gauge field equation becomes
\Eqr{
d\left[\pd_aH_I \left(\ast_{\Zsp}dz^a\right)\wedge
\ast_{\Xsp}\Omega(\Xsp_I)\right]=0,
\label{cn:gauge2:Eq}
}
where we used Eqs.~(\ref{cn:extremal:Eq}), (\ref{cn:warp:Eq}), and
$\ast_{\Xsp}$, $\ast_{\Zsp}$ are the Hodge dual
operators on $\Xsp(\equiv \cup_{I}X_I)$ and $\Zsp$, respectively.
Hence, \eqref{cn:gauge2:Eq} gives (\ref{cn:warp:Eq})
and we find
\Eq{
\lap_{\Zsp}H_I=0.
\label{cn:gauge3:Eq}
}
The roles of the Bianchi identity and field equations are
interchanged for magnetic ansatz~\cite{Ohta:1997gw, Youm:1997hw,
Argurio:1998cp}. Then the net result is the same.
In order to complete the system of equations,
we must also consider the scalar field equation.
Substituting the ansatz for fields
\eqref{cn:fields:Eq},
and the metric (\ref{cn:metric:Eq}), (\ref{cn:warp:Eq}),
the equation of motion for the scalar field
\eqref{cn:scalar:Eq} reduces to
\Eqr{
&&-\prod_{I''}h_{I''}^{-a_{I''}}
\sum_{I}\frac{1}{N_I}\epsilon_Ic_I\left[h_I^{-1}\frac{d^2K_I}{dt^2}
-\left\{\left(1-\frac{4}{N_I}\right)\pd_t\ln h_I-
\sum_{I'\ne I}\frac{4}{N_{I'}}\pd_t\ln h_{I'}\right\}\pd_t\ln h_I
\right.\nn\\
&&\left.~~~~~-N_I\Lambda_Ih_I^{-2}\right]
+\prod_{I''}h_{I''}^{-b_{I''}}\sum_I\frac{1}{N_I}h_I^{-1}
\epsilon_Ic_I\lap_{\Zsp} H_I=0.
\label{cn:scalar2:Eq}
}
Furthermore, \eqref{cn:scalar2:Eq} reads
\Eqrsubl{cn:enough:Eq}{
&&\hspace{-1cm}\frac{d^2K_I}{dt^2}=0,
\label{cn:KI:Eq}
\\
&&\hspace{-1cm}\triangle_{\Zsp}H_I=0,
\label{5}\\
&&\hspace{-1cm}\sum_{I}\frac{\epsilon_Ic_I}{N_I}\left[
\left\{\left(1-\frac{4}{N_I}\right)\pd_t\ln h_I-
\sum_{I'\ne I}\frac{4}{N_{I'}}\pd_t\ln h_{I'}\right\}\pd_t\ln h_I
+N_I\Lambda_Ih_I^{-2}\right]=0\,.
\label{cn:scalar3:Eq}
}
From Eq.~(\ref{cn:KI:Eq}), we obtain
\Eq{
K_I=A_I t+B_I,
}
where $A_I$ and $B_I$ are constants.
\subsubsection{The intersection involving the same brane}
Let us first consider the case that all cosmological constants become
nonvanishing. If we set $\Lambda_I\ne 0$, the field equations imply
that all functions are equal:
\Eq{
h_I(t,z)=K(t, z)=K_0(t)+K_1(z),~~~~N_I=N_{I'}\equiv N.
\label{sa:h:Eq}
}
We can find the solutions if the function $h$ and $N$ satisfy
\Eq{
K_0(t)=A\, t+B,~~~~A=\pm\sqrt{N_I\Lambda_I/
\sum_I\left(\frac{4}{N_I}-1\right)}\,,
\label{sa:K:Eq}
}
where $B$ denotes a constant.
Then the remaining
Einstein equations~(\ref{cn:cEinstein:Eq}) are
\Eq{
R_{ab}(\Zsp)=0.
\label{sa:Ricci:Eq}
}
Now we set
\Eq{
\quad u_{ab}=\delta_{ab}\,,
\label{sa:flat:Eq}
}
where $\delta_{ab}$ is the $(D-p-1)$-dimensional Euclidean metric.
In this case, the solution for $h_I$ can be obtained explicitly as
\Eq{
K(t, z)=At+B +\sum_{k}\frac{M_{k}}{|z^a- z^a_{k}|^{D-p-3}},
\label{sa:exact:Eq}
}
where $M_{k}$'s are constant parameters and $z^a_{k}$
represents the positions of the branes in Z space.
If the functions $h_I$ coincide, the locations of the
$p_I$-brane will also coincide
In this case, all branes have the same
total amount of charge at the same position.
Let us consider the intersection rule in the $D$-dimensional
gravity theory. If we set $p_I=\tilde{p}$ for all $p_I$,
the intersection rule (\ref{cn:rule:Eq}) leads to
\Eq{
\bar{p}=\tilde{p}-\frac{N}{2}.
\label{sa:chi2:Eq}
}
Then, we find the intersection involving two $\tilde{p}$-branes:
\Eq{
\tilde{p}\cap \tilde{p}=\tilde{p}-\frac{N}{2}.
\label{sa:int:Eq}
}
Since the number of intersections for $\tilde{p}<\frac{N}{2}$ is negative,
there is no solution in these brane backgrounds.
If we choose $K_0=0 ~(A=B=0$),
the metric describes the known static and extremal multi-black-hole
solution with black hole charges
$M_{k}$~\cite{Argurio:1997gt, Ohta:1997gw, Youm:1997hw, Argurio:1998cp}.
\subsubsection{A dynamical brane in the intersecting brane system}
In the following, we consider the case
that there is only one function $h_I$ which depends on both
$z^a$ and $t$. We denote it with the
subscript $\tilde{I}$,
while other functions of $I'\neq \tilde{I}$
are either dependent on $z^a$ or constant.
If we assume $N_{\tilde{I}}\ne 4$\,, we have
\Eqr{
\pd_th_{I'}=0\,,~~~p_{\tilde{I}}=0\,,~~~\Lambda_{I'}=0\,,~~~
{\rm for}~I'\ne {\tilde{I}}.
\label{cn:condition:Eq}
}
We can find the solutions if the function $h_{\tilde{I}}$ and
$N_{\tilde{I}}$ satisfy
\Eqr{
\hspace{-0.3cm}
h_{\tilde{I}}(t, z)= K_{\tilde{I}}(t)+H_{\tilde{I}}(z),~~~
K_{\tilde{I}}(t)=\pm\left[\left(\frac{4}{N_{\tilde{I}}}-1\right)^{-1}
N_{\tilde{I}}\Lambda_{\tilde{I}}\right]^{\frac{1}{2}}\,t+c_{\tilde{I}}\,,
~~~N_{\tilde{I}}\ne 4,
\label{cn:h:Eq}
}
where $c_{\tilde{I}}$ is constant.
Then the remaining
Einstein equations~\eqref{cn:cEinstein:Eq} are
\Eq{
R_{ab}(\Zsp)=0.
\label{cn:Ricci:Eq}
}
Now we set
\Eq{
\quad u_{ab}=\delta_{ab}\,,
\label{cn:flat metric:Eq}
}
where $\delta_{ab}$ is the $(D-p-1)$-dimensional Euclidean metric.
In this case, the solution for $h_I$ can be written explicitly as
\Eqrsubl{cn:hI:Eq}{
h_{\tilde{I}}(t, z)&=&\pm\left[\left(\frac{4}{N_{\tilde{I}}}-1\right)^{-1}
N_{\tilde{I}}\Lambda_{\tilde{I}}\right]^{\frac{1}{2}}t
+\tilde{c}_{\tilde{I}}
+\sum_{k}\frac{M_{\tilde{I}, k}}{|z^a- z^a_{k}|^{D-p-3}}\,,\\
h_{I'}(z)&=&\tilde{c}_{I'}
+\sum_{l}\frac{M_{I',\,l}}{|z^a-z^a_{l}|^{D-p-3}}\,,
}
where $\tilde c_{\tilde{I}}$, $\tilde{c}_{\tilde{I}}$, $M_{\tilde{I}, k}$,
and $M_{I',\,l}$ are constant parameters and
$z^a_{k}$ and $z^a_{l}$ denote the positions of the branes in Z space.
$N_{\tilde I}<4$ leads to $\Lambda_{\tilde I}>0$
and vice versa.
Since the functions $h_I$ coincide, the locations of the
$p_I$-brane will also coincide
This physically means that all branes have the same
total amount of charge at the same position.
Here we have discussed the solution without compactification of $\Zsp$ space.
If we consider the case that $q$ dimensions of $\Zsp$ space are
smeared, we can find the different power of harmonics, i.e.
${|z^a-z^a_{k}|^{-(D-p-3-q)}}$ ($q\leq D-p-2$).
For $K_{\tilde{I}}=0 ~(A=B=0$),
the solution describes the known static and
extremal multi-black-hole
solution with black hole charges
$M_{\tilde{I}, k}$~\cite{Ohta:1997gw, Youm:1997hw, Argurio:1998cp}.
We can find the solution (\ref{cn:hI:Eq}) for any $N_I\ne 4$.
If we choose ${N}_I=4$,
the solutions have already discussed in
Ref.~\cite{Maeda:2009zi}.
Let us consider the intersection rule in the $D$-dimensional gravity theory.
If we choose $p_{\tilde{I}}=\tilde{p}=0$ for all
$p_{\tilde{I}}\ne p_{I'}$,
the intersection rule (\ref{cn:rule:Eq}) leads to
\Eq{
\frac{p_{I'}+1}{D-2}-1-
\frac{1}{2}\epsilon_{\tilde{I}}c_{\tilde{I}}\epsilon_{I'}c_{I'}=0.
\label{cn:chi2:Eq}
}
Now we discuss the application of
the time-dependent solutions to study
the cosmology. We assume an isotropic and homogeneous three-space
in the Friedmann-Robertson-Walker (FRW) universe
after compactification.
We set the $(D-p-1)$-dimensional
Euclidean space with $u_{ab}(\Zsp)=\delta_{ab}(\Zsp)$ and
consider the case that there is only one function $h_I$
depending on both $z^a$ and $t$, which we denote it with the
subscript $\tilde{I}$, and other functions are either dependent on
$z^a$ or constant.
If we assume $N_{\tilde{I}}\ne 4$,
the $D$-dimensional metric can be expressed as
\begin{eqnarray}
ds^2&=&-\prod_{I\ne \tilde{I}}h_I^{a_I}
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{2}{a_{\tilde{I}}+2}}
H_{\tilde{I}}\right]^{a_{\tilde{I}}}d\tau^2\nn\\
&&
+\sum_{\alpha=1}^p\prod_{I\ne \tilde{I}}h_I^{\delta_I^{(\alpha)}}
\left[1+\left(\frac{\tau}{\tau_0}\right)
^{-\frac{2}{a_{\tilde{I}}+2}}H_{\tilde{I}}\right]^
{\delta_{\tilde{I}}^{(\alpha)}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2\delta_{\tilde{I}}^{(\alpha)}}
{a_{\tilde I}+2}}
\left(dx^{\alpha}\right)^2\nn\\
&&+\prod_{I\ne \tilde{I}}h_I^{b_I}
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{2}{a_{\tilde{I}}+2}}
H_{\tilde{I}}\right]^{b_{\tilde{I}}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2b_{\tilde{I}}}{a_{\tilde{I}}+2}}
\delta_{ab}(\Zsp)dz^adz^b,
\label{cn:metric-J:Eq}
\end{eqnarray}
where the function $H_{\tilde{I}}$ is given by
\Eq{
H_{\tilde{I}}=\sum_{k}\frac{M_{\tilde{I},\,k}}
{|z^a- z^a_{k}|^{D-p-3}}\,,
\label{cn:HI:Eq}
}
and the cosmic time $\tau$ defined by
\Eq{
\frac{\tau}{\tau_0}=\left(At\right)^{(a_{\tilde{I}}+2)/2},~~~~\tau_0=
\frac{2}{\left(a_{\tilde{I}}+2\right)A}.
\label{cn:cosmic:Eq}
}
If we can regard the three-dimensional part of the overall
transverse space Z as our Universe,
the power of the scale factor in the fastest expanding case is
expressed as
\Eq{
\lambda=\frac{b_{\tilde{I}}}{a_{\tilde{I}}+2}=\left[-D+3
+\frac{N_{\tilde{I}}}{2}(D-2)\right]^{-1},
~~~~~~{\rm for}~~D>2,
\label{cn:power:Eq}
}
where we used the $D$-dimensional metric
(\ref{cn:metric-J:Eq}).
Hence, we cannot find the cosmological model which
exhibits an accelerating expansion
of our Universe.
On the other hand, if our three-space is given by a
three-dimensional subspace in relative transverse space,
the power of the scale factor in the fastest expanding case is also
given by \eqref{cn:power:Eq}.
By taking $\tau\to \tau_\infty-\tau$, where $\tau_\infty$ is constant,
we have accelerated expansion for $\tau_\infty>\tau$ and $\lambda<0$.
This is equivalent to
\Eq{
N_{\tilde{I}}>2\,,~~~D>2-\frac{2}{N_{\tilde{I}}-2}\,,~~~~
{\rm or}~~~~
N_{\tilde{I}}<2\,,~~~3<D<2-\frac{2}{N_{\tilde{I}}-2}\,,
\label{cn:condition1:Eq}
}
for $D>3$\,.
However, the scale factor of our Universe diverges at
$\tau=\tau_\infty$.
On the other hand, the power of the scale factor in the fastest
expanding case is automatically positive for
$D=3$ and $N_{\tilde{I}}>0$\,.
Next we discuss the cosmological solution in the lower-dimensional
effective theories.
We compactify $d(\equiv \sum_{\alpha}d_{\alpha}+d_z)$
dimensions to give our Universe,
where $d_{\alpha}$ and $d_z$ denote the compactified dimensions with
respect to the relative and overall transverse space, respectively.
The $D$-dimensional metric (\ref{cn:metric:Eq}) is written by
\Eq{
ds^2=ds^2(\Msp)+ds^2(\Nsp)
\label{cn:metric2:Eq},
}
where $ds^2(\Msp)$ is a $(D-d)$-dimensional metric and
$ds^2(\Nsp)$ is a metric of compactified dimensions.
In order to discuss the dynamics of the $(D-d)$-dimensional universe
in the Einstein frame, we use the conformal transformation
\Eq{
ds^2(\Msp)=h_{\tilde{I}}^{B_{\tilde{I}}}\prod_{I\ne \tilde{I}}
h_I^{C_I}ds^2(\bar{\Msp}),
}
where $B_{\tilde{I}}$ and $C_I$ are expressed, respectively, as
\Eq{
B_{\tilde{I}}=-\frac{\sum_{\alpha}d_{\alpha}
{\delta_{\tilde{I}}}^{(\alpha)}+d_zb_{\tilde{I}}}{D-d-2},~~~~
C_I=-\frac{\sum_{\alpha}d_{\alpha}\delta_I^{(\alpha)}+d_zb_I}{D-d-2}.
}
The $(D-d)$-dimensional metric in the Einstein frame is thus
given by
\begin{eqnarray}
ds^2(\bar{\Msp})&=&h_{\tilde{I}}^{-B_{\tilde{I}}}\prod_{J\ne \tilde{I}}
{h_J}^{-C_J}\left[
-h^{a_{\tilde{I}}}\prod_{I\ne \tilde{I}}h_I^{a_I}
dt^2+\sum_{{\alpha}'}h_{\tilde{I}}^{\delta_{\tilde{I}}^{({\alpha}')}}
\prod_{I\ne \tilde{I}}h_I^{\delta_I^{({\alpha}')}}
\left(dx^{{\alpha}'}\right)^2\right.\nn\\
&&\left.+h_{\tilde{I}}^{b_{\tilde{I}}}\prod_{I\ne \tilde{I}}h_I^{b_I}
\delta_{a'b'}({\Zsp}')dz^{a'}dz^{b'}\right],
\label{cn:metric-E:Eq}
\end{eqnarray}
where $x^{{\alpha}'}$ denotes the coordinate of $(p-d_{\alpha})$-dimensional
relative transverse space and ${\Zsp}'$ is
$(D-p-1-d_z)$-dimensional spaces.
If we set $K_{\tilde{I}}=At$, the $(D-d)$-dimensional metric
(\ref{cn:metric:Eq}) in the Einstein frame can be expressed as
\begin{eqnarray}
ds^2(\bar{\Msp})&=&\prod_{I\ne \tilde{I}}h_I^{-C_I}
\left[-\prod_{I\ne \tilde{I}}h_I^{a_I}
\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{2}{{B'}_{\tilde{I}}+2}}
H_{\tilde{I}}\right\}^{{B'}_{\tilde{I}}}d\tau^2\right.\nn\\
&&\hspace{-0.4cm}+
\sum_{\alpha'}\prod_{I\ne \tilde{I}}h_I^{\delta^{(\alpha')}_I}
\left\{1+\left(\frac{\tau}{\tau_0}\right)
^{-\frac{2}{{B'}_{\tilde{I}}+2}}H_{\tilde{I}}\right\}^
{-B_{\tilde{I}}+\delta_{\tilde{I}}^{(\alpha')}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2\left(-B_{\tilde{I}}
+\delta_{\tilde{I}}^{(\alpha')}\right)}
{{B'}_{\tilde{I}}+2}}
\left(dx^{\alpha'}\right)^2\nn\\
&&\left.\hspace{-0.4cm}+\prod_{I\ne \tilde{I}}h_I^{b_I}
\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{2}{{B'}_{\tilde{I}}+2}}
H_{\tilde{I}}\right\}^{{B'}_{\tilde{I}}+1}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2\left({B'}_{\tilde{I}}+1\right)}
{{B'}_{\tilde{I}}+2}}
\delta_{a'b'}({\Zsp}')dz^{a'}dz^{b'}\right],
\label{cn:metric-E2:Eq}
\end{eqnarray}
where ${B'}_{\tilde{I}}$ is given by
${B'}_{\tilde{I}}=-B_{\tilde{I}}+a_{\tilde{I}}$ and we define
the cosmic time $\tau$:
\Eq{
\frac{\tau}{\tau_0}=\left(At\right)^{({B'}_{\tilde{I}}+2)/2},~~~~\tau_0=
\frac{2}{\left({B'}_{\tilde{I}}+2\right)A}\,.
}
Hence, in the Einstein frame, the power of the scale factor
in the fastest expanding case is given by
\Eq{
0<\frac{{B'}_{\tilde{I}}+1}{{B'}_{\tilde{I}}+2}<1,
~~~~~{\rm for}~~D-d-2>0
\,.
\label{cn:power-E:Eq}
}
If the physical parameters satisfy \eqref{cn:power-E:Eq},
the solutions do not give
an accelerating expansion in our Universe.
These are the similar results with the case of the other partially
localized and delocalized intersecting brane backgrounds.
Although we find the exact time-dependent brane solution,
the power exponent of the scale factor is too small.
Furthermore, in order to
discuss a de Sitter solution in an intersecting brane background,
one has to consider the trivial dilaton,
which will be discussed in the next subsection.
\subsection{de Sitter universe}
In this subsection, we consider the Einstein
equations (\ref{cn:Einstein:Eq}) with $c_{\tilde{I}}=0$.
Equation (\ref{cn:c:Eq}) gives
\Eq{
N_{\tilde{I}}=\frac{2(D-p_{\tilde{I}}-3)(p_{\tilde{I}}+1)}{(D-2)}.
\label{ds2:N:Eq}
}
If we assume
\Eqr{
p=p_{\tilde{I}}=0\,,~~~
N_{\tilde{I}}=\frac{2(D-3)}{(D-2)}\,,~~~
\alpha_{I'}=-\frac{N_{I'}a_{I'}}{2\epsilon_{I'}c_{I'}}\,,
~~~\Lambda_{I'}=0\,,~~~
{\rm for}~~I'\ne \tilde{I}\,,
}
the field equations reduce to
\Eqrsubl{ds2:equations:Eq}{
&&R_{ij}(\Zsp)=0\,,\\
&&h_{\tilde{I}}(t, z)=K_{\tilde{I}}(t)+H_{\tilde{I}}(z),~~~~~
\left(\frac{dK_{\tilde{I}}}{dt}\right)^2
-\frac{2(D-3)^2}{(D-2)(D-1)}\Lambda_{\tilde{I}}=0,
\label{ds2:warp1:Eq}\\
&&
\triangle_{\Zsp}H_{\tilde{I}}=0\,,~~~~\triangle_{\Zsp}h_{I'}=0.
\label{ds2:warp2:Eq}
}
Then Eq.~(\ref{ds2:warp1:Eq}) gives
\Eq{
K_{\tilde{I}}(t)=c_0t+\tilde{c}\,,
\label{ds2:exact:Eq}
}
where $\tilde{c}$ is an integration constant and
$c_0$ is given by
\Eq{
c_0=\pm(D-3)\sqrt{\frac{2}{(D-2)(D-1)}\Lambda_{\tilde{I}}}\,.
\label{ds2:c0:Eq}
}
Thus, there is no solution for $\Lambda_{\tilde{I}}<0$.
If the metric $u_{ab}({\rm Z})$
is assumed to be Eq.~(\ref{cn:flat metric:Eq}),
the function $H_{\tilde{I}}$ is given by Eq.~(\ref{cn:HI:Eq}).
Now we introduce a new time coordinate $\tau$ by
\Eq{
c_0\tau=\ln t\,.
\label{ds2:cosmic:Eq}
}
The $D$-dimensional metric
(\ref{cn:metric:Eq}) is then rewritten as
\Eqr{
ds^2&=&-\prod_{I'\ne \tilde{I}}h_{I'}^{a_{I'}}(z)
\left(1+c_0^{-1}\e^{-c_0\tau}H_{\tilde{I}}\right)^{-2}
d\tau^2+\left(1+c_0^{-1}
\e^{-c_0\tau}H_{\tilde{I}}\right)^{\frac{2}{D-3}}
\left(c_0\e^{c_0\tau}\right)^{\frac{2}{D-3}}\nn\\
&&\times\left[\sum_{\alpha=1}^p\prod_{I'\ne \tilde{I}}
\left\{h_{I'}(z)\right\}^{\delta^{(\alpha)}_{I'}}(dx^{\alpha})^2
+\prod_{I'\ne \tilde{I}}\left\{h_{I'}(z)\right\}^{b_I}
u_{ab}(\Zsp)dz^adz^b\right].
\label{ds2:metric:Eq}
}
The $D$-dimensional metric (\ref{ds2:metric:Eq}) implies
that the spacetime
describes an isotropic and homogeneous universe
if $H_{\tilde{I}}=0$.
In the
region where
the terms with
$H_{\tilde{I}}$ are negligible
and $h_{I'}$ approaches a constant,
which is realized in the limit $\tau\rightarrow\infty$
and for $c_0>0$,
the $D$-dimensional spacetime becomes de Sitter universe.
If we set $h_{I'}(z)=$const and $u_{ab}=\delta_{ab}$,
Eq.~(\ref{ds2:metric:Eq}) becomes the solution which has been discussed by
Ref.~\cite{Maki:1992tq} (see also \cite{Ivashchuk:1996zv}).
Furthermore, for $D=4$
and by setting all $h_{I'}=1$,
the solution is
the Kastor-Traschen one
\cite{Kastor:1992nn}.
\subsection{The behavior of the solutions}
Now we will study the spacetime structure.
The metric has singularities at $h_{\tilde{I}}=0$ or $h_{I'}=0$\,.
The spacetime is thus not singular when it is restricted inside the domain
specified by the conditions
\Eq{
h_{\tilde{I}}(t, z) = a_0+a_1\,t+K_{\tilde{I}}(z)>0\,,~~~~~
h_{I'}(z)>0\,,
}
where the function $K_{\tilde{I}}$ is defined in (\ref{cn:HI:Eq}).
The $D$-dimensional
spacetime cannot be extended beyond this region, because
a curvature singularity appears in the $D$-dimensional spacetime.
The regular spacetime with branes ends up with the
singularities.
Since the system with $a_1>0$ has the time reversal one of $a_1<0$,
the dynamics of the spacetime depends on the signature of $a_1$\,.
Here we will consider the case with $a_1>0$.
Then the function $h_{\tilde{I}}$ is positive everywhere for $t>0$ and
the spacetime is nonsingular.
In the limit of $t\rightarrow \infty$ and apart from a position of the branes,
near which the geometry takes a cylindrical form of an infinite throat,
the solution is approximately described by
a time-dependent uniform spacetime.
Now we discuss the time evolution for $t\le 0$\,.
The spacetime is regular everywhere and
has a cylindrical topology near each brane at $t=0$\,.
As time slightly decreases, a curvature singularity
appears as $|z^a-z^a_{\alpha}|\rightarrow\infty$\,.
The singular hypersurface cuts off more and more of the space
as time decreases further.
When $t$ continues to decrease, the singular hypersurface eventually splits
and surrounds each of the $p$-brane throats individually.
The spatial surface is finally composed of two isolated throats.
For $t>0$, the time evolution of the $D$-dimensional spacetime
is the time reversal of $t<0$.
For any values of fixed $z^a$ in the regular domain in the
$D$-dimensional spacetime (\ref{cn:metric:Eq}),
the overall transverse space tends to expand asymptotically like
$t^{b_{\tilde{I}}}$\,.
Thus, the solutions describe static intersecting brane systems
composed of $p$-branes near the positions of the branes, while,
in the far region as $|z^a-z^a_{\alpha}| \rightarrow \infty$\,,
the solutions approach de Sitter or FRW universes with
the power-law expansion $t^{b_{\tilde{I}}}$\,.
The emergence of time-dependent universes is an important feature of the
dynamical brane solutions.
\subsubsection{Asymptotic structure}
We study the asymptotic behavior of the solutions.
The solution describes a charged black
hole in the FRW or de Sitter universe in the limit of
$|z^a|\rightarrow \infty$, and $H_{\tilde{I}}$ vanishes.
First we consider the case of a power-law expanding universe.
The function $h_{\tilde{I}}$ depends only on time $t$ in the far region
from branes, and the resulting metric (\ref{cn:metric-J:Eq}) can be expressed
as
\Eqr{
ds^2&=&-d\tau^2+\sum_{\alpha=1}^p
\left(\frac{\tau}{\tau_0}\right)^{\frac{2\delta_{\tilde{I}}^{(\alpha)}}
{a_{\tilde I}+2}}\left(dx^{\alpha}\right)^2
+\left(\frac{\tau}{\tau_0}\right)^{\frac{2b_{\tilde{I}}}{a_{\tilde{I}}+2}}
\delta_{ab}(\Zsp)dz^adz^b\,.
\label{a:metric:Eq}
}
The scale factor of the relative transverse space is given by
$a_{\rm r}(\tau)=\left(\tau/\tau_0\right)
^{b_{\tilde{I}}/a_{\tilde{I}}+2}$, while the expansion low for
the overall transverse space is written by
$a_{\rm t}(\tau)=\left(\tau/\tau_0\right)
^{b_{\tilde{I}}/a_{\tilde{I}}+2}$\,.
On the other hand, for $c_{\tilde{I}}=0$
corresponding to de Sitter universe (\ref{ds2:metric:Eq}),
the metric of $D$-dimensional spacetime in the far region
from branes becomes
\Eqr{
ds^2&=&-d\tau^2+\left(c_0\e^{c_0\tau}\right)^{\frac{2}{D-3}}
\left[\sum_{\alpha=1}^p(dx^{\alpha})^2
+u_{ab}(\Zsp)dz^adz^b\right].
\label{a:metric2:Eq}
}
Figure \ref{fig:p} depicts the
conformal diagrams of the FRW and de Sitter universes.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.28, angle=-90]{frw00.eps}
\hskip 0.5cm
\includegraphics[keepaspectratio, scale=0.28, angle=-90]{ds00.eps}
\\
(a) \hskip 8cm (b) ~~~~~~
\caption{\baselineskip 14pt
Conformal diagrams of the $D$-dimensional spacetime for $p_{\tilde{I}}=0$.
The regions corresponding to $\tilde{r}\rightarrow\infty$ give
the original spacetime, where
$\tilde{r}^2=\sum_\alpha(x^\alpha)^2+\delta_{ab}z^az^b$.
(a) For the case of $a_{\tilde{I}}+2\ne 0$, the metric
(\ref{cn:metric:Eq}) approaches in the limit $\tilde{r}\rightarrow\infty$
to the $D$-dimensional flat FRW spacetime.
(b) We also depict the conformal diagrams in the case of
$a_{\tilde{I}}+2=0$. One can recognize that the asymptotic region of
the spacetime is the de Sitter universe.}
\label{fig:p}
\end{center}
\end{figure}
\subsubsection{Near-horizon geometry}
Next we discuss the near-horizon geometry of the solutions. We set the
metric of $(D-p-1)$-dimensional overall transverse space:
\Eq{
\delta_{ab}(\Zsp)dz^adz^b=dr^2+r^2d\Omega^2_{(D-p-2)}\,,
}
where $\delta_{ab}$ denotes the metric of $(D-p-1)$-dimensional flat space
and the line elements of a $(D-p-2)$-sphere (${\rm S}^{D-p-2}$)
are given by $d\Omega^2_{(D-p-2)}$\,.
The harmonic function $K_{\tilde{I}}$ dominates in the limit of
$r\rightarrow 0$\,, and the time dependence
can be ignored. Thus the system becomes static near a position of branes.
When all of the branes are located at the origin of the Z spaces,
the solutions are rewritten as
\Eqrsubl{nh:h:Eq}{
h_{\tilde{I}}(t, r)&=&a_0+a_1\,t +\frac{M_{\tilde{I}}}{r^{D-p-3}}\,,
\label{nh:hI:Eq}\\
h_{I'}(r)&=&1+\frac{L_{I'}}{r^{D-p-3}}\,.
\label{nh:h':Eq}
}
Here $M_{\tilde{I}}$ and $L_{I'}$ are the mass of $p_{\tilde{I}}$-, and
$p_{I'}$-branes, respectively.
In the near-horizon region $r\rightarrow 0$\,,
the dependence on $t$ in (\ref{nh:h:Eq}) is negligible.
Then the metric is reduced to the following form:
\Eqr{
\hspace{-1cm}
ds^2&=&r^2\left(\frac{M_{\tilde{I}}}{r^{D-p-3}}\right)^{b_{\tilde{I}}}
\prod_{I'}\left(\frac{L_{I'}}{r^{D-p-3}}\right)^{b_{I'}}
\left[-r^{-2}
\left(\frac{M_{\tilde{I}}}{r^{D-p-3}}\right)^{-\frac{4}{N_{\tilde{I}}}}
\prod_{I'}\left(\frac{L_{I'}}{r^{D-p-3}}\right)^{-\frac{4}{N_{I'}}}dt^2
\right.\nn\\
&&\left.\hspace{-1cm}+r^{-2}\sum^p_{\alpha=1}\prod_{I'}
\left(\frac{M_{\tilde{I}}}{r^{D-p-3}}\right)^{-b_{\tilde{I}}
+\delta^{(\alpha)}_{\tilde{I}}}
\left(\frac{L_{I'}}{r^{D-p-3}}\right)^{-b_{I'}
+\delta^{(\alpha)}_I}(dx^{\alpha})^2
+\left(\frac{dr^2}{r^2}+d\Omega^2_{(D-p-2)}\right)\right].
\label{nh:nh:Eq}
}
Thus the metric (\ref{nh:nh:Eq}) describes a warped product of
$(p+2)$-dimensional spacetime M${}_{p+2}$ and $(D-p-2)$-dimensional
sphere ${\rm S}^{D-p-2}$\,.
Hence, the near-brane geometry has the same metric form as the static one.
If it has a horizon geometry,
we can obtain a black hole solution in the time-dependent background.
In fact, some solutions, for instance, the
M2$-$M2$-$M2, M2$-$M2$-$M5$-$M5 intersecting solution in eleven dimensions,
give regular black hole spacetimes in the static limit
\cite{Maeda:2009zi}.
Our solution approaches asymptotically the dynamical universe with
the scale factor $a(\tau)$\,, while the static solution gives a black hole.
Then we can regard the present solution as a black hole in the expanding
universe.
\section{Collision of 0-branes}
\label{sec:c}
In this section, we apply our dynamical intersecting brane
solutions found in the previous section to brane collisions.
The functions
$h_{\tilde{I}}$ and $h_{I'}$ are assumed to be
\Eq{
h_{\tilde{I}}(t, z)=c_0 t +\tilde c_{\tilde{I}}+H_{\tilde{I}}(z)\,,~~~~
h_{I'}=h_{I'}(z)\,.
\label{cp:h:Eq}
}
Here $c_0$ and $\tilde c_{\tilde{I}}$ are constants, and
the function $H_{\tilde{I}}$ and $h_{I'}$ are expressed, respectively, as
\Eqr{
H_{\tilde{I}}(z)&=&\sum_{k=1}^{m}\frac{M_{\tilde{I}, k}}
{|z^a-z^a_k|^{D-p-3-d}}\,,
~~~h_{I'}(z)=\tilde{c}_{I'}
+\sum_{l=1}^{m}\frac{Q_{{I'},\,l}}{|z^a-z^a_l|^
{D-p-3-d_{I'}}}\,,
\label{cp:h2:Eq}
}
where $\tilde{c}_{I'}$ is constant,
$d$ and $d_{I'}$ denote the number of smeared dimension for
$0$-brane and $p_{I'}$-brane, respectively,
we assume $D\neq p+3+d$ and $D\neq p+3+d_{I'}$, and
$M_{\tilde{I}, k}~ (k=1\,,\cdots\,, m)$ and
$Q_{{I'},\,l}~ (l=1\,,\cdots\,, m)$ are mass constants of 0-brane and
$p_{I'}$-branes located at $z^a_k$ and $z^a_l$, respectively.
Since $h_{I'}$ is the harmonic function on the
$(D-p-1-d_{I'})$-dimensional Euclidean subspace in $\Zsp$,
we define
\Eqrsubl{cp:z:Eq}{
&&\hspace{-0.5cm}|z^a-z^a_k|
=\sqrt{\left(z^1-z^1_k\right)^2+\left(z^2-z^2_k\right)^2+\cdots+
\left(z^{D-p-1-d}-z^{D-p-1-d}_k\right)^2}\,,
\label{cp:zk:Eq}\\
&&\hspace{-0.5cm}|z^a-z^a_l|
=\sqrt{\left(z^1-z^1_l\right)^2+\left(z^2-z^2_l\right)^2+\cdots+
\left(z^{D-p-1-d_{I'}}-z^{D-p-1-d_{I'}}_l\right)^2}\,.
\label{cp:zl:Eq}
}
The metric, scalar, and gauge fields are
given by Eqs.~(\ref{cn:metric:Eq}) and
(\ref{cn:fields:Eq}), respectively.
For $D=p+3+d$ and $D=p+3+d_{I'}$, these become
\Eqr{
H_{\tilde{I}}(z)&=&\sum_{k=1}^{m}M_{\tilde{I}, k}\,
\ln |z^a-z^a_k|\,,
~~~h_{I'}(z)=\tilde{c}_{I'}
+\sum_{l=1}^{m}Q_{{I'},\,l}\, \ln |z^a-z^a_l|\,.
\label{cp:h3:Eq}
}
Since the time dependence allows only for the 0-brane,
we see that the $(D-3-d_{I'})$-brane background is critical case.
If we consider the $(D-2-d_{I'})$-brane,
the functions $h_{\tilde{I}}$ and $h_{I'}$ are
written by the sum of linear functions of $z$.
The possibility of brane collisions comes from the difference in
the overall transverse dimension.
From the solution (\ref{cp:h2:Eq}), there are curvature
singularities at $h_{\tilde{I}}=0$ or at $h_{I'}=0$ in the
$D$-dimensional background.
Note that the regular $D$-dimensional spacetime
is restricted to the region of $h_{\tilde{I}}>0$ and $h_{I'}>0$, which is
bounded by curvature singularities.
Hence, the $D$-dimensional metric (\ref{cn:metric:Eq}) is regular
if and only if $h_{\tilde{I}}>0$ and $h_{I'}>0$.
The solution with $0-p_{I'}$ branes
takes the form (\ref{cn:metric-J:Eq}),
where we set $K_{\tilde{I}}=c_0 t$
and the function $H_{\tilde{I}}$ is given by (\ref{cn:HI:Eq}).
We classify the behavior of the harmonic function $h_{I'}$ into
two classes: $p_{I'}\le (D-4-d_{I'})$ and $p_{I'}= (D-2-d_{I'})$.
Since these depend on the dimensions of the $p_{I'}$-brane,
we discuss them below separately.
In the case of the $(D-3-d_{I'})$-brane, the harmonic function
$h_{I'}$ diverges both at infinity and near $(D-3-d_{I'})$-branes.
Since there is no regular spacetime region near branes due to
$h_{I'}\rightarrow -\infty$, these solutions are not physically relevant.
In the following, we discuss the collision involving the $0-p_{I'}$ brane
in $D$-dimensional spacetime.
\subsection{Collision of the $0-p_I$-brane in the
asymptotically power-law expanding universe}
The harmonic function $H_{\tilde{I}}$ becomes dominant in
the limit of $z^a\rightarrow \,z^a_k$\,, while
the function $h_{\tilde{I}}$ depends only on time $\tau$
in the limit of $|z^a|\rightarrow \infty$.
Hence, we find a static structure of the $0-p_{I'}$-brane system
near branes. In the far region from branes, the function
$H_{\tilde{I}}$ vanishes.
Therefore, the metric can be written by
\Eqr{
ds^2&=&-\prod_{I'\ne \tilde{I}}h_{I'}^{a_{I'}}
\bar{h}_{\tilde{I}}^{a_{\tilde{I}}}d\tau^2+\sum_{\alpha=1}^p
\prod_{I'\ne \tilde{I}}h_{I'}^{\delta_{I'}^{(\alpha)}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2\delta_{\tilde{I}}^{(\alpha)}}
{a_{\tilde I}+2}}\bar{h}_{\tilde{I}}^{\delta_{\tilde{I}}^{(\alpha)}}
\left(dx^{\alpha}\right)^2\nn\\
&&+\prod_{I'\ne \tilde{I}}h_{I'}^{b_{I'}}
\left(\frac{\tau}{\tau_0}\right)^{\frac{2b_{\tilde{I}}}{a_{\tilde{I}}+2}}
\bar{h}_{\tilde{I}}^{b_{\tilde{I}}}\delta_{ab}(\Zsp)dz^adz^b\,,
\label{co:sufrace2:Eq}
}
where $\bar{h}_{\tilde{I}}$ is defined by
\Eq{
\bar{h}_{\tilde{I}}=
1+\left(\frac{\tau}{\tau_0}\right)^{-\frac{2}{a_{\tilde{I}}+2}}
H_{\tilde{I}}\,.
}
In order to analyze the brane collision,
we consider a concrete example, in which
two $0-p_{I'}$ branes are located at $z^a=(\pm L, 0, \ldots, 0)$.
We will discuss the time evolution separately with respect to the signature of
a constant $\tau_0$\,, because the behavior of spacetime strongly
depends on it. Since the metric function is singular
at $h_{\tilde{I}}(\tau, z)=0$ and $h_{I'}=0$, one can note that
the regular spacetime exists inside the domain restricted by
\Eq{
h_{\tilde{I}}(\tau, z) = \left(\frac{\tau}{\tau_0}\right)^
{\frac{2}{a_{\tilde{I}}+2}}+H_{\tilde{I}}(z)>0\,,~~~~~
h_{I'}=h_{I'}(z)>0\,,
}
where the functions $H_{\tilde{I}}$ and
$h_{I'}$ are defined in (\ref{cp:h:Eq}).
The brane background evolves into a curvature singularity,
because the dilaton $\phi$ diverges.
Since the $D$-dimensional spacetime cannot be extended beyond this region,
the regular spacetime with two 0-branes ($p+d\le 6$) ends on these singular
hypersurfaces.
The solution with $(\tau_0)^{-2/(a_{\tilde{I}}+2)}>0$ is the time
reversal one of $(\tau_0)^{-2/(a_{\tilde{I}}+2)}<0$,
because the time dependence appears only in the form of
$(\tau/\tau_0)^{2/(a_{\tilde{I}}+2)}$.
In the following, we consider
the case with $(\tau_0)^{-2/(a_{\tilde{I}}+2)}<0$.
For $(\tau)^{2/(a_{\tilde{I}}+2)}<0$,
the $D$-dimensional spacetime is nonsingular, because the function
$h_{\tilde{I}}$ is positive everywhere.
In the limit of $(\tau)^{2/(a_{\tilde{I}}+2)}\rightarrow -\infty$,
the $D$-dimensional spacetime becomes asymptotically
a time-dependent uniform background,
while the cylindrical forms of infinite
throats exist near branes.
For $\tau>0$, the spatial metric is initially
regular everywhere. The $D$-dimensional
spacetime has a cylindrical topology
near each brane.
As $\tau$ increases slightly, a singular hypersurface appears from
the spatial infinity
($|z^a-z^a_k|\rightarrow\infty$).
As $\tau$ increases further, the singularity cuts the space off
more and more.
Since the singular hypersurface eventually splits
and surrounds each of the brane throats,
the spatial surface is finally composed of two isolated throats.
One notes that the transverse
dimensions in the metric (\ref{co:sufrace2:Eq})
expand asymptotically as $\tau^{b_{\tilde{I}}/(a_{\tilde{I}}+2)}$
for fixed spatial coordinates $z^a$.
The $D$-dimensional spacetime becomes static near branes, while the
background approaches a FRW universe in the far region
($|z^a-z^a_k|\rightarrow \infty$).
Hence, the time evolution of the four-dimensional universe depends on the
position of the observer.
For $(\tau/\tau_0)^{2/(a_{\tilde{I}}+2)}<0$, the behavior of
$D$-dimensional spacetime is
the time reversal of the period of $(\tau/\tau_0)^
{2/(a_{\tilde{I}}+2)}>0$.
Now we define
\Eq{
z_{\perp}=\sqrt{\left(z^2\right)^2+\left(z^3\right)^2+\cdots
+\left(z^{D-1-p-d}\right)^2}\,.
\label{cp:zp:Eq}
}
By using the above equation, the proper distance at $z_{\perp}=0$
between two branes can be written by
\begin{eqnarray}
d(\tau)&=&\int_{-L}^L dz^1
\left[\left(\frac{\tau}{\tau_0}\right)^{2/(a_{\tilde{I}}+2)}
+\frac{M_1}{|z^1+L|^{D-3-p_{I'}-d_{I'}}}
+\frac{M_2}{|z^1-L|^{D-3-p_{I'}-d_{I'}}}\right]^
{b_{\tilde{I}}/2}
\nn\\
&&\times
\left(1+\frac{Q_1}{|z^1+L|^{D-3-p_{I'}-d}}
+\frac{Q_2}{|z^1-L|^{D-3-p_{I'}-d}}\right)^{b_{I'}/2}
\,.
\label{cp:distance:Eq}
\end{eqnarray}
The proper distance is a monotonically increasing function of $\tau$.
We illustrate $d(\tau)$
for the case of the $0-p_{I'}$ brane system in Fig. \ref{fig:mil}.
We consider the case of $d=d_{I'}=0$,
$\tau_0=-1$, $Q_1=Q_2=M_1=M_2=1$, $L=1$\,, and $D=10$ or $D=8$.
It shows that two 0-branes are initially ($\tau< 0$) approaching,
the distance $d(\tau)$ takes the minimum finite value at $\tau=0$,
and then two 0-branes segregate each other.
Thus they will never collide.
Hence, we cannot discuss a brane collision in this case.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.6]{w1.eps}
\put(-100, 135){$d(\tau)$}
\put(10,10){$\tau$}
\hskip 2cm
\includegraphics[keepaspectratio, scale=0.6]{v1.eps}
\put(-105,135){$d(\tau)$}
\put(10,10){$\tau$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
(a) For the case of $M_1=M_2$ in the asymptotically power-law
expanding universe, the proper distance between two dynamical
0-branes given in (\ref{cp:distance:Eq}) is depicted. We fix $d=d_{I'}=0$,
$D=10$, $\tau_0=-1$, $M_1=1,
M_2=1$, $N=2$\,, and $L=1$ for the $0-8$- (bold curve) and
$0-6$- (solid curve) branes.
The distance decreases initially ($\tau<0$)
but turns to increase at $\tau=0$, and
then two 0-branes segregate each other.
(b) We also show the proper distance between two dynamical
0-branes for $0-8$- (bold curve) and $0-6$- (solid curve)
brane systems from the bottom in the case of $d=d_{I'}=0$,
$M_1=10$, $M_2=1$, $N=2$, $L=1$, and $D=10$ in the
asymptotically power-law expanding universe.
Although the proper distance initially decreases as $\tau(<0)$ increases,
the distance increases as $\tau(>0)$ increases.}
\label{fig:mil}
\end{center}
\end{figure}
\subsection{Collision of the $0-p_{I'}$-brane
in the asymptotically de Sitter universe}
Let us next discuss the collision in the $0-p_{I'}$-brane with a
trivial dilaton system. We consider the case that
the harmonic function $H_{\tilde{I}}$ and $h_{I'}$ are linear in $z$
and discuss in detail the
$0-(D-2)$-brane in $D$ dimensions as a example.
In this case, we have one extra dimension $z$ in $\Zsp$ space.
The $D$-dimensional metric (\ref{ds2:metric:Eq})
can be rewritten by
\Eqr{
ds^2&=&-h_{I'}^{a_{I'}}(z)
\left(1+c_0^{-1}\e^{-c_0\tau}H_{\tilde{I}}\right)^{-2}
d\tau^2+h_{I'}^{a_{I'}}(z)
\left(1+c_0^{-1}\e^{-c_0\tau}H_{\tilde{I}}\right)^{\frac{2}{D-3}}
\left(c_0\e^{c_0\tau}\right)^{\frac{2}{D-3}}\nn\\
&&\times\left[\sum_{\alpha=1}^p(dx^{\alpha})^2
+h_{I'}^{4/N_{I'}}(z)dz^2\right]\,,
\label{cods:metric:Eq}
}
where the function $H_{\tilde{I}}(z)$ is written by
\Eqr{
H_{\tilde{I}}(z)=\sum_{k=1}^{m}M_{\tilde{I}\,, k}\,|z-z_k|\,.
\label{cods:L:Eq}
}
We consider the collision in the $0-p_{I'}$-brane system
with charges $M_1$ and $Q_1$ at $z^1=-L$
and the other with charges $M_2$ and $Q_2$ at $z^1=L$.
The proper distance at $z_{\perp}=0$ between the two $0$-branes can be
expressed as
\Eqr{
d(\tau)&=&\int^L_{-L} dz
\left(c_0\e^{c_0\tau}+M_1\,|\,z^1+\,L|
+M_2\,|\,z^1-\,L|\right)^{1/(D-3)}\nn\\
&&\times
\left(1+Q_1\,|\,z^1+\,L|
+Q_2\,|\,z^1-\,L|\right)^{b_{I'}/2}\,.
\label{ds:length:Eq}
}
In the period of $c_0<0$,
the proper distance increases as
$\tau$ increases.
If $M_1\ne M_2$, a singular hypersurface appears at
$\tau=\tau_s\equiv\ln\left[-(M_1|z^1+L|+M_2|z^1-L|)c_0^{-1}\right]c_0^{-1}<0$
when the distance is still finite.
However, in the case of the equal charges $Q_1=Q_2=M_1=M_2=M$, the
situation is completely different, because
the proper distance finally vanishes at $\tau_s=
\ln\left(-2MLc_0^{-1}\right)c_0^{-1}<0$
as
\Eqr{
d(\tau)= 2L\left(c_0\e^{c_0\tau}+2LM\right)^{1/(D-3)}
\left(1+2LM\right)^{b_{I'}/2}
\,.}
Then two branes can collide.
A singularity is formed
at the same point and time.
Let us consider the case $p_{I'}\ne D-2$.
The $D$-dimensional metric (\ref{ds2:metric:Eq})
can be written as
\Eqr{
ds^2&=&-h_{I'}^{a_{I'}}(z)
\left(1+c_0^{-1}\e^{-c_0\tau}H_{\tilde{I}}\right)^{-2}
d\tau^2+h_{I'}^{a_{I'}}(z)
\left(1+c_0^{-1}\e^{-c_0\tau}H_{\tilde{I}}\right)^{\frac{2}{D-3}}
\left(c_0\e^{c_0\tau}\right)^{\frac{2}{D-3}}\nn\\
&&\times\left[\sum_{\alpha=1}^p(dx^{\alpha})^2
+h_{I'}^{4/N_{I'}}(z)\delta_{ab}(\Zsp)dz^adz^b\right].
\label{cods:metric2:Eq}
}
Since the proper distance at $z_{\perp}=0$
between two branes is given by
\Eqr{
d(\tau)&=&\int^L_{-L} dz
\left(c_0\e^{c_0\tau}+\frac{M_1}{|\,z^1+\,L|^{D-p_{I'}-3-d}}
+\frac{M_2}{|\,z^1-\,L|^{D-p_{I'}-3-d}}\right)^{1/(D-3)}
\nn\\
&&\times
\left(1+\frac{Q_1}{|z^1+L|^{D-3-p_{I'}-d}}
+\frac{Q_2}{|z^1-L|^{D-3-p_{I'}-d}}\right)^{b_{I'}/2}
\,,
\label{distance2}
}
the distance increases monotonically with respect to $\tau$.
In the case of $c_0<0$, initially ($\tau=0$), $D$-dimensional space is regular
except at $|z^a-z^a_k|\rightarrow 0$\,, while this is an asymptotically
time-dependent spacetime and has the cylindrical form of an infinite
throat near the 0-brane.
At $\tau=\tau_s<0$,
a singularity appears
from the spatial infinity ($|z^a-z^a_k|\rightarrow\infty$).
As time decreases ($\tau<0$),
the singular hypersurface erodes the region with
the large values of $|z^a-z^a_k|$.
Since only the region near 0-branes remains regular, eventually
it splits and each fragment surrounds each 0-brane
individually.
Figure \ref{fig:ds} shows that
this singularity appears before the proper distance $d(\tau)$ vanishes.
Hence, the $D$-dimensional spacetime has the singularity
before two branes collide.
Although two 0-branes approach very slowly,
a singularity suddenly appears
at a finite distance. Then,
the spacetime splits into two isolated 0-brane throats.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.58]{f.eps}
\put(-165, 125){$d(\tau)$}
\put(10,10){$\tau$}
\hskip 2cm
\includegraphics[keepaspectratio, scale=0.58]{rr1.eps}
\put(-165,125){$d(\tau)$}
\put(10,10){$\tau$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
(a) For the case of $M_1=M_2$ in the asymptotically de Sitter universe,
we show the proper distance between two dynamical 0-branes given in
(\ref{ds:length:Eq}). We set $D=10$, $c_0=-1$, $M_1=1,
M_2=1$, $N=2$, and $L=1$ for the $0-8$-brane. The proper distance rapidly
vanishes near where two branes collide.
(b) We also show the proper distance between two dynamical
0-branes for the $0-8$- (bold curve) and
$0-6$- (solid curve) brane systems
from the bottom in the case of $M_1=10$, $M_2=1$, $N=2$,
and $D=10$ in the asymptotically de Sitter universe.
The proper distance initially decreases as $\tau$ decreases
and remains still finite when a singularity appears.}
\label{fig:ds}
\end{center}
\end{figure}
We show $d(\tau)$ integrated numerically in Fig. \ref{fig:ds}
for the case of $c_0<0$.
In the future direction, the proper distance $d$ increases.
Then for $\tau>0$, each brane gradually
separates as $\tau$ increases.
\section{Applications to supergravities}
\label{sec:sug}
In the case of ten or eleven dimensions with $N=4$ and $\Lambda_I=0$,
Eq.~(\ref{pl:action:Eq}) gives the action of supergravities.
For instance, the bosonic part of the action of $D=11$ supergravity
includes only 4-form field strength, while, for $D=10$,
the constant $c$ is precisely the dilaton coupling for the
Ramond-Ramond $\left(p+2\right)$-form in the type II supergravities.
The dynamical solutions for the case of $N=4$
have been already discussed in Ref.~\cite{Maeda:2010aj}. In this section,
we will discuss the time-dependent solution in six-dimensional
Nishino-Salam-Sezgin (NSS) gauged supergravity
and Romans' gauged supergravity models.
The bosonic part of the six-dimensional
NSS model~\cite{Salam:1984cj, Nishino:1984gk, Nishino:1986dc, Burgess:2003mk}
is given by the expression (\ref{pl:action:Eq}) with $\Lambda_r>0$,
$\Lambda_s=0$, while Romans' six-dimensional
gauged supergravity \cite{Romans:1985tw} is expressed by the action
(\ref{pl:action:Eq}) with $\Lambda_r<0$, $\Lambda_s=0$.
\subsection{Nishino-Salam-Sezgin gauged supergravity}
Now we consider the NSS model among the theories of $D=6$.
The couplings of the 2-form ($p_r=0$)
and the 3-form ($p_s=1$) field strengths to the dilaton
are given by $\epsilon_rc_r=-\frac{1}{\sqrt{2}}$ and
$\epsilon_sc_s=-\sqrt{2}$, respectively:
\Eqr{
S&=&\frac{1}{2\kappa^2}\int \left[\left(R-2\e^{\phi/\sqrt{2}}
\Lambda\right)\ast{\bf 1}
-\frac{1}{2}\ast d\phi \wedge d\phi\right.\nn\\
&&\left.\hspace{-0.5cm}
-\frac{1}{2\cdot 2!}
\e^{-\phi/\sqrt{2}}\ast F_{(2)}\wedge F_{(2)}
-\frac{1}{2\cdot 3!}
\e^{-\sqrt{2}\phi}\ast F_{(3)}\wedge F_{(3)}
\right],
\label{nss:action:Eq}
}
where $R$ denotes the Ricci scalar constructed from the six-dimensional
metric $g_{MN}$\,, $\kappa^2$ is the six-dimensional gravitational constant,
$\ast$ is the Hodge operator in the six-dimensional spacetime,
$\phi$ denotes the scalar field, $\Lambda>0$ is the cosmological
constant, and $F_{\left(2\right)}$ and $F_{\left(3\right)}$
are 2-form and 3-form field strengths, respectively.
From Eq.~(\ref{pl:c:Eq}), the NSS model
is realized by choosing $\Lambda_r=\Lambda>0$,
$\Lambda_s=0$, $N_r=2$, and $N_s=4$.
The six-dimensional action (\ref{nss:action:Eq}) gives the field equations
\Eqrsubl{nss:equations:Eq}{
&&
R_{MN}=\frac{1}{2}\e^{\phi/\sqrt{2}}\Lambda g_{MN}
+\frac{1}{2}\pd_M\phi \pd_N \phi
+\frac{\e^{-\phi/\sqrt{2}}}{2\cdot 2!}
\left(2F_{MA}{F_N}^A
-\frac{1}{4} g_{MN} F^2_{\left(2\right)}\right)\nn\\
&&~~~~~~~~~+\frac{\e^{-\sqrt{2}\phi}}{2\cdot 3!}
\left(3F_{MAB} {F_N}^{AB}
-\frac{1}{2} g_{MN} F_{\left(3\right)}^2\right),
\label{nss:Einstein:Eq}\\
&&
\lap\phi+
\frac{\sqrt{2}}{4\cdot 2!}\,\e^{-\phi/\sqrt{2}}\,
F^2_{\left(2\right)}+\frac{\sqrt{2}}{2\cdot 3!}\,\e^{-\sqrt{2}\phi}
F^2_{\left(3\right)}-\sqrt{2}\,\e^{\phi/\sqrt{2}}\,\Lambda=0\,,
\label{nss:scalar:Eq}\\
&&
d\left[\e^{-\phi/\sqrt{2}}\ast F_{\left(2\right)}\right]=0\,,
\label{nss:gauge-r:Eq}\\
&&
d\left[\e^{-\sqrt{2}\phi}\ast F_{\left(3\right)}\right]=0\,,
\label{nss:gauge-s:Eq}
}
where $\lap$ denotes the Laplace operator with respect to the six-dimensional
metric $g_{MN}$\,.
We look for solutions whose spacetime
metric has the form
\Eqr{
ds^2&=&h^{1/2}_2(t, y, z)h_3^{1/2}(t, y, z)\left[-h^{-2}_2(t, y, z)
h^{-1}_3(t, y, z)dt^2\right.\nn\\
&&\left.+h^{-1}_3(t, y, z)dy^2
+u_{ab}(\Zsp)dz^adz^b\right],
\label{nss:metric:Eq}
}
where $u_{ab}(\Zsp)$ is the four-dimensional metric which
depends only on the four-dimensional coordinates $z^a$.
The scalar field $\phi$ and field
strengths $F_{\left(2\right)}$, and $F_{\left(3\right)}$ are written,
respectively, by
\Eqrsubl{nss:ansatz:Eq}{
\e^{\phi}&=&\left(h_2h_3\right)^{-\sqrt{2}/2}\,,
\label{nss:phi:Eq}\\
F_{\left(2\right)}&=&d\left[\sqrt{2}\,h^{-1}_2(t, y, z)\right]\wedge dt,
\label{nss:fr:Eq}\\
F_{\left(3\right)}&=&d\left[h^{-1}_3(t, y, z)\right]\wedge dt\wedge dy,
\label{nss:fs:Eq}
}
First we consider the Einstein equation (\ref{nss:Einstein:Eq}).
By using the ansatz (\ref{nss:metric:Eq}) and (\ref{nss:ansatz:Eq}),
the Einstein equations become
\Eqrsubl{nss:cEinstein:Eq}{
&&\frac{5}{4}h_2^{-1}\pd_t^2h_2
+\frac{3}{4}h_3^{-1}\pd_t^2h_3+
\frac{1}{4}h_2^{-2}\left(3h_2^{-1}\pd_y^2h_2+h_3^{-1}\pd_y^2h_3\right)
+\frac{1}{4}h_2^{-2}h_3^{-1}
\left(3h_2^{-1}\lap_{\Zsp}h_2+h_3^{-1}\lap_{\Zsp}h_3\right)\nn\\
&&~~~~-\frac{1}{2}h_2^{-2}h_3^{-1}\Lambda
+\frac{1}{4}\left(\pd_t\ln h_2\right)^2
+\frac{7}{4}\pd_t\ln h_2\pd_t\ln h_3
-\frac{3}{4}h_2^{-2}(1-h_3)\left(\pd_y\ln h_2\right)^2\nn\\
&&~~~~+\frac{3}{4}h_2^{-2}\pd_y\ln h_2\pd_y\ln h_3
-\frac{3}{4}h_2^{-2}h_3^{-1}(1-h_3)u^{ab}\pd_a\ln h_2\pd_b\ln h_2\nn\\
&&~~~~-\frac{1}{4}h_2^{-2}h_3^{-1}(1-h_2^2)u^{ab}\pd_a\ln h_3\pd_b\ln h_3=0,
\label{nss:cEinstein-tt:Eq}\\
&&2h_2^{-1}\pd_t\pd_y h_2+2h_3^{-1}\pd_t\pd_y h_3
+\pd_t\ln h_2\pd_y\ln h_3+3\pd_t\ln h_3\pd_y\ln h_2=0,
\label{nss:cEinstein-ty:Eq}\\
&&2h_2^{-1}\pd_t\pd_a h_2+h_3^{-1}\pd_t\pd_a h_3
+\pd_t\ln h_2\pd_a\ln h_3+\pd_t\ln h_3\pd_a\ln h_2=0,
\label{nss:cEinstein-ta:Eq}\\
&&\frac{1}{4}h_2^2\left(h_2^{-1}\pd_t^2h_2-h_3^{-1}\pd_t^2h_3\right)
-\frac{1}{4}\left(h_2^{-1}\pd_y^2h_2+3h_3^{-1}\pd_y^2h_3\right)
-\frac{1}{4}h_3^{-1}
\left(h_2^{-1}\lap_{\Zsp}h_2-h_3^{-1}\lap_{\Zsp}h_3\right)\nn\\
&&~~~-\frac{1}{2}h_3^{-1}\Lambda+\frac{1}{4}\left(\pd_th_2\right)^2
-\frac{1}{4}h_2^2\pd_t\ln h_2\pd_t\ln h_3
-\frac{3}{4}(1-h_3)\left(\pd_y\ln h_2\right)^2
-\frac{5}{4}\pd_y\ln h_2\pd_y\ln h_3\nn\\
&&~~~+\frac{1}{4}h_3^{-1}(1-h_3)u^{ab}\pd_a\ln h_2\pd_b\ln h_2
-\frac{1}{4}h_3^{-1}(1-h_2^2)u^{ab}\pd_a\ln h_3\pd_b\ln h_3=0,
\label{nss:cEinstein-yy:Eq}\\
&&h_3^{-1}\pd_y\pd_a h_3+2\pd_y\ln h_2\pd_a\ln h_2
+\pd_y\ln h_2\pd_a\ln h_3+\pd_y\ln h_3\pd_a\ln h_2=0,
\label{nss:cEinstein-ya:Eq}\\
&&R_{ab}(\Zsp)+\frac{1}{4}h_2^2h_3u_{ab}
\left(h_2^{-1}\pd_t^2 h_2+h_3^{-1}\pd_t^2 h_3\right)
-\frac{1}{4}h_3u_{ab}
\left(h_2^{-1}\pd_y^2 h_2+h_3^{-1}\pd_y^2 h_3\right)\nn\\
&&~~~-\frac{1}{4}u_{ab}
\left(h_2^{-1}\lap_{\Zsp}h_2+h_3^{-1}\lap_{\Zsp}h_3\right)
+\frac{1}{4}h_2^2h_3u_{ab}\left[\left(\pd_t\ln h_2\right)^2
+3\pd_t\ln h_2\pd_t\ln h_3\right]\nn\\
&&~~~+\frac{1}{4}h_3\left(1-h_3\right)u_{ab}\left(\pd_y\ln h_2\right)^2
-\frac{1}{4}h_3u_{ab}\pd_y\ln h_2\pd_y\ln h_3
+\frac{1}{4}(1-h_3)u_{ab}u^{cd}\pd_c\ln h_2\pd_d\ln h_2\nn\\
&&~~~+\frac{1}{4}(1-h_2^2)u_{ab}u^{cd}\pd_c\ln h_3\pd_d\ln h_3
-(1-h_3)\pd_a\ln h_2\pd_b\ln h_2\nn\\
&&~~~-\frac{1}{2}(1-h_2^2)\pd_a\ln h_3\pd_b\ln h_3
-\frac{1}{2}\left(\pd_a\ln h_2\pd_b\ln h_3
+\pd_a\ln h_3\pd_b\ln h_2\right)-\frac{1}{2}u_{ab}\,\Lambda=0,
\label{nss:cEinstein-ab:Eq}
}
where $\triangle_{\Zsp}$ denotes the Laplace operator on $\Zsp$ space and
$R_{ab}(\Zsp)$ is the Ricci tensor
constructed from the metric $u_{ab}(\Zsp)$\,.
We next consider the gauge field equations
(\ref{nss:gauge-r:Eq}) and (\ref{nss:gauge-s:Eq}).
Under the assumption (\ref{nss:ansatz:Eq}),
the gauge field equations are written by
\Eqrsubl{nss:gauge2:Eq}{
&&d\left[h_3^2\,\pd_y h_2\,\Omega(\Zsp)
+h_3\,\pd_a h_2 dy\wedge\left(\ast_{\Zsp}dz^a\right)\right]=0,
\label{nss:gauge2-r:Eq}\\
&&d\left[h_2\,\pd_a h_3\left(\ast_{\Zsp}dz^a\right)\right]=0,
\label{nss:gauge2-s:Eq}
}
where $\ast_{\rm Z}$ denotes the Hodge operator on Z and
$\Omega(\Zsp)$ is the volume 4-form on Z space:
\Eq{
\Omega(\Zsp)=\sqrt{u}\,dz^1\wedge dz^2\wedge dz^3 \wedge dz^4\,.
\label{nss:volume:Eq}
}
Here, $u$ is the determinant of the metric $u_{ab}$\,.
Finally we consider the equation of motion for the scalar field.
Substituting the ansatz (\ref{nss:ansatz:Eq})
into Eq.~(\ref{nss:scalar:Eq}), we have
\Eqr{
&&h_2^2h_3\left(h_2^{-1}\pd_t^2h_2+h_3^{-1}\pd_t^2h_3\right)
+h_3\left(\pd_th_2\right)^2
+3h_2\pd_th_2\pd_th_3-h_3\left(h_2^{-1}\pd_y^2h_2
+h_3^{-1}\pd_y^2h_3\right)\nn\\
&&~~~~+h_3\left(1-h_3\right)\left(\pd_y\ln h_2\right)^2
-h_2^{-1}\pd_yh_2\pd_yh_3-h_2^{-1}\lap_{\Zsp}h_2-h_3^{-1}\lap_{\Zsp}h_3\nn\\
&&~~~~+\left(1-h_3\right)u^{ab}\pd_a\ln h_2\pd_b\ln h_2
+\left(1-h_2^2\right)u^{ab}\pd_a\ln h_3\pd_b\ln h_3-2\Lambda=0\,.
\label{nss:scalar-e:Eq}
}
Now we consider the two cases.
One is $\pd_th_2\ne 0$ and $\pd_th_3=0$. The other is
$\pd_th_2= 0$ and $\pd_th_3\ne 0$.
Upon setting $h_2=1$, the field equations reduce to
\Eqrsubl{nss:equation:Eq}{
&&R_{ab}(\Zsp)=0\,,
\label{nss:Ricci:Eq}\\
&&h_2=1\,,~~~~h_3=k_0(t)+k_1(y)+k_2(z),
\label{nss:h1:Eq}\\
&&\frac{d^2k_0}{dt^2}=\Lambda, ~~~~~
\frac{d^2k_0}{dt^2}=-\frac{d^2k_1}{dy^2}\,,~~~~\triangle_{\Zsp}k_2=0\,.
\label{nss:h2:Eq}
}
We can also choose the solution in which the
0-brane part depends on $t$. Then, we have
\Eqrsubl{nss:solution2:Eq}{
&&R_{ab}(\Zsp)=0,
\label{nss:Ricci2:Eq}\\
&&h_2=h_2(t, v),~~~~h_2=K_0(t)+K_1(v)\,,~~~~~h_3=1\,,
\label{nss:h2-2:Eq}\\
&&\left(\frac{dK_0}{dt}\right)^2=2\Lambda,~~~~\triangle_{\rm W}K_1=0\,,
\label{nss:warp2:Eq}
}
where $\lap_{\rm W}$ denotes Laplace operator with respect to the metric
$w_{mn}$\,:
\Eq{
w_{mn}dv^mdv^n=dy^2+u_{ab}(\Zsp)dz^adz^b\,,~~~~
\triangle_{\rm W}K_1=\pd_y^2K_1+\triangle_{\Zsp}K_1\,.
\label{nss:lapW:Eq}
}
Here, $w_{mn}$ is the five-dimensional metric, and $v^m$
denotes the five-dimensional coordinate.
As a special example, we consider the case
\Eq{
u_{ab}=\delta_{ab}\,,~~~~~~h_2=1\,,
\label{nss:flat:Eq}
}
where $\delta_{ab}$ the four-dimensional Euclidean metric.
Then, the solution for $h_3$ can be obtained explicitly as
\cite{Minamitsuji:2010fp}
\Eqrsubl{nss:solution:Eq}{
h_2&=&1,
\label{nss:solution-h2:Eq}\\
h_3(t, y, z)&=&\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1t+c_2y
+c_3+\sum_{l=1}^N\frac{M_l}{|z^a-z^a_l|^2},
\label{nss:solution-h3:Eq}
}
where $c_i~(i=1,~2,~3)$ and $z^a_l$ are constants and
the parameter $M_l$ is the mass constant of 1-branes\,,
which is located at $z^a=z^a_l$\,.
We can obtain the solution
for $h_3=1$ and $\pd_th_2\ne 0$
if the roles of $h_2$ and $h_3$ are exchanged.
The solution of the field equations is then written as
\Eqrsubl{nss:solution3:Eq}{
h_2(t, v)&=&\epsilon\sqrt{2\Lambda}\,t+c_4+\sum_{\alpha=1}^{N'}
\frac{L_{\alpha}}
{|v^m-v^m_{\alpha}|^3}\,,
\label{nss:solution3-r:Eq}\\
h_3&=&1\,,
\label{nss:solution3-s:Eq}
}
where $c_4$, $v^m_\alpha$, and $L_\alpha$ are constants and
$\epsilon=\pm 1$.
The delocalized brane solutions
in the six-dimensional NSS supergravity
\cite{Nishino:1984gk, Salam:1984cj, Nishino:1986dc,
Gibbons:2003di,Aghababaie:2003ar} have been investigated
in Refs.~\cite{Maeda:1984gq,Maeda:1985es, Cline:2003ak, Vinet:2004bk,
Vinet:2005dg,
Peloso:2006cq, Tolley:2006ht, Tolley:2007et,Minamitsuji:2010fp},
including applications to cosmological models.
According to the intersection rule,
the number of the intersections dimensions involving the 0-brane and
1-brane is $-1$. Although meaningless in ordinary spacetime,
these configurations are relevant in the Euclidean space,
for instance, representing instantons.
In the following, we consider cosmological aspects of the solution describing
time-dependent branes. We first study the time dependence of the
scale factors in the 0-brane solutions after compactifying the extra
directions, and our Universe is discussed. Next we discuss the
dynamical 1-brane solution and apply it to the cosmology.
\subsubsection{Cosmology in the 0-brane system}
For the solution (\ref{nss:solution3:Eq}),
we introduce a new time coordinate $\tau$ as
\Eq{
\left(\frac{\tau}{\tau_0}\right) \equiv
\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{1/4}\,,~~~~
\tau_0 \equiv \frac{4}{\epsilon\sqrt{2\Lambda}}\,.
\label{nss:ct:Eq}
}
The six-dimensional metric is thus given by
\Eqr{
&&\hspace{-1.2cm}ds^2=
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-4}\bar{h}_2(v)
\right]^{-\frac{3}{2}}
\left[-d\tau^2
+\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-4}
\bar{h}_2(v)\right\}
\left(\frac{\tau}{\tau_0}\right)^2
\delta_{mn}({\rm W})dv^mdv^n
\right],
\label{nss:metric3:Eq}
}
where $\delta_{mn}$ is the five-dimensional Euclidean metric
and $\bar{h}_2(v)$ is defined by
\Eq{
\bar{h}_2(v) \equiv \sum_{\alpha=1}^{N'}\frac{L_{\alpha}}
{|v^m-v^m_{\alpha}|^{3-d_s}}\,.
\label{nss:bh:Eq}
}
Here $d_{s}$ denotes the number of smeared dimensions and
should satisfy $0\le d_{s}\le 4$\,.
The six-dimensional spacetime implies $(\tau/\tau_0)^{-4}\,\bar{h}_2(v)=0$
in the limit $\tau\rightarrow\infty$\,.
Then the scale factor of the six-dimensional space is
proportional to $\tau$\,.
Although the dynamical 0-brane solutions cannot give a realistic universe
such as an accelerating expansion, a matter- or a
radiation-dominated era, there is a possibility that appropriate
compactification and smearing of the transverse space to the 0-brane
may lead to a realistic expansion. Now we will discuss
this possibility.
We consider some compactification and smearing of the extra
directions of the solutions.
Our Universe has to be described by the 0-brane solution
with six directions. Since
the time direction is expressed as $t$\,,
the remaining task is to identify
the three spatial directions from the coordinates $v^m$.
In an approach such as the construction of the
cosmological scenario on the basis of a dynamical brane background,
three spatial directions are supposed to be on the overall
transverse space to branes.
If the spatial directions are specified with $v^m$\,,
it also works in the present case.
Then space is isotropic from the expression of the metric.
Now we look for a way to realize an isotropic
and homogeneous three-dimensional space in the 0-brane solutions.
Since we set the coordinates $(t, v^2, v^3, v^4)$ which describes our
Universe, it is convenient to decompose the six-dimensional metric of
the solutions into the following form:
\Eq{
ds^2=ds_4^2+ds^2_{\rm i}\,,
\label{nss:c-metric:Eq}
}
where each part of the six-dimensional metric is given by
\Eqrsubl{nss:c-metric2:Eq}{
ds^2_4&=&-h^{-3/2}_2(t, v)dt^2+h^{1/2}_2(t, v)
\delta_{\alpha\beta}dv^\alpha dv^\beta\,, \\
ds^2_{\rm i}&=&h^{1/2}_2(t, v)\delta_{ij}dv^idv^j\,.
}
Here $ds^2_4$ is the metric of the four-dimensional spacetime
with $t,~v^{\alpha}$~$(\alpha=3, 4, 5)$\,, while $ds^2_{\rm i}$ denotes the
metric of the internal space.
We can obtain the compactifications of the solutions depending
on the internal space.
The internal space is described by the coordinates $v^i~(i=1,2)$, and
the spatial part of our Universe $\delta_{\alpha\beta}$ is three-dimensional
with $v^\alpha~(\alpha=3, 4, 5)$\,.
Then $\delta_{\alpha\beta}$ and $\delta_{ij}$ are
the three- and two-dimensional Euclidean metrics, respectively.
Now we derive the lower-dimensional effective theory by
compactifying the extra directions.
In order to find a realistic universe, we compactify the
$d$-dimensional space to be a $d$-dimensional torus,
where $d$ is the compactified dimensions for the direction
of internal space. The remaining noncompact space
is the external space.
The range of $d$ is given by $0\le d \le 1$\,,
because the $v^1$ direction is preserved to measure the position of the
universe in the overall transverse space.
Hence the $v^2$ direction will be compactified,
where the compactified direction has to be smeared out
before the compactification.
Then the metric (\ref{nss:metric:Eq}) with $h_3=1$
is recast into the following form:
\Eq{
ds^2=ds^2_{\rm e}+ds^2_{\rm i}\,,
\label{nss:c-metric3:Eq}
}
where $ds^2_{\rm e}$ is the metric of $(6-d)$-dimensional external spacetime
and $ds^2_{\rm i}$ is the metric of compactified dimensions.
Upon setting $d=1$, the compactified metric in the Einstein frame is
\Eqr{
d\bar{s}^2_{\rm e}&=&
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-6}\bar{h}_2(v)
\right]^{-\frac{5}{3}}
\left[-d\tau^2+\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-6}
\bar{h}_2(v)\right\}^2
\left(\frac{\tau}{\tau_0}\right)^{2}\right.\nn\\
&&\left.\times
\left\{\delta_{\alpha\beta}dv^{\alpha}dv^{\beta}
+\left(dv^1\right)^2\right\}\right],
\label{nss:Emetric:Eq}
}
where $d\bar{s}^2_{\rm e}$ is the five-dimensional metric in
the Einstein frame and the constant parameters $\tau_0$ and
the cosmic time $\tau$ are defined, respectively, as
\Eq{
\frac{\tau}{\tau_0} \equiv \left(\epsilon\sqrt{2\Lambda}
t\right)^{1/6},~~~~
\tau_0 \equiv \frac{6}{\epsilon\sqrt{2\Lambda}}\,.
\label{nss:pa:Eq}
}
Since the power exponent of the scale factor is given by 1\,,
the metric of four-dimensional spacetime
in the Einstein frame implies that
the solutions gives rise to a Milne universe.
To construct a realistic cosmological model such as in the inflationary
scenario, it would be necessary to add some new ingredients in the
background. Figure \ref{fig:milne} depicts the
conformal diagrams of the five-dimensional spacetime in the limit
$\tau\rightarrow\infty$. Hence, the asymptotic
regions of the present spacetime (\ref{nss:Emetric:Eq}) resemble the
five-dimensional Milne universe.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.28, angle=-90]{milne30.eps}
\hskip 0.5cm
\includegraphics[keepaspectratio, scale=0.28, angle=-90]{AdS2.eps}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
(a)Conformal diagrams of the Milne universe are depicted.
The solid line denotes the trapping horizon $\bar{r}_{\rm T}\equiv
da(\tau)/d\tau$, where $a=(\tau/\tau_0)$ \cite{Maeda:2010ja}.
One can recognize that the asymptotic region of
the spacetime in (\ref{nss:Emetric:Eq}) corresponding to
$\bar{r}\rightarrow\infty$ approximates the five-dimensional
Milne universe, where $\bar{r}^2=\delta_{\alpha\beta}v^{\alpha}
v^{\beta}+\left(v^1\right)^2$. (b) The geometry of the 0-brane system
(\ref{nss:solution3:Eq}) in the limit
$r\rightarrow 0$ is depicted. The domain corresponding to
$r\rightarrow 0$ with finite $t$ describes warped
AdS${}_2\times{\rm S}^4$ spacetime \cite{Maeda:2010ja}.}
\label{fig:milne}
\end{center}
\end{figure}
Finally, we discuss the near-horizon geometry of the 0-brane solution.
When all of the 0-branes are located at the origin of the overall transverse
space, the solution can be expressed as
\Eq{
h_2(t, r)=\epsilon\sqrt{2\Lambda}\,t+c_4+\frac{L}{r^3}\,,~~~~~~~~
r^2 \equiv \delta_{mn}\, v^m v^n\,,
\label{nh:h2:Eq}
}
where $L$ is the total mass of 0-branes
\Eq{
L \equiv \sum_{\alpha=1}^{N'} L_{\alpha}\,.
}
In the near-horizon limit $r\rightarrow 0$\,,
the dependence on $t$ in (\ref{nh:h2:Eq}) is negligible.
The six-dimensional metric is thus reduced to the following form:
\Eqrsubl{nh:nh2-metric:Eq}{
ds^2&=&\left(\frac{L}{r^3}\right)^{-1/6}
\left[ds^2_{\rm AdS_2}+L^{2/3}d\Omega^2_{(4)}\right], \\
ds^2_{\rm AdS_2}&\equiv&-\left(\frac{L^{4/3}}{r^4}\right)^{-1}dt^2
+\frac{L^{2/3}}{r^2}\,dr^2\,,
}
where $\delta_{mn}\, dv^m dv^n=dr^2+r^2d\Omega^2_{(4)}$ has been performed.
The line elements of a two-dimensional AdS space (AdS$_2$)
and a four-sphere with the unit radius (${\rm S}^4$)
are given by $ds^2_{\rm AdS_2}$ and $d\Omega^2_{(4)}$\,, respectively.
Then the six-dimensional metric (\ref{nh:nh2-metric:Eq}) in
the near-horizon limit of the
0-brane system describes a warped product of AdS${}_2$ and ${\rm S}^4$\,.
Figure \ref{fig:milne} shows the geometry of the
AdS${}_2$ and ${\rm S}^4$\,.
\subsubsection{Cosmology in the 1-brane system}
Now we discuss the cosmological evolution for the time-dependent 1-brane
solution (\ref{nss:solution:Eq}).
We define the cosmic time $\tau$, which is given by
\Eq{
\left(\frac{\tau}{\tau_0}\right) \equiv
\left(\frac{\Lambda}{2}\,t^2\right)^{1/4}\,,~~~~
\tau_0 \equiv \frac{2\sqrt{2}}{\sqrt{\Lambda}}\,,~~~~c_1=0\,.
\label{nss:ct2:Eq}
}
The six-dimensional metric is expressed as
\Eqr{
&&\hspace{-0.8cm}ds^2=
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-4}\bar{h}_3(y, z)
\right]^{-\frac{1}{2}}
\left[-d\tau^2
+\left(\frac{\tau}{\tau_0}\right)^{-2}dy^2
\right.\nn\\
&&\left.+\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-4}
\bar{h}_3(y, z)\right\}
\left(\frac{\tau}{\tau_0}\right)^2
\delta_{ab}(\Zsp)dz^adz^b\right],
\label{nss:metric4:Eq}
}
where $\bar{h}_3(y, z)$ is defined by
\Eq{
\bar{h}_3(y, z) \equiv -\frac{\Lambda}{2} y^2+c_2y
+c_3+\sum_{l=1}^N\frac{M_l}{|z^a-z^a_l|^{2-d_{s}}}\,.
}
Here $d_{s}$ is the number of smeared dimensions and
should satisfy $0\le d_{s}\le 3$\,.
In order to fix the location of our Universe in the transverse space,
let us assume that at least one direction of $z^a~(a=1,\ldots,4)$ is
not smeared.
Now we apply the 1-brane solution to the lower-dimensional effective theory.
Let us consider a compactification and smearing of the extra
directions of the 1-brane solution.
First of all, our Universe is described by the solutions with the
six-dimensional coordinates $t,~y,~z^a~(a=1,\ldots,4)$\,.
The time direction is written by $t$.
Our choice is to take the three-dimensional
from the overall transverse space with $z^a$.
The four-dimensional universe is spanned by $t$, $z^2$, $z^3$, and $z^4$\,,
for instance. The $z^1$ direction is preserved to measure the position
of our Universe in the overall transverse space of the 1-brane.
Since the metric depends on $z^a$ explicitly, we have to smear
out ${z}^2$\,, ${z}^3$\,, and ${z}^4$ so as to define our Universe.
Then the number of the smeared directions
$d_{s}$ should satisfy the condition $d_{s}=3$\,.
It is necessary to take that $c_2=0$ and $\Lambda=0$ in (\ref{nss:metric:Eq})
to compactify the $y$ direction.
We compactify the $y$ direction to fit our Universe, where $y$ denotes the
compactified dimensions with respect to the world volume of the 1-brane.
The metric (\ref{nss:metric:Eq}) with $h_2=1$ is then described by
(\ref{nss:c-metric3:Eq}).
In terms of the conformal transformation
\Eq{
ds^2_{\rm e}=h_3^{1/6}d\bar{s}^2_{\rm e}\,,
}
we can rewrite the $(6-d)$-dimensional metric in the Einstein frame.
If we set $d=1$, the five-dimensional metric in the Einstein frame is
\Eqr{
d\bar{s}^2_{\rm e}&=&-h_3^{-2/3}(t, z)dt^2
+h_3^{1/3}(t, z)\delta_{ab}(\Zsp)dz^adz^b\,,
\label{nss:metric5:Eq}
}
where $d\bar{s}^2_{\rm e}$ is the metric of five-dimensional external
spacetime in the Einstein frame.
For $h_3=c_1t+\bar{h}_3(z)$, the metric (\ref{nss:metric5:Eq})
is thus rewritten as
\Eqr{
d\bar{s}^2_{\rm e}&=&-
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-3/2}\bar{h}_3(z)
\right]^{-2/3}d\tau^2\nn\\
&&+\left[1+\left(\frac{\tau}{\tau_0}\right)^{-3/2}\bar{h}_3(z)
\right]^{1/3}\left(\frac{\tau}{\tau_0}\right)^{1/2}
\left[\delta_{\alpha\beta}dz^{\alpha}dz^{\beta}
+\left(dz^1\right)^2\right],
\label{nss:metric6:Eq}
}
where the spatial part of our Universe $\delta_{\alpha\beta}$
is three-dimensional with $z^\alpha~(\alpha=2, 3, 4)$\,, and
the constant parameters $\tau_0$ and
the cosmic time $\tau$ are defined, respectively, as
\Eq{
\frac{\tau}{\tau_0} \equiv \left(c_1t\right)^{2/3},~~~~
\tau_0 \equiv \frac{3}{2c_1}\,.
\label{nss:pa2:Eq}
}
Unfortunately, the power exponent of the four-dimensional
universe in the Einstein frame becomes 1/4.
Hence, we have to conclude that, in order to obtain a realistic expansion
of the universe in this type of models, one has to include additional
fields on the background.
Let us finally consider the case of the near-horizon limit that
the spacetime metric and the functions $h_2$ and
$h_3$ satisfy (\ref{nss:solution:Eq}).
If we consider the case where $N$ 1-branes are located at the origin of
the Z space, we have
\Eq{
h_3(t, r)=\frac{\Lambda}{2} \left(t^2-y^2\right)+c_1t+c_2y
+c_3+\frac{M}{r^2}\,,~~~~~~~~
r^2 \equiv \delta_{ab}\, z^a z^b\,,
\label{nh:h3:Eq}
}
where $M$ is the total mass of 1-branes
\Eq{
M \equiv \sum_{l=1}^{N} M_{l}\,.
}
Since the dependence on $t$ and $y$ in (\ref{nh:h3:Eq}) is negligible
in the near-horizon limit $r\rightarrow 0$\,,
the six-dimensional metric is reduced to the following form:
\Eq{
ds^2=\left(\frac{M}{r^2}\right)^{-1/2}
\left[-dt^2+dy^2+M\left\{\frac{dr^2}{r^2}
+d\Omega^2_{(3)}\right\}\right],
}
where $\delta_{ab}\, dz^a dz^b=dr^2+r^2d\Omega^2_{(3)}$ has been used.
The line elements of a three-dimensional space (M${}_3$)
and a three-sphere
are given by $ds^2_{\rm M_3}$, $d\Omega^2_{(3)}$\,, respectively.
Thus we see that the near-horizon limit of the 1-brane system is a
warped product of M${}_3$ with a certain internal 3-space with a circle.
\subsection{Collision of the 0-brane in Nishino-Salam-Sezgin
gauged supergravity}
\label{sec:ns-c}
We next study the behavior of the time-dependent 0-brane
solution (\ref{nss:solution3:Eq})\,.
Substituting (\ref{nss:solution3:Eq}) into the metric
(\ref{nss:metric:Eq})\,, the six-dimensional metric is expressed as
\Eq{
ds^2=-\left[\epsilon\sqrt{2\Lambda}\,t+c_4+\bar{h}_2(v)\right]^{-3/2}
dt^2+\left[\epsilon\sqrt{2\Lambda}\,t+c_4+\bar{h}_2(v)\right]^{1/2}
w_{mn}dv^mdv^n\,,
\label{nss:surface:Eq}
}
where $w_{mn}$ is given by (\ref{nss:lapW:Eq}), and
the function $\bar{h}_2(v)$ is defined by (\ref{nss:bh:Eq}).
Since the time dependence appears through the function $h_2$,
the next task is to study the time evolution of the solutions carefully.
Hereafter we will consider it by focusing upon collision of
0-branes. We also discuss smearing out some of the directions
in the transverse space to decrease the number of transverse dimensions
to 0-brane effectively.
Now we consider the case that the number of the smeared direction
is given by $d_{s}$\,. Then the function $\bar{h}_2(v)$
can be expressed as
\Eq{
\bar{h}_2(v) \equiv \sum_{\alpha}\frac{L_{\alpha}}
{|v^m-v^m_{\alpha}|^{3-d_{s}}}\,,
\label{nss:sme:Eq}
}
where $d_{s}$ is the number of smeared dimensions and
should satisfy $0\le d_{s}\le 4$\,, and we assume that
at least one direction of $v^m~(m=1,\ldots,5)$ is not smeared
in order to fix the location of our universe in the transverse space.
In the following,
we will use the function (\ref{nss:sme:Eq}).
We will discuss the asymptotic behavior of the time-dependent solutions.
In the limit of $v^m\rightarrow \,v^m_{\alpha}$,
the time dependence in the function $h_2$ can be ignored because
the harmonic function $\bar{h}_2(v)$ dominates near a position of 0-brane.
On the other hand,
function $\bar{h}_2(v)$ vanishes in the limit of $v^m\rightarrow \infty$.
Then the system becomes static near 0-brane
while $h_2$ depends only on time $t$ in the far region from 0-branes.
Thus the six-dimensional metric in the limit of $v^m\rightarrow \infty$\,,
is rewritten by
\Eqr{
ds^2=-\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{-3/2}
dt^2+\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{1/2}
w_{mn}dv^mdv^n\,.
\label{nss:surface2:Eq}
}
The metric has singularity at $h_2=0$\,. Then the spacetime is regular
if it is restricted inside the domain specified by the conditions,
\Eq{
h_2(t, v)=\epsilon\sqrt{2\Lambda}\,t+c_4+\bar{h}_2(v)>0\,,
}
where the function $\bar{h}_2(v)$ is defined in (\ref{nss:sme:Eq}).
Since the spacetime evolves into a curvature singularity,
the six-dimensional spacetime cannot be extended beyond this region.
The regular spacetime with 0-branes ends up with the singularities.
The evolution of the spacetime highly depends on the signature of
$\tilde{\Lambda}\left(\equiv \epsilon\sqrt{2\Lambda}\right)$\,.
The system with $\tilde{\Lambda}>0$ has the time
reversal one of $\tilde{\Lambda}<0$.
Now we will discuss the case with $\tilde{\Lambda}<0$.
For $t<0$, the spacetime is not singular because
the function $h_2$ is positive everywhere.
In the limit of $t\rightarrow -{\infty}$, the solution is approximately
given by a time-dependent uniform spacetime apart
from a position of 0-branes. In the vicinity of branes, the geometry
takes a cylindrical form of infinite throat.
We study the time evolution for $t>0$ and $c_4=0$\,.
At $t=0$\,, the spacetime is regular everywhere and
has a cylindrical topology near each 0-brane.
As time slightly evolves, a curvature singularity
appears as $|v^m-v^m_{\alpha}|\rightarrow\infty$\,.
The singular hypersurface cuts off more and more of the
space as time increases further.
When time continues to evolve, the singular hypersurface eventually splits
and surrounds each of the 0-brane throats individually.
Hence the spatial surface is composed of each isolated throats.
For $t<0$, the time evolution of the six-dimensional spacetime
is the time reversal of $t>0$.
Since the metric (\ref{nss:surface2:Eq}) in the regular domain implies that
the overall transverse space tends to expand asymptotically like $t^{1/4}$\,,
for any values of fixed $v^m$, the solutions describe static
0-branes near the positions of the branes.
In the far region as $|v^m-v^m_{\alpha}|
\rightarrow \infty$\,, the solutions approach
FRW universes with the power law expansion $t^{1/4}$\,. The emergence of
FRW universes is an important feature of the time-dependent
0-brane solutions.
We will discuss whether two 0-branes can collide or not.
We put the two 0-branes at $\vect{v}_1=(0, 0, \ldots, 0)$
and $\vect{v}_2=(\xi, 0, \ldots, 0)$\,,
where $\xi$ is a constant.
If we introduce the following quantity
\Eq{
\tilde{v}=\sqrt{\left(v^2\right)^2+\left(v^3\right)^2+\cdots
+\left(v^{5-d_{s}}\right)^2}\,,
}
the proper distance at $\tilde{v}=0$ between the two 0-branes
is given by
\Eqr{
d(t)&=&\int^{\xi}_0 dv^1
\left(\epsilon\sqrt{2\Lambda}\,t+c_4+\frac{L_1}{|v^1|^{3-d_{s}}}
+\frac{L_2}{|v^1-\xi|^{3-d_{s}}}\right)^{1/4}\,,
\label{nss:distance:Eq}
}
where $L_1$ and $L_2$ are the charges of the 0-brane.
For $\epsilon=-1$, this is a monotonically decreasing function of $t$.
The behavior of the proper length
is different depending on the number of the smeared directions $d_{s}$\,.
We will discuss it for each of the values of $d_{s}$ below.
First we consider the case with $d_s \le 3$.
The proper length is plotted in Fig.~\ref{fig:nss0} for the cases with
$d_s=0$ and $d_s=2$. Since both cases show that a singularity appears
before the proper distance becomes zero, the singularity between
two 0-branes appears before collision. The two 0-branes approach
very slowly, and then the singular hypersurface suddenly appears at a
finite proper distance. The spacetime finally splits into two isolated
0-brane throats. Therefore one cannot see collision of the 0-branes in these
examples. For the other case with $d_s=1$, the result is the same.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ns3.eps}
\put(-175,120){$d(t)$}
\put(10,10){$t$}
\hskip 2.0cm
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ns4.eps}
\put(-175,120){$d(t)$}
\put(10,10){$t$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The time evolution of the proper distance between two dynamical 0-branes
for $d_s=0$ (a) and $d_s=2$ (b) in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity. For both cases,
the two 0-brane charges are identical,
$L_1=L_2=1$ and the parameters are taken as
$c_4=0$\,, $\Lambda=0.5$\,, $\epsilon=-1$\,, and $\xi=1$.
The result is also the same and a singularity develops before
collision of 0-branes.}
\label{fig:nss0}
\end{center}
\end{figure}
Next we consider the case with $d_{s}=4$\,, and
assume that the $v^m$ directions apart from $v^1$ are smeared.
Since the function $\bar{h}_2$ is linear in $v$\,,
the behavior of the proper distance is different from the previous case.
The six-dimensional metric is now given by (\ref{nss:surface:Eq})\,.
By choosing $v=v^1$\,, the harmonic function $\bar{h}_2$ is written by
\Eq{
\bar{h}_2(v)=\sum_{\alpha=1}^{N'}L_{\alpha}
|\,v-\,v_{\alpha}|\,.
\label{nss:bh2:Eq}
}
We discuss the time-dependent solutions in the case that one 0-brane
charge $L_1$ is located at $v=0$ and the other $L_2$ at $v=\xi$\,.
The proper length between the two 0-branes is given by
\Eqr{
d(t)&=&\int^{\xi}_0 dv \left[\epsilon\sqrt{2\Lambda}\,t+c_4
+\left(L_1|v|+L_2|v-\xi|\right)\right]^{1/4}\,.
\label{nss:length:Eq}
}
For $\epsilon=-1$, the proper distance decreases with time.
If we set $L_1\ne L_2$\,, a singularity appears again at
a certain finite time $t=t_{\rm s}$\,,
while the proper distance is still finite, where $t_{\rm s}$ is defined as
\Eq{
t_{\rm s}\equiv\frac{c_4+L_1|v|+L_2|v-\xi|}{\sqrt{2\Lambda}}\,.
}
This is the same result as the case with $d_{s}\leq 3$\,.
On the other hand, two 0-branes have the same brane charge $L_1=L_2=L$\,,
the proper distance vanishes at a certain finite time $t=t_{\rm c}$\,,
where $t_{\rm c}$ is defined by
\Eq{
t_{\rm c} \equiv \frac{c_4+L\xi}{\sqrt{2\Lambda}}\,.
}
Hence two 0-branes can collide completely.
In terms of $t_{\rm c}$\,, the proper length is expressed as
\Eqr{
d(t)=L\left[-\sqrt{2\Lambda}(t-t_{\rm c})\right]^{1/4}\,.
}
If we choose the values as
$c_4=0$\,, $\Lambda=0.5$\,, $\xi=1$ and
$\epsilon=-1$\,, the proper distance $d(t)$ is plotted in
Fig.~\ref{fig:nss}
for the two cases (a) the same 0-brane charges $L_1=L_2=1$ and (b) different
charges $L_1=2$, $L_2=1$\,.
In the case (a) the two 0-branes can collide completely.
However, in the case (b) a singularity appears before collision,
as we have already discussed analytically.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ns1.eps}
\put(-175,115){$d(t)$}
\put(10,10){$t$}
\hskip 2.0cm
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ns2.eps}
\put(-175,115){$d(t)$}
\put(10,10){$t$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The time evolution of the proper distance between two dynamical 0-branes
for $L_1=L_2=1$ (a) and $L_1=2$, $L_2=1$ (b) in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity. We fix $d_s=4$\,, $c_4=0$\,,
$\Lambda=0.5$\,,
$\epsilon=-1$\,, and $\xi=1$.
The proper distance rapidly vanishes near two
0-branes collide for the case of $L_1=L_2=1$.
While for the case of $L_1=2$, $L_2=1$, it
is still finite when a curvature singularity appears.}
\label{fig:nss}
\end{center}
\end{figure}
\subsection{Collision of the 1-brane in Nishino-Salam-Sezgin
gauged supergravity}
Now we apply our time-dependent solutions to
a collision of 1-brane systems.
In the case of $h_2=1$, the function $h_3$ is assumed to be
\Eq{
h_3(t, y, z)=\frac{\Lambda}{2}\left(t^2-y^2\right)+c_3+\tilde{h}(z)\,,
\label{nsc:h:Eq}
}
where $c_3$ is a constant parameter, we choose $c_1=c_2=0$, and
the harmonic function $\tilde{h}$ is expressed as
\Eqrsubl{nsc:h2:Eq}{
\tilde{h}(z)&=&\sum_{l=1}^{N}\frac{M_l}{|z^a-z^a_l|^{2-d_s}}\,,
~~~~~~{\rm for}~~d_s\neq 2\,,
\label{nsc:h2-1:Eq}
\\
\tilde{h}(z)&=&\sum_{l=1}^{N}M_l\, \ln |z^a-z^a_l|\,,
~~~~{\rm for}~~ d_s=2\,.
\label{nsc:h2-2:Eq}
}
Here $M_l$ are charges of 1-branes located at $z^a=z^a_l$ and
\Eq{
|z^a-z^a_l|
=\sqrt{\left(z^1-z^1_l\right)^2+\left(z^2-z^2_l\right)^2+\cdots+
\left(z^{4-d_s}-z^{4-d_s}_l\right)^2}\,,
}
because the harmonic function $\tilde{h}$ is defined on the
$(4-d_s)$-dimensional Euclidean subspace in $\Zsp$.
The six-dimensional metric, scalar field,
and gauge field of the solution are
given by Eqs.~(\ref{nss:metric:Eq}), and
(\ref{nss:ansatz:Eq}), respectively.
We see that $d_s=2$ case is critical.
For $d_s=3$, the function $\tilde{h}$ is written by
the sum of linear functions of $z$.
The possibility of 1-brane collisions depends on the difference
in the transverse dimensions because the behaviors of the gravitational
field in the transverse space depends on the number of the transverse
dimensions.
Although the six-dimensional metric (\ref{nss:metric:Eq}) is regular
if and only if $h_3>0$, the spacetime
shows curvature singularities at $h_3=0$.
Hence the regular six-dimensional spacetime
is restricted to the region of $h_3>0$, which is
bounded by curvature singularities.
Let us study the time evolution for time-dependent 1-brane
solution (\ref{nss:solution:Eq}).
We perform the following coordinate transformation:
\begin{eqnarray}
&&t=\sqrt{\frac{2}{\Lambda}}\,
\tilde{t} \cosh \tilde{y}~~,~~~y=\sqrt{\frac{2}{\Lambda}}\,
\tilde{t} \sinh \tilde{y}
\,.
\end{eqnarray}
If we choose $c_1=c_2=0$, we find
\begin{eqnarray}
\hspace{-0.5cm}ds^2&=&\frac{2}{\Lambda}\,
\left[1+\tilde{t}^{-2}\tilde{h}(\tilde{z})\right]^{-1/2}\tilde{t}^{-1}
\left[-d\tilde{t}^2+\tilde{t}^2\left\{d\tilde{y}^2
+\left(1+\tilde{t}^{-2}\tilde{h}(\tilde{z})\right)
\delta_{ab}d\tilde{z}^ad\tilde{z}^b\right\}
\right]
\nonumber \\
\hspace{-0.5cm}&=&\frac{2}{\Lambda}\,
\left[1+16\bar{t}^{-4}\tilde{h}(\tilde{z})\right]^{-1/2}
\left[-d\bar{t}^2+\frac{1}{4}\,\bar{t}^2
\left\{d\tilde{y}^2+
\left(1+16\bar{t}^{-4}\tilde{h}(\tilde{z})\right)
\delta_{ab}d\tilde{z}^ad\tilde{z}^b\right\}\right],
\label{nsc:sufrace:Eq_m}
\end{eqnarray}
where $\tilde{h}(\tilde{z})$, $\bar{t}$, and $\tilde{z}^a$ are defined,
respectively, by
\begin{eqnarray}
\tilde{h}(\tilde{z})=c_3
+\frac{\Lambda}{2}\sum_{l=1}^N\frac{M_l}{|\tilde{z}^a-\tilde{z}^a_l|^2}\,,
~~~~~~\bar{t}=2\tilde{t}^{1/2},~~~~~~~
\tilde{z}^a=\sqrt{\frac{\Lambda}{2}}~z^a\,.
\end{eqnarray}
Here, $\tilde{t}$ obeys
$\Lambda(t^2-y^2)/2=\tilde{t}^2$.
The six-dimensional metric (\ref{nsc:sufrace:Eq_m}) represents
a homogeneous and isotropic spacetime whose scale factor evolves
as the cosmic time
$\bar{t}$, which is described as the Milne universe.
Hence, we can consider
that the present solution with $\Lambda>0$ gives a system of
1-branes in the Milne universe. The existence of the expanding Milne
universe is guaranteed by the scalar field with the exponential potential
in the six-dimensional action (\ref{nss:action:Eq}).
Now let us consider the collision of 1-branes.
The solution (\ref{nss:metric:Eq}) without 0-branes can be written
in the form
\Eq{
ds^2=\left[\frac{\Lambda}{2}\left(t^2-y^2\right)
+c_3+\tilde{h}(z)\right]^{-\frac{1}{2}}
\left(-dt^2+dx^2\right)+\left[\frac{\Lambda}{2}\left(t^2-y^2\right)
+c_3+\tilde{h}(z)\right]^{\frac{1}{2}}
u_{ab}dz^adz^b\,,
\label{nsc:surface:Eq}
}
where we choose $c_1=c_2=0$, $u_{ab}$ denotes the four-dimensional metric,
and the function $\tilde{h}(z)$ is given by (\ref{nsc:h2:Eq}).
The behavior of the harmonic function $\tilde{h}(z)$ is divided into two
classes depending on the dimensions of the 1-brane, that is,
$d_s\ne 2$ and $d_s=2$, which we will study below separately.
For $d_s=2$, the harmonic function $\tilde{h}(z)$ diverges
both at infinity and near 1-branes.
In particular, there is no regular spacetime region near 1-branes,
because $\tilde{h}(z)\rightarrow -\infty$\,.
Hence, such a 1-brane solution is not physically relevant.
Since the harmonic function $\tilde{h}(z)$ becomes dominant
in the limit of $z^a\rightarrow z^a_l$ (near 1-branes),
we find a static structure of the 1-brane system.
In the far region from 1-branes, that is,
in the limit of $|z^a-z^a_l|\rightarrow \infty$,
the function $h_3$ depends only on time $t$, because
$\tilde{h}(z)$ vanishes. The metric is thus written by
\Eqr{
ds^2=\left[\frac{\Lambda}{2}\left(t^2-y^2\right)+c_3\right]^{-\frac{1}{2}}
\left(-dt^2+dx^2\right)
+\left[\frac{\Lambda}{2}\left(t^2-y^2\right)+c_3\right]^{\frac{1}{2}}
u_{ab}dz^adz^b\,.
\label{nsc:surface2:Eq}
}
In the following, we will analyze one concrete example, in which
two 1-branes are located at $\,\vect{z}_1=(0, 0,\ldots,0)$ and
$\,\vect{z}_2=(z_0, 0,\ldots,0)$\, in order to study in more detail.
Since the metric function is singular at $h_3=0$\,,
the regular spacetime exists inside the domain restricted by
\Eq{
h_3(t, z) = \frac{\Lambda}{2}\left(t^2-y^2\right)+c_3+\tilde{h}(z)>0\,,
}
where the function $\tilde{h}(z)$ is given by (\ref{nsc:h2:Eq}).
The six-dimensional
spacetime cannot be extended beyond this region, because
not only does the dilaton $\phi$ diverge but also the spacetime
evolves into a curvature singularity.
The regular spacetime with two 1-branes ends on these singularities.
The time dependence appears in the form of $\frac{\Lambda}{2} t^2$.
For $t^2>y^2$ and $c_3=0$, the function $h$ is positive everywhere and
the six-dimensional spacetime is not singular. It is asymptotically
a time-dependent uniform spacetime except for
near branes in the limit of $z^a\rightarrow z^a_l$\,,
where the background geometry becomes the cylindrical
forms of infinite throats
When $t\le 0$, the spatial metric is initially ($t\rightarrow-\infty$)
regular apart from $y\rightarrow\pm\infty$\,, and
the spacetime has a cylindrical topology near each 1-brane.
As $t$ evolves slightly, a curvature singularity appears at
$y\rightarrow\pm\infty$ and from a
far region ($|z^1|\rightarrow\infty$).
As $t$ evolves further, the singularity cuts off the
space. As the time continues to increase, the singular hypersurface
eventually splits and surrounds each of the 1-brane throats individually.
Then the spatial surface is composed of two isolated throats.
The six-dimensional metric (\ref{nsc:surface2:Eq}) implies that the
transverse dimensions expand asymptotically as $t^{1/2}$
for fixed spatial coordinates $y$ and $z^a$\,.
However, this is observer dependent, because it becomes static near branes,
and the spacetime approaches a Friedmann-Robertson-Walker
universe in the far region ($|z^1|\rightarrow \infty$),
which expands in the background isotropically.
If we define
\Eq{
\bar{z}=\sqrt{\left(z^2\right)^2+\left(z^3\right)^2+\cdots
+\left(z^{4-d_s}\right)^2}\,,
}
the proper length at $\bar{z}=0$
between two 1-branes is written by
\begin{eqnarray}
d(t, y)&=&\int_{0}^{z_0} dz^1
\left[\frac{\Lambda}{2}\left(t^2-y^2\right)+c_3+\frac{M_1}{|z^1|^{2-d_s}}
+\frac{M_2}{|z^1-z_0|^{2-d_s}}\right]^{\frac{1}{4}}
\,.
\label{nsc:distance:Eq}
\end{eqnarray}
This is a monotonically decreasing function of time for $t\le 0$.
In Fig.~\ref{fig:ns}, we show $d(t, y)$
for the case of the 1-brane system.
We set $\Lambda=2$, $z_0=1$, $c_3=0$, and $M_1=M_2=1$.
All of the six-dimensional space is initially ($t=-\infty$) regular
except at $|y|\rightarrow \infty$ and $|z^1-z_0|\rightarrow\infty$.
Although the spacetime becomes asymptotically
time dependent and has the cylindrical form of an infinite
throat near the 1-brane, the singularity appears
from a far region ($|z^1-z_0|\rightarrow\infty$)
and $|y|\rightarrow \infty$\,.
As time increases ($t<0$),
the singularity erodes the region with the large $|y|$ region.
The region of transverse space is also invaded in time.
As a result, only the region of small $|y|$ and near 1-branes remains regular.
When we study the evolution on the $y$ and $z^a$ plane, the singularity
appears at infinity $|z^1|\rightarrow\infty$\,,
$|y|\rightarrow \infty$\,, and comes to the region of two 1-branes.
A singular hypersurface eventually surrounds each 1-brane
individually, and then the regular regions near 1-branes split
into two isolated throats.
For the period of $t>0$, we find the time-reversed
behavior of the case of $t<0$.
Figures \ref{fig:ns} and \ref{fig:ns1} show that this singularity appears
before the distance $d$ vanishes.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss1.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}
\hskip 2.cm
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss3.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The time evolution of the proper distance between two dynamical
1-branes for (a) $d_s=3$ and (b) $d_s=1$ in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity.
We fix $c_3=0$\,, $M_1=M_2=1$\,, $z_0=1$\,, and $\Lambda=2$.
The proper distance rapidly vanishes near where two
1-branes collide for the case of $d_s=3$,
while for the case of $d_s=1$, it
is still finite when a curvature singularity appears.}
\label{fig:ns}
\end{center}
\end{figure}
Then a singularity between two branes forms before their collision
except for $d_s=3$.
Two 1-branes approach very slowly, a singularity suddenly appears
at a finite distance, and the six-dimensional spacetime splits
into two isolated 1-brane throats.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss8.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}
\hskip 2.cm
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss7.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The proper distance between two dynamical 1-branes given in
(\ref{nsc:distance:Eq}) is depicted for (a) $M_1=10$, $M_2=1$
and (b) $M_1=2$\,, $M_2=1$ in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity.
We fix $c_3=0$\,, $d_s=0$\,, $z_0=1$, and $\Lambda=2$.
In both cases, a singularity appears at $t=t_{\rm s}<0$ when the distance is
still finite.}
\label{fig:ns1}
\end{center}
\end{figure}
On the other hand, we can discuss a brane collision for $d_s=3$ and $t<0$.
If $M_1\ne M_2$, a singularity appears at $t=t_{\rm s}<0$ when
the distance is still finite (see Fig.~\ref{fig:ns2}).
This is just the same as the case in
Sec.~\ref{sec:ns-c}. However, if $M_1=M_2=M$, the result completely changes
(Fig.~\ref{fig:ns}).
Since the distance eventually vanishes at $t=t_{\rm c}$,
two 1-branes collide with each other.
The proper length for fixed $y$ decreases as time increases from $t=-\infty$,
and it eventually vanishes at $t=t_{\rm c}$. Hence, one 1-brane approaches
the other as time evolves, causing the complete collision at $t=t_{\rm c}$.
If we fix the 1-brane charges such that $M_1=M_2=M$, the branes first
collide at larger $|y|$, and as time progresses, the subsequent collisions
occur at the smaller $|y|$.
We show $d(t, y)$ integrated numerically in
Figs.~\ref{fig:ns} $-$ \ref{fig:ns2}.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss4.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}
\hskip 2.cm
\includegraphics[keepaspectratio, scale=0.5, angle=0]{nss6.eps}
\put(-195, 80){$d(t, y)$}
\put(-105, 0){$t$}
\put(-5,40){$y$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The time evolution of the proper distance between two dynamical 1-branes
for (a) $M_1=10$, $M_2=1$ and (b) $M_1=2$\,, $M_2=1$ in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity.
We fix $c_3=0$\,, $d_s=3$\,, $z_0=1$, and $\Lambda=2$.
For $M_1\ne M_2$, a singularity appears at $t=t_{\rm s}<0$
when the distance is
still finite. Then, the solution does not describe the collision
of two 0-branes.}
\label{fig:ns2}
\end{center}
\end{figure}
We also calculate the distance $d(t, y)$ at $y=0$ and $\bar{z}=0$
between two branes before the singularity appears except for the case
of $d_s=3$ if $M_1=M_2$.
The proper length is also given by Eq. (\ref{nsc:distance:Eq}).
In the present case, $d$ is a monotonically decreasing function
of $t^2$ when $t<0$.
We show the time evolution of the distance in Fig.~\ref{fig:ns3}
for the case of $M_1=M_2$.
On the other hand, for the case of ${M}_1\ne {M}_2$,
a singularity appears, when the proper distance is still finite.
For the period of $t>0$, the behavior of six-dimensional
spacetime is the time reversal of the period of $t<0$.
We show the proper distance $d(t)$ integrated numerically
in Fig.~\ref{fig:ns3} for the cases of $d_s=3$ and $d_s\ne 3$.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.55, angle=0]{nss10.eps}
\put(-95,120){$d(t)$}
\put(10,10){$t$}
\hskip 2.0cm
\includegraphics[keepaspectratio, scale=0.55, angle=0]{nss12.eps}
\put(-95,120){$d(t)$}
\put(10,10){$t$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
(a) The proper distance between two dynamical 1-branes at $y=0$ and
$\bar{z}=0$ for the case of $d_s=0$ in the six-dimensional
Nishino-Salam-Sezgin gauged supergravity is depicted.
We fix $c_4=0$\,, $z_0=1$\,, and $\Lambda=2$\,. For $t<0\,,$
the proper length decreases as time increases.
The bold line denotes the case of $M_1=M_2=1$, while the solid
one corresponds to the $M_1=10$, $M_2=1$ case. (b) For the case of
$M_1=M_2$ in the six-dimensional Nishino-Salam-Sezgin gauged supergravity,
the time evolution of the proper distance between two dynamical 1-branes at
$y=3$ and $\tilde{z}=0$ given in (\ref{nsc:distance:Eq}) is depicted.
We fix $c_4=0$\,, $z_0=1$\,, and $\Lambda=2$. We show the lengths
for $d_s=0$ (bold line), $d_s=1$ (solid line), and $d_s=3$ (dashed line).
The proper distance rapidly vanishes near where two
1-branes collide in the case of $d_s=3$,
while in the case of $d_s\ne 3$,
is still finite when a curvature singularity appears.}
\label{fig:ns3}
\end{center}
\end{figure}
\subsection{Romans' six-dimensional gauged supergravity}
Similarly, for the six-dimensional Romans' theory \cite{Romans:1985tw},
following the discussion in Ref.~\cite{Nunez:2001pt},
the coupling
of the 3-form and of the 2-form field strengths to the dilaton are
given by $\epsilon_rc_r=-1/\sqrt{2}$ and $\epsilon_sc_s=\sqrt{2}$,
respectively:
\Eqr{
S&=&\frac{1}{2\kappa^2}\int \left[\left(R+2\e^{\phi/\sqrt{2}}
\lambda\right)\ast{\bf 1}
-\frac{1}{2}\ast d\phi \wedge d\phi\right.\nn\\
&&\left.\hspace{-0.5cm}
-\frac{1}{2\cdot 2!}
\e^{-\phi/\sqrt{2}}\ast F_{(2)}\wedge F_{(2)}
-\frac{1}{2\cdot 3!}
\e^{\sqrt{2}\phi}\ast F_{(3)}\wedge F_{(3)} \right],
\label{ro:action:Eq}
}
where $R$ denotes the Ricci scalar constructed from the six-dimensional
metric $g_{MN}$\,, $\kappa^2$ is the six-dimensional gravitational constant,
$\ast$ is the Hodge operator in the six-dimensional spacetime,
$\phi$ denotes the scalar field, $\lambda>0$ is cosmological
constants, and $F_{\left(3\right)}$ and $F_{\left(2\right)}$
are 3-form and 2-form field strengths, respectively.
This has a negative scalar potential.
In terms of Eq.~(\ref{pl:c:Eq}), Romans' model
is given by choosing $\Lambda_r=-\lambda$,
$\Lambda_s=0$, $N_r=2$, and $N_s=4$.
From the six-dimensional action (\ref{ro:action:Eq}),
we find the field equations:
\Eqrsubl{ro:equations:Eq}{
&&R_{MN}=-\frac{1}{2}\e^{\phi/\sqrt{2}}\lambda g_{MN}
+\frac{1}{2}\pd_M\phi \pd_N \phi
+\frac{\e^{-\phi/\sqrt{2}}}{2\cdot 2!}
\left(2F_{MA}{F_N}^A
-\frac{1}{4} g_{MN} F^2_{\left(2\right)}\right)\nn\\
&&~~~~~~~~~+\frac{\e^{\sqrt{2}\phi}}{2\cdot 3!}
\left(3F_{MAB} {F_N}^{AB}
-\frac{1}{2} g_{MN} F_{\left(3\right)}^2\right)\,,
\label{ro:Einstein:Eq}\\
&&\lap\phi+
\frac{\sqrt{2}}{4\cdot 2!}\,\e^{-\phi/\sqrt{2}}\,F^2_{\left(2\right)}
-\frac{\sqrt{2}}{2\cdot 3!}\,\e^{\sqrt{2}\phi}\,F^2_{\left(3\right)}
+\sqrt{2}\,\e^{\phi/\sqrt{2}}\,\lambda=0\,,
\label{ro:scalar:Eq}\\
&&d\left[\e^{-\phi/\sqrt{2}}\ast F_{\left(2\right)}\right]=0\,,
\label{ro:gauge-r:Eq}\\
&&d\left[\e^{\sqrt{2}\phi}\ast F_{\left(3\right)}\right]=0\,,
\label{ro:gauge-s:Eq}
}
where $\lap$ denotes the Laplace operator with respect to the six-dimensional
metric $g_{MN}$\,.
We assume the six-dimensional metric of the form (\ref{nss:metric:Eq}).
The scalar field and the gauge field strengths are assumed to be
\Eqrsubl{ro:ansatz:Eq}{
\e^{\phi}&=&h_2^{-\sqrt{2}/2}\,h_3^{\sqrt{2}/2},
\label{ro:phi:Eq}\\
F_{\left(2\right)}&=&d\left[\sqrt{2}\,h^{-1}_2(t, y, z)\right]\wedge dt,
\label{ro:fr:Eq}\\
F_{\left(3\right)}&=&d\left[h^{-1}_3(t, y, z)\right]\wedge dt\wedge dy\,.
\label{ro:fs:Eq}
}
The Einstein equations (\ref{ro:Einstein:Eq}) then reduce to
\Eqrsubl{ro:cEinstein:Eq}{
&&\frac{5}{4}h_2^{-1}\pd_t^2h_2
+\frac{3}{4}h_3^{-1}\pd_t^2h_3+
\frac{1}{4}h_2^{-2}\left(3h_2^{-1}\pd_y^2h_2+h_3^{-1}\pd_y^2h_3\right)
+\frac{1}{4}h_2^{-2}h_3^{-1}
\left(3h_2^{-1}\lap_{\Zsp}h_2+h_3^{-1}\lap_{\Zsp}h_3\right)\nn\\
&&~~~+\frac{1}{2}h_2^{-2}\lambda
+\frac{1}{4}\left(\pd_t\ln h_2\right)^2
+\frac{7}{4}\pd_t\ln h_2\pd_t\ln h_3
+\frac{3}{4}h_2^{-2}\pd_y\ln h_2\pd_y\ln h_3=0,
\label{ro:cEinstein-tt:Eq}\\
&&h_2^{-1}\pd_t\pd_y h_2+h_3^{-1}\pd_t\pd_y h_3
+\pd_t\ln h_3\pd_y\ln h_2=0\,,
\label{ro:cEinstein-ty:Eq}\\
&&2h_2^{-1}\pd_t\pd_a h_2+h_3^{-1}\pd_t\pd_a h_3=0\,,
\label{ro:cEinstein-ta:Eq}\\
&&\frac{1}{4}h_2^2\left(h_2^{-1}\pd_t^2h_2-h_3^{-1}\pd_t^2h_3\right)
-\frac{1}{4}\left(h_2^{-1}\pd_y^2h_2+3h_3^{-1}\pd_y^2h_3\right)
-\frac{1}{4}h_3^{-1}
\left(h_2^{-1}\lap_{\Zsp}h_2-h_3^{-1}\lap_{\Zsp}h_3\right)\nn\\
&&~~~+\frac{1}{2}\lambda+\frac{1}{4}\left(\pd_th_2\right)^2
-\frac{1}{4}h_2^2\pd_t\ln h_2\pd_t\ln h_3
-\frac{5}{4}\pd_y\ln h_2\pd_y\ln h_3=0\,,
\label{ro:cEinstein-yy:Eq}\\
&&h_3^{-1}\pd_y\pd_a h_3+2\pd_y\ln h_2\pd_a\ln h_2=0\,,
\label{ro:cEinstein-ya:Eq}\\
&&R_{ab}(\Zsp)+\frac{1}{4}h_2^2h_3u_{ab}
\left(h_2^{-1}\pd_t^2 h_2+h_3^{-1}\pd_t^2 h_3\right)
-\frac{1}{4}h_3u_{ab}
\left(h_2^{-1}\pd_y^2 h_2+h_3^{-1}\pd_y^2 h_3\right)\nn\\
&&~~~-\frac{1}{4}u_{ab}
\left(h_2^{-1}\lap_{\Zsp}h_2+h_3^{-1}\lap_{\Zsp}h_3\right)
+\frac{1}{4}h_2^2h_3u_{ab}\left[\left(\pd_t\ln h_2\right)^2
+3\pd_t\ln h_2\pd_t\ln h_3\right]\nn\\
&&~~~-\frac{1}{4}h_3u_{ab}\pd_y\ln h_2\pd_y\ln h_3
+\frac{1}{2}u_{ab}h_3\,\lambda=0,
\label{ro:cEinstein-ab:Eq}
}
where $\triangle_{\Zsp}$ denotes the Laplace operator on Z space and
$R_{ab}(\Zsp)$ is the Ricci tensor
constructed from the metric $u_{ab}(\Zsp)$\,.
We next consider the gauge field.
Under the ansatz (\ref{ro:ansatz:Eq}),
the Bianchi identity is automatically satisfied. Also the equation of motion
for the gauge field becomes
\Eqrsubl{ro:gauge2:Eq}{
&&d\left[h_3^{-1}\pd_y h_2\,\Omega(\Zsp)
+\pd_a h_2\,dy\wedge\left(\ast_{\Zsp}dz^a\right)\right]=0,
\label{ro:gauge2-r:Eq}\\
&&d\left[\pd_a h_3\left(\ast_{\Zsp}dz^a\right)\right]=0,
\label{ro:gauge2-s:Eq}
}
where $\ast_{\rm Z}$ denotes the Hodge operator on Z.
Although the roles of the Bianchi identity and field
equations are interchanged, the net result is the same.
Finally, we consider the equation of motion for the scalar field.
Substituting the scalar
field and the gauge field in (\ref{ro:ansatz:Eq}) into the equation
of motion for the scalar field (\ref{ro:scalar:Eq}), we have
\Eqr{
&&h_2^2h_3\left(h_2^{-1}\pd_t^2h_2-h_3^{-1}\pd_t^2h_3\right)
+h_3\left(\pd_th_2\right)^2
-h_2\pd_th_2\pd_th_3-h_3\left(h_2^{-1}\pd_y^2h_2
-h_3^{-1}\pd_y^2h_3\right)\nn\\
&&~~~~-h_2^{-1}\pd_yh_2\pd_yh_3
-h_2^{-1}\lap_{\Zsp}h_2+h_3^{-1}\lap_{\Zsp}h_3+2h_3\lambda=0\,.
\label{ro:scalar-e:Eq}
}
Then, the functions $h_2$ and $h_3$ satisfy the equations
\Eqrsubl{ro:scalar-s:Eq}{
&&
\left(\pd_th_2\right)^2+2\lambda+h_2\pd_t^2h_2-h_2^{-1}\lap_{\rm W}h_2=0,
~~~~{\rm For}~~h_3=1\,,
\label{ro:scalar-s1:Eq}\\
&&-\pd_t^2h_3+\pd_y^2h_3+2\lambda h_3+h_3^{-1}\lap_{\Zsp}h_3=0\,,
~~~~{\rm For}~~h_2=1\,,
\label{ro:scalar-s2:Eq}
}
where Laplace operator $\lap_{\rm W}$ is defined in Eq.~(\ref{nss:lapW:Eq}).
If we set $h_2=1$, the field equations give
\Eqrsubl{ro:equation:Eq}{
&&R_{ab}(\Zsp)=0\,,
\label{ro:Ricci:Eq}\\
&&h_2=1\,,~~~~\lambda=0,~~~~\pd_t^2h_3=\pd_y^2h_3=0\,,
~~~~\triangle_{\Zsp}h_3=0\,.
\label{ro:h1:Eq}
}
Now we will focus upon a case by imposing the conditions
\Eq{
u_{ab}=\delta_{ab}\,,~~~~~~h_2=1\,,~~~~~~\lambda=0\,,
\label{ro:flat:Eq}
}
where $\delta_{ab}$ is the four-dimensional Euclidean metric.
Then, the solution for $h_3$ can be obtained explicitly as
\Eq{
h_3(t, y, z)=c_1t+c_2y+c_3+\sum_{l=1}^N\frac{M_l}{|z^a-z^a_l|^2},
\label{ro:solution-s:Eq}
}
where $c_i~(i=1,~2,~3)$ are constants\,.
One can easily get the solution
for $h_3=1$\,, $\lambda\ne 0$ and $\pd_th_2\ne 0$
if the roles of $h_2$ and $h_3$ are exchanged.
The solution of field equations is thus expressed as
\Eqrsubl{ro:solution3:Eq}{
h_2(t, v)&=&\pm\sqrt{2i\lambda}\,t+c_5+\sum_{\alpha=1}^{N'}\frac{L_{\alpha}}
{|v^m-v^m_{\alpha}|^3}\,,
\label{ro:solution3-r:Eq}\\
h_3&=&1\,,
\label{ro:solution3-s:Eq}
}
where $c_5$, $v^m_\alpha$, and $L_\alpha$ are constants, and
the five-dimensional coordinate $v^m$ defined by \eqref{nss:lapW:Eq}.
Hence, there is no cosmological 0-brane solution in terms of the
ansatz of fields (\ref{nss:metric:Eq}) and (\ref{ro:ansatz:Eq}) if
$\lambda\ne 0$.
\subsection{Romans' five-dimensional gauged supergravity}
Finally, we consider the five-dimensional Romans' theory \cite{Romans}.
The five-dimensional action is given by
\Eqr{
S&=&\frac{1}{2\kappa^2}\int \left[\left(R+2\e^{2\phi/\sqrt{6}}
\bar{\lambda}\right)\ast{\bf 1}
-\frac{1}{2}\ast d\phi \wedge d\phi\right.\nn\\
&&\left.\hspace{-0.5cm}
-\frac{1}{2\cdot 2!}
\e^{-2\phi/\sqrt{6}}\ast F_{(2)}\wedge F_{(2)}
-\frac{1}{2\cdot 2!}
\e^{4\phi/\sqrt{6}}\ast H_{(2)}\wedge H_{(2)} \right],
\label{ro5:action:Eq}
}
where the expectation values of Yang-Mills potential is assumed to vanish,
$R$ denotes the Ricci scalar constructed from the five-dimensional metric
$g_{MN}$\,,
$\kappa^2$ is the five-dimensional gravitational constant,
$\ast$ is the Hodge operator in the five-dimensional space-time,
$\phi$ denotes the scalar field, $\bar{\lambda}>0$ is cosmological
constant, $F_{\left(2\right)}$ and $H_{\left(2\right)}$
are two-form field strengths,
and the couplings of the 2-form field strengths and cosmological constant
to the dilaton are given by $\epsilon_rc_r=-2/\sqrt{6}$,
$\epsilon_sc_s=4/\sqrt{6}$, $\alpha_r=2/\sqrt{6}$\,,
in the action (\ref{pl:action:Eq}), respectively.
This has also a negative scalar potential.
In terms of Eq.~(\ref{pl:c:Eq}), Romans' five-dimensional model
is given by setting $\Lambda_r=-\bar{\lambda}$,
$\Lambda_s=0$, $N_r=2$ and $N_s=4$.
The five-dimensional action (\ref{ro5:action:Eq}),
gives the field equations:
\Eqrsubl{ro5:equations:Eq}{
&&R_{MN}=-\frac{1}{2}\e^{2\phi/\sqrt{6}}\,\bar{\lambda}\,g_{MN}
+\frac{1}{2}\pd_M\phi \pd_N \phi
+\frac{\e^{-2\phi/\sqrt{6}}}{2\cdot 2!}
\left(2F_{MA}{F_N}^A
-\frac{1}{3} g_{MN} F^2_{\left(2\right)}\right)\nn\\
&&~~~~~~~~+\frac{\e^{4\phi/\sqrt{6}}}{2\cdot 2!}
\left(2H_{MA} {H_N}^{A}
-\frac{1}{3} g_{MN} H_{\left(2\right)}^2\right),
\label{ro5:Einstein:Eq}\\
&&\lap\phi+\frac{\sqrt{6}}{6\cdot 2!}\,\e^{-2\phi/\sqrt{6}}\,
F^2_{\left(2\right)}
-\frac{\sqrt{6}}{3\cdot 2!}\,\e^{4\phi/\sqrt{6}}
H^2_{\left(2\right)}+\frac{2\sqrt{6}}{3}\,
\e^{2\phi/\sqrt{6}}\,\bar{\lambda}=0\,,
\label{ro5:scalar:Eq}\\
&&d\left[\e^{-2\phi/\sqrt{6}}\ast F_{\left(2\right)}\right]=0\,,
\label{ro5:gauge-r:Eq}\\
&&
d\left[\e^{4\phi/\sqrt{6}}\ast H_{\left(2\right)}\right]=0\,,
\label{ro5:gauge-s:Eq}
}
where $\lap$ denotes the Laplace operator with respect to the
five-dimensional metric
$g_{MN}$\,.
We assume the five-dimensional metric of the form
\Eqr{
ds^2&=&h^{2/3}_2(t, z)k_2^{1/3}(t, z)\left[-h^{-2}_2(t, z)
k^{-1}_2(t, z)dt^2+u_{ab}(\Zsp)dz^adz^b\right],
\label{ro5:metric:Eq}
}
where $u_{ab}(\Zsp)$ is the four-dimensional metric which
depends only on the four-dimensional coordinates $z^a$.
The scalar field and the gauge field strengths are assumed to be
\Eqrsubl{ro5:ansatz:Eq}{
\e^{\phi}&=&h_2^{-2/\sqrt{6}}\,k_2^{2/\sqrt{6}},
\label{ro5:phi:Eq}\\
F_{\left(2\right)}&=&d\left[\sqrt{2}\,h^{-1}_2(t, z)\right]\wedge dt,
\label{ro5:fr:Eq}\\
H_{\left(2\right)}&=&d\left[k^{-1}_2(t, z)\right]\wedge dt\,.
\label{ro5:fs:Eq}
}
Then, the field equations are reduced to
\Eqrsubl{ro5:equation:Eq}{
&&
R_{ab}(\Zsp)=0\,,
\label{ro5:eq-ab:Eq}\\
&&
h_2(t, z)=h_0(t)+\bar{h}(z)\,,~~~~~
k_2(t, z)=k_0(t)+\bar{k}(z)\,,
\label{ro5:eq-h:Eq}\\
&&
\left(\frac{dh_0}{dt}\right)^2+2\bar{\lambda}=0\,,~~~~~~
\frac{dh_0}{dt}\frac{dk_0}{dt}=0\,,~~~~~~\frac{d^2h_0}{dt^2}=0\,,~~~~~~
\frac{d^2k_0}{dt^2}=0\,,\\
&&\lap_{\Zsp}\bar{h}=0\,,~~~~~~\lap_{\Zsp}\bar{k}=0\,,
\label{ro5:eq-h2:Eq}
}
where $\triangle_{\Zsp}$ is
the Laplace operator on $\Zsp$ space, and $R_{ab}(\Zsp)$ is the Ricci tensor
with respect to the metric $u_{ab}(\Zsp)$\,.
Setting $\bar{\lambda}\ne 0$, there is no cosmological solution because of
Eq.~(\ref{ro5:eq-h2:Eq}).
Let us consider the case
\Eq{
u_{ab}=\delta_{ab}\,,~~~~~~\bar{\lambda}=0\,,~~~~~~\frac{dh_0}{dt}=0\,,
\label{ro5:flat:Eq}
}
where $\delta_{ab}$ the four-dimensional Euclidean metric.
Then we can construct the solution
\Eqrsubl{ro5:solution:Eq}{
h_2(z)&=&c_1+\sum_{\alpha=1}^{N'}\frac{L_{\alpha}}{|z^a-z^a_{\alpha}|^2}\,,\\
k_2(t, z)&=&c_2\,t+c_3+\sum_{l=1}^{N}\frac{M_l}{|z^a-z^a_l|^2}\,,
}
where $c_i~(i=1,~2,~3)$, $L_{\alpha}$, and $M_l$
are the constants.
Now we discuss the cosmological evolution for time-dependent
solution (\ref{ro5:solution:Eq}).
We define the cosmic time $\tau$, which is given by
\Eq{
\left(\frac{\tau}{\tau_0}\right) \equiv
\left(c_2\,t\right)^{2/3}\,,~~~~
\tau_0 \equiv \frac{3}{2c_2}\,,~~~~c_3=0\,.
\label{ro5:ct2:Eq}
}
The five-dimensional metric can be expressed as
\Eqr{
&&\hspace{-0.8cm}ds^2=h_2^{-4/3}(z)
\left[1+\left(\frac{\tau}{\tau_0}\right)^{-3/2}\bar{k}(z)
\right]^{-\frac{2}{3}}
\left[-d\tau^2
\right.\nn\\
&&\left.+h_2^2(z)\left\{1+\left(\frac{\tau}{\tau_0}\right)^{-3/2}
\bar{k}(z)\right\}
\left(\frac{\tau}{\tau_0}\right)^{1/2}
\delta_{ab}(\Zsp)dz^adz^b\right],
\label{ro5:metric4:Eq}
}
where the functions $h_2(z)$ and $\bar{k}(z)$ are given by
\Eq{
h_2(z)=c_1+\sum_{\alpha=1}^{N'}
\frac{L_{\alpha}}{|z^a-z^a_{\alpha}|^{2-d_{s}}}\,,~~~~~
\bar{k}(z) \equiv c_3+\sum_{l=1}^{N}\frac{M_l}{|z^a-z^a_l|^{2-d_{s}}}\,.
\label{ro5:sme:Eq}
}
Here $d_{s}$ is the number of smeared dimensions and
should satisfy $0\le d_{s}\le 3$\,. Here we assume that
one direction of $z^a~(a=1,\ldots,4)$ is not smeared
in order to fix the location of our universe in the transverse space.
Our Universe is given by the solutions with the
five-dimensional coordinates $t,~z^a~(a=1,\ldots,4)$\,.
The time direction is written by $t$\,.
Our choice is to take the three-dimensional
from the overall transverse space with $z^a$.
The four-dimensional universe is spanned by $t$, $z^2$, $z^3$ and $z^4$\,,
for instance. The $z^1$ direction is preserved to measure the position
of our universe in the overall transverse space of 0-branes.
Since the metric depends on $z^a$ explicitly, we have to smear
out ${z}^2$\,, ${z}^3$ and ${z}^4$ so as to define our Universe.
Then the number of the smeared directions
$d_{s}$ should satisfy the condition $d_{s}=3$\,.
Unfortunately, the power exponent of four-dimensional
universe becomes 1/4.
Hence, we have to conclude that in order to obtain a realistic expansion
of the universe in this type of models, one has to include additional
fields on the background.
We study the asymptotic behavior of the dynamical 0-brane background.
The time dependence in the function $h_2$ can be ignored
in the limit of $z^a\rightarrow \,z^a_l$, because
the harmonic function $\bar{k}(z)$ dominates near a position of 0-brane.
In the limit of $z^a\rightarrow \infty$\,, as function $\bar{k}(z)$
vanishes, the system becomes static near 0-brane.
Then, the function $k_2$ depends only on time in the far region from 0-branes.
The five-dimensional metric in the limit of $z^a\rightarrow \infty$\,,
is thus given by
\Eqr{
ds^2=-\left(c_2\,t+c_3\right)^{-2/3}
dt^2+\left(c_2\,t+c_3\right)^{1/3}
u_{ab}dz^adz^b\,.
\label{ro5:surface:Eq}
}
The metric has singularity at $t=-c_3/c_2$\,.
Then the five-dimensional spacetime does not have any singularity
if it is restricted inside the domain satisfied by the conditions,
\Eq{
h_2(z)=c_1+\sum_{\alpha=1}^{N'}
\frac{L_{\alpha}}{|z^a-z^a_{\alpha}|^{2-d_{s}}}>0\,,~~~~~~
k_2(t, z)=c_2\,t+\bar{k}(z)>0\,,
\label{ro5:do:Eq}
}
where the function $\bar{k}(z)$ is defined in (\ref{ro5:sme:Eq}).
The five-dimensional spacetime cannot be extended beyond this region.
Since the spacetime evolves into a curvature singularity,
the regular spacetime with dynamical 0-branes ends up with
the singularities.
Although the evolution of dynamical 0-brane with $c_2>0$ has the time
reversal one of $c_2<0$, the behavior of the background spacetime strongly
depends on the signature of $c_2$\,.
In the following, we will focus on the case with $c_2<0$.
For $t<0$, the function $h_2$ is positive everywhere. Then
the spacetime is not singular.
In the limit of $t\rightarrow -{\infty}$, the solution becomes
a time-dependent uniform spacetime apart from a position of 0-branes.
The five-dimensional background geometry can be described as
a cylindrical form of infinite throat near the dynamical 0-branes,
Let us consider the time evolution the five-dimensional spacetime.
At $t=0$\,, the five-dimensional spacetime does not have any curvature
singularity in the background.
The background geometry has a cylindrical topology near each 0-brane.
As time slightly increases, a curvature singularity
appears far from 0-branes $|z^a-z^a_\alpha|\rightarrow\infty$\,.
After that, the singular hypersurface cuts off more and more of the
space as time increases further.
The singular hypersurface splits and surrounds each of
the 0-brane throats individually after time continues to evolve.
The spatial surface is finally composed of two isolated throats.
The time evolution of the five-dimensional
spacetime for $t<0$ is the time reversal of $t>0$.
We find that the overall transverse space tends to expand asymptotically
like $t^{1/6}$\,, for any values of fixed $z^a$, in the regular domain
of the five-dimensional metric (\ref{ro5:surface:Eq}),
while the solutions describe static
0-branes near the positions of the branes.
In the far from 0-branes, where $|z^a-z^a_{\alpha}|
\rightarrow \infty$\,, the background geometry becomes
FRW universes with the power law expansion $t^{1/6}$\,.
Next we consider the case of the near-horizon limit that
the spacetime metric and the functions $h_2$,
$k_2$ are given by (\ref{ro5:solution:Eq}).
If we consider the case where all 0-branes are located at the origin of
the Z space, we have
\Eqrsubl{ro5:hk2:Eq}{
h_2(r)&=&c_1+\frac{L}{r^2}\,,\\
k_2(t, r)&=&c_2t+c_3+\frac{M}{r^2}\,,~~~~~~~~
r^2 \equiv \delta_{ab}\, z^a z^b\,,
}
where $L$, $M$ are the total mass of 0-branes
\Eq{
L \equiv \sum_{\alpha=1}^{N'} L_{\alpha}\,,~~~~~~
M \equiv \sum_{l=1}^{N} M_{l}\,.
}
Since the dependence on $t$ in (\ref{ro5:hk2:Eq}) is negligible
in the near-horizon limit $r\rightarrow 0$\,,
the five-dimensional metric is reduced to the following form
\Eqrsubl{ro5:nh3-metric:Eq}{
ds^2&=&ds^2_{\rm AdS_2}+L^{2/3}M^{1/3}d\Omega^2_{(3)}\,, \\
ds^2_{\rm AdS_2}&\equiv&L^{-4/3}M^{-2/3}
\left(-r^4dt^2+\frac{L^2M}{r^2}\,dr^2\right)\,,
}
where $\delta_{ab}\, dz^a dz^b=dr^2+r^2d\Omega^2_{(3)}$ has been used.
The line elements of a two-dimensional AdS space (AdS$_2$)
and a three-sphere with the unit radius (${\rm S}^3$)
are given by $ds^2_{\rm AdS_2}$ and $d\Omega^2_{(3)}$\,, respectively.
Thus we see that the near-horizon limit of the 0-brane system is a
AdS$_2$ with a certain internal 3-space.
Before closing this subsection, we discuss the collision of 0-brane.
There are two kinds of 0-brane in the five-dimensional spacetime.
One is static 0-brane coming from the function $h_2(z)$\,. The other is
dynamical 0-brane given by $k_2(t, z)$\,.
We set the two dynamical 0-branes at $\,\vect{z}_1=(0, 0,\cdots, 0)$
and $\,\vect{z}_2=(P, 0,\cdots, 0)$\,,
where $P$ is a constant.
On the other hand, we suppose that $N'$ static
0-branes are sitting at a point,
\Eq{
\vect{z}_{1} = \cdots = \,\vect{z}_{N'} \equiv \,\vect{z}_0
=(z_0^1\,,\,0\,,\,\cdots\,,\,0)\,.
\label{ro5:point:Eq}
}
Now we consider the following quantity
\Eq{
\tilde{z}=\sqrt{\left(z^2\right)^2+\left(z^3\right)^2+\cdots
+\left(z^{4-d_{s}}\right)^2}\,.
}
Then the proper length at $\tilde{z}=0$ between the two dynamical 0-branes
is given by
\Eqr{
d(t)&=&\int^{P}_0 dz^1\left(c_1+\frac{L}{|z^1-z^1_0|^{2-d_{s}}}\right)^{1/3}
\left(c_2\,t+c_3+\frac{M_1}{|z^1|^{2-d_{s}}}
+\frac{M_2}{|z^1-P|^{2-d_{s}}}\right)^{1/6}\,,
\label{ro5:distance:Eq}
}
where $M_1$ and $M_2$ are the charges of the dynamical
0-brane and $L$ is defined by
\Eq{
L=\sum_{\alpha=1}^{N'}L_\alpha\,.
\label{ro5:L:Eq}
}
For $c_2=-1$, the length $d(t)$ is a monotonically decreasing
function of time. Since the time evolution of the proper length
depends on the number of the smeared directions $d_{s}$\,,
we shall analyze it for each of the values of $d_{s}$ below.
First we consider the case with $d_s \le 2$.
For $d_2=2$, the harmonic functions $h_2$ and $k_2$ diverges both at
infinity and near 0-branes. In particular,
there is no regular spacetime region near 0-branes
because of $h_2\rightarrow\infty$ and $\bar{k}\rightarrow\infty$\,.
Then, these are not physically relevant. Hence, we show
the proper length in Fig.~\ref{fig:ro0} for the cases with
$d_s=0$ and $d_s=1$. For both cases the singularity between
two dynamical 0-branes appears before collision because a singularity appears
before the proper distance becomes zero. Although two dynamical 0-branes
initially approach very slowly, the singular hypersurface suddenly
appears at a finite distance, and the spacetime finally splits into
two isolated 0-brane throats.
Therefore we cannot analyze the collision of the dynamical 0-branes in these
examples.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ro3.eps}
\put(-175,120){$d(t)$}
\put(10,10){$t$}
\hskip 2.0cm
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ro4.eps}
\put(-175,120){$d(t)$}
\put(10,10){$t$}\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The behavior of the proper distance between two dynamical 0-branes
for $d_s=0$ (a) and $d_s=1$ (b) in
the five-dimensional Romans' theory. For both cases,
the two dynamical 0-brane charges
are identical, $M_1=M_2=1$ and the parameters are taken as $c_1=0$\,,
$c_2=-1$\,, $c_3=0$\,, $L=1$\,, $z_0^1=0$\,,
and $P=1$.
The result is also the same and a singularity appears before
collision of dynamical 0-branes.}
\label{fig:ro0}
\end{center}
\end{figure}
However, for the case with $d_{s}=3$\,,
the function $h_2$ and $\bar{k}$ are written by the linear function of
$z^a$\,.
If we assume that the $z^a$ directions apart from $z^1$ are smeared,
the time evolution of the proper length is different from the previous case.
Hence, the harmonic functions $h_2$ and $\bar{k}$ are expressed as
\Eq{
h_2\left(z^1\right)=c_1+\sum_{\alpha=1}^{N'}L_{\alpha}
|z^1-z^1_{\alpha}|\,,~~~~~~~
\bar{k}\left(z^1\right)=c_3+\sum_{l=1}^{N}M_l
|z^1-z^1_l|\,.
\label{ro5:bh2:Eq}
}
We study the dynamics of 0-branes, where one 0-brane
charge $M_1$ is located at $z^1=0$ and the other $M_2$ at $z^1=P$\,.
The proper distance between the two dynamical 0-branes is given by
\Eqr{
d(t)&=&\int^P_0 dz^1 \left(c_1+L|z^1-z_0^1|\right)^{1/3}
\left[c_2\,t+c_3+\left(M_1|z^1|+M_2|z^1-P|\right)\right]^{1/6}\,,
\label{ro5:length:Eq}
}
where we assume again that $N'$ static
0-branes are sitting at a point $z^1=z_0^1$\,, and $L$ is defined by
(\ref{ro5:L:Eq}).
For $c_2<0$, the proper distance decreases with time.
Setting $M_1\ne M_2$\,, a curvature singularity appears again at
a certain finite time $t=t_{\rm s}$\, before the dynamical 0-branes
collide. Then, $t_{\rm s}$ is written by
\Eq{
t_{\rm s}\equiv-\frac{c_3+M_1|z^1|+M_2|z^1-P|}{c_2}\,.
}
This is the same result as the case with $d_{s}\leq 2$\,.
On the other hand, two 0-branes have the same brane charge $M_1=M_2=M$\,,
the proper distance vanishes at a certain finite time $t=t_{\rm c}$\,,
where $t_{\rm c}$ is defined by
\Eq{
t_{\rm c} \equiv -\frac{c_3+MP}{c_2}\,.
}
Then two dynamical 0-branes collide completely.
If we set $z^1_0=0$\,,
for simplicity, the proper length between two dynamical 0-branes
can be written by
\Eqr{
d(t)=\frac{3}{4L}\left[-c_1^{4/3}+\left(c_1+LP\right)^{4/3}\right]
\left[c_2(t-t_{\rm c})\right]^{1/6}\,.
}
If we choose the physical parameters as $c_1=0$\,,
$c_2=-1$\,, $c_3=0$\,, $P=1$\,, $z^1_0=0$\,, and
$L=1$\,, the proper distance $d(t)$ is depicted in
Fig.~\ref{fig:ro} for the two cases (a) the same 0-brane charges
$M_1=M_2=1$ and (b) different
charges $M_1=2$, $M_2=1$\,.
For the case (a), the two dynamical 0-branes can collide completely.
On the other hand, in the case (b) a singularity appears before the
collision of dynamical 0-branes, as we have already discussed in
Sec.\,\ref{sec:ns-c}.
\begin{figure}[h]
\begin{center}
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ro1.eps}
\put(-175,115){$d(t)$}
\put(10,10){$t$}
\hskip 2.0cm
\includegraphics[keepaspectratio, scale=0.55, angle=0]{ro2.eps}
\put(-175,115){$d(t)$}
\put(10,10){$t$}
\\
(a) \hskip 7cm (b) ~~~~~~
\caption{\baselineskip 14pt
The behavior of the proper distance between two dynamical 0-branes
for $M_1=M_2=1$ (a) and $M_1=2$, $M_2=1$ (b) in
the five-dimensional Romans' theory. We fix $d_s=3$\,, $c_1=0$\,,
$c_2=-1$\,, $c_3=0$\,, $z^1_0=0$\,,
$L=1$\,, and $P=1$.
The proper length rapidly vanishes near two
0-branes collide for the case of $M_1=M_2=1$.
For the case of $M_1=2$, $M_2=1$, it
is still finite when a curvature singularity appears.}
\label{fig:ro}
\end{center}
\end{figure}
\section{The Instability of the dynamical brane background}
\label{sec:in}
In this section, we briefly discuss the nature of the singularities
appearing in the time-dependent solutions and present the stability
analysis for the dynamical brane background. We follow the method
used by \cite{oai1, Quevedo:2002tm, oai2, Cornalba:2003ze} (See also
\cite{Grojean:2001pv, Cornalba:2002fi, Cornalba:2002nv}) and
present the preliminary analysis performed,
where the Klein-Gordon modes are analyzed.
An analysis of such possibility will definitely make the property of
singularity more clear, even if it is just a simple preliminary
study to asses this issue.
\subsection{The dynamical 0-brane background in
Nishino-Salam-Sezgin gauged supergravity}
Let us first consider the stability for the 0-brane solution in
Nishino-Salam-Sezgin gauged supergravity.
The six-dimensional metric becomes static space near 0-brane
while the background depends only on the times far
from 0-brane. We will study the stability in the
0-brane solution far from the branes.
For the limit $r\rightarrow\infty$ in the solution (\ref{nh:h2:Eq}),
the six-dimensional metric is expressed as
\Eqrsubl{st:metric:Eq}{
ds^2&=&-\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{-3/2}
dt^2+\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{1/2}
\delta_{mn}(\Wsp)dv^mdv^n\,,\\
\delta_{mn}(\Wsp)dv^mdv^n&\equiv&dr^2+r^2\omega_{ij}(\Ssp^4)d\xi^id\xi^j\,,
}
where $\omega_{ij}(\Ssp^4)$ denotes the metric of four-dimensional sphere.
The six-dimensional metric has curvature singularity at
$t=-c_4/\epsilon\sqrt{2\Lambda}$\,.
In order to study the stability, we consider the Klein-Gordon equation for
a massive scalar field propagating in the background (\ref{st:metric:Eq}):
\Eq{
-\frac{1}{\sqrt{-g}}\pd_M\left(\sqrt{-g}g^{MN}\pd_N\varphi\right)
+m^2\varphi=0\,,
\label{st:KG:Eq}
}
where $g$ denotes the determinant of the six-dimensional metric
(\ref{st:metric:Eq}).
In terms of the metric (\ref{st:metric:Eq}),
the Klein-Gordon equation can be written by
\Eq{
\pd_t\left[\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^2\pd_t\varphi\right]
-r^{-4}\pd_r\left(r^4\pd_r\varphi\right)-\frac{1}{r^2}\lap_{\Ssp^4}\varphi
+\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{1/2}m^2\varphi=0\,,
}
where $\lap_{\Ssp^4}$ denotes the Laplace operator on the $\Ssp^4$.
The six-dimensional metric involved permits separation of variables,
so we take
\Eq{
\varphi=\varphi_0(t)\varphi_1(r)\varphi_2(\xi)\,,
\label{st:ansatz:Eq}
}
where the functions $\varphi_1(r)$, $\varphi_2(\xi)$ obeys the
eigenvalue equations
\Eq{
\lap_\Wsp\varphi_1(r)\varphi_2(\xi)
=-\lambda_{\Wsp}^2\,\varphi_1(r)\varphi_2(\xi)\,.
\label{st:e-y:Eq}
}
Here $\lap_\Wsp$ is the Laplace operator on the W space,
$\lambda_{\Wsp}$ is eigenvalue and
$\varphi_1(r)$, $\varphi_2(\xi)$ satisfy
\cite{Dowker:1988pp}
\Eq{
\varphi_1(r)=\frac{1}{r}\left[b_1J_{\nu}(\lambda_{\Wsp} r)
+b_2Y_{\nu}(\lambda_{\Wsp} r)
\right],~~~~~
\lap_{\Ssp^4}\varphi_2(\xi)=-\lambda^2_{\Ssp^4}\varphi_2(\xi)\,,
}
where $b_1$, $b_2$ are constants, and
$J_{\nu}$, $Y_{\nu}$ denote the Bessel functions and $\nu$ is
related to the eigenvalue $\lambda^2_{\Ssp^4}$ as
\Eq{
\nu^2=\lambda^2_{\Ssp^4}+\frac{9}{4}\,.
}
The Klein-Gordon equation thus is rewritten by
\Eq{
\frac{d^2\varphi_0}{dt^2}
+\frac{2\epsilon\sqrt{2\Lambda}}{\left(\epsilon\sqrt{2\Lambda}t+c_4\right)}
\frac{d\varphi_0}{dt}
+\frac{1}{\left(\epsilon\sqrt{2\Lambda}t+c_4\right)^2}\left[\lambda_{\Wsp}^2
+\left(\epsilon\sqrt{2\Lambda}t+c_4\right)^{1/2}m^2\right]\varphi_0=0\,.
}
Then the solution for $\varphi_0$ is oscillatory, having the form
\Eqr{
\varphi_0(t)&=&\frac{\Lambda}
{2m^2\left(\epsilon\sqrt{2\Lambda}t+c_4\right)^{1/2}}\,\left[\eta_1\,
\Gamma\left(1-\gamma\right)J_{-\gamma}
\left\{\frac{4m\left(\epsilon\sqrt{2\Lambda}t+c_4\right)^{1/4}}
{\epsilon\sqrt{2\Lambda}}\right\}\right.\nn\\
&&\left.+\eta_2\,\Gamma\left(1+\gamma\right)J_{\gamma}
\left\{\frac{4m\left(\epsilon\sqrt{2\Lambda}t+c_4\right)^{1/4}}
{\epsilon\sqrt{2\Lambda}}\right\}\right]\,,
\label{st:solutions:Eq}
}
where $\eta_1$, $\eta_2$ are constants, $J_{-\gamma}$, $J_{\gamma}$ denote
the Bessel functions and $\gamma$ is given by
\Eq{
\gamma=2\sqrt{1-\frac{2\lambda_{\Wsp}^2}{\Lambda}}\,.
\label{st:gamma:Eq}
}
Let us consider the energy of the Klein-Gordon modes to study whether
the instability occurs or not.
Using the asymptotic solution
(\ref{st:solutions:Eq}),
we will see that $E\rightarrow\infty$ as the singularity is approached, where
there is a curvature singularity at $t=-c_4/\epsilon\sqrt{2\Lambda}$
in the six-dimensional background (\ref{st:metric:Eq}).
Since the velocity is well-behaved besides the singularity,
the energy of the Klein-Gordon modes can estimate as
\Eq{
E=-u^M\pd_M\varphi\,,~~~~~u=\alpha\pd_t+\beta\pd_r\,,
\label{st:e:Eq}
}
where $u$ is velocity.
In terms of the normalization condition $u^2=-1$, the behavior of the
$\alpha$ and $\beta$ are determined in order to remain
non-singular.
Then, we find
\Eq{
-\alpha^2\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{-3/2}
+\beta^2\left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{1/2}=-1\,.
}
As $t\rightarrow 0$, $\alpha$ and $\beta$ have to behave
\Eq{
\alpha\sim \left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{3/4}\,,~~~~~
\beta\sim \left(\epsilon\sqrt{2\Lambda}\,t+c_4\right)^{-1/4}\,,
\label{st:coe:Eq}
}
in the limit $r\rightarrow\infty$ for the dynamical 0-brane background.
Upon setting (\ref{st:coe:Eq}), one then finds
\Eq{
-E=\alpha\pd_t\varphi+\beta\pd_r\varphi\,.
}
In terms of the asymptotic solution (\ref{st:solutions:Eq})
with $\gamma$ (\ref{st:gamma:Eq}), we find that $E\rightarrow\infty$
as the singularity is approached if we set $\epsilon=1$ and
$\Lambda>0$.
Hence the 0-brane solution implies
that the energy momentum tensor of the scalar field mode diverges
far from 0-brane.
However, it is necessary to study a full analysis of the metric perturbations
whether the mode of Klein Gordon field is not likely to
destabilize the metric modes near the singularity or not.
\subsection{The dynamical 1-brane background in
Nishino-Salam-Sezgin gauged supergravity}
Next we consider the stability for the 1-brane solution in
Nishino-Salam-Sezgin gauged supergravity.
For the metric (\ref{nsc:h:Eq}), the harmonic function
$\tilde{h}(z)$ dominates in the limit of $z^a\rightarrow z^a_l$
(near a position of 1-branes) and the time dependence can be ignored.
Thus the background becomes static.
On the other hand, in the limit of
$|z^a|\rightarrow \infty$, $\tilde{h}(z)$ vanishes.
Then $h_3$ depends on
$t$ and $y$ in the far region from 1-branes and the
resulting metric is given by
\Eqrsubl{st1:metric:Eq}{
ds^2&=&\left[\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t+
c_2y+c_3\right]^{-1/2}\left(-dt^2+dy^2\right)\nn\\
&&+\left[\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t+
c_2y+c_3\right]^{1/2}\delta_{ab}(\Zsp)dz^adz^b\,,\\
\delta_{ab}(\Zsp)dz^adz^b&\equiv&dr^2
+r^2\tilde{\omega}_{ij}(\Ssp^3)d\xi^id\xi^j\,,
}
where $\tilde{\omega}_{ij}(\Ssp^3)$ is
the metric of three-dimensional sphere.
For the six-dimensional metric, a curvature singularity may appear at
\Eq{
\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t+
c_2y+c_3=0\,.
}
In the following, we will again discuss the stability in the
1-brane solution far from the branes.
Let us consider the Klein-Gordon equation for
a massive scalar field in the six-dimensional background
(\ref{st1:metric:Eq}):
\Eq{
-\frac{1}{\sqrt{-g}}\pd_M\left(\sqrt{-g}g^{MN}\pd_N\psi\right)
+m^2\psi=0\,,
\label{st1:KG:Eq}
}
where $g$ denotes the determinant of the six-dimensional metric
(\ref{st1:metric:Eq}).
Substituting the six-dimensional metric (\ref{st1:metric:Eq})
into the Klein-Gordon equation for
a massive scalar field (\ref{st1:KG:Eq}), we find
\Eqr{
&&\pd_t\left[\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t\right\}
\pd_t\psi\right]
-\pd_y\left[\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t\right\}
\pd_y\psi\right]
-r^{-3}\pd_r\left(r^3\pd_r\psi\right)\nn\\
&&~~~~~-\frac{1}{r^2}\lap_{\Ssp^3}\psi
+\left[\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t\right\}\right]^{1/2}
m^2\psi=0\,,
\label{st1:field:Eq}
}
where we set $c_2=c_3=0$\,, and
$\lap_{\Ssp^3}$ denotes the Laplace operator on the $\Ssp^3$.
The six-dimensional metric involved permits separation of variables,
so we take
\Eq{
\psi=\psi_0(t)\psi_1(y)\psi_2(r)\psi_3(\xi)\,,
\label{st1:ansatz:Eq}
}
where the functions $\psi_2(r)$, $\psi_3(\xi)$ obeys the
eigenvalue equations
\Eq{
\lap_\Zsp\psi_2(r)\psi_3(\xi)=
-\lambda_{\Zsp}^2\,\psi_2(r)\psi_3(\xi)\,.
\label{st1:eigen:Eq}
}
Here $\triangle_{\rm Z}$ is the Laplace operator on the Z space,
$\lambda_{\Zsp}$ is eigenvalue for the equation (\ref{st1:eigen:Eq})\,,
and $\psi_2(r)$, $\psi_3(\xi)$ are satisfy
\Eq{
\psi_2(r)=\frac{1}{r}\left[b_1J_{\nu}(\lambda_{\Zsp} r)
+b_2Y_{\nu}(\lambda_{\Zsp} r)
\right],~~~~~
\lap_{\Ssp^3}\psi_3(\xi)=-\lambda^2_{\Ssp^3}\psi_3(\xi)\,,
\label{st1:eigen2:Eq}
}
where $b_1$, $b_2$ are constants,
$J_{\nu}$, $Y_{\nu}$ denote the Bessel functions and $\nu$ is
related to $\lambda^2_{\Ssp^3}$\,:
\Eq{
\nu^2=\lambda^2_{\Ssp^3}+1\,.
\label{st1:value:Eq}
}
Hence, the Klein-Gordon equation is reduced to
\Eqr{
&&\psi_1\pd_t\left[\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t\right\}
\frac{d\psi_0}{dt}\right]
-\psi_0\pd_y\left[\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1 t\right\}
\frac{d\psi_1}{dy}\right]\nn\\
&&~~~~~+\left[\lambda_{\Zsp}^2
+\left\{\frac{\Lambda}{2}\left(t^2-y^2\right)+c_1t\right\}^{1/2}
m^2\right]\psi_0\psi_1=0\,.
\label{st1:field2:Eq}
}
We shall discuss the massless cases in the following. In terms of
$c_1=0$ and $m=0$\,,
the particular solutions of $\psi_0$ and $\psi_1$ are given, respectively, by
\Eq{
\psi_0(t)=\zeta_1\,t^{-\frac{1}{2}+\rho}+\zeta_2\,t^{-\frac{1}{2}-\rho}\,,~~~~~
\psi_1(y)=\sigma_1\,y^{-\frac{1}{2}+\rho}+\sigma_2\,y^{-\frac{1}{2}-\rho}\,,
\label{st1:solutions:Eq}
}
where $\zeta_i~(i=1,\,2)$ and $\sigma_i~(i=1,\,2)$ are constants and
$\rho$ is defined by
\Eq{
\rho=\frac{1}{2}\sqrt{1-\frac{4\lambda_{\Zsp}^2}{\Lambda}}\,.
\label{st1:rho:Eq}
}
We study the stability in terms of the energy of the Klein-Gordon modes.
Using the asymptotic solution given by Eq.~(\ref{st1:solutions:Eq}),
we can estimate the energy
near the singularity,
where there is a curvature singularity at $t=\pm y$
in the six-dimensional background (\ref{st1:metric:Eq}).
Since the velocity is well defined besides the singularity,
the energy of the Klein-Gordon modes can be written as
\Eq{
E=-v^M\pd_M\psi\,,~~~~~v=\alpha_t\pd_t+\alpha_y\pd_y+\alpha_r\pd_r\,,
\label{st1:e:Eq}
}
where $v$ is velocity.
By using the normalization condition $v^2=-1$, the behavior of the
$\alpha_t$\,, $\alpha_y$\,, and $\alpha_r$ are determined
in order to remain nonsingular. Then, we find
\Eq{
\left(-\alpha_t^2+\alpha_y^2\right)
\left[\frac{\Lambda}{2}\left(t^2-y^2\right)\right]^{-1/2}
+\alpha_r^2\left[\frac{\Lambda}{2}\left(t^2-y^2\right)\right]^{1/2}=-1\,.
}
In the limit $r\rightarrow\infty$ and $t\rightarrow 0$,
for the dynamical 1-brane background (\ref{st1:metric:Eq}),
$\alpha_t$, $\alpha_y$, and $\alpha_r$ provided
\Eq{
\left(-\alpha_t^2+\alpha_y^2\right)^{1/2}
\sim \left[\frac{\Lambda}{2}\left(t^2-y^2\right)\right]^{1/4}\,,~~~~~
\alpha_r\sim \left[\frac{\Lambda}{2}\left(t^2-y^2\right)\right]^{-1/4}\,.
\label{st1:coe:Eq}
}
If we set the parameters (\ref{st1:coe:Eq}), one then finds
\Eq{
-E=\alpha_t\pd_t\psi+\alpha_y\pd_y\psi+\alpha_r\pd_r\psi\,.
}
In terms of the asymptotic solution (\ref{st1:solutions:Eq}),
we find that $E\rightarrow\infty$
as the singularity is approached at $t=\pm y$\,.
Let us next consider the case $m=0$ and $\Lambda=0$\,.
In the limit $r\rightarrow\infty$ for the solution (\ref{nh:h3:Eq}),
the six-dimensional metric becomes
\Eqrsubl{st1:metric2:Eq}{
ds^2&=&\left(c_1\,t+c_4\right)^{-1/2}
\left(-dt^2+dy^2\right)
+\left(c_1\,t+c_4\right)^{1/2}
\delta_{ab}(\Zsp)dz^adz^b\,,\\
\delta_{ab}(\Zsp)dz^adz^b&\equiv&dr^2+r^2\omega_{ij}(\Ssp^3)d\xi^id\xi^j\,,
}
where we set $c_2=c_3=0$ and $\omega_{ij}(\Ssp^3)$ denotes
the metric of the three-dimensional sphere.
There is a curvature singularity at $t=-c_4/c_1$\,.
Now we study the behavior of the Klein-Gordon field.
The scalar field equation (\ref{st1:KG:Eq}) in the six-dimensional background
(\ref{st1:metric2:Eq}) reads
\Eq{
\pd_t\left(c_1 t\,\pd_t\psi\right)
-\pd_y\left(c_1 t\pd_y\psi\right)
-r^{-3}\pd_r\left(r^3\pd_r\psi\right)
-\frac{1}{r^2}\lap_{\Ssp^3}\psi+\left(c_1 t\right)^{1/2}
m^2\psi=0\,.
\label{st1:field3:Eq}
}
If we assume that the scalar field $\psi$ is given by
(\ref{st1:ansatz:Eq}), where $\psi_2(r)$ and $\psi_3(\xi)$ can be
written by (\ref{st1:eigen2:Eq}), the function $\psi_1(y)$ is
determined by the eigenvalue equation:
\Eq{
\frac{d^2\psi_1}{dy^2}=-\lambda_y^2\,\psi_1\,.
}
Here, $\lambda_y$ is constant. Then, the equation for $\psi_0$ becomes
\Eq{
\frac{d}{dt}\left(c_1 t\,\frac{d\psi_0}{dt}\right)
+\left[\lambda_y^2\,c_1 t+\lambda_{\Zsp}^2
+\left(c_1 t\right)^{1/2}m^2\right]\psi_0=0\,.
}
For the massless case, the solution of $\psi_0$ is given by the
oscillatory form
\Eq{
\psi_0(t)=\e^{-i\lambda_y t}\left[
f_1U\left(\vartheta\,, 1\,, 2i\lambda_y t\right)
+f_2L_{-\vartheta}\left(2i\lambda_y t\right)\right],
\label{st1:solution:Eq}
}
where $U$ denotes the hypergeometric function, $L_{-\vartheta}$ is the
Laguerre polynomial, $f_1$ and $f_2$ are constants,
and $\vartheta$ is defined by
\Eq{
\vartheta=\frac{1}{2}+\frac{i \lambda^2_\Zsp}{2c_1\lambda_y}\,.
}
We estimate the energy of the Klein-Gordon modes whether
the instability exists or not.
In terms of the asymptotic solution (\ref{st1:solution:Eq}),
we can present the behavior of the energy
as the singularity is approached.
Since the velocity is well behaved except for the singularity,
the energy of the Klein-Gordon modes is given by (\ref{st1:e:Eq})\,.
By using the normalization condition $v^2=-1$,
$\alpha_t$, $\alpha_y$, and $\alpha_r$ are determined by
\Eq{
\left(-\alpha_t^2+\alpha_y^2\right)\left(c_1\,t\right)^{-1/2}
+\alpha_r^2\left(c_1\,t\right)^{1/2}=-1\,.
}
As $t\rightarrow 0$ and $r\rightarrow\infty$, the functions
$\alpha_t$, $\alpha_y$, and $\alpha_r$ are set to be
\Eq{
\left(-\alpha_t^2+\alpha_y^2\right)^{1/2}\sim \left(c_1\,t\right)^{1/4}\,,
~~~~~\alpha_r\sim \left(c_1\,t\right)^{-1/4}\,.
\label{st1:coe2:Eq}
}
If we use Eq.~(\ref{st1:coe2:Eq}), the energy can be expressed as
\Eq{
-E=\alpha_t\pd_t\psi+\alpha_y\pd_y\psi+\alpha_r\pd_r\psi\,.
\label{st1:energy2:Eq}
}
Then, for the asymptotic solution (\ref{st1:solution:Eq}),
the energy becomes $E\rightarrow\infty$
as the singularity is approached, that is, $t\rightarrow 0$.
Since the 1-brane solution gives
that the energy-momentum tensor of the Klein-Gordon field mode diverges
in this limit, the mode of the scalar field cannot
stabilize the metric modes near the singularity.
\subsection{The dynamical 0-brane background in five-dimensional
Romans' gauged supergravity}
In this subsection, we analyze the stability of the
dynamical 0-brane background
in the five-dimensional Romans' gauged supergravity. We will study the
stability of the scalar field far from 0-branes.
For $r\rightarrow\infty$ in the dynamical 0-brane background
(\ref{ro5:hk2:Eq}), the five-dimensional metric becomes
\Eqrsubl{r5:metric:Eq}{
ds^2&=&-\left(c_2\,t+c_3\right)^{-2/3}dt^2
+\left(c_2\,t+c_3\right)^{1/3}
\delta_{ab}(\Zsp)dz^adz^b\,,\\
\delta_{ab}(\Zsp)dz^adz^b&\equiv&dr^2+r^2\omega_{ij}(\Ssp^3)d\xi^id\xi^j\,,
}
where we set $c_1=1$ and
$\omega_{ij}(\Ssp^3)$ denotes the metric of the three-dimensional sphere.
There is a curvature singularity at $t=-c_3/c_2$\,.
Let us consider the behavior of the Klein-Gordon field:
\Eq{
-\frac{1}{\sqrt{-g}}\pd_M\left(\sqrt{-g}g^{MN}\pd_N\varphi\right)
+m^2\varphi=0\,,
\label{r5:KG:Eq}
}
where $g$ denotes the determinant of the five-dimensional metric
(\ref{r5:metric:Eq}).
The scalar field equation (\ref{r5:KG:Eq}) in the five-dimensional background
(\ref{r5:metric:Eq}) reads
\Eq{
\pd_t\left[\left(c_2\,t+c_3\right)\,\pd_t\varphi\right]
-r^{-3}\pd_r\left(r^3\pd_r\varphi\right)
-\frac{1}{r^2}\lap_{\Ssp^3}\varphi+\left(c_2\,t+c_3\right)^{1/3}
m^2\varphi=0\,.
\label{r5:field3:Eq}
}
Here, $\lap_{\Ssp^3}$ denotes the Laplace operator on the $\Ssp^3$\,,
and the scalar field $\varphi$ is assumed to be
\Eq{
\varphi=\varphi_0(t)\varphi_1(r)\varphi_2(\xi)\,,
\label{r5:ansatz:Eq}
}
where $\varphi_1(r)$ and $\varphi_2(\xi)$ are
determined by the eigenvalue equation:
\Eq{
\lap_\Zsp\varphi_1(r)\varphi_2(\xi)
=-\lambda_{\Zsp}^2\,\varphi_1(r)\varphi_2(\xi)\,.
\label{r5:e-z:Eq}
}
Here $\triangle_{\rm Z}$ is the Laplace operator on the Z space,
$\lambda_{\Zsp}$ is the eigenvalue, and functions
$\varphi_1(r)$ and $\varphi_2(\xi)$ obey
\cite{Dowker:1988pp}
\Eq{
\varphi_1(r)=\frac{1}{r}\left[\bar{b}_1J_{\nu}(\lambda_{\Zsp} r)
+\bar{b}_2Y_{\nu}(\lambda_{\Zsp} r)
\right],~~~~~
\lap_{\Ssp^3}\varphi_2(\xi)=-\lambda^2_{\Ssp^3}\varphi_2(\xi)\,,
}
where $\bar{b}_1$ and $\bar{b}_2$ are constants,
$J_{\nu}$ and $Y_{\nu}$ denote the Bessel functions, and $\nu$ is
related to the eigenvalue $\lambda^2_{\Ssp^3}$ as
\Eq{
\nu^2=\lambda^2_{\Ssp^3}+1\,.
}
By using Eq.~(\ref{r5:e-z:Eq}), the equation for $\varphi_0$ becomes
\Eq{
\frac{d}{dt}\left[\left(c_2\,t+c_3\right)\,\frac{d\varphi_0}{dt}\right]
+\left[\lambda_{\Zsp}^2
+\left(c_2\,t+c_3\right)^{1/3}m^2\right]\varphi_0=0\,.
}
For $m=0$, the solution of $\varphi_0$
is given by the oscillatory form
\Eq{
\varphi_0(t)=\bar{f}_1J_0\left(\frac{2\lambda_{\Zsp}}{c_2}
\sqrt{c_2\,t+c_3}\right)
+\bar{f}_2Y_0\left(\frac{2\lambda_{\Zsp}}{c_2}\sqrt{c_2\,t+c_3}\right)\,,
\label{r5:solution:Eq}
}
where $\bar{f}_1$ and $\bar{f}_2$ are constants.
We calculate the energy of the Klein-Gordon modes to study the
stability of the dynamical 0-brane background.
By using the asymptotic solution (\ref{r5:solution:Eq}),
we can present the behavior of the energy
as the singularity is approached.
Since it is possible to calculate the velocity except for the singularity,
the energy of the Klein-Gordon modes is given by
\Eq{
E=-u^M\pd_M\varphi\,,~~~~~u=\alpha_t\pd_t+\alpha_r\pd_r\,,
\label{r5:e:Eq}
}
where $u$ denotes the velocity in the five-dimensional spacetime.
In terms of the normalization condition $u^2=-1$,
$\alpha_t$ and $\alpha_r$ are given by
\Eq{
-\alpha_t^2\left(c_2\,t+c_3\right)^{-2/3}
+\alpha_r^2\left(c_2\,t+c_3\right)^{1/3}=-1\,.
}
As $t\rightarrow -c_3/c_2$ and $r\rightarrow\infty$, the functions
$\alpha_t$ and $\alpha_r$ are described as
\Eq{
\alpha_t\sim \left(c_2\,t+c_3\right)^{1/3}\,,
~~~~~\alpha_r\sim \left(c_2\,t+c_3\right)^{-1/6}\,.
\label{r5:coe2:Eq}
}
By using Eq.~(\ref{r5:coe2:Eq}), the energy of the scalar field
can be expressed as
\Eq{
-E=\alpha_t\pd_t\varphi+\alpha_r\pd_r\varphi\,.
\label{r5:energy2:Eq}
}
For the asymptotic solution (\ref{r5:solution:Eq}), one can note that
the energy becomes $E\rightarrow\infty$
as the singularity is approached.
The dynamical 0-brane solution gives
that the energy-momentum tensor of the scalar field mode diverges
in the limit $t\rightarrow-c_3/c_2$\,.
Hence, the mode of the scalar field cannot
stabilize the metric modes near the singularity.
\subsection{Intersection involving the $0-p_I$-brane background in the
$D$-dimensional asymptotically power-law expanding universe}
Now we investigate the stability
analysis for the dynamical $0-p_I$-brane background.
The geometry of the $0-p_{I'}$-brane system becomes a static structure
near branes, while the background geometry depends only on the time
in the far region from branes. By setting $B=0$ in the $D$-dimensional
background (\ref{sa:exact:Eq}), the metric in the limit
$z^a\rightarrow\infty$ is thus given by
\Eqrsubl{sp:metric:Eq}{
ds^2&=&-\left(At\right)^{a_0}dt^2
+\left(At\right)^{b_0}\delta_{ab}(\Zsp)dz^adz^b\,,\\
\delta_{ab}(\Zsp)dz^adz^b&=&dr^2
+r^2\bar{\omega}_{ij}(\Ssp^{D-2})d\xi^id\xi^j\,,
}
where $\bar{\omega}_{ij}(\Ssp^{D-2})$ denotes the metric of the
$(D-2)$-dimensional sphere and $a_0$ and $b_0$ are defined, respectively, by
\Eq{
a_0=-\frac{D-3}{D-2}\,,~~~~b_0=\frac{1}{D-2}\,.
\label{sp:para:Eq}
}
The $D$-dimensional spacetime has singularities at $t=0$\,.
Let us consider the Klein-Gordon equation
to discuss
the stability analysis
\Eq{
-\frac{1}{\sqrt{-g}}\pd_M\left(\sqrt{-g}g^{MN}\pd_N\varphi\right)
+m^2\varphi=0\,,
\label{sp:KG:Eq}
}
where $g$ denotes the determinant of the $D$-dimensional metric
(\ref{sp:metric:Eq}). Equation (\ref{sp:KG:Eq}) on the
$D$-dimensional background (\ref{sp:metric:Eq}) becomes
\Eq{
\pd_t\left(At\pd_t\varphi\right)
-r^{-(D-2)}\pd_r\left(r^{D-2}\pd_r\varphi\right)
-\frac{1}{r^2}\lap_{\Ssp^{D-2}}\varphi+\left(At\right)^{b_0}
m^2\varphi=0\,,
\label{sp:KG2:Eq}
}
where $\lap_{\Ssp^{D-2}}$ denotes the Laplace operator on the $\Ssp^{D-2}$\,.
We assume that the scalar field $\varphi$ is
given by
\Eq{
\varphi=\varphi_0(t)\varphi_1(r)\varphi_2(\xi)\,,
\label{sp:ansatz:Eq}
}
where the functions $\varphi_1(r)$ and $\varphi_2(\xi)$ obey the
eigenvalue equations
\Eq{
\lap_\Zsp\varphi_1(r)\varphi_2(\xi)
=-{\lambda}_{\Zsp}^2\,\varphi_1(r)\varphi_2(\xi)\,.
\label{sp:eigen:Eq}
}
Here $\lap_{\Zsp}$ denotes the Laplace operator on the $\Zsp$ space, and
$\lambda_{\Ysp}$ is the eigenvalue for the equation.
The functions $\varphi_1(r)$ and $\varphi_2(\xi)$ also satisfy the
equations \cite{Dowker:1988pp}
\Eq{
\varphi_1(r)=\frac{1}{r}\left[b_3J_{\bar{\nu}}(\lambda_{\Zsp} r)
+b_4Y_{\bar{\nu}}(\lambda_{\Zsp} r)
\right],~~~~~
{\lap}_{\Ssp^{D-2}}\varphi_2(\xi)
=-{\lambda}_{\Ssp^{D-2}}^2\,\varphi_2(\xi)\,,
\label{sp:eigen2:Eq}
}
where $b_3$ and $b_4$ are constants,
$J_{\bar{\nu}}$ and $Y_{\bar{\nu}}$ denote the Bessel functions, and
$\bar{\nu}$ is related to the eigenvalue $\lambda^2_{\Ssp^{D-2}}$ as
\Eq{
\bar{\nu}^2=\lambda^2_{\Ssp^{D-2}}+\frac{\left(D-3\right)^2}{4}\,.
}
By using Eqs.~(\ref{sp:metric:Eq}), (\ref{sp:ansatz:Eq}),
and (\ref{sp:eigen:Eq}), the field equation for $\varphi_0$ becomes
\Eq{
\frac{d}{dt}\left(A t\,\frac{d\psi_0}{dt}\right)
+\left[\lambda_{\Zsp}^2
+\left(At\right)^{b_0}m^2\right]\varphi_0=0\,.
\label{sp:field2:Eq}
}
Let us consider the case of $m=0$\,.
The solution of $\varphi_0$ is given by the oscillating form
\Eq{
\varphi_0(t)=f_3J_0\left(2\lambda_{\Zsp}\sqrt{A^{-1}\,t}\right)
+f_4Y_0\left(2\lambda_{\Zsp}\sqrt{A^{-1}\,t}\right)\,,
\label{sp:solution:Eq}
}
where $f_3$ and $f_4$ are constants and $J_0$ and $Y_0$
are the Bessel functions.
The energy of the Klein-Gordon modes can be calculated by
\Eq{
E=-u^M\pd_M\varphi\,,~~~~~u=\alpha\pd_t+\beta\pd_r\,,
\label{sp:e:Eq}
}
where $u$ is velocity. Then, $\alpha$ and $\beta$ are determined by
\Eq{
-\alpha^2\left(A\,t\right)^{a_0}+\beta^2\left(A\,t\right)^{b_0}=-1\,,
\label{sp:ab:Eq}
}
where we used the normalization condition $u^2=-1$\,.
In the case of $t\rightarrow 0$ and $r\rightarrow\infty$,
$\alpha$ and $\beta$ must behave as
\Eq{
\alpha\sim \left(A\,t\right)^{-a_0/2}\,,~~~~~
\beta\sim \left(A\,t\right)^{-b_0/2}\,,
\label{sp:coe:Eq}
}
in order to remain nonsingular.
If we use the expression (\ref{sp:coe:Eq}), the energy of the scalar field
is given by
\Eq{
-E=\alpha\,\pd_t\varphi+\beta\,\pd_r\varphi\,.
\label{sp:e2:Eq}
}
For the asymptotic solution (\ref{sp:solution:Eq}),
one can note that the energy becomes $E\rightarrow\infty$
as the singularity is approached. Hence,
the energy-momentum tensor of the Klein-Gordon field mode diverges.
The classical solution gives the mode of the scalar field which cannot
stabilize the metric modes near the singularity.
\subsection{Intersection involving the $0-p_{I'}$-brane background
in the $D$-dimensional asymptotically de Sitter universe}
Finally, we discuss the stability analysis for the $0-p_{I'}$-brane
in the asymptotically de Sitter universe. If we set $\tilde{c}=0$
and take $z^a\rightarrow\infty$ in the
background (\ref{ds2:exact:Eq}), the $D$-dimensional metric becomes
\Eqrsubl{sd:metric:Eq}{
ds^2&=&-d\tau^2
+\left(c_0\,\e^{c_0\tau}\right)^{2/(D-3)}\delta_{ab}(\Zsp)dz^adz^b\,,\\
\delta_{ab}(\Zsp)dz^adz^b
&=&dr^2+r^2\bar{\omega}_{ij}(\Ssp^{D-2})d\xi^id\xi^j\,,
}
where $\bar{\omega}_{ij}(\Ssp^{D-2})$
denotes the metric of the $(D-2)$-dimensional sphere,
$c_0$ is given by (\ref{ds2:c0:Eq}), and the cosmic time $\tau$ is
defined by (\ref{ds2:cosmic:Eq}).
There is a curvature singularity at $\tau\rightarrow-\infty$
in the $D$-dimensional spacetime. In the following, we set $c_0>0$\,.
Otherwise, the scale factor of $D$-dimensional spacetime becomes
complex or negative.
We consider the Klein-Gordon field
to analyze the stability
\Eq{
-\frac{1}{\sqrt{-g}}\pd_M\left(\sqrt{-g}g^{MN}\pd_N\varphi\right)
+m^2\varphi=0\,,
\label{sd:KG:Eq}
}
where $g$ is the determinant of the six-dimensional metric
(\ref{sd:metric:Eq}).
Substituting the $D$-dimensional metric (\ref{sd:metric:Eq}) into
Eq.~(\ref{sd:KG:Eq}), we obtain
\Eq{
\left(c_0\,\e^{c_0\tau}\right)^{-1}
\pd_\tau\left[\left(c_0\,\e^{c_0\tau}\right)^{\frac{D-1}{D-3}}
\pd_\tau\varphi\right]
-r^{-(D-2)}\pd_r\left(r^{D-2}\pd_r\varphi\right)
-\frac{1}{r^2}\lap_{\Ssp^{D-2}}\varphi
+\left(c_0\,\e^{c_0\tau}\right)^{\frac{2}{D-3}}
m^2\varphi=0\,,
\label{sd:KG2:Eq}
}
where $\lap_{\Ssp^{D-2}}$ denotes the Laplace operator on the $\Ssp^{D-2}$\,.
We assume an ansatz for the scalar field $\varphi$\,:
\Eq{
\varphi=\varphi_0(\tau)\varphi_1(r)\varphi_2(\xi)\,,
\label{sd:ansatz:Eq}
}
where the functions $\varphi_1(r)$ and $\varphi_2(\xi)$ satisfy the
eigenvalue equation
\Eq{
\lap_\Zsp\varphi_1(r)\varphi_2(\xi)
=-{\lambda}_{\Zsp}^2\,\varphi_1(r)\varphi_2(\xi)\,,
\label{sd:eigen:Eq}
}
and obey the equations
\Eq{
\varphi_1(r)=\frac{1}{r}\left[b_5J_{\tilde{\nu}}(\lambda_{\Zsp} r)
+b_6Y_{\tilde{\nu}}(\lambda_{\Zsp} r)
\right],~~~~~
{\lap}_{\Ssp^{D-2}}\varphi_2(\xi)
=-{\lambda}_{\Ssp^{D-2}}^2\,\varphi_2(\xi)\,.
\label{sd:eigen2:Eq}
}
Here $\triangle_{\rm Z}$ is the Laplace operator on the Z space, and
$b_5$ and $b_6$ denote constants,
$J_{\tilde{\nu}}$ and $Y_{\tilde{\nu}}$ are the Bessel functions, and
$\tilde{\nu}$ is written by the eigenvalue $\lambda^2_{\Ssp^{D-2}}$ as
\Eq{
\tilde{\nu}^2=\lambda^2_{\Ssp^{D-2}}+\frac{\left(D-3\right)^2}{4}\,.
}
In terms of Eqs.~(\ref{sp:eigen:Eq}), (\ref{sd:metric:Eq}),
and (\ref{sd:ansatz:Eq}), the field equation for $\varphi_0$ becomes
\Eq{
\left(c_0\,\e^{c_0\tau}\right)^{-1}
\frac{d}{d\tau}\left[\left(c_0\,\e^{c_0\tau}\right)^{(D-1)/(D-3)}\,
\frac{d\varphi_0}{d\tau}\right]+\left[\lambda_{\Zsp}^2
+\left(c_0\,\e^{c_0\tau}\right)^{2/(D-3)}m^2\right]\varphi_0=0\,.
\label{sd:field:Eq}
}
Let us first consider the solution of $\varphi_0$ for $D=5$ and $D=6$\,.
In the case of $D=5$, the solution of $\varphi_0$ can be expressed as
\Eqr{
\varphi_0(\tau)&=&
\lambda_{\Zsp}^2\,c_0^{-3}\,\e^{-c_0\tau}\left[
f_5\,\Gamma\left(1-\ell_1\right)\,
J_{-{\ell_1}}
\left(2\lambda_{\Zsp}\,c_0^{-3/2}\,\e^{-c_0\tau/2}\right)\right.\nn\\
&&\left.+f_6\,\Gamma\left(1+\ell_1\right)\,
J_{\ell_1}\left(2\lambda_{\Zsp}\,c_0^{-3/2}\,\e^{-c_0\tau/2}\right)\right],
\label{sd:solution:Eq}
}
where $f_5$ and $f_6$ are constants, $J_{\ell_1}$ and $J_{-{\ell_1}}$
are the Bessel functions, and $\ell_1$ is defined by
\Eq{
\ell_1=2\sqrt{1-\left(\frac{m}{c_0}\right)^2}\,.
}
On the other hand, setting $D=6$\,,
we can also find the solution of $\varphi_0$\,:
\Eqr{
\varphi_0(\tau)&=&\frac{9}{4}\,\sqrt{\frac{3}{2}}\,
\lambda_{\Zsp}^{5/2}\,c_0^{-10/3}\,\e^{-5c_0\tau/6}\left[
f_7\,\Gamma\left(1-\ell_2\right)\,
J_{-\ell_2}\left(3\lambda_{\Zsp}\,c_0^{-4/3}\,
\e^{-c_0\tau/3}\right)\right.\nn\\
&&\left.+f_8\,\Gamma\left(1+\ell_2\right)\,
J_{\ell_2}\left(3\lambda_{\Zsp}\,c_0^{-4/3}\,\e^{-c_0\tau/3}\right)\right].
\label{sd:solution2:Eq}
}
Here $f_7$ and $f_8$ denote constants, $J_{\ell_2}$ and $J_{-{\ell_2}}$
are the Bessel functions, and the constant $\ell_2$ is given by
\Eq{
\ell_2=\frac{1}{2}\,\sqrt{25-36\left(\frac{m}{c_0}\right)^2}\,.
}
For $D\ge 4$, the solution of $\varphi_0$
can be written in the following form:
\Eqr{
\varphi_0(\tau)&=&\left(\frac{D-3}{2}\right)^{\frac{D-1}{2}}\,
\lambda_{\Zsp}^{\frac{D-1}{2}}\,c_0^{-\frac{(D-1)(D-2)}{2(D-3)}}\,
\e^{-\frac{D-1}{2(D-3)}c_0\tau}\nn\\
&&\times\left[
\bar{f}\,\Gamma\left(1-\ell\right)\,
J_{-\ell}\left((D-3)\,\lambda_{\Zsp}\,c_0^{-\frac{D-2}{D-3}}\,
\e^{-\frac{1}{D-3}c_0\tau}\right)\right.\nn\\
&&\left.+\tilde{f}\,\Gamma\left(1+\ell\right)\,
J_{\ell}\left((D-3)\,\lambda_{\Zsp}\,c_0^{-\frac{D-2}{D-3}}\,
\e^{-\frac{1}{D-3}c_0\tau}\right)\right],
\label{sd:solution-g:Eq}
}
where $\bar{f}$ and $\tilde{f}$ are constants, $J_{\ell}$ and $J_{-{\ell}}$
denote the Bessel functions, and $\ell$ is defined by
\Eq{
\ell=\frac{1}{2}\,\sqrt{(D-1)^2-4(D-3)^2\left(\frac{m}{c_0}\right)^2}\,.
}
Since the energy of the scalar field
can be written as (\ref{sp:e:Eq}), the energy of the Klein-Gordon modes can
be given by the expression (\ref{sp:e2:Eq})\,. Then, we can find
$\alpha$ and $\beta$ in this way:
\Eq{
-\alpha^2+\beta^2\left(c_0\,\e^{c_0\tau}\right)^{2/(D-3)}=-1\,,
\label{sd:ab:Eq}
}
where we used the normalization condition $u^2=-1$\,.
In the case of $\tau\rightarrow -\infty$ and $r\rightarrow\infty$,
$\alpha$ and $\beta$ have to behave as
\Eq{
\alpha\sim {\rm constant}\,,~~~~~
\beta\sim \left(c_0\,\e^{c_0\tau}\right)^{-1/(D-3)}\,.
\label{sd:coe:Eq}
}
For the solution (\ref{sd:solution-g:Eq})\,,
the energy becomes $E\rightarrow\infty$ in the limit
$\tau\rightarrow -\infty$\,.
Since the energy is not convergent with the asymptotic solution,
the mode of the scalar field does not stabilize the metric modes
near the singularity.
\section{Conclusion and Discussion}
\label{sec:cd}
In this paper, we have discussed the
time-dependent intersecting branes
with cosmological constants for not only the delocalized case but also the
partially localized one in $D$-dimensional gravitational theory.
We are everywhere brief, and on some points, we simply call attention
to questions that might be investigated in the future.
The function $h_{I}$ depends on time as well as the coordinate
of the relative and overall transverse spaces.
The coupling constants between the field strengths
and the dilaton are given by the assumptions (\ref{pl:ansatz:Eq}) or
(\ref{cn:fields:Eq}) and depend on the parameter $N$.
In the case of the eleven- or ten-dimensional supergravity
theory, the dilaton coupling requires $N=4$.
The power of the time dependence depends
on the number of the brane and total dimensions with
the parameter $N$ of the dilaton coupling constant.
An exceptional case arises if the parameter $N$ in the dilaton coupling
takes another value than 4. There are static solutions with $N\ne 4$
in the lower-dimensional supergravity theories as well as
Einstein-Maxwell theory \cite{Lu:1995cs, Cvetic:2000dm}.
In this case one gets asymptotically power-law expanding solutions
if the dilaton is nontrivial. For the trivial dilaton,
the Einstein equations give an asymptotically de Sitter solution
for a single 2-form field strength.
Since the cosmological constant is related to
the field strength,
the time derivative of the warp factor arises only
from the Ricci tensor and can be compensated
by the cosmological constant in the Einstein equations.
This is the same structure as in Refs.
\cite{Kastor:1992nn, Maeda:2010aj, Minamitsuji:2010kb} and
the generalization of the solutions
\cite{Kastor:1992nn, Maki:1992tq, Maki:1994wc}.
In $N=4$ case, the equation of motion in the presence of the
cosmological constant gives the static delocalized or partially localized
intersecting brane solution because of the ansatz of the fields.
Thus, one expects that the recipe for picking an accelerating expansion
from the dynamical intersecting brane
solutions depends on the dilaton coupling
constant, and this is the case for the proposal in Ref.~
\cite{Minamitsuji:2010kb}.
Once the de Sitter solution is obtained in the single $p$-brane solutions,
it is possible to apply it to the intersecting brane systems.
An immediate point is that the time-dependent solutions make dynamical
compactification more or less obvious,
since cosmological evolution is a general property of the solution
(with constant parameters) once the function $h_I$
is properly endowed with the time dependence.
The power of the scale factor in some solutions gives an accelerating
expansion law even in the case that functions $h_I$ depend on
both the time and coordinates of overall transverse space, while the
extra dimension will shrink as cosmic time increases.
However, something is still missing, because the scale factor of our
Universe diverges at $\tau=\tau_\infty$. At the moment, it is not
clear how to do this.
We have discussed the dynamics of the brane collisions.
As the spacetime is contracting in the $D$-dimensional spacetime,
each 0-brane approaches others as the time evolves for $\tau<0$
but separates for $\tau>0$ in the asymptotically
power-law expanding solutions. Thus 0-branes never collide.
In the case of asymptotically de Sitter solutions,
all domains between branes are connected at $\tau=0$ ($c_0<0$).
The domain shrinks as the time decreases, while
the proper distance becomes constant as $\tau$ increases.
For the $0-p_{I'}$-brane system ($p\le 7$),
a singularity appears before 0-branes collide, and eventually the
topology of the spacetime changes so that
branes are separated by singular hypersurfaces surrounding each brane
if branes are not smeared.
Thus, we cannot describe the collision of two 0-branes
in terms of these solutions.
On the other hand, the $0-8$-brane system in ten dimensions or
the smeared $0-p_{I}$-brane system in $D$-dimensional theory
can provide examples of colliding branes if they have the
same brane charges and only one overall transverse space.
We have analyzed the collision of the brane where
the $p_0-p_I$-branes are localized at the same position along
the overall transverse directions, in the case of equal charges.
The brane collision would not occur
if the brane charges are different.
Moreover, if these branes are localized at different positions,
it raises the possibility that the curvature singularities appear.
We have also studied the dynamics of the five- or
six-dimensional supergravity
models with applications to cosmology and collision of branes.
First we have discussed the brane solutions
to study the time evolution in the NSS model.
In the case of vanishing 3-form field strength in the five-dimensional
effective theory, the scale factor of our four-dimensional spacetime
is a linear function of the cosmic time which is the same evolution
as the Milne universe. On the other hand, for the dynamical
1-brane without 2-form field strength, the solution tells us
that the function $h$ depends on all the world-volume coordinates of
the 1-brane. Hence, the contribution of the field strength except for
the 2-form leads to an inhomogeneous universe.
We have investigated the dynamics of 0-branes and found that, when
the spacetime is contracting in six dimensions, each 0-brane approaches
the others as the time evolves. All domains between branes connected
initially ($t=0$), but it shrinks as $t$ increases.
However, for the 0-brane system without smearing branes, a singularity
appears before 0-branes collide, and eventually the topology of
the spacetime changes such that parts of the branes are separated
by a singular region surrounding each brane. Thus, the
solution cannot describe the collision of two 0-branes.
In contrast, the smeared 0-brane system with $d_s=4$ can provide
an example of colliding 0-branes and collision of the universes,
if they have the same brane charges.
We have next constructed the time-dependent 1-brane solution in the NSS
supergravty model. In the asymptotic far 1-brane region,
the 1-brane spacetime
in the NSS model approaches the six-dimensional Milne universe.
In regions close to the 1-branes, for concreteness, we have studied
the case of two 1-branes in detail. The 1-brane is approaching the
other as the time progress for $t<0$.
We have found that, in the case of $t<0$,
all of the domains between the 1-branes are initially
connected, but some region (near small $y$) shrinks as the time
increases, and eventually the topology of the spacetime changes such
that parts of the branes are separated by a singular region surrounding each
1-brane. Thus, in the case of $d_s\ne 3$, 1-branes never collide.
On the other hand, the case of $d_s=3$, for $t<0$, could provide an
example of colliding 1-branes. We found that the collision time
depends on both brane charges and the place in the world volume of
the 1-brane. Since this case has the time-reversal symmetry, the evolution
for $t>0$ is obtained by the time-reversal transformation.
We also investigated the time-dependent solution in the five-dimensional
supergravity model. The power of the scale factor is so small
that the solutions cannot give a realistic expansion law.
Then, it is necessary to include additional matter on the background
in order to obtain a realistic expanding universe.
We finally analyzed the classical instability of the dynamical
brane background towards singularity. In order to present the instability
of the dynamical brane background, we have estimated whether an
instability does exist by computing the energy of the Klein-Gordon
modes. One can find that the energy seen by an observer diverges
as the curvature singularity is approached. This implies that
the mode of the
scalar field is likely to destabilize the background metric modes
near the singularity. Although this result has been given by preliminary
analysis, it has made the property of singularity in the dynamical
brane background more clear. It is also necessary for us to perform a more
rigorous analysis by considering in detail the metric perturbation
whether the stability analysis arrives to the same conclusion or not.
A recent study of intersecting
systems depending on the time coordinate and overall transverse space
shows that all warp factors in the solutions can depend on time
\cite{Maeda:2012xb}.
It is interesting to study if similar more general solutions
can be obtained by relaxing some of our assumptions.
We hope to report on this subject in the near future elsewhere.
\section*{Acknowledgments}
We thank M. Minamitsuji for numerous valuable
discussions and careful reading of the manuscript and also
thank M. Cvetic and G.W. Gibbons for discussions and valuable comments.
|
\section{Introduction}
{\noindent}Quantum technologies exploit intense interactions between field and matter degrees of freedom~\cite{MonroeNATURE}, and it is a typical experimental goal in this context to maximize the coupling between the two. Traditional cavity QED setups have been extremely successful in this regard, yet they result in coupling frequencies that are only a tiny fraction of that of the system components~\cite{HarocheREV}. Experimental advances, for example in semiconductor microcavities and circuit QED, have now pushed the strength of light-matter interactions into the ultrastrong-coupling regime (USC)~\cite{USexp1, USexp2, USexp3, USexp4, USexp5}. This regime is characterized by the coupling frequency $\lambda$ being a non-negligible fraction of the bare frequency of the matter degree of freedom, say $\omega_b$. The theoretical description of the USC goes beyond the rotating wave approximation (RWA), demanding the inclusion in the Hamiltonian of terms that do not conserve the excitation numbers of the individual components --- the `counter-rotating terms' (CR) ~\cite{BraakPRL, CiutiPRA, SolanoPRL}. This regime has been studied extensively due to the lure of exotic phenomena such as the existence of virtual excitations in the ground state~\cite{CiutiPRA}, dynamical Casimir effects~\cite{DeLiberatoCASIMIR}, quantum phase transitions~\cite{CiutiNATURE, BrandesPRA}, and counter-intuitive radiation statistics~\cite{RidolfoPRL,Ridolfo2}.
{\noindent}In this regime, however, the sole inclusion of the CR terms may not be sufficient to correctly describe the new physics. Another important ingredient is the diamagnetic -- or `$A^2$' -- term, which is proportional to the square of the vector potential $\hat{\mathbf{A}}$ and ensures gauge invariance in the non-relativistic minimal coupling Hamiltonian~\cite{Woolley}. The effects associated with this term, and the related Thomas-Reiche-Kuhn (TRK) sum rules, are of crucial importance in the research on the `Dicke phase transition', and are still under active investigation and debate~\cite{ZakowiczPRL, RzazewskiPRA,Keeling, KnightPRA,BaumannNATURE,CiutiNATURE,ciuticritics,ciutireply,BirulaPRA,ChirolliPRL,DomokosPRL,BambaARXIV}. A further point deserving attention is that the two-level approximation, useful to simplify the description of quantum emitters, may fail in the USC~\cite{threelevel}. Finally, even the multi-mode nature of the cavity field is known to play a role in the `deep strong coupling' regime $\lambda\gtrsim\omega_b$. \cite{DeLiberatoPRL}.
{\noindent}In most of the above examples the physics beyond the CR terms, for example due to $A^2$, becomes crucial as the strength of light-matter interactions is pushed towards the extreme regime $\lambda\!\gtrsim\!\omega_b$. In contrast, to the best of our knowledge, clear-cut {\it qualitative} signatures of these extra terms have not been discussed in the currently experimentally relevant regime $\lambda/\omega_b\!\sim\!0.2$. The present work aims at giving a contribution in this direction. We begin by studying a common Hamiltonian model of light-matter interaction, in which a bosonic matter excitation is ultrastrongly coupled to a single-mode cavity field. We find that the $A^2$ term imposes an interesting constraint on the structure of the normal modes of the system, the upper and lower polariton \cite{HopfieldPHYSREV}. This implies that the bare vacuum of the matter and field modes, which is no longer the true ground state of the system, contains equal populations of the two polaritons. Interestingly, this observation is independent of the specific choice of the various model parameters, provided that they are chosen compatibly with the TRK sum rule. To test this finding, one needs to design an experiment that explicitly relies on the relationship between polaritons and bare modes. We show that a possible option is to perform a `quench' of the coupling, a rapid switch-on of $\lambda$ from an initially negligible value, followed by a spectral analysis of the resulting `quantum vacuum radiation' exiting the cavity (that is, radiation that is due to a non-adiabatic change of the ground state of the system) \cite{DeLiberatoCASIMIR}.\\
In the second part of the paper, we investigate the robustness of the considered model by interpreting it as a low-energy approximation to a Coulomb gauge minimal coupling Hamiltonian. This allows us to clarify the role of the TRK sum rule in the considered system, as well as to identify some extra terms -- besides $A^2$ -- that one may need to include in the effective low-energy Hamiltonian to accurately model the USC. In a nutshell, one should include an effective self-interaction term for the matter, mediated by the higher cavity harmonics, plus a term describing the electrostatic interaction between the dipole and its `images' on the cavity walls \cite{DomokosPRL}. While the remarkable symmetry between the two populations is in general lost, we are able to gain an analytical understanding of this more complete model and the consequences of the new terms.
{\noindent}The paper is organized as follows. In section~\ref{basic} we discuss our main result in its simplest form, by analyzing the effect of an $A^2$-like term on a common model of coupled oscillators. Section~\ref{quench} illustrates the quenching experiment that allows to investigate the relationship between bare modes and polaritons. In Section~\ref{micro} we illustrate the microscopic model that is assumed to underlie our theory, clarifying the role of the TRK sum rule and deriving a refined effective Hamiltonian for the two modes of interest. In section~\ref{Dicke} we briefly discuss the extension of our results to Dicke-like models, and in section~\ref{fine} we draw our conclusions. Some additional technical details and derivations are provided in three appendixes.
\section{Basic Model}\label{basic}
{\noindent}We start with a common effective model of light-matter interaction, featuring a photonic mode $\hat a$ of bare frequency $\omega_a$ (cavity mode for brevity) coupled to a bosonic matter mode $\hat b$ of bare frequency $\omega_b$. The latter can be thought of as a quantized oscillating dipole. The Hamiltonian reads ($\hbar\! =\! 1$)
\begin{align}
H &= \omega_a\hat a^{\dagger}\hat a\!+\!\omega_b\hat b^{\dagger}\hat b\!+\!\lambda(\hat a\!+\!\hat a^{\dagger})(\hat b\!+\!\hat b^{\dagger}) \!+\! D(\hat a \!+ \!\hat a^{\dagger})^2\label{Hsys},
\end{align}
where $\lambda$ quantifies the light-matter coupling strength, while the contribution proportional to $D$ is due to the $A^2$ term. As shown in section \ref{micro} below, the TRK sum rule imposes $D\!=\!\lambda^2/\omega_b$~\cite{sumrule}. Nevertheless, we shall keep $D$ implicit for later convenience. Being bilinear in the bosonic operators, Hamiltonian \eqref{Hsys} can be diagonalized exactly. Following Hopfield, we shall refer to the normal modes of the system as the upper (U) and lower (L) polariton \cite{HopfieldPHYSREV}, with associated bosonic operators $\hat p_U$, $\hat p_L$ and eigenfrequencies $\omega_U>\omega_L$. In terms of the normal modes, the Hamiltonian assumes the simple form
\begin{equation}
H=\omega_U\hat p_{U}^\dagger\hat p_U+\omega_L\hat p_L^\dagger\hat p_L,
\end{equation}
up to a constant. Explicit expressions for the eigenfrequencies $\omega_U,\omega_L$ and the polaritonic operators will be shown in the subsection below.
\begin{figure}[t!]
\begin{center}\includegraphics[width=.98\linewidth]{FIG1}
\end{center}
\caption{(Color online). Mean polaritonic populations $n_U,n_L$ in the bare ground state $\ket 0$. ({\bf a}): Mean populations versus normalized coupling strength at bare mode resonance $\omega_a\!=\!\omega_b$. ({\bf b}): Mean populations versus bare detuning $\Delta \!=\! \omega_a\!-\!\omega_b$, fixing $\lambda \!=\! 0.2\omega_b$. In all plots, the blue solid line refers to $D\!=\! \lambda^2/\omega_b$, while the red (dotted, dot-dashed) lines to $D\!=\!0$.
\label{fig:Energy}}
\end{figure}
\subsection{A simple signature of the $A^2$ term}\label{signature}
{\noindent}To investigate the impact of $A^2$ on the physics of our system, we shall study in detail the relationship between the bare modes $\hat a,\hat b$ and the polaritons $\hat p_k$, where $k\in\{U,L\}$ hereafter. We start by noting that the bare modes vacuum $\ket0$, defined by $\hat a\ket0=\hat b\ket0=0$, does not in general coincide with the ground state of the polaritons: $\hat p_k\ket0\neq0$. We thus turn our attention to the mean populations
\begin{equation}
n_k\!\equiv\!\bra0\hat{p}_k^{\dagger}\hat{p}_k\ket0,\label{nk}
\end{equation}
whose nonzero value is perhaps the simplest signature of the USC. Fig.~\ref{fig:Energy} illustrates the behaviour of $n_k$ as a function of the coupling strength $\lambda$, the bare frequency difference $\Delta \!\equiv\!\omega_a\!-\omega_b$ and, most importantly, the parameter $D$. When the TRK value $D \!=\! {\lambda^2}/{\omega_b}$ is taken, we observe that the excitations are distributed equally between $\hat{p}_U$ and $\hat{p}_L$. Setting instead $D\!=\!0$, which corresponds to neglecting $A^2$, predicts a significantly higher population for the lower frequency mode $\hat p_L$. We note that this holds in all the explored range of the remaining parameters $\lambda,\omega_a,\omega_b$.
{\noindent}This finding can be confirmed analytically. We present here a derivation outlined by an anonymous referee and which is particularly transparent. To find the normal modes of the Hamiltonian, we write $H \!=\!\tfrac12\hat{\boldsymbol{a}}^\dagger\mathbf{H}\hat{\boldsymbol{a}}$ (up to a constant), where $\hat{\boldsymbol{a}}=(\hat a,\hat b,\hat a^\dagger,\hat b^\dagger)$ and $\mathbf{H}$ is a positive matrix that can be easily inferred from Eq.~\eqref{Hsys}. By Williamson's theorem we have $\mathbf{H}=\mathbf{S}^\dagger{\sf diag}(\omega_U,\omega_L,\omega_U,\omega_L)\mathbf{S}$, where $\mathbf{S}$ is a symplectic matrix~\cite{Williamson}. Hence the polaritonic modes $\hat{\boldsymbol{p}}\!=\!({\hat p}_U,{\hat p}_L,{\hat p}_U^\dagger,{\hat p}_L^\dagger)$ are given by $\hat{\boldsymbol{p}}=\mathbf{S}\hat{\boldsymbol a}\label{polaritons}$, and the bosonic commutation relations $[\hat p_k,\hat p^\dagger_{k'}]=\delta_{kk'}$ are guaranteed by construction. For our specific system, we get (see appendix~\ref{diagonalizza})
\begin{align}
\!\!\!&\omega_{U,L}^2\!=\!\tfrac{{\omega}_a^2+4D\omega_a+{\omega}_b^2}{2}\!\pm\!\sqrt{\left(\tfrac{{\omega}_a^2+4D\omega_a-{\omega}_b^2}{2}\right)^{\!2}\!+\!4\lambda^{2}\omega_a\omega_b},\label{freqs}\\
\hat p_{U}\!&=\!\cos\theta[\mu(\!\tfrac{\omega_U}{\omega_a}\!)\hat a\!+\!\nu(\!\tfrac{\omega_U}{\omega_a}\!)\hat a^\dagger]\!-\!\sin\theta[\mu(\!\tfrac{\omega_U}{\omega_b}\!)\hat b\!+\!\nu(\!\tfrac{\omega_U}{\omega_b}\!)\hat b^\dagger],\label{pU}\\
\!\!\hat p_{L}\!&=\!\sin\theta[\mu(\!\tfrac{\omega_L}{\omega_a}\!)\hat a\!+\!\nu(\!\tfrac{\omega_L}{\omega_a}\!)\hat a^\dagger]\!+\!\cos\theta[\mu(\!\tfrac{\omega_L}{\omega_b}\!)\hat b\!+\!\nu(\!\tfrac{\omega_L}{\omega_b}\!)\hat b^\dagger],\label{pL}
\end{align}
where $\mu(x)\!\equiv\!\tfrac12(\sqrt x\!+\!1/\sqrt x)$, $\nu(x)\!\equiv\!\tfrac12(\sqrt x\!-\!1/\sqrt x)$ and $\theta$ is defined by $\cos2\theta\!\equiv\!({{\omega}_a^2\!+\!4D\omega_a\!-\!{\omega}_b^2})/{(\omega_U^2\!-\!\omega_L^2})$; $\theta\!<\!0$. These choices are consistent with the ordering $\omega_U\!>\!\omega_L$. It is easy to check that equations \eqref{pU} and \eqref{pL}, together with their Hermitian conjugates, implicitly define a symplectic matrix ${\mathbf S}$ in accordance with the general discussion above. We can now evaluate $n_k$ by substituting Eqs.~(\ref{pU},\ref{pL}) in Eq.\eqref{nk}, obtaining
\begin{align}
n_U&=\tfrac14\cos^2\theta\,\big(\tfrac{\omega_U}{\omega_a}+\tfrac{\omega_a}{\omega_U}\big)+\tfrac14\sin^2\theta\,\big(\tfrac{\omega_U}{\omega_b}+\tfrac{\omega_b}{\omega_U}\big)-\tfrac12,\label{nU}\\
n_L&=\tfrac14\sin^2\theta\,\big(\tfrac{\omega_L}{\omega_a}+\tfrac{\omega_a}{\omega_L}\big)+\tfrac14\cos^2\theta\,\big(\tfrac{\omega_L}{\omega_b}+\tfrac{\omega_b}{\omega_L}\big)-\tfrac12.\label{nL}
\end{align}
We further notice that the product of the polaritonic frequencies obeys the equation
\begin{align}
{\omega_U\omega_L}\!=\!{\omega_a\omega_b}\sqrt{1\!+\!\tfrac{4}{\omega_a}\big(D\!-\!\tfrac{\lambda^2}{\omega_b}\big)},\label{product}
\end{align}
Choosing the TRK value $D\!=\!\lambda^2/\omega_b$, Eq.~\eqref{product} reduces to $\omega_U\omega_L\!=\!\omega_a\omega_b$, which can be rearranged as ${\omega_U}/{\omega_a}\!\!=\!\!{\omega_b}/{\omega_L}$ and ${\omega_U}/{\omega_b}\!\!=\!\!{\omega_a}/{\omega_L}$. This is easily shown to yield $n_U\!=\!n_L$ via Eqs.~\eqref{nU} and \eqref{nL}. Taking a step further, we find for generic $D\!>\!0$ that the sign of $n_U\!-\!n_L$ is always the same as that of $D\!-\!\lambda^2/\omega_b$ in a broad range of parameters (see appendix~\ref{popstudy}). We thus have a sufficiently general scenario in which the two populations are equal if and only if $D$ assumes the appropriate TRK value, regardless of the specific arrangement of the remaining model parameters. Note that the equality $n_U\!=\!n_L$ can also be stated as a constraint on the matrix elements of ${\mathbf{S}}$, without making reference to a particular quantum state of the system. This simple and yet striking signature of the $A^2$ term on the structure of the polaritons constitutes our main result.
{\noindent}In passing we mention that, in the case of many matter modes interacting with the same single-mode field, a relationship analogue to Eq.~\eqref{product} was derived, implying that the product of polaritonic frequencies equals that of the bare frequencies under the TRK rule \cite{ciuticritics}. It will be interesting to investigate what constraints this may pose to the behaviour of these more general systems.
\section{Detecting $A^2$ via vacuum emission}\label{quench}
{\noindent}In principle, the relationship between bare and polaritonic modes could be investigated via a quenching experiment. The idea is to `switch on' the coupling $\lambda$ and the associated parameter $D$ non-adiabatically, starting from an initially negligible value. If the modulation is applied fast enough, the system remains in its initial state: at sufficiently low temperatures we could assume it to be the bare vacuum $\ket 0$. Since this is no longer the ground state of the Hamiltonian for $t\!>\!0$, the system will relax towards the vacuum of the polaritons, and to do so it must radiate photons outside the cavity. This process is a particular instance of quantum vacuum radiation \cite{DeLiberatoCASIMIR}. In absence of other relaxation mechanisms, we expect $n_k$ photons to be emitted at each frequency $\omega_k$ (on average), so that a simple spectral analysis of the cavity output field could be used to test the equality $n_U\!=\!n_L$ -- see Fig.~\ref{fig:concept}. This intuition is substantiated by the more quantitative discussion below. Before proceeding, we note that the non-adiabatic modulation of light-matter interactions has been experimentally demonstrated in solid state setups, by inducing a fast change in the density of the available charge carriers and hence in the relevant dipole moment matrix elements~\cite{Quench1, Quench2, Quench3}.
{\noindent}To model the radiative relaxation of the system following the quench, we couple the cavity to a continuum of external modes $\hat \alpha_\omega$ -- with $[\hat\alpha_\omega,\hat\alpha_{\omega'}^\dagger]=\delta(\omega\!-\!\omega')$. For simplicity we neglect matter losses and assume that all modes $\hat \alpha_\omega$ are accessible for measurement. The total Hamiltonian is now
\begin{equation}
H_{\sf tot} =H + H_{\sf ext}+H_I,
\end{equation}
where $H$ is given by Eq.~\eqref{Hsys}, while $H_{\sf ext}\!=\!\!\int d\omega\,\omega\,\hat \alpha_{\omega}^{\dagger}\hat \alpha_{\omega}$ and $H_I\!=\!\! \int d\omega\, J(\omega)(\hat a+\hat a^{\dagger})(\hat \alpha_{\omega} \!+ \hat \alpha^{\dagger}_{\omega})$ model the free evolution of the external modes and their coupling to the cavity. In the USC, the open dynamics of the system is better described in terms of polaritons. We thus recast $H_{\sf tot}$ in terms of the operators $\hat p_k$, and we assume that the coupling $J(\omega)$ is weak enough for us to perform a RWA in the interaction term $H_{I}$. We obtain
\begin{align}
\!\!H_{\sf tot} \simeq& \!\sum_{k=L, U}\!\!\omega_k\hat p^{\dagger}_k\hat p_k\!+H_{\sf ext}+\!\int\! \!{\rm d}\omega\!\! \sum_{k=L, U}\![J_k(\omega)\hat \alpha^{\dagger}_{\omega}\hat p_k \!+\! {\sf H.c.}]\label{RWA}
\end{align}
where $J_{U}(\omega)\!=\!J(\omega)\!\cos\theta\sqrt{\omega_a/\omega_U}$, $J_{L}(\omega)\!=\!J(\omega)\!\sin\theta\sqrt{\omega_a/\omega_L}$. Note that the RWA must be performed in the polaritonic basis~\cite{Bambareply,BreuerBOOK,ZollerBOOK,Bamba2}, since it is the operators $\hat p_{U},\hat p_L$ that oscillate harmonically in the interaction picture. As we are investigating photon emission in a non-stationary regime, we aim to determine the statistics of the external field modes as a function of the system conditions immediately after the quench. It is convenient to do so by a somewhat unusual application of the Heisenberg equations of motion. We note that the {\it initial} system operators can be expressed as a linear combination of polaritons and external modes {\it at any later time}:
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{FIG2.pdf}
\caption{(Color online). (a) The coupling between a dipole and a cavity field is suddenly switched on. The system relaxes to the ground state by radiating into the output modes $\hat f_U,\hat f_L$.
(b) Frequency distribution of the output modes (arbitrary units). We have fixed $\omega_a\!=\!\omega_b$, $\lambda\!=\!0.1\omega_b$, $D\!=\!\lambda^2/\omega_b$ in Eq.~\eqref{Hsys}, and a frequency-independent coupling to the external modes $J(\omega)\!=\!\sqrt{\gamma/2\pi}$, where $\gamma=0.01\omega_a$ would be the decay rate of the cavity in absence of the matter mode.}
\label{fig:concept}
\end{figure}
\begin{equation}
\hat p_k(0)=\sum_{k'}v_{kk'}(t)\hat p_{k'}(t)+\int{\rm d}\omega\, \phi_k(\omega,t) \hat \alpha_\omega(t),\label{linear}
\end{equation}
{\noindent}where the functions $v_{kk'}(t)$ and $\phi_k(\omega,t)$
can be determined as follows. Since the total time derivative must vanish on both sides, the differential equations $\dot v_{kk'}\!=\!i\omega_{k'} v_{kk'}\!+\!i\int{\rm d}\omega\,J_k\phi_k$ and $\partial_t\phi_k\!=\!i\omega\phi_k\!+\!\sum_{k'}J_{k'}v_{kk'}$ must hold, with initial conditions $v_{kk'}(0)\!=\!\delta_{kk'},\phi_k(\omega,0)\!=\!0$. The preservation of commutation relations imposes the normalization $ \sum_l v_{kl}v^*_{k'l}\!+\!\int{\rm d}\omega\,\phi_k\phi_{k'}^*\!=\!\delta_{kk'}$ at all times. For a given form of $J(\omega)$, $v_{kk'}$ and $\phi_k$ could be calculated in principle, e.g. numerically, by Fano-like techniques or Laplace transforms \cite{FanoREV,barnettbook}. Such details, however, are largely unimportant for our purposes. In standard scenarios, Eq.~\eqref{RWA} will induce a dissipative dynamics of the polaritonic system, such that one has ${v}_{kk'}\!\to\!0$ for sufficiently long times, and the full quantum statistics of the initial system state will be retrieved in specific combinations of the external field modes. These can be formally expressed as
\begin{equation}
\hat f_k\equiv\lim_{t\to\infty}\int{\rm d}\omega\, \phi_k(\omega,t) \hat \alpha_\omega(t)=\hat p_k(0).\label{outputmodes}
\end{equation}
By looking at the output modes $\hat f_k$, we can thus access {\it the full quantum statistics} of the polaritons immediately after the quench: the mean populations are for example given by $n_k\!=\!\langle \hat f_k^\dagger\hat f_k\rangle$. We note that the main message expressed by Eq.~\eqref{outputmodes} does not depend on the details of the interaction between cavity and external fields, but each asymptotic amplitude $\tilde\phi_k(\omega)\!\equiv\!\lim_{t\to\infty}\phi_k(\omega,t){\rm e}^{i\omega t}$ does, and needs to be evaluated on a case-by-case basis. Typically, $|\tilde\phi_k(\omega)|^2$ is sharply peaked around the corresponding polaritonic frequency $\omega_k$, and $\hat f_U$ and $\hat f_L$ can be spectrally resolved (an equivalent statement is that the timescales of emission are long as compared to $\omega_k^{-1}$). As an example, in Fig.~\ref{fig:concept} we plot $|\tilde\phi_{k}|^2$ for the simplest case of a frequency-independent coupling to the continuum; we can expect qualitatively similar results when considering more realistic profiles for $J(\omega)$. We remark that the neglect of losses and thermal noise allowed us to derive particularly straightforward relationships between intra- and extra-cavity observables. Still we can expect that, in a realistic system, Eq.~\eqref{outputmodes} can hold to a good approximation if the emission of detectable photons is the dominant relaxation process of the system. A quantitative study of these issues in lossy systems will be presented in future work.
\section{Microscopic model}\label{micro}
{\noindent}In this section we investigate the validity of Hamiltonian \eqref{Hsys} as a low-energy approximation to a more complete microscopic model. This gives us the opportunity to clarify the role of the TRK sum rule in our system, and to discuss some of the possible refinements of our basic model. In particular, we shall investigate the role of the following contributions to the matter-field interaction: (i) higher harmonics of the cavity field; (ii) the multimode nature of matter excitations; (iii) the electrostatic interaction between the dipole moment of the matter mode and the cavity boundaries. The derivations below are based on the assumption that matter excitations are well described by a collection of quantized harmonic oscillators, in analogy to the Hopfield model \cite{HopfieldPHYSREV}. Our calculations could also be applied to Dicke-like models in the Holstein-Primakoff regime \cite{Holstein,Holstein-multi,CiutiNATURE}, although in that case we are unable to fully take into account the electrostatic dipole-dipole interactions between different atoms. Only the spatially homogeneous contribution of these interactions can be included in our model, by appropriately rescaling the parameter $u$ defined below \cite{Keeling}.
\subsection{The minimal coupling Hamiltonian}
{\noindent}We assume that our matter mode can be microscopically described as a collection of non-relativistic particles of mass $m_j$ and charge $q_j$, subject to a potential $V$ that includes trapping forces as well as inter-particle interactions (in absence of the cavity). The interaction with the electromagnetic field in the cavity is modeled via a minimal coupling Hamiltonian in the Coulomb gauge, as per
\begin{equation}
H_{\sf mic}=\sum_{j}\frac{(\hat{\mathbf{p}}_j-q_j\hat{\mathbf{A}})^2}{2 m_j}+V(\hat{\mathbf{x}}_1,\hat{\mathbf{x}}_2,...)+V_{\sf img}+H_{\sf EM},\label{Hmic}
\end{equation}
where $\hat{\mathbf{p}}_j$ is the momentum of the $j-$th particle, $\hat{\mathbf{x}}_j$ its position, $\hat{\mathbf{A}}$ is the vector potential operator, $V_{\sf img}$ is the electrostatic interaction between matter and cavity walls \cite{DomokosPRL}, and $H_{\sf EM}$ is the free Hamiltonian of the field. We adopt the dipole approximation: the effective linear size of our emitter is assumed to be much smaller than the wavelength of light under consideration, hence the spatial dependence of $\hat{\mathbf{A}}$ across the emitter is neglected. In a similar spirit, we shall assume that $V_{\sf img}$ depends only on the total dipole moment of the matter excitations (in simple geometries, it can thus be calculated with the {\it method of images}). Since the components of $\hat{\mathbf{A}}$ commute with all particle operators, we can expand the Hamiltonian as
\begin{align}
H_{\sf mic}&=H_{\sf mic}^0+H_{\sf int}+H_{\sf EM},\\
H_{\sf mic}^0&=\sum_{j}\frac{\hat{\mathbf{p}}^2_j}{2 m_j}+\hat V,\\
H_{\sf int}&=-\sum_j\frac{q_j\hat{\mathbf{p}}_j\cdot\hat{\mathbf{A}}}{m_j}+\sum_j\frac{q_j^2}{2m_j}\hat{\mathbf{A}}^2+V_{\sf img},
\end{align}
Note that $H_{\sf int}$ includes all the Hamiltonian terms that would be suddenly switched on in the quenching experiment described earlier: indeed, all these terms depend on the effective dipole moment of matter. It is now useful to define
\begin{align}
\hat{\mathbf{d}}&\equiv\sum_j q_j\hat{\mathbf{x}}_j,\\
\hat{\mathbf{j}}&\!\equiv\!\sum_jq_j\frac{\hat{\mathbf{p}}_j}{m_j},
\end{align}
where $\hat{\mathbf{d}}$ is the {\it electric dipole operator}, while $\hat{\mathbf{j}}$ resembles a current operator (note however that $\hat{\mathbf{p}}_j$ is the canonical momentum, not the kinetic one). We can thus rewrite the microscopic Hamiltonian as
\begin{align}
H_{\sf mic}=H_{\sf mic}^0-\hat{\mathbf{j}}\cdot\hat{\mathbf{A}}+\sum_j\frac{q_j^2}{2m_j}\hat{\mathbf{A}}^2+V_{\sf img}+H_{\sf EM}\label{Hmic1}.
\end{align}
We now recall the TRK sum rule. Let $H_{\sf mic}^0$ have a complete set of eigenstates $\ket{E_n}$ with associated eigenvalues $E_n$ (these would be the eigenstates of matter in absence of interaction with radiation). We recall that the completeness relation $\sum_n\proj{E_n}=\mathbb I$ holds in the Hilbert space of the matter degrees of freedom, and we set the energy of the bare ground state $\ket {E_0}$ to zero for convenience. Exploiting the commutation relations
\begin{align}
[H_{\sf mic}^0,\hat{\mathbf{d}}]&=-i\hat{\mathbf{j}},\label{Hdj}\\
[\hat d_{k},\hat j_l]&=i\delta_{kl}\sum_jq_j^2/m_j,
\end{align}
we can derive the equality \cite{sumrule}:
\begin{align}
\sum_n\frac{\bra{E_0}\hat{\mathbf{j}}\cdot\hat{\mathbf{A}}\kebra{E_n}{E_n}\hat{\mathbf{j}}\cdot\hat{\mathbf{A}}\ket{E_0}}{E_n}=\sum_j\frac{q_j^2}{2m_j}\hat{\mathbf{A}}^2. \label{TRK}
\end{align}
Eq.~\eqref{TRK} is a possible formulation of the TRK sum rule for the ground state. Note that it is an equality for {\it field operators}, since the matrix elements are only taken in the Hilbert space of the matter degrees of freedom. We anticipate that from Eq.~\eqref{TRK} it is possible to derive the crucial equality $D=\lambda^2/\omega_b$ by a somewhat crude two-mode approximation, in which one substitutes $\hat{\mathbf{j}}\simeq-\mathbf{j}_1(\hat b+\hat b^\dagger)$ and $\hat {\mathbf A}\simeq\mathbf A_1(\hat a+\hat a^\dagger)$ ($\mathbf{j}_1$ and $\mathbf A_1$ are constant vectors of the appropriate units -- see below). Consistency with Eq.~\eqref{TRK} then implies ${(\mathbf{j}_1\cdot\mathbf A_1)^2}/{\omega_b}=\mathbf A_1^2\sum_j{q_j^2}/{2m_j}$. Finally, the equality of interest is obtained if one notices that the relevant coupling constants, in our notation, are given by $\lambda=\mathbf{j}_1\cdot\mathbf A_1$ and $D=\mathbf A_1^2\sum_jq_j^2/2m_j$.
{\noindent}In what follows, we shall show that the above reasoning is indeed correct under certain approximations. Inspired by the Hopfield model \cite{HopfieldPHYSREV}, we now assume that the matter degrees of freedom are well-approximated by a collection of harmonic excitations $\hat b_l$ of frequency $\omega_{b,l}$, and that the relevant matter operators can be expanded as
\begin{align}
H_{\sf mic}^0&\simeq\sum_l\omega_{b,l}\hat b_l^\dagger\hat b_l,\label{H0mic}\\
\hat{\mathbf{j}}&\simeq-\sum_l\mathbf{j}_l(\hat b_l+\hat b_l^\dagger),\label{jmic}\\
\hat{\mathbf{d}}&\simeq-i\sum_l\frac{\mathbf{j}_l}{\omega_{b,l}}(\hat b_l-\hat b_l^\dagger)\label{dmic},
\end{align}
where the constant vectors $\mathbf{j}_l$ encode information about the amplitude and polarization of matter excitations, and we have maintained consistency with Eq.~\eqref{Hdj}. The meaning of the approximation signs in Eqs.~\eqref{H0mic}-\eqref{dmic} is discussed in more detail in appendix~\ref{appb}; the bottom line is that the additional excitations of matter not captured by the modes $\hat b_l$ can be adiabatically eliminated, and their contribution drops out from both sides of the equal sign in the TRK sum rule \eqref{TRK}. We can now expand the vector potential of the field, at the location of the matter mode, as
\begin{align}
\hat{\mathbf{A}}\!=\!\sum_k\mathbf{A}_k(\hat a_k\!+\!\hat a_k^\dagger),
\end{align}
where each $\mathbf{A}_k$ is a constant vector characterising the single-photon field amplitude and polarization of the $k$-th cavity mode, with associated bosonic annihilation operator $\hat a_k$. The full Hamiltonian \eqref{Hmic1} can thus be recast as
\begin{align}
H_{\sf mic} & \!=\! \sum_l\omega_{b,l}\hat b_l^\dagger \hat b_l\!+\!\!\sum_{k}\!\omega_{a,k}\hat a_k^\dagger\hat a_k\!+\!\sum_{k,l}\! \lambda_{l,k}(\hat b_l\!+\!\hat b_l^\dagger)(\hat a_k\!+\!\hat a_k^\dagger)\nonumber\\
&+\sum_{k,n}\!D_{kn}(\hat a_k\!+\!\hat a_k^\dagger)(\hat a_n\!+\!\hat a_n^\dagger)+V_{\sf img},
\label{Hsys1}
\end{align}
where $\omega_{a,k}$ is the bare frequency of the $k$-th cavity mode, $\lambda_{lk}=\mathbf{j}_l\cdot\mathbf{A}_k$ quantifies the coupling strength between the $l$-th matter excitation and $k$-th cavity mode, and consistency with the sum rule in Eq.~\eqref{TRK} fixes $D_{kn}\!=\!\sum_l\lambda_{l,k}\lambda_{l,n}/\omega_{b,l}$ ~\cite{sumrule}. Note that we have not specified yet the electrostatic contribution $V_{\sf img}$: here we shall not attempt to study the structure of this term from first principles, rather we will assume that it is a quadratic function of the dipole moment $\hat{\mathbf{d}}$; this corresponds to the assumption that the induced charge densities on the cavity walls will depend linearly on the dipole moment components.
\subsection{Reduction to a two-mode model}
{\noindent}To recover a simpler Hamiltonian resembling Eq.~\eqref{Hsys}, we shall now assume that the lowest-frequency modes $\hat a\!\equiv\!\hat a_1$ and $\hat b\!\equiv\!\hat b_1$ are dominant in the interaction. Intuitively, this should hold when the two frequencies $\omega_a\!\equiv\!\omega_{a,1}$ and $\omega_b\!\equiv\!\omega_{b,1}$ are of the same order, the coupling $\lambda\!\equiv\!\lambda_{1,1}$ is not too large (as compared to $\omega_a,\omega_b$), and the TRK sum rule is approximately saturated by the considered transition: $D_{1,1}\!\sim\!\lambda^2/\omega_{b}$ (this also implies that $\mathbf{j}_1$ and $\mathbf{A}_1$ should be approximately parallel to each other). All the remaining parameters should conspire in such a way that the other light and matter modes will either stay close to their ground state in the dynamics of interest, or they will decouple from $\hat a,\hat b$ (e.g. by featuring polarizations orthogonal to both $\mathbf{j}_1$ and $\mathbf{A}_1$). Under these conditions we can obtain an effective Hamiltonian for the two modes $\hat a,\hat b$ by adiabatically eliminating all other modes from Eq.~\eqref{Hsys1}. Following standard procedures to eliminate weakly coupled excitations (See Refs.~\cite{james,sorensen} and appendix~\ref{appb}) we thus obtain the effective Hamiltonian
\begin{align}
H_{\sf eff}&= \omega_a\hat a^{\dagger}\hat a\!+\!\omega_b\hat b^{\dagger}\hat b\!+\!\lambda(\hat a\!+\!\hat a^{\dagger})(\hat b\!+\!\hat b^{\dagger}) \!+\! D(\hat a \!+ \!\hat a^{\dagger})^2\nonumber\\
&-\eta(\hat b\!+\!\hat b^\dagger)^2+u[i(\hat b\!-\!\hat b^\dagger)]^2
\label{adiabatic}
\end{align}
where $\eta\!\equiv\!\sum_{k>1}{\lambda_{1,k}^2}/{\omega_{a,k}}$, $u$ is a phenomenological coupling parameter arising from $V_{\sf img}$ (see below), and we have kept only second order terms in the couplings $\lambda_{l,k}$ with $l>1$. We observe that the elimination of the off-resonant matter modes induces the rescaling $D_{1,1}\!\to\! D_{1,1}\!-\!\sum_{l>1}\lambda_{l,1}^2/\omega_{b,l}$, hence retrieving the same value $D$ that we used in Hamiltonian \eqref{Hsys}. This suggests that one should {\it not} include the contributions of neglected transitions in the $A^2$ term, and justifies our slight abuse in terminology in referring to $D=\lambda^2/\omega_b$ as the `TRK sum rule'. The terms proportional to $\eta$ and $u$ are qualitatively new contributions that were not present in Eq.~\eqref{Hsys}. While the coefficient $\eta$ depends on the specific cavity structure, it is in general positive and of second order in the coupling, such that it may not be negligible with respect to the other terms. We emphasize that the $\eta$-term has {\it not} been obtained through a canonical transformation of $H_{\sf mic}$, even though its form may be reminiscent of the ``$P^2$ term" arising in the Power-Zienau-Woolley representation \cite{Cohen}. Finally, the term proportional to $u$ is simply the contribution of the matter mode $\hat b$ to the electrostatic energy $V_{\sf img}$ (recall that we are assuming $V_{\sf img}$ to be quadratic in the dipole moment -- see also appendix~\ref{appb}). Since we are not aware of a general method to determine the parameter $u$ as a function of the others, as we did for example with $D$, we shall study its effect as it is varied in the range $u\in[-D,D]$.
\subsection{Reliability of the effective Hamiltonian}
{\noindent}To confirm the validity of the effective Hamiltonian \eqref{adiabatic}, we compare its predictions to those of Eq.~\eqref{Hsys1} in a concrete example. For definiteness we assume that all the vectors $\mathbf{j}_l$ and $\mathbf{A}_k$ characterizing the modes of interest lie along the same axis, and we fix the structure of cavity modes and matter excitations such that $\omega_{a,k}\!=\!(2k\!-\!1)\omega_a$, $\omega_{b,j}\!=\!\tfrac13(4j^2\!-\!1)\omega_b$ and $\lambda_{j,k}\!=\!\lambda \tfrac{3j}{(4j^2\!-\!1)\sqrt{2k\!-\!1}}$, mimicking the relevant frequencies and coupling constants for a deep rectangular well placed in the middle of a Fabry-Perot resonator (with the important difference that, for us, each matter excitation is associated with a different harmonic oscillator). As a result we obtain $\eta\simeq0.23\lambda^2/\omega_a$. For the purposes of this section we simply take $u=0$, as our primary objective is to show that the introduction of the $\eta$ term and the rescaling of the $A^2$ term capture well the effect of higher-frequency cavity modes and matter excitations.
Fig.~\ref{fig:More} displays a comparison of the populations $n_U,n_L$ in the bare ground state, as predicted by either $H_{\sf mic}$ or $H_{\sf eff}$. In the former case, the lower (upper) polariton can be defined as the eigenmode of $H_{\sf mic}$ with the lowest (second lowest) frequency. In both cases we can observe a small deviation from the results of Fig.~\ref{fig:Energy}, such that $n_L\!\gtrsim\!n_U$. The important point is that the impact of the fuller matter-field interaction model is well captured by the simple effective Hamiltonian ~\eqref{adiabatic}: the discrepancy between Eqs.~(\ref{Hsys1}) and (\ref{adiabatic}) ranges from $1\%$ to $7\%$ of the plotted quantities. Interestingly, the best agreement is observed when $\omega_a\!\sim\!\omega_b$ and $\lambda/\omega_b\sim\!0.25$.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{FIG3}
\end{center}
\caption{(Color online). Comparison between the polaritonic populations predicted by the multimode Hamiltonian ~\eqref{Hsys1} and the effective model \eqref{adiabatic}, in the case $u\!=\!0$. {\bf (a)}: Populations versus coupling strength, when $\omega_a\!=\!\omega_b$. {\bf (b)}: Populations versus bare detuning $\Delta=\omega_a-\omega_b$, with $\lambda\!=\!0.2\omega_b$. To perform the simulations it was sufficient to include 5 matter transitions and 25 cavity modes in Hamiltonian \eqref{Hsys1}.}
\label{fig:More}
\end{figure}
\subsection{Effective Hamiltonian analysis: distribution of populations}
{\noindent}One of the advantages of a few-mode model is that it can be amenable to analytical investigations. Having provided some evidence for the reliability of the Hamiltonian $H_{\sf eff}$, here we exploit its relatively simple form to generalize the results of Section \ref{signature}, and discuss how the parameters $\eta,u$ influence the balance of polaritonic populations in the bare ground state. As we will see shortly, one can still determine a simple analytical condition on the model parameters that results in equal populations. Following similar steps as in Section~\ref{signature}, it is possible to obtain explicit expressions for the eigenfrequencies $\omega_{U,L}$ and the corresponding polaritonic operators. For brevity we shall report the expressions of the bare ground state populations, referring the reader to appendix~\ref{diagonalizza} for a full diagonalization of the Hamiltonian. The quantities of interest read
\begin{align}
n_U&=\tfrac14\cos^2\theta\,\big(\tfrac{\omega_U}{\omega_a}\!+\!\tfrac{\omega_a}{\omega_U}\big)\!+\!\tfrac14\sin^2\theta\,\big(\tfrac{\omega_U}{\omega_b+4u}\!+\!\tfrac{\omega_b+4u}{\omega_U}\big)-\tfrac12,\label{nU1}\\
n_L&=\tfrac14\sin^2\theta\,\big(\tfrac{\omega_L}{\omega_a}\!+\!\tfrac{\omega_a}{\omega_L}\big)\!+\!\tfrac14\cos^2\theta\,\big(\tfrac{\omega_L}{\omega_b+4u}\!+\!\tfrac{\omega_b+4u}{\omega_L}\big)-\tfrac12.\label{nL1}
\end{align}
The specific forms of $\theta,\omega_U,\omega_L$ do not enter the current discussion, but it is important to point out that they will be different from what reported in Section \ref{signature}, except for the `trivial' case $\eta=u=0$. Of great use to us is the following product rule obeyed by the polaritonic frequencies:
\begin{align}
\omega_U^2\omega_L^2=\omega_a(\omega_b\!+\!4u)[(\omega_a\!+\!4D)(\omega_b\!-\!4\eta)-4\lambda^2]\label{prod2},
\end{align}
which can be exploited as follows. By inspecting Eqs.~\eqref{nU1} and \eqref{nL1} we see that a {\it sufficient} condition to achieve equal populations is now given by $\omega_U\omega_L=\omega_a(\omega_b\!+\!4u)$. Comparing this with Eq.~\eqref{prod2}
we can then derive the following condition:
\begin{align}
u=-\frac{\omega_a+4D}{\omega_a}\eta+\frac{\omega_b}{\omega_a}\left(D-\frac{\lambda^2}{\omega_b}\right)\quad\Rightarrow\quad n_U\!=\!n_L\label{equalpops}
\end{align}
Obviously, $D=\lambda^2/\omega_b,\eta=u=0$ represents a possible solution, which corresponds to what we found in Section~\ref{signature}. In general, we can see that assigning the TRK value to the parameter $D$ is no longer sufficient to ensure equal populations. As shown in Fig~\ref{fig:Heff}, the distribution of populations will be ultimately determined by the additional model parameters. It is interesting to note that, while Eq.~\eqref{equalpops} is only a sufficient condition to have $n_U\!=\!n_L$, it is both necessary and sufficient in the examples reported in Fig.~\ref{fig:Heff}, where we allow $\eta$ and $u$ to be of the same order as the parameter $D$. In a rather broad range of parameters, we thus have a simple analytical criterion to determine which polariton will be most populated.
\begin{figure}[t!]
\begin{center}
\hspace{-.5cm}\includegraphics[width=\linewidth]{FIG4}
\end{center}
\caption{(Color online). Contour plots of the relative population difference $(n_U\!-\!n_L)/(n_U\!+\!n_L)$ predicted via the few-mode Hamiltonian $H_{\sf eff}$. We take $D=\lambda^2/\omega_b,\lambda=0.25\omega_b$ in all plots. In plot {\bf (a)} we fix $\omega_a=\omega_b$ and vary the two parameters $u,\eta$, while in plot {\bf (b)} we fix $\eta=0.23\lambda^2/\omega_a$ (as obtained earlier for the Fabry-Perot modes) and vary $u$ together with the bare-mode detuning $\Delta$. The red dashed line corresponds to the model parameters obeying Eq.~\eqref{equalpops}, resulting in $n_U\!=\!n_L$. We have obtained qualitatively similar plots for $\lambda\in\{0.2\omega_b,0.15\omega_b,0.1\omega_b,0.05\omega_b\}$ (Not shown).\label{fig:Heff}}
\end{figure}
\section{Extension to Dicke models}\label{Dicke}
{\noindent}Before concluding, it is useful to illustrate the modification of our predictions when the behaviour of matter deviates significantly from a simple harmonic oscillator. To this end, we consider a generalized Dicke model that closely mimics Eq.~\eqref{adiabatic}:
\begin{align}
\!\!\!H_{\sf Dicke}&=\omega_a\hat a^\dagger\hat a+\omega_b \frac{\hat J_z}{2}+\lambda\frac{\hat J_x}{\sqrt{n}}(\hat a\!+\!\hat a^\dagger)\!+\!D(\hat a\!+\!\hat a^\dagger)^2\nonumber\\
&-\eta\frac{(2\hat J_x)^2}{n}+u\frac{(2\hat J_y)^2}{n},\label{angolone}
\end{align}
where $\hat J_x,\hat J_y,\hat J_z$ are spin-$n/2$ operators. Note that, through the Holstein-Primakoff mapping, it is possible to recover the Hamiltonian $H_{\sf eff}$ as the limit of Eq.~\eqref{angolone} for $n\to\infty$ \cite{Holstein,Holstein-multi}. In Dicke models the integer $n$ is typically interpreted as the number of two-level atoms that collectively interact with the same field. However, as discussed in section~\ref{micro}, in this case the Hamiltonian $H_{\sf Dicke}$ does not fully take into account the impact of electrostatic dipole-dipole interactions, an approximation that might not be well justified in the ultrastrong coupling regime \cite{DomokosPRL}. While these issues certainly deserve further study, here we shall simply adopt Eq.~\eqref{angolone} as our starting point. For our scopes the integer $n$ quantifies the importance of anharmonic effects in matter, such that $H_{\sf Dicke}$ can be interpreted as $H_{\sf eff}$ plus a anharmonic perturbation (this could be formalized via the Holstein-Primakoff mapping, if desired). Exploiting this interpretation we shall draw a comparison between the two models, making use of concepts that are rigorously defined only for the bilinear Hamiltonian $H_{\sf eff}$. We diagonalize Eq.~\eqref{angolone} numerically by truncating the Hilbert space of the cavity, and for our scopes it is sufficient to represent $\hat a$ and $\hat a^\dagger$ as 10-dimensional matrices. Fig.~\ref{fig:HDicke} compares the low-energy spectrum of $H_{\sf Dicke}$, with $n=5$, with that of $H_{\sf eff}$. The qualitative agreement between the two encourages us to label the ground and excited states respectively as $\ket G$, $\ket{N_L,\!N_U}$ in both models. Note that the definition $\ket{N_L,\!N_U}\!\propto\!(\hat p_L^\dagger)^{N_L}\!(\hat p_U^\dagger)^{N_U}\!\ket{G}$ holds in the case of $H_{\sf eff}$, while no simple explicit expression is available for the eigenstates of $H_{\sf Dicke}$. Both Hamiltonians commute with the parity operator: it can be directly verified that each term in either Eq.~\eqref{adiabatic} or Eq.~\eqref{angolone} can only leave the number of bare excitations unchanged, raise it by two or lower it by two. In both models, it can be shown that the ground state $\ket{G}$ is even, while the parity of the excited states is $(N_U\!+\!N_L)\,{\sf mod}\,2$. This symmetry implies that we can expand the bare ground state as
\begin{align}
\ket0\simeq c_0\ket{G}+c_{2_U}\ket{2_U}+c_{2_L}\ket{2_L}+c_{1_L1_U}\ket{1_L1_U}\label{groundo},
\end{align}
where the $c$'s are appropriate complex coefficients, and the overlaps with higher excited states are found to be negligible in all the explored examples. It is understood that the various coefficients and states appearing in Eq.~\eqref{groundo} assume different forms depending on whether $H_{\sf eff}$ or $H_{\sf Dicke}$ is being considered. From Eq.~\eqref{groundo} it follows that the polaritonic populations in the bare ground state are well approximated by
\begin{align}
n_U&\simeq2|c_{2_U}|^2+|c_{1_L1_U}|^2,\label{nU5}\\
n_L&\simeq2|c_{2_L}|^2+|c_{1_L1_U}|^2.\label{nL5}
\end{align}
While of no particular use in the study of $H_{\sf eff}$, where exact expressions are readily available, we can exploit Eqs.~\eqref{nU5} and \eqref{nL5} to calculate (and in fact, {\it define}) the populations $n_U,n_L$ for the Dicke model. Fig.~\ref{fig:HDicke} displays the behaviour of the populations of interest for an arrangement of parameters satisfying Eq.~\eqref{equalpops}. A good qualitative agreement can be observed between the two models in the range of coupling strengths $\lambda\lesssim0.25\omega_b$. As it can be expected the discrepancy between the two tends to grow with increasing $\lambda$: differently from $H_{\sf eff}$, which predicts equal populations, $H_{\sf Dicke}$ results in $n_U\gtrsim n_L$. A detailed explanation of this result goes beyond the scopes of this manuscript. In future studies of the ultrastrong coupling regime, it will be certainly interesting to delve deeper in the study of similarities and differences between truly nonlinear Hamiltonians, such as $H_{\sf Dicke}$, and bilinear interaction models such as those studied here.
\vspace{1cm}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{FIG5}
\end{center}
\caption{(Color online) {\bf (a)}: Low-energy spectra of $H_{\sf Dicke}$ with $n=5$ (empty black circles) and $H_{\sf eff}$ (blue, continuous) as a function of coupling strength. We have displayed the rescaled energies $E/\omega_b$ of the ground state and the lowest five excited states (see labels in the plot). We fixed $\omega_a=\omega_b$, $D=\lambda^2/\omega_b$, $\eta=0.23\lambda^2/\omega_a$. The parameter $u$ is chosen to satisfy Eq.~\eqref{equalpops}, such that equal populations are predicted by $H_{\sf eff}$. {\bf (b)}: Comparison of the polaritonic populations in the bare ground state, for the two Hamiltonias $H_{\sf eff}$ and $H_{\sf Dicke}$. The discrepancy between the two increases with the coupling strength $\lambda$.}
\label{fig:HDicke}
\end{figure}
\section{Discussion and Conclusions}\label{fine}
{\noindent}We have identified a qualitative signature of the $A^2$ term in what is arguably the simplest model of ultrastrong coupling between a single-mode field and a matter excitation. Our finding is a consequence of the TRK sum rule, and in terms of bare vacuum populations it can be expressed in the elegant form $n_U\!=\!n_L$. We have shown how this prediction can be verified by a quenching experiment, assuming that the dominant decay mechanism of the system is the emission of detectable photons. Taking one step further, we have then questioned the validity of the model itself, by interpreting it as a low-energy approximation to a multimode minimal-coupling Hamiltonian. Our analysis gives rise to an effective Hamiltonian for the two modes of interest, featuring two extra terms as compared to our initial interaction model. The effect of these new terms on the quantities $n_U,n_L$ has been discussed. In fact, the information provided in this manuscript makes it straightforward to characterize the full {\it covariance matrix} of the polaritonic modes in the bare ground state.
{\noindent}The results of our study are relevant to a regime of ultrastrong coupling that is accessible in state-of-the-art experiments, and can be used to check the validity of various common approximations and assumptions in the interaction model. For example, if the relationship between bare and polaritonic modes could be experimentally investigated (e.g. via the quenching experiment described here), one would be able to estimate the most appropriate values of the various coupling constants appearing in the effective model $H_{\sf eff}$ (obviously, also the two-mode assumption should be verified in parallel). This could be particularly valuable in systems such as circuit QED, where the influence of the TRK sum rule on the model parameters is under debate (for example, it has been suggested that $D\ll\lambda^2/\omega_b$ could be obtained \cite{CiutiNATURE}).
{\noindent}A rather broad and intriguing open question is whether simple signatures of $A^2$ and other Hamiltonian terms can be identified in more general models, for example featuring a larger number of matter and field modes and/or strong anharmonicities. Despite the theoretical challenge, including one or more of these generalizations may become necessary in attempting to model ever increasing light-matter couplings.
\subsection*{Acknowledgements}
{\noindent}This work was supported by the Leverhulme Trust, the Qatar National Research Fund (Grant No. NPRP 4-554-1-084), the UK EPSRC (Active Plasmonics Programme and Grant No. EP/K034480/1). S.A.M. and M.S.K. acknowledge support from the Royal Society Wolfson Research Merit Awards. We thank F. Armata, S. Barnett, F. Ciccarello, G. M. Palma, J. Iles-Smith, M.-J. Hwang, R. Passante, P. L. Knight, and S. De Liberato for fruitful discussions.
\noindent\textit{Note added} --- During the revision of this manuscript, a preprint was published dealing with the detection of the $A^2$ term in circuit QED \cite{Apero}.
\newline
|
\section{Introduction}
Crowdsourcing data to make better, more informed decisions is becoming increasingly commonplace.
For any such crowdsourcing,
privacy-preservation mechanisms should be applied
to reduce and control the privacy risks introduced by the data collection process,
and balance that risk against the beneficial utility of the collected data.
For this purpose we introduce
\RAPPORSPELLEDOUT{},
or \RAPPOR{},
a widely-applicable, practical new mechanism that provides strong privacy guarantees
combined with high utility,
yet is not founded on the use of trusted third parties.
\RAPPOR{} builds on the ideas of \textit{randomized response}, a surveying technique developed in the 1960s for collecting statistics on sensitive topics where survey respondents wish to retain confidentiality~\citep{warner}.
An example commonly used to describe this technique involves a question on a sensitive topic, such as ``Are you a member of the Communist party?''~\cite{WikipediaRR}.
For this question, the survey respondent is asked to flip a fair coin, in secret, and answer ``Yes'' if it comes up heads, but tell the truth otherwise (if the coin comes up tails).
Using this procedure, each respondent retains very strong deniability for any ``Yes'' answers, since such answers are most likely attributable to the coin coming up heads; as a refinement, respondents can also choose the untruthful answer by flipping another coin in secret, and get strong deniability for both ``Yes'' and ``No'' answers.
Surveys relying on randomized response enable easy computations of accurate population statistics while preserving the privacy of the individuals. Assuming absolute compliance with the randomization protocol (an assumption that may not hold for human subjects, and can even be non-trivial for algorithmic implementations~\cite{mironov-CCS12}), it is easy to see that in a case where both ``Yes'' and ``No'' answers can be denied (flipping two fair coins), the true number of ``Yes'' answers can be accurately estimated by $2(Y - 0.25)$, where $Y$ is the proportion of ``Yes'' responses.
In expectation,
respondents will provide the true answer 75\% of the time,
as is easy to see by
a case analysis of the two fair coin flips.
Importantly,
for one-time collection,
the above randomized survey mechanism
will protect the privacy
of any specific respondent,
irrespective of any attacker's prior knowledge,
as assessed via the $\epsilon$-differential privacy guarantee~\cite{dwork06}.
Specifically,
the respondents will have
differential privacy at the level $\epsilon = \ln\bigl(0.75 / (1 - 0.75)\bigr) = \ln(3)$.
This said,
this privacy guarantee degrades
if the survey is repeated---e.g.,
to get fresh, daily statistics---and
data is collected multiple times from the same respondent.
In this case,
to maintain both differential privacy and utility,
better mechanisms are needed,
like those we present in this paper.
Privacy-Preserving Aggregatable Randomized Response,
or \RAPPOR{}s, is a new mechanism for collecting statistics from end-user, client-side software,
in a manner that provides strong privacy protection
using randomized response techniques.
\RAPPOR{} is designed to permit collecting, over large numbers of clients,
statistics on client-side values and strings,
such as their categories, frequencies, histograms,
and other set statistics.
For any given value reported,
\RAPPOR{}
gives
a strong deniability guarantee for the reporting client,
which strictly limits private information disclosed,
as measured by an $\epsilon$-differential privacy bound,
and holds even for
a single client that reports often on the same value.
A distinct contribution is \RAPPOR{}'s ability
to collect statistics about an arbitrary set of strings
by applying randomized response to Bloom filters~\cite{bloom}
with strong $\epsilon$-differential privacy guarantees.
Another contribution is the elegant manner
in which \RAPPOR{}
protects the privacy of clients
from whom data is collected repeatedly (or even infinitely often), and
how \RAPPOR{}
avoids addition of privacy externalities,
such as those that might be created by
maintaining a database of contributing respondents (which might be breached),
or repeating a single, memoized response (which would be linkable, and might be tracked).
In comparison,
traditional randomized response does not provide any longitudinal privacy
in the case when multiple responses are collected from the same participant.
Yet another contribution is that the \RAPPOR{} mechanism is performed locally on the client, and does not require a trusted third party.
Finally, \RAPPOR{}
provides a novel, high-utility
decoding framework for learning statistics
based on a sophisticated combination of
hypotheses testing,
least-squares solving, and LASSO regression~\cite{lasso}.
\subsection{The Motivating Application Domain}\label{sec:motivation}
\RAPPOR{} is
a general technology for privacy-preserving data collection and crowdsourcing of statistics,
which could be applied in a broad range of contexts.
In this paper, however, we focus on the specific application domain
that motivated the development of \RAPPOR{}:
the need for Cloud service operators to collect up-to-date statistics
about the activity of their users and their client-side software.
In this domain,
\RAPPOR{} has already seen limited deployment in Google's Chrome Web browser,
where it has been used
to improve the data sent by users that have opted-in to reporting statistics~\cite{ChromeRAPPORpage}.
Section~\ref{sec:chromehome} briefly describes this real-world application,
and the benefits \RAPPOR{} has provided
by shining a light on the unwanted or malicious hijacking of user settings.
For a variety of reasons,
understanding population statistics is a key part
of an effective, reliable operation of online services
by Cloud service and software platform operators.
These reasons are often as simple as
observing how frequently certain software features are used,
and measuring their performance and failure characteristics.
Another, important set of reasons
involve
providing better security and abuse protection to the users, their clients, and the service itself.
For example, to assess the prevalence of botnets or hijacked clients,
an operator may wish to monitor how many clients
have---in the last 24 hours---had critical preferences overridden,
e.g., to redirect
the users' Web searches to the URL of a known-to-be-malicious search provider.
The collection of up-to-date crowdsourced statistics
raises a dilemma for service operators.
On one hand,
it will likely be detrimental to the end-users' privacy
to directly collect their information.
(Note that even the search-provider preferences of a user
may be uniquely identifying, incriminating,
or otherwise compromising for that user.)
On the other hand,
not collecting any such information
will also be to the users' detriment:
if operators cannot
gather the right statistics,
they cannot make
many software and service improvements that benefit users
(e.g., by detecting or preventing malicious client-side activity).
Typically, operators resolve this dilemma
by using techniques
that derive only the necessary high-order statistics,
using mechanisms
that limit the users' privacy risks---for example,
by collecting only coarse-granularity data,
and by eliding data that is not shared by a certain number of users.
Unfortunately,
even for careful operators,
willing to utilize state-of-the-art techniques,
there are few existing, practical mechanisms
that offer both privacy and utility,
and even fewer that provide clear privacy-protection guarantees.
Therefore,
to reduce privacy risks,
operators rely to a great extent on pragmatic means and processes,
that, for example, avoid the collection of data,
remove unique identifiers, or otherwise systematically scrub data,
perform mandatory deletion of data after a certain time period,
and, in general, enforce access-control and auditing policies on data use.
However, these approaches are limited in their ability to provide provably-strong privacy guarantees.
In addition, privacy externalities from individual data collections,
such as timestamps or linkable identifiers, may arise;
the privacy impact of those externalities may
be even greater than that of the data collected.
\RAPPOR{} can help operators
handle the significant challenges, and potential privacy pitfalls,
raised by this dilemma.
\subsection{Crowdsourcing Statistics with \RAPPOR{}}
Service operators may apply \RAPPOR{}
to crowdsource statistics
in a manner that protects their users' privacy,
and thus address the challenges described above.
As a simplification,
\RAPPOR{} responses can be assumed to be \emph{bit strings},
where each bit corresponds to a randomized response
for some logical predicate on the reporting client's properties, such as its values, context, or history.
(Without loss of generality,
this assumption
is used for the remainder of this paper.)
For example, one bit in a \RAPPOR{} response may correspond to a predicate
that indicates the stated gender, male or female, of the client user,
or---just as well---their membership in the Communist party.
The structure of a \RAPPOR{} response need not be otherwise constrained;
in particular,
(i) the response bits may be sequential, or unordered,
(ii) the response predicates may be independent, disjoint, or correlated,
and (iii) the client's properties may be immutable, or changing over time.
However,
those details (e.g., any correlation of the response bits)
must be correctly accounted for,
as they impact both the utilization and privacy guarantees of \RAPPOR{}---as
outlined in the next section, and detailed in later sections.
In particular,
\RAPPOR{} can be used to
collect
statistics on
categorical client properties,
by having each bit in a client's response
represent whether, or not, that client belongs to a category.
For example, those categorical predicates
might represent
whether, or not, the client is utilizing a software feature.
In this case,
if each client can use only one of three disjoint features, $X, Y$, and $Z$,
the collection of a three-bit \RAPPOR{} response from clients
will allow measuring
the relative frequency
by which the features are used by clients.
As regards to privacy,
each client will be protected
by the manner in which
the three bits are derived from a single (at most) true predicate;
as regards to utility,
it will suffice to count how many responses had the bit set, for each distinct response bit,
to get a good statistical estimate of the empirical distribution of the features' use.
\RAPPOR{}
can also be used to
gather
population statistics on
numerical and ordinal values,
e.g.,
by associating response bits
with predicates for different ranges of numerical values,
or by reporting on
disjoint categories for different logarithmic magnitudes of the values.
For such numerical \RAPPOR{} statistics,
the estimate may be improved
by collecting and utilizing
relevant information about the priors and shape of
the empirical distribution, such as its smoothness.
Finally,
\RAPPOR{}
also allows collecting statistics on
non-categorical domains, or categories that cannot be enumerated ahead of time,
through the use of Bloom filters~\cite{bloom}.
In particular, \RAPPOR{}
allows collection of compact Bloom-filter-based randomized responses on strings,
instead of
having clients report
when they match a set of hand-picked strings, predefined by the operator.
Subsequently,
those responses can be matched against candidate strings, as they become known to the operator,
and used to estimate both known and unknown strings in the population.
Advanced statistical decoding techniques must be applied
to accurately interpret the randomized, noisy data in Bloom-filter-based \RAPPOR{} responses.
However, as in the case of categories,
this analysis needs only consider the aggregate counts of distinct bits set in \RAPPOR{} responses
to provide good estimators for population statistics,
as detailed in Section~\ref{sec:decoding}.
Without loss of privacy,
\RAPPOR{} analysis can be re-run on a collection of responses, e.g.,
to consider new strings and cases missed in previous analyses,
without the need to re-run the data collection step.
Individual responses
can be especially useful for exploratory or custom data analyses.
For example, if the geolocation of clients' IP addresses are collected
alongside the \RAPPOR{} reports of their sensitive values,
then the observed distributions
of those values
could be compared across different geolocations,
e.g., by analyzing different subsets separately.
Such analysis
is compatible with \RAPPOR{}'s privacy guarantees,
which hold true even in the presence of auxiliary data,
such as geolocation.
By limiting the number of correlated categories, or Bloom filter hash functions,
reported by any single client,
\RAPPOR{} can maintain its differential-privacy guarantees even when
statistics are collected on multiple aspects of clients,
as outlined next, and detailed in Sections~\ref{sec:diffprivacy} and~\ref{sec:attacks}.
\subsection{\RAPPOR{} and (Longitudinal) Attacks}
Protecting privacy for both one-time and multiple collections requires consideration of several distinct attack models.
A basic attacker is assumed to have access to a single report and can be stopped with a single round of randomized response.
A windowed attacker has access to multiple reports over time from the same user.
Without careful modification of the traditional randomized response techniques, almost certainly full disclosure of private information would happen.
This is especially true if the window of observation is large and the underlying value does not change much.
An attacker with complete access to all clients' reports (for example, an insider with unlimited access rights),
is the hardest to stop, yet such an attack is also the most difficult to execute in practice. \RAPPOR{} provides explicit trade-offs between
different attack models in terms of tunable privacy protection for all three types of attackers.
\RAPPOR{} builds on the basic idea of memoization and provides a framework for one-time and longitudinal privacy protection by playing the randomized response game twice with a memoization step in between. The first step, called a Permanent randomized response, is used to create a ``noisy'' answer which is memoized by the client and permanently reused in place of the real answer. The second step, called an Instantaneous randomized response, reports on the ``noisy'' answer over time, eventually completely revealing it.
Long-term, longitudinal privacy is ensured by the use of the Permanent randomized response, while the use of an Instantaneous randomized response provides protection against possible tracking externalities.
The idea of \emph{underlying memoization} turns out to be crucial for privacy protection in the case where multiple responses are collected from the same participant over time. For example, in the case of the question about the Communist party from the start of the paper, memoization can allow us to provide $\ln(3)$-differential privacy even with an \emph{infinite} number of responses, as long as the underlying memoized response has that level of differential privacy.
On the other hand,
without memoization or other limitation on responses,
randomization is not sufficient to
maintain plausible deniability
in the face of multiple collections.
For example,
if 75 out of 100 responses are ``Yes'' for a single client
in the randomized-response scheme
at the very start of this paper,
the true answer will have been ``No''
in a vanishingly unlikely $1.39 \times 10^{-24}$
fraction of cases.
Memoization is absolutely effective in providing longitudinal privacy only in cases when the underlying true value does not change
or changes in an uncorrelated fashion. When users' consecutive reports are temporally correlated, differential privacy guarantees
deviate from their nominal levels and become progressively weaker as correlations increase. Taken to the extreme, when asking users
to report daily on their age in days, additional measures are required to prevent full disclosure over time, such as stopping collection after
a certain number of reports or increasing the noise levels exponentially,
as discussed further in Section~\ref{sec:attacks}.
For a client that reports on a property that strictly alternates between two true values, ($a, b, a, b, a, b, a, b, \ldots$), the two memoized Permanent randomized
responses for $a$ and $b$ will be reused, again and again, to generate \RAPPOR{} report data. Thus, an attacker that obtains a large enough number of reports,
could learn those memoized ``noisy" values with arbitrary certainty---e.g.,
by separately analyzing the even and odd subsequences.
However, even in this case,
the attacker cannot be certain of the values of $a$ and $b$
because of memoization.
This said,
if $a$ and $b$ are correlated,
the attacker may still learn more
than they otherwise would have;
maintaining privacy in the face of
any such correlation
is discussed further in Sections~\ref{sec:diffprivacy} and~\ref{sec:attacks} (see also~\cite{KiferM11}).
\begin{figure*}[!t]
\centering
\includegraphics[trim=0 0.2in 0 0,clip=true,width=2\columnwidth]{life.pdf}
\caption{Life of a \RAPPOR{} report: The client value of the string ``The number 68'' is hashed onto the Bloom filter $B$ using $h$ (here 4) hash functions. For this string, a Permanent randomized response $B'$ is produces and memoized by the client, and this $B'$ is used (and reused in the future) to generate Instantaneous randomized responses $S$ (the bottom row), which are sent to the collecting service.}\vspace*{-2ex}
\label{fig:life}
\end{figure*}
In the next section we will describe the \RAPPOR{} algorithm in detail. We then provide intuition and formal justification for the reasons why the proposed algorithm satisfies the rigorous privacy guarantees of differential privacy. We then devote several sections to discussion of the additional technical aspects of \RAPPOR{} that are crucial for its potential uses in practice, such as parameter selection, interpretation of results via advanced statistical decoding, and experiments illustrating what can be learned in practice. The remaining sections discuss our experimental evaluation, the attack models we consider, the limitations of the \RAPPOR{} technique, as well as related work.
\section{The Fundamental \RAPPOR{} Algorithm}
Given a client's value $v$, the \RAPPOR{} algorithm executed by the client's machine, reports to the server a bit array of size $k$, that encodes a ``noisy" representation of its true value $v$. The noisy representation of $v$ is chosen in such a way so as to reveal a \emph{controlled} amount of information about $v$, limiting the server's ability to learn with confidence what $v$ was. This remains true even for a client that submits an infinite number of reports on a particular value $v$.
To provide such strong privacy guarantees, the \RAPPOR{} algorithm implements two separate defense mechanisms, both of which are based on the idea of randomized response and can be separately tuned depending on the desired level of privacy protection at each level.
Furthermore, additional uncertainty is added through the use of Bloom filters which serve not only to make reports compact, but also to complicate the life of any attacker (since any one bit in the Bloom filter may have multiple data items in its pre-image).
The \RAPPOR{} algorithm takes in the client's true value $v$ and parameters of execution $k, h, f, p, q$, and is executed locally on the client's machine performing the following steps:
\begin{enumerate}
\item {\bf Signal.} Hash client's value $v$ onto the Bloom filter $B$ of size $k$ using $h$ hash functions.
\item {\bf Permanent randomized response.} For each client's value $v$ and bit $i, 0 \leq i < k$ in $B$, create a binary reporting value $B'_i$ which equals to
$$
B'_i = \begin{cases}
1, & \text{with probability $\frac{1}{2}f$} \\
0, & \text{with probability $\frac{1}{2}f$} \\
B_i, & \text{with probability $1 - f$}
\end{cases}
$$
where $f$ is a user-tunable parameter controlling the level of longitudinal privacy guarantee.\\[1ex]
Subsequently, this $B'$ is memoized and reused as the basis for all future reports on this distinct value $v$.
\item {\bf Instantaneous randomized response.} Allocate a bit array $S$ of size $k$ and initialize to 0. Set each bit $i$ in $S$ with probabilities
$$
P(S_i = 1) = \begin{cases}
q, & \text{if $B'_i = 1$}. \\
p, & \text{if $B'_i = 0$}.
\end{cases}
$$
\item {\bf Report.} Send the generated report $S$ to the server.
\end{enumerate}
There are many different variants of the above randomized response mechanism. Our main objective for selecting these two particular versions was to make the scheme intuitive and easy to explain.
The Permanent randomized response (step 2) replaces the real value $B$ with a derived randomized noisy value $B'$. $B'$ may or may not contain any information about $B$ depending on whether signal bits from the Bloom filter are being replaced by random 0's with probability $\frac{1}{2}f$. The Permanent randomized response ensures privacy because of the adversary's limited ability to differentiate between true and ``noisy'' signal bits. It is absolutely critical that all future reporting on the information about $B$ uses the same randomized $B'$ value to avoid an ``averaging" attack, in which an adversary estimates the true value from observing multiple noisy versions of it.
The Instantaneous randomized response (step 3) plays several important functions. Instead of directly reporting $B'$ on every request, the client reports a randomized version of $B'$. This modification significantly increases the difficulty of tracking a client based on $B'$, which could otherwise be viewed as a unique identifier in longitudinal reporting scenarios. It also provides stronger short-term privacy guarantees (since we are adding more noise to the report) which can be independently tuned to balance short-term vs long-term risks. Through tuning of the parameters of this mechanism we can effectively balance utility against different attacker models.
Figure~\ref{fig:life} shows a random run of the \RAPPOR{} algorithm. Here, a client's value is $v = ``68"$, the size of the Bloom filter is $k = 256$, the number of hash functions is $h = 4$, and the tunable randomized response parameters are: $p = 0.5$, $q = 0.75$, and $f = 0.5$. The reported bit array sent to the server is shown at the bottom of the figure. 145 out of 256 bits are set in the report. Of the four Bloom filter bits in $B$ (second row), two are propagated to the noisy Bloom filter $B'$. Of these two bits, both are turned on in the final report. The other two bits are never reported on by this client due to the permanent nature of $B'$. With multiple collections from this client on the value ``68'', the most powerful attacker would eventually learn $B'$ but would continue to have limited ability to reason about the value of $B$, as measured by differential privacy guarantee. In practice, learning about the actual client's value $v$ is even harder because multiple values map to the same bits in the Bloom filter \citep{bloom_privacy}.
\subsection{\RAPPOR{} Modifications}
The \RAPPOR{} algorithm can be modified in a number of ways depending on the particulars of the scenario in which privacy-preserving data collection is needed. Here, we list three common scenarios where omitting certain elements from the \RAPPOR{} algorithm leads to a more efficient learning procedure, especially with smaller sample sizes.
\begin{itemize}
\item {\bf One-time \RAPPOR{}.} One time collection, enforced by the client, does not require longitudinal privacy protection. The Instantaneous randomized response step can be skipped in this case and a direct randomization on the true client's value is sufficient to provide strong privacy protection.
\item {\bf Basic \RAPPOR{}.} If the set of strings being collected is relatively small and well-defined, such that each string can be deterministically mapped to a single bit in the bit array, there is no need for using a Bloom filter with multiple hash functions. For example, collecting data on client's gender could simply use a two-bit array with ``male'' mapped to bit 1 and ``female'' mapped to bit 2. This modification would affect step 1, where a Bloom filter would be replaced by a deterministic mapping of each candidate string to one and only one bit in the bit array. In this case, the effective number of hash functions, $h$, would be 1.
\item {\bf Basic One-time \RAPPOR{}.} This is the simplest configuration of the \RAPPOR{} mechanism, combining the first two modifications at the same time: one round of randomization using a deterministic mapping of strings into their own unique bits.
\end{itemize}
\section{Differential Privacy of \RAPPOR{}}\label{sec:diffprivacy}
The scale and availability of data in today's world makes increasingly sophisticated attacks feasible, and any system that hopes to withstand such attacks should aim to ensure rigorous, rather than merely intuitive privacy guarantees. For our analysis, we adopt the rigorous notion of privacy, \textit{differential privacy}, which was introduced by Dwork et al~\citep{dwork06} and has been widely adopted~\cite{DworkCACM}. The definition aims to ensure that the output of the algorithm does not significantly depend on any particular individual's data. The quantification of the increased risk that participation in a service poses to an individual can, therefore, empower clients to make a better informed decision as to whether they want their data to be part of the collection.
Formally, a randomized algorithm $A$ satisfies $\epsilon$-differential privacy~\citep{dwork06} if for all pairs of client's values $v_1$ and $v_2$ and for all $R\subseteq Range(A)$,
$$
P(A(v_1) \in R) \le e^\epsilon P(A(v_2) \in R).
$$
We prove that the \RAPPOR{} algorithm satisfies the definition of differential privacy next.
Intuitively, the Permanent randomized response part ensures that the ``noisy" value derived from the true value protects privacy, and the Instantaneous randomized response provides protection against usage of that response by a longitudinal tracker.
\subsection{Differential Privacy of the Permanent Randomized Response}
\begin{theorem}\label{thm-one}
The Permanent randomized response (Steps 1 and 2 of \RAPPOR{}) satisfies $\epsilon_{\infty}$-differential privacy where $\epsilon_{\infty} = 2h\ln\left(\frac{1 - \frac{1}{2}f}{\frac{1}{2}f}\right)$.
\end{theorem}
\begin{proof}
Let $S = s_1, \ldots, s_k$ be a randomized report generated by the \RAPPOR{} algorithm. Then the probability of observing any given report $S$ given the true client value $v$ and assuming that $B'$ is known is
\begin{eqnarray*}
P(S = s | V = v) & = & P(S = s | B, B', v)\cdot P(B' | B, v)\cdot P(B | v) \\
& = & P(S = s | B')\cdot P(B' | B)\cdot P(B | v) \\
& = & P(S = s | B')\cdot P(B' | B).
\end{eqnarray*}
Because $S$ is conditionally independent of $B$ given $B'$, the first probability provides no additional information about $B$.
$P(B' | B)$ is, however, critical for longitudinal privacy protection. Relevant probabilities are
\begin{eqnarray*}
P(b'_i = 1 | b_i = 1) & = & \frac{1}{2}f + 1 - f = 1 - \frac{1}{2}f \;\;\;\text{~and~}\\
P(b'_i = 1 | b_i = 0) & = & \frac{1}{2}f.
\end{eqnarray*}
Without loss of generality, let the Bloom filter bits $1, \ldots, h$ be set, i.e., $b^* = \{b_1 = 1, \ldots, b_h = 1, b_{h+1} = 0, \ldots, b_{k} = 0\}$. Then,
\begin{eqnarray*}
P(B' = b' | B = b^*) & = & \left(\frac{1}{2}f\right)^{b'_1}\left(1 - \frac{1}{2}f\right)^{1 - b'_1} \times \ldots \\
& & \times \left(\frac{1}{2}f\right)^{b'_h}\left(1 - \frac{1}{2}f\right)^{1 - b'_h} \times \ldots \\
& & \times \left(1 - \frac{1}{2}f\right)^{b'_{h + 1}}\left(\frac{1}{2}f\right)^{1 - b'_{h + 1}} \times \ldots \\
& & \times \left(1 - \frac{1}{2}f\right)^{b'_k}\left(\frac{1}{2}f\right)^{1 - b'_k}.
\end{eqnarray*}
Let $RR_{\infty}$ be the ratio of two such conditional probabilities with distinct values of $B$, $B_1$ and $B_2$, i.e., $RR_{\infty} = \frac{P(B'\in R^* | B = B_1)}{P(B' \in R^* | B = B_2)}$. For the differential privacy condition to hold, $RR_{\infty}$ needs to be bounded by $\exp(\epsilon_{\infty})$.
\begin{eqnarray*}
RR_{\infty} & = & \frac{P(B'\in R^* | B = B_1)}{P(B' \in R^* | B = B_2)} \\
& = & \frac{\sum_{B'_i \in R^*}P(B' = B'_i |B = B_1)}{\sum_{B'_i \in R^*} P(B' = B'_i | B = B_2)} \\
& \le & \max_{B'_i \in R^*} \frac{P(B' = B'_i | B = B_1)}{P(B' = B'_i| B = B_2)} \;\;\;\;\; \text{(by Observation~\ref{obs-ratios})} \\
& = & \left(\frac{1}{2}f\right)^{2(b'_1 + b'_2 + \ldots + b'_h - b'_{h+1} - b'_{h+2} - \ldots - b'_{2h})} \\
& & \times \left(1 - \frac{1}{2}f\right)^{2(b'_{h+1} + b'_{h+2} + \ldots + b'_{2h} - b'_1 - b'_2 - \ldots - b'_h)}.
\end{eqnarray*}
Sensitivity is maximized when $b'_{h+1} = b'_{h+2} = \ldots = b'_{2h} = 1$ and $b'_1 = b'_2 = \ldots = b'_h = 0$. Then,\\
$
RR_{\infty} = \left(\frac{1 - \frac{1}{2}f}{\frac{1}{2}f}\right)^{2h} \text{~and~}
\epsilon_{\infty} = 2h\ln\left(\frac{1 - \frac{1}{2}f}{\frac{1}{2}f}\right).
$~\end{proof}
Note that $\epsilon_{\infty}$ is not a function of $k$. It is true that a smaller $k$, or a higher rate of Bloom filter bit collision, sometimes improves privacy protection, but, on its own, it is not sufficient nor necessary to provide $\epsilon$-differential privacy.
\subsection{Differential Privacy of the Instantaneous Randomized Response}
With a single data collection from each client, the attacker's knowledge of $B$ must come directly from a single report $S$ generated by applying the randomization twice, thus, providing a higher level of privacy protection than under the assumption of complete knowledge of $B'$.
Because of a two-step randomization, probability of observing a 1 in a report is a function of both $q$ and $p$ as well as $f$.
\begin{lemma}\label{lem-one}
Probability of observing 1 given that the underlying Bloom filter bit was set is given by
$$
q^* = P(S_i = 1 | b_i = 1) = \frac{1}{2}f(p + q) + (1 - f)q.
$$
Probability of observing 1 given that the underlying Bloom filter bit was \emph{not} set is given by
$$
p^* = P(S_i = 1 | b_i = 0) = \frac{1}{2}f(p + q) + (1 - f)p.
$$
\end{lemma}
We omit the proof as the reasoning is straightforward that probabilities in both cases are mixtures of random and true responses with the mixing proportion $f$.
\begin{theorem}
The Instantaneous randomized response (Step 3 of \RAPPOR{}) satisfies $\epsilon_1$-differential privacy, where $\epsilon_1 = h\log\left(\frac{q^*(1 - p^*)}{p^*(1 - q^*)}\right)$ and $q^*$ and $p^*$ as defined in Lemma~\ref{lem-one}.
\end{theorem}
\begin{proof}
The proof is analogous to Theorem~\ref{thm-one}. Let $RR_1$ be the ratio of two conditional probabilities, i.e., $RR_1 = \frac{P(S \in R | B = B_1)}{P(S \in R | B = B_2)}$. To satisfy the differential privacy condition, this ratio must be bounded by $\exp(\epsilon_1)$.
\begin{eqnarray*}
RR_1 &=& \frac{P(S \in R | B = B_1)}{P(S \in R | B = B_2)} \\
&=& \frac{\sum_{s_j \in R}P(S = s_j | B = B_1)}{\sum_{s_j \in R}P(S = s_j | B = B_2)} \\
&\le& \max_{s_j \in R} \frac{P(S= s_j | B = B_1)}{P(S = s_j | B = B_2)} \\
&=& \left[\frac{q^*(1 - p^*)}{p^*(1 - q^*)}\right]^h
\end{eqnarray*}
and
$$
\epsilon_1 = h\log\left(\frac{q^*(1 - p^*)}{p^*(1 - q^*)}\right).
$$\end{proof}
The above proof naturally extends to $N$ reports, since each report that is not changed contributes a fixed amount to the total probability of observing \emph{all} reports and enters both nominator and denominator in a multiplicative way (because of independence). Since our differential privacy framework considers inputs that differ only in a single record, $j$, (reports set $D_1$ becomes $D_2$ differing in a single report $S_j$), the rest of the product terms end up canceling out in the ratio
\begin{eqnarray*}
\frac{P(S_1 = s_1, S_2 = s_2, \ldots, S_j = s_j, \ldots, S_N = s_N | B_1)}{P(S_1 = s_1, S_2 = s_2, \ldots, S_j = s_j, \ldots, S_N = s_N | B_2)} =\\
\frac{\prod_{i=1}^N P(S_i = s_i | B_1)}{\prod_{i=1}^N P(S_i = s_i | B_2)} = \frac{P(S_j = s_j | B_1)}{P(S_j = s_j | B_2)}.
\end{eqnarray*}
Computing $\epsilon_n$ for the $n$th collection cannot be made without additional assumptions about how effectively the attacker can learn $B'$ from the collected reports. We continue working on providing these bounds under various learning strategies. Nevertheless, as $N$ becomes large, the bound approaches $\epsilon_{\infty}$ but always remains strictly smaller.
\begin{figure*}[!t]
\centering
\includegraphics[width=2\columnwidth]{sim2.pdf}
\caption{Recall versus precision depending on choice of parameters $k$, $h$, and $m$. The first panel shows the true population distribution from which \RAPPOR{} reports were sampled. The other three panels vary one of the parameters while keeping the other two fixed. Best precision and recall are achieved with using 2 hash functions, while the choices of $k$ and $m$ do not show clear preferences.}
\label{fig:sim}
\end{figure*}
\section{High-utility Decoding of Reports}\label{sec:decoding}
In most cases, the goal of data collection using \RAPPOR{} is to learn which strings are present in the sampled population and what their corresponding frequencies are. Because we make use of the Bloom filter (loss of information) and purposefully add noise for privacy protection, decoding requires sophisticated statistical techniques.
To facilitate learning, before any data collection begins each client is randomly assigned and becomes a permanent member of one of $m$ \emph{cohorts}. Cohorts implement different sets of $h$ hash functions for their Bloom filters, thereby reducing the chance of accidental collisions of two strings across all of them. Redundancy introduced by running $m$ cohorts simultaneously greatly improves the false positive rate. The choice of $m$ should be considered carefully, however. When $m$ is too small, then collisions are still quite likely, while when $m$ is too large, then each individual cohort provides insufficient signal due to its small sample size (approximately $N/m$, where $N$ is the number of reports). Each client must report its cohort number with every submitted report, i.e., it is not private but made private.
We propose the following approach to learning from the collected reports:
\begin{itemize}
\item Estimate the number of times each bit $i$ within cohort $j$, $t_{ij}$, is truly set in $B$ for each cohort. Given the number of times each bit $i$ in cohort $j$, $c_{ij}$ was set in a set of $N_j$ reports, the estimate is given by
$$
t_{ij} = \frac{c_{ij} - (p + \frac{1}{2}fq - \frac{1}{2}fp)N_j}{(1 - f)(q - p)}.
$$
Let $Y$ be a vector of $t_{ij}$'s, $i \in [1, k], j \in[1, m]$.
\item Create a design matrix $X$ of size $km \times M$ where $M$ is the number of candidate strings under consideration. $X$ is mostly 0 (sparse) with 1's at the Bloom filter bits for each string for each cohort. So each column of $X$ contains $hm$ 1's at positions where a particular candidate string was mapped to by the Bloom filters in all $m$ cohorts. Use Lasso \citep{lasso} regression to fit a model $Y \sim X$ and select candidate strings corresponding to non-zero coefficients.
\item Fit a regular least-squares regression using the selected variables to estimate counts, their standard errors and p-values.
\item Compare p-values to a Bonferroni corrected level of $\alpha / M = 0.05 / M$ to determine which frequencies are statistically significant from 0. Alternatively, controlling the False Discovery Rate (FDR) at level $\alpha$ using the Benjamini-Hochberg procedure \cite{Benjamini1995}, for example, could be used.
\end{itemize}
\subsection{Parameter Selection}
Practical implementation of the \RAPPOR{} algorithm requires specification of a number of parameters. $p$, $q$, $f$ and the number of hash functions $h$ control the level of privacy for both one-time and longitudinal collections. Clearly, if no longitudinal data is being collected, then we can use One-time \RAPPOR{} modification. With the exception of $h$, the choice of values for these parameters should be driven exclusively by the desired level of privacy $\epsilon$. $\epsilon$ itself can be picked depending on the circumstances of the data collection process; values in the literature range from $0.01$ to $10$ (see Table 1 in \cite{HsuGHKNPR14}).
Bloom filter size, $k$, the number of cohorts, $m$, and $h$ must also be specified \emph{a priori}. Besides $h$, neither $k$ nor $m$ are related to the worst-case privacy considerations and should be selected based on the efficiency properties of the algorithm in reconstructing the signal from the noisy reports.
We ran a number of simulations (averaged over 10 replicates) to understand how these three parameters effect decoding; see Figure \ref{fig:sim}. All scenarios assumed $\epsilon = \ln(3)$ privacy guarantee. Since only a single report from each user was simulated, One-time \RAPPOR{} was used. Population sampled is shown in the first panel and contains 100 non-zero strings with 100 strings that had zero probability of occurring. Frequencies of non-zero strings followed an exponential distribution as shown in the figure.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{learnable.pdf}
\caption{Sample size vs the upper limit on the strings whose frequency can be learned. Seven colored lines represent different cardinalities of the candidate string set. Here, $p = 0.5$, $q = 0.75$ and $f = 0$.}
\label{fig:learn}
\end{figure}
In the other three panels, the x-axis shows the recall rate and the y-axis shows the precision rate. In all three panels, the same set of points are plotted and are only labeled differently depending on which parameter changes in a particular panel. Each point represents an average recall and precision for a unique combination of $k$, $h$, and $m$. For example, the second panel shows the effect of the Bloom filter size on both precision and recall while keeping both $h$ and $m$ fixed. It is difficult to make definitive conclusions about the optimal size of the Bloom filter as different sizes perform similarly depending on the values of $h$ and $m$. The third panel, however, shows a clear preference for using only two hash functions from the perspective of utility, as the decrease in the number of hash functions used increases the expected recall. The fourth panel, similarly to the second, does not definitively indicate the optimal direction for choosing the number of cohorts.
\subsection{What Can We Learn?}
In practice, it is common to use thresholds on the number of unique submissions in order to ensure some privacy. However, arguments as to how those thresholds should be set abound, and most
of the time they are based on a `feel' for what is accepted and lack any objective justification. \RAPPOR{} also requires $\epsilon$, a user-tunable parameter, which by the design of the algorithm translates into limits on frequency domain, i.e., puts a lower limit on the number of times a string needs to be observed in a sample before it can be reliably identified and its frequency estimated. Figure~\ref{fig:learn} shows the relationship between the sample size (x-axis) and the theoretical upper limit (y-axis) on how many strings can be detected at that sample size for a particular choice of $p = 0.5$ and $q = 0.75$ (with $f = 0$) at a given confidence level $\alpha = 0.05$.
It is perhaps surprising that we do not learn more at very large sample sizes (e.g., one billion). The main reason is that as the number of strings in the population becomes large, their frequencies proportionally decrease and they become hard to detect at those low frequencies.
We can only reliably detect about 10,000 strings in a sample of ten billion and about 1,000 with a sample of one hundred million. A general rule of thumb is $\sqrt{N}/10$, where $N$ is the sample size. These theoretical calculations are based on the Basic One-time \RAPPOR{} algorithm (the third modification) and are the upper limit on what can be learned since there is no additional uncertainty introduced by the use of Bloom filter. Details of the calculations are shown in the Appendix.
While providing $\ln(3)$-differential privacy for one time collection, if one would like to detect items with frequency 1\%, then one million samples are required,
0.1\% would require a sample size of 100 million and 0.01\% items would be identified only in a sample size of 10 billion.
Efficiency of the unmodified \RAPPOR{} algorithm is significantly inferior when compared to the Basic One-time \RAPPOR{} (the price of compression). Even for the Basic One-time \RAPPOR{}, the provided bound can be theoretically achieved only if the underlying distribution of the strings' frequencies is uniform (a condition under which the smallest frequency is maximized). With the presence of several high-frequency strings, there is less probability mass left for the tail and, with the drop in their frequencies, their detectability suffers.
\section{Experiments and Evaluation}
We demonstrate our approach using two simulated and two real-world collection examples. The first simulated one uses the Basic One-time \RAPPOR{} where we learn the shape of the underlying Normal distribution. The second simulated example uses unmodified \RAPPOR{} to collect strings whose frequencies exhibit exponential decay. The third example is drawn from a real-world dataset on processes running on a set of Windows machines. The last example is based on the Chrome browser settings collections.
\subsection{Reporting on the Normal Distribution}
To get a sense of how effectively we can learn the underlying distribution of values reported through the Basic One-time \RAPPOR{}, we simulated learning the shape of the Normal distribution (rounded to integers) with mean 50 and standard deviation 10. The privacy constraints were: $q = 0.75$ and $p = 0.5$ providing $\epsilon = \ln(3)$ differential privacy ($f = 0$). Results are shown in Figure~\ref{fig:normal} for three different sample sizes. With 10,000 reports, results are just too noisy to obtain a good estimate of the shape. The Normal bell curve begins to emerge already with 100,000 reports and at one million reports it is traced very closely. Notice the noise in the left and right tails where there is essentially no signal. It is required by the differential privacy condition and also gives a sense of how uncertain our estimated counts are.
\begin{figure*}[!t]
\centering
\includegraphics[scale=.5]{normal.pdf}
\caption{Simulations of learning the normal distribution with mean 50 and standard deviation 10. The \RAPPOR{} privacy parameters are $q = 0.75$ and $p = 0.5$, corresponding to $\epsilon = \ln(3)$. True sample distribution is shown in black; light green shows the estimated distribution based on the decoded \RAPPOR{} reports. We do not assume \emph{a priori} knowledge of the Normal distribution in learning. If such prior information were available, we could significantly improve upon learning the shape of the distribution via smoothing.}
\label{fig:normal}
\end{figure*}
\subsection{Reporting on an Exponentially-distributed Set of Strings}
The true underlying distribution of strings from which we sample is shown in Figure~\ref{fig:detected}. It shows commonly encountered exponential decay in the frequency of strings with several ``heavy hitters'' and the long tail. After sampling 1 million values (one collection event per user) from this population at random, we apply \RAPPOR{} to generate 1 million reports with $p = 0.5$, $q = 0.75$, $f = 0.5$, two hash functions, Bloom filter size of 128 bits and 16 cohorts.
After the statistical analysis using the Bonferroni correction discussed above, 47 strings were estimated to have counts significantly different from 0. Just 2 of the 47 strings were false positives, meaning their true counts were truly 0 but estimated to be significantly different. The top-20 detected strings with their count estimates, standard errors, p-values and z-scores (SNR) are shown in Table~\ref{tab:results20}. Small p-values show high confidence in our assessment that the true counts are much larger than 0 and, in fact, comparing columns 2 and 5 confirms that. Figure~\ref{fig:detected} shows all 47 detected strings in dark red. All common strings above the frequency of approximately $1\%$ were detected and the long tail remained protected by the privacy mechanism.
\begin{figure}[!t]
\centering
\includegraphics[trim=0 0.8in 0 0.6in,clip=true,width=\columnwidth]{detected.pdf}
\caption{Population of strings with their true frequencies on the vertical axis (0.01 is 1\%). Strings detected by \RAPPOR{} are shown in dark red.}
\label{fig:detected}
\end{figure}
\begin{table}[!h]
\centering
\begin{tabular}{lcccccr}
\hline
String & Est.\ & Stdev\ & P.value & Truth & Prop.\ & SNR \\
\hline
V\_1 & 48803 & 2808 & 5.65E-63 & 49884 & 0.05 & 17.38 \\
V\_2 & 47388 & 2855 & 5.82E-58 & 47026 & 0.05 & 16.60 \\
V\_5 & 41490 & 2801 & 4.30E-47 & 40077 & 0.04 & 14.81 \\
V\_7 & 40682 & 2849 & 4.58E-44 & 36565 & 0.04 & 14.28 \\
V\_4 & 40420 & 2811 & 1.31E-44 & 42747 & 0.04 & 14.38 \\
V\_3 & 39509 & 2882 & 7.03E-41 & 44642 & 0.04 & 13.71 \\
V\_8 & 36861 & 2842 & 5.93E-37 & 34895 & 0.03 & 12.97 \\
V\_6 & 36220 & 2829 & 4.44E-36 & 38231 & 0.04 & 12.80 \\
V\_10 & 34196 & 2828 & 1.72E-32 & 31234 & 0.03 & 12.09 \\
V\_9 & 32207 & 2805 & 1.45E-29 & 33106 & 0.03 & 11.48 \\
V\_12 & 30688 & 2822 & 9.07E-27 & 28295 & 0.03 & 10.87 \\
V\_11 & 29630 & 2831 & 5.62E-25 & 29908 & 0.03 & 10.47 \\
V\_14 & 27366 & 2850 & 2.33E-21 & 25984 & 0.03 & 9.60 \\
V\_19 & 23860 & 2803 & 3.41E-17 & 20057 & 0.02 & 8.51 \\
V\_13 & 22327 & 2826 & 4.69E-15 & 26913 & 0.03 & 7.90 \\
V\_15 & 21752 & 2825 & 2.15E-14 & 24653 & 0.02 & 7.70 \\
V\_20 & 20159 & 2821 & 1.26E-12 & 19110 & 0.02 & 7.15 \\
V\_18 & 19521 & 2835 & 7.74E-12 & 20912 & 0.02 & 6.89 \\
V\_17 & 18387 & 2811 & 7.86E-11 & 22141 & 0.02 & 6.54 \\
V\_21 & 18267 & 2828 & 1.33E-10 & 17878 & 0.02 & 6.46 \\
\hline
\end{tabular}
\caption{Top-20 strings with their estimated frequencies, standard deviations, p-values, true counts and signal to noise ratios (SNR or z-scores).}
\label{tab:results20}
\end{table}
\subsection{Reporting on Windows Process Names}\label{sec:windows}
We collected 186,792 reports from 10,133 different Windows computers, sampling actively running
processes on each machine. On average, just over 18 process names were collected from each machine with the goal of recovering the most common ones and estimating the frequency of a particularly malicious binary named ``BADAPPLE.COM''.
This collection used 128 Bloom filter with 2 hash functions and 8 cohorts. Privacy parameters were chosen such that $\epsilon_1 = 1.0743$ with $q = 0.75$, $p = 0.5,$ and $f = 0.5$. Given this configuration, we optimistically expected to discover processes with frequency of at least 1.5\%.
We identified 10 processes shown in Table \ref{tab:proc} ranging in frequency between 2.5\% and 4.5\%. They were identified by controlling the False Discovery Rate at 5\%. The ``BADAPPLE.COM'' process was estimated to have frequency of 2.6\%. The other 9 processes were common Windows tasks we would expect to be running on almost every Windows machine.
\begin{table}[h]
\centering
\caption{Windows processes detected.}
\begin{tabular}{lcccc}
\hline
Process Name & Est.\ & Stdev & P.value & Prop.\ \\
\hline
RASERVER.EXE & 8054 & 1212 & 1.56E-11 & 0.04 \\
RUNDLL32.EXE & 7488 & 1212 & 3.32E-10 & 0.04 \\
CONHOST.EXE & 7451 & 1212 & 4.02E-10 & 0.04 \\
SPPSVC.EXE & 6363 & 1212 & 7.74E-08 & 0.03 \\
AITAGENT.EXE & 5579 & 1212 & 2.11E-06 & 0.03 \\
MSIEXEC.EXE & 5147 & 1212 & 1.10E-05 & 0.03 \\
SILVERLIGHT.EXE & 4915 & 1212 & 2.53E-05 & 0.03 \\
BADAPPLE.COM & 4860 & 1212 & 3.07E-05 & 0.03 \\
LPREMOVE.EXE & 4787 & 1212 & 3.95E-05 & 0.03 \\
DEFRAG.EXE & 4760 & 1212 & 4.34E-05 & 0.03 \\
\hline
\end{tabular}
\label{tab:proc}
\end{table}
\subsection{Reporting on Chrome Homepages}\label{sec:chromehome}
The Chrome Web browser has implemented and deployed \RAPPOR{}
to collect data about Chrome clients~\cite{ChromeRAPPORpage}.
Data collection has been limited to
some of the Chrome users who have opted in to send usage statistics to Google,
and to certain Chrome settings,
with daily collection
from approximately $\sim$14 million respondents.
Chrome settings, such as homepage, search engine and others, are often targeted by malicious software and changed without users' consent. To understand who the main players are, it is critical to know the distribution of these settings on a large number of Chrome installations. Here, we focus on learning the distribution of homepages and demonstrate what can be learned from a dozen million reports with strong privacy guarantees.
This collection used 128 Bloom filter with 2 hash functions and 32 cohorts. Privacy parameters were chosen such that $\epsilon_1 = 0.5343$ with $q = 0.75$, $p = 0.5,$ and $f = 0.75$.
Given this configuration, optimistically, \RAPPOR{} analysis can discover homepage URL domains, with statistical confidence, if their frequency exceeds 0.1\% of the responding population.
Practically, this means that more than $\sim$14 thousand clients must report on the same URL domain, before it can be identified in the population by \RAPPOR{} analysis.
Figure~\ref{fig:detectedtwo} shows the relative frequencies of 31 unexpected homepage domains discovered by \RAPPOR{} analysis.
(Since not all of these are necessarily malicious, the figure does not include the actual URL domain strings that were identified.)
As one might have expected, there are several popular homepages, likely intentionally set by users, along with a long tail of relatively rare URLs. Even though less than 0.5\% out of 8,616 candidate URLs provide enough statistical evidence for their presence (after the FDR correction), they collectively account for about 85\% of the total probability mass.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{homepages.pdf}
\caption{Relative frequencies of the top 31 unexpected Chrome homepage domains found by analyzing $\sim$14 million \RAPPOR{} reports, excluding expected domains
(the homepage ``google.com'', etc.).}
\label{fig:detectedtwo}
\end{figure}
\section{Attack Models and Limitations}\label{sec:attacks}
We consider three types of attackers with different capabilities for collecting \RAPPOR{} reports.
The least powerful attacker has access to a single report from each user and is limited by one-time differential privacy level $\epsilon_1$ on how much knowledge gain is possible.
This attacker corresponds to an eavesdropper that has temporary ability to snoop on the users' reports.
A windowed attacker is presumed to have access to one client's data over a well-defined period of time. This attacker, depending on the sophistication of her learning model, could learn more information about a user than the attacker of the first type. Nevertheless, the improvement in her ability to violate privacy is strictly bounded by the longitudinal differential privacy guarantee of $\epsilon_{\infty}$.
This more powerful attacker may correspond to an adversary such as a malicious Cloud service employee, who may have temporary access to reports, or access to a time-bounded log of reports.
The third type of attacker is assumed to have unlimited collection capabilities and can learn the Permanent randomized response $B'$ with absolute certainty. Because of the randomization performed to obtain $B'$ from $B$, she is also bounded by the privacy guarantee of $\epsilon_{\infty}$ and cannot improve upon this bound with more data collection.
This corresponds to a worst-case adversary, but still one that doesn't have direct access to the true data values on the client.
Despite envisioning a completely local privacy model, one where users themselves release data in a privacy-preserving fashion,
operators of \RAPPOR{} collections, however, can easily manipulate the process to learn more information than warranted by the
nominal $\epsilon_{\infty}$. Soliciting users to participate more than once in a particular collection results in multiple Permanent
randomized responses for each user and partially defeats the benefits of memoization. In the web-centric world, users use multiple accounts
and multiple devices and can unknowingly participate multiple times, releasing more information than what they expected.
This problem could be mitigated to some extent by running collections per account and sharing a common Permanent randomized response.
Notice the role of the operator to ensure that such processes are in place and the required or assumed trust on the part of the user.
It is likely that some attackers will aim to target specific users
by isolating and analyzing reports from that user, or a small group of users that includes them.
Even so, some randomly-chosen users need not fear such attacks at all:
with probability $\left(\frac{1}{2}f\right)^h$, clients will generate a Permanent randomized response $B'$ with all 0s at the positions of set Bloom filter bits. Since these clients are not contributing any useful information to the collection process, targeting them individually by an attacker is counter-productive. An attacker has nothing to learn about this particular user.
Also, for all users, at all times, there is plausible deniability proportional to the fraction of clients providing no information.
\begin{figure}[t]
\begin{center}
\includegraphics[trim=0 0.2in 0 0,clip=true,width=\columnwidth]{fdr.pdf}
\caption{False Discovery Rate (FDR) as a function of string frequency and $f$. Identifying rare strings in a population without introducing a large number of false discoveries is infeasible. Also, FDR is proportional to $f$.}
\label{fig:fdr}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[trim=0 0.2in 0 0,clip=true,width=\columnwidth]{fdr2.pdf}
\caption{Exact probabilities for inferring the true value $v$ given the two bits observed in a \RAPPOR{} report $S$ corresponding to the two bits set by string $v$. For rare strings, even when both bits are set to 1 (green lines), it is still much more likely that the client \emph{did not} report $v$, but some other value.}
\label{fig:fdr2}
\end{center}
\end{figure}
In one particular attack scenario,
imagine an attacker that is interested in learning whether a given client has a particular value $v$,
whose population frequency is known to be $f_v$.
The strongest evidence in support of $v$ comes in the form of both Bloom filter bits for $v$ being set in the client's report (if two hash functions are used).
The attacker can formulate its target set by selecting all reports with these two bits set.
However, this set will miss some clients with $v$ and include other clients who did not report $v$.
False discovery rate (FDR) is the proportion of clients in the target set who reported a value different from $v$.
Figure \ref{fig:fdr} shows FDR as a function of $f_v$, the frequency of the string $v$.
Notably,
for relatively rare values, most clients in the target set will, in fact, have a value that is different from $v$,
which will hopefully deter any would-be attackers.
The main reason for the high FDR rate at low frequencies $f_v$ stems from the limited evidence provided by the observed bits in support of $v$.
This is clearly illustrated by
Figure \ref{fig:fdr2} where the probability that $v$ was reported (1) or not reported (0) by the client is plotted as a function of $f_v$.
For relatively rare strings (those with less than 10\% frequency), even when both bits corresponding to $v$ are set in the report,
the probability of $v$ being reported is much smaller than of it not being reported.
Because the prior probability $f_v$ is so small, a single client's reports cannot provide sufficient evidence in favor of $v$.
\subsection{Caution and Correlations}
Although it advances the state of the art,
\RAPPOR{} is not a panacea, but rather simply a tool
that can provide significant benefits
when used cautiously, and correctly,
using parameters appropriate to its application context.
Even then,
\RAPPOR{} should be used only
as part of a comprehensive privacy-protection strategy,
which should include
limited data retention and
other pragmatic processes mentioned in Section~\ref{sec:motivation},
and already in use by Cloud operators.
As in previous work on differential privacy for database records, \RAPPOR{} provides privacy guarantees for the responses from individual clients. One of the limitations of our approach has to do with ``leakage'' of additional information when respondents use several clients that participate in the same collection event. In the real world, this problem is mitigated to some extent by intrinsic difficulty of linking different clients to the same participant. Similar issues occur when highly correlated, or even exactly the same, predicates are collected at the same time. This issue, however, can be mostly handled with careful collection design.
Such inadvertent correlations can arise in many different ways in \RAPPOR{} applications, in each case possibly leading to the collection of too much correlated information from a single client, or user, and a corresponding degradation of privacy guarantees.
Obviously, this may be more likely to happen if \RAPPOR{} reports are collected, from each client, on too many different client properties.
However, it may also happen in more subtle ways.
For example, the number of cohorts used in the collection design must be carefully selected and changed over time,
to avoid privacy implications;
otherwise, cohorts may be so small as to facilitate the tracking of clients,
or clients may report as part of different cohorts over time,
which will reduce their privacy.
\RAPPOR{} responses can even affect client anonymity,
when they are collected on immutable client values that are the same across all clients:
if the responses contain too many bits (e.g., the Bloom filters are too large),
this can facilitate tracking clients,
since the bits of the Permanent randomized responses are correlated.
Some of these concerns may not apply in practice
(e.g., tracking responses may be infeasible, because of encryption),
but all must be considered in \RAPPOR{} collection design.
In particular, longitudinal privacy protection guaranteed by the Permanent randomized response assumes that client's value does not change over time. It is only slightly violated if the value changes very slowly. In a case of rapidly changing, correlated stream of values from a single user, additional measures must be taken to guarantee longitudinal privacy. The practical way to implement this would be to budget $\epsilon_{\infty}$ over time, spending a small portion on each report. In the \RAPPOR{} algorithm this would be equivalent to letting $q$ get closer and closer to $p$ with each collection event.
Because differential privacy deals with the worst-case scenario, the uncertainty introduced by the Bloom filter does not play any role in the calculation of its bounds. Depending on the random draw, there may or may not be multiple candidate strings mapping to the same $h$ bits in the Bloom filter. For the average-case privacy analysis, however, Bloom filter does provide additional privacy protection (a flavor of $k$-anonymity) because of the difficulty in reliably inferring a client's value $v$ from its Bloom filter representation $B$ \citep{bloom_privacy}.
\section{Related Work}
Data collection from clients in a way that preserves their privacy and at the same time enables meaningful aggregate inferences is an active area of research both in academia and industry.
Our work fits into the category of problems recently explored by \cite{Hsu2012, SafeZones, Chan2012, Liu2012}, where an untrusted aggregator wishes to learn the ``heavy hitters" in the clients' data or run certain types of learning algorithms on the aggregated data, while guaranteeing the privacy of each contributing client and, in some cases, restricting the size of the communication from the client to the untrusted aggregator. Our contribution is to suggest an alternative to those already explored that is intuitive, easy-to-implement, and potentially more suitable to certain learning problems, and to provide a detailed statistical decoding methodology for our approach, as well as experimental data on its performance. Furthermore, in addition to guaranteeing differential privacy, we make explicit algorithmic steps towards protection against linkability across reports from the same user.
It is natural to ask why we built our mechanisms upon randomized response, rather than upon two primitives most commonly used to achieve differential privacy: the Laplace and Exponential mechanisms~\cite{dwork06,McSherryT07}.
The Laplace mechanism is not suitable because the client's reported values may be categorical, rather than numeric, in which case direct noise addition does not make semantic sense. The Exponential mechanism is not applicable due to our desire to implement the system in a local model, where the privacy is ensured by each client individually without a need for a trusted third party. In that case, the client does not have sufficient information about the data space in order to do the necessary biased sampling required by the Exponential mechanism. Finally, randomized response has the additional benefit of being relatively easy to explain to the end user, making the reasoning about the algorithm used to ensure privacy more accessible than other mechanisms implementing differential privacy.
Usage of various dimensionality reduction techniques in order to improve the privacy properties of algorithms while retaining utility is also fairly common~\cite{Liu2012, JL, AggarwalY07, MirMNW11}. Although our reliance on Bloom filters is driven by a desire to obtain a compact representation of the data in order to lower each client's potential transmission costs and the desire to use technologies that are already widely adopted in practice~\cite{BroderM03}, the related work in this space with regards to privacy~\cite{bloom_privacy} may be a source for optimism as well.
It is conceivable that through a careful selection of hash functions,
or choice of other Bloom filter parameters,
it may be possible to further raise privacy defenses against attackers,
although we have not explored that direction in much detail.
The work most similar to ours is by Mishra and Sandler~\cite{MishraS06}. One of the main additional contributions of our work is the more extensive decoding step, that provides both experimental and statistical analyses of collected data for queries that are more complex than those considered in their work. The second distinction is our use of the second randomization step, the Instantaneous randomized response, in order to make the task of linking reports from a single user difficult, along with more detailed models of attackers' capabilities.
The work of~\cite{OurData} approaches the challenge of eliminating the need for a trusted aggregator with a distributed solution that places trust in other clients instead. \cite{chen2012towards} and \cite{Akkus2012} implement a differentially private protocol over distributed user data by relying on an honest-but-curious proxy or data aggregator bound by certain commitments.
Several lines of work aim to address the question of longitudinal data collection with privacy. The work of \cite{MedianMech} considers scenarios when many predicate queries are asked against the same dataset, and it uses an approach that, rather than providing randomization for each answer separately, attempts to reconstruct the answer to some queries based on the answers previously given to other queries. The high-level idea of \RAPPOR{} bears some resemblance to this technique--the Instantaneous randomized response is reusing the result of the Permanent randomized response step. However, the overall goal is different---rather than answering a diverse number of queries, \RAPPOR{} collects reports to the same query over data that may be changing over time.
Although it does not operate under the same local model as \RAPPOR{}, recent work by \cite{DworkNPRY10} on pan-private streaming and by~\cite{Dwork2010} on privacy under continual observation introduces additional ideas relevant for the longitudinal data collection with privacy.
\section{Summary}
\RAPPOR{} is a flexible, mathematically rigorous and practical platform for anonymous data collection
for the purposes of privacy-preserving crowdsourcing of population statistics on client-side data.
\RAPPOR{} gracefully handles multiple data collections from the same client by providing well-defined longitudinal differential privacy guarantees. Highly tunable parameters allow to balance risk versus utility over time, depending on one's needs and assessment of likelihood of different attack models. \RAPPOR{} is purely a client-based privacy solution. It eliminates the need for a trusted third-party server and puts control over client's data back into their own hands.
\paragraph{Acknowledgements.}
The authors would like to thank our many colleagues at Google and its Chrome team who have helped with this work, with special thanks due to Steve Holte and Moti Yung.
Thanks also to the CCS reviewers, and many others who have provided insightful feedback on the ideas, and this paper,
in particular, Frank McSherry, Arvind Narayanan, Elaine Shi, and Adam D.\ Smith.
\bibliographystyle{plainnat}
|
\section{Introduction}
This paper complements and generalizes the results obtained in
\cite{alessio} on the microscopic foundations of the
optics of materials. The main new result of that paper was a proof
of the existence of polaritons in ionic crystals,
that was obtained by calculating the normal modes and
exhibiting the explicit form of the dispersion curves.
Apparently, the existence
of polaritons, whose qualitative importance is evident since it explains
why crystals are transparent to visible light, was
previously understood only at a phenomenological level,
in terms of a macroscopic
polarization field (see for example \cite{gpp}, page 239).
An interesting point is that the microscopic proof was
obtained in \cite{alessio} in a completely classical framework,
on the basis of
the system of Newton's equations for each charge,
in which the full electrodynamic forces are taken into account,
both the mutual retarded ones and the individual radiation reaction
forces. For example it is just retardation that makes it possible
that the new polaritonic branches occur, and
Born and Huang \cite{bh} couldn't get this result just because
they didn't fully take the role of retardation into account.
The result
was obtained in \cite{alessio} by previously reducing
the original electrodynamic model to a Hamiltonian conservative
one. This in turn was made possible by exploiting two
global properties of the original microscopic electrodynamic system, namely,
the Wheeler-Feynman identity \cite{wf} and the Ewald--Oseen resummation of
the far fields (see \cite{ewald}\cite{oseen2} and \cite{bw}, page 101),
which, jointly used, provide both a cancellation of the
radiation reaction force acting on each charge, and an elimination
of the problems related to delay. Both properties
were proven in \cite{alessio} (following \cite{cg} and \cite{mcg}),
for the case of ionic crystals.
It is then natural to ask whether such a result concerning the
dispersion curves may be complemented by providing a microscopic
expression for the electric susceptibility of the system, which
would allow one to determine the expected
absorption and emission spectra. Moreover one might also look for
an extension of the methods,
formulating the theory in such a general frame that it can apply to
disordered dielectric systems such as glasses, or even liquids or gases.
In the present paper we show how a microscopic expression for
susceptibility is obtained for ordered systems, and how the
result can be extended,
at least partly, to cover the case of disordered systems.
Indeed we will show how, if
the two mentioned global properties hold (so that the system can be
reduced to a conservative Hamiltonian one), then
the statistical mechanical methods of
Green--Kubo type \cite{gc}\cite{gc2} can be used to provide a microscopic
expression for macroscopic polarization, and so for susceptibility.
In particular, it will be explicitly exhibited that
the phenomena of absorption and
emission are not related, at least in a any direct way, to the
radiation reaction force, and can in fact be understood as symmetrical
features of a time reversible dynamics.
In order to obtain such results, we have to overcome a
difficulty which arises if one tries to imitate in a strict way the
Green--Kubo type methods generally used in the quantum case.
Indeed, the available procedure makes use, in an apparently
essential way, of the Gibbs measure in phase space, whereas Gibbs' measure
does not even exist in the classical case, due to the divergence
induced by the attractive Coulomb potentials. We however show how
susceptibility can actually be proven to exist, obtaining for it an
expression in terms of time correlations. Then we
study its properties, and in particular deduce the $f$--sum rule,
the essentially
classical character of which was already pointed out by Van Vleck and
Huber \cite{vanpaper}.
The existence of susceptibility and its general expression are
completely independent of the qualitative nature of the motions of the system.
It is instead the form of the spectrum that depends on the
stability properties of the motions. We show how
a pure line spectrum occurs for stable (almost periodic) motions
of the system, while a broadening of the lines or even a
continuous spectrum occur when chaoticity sets in. We also
discuss the relevance that in this connection have some quite recent
results on the theory of dynamical systems, in particular the results
that made possible to extend to systems of interest
for statistical mechanics the methods of perturbation
theory \cite{andrea}\cite{maiocchi}\cite{fpu} which allow
one to estimate when a
transition from ordered to chaotic motions should occur
(see the numerical works \cite{cggp}\cite{plasmi2}).
For what concerns the extension to disordered systems, all depends on
proving the two mentioned global electrodynamic properties.
For the Wheeler-Feynman identity, we do here more than required,
because we give a proof
which applies to completely general
systems, and not just to dielectrics.
In fact, the identity is shown to be equivalent to
a general form of causality of
electrodynamics, which is reminiscent of a general property
assumed in quantum
electrodynamics. The properties related to the Ewald
resummation methods are instead assumed to hold for dielectrics,
just by analogy with the case of ordered systems.
In section \ref{2} it is recalled how a first step in passing
from microscopic to
macroscopic electromagnetism consists in performing a local
space--average.
Our treatment is standard, apart from a minor point.
In section \ref{3} the second step is performed, which involves a
phase space (or ensemble) average, and leads to a
Green--Kubo type formula for macroscopic susceptibility,
in a completely symmetrical
way for absorption and emission. The proof
is obtained without using the Gibbs measure.
Preliminarily, it is recalled how the reduction to a
conservative Hamiltonian system is obtained through the Ewald
resummation methods, making use of the
Wheeler--Feynman identity.
In section \ref{4} the analyticity properties of susceptibility are
recalled, and the $f$--sum rule is proven. In section \ref{5} it is
shown how under quite general conditions susceptibility is
expressed in terms of equilibrium time--correlation functions between
positions and velocities of the charges.
In section \ref{6} it is discussed how the spectrum depends on the
qualitative stability properties of the motions of the system, and in
particular how a pure spectrum arises in the presence of suitable
stability properties (almost periodicity) of the
orbits. Instead, a broadening of the lines, or even a continuous spectrum
are expected to occur as chaoticity sets in. In section \ref{7}
this is illustrated by studying the particular case of
ionic crystals.
Some final comments are added in Section \ref{8} .
An Appendix is devoted to a proof of the
Wheeler--Feynman identity (and of the consequent cancellation of the
radiation reaction forces), which applies in a completely general
situation, irrespective of the ordered or disordered structure of the
system.
\section{From microscopic to
macroscopic electromagnetism. First step:
local space--averages and the microscopic polarization field}\label{2}
As we know,
\cite{drude}\cite{lorentz}\cite{born}\cite{vanbook}\cite{degroot}\cite{degroot2}
macroscopic electromagnetism is characterized by four fields:
the electric field $ {\boldsymbol{\mathcal{E}}} $,
the magnetic induction field $ {\boldsymbol{\mathcal{B}}} $, the electric induction
field $\boldsymbol{\mathcal{D}}$
and the magnetic field $\boldsymbol{\mathcal{H}}$.
Since the times of Lorentz, the first two
are thought of as local space--averages of corresponding microscopic fields
$\vett{E}$, $\vett{B}$, while the latter ones are defined as
$\boldsymbol{\mathcal{D}}= {\boldsymbol{\mathcal{E}}} +4\pi {\boldsymbol{\mathcal{P}}} $ and $\boldsymbol{\mathcal{H}}
= {\boldsymbol{\mathcal{B}}} -4\pi {\boldsymbol{\mathcal{M}}} $, where the polarization vector $ {\boldsymbol{\mathcal{P}}} $ and
the magnetization vector $ {\boldsymbol{\mathcal{M}}} $ are the response
of a material body to the presence of an external electric or
magnetic field. In the macroscopic treatments one assumes
that there hold the constitutive relations
$\boldsymbol{\mathcal{D}}=\varepsilon {\boldsymbol{\mathcal{E}}} $ and
$\boldsymbol{\mathcal{H}}=\mu {\boldsymbol{\mathcal{B}}} $,
or rather that analogous relations hold frequency by frequency,
i.e., that one has
$$
\hat{\boldsymbol{\mathcal{D}}} (\vett x,\omega)=\varepsilon(\omega)\hat {\boldsymbol{\mathcal{E}}} (\vett
x,\omega) \ , \quad
\hat{\boldsymbol{\mathcal{H}}} (\vett x,\omega)= \mu(\omega)
\hat {\boldsymbol{\mathcal{B}}} (\vett x,\omega) \ ,
$$
where $\hat {\boldsymbol{\mathcal{E}}} $, $\hat{\boldsymbol{\mathcal{D}}}$, $\hat {\boldsymbol{\mathcal{B}}} $ and
$\hat{\boldsymbol{\mathcal{H}}}$, are the time Fourier transforms of
the corresponding fields. In this section we recall how, in order to obtain
a macroscopic expression for polarization, a first step
is accomplished through a local space--averaging
procedure. This is a completely standard passage, and only a minor modification
to the familiar procedure will be introduced.
Consider a dielectric body, thought of as
microscopically constituted of a certain number $N$ of
neutral molecules or atoms, each containing a stable aggregate of point
charges. In such a case the microscopic Maxwell equations
read
\begin{equation*}
\begin{split}
\Div \vett E & = 4\pi \sum_{k=1}^N \sum_{j=0}^{n_k}
e_j\delta(\vett x-\vett {x}_{j,k}) \\
\Rot \vett E & = -\, \frac {1}c\partial_t \vett B \\
\Div \vett B & = 0 \\
\Rot \vett B & = \frac {4\pi}c \sum_{k=1}^N \sum_{j=0}^{n_k} e_j
{\dot{\vett x}}_{j,k} \delta(\vett x - \vett{x}_{j,k}) +
\frac {1}c\partial_t \vett E \ ,
\end{split}
\end{equation*}
where $\vett{x}_{j,k}$ is the position of the $j$--the particle (of charge
$e_j$) in the $k$--th molecule or atom.
\subsection*{The local space--averaging procedure. Space--averaged fields
and sources}
Now, following Lorentz, the
macroscopic fields $ {\boldsymbol{\mathcal{E}}} $ and
$ {\boldsymbol{\mathcal{B}}} $ are defined as local space--averages
of the values the microscopic fields take in
what is sometime called
a ``physically infinitesimal domain''\cite{degroot}, or a ``physically
small volume element'' \cite{kirk}, of volume $\Delta V$
located about the considered
point $\vett x$. Think for example of a cubic volume element with side
$100$ \"Amstrong, which, in a solid or in a liquid, in ordinary
conditions contains about one million molecules.
Due to the linearity of the Maxwell equations, the
space--averaged fields are expected to be solutions of those
same equations, having as sources
the averaged charge and current densities.
This becomes a rather simple theorem if
the space--averaging procedure at $\vett x$ is mathematically implemented
through a convolution with a suitable smooth
($C^\infty$ class) function $N(\cdot )$ centered at $\vett x$, which
essentially vanishes outside the chosen volume element, while
having inside it essentially a constant normalizing
value, namely, $1/\Delta V $.
The macroscopic fields are thus defined as
\begin{equation*}
\begin{split}
{\boldsymbol{\mathcal{E}}} (\vett x,t) &= N\ast \vett E\;(\vett x,t) \stackrel {\mathrm{def}}{=}
\int_{\mathbb{R}^3} \mathrm{d}\vett y N(\vett x -\vett y)\vett E(\vett y,t) \\
{\boldsymbol{\mathcal{B}}} (\vett x,t) &= N\ast \vett B\;(\vett x,t) \stackrel {\mathrm{def}}{=}
\int_{\mathbb{R}^3} \mathrm{d}\vett y N(\vett x -\vett y)\vett B(\vett y,t) \ .
\end{split}
\end{equation*}
As the microscopic fields are distributions
(because $\delta$ functions occur in the sources), it turns out
that the differential operators commute with the
convolution, i.e., one has
\begin{equation*}
\begin{split}
\Div {\boldsymbol{\mathcal{E}}} &= N \ast \Div \vett E \ , \quad
\Rot {\boldsymbol{\mathcal{E}}} = N \ast \Rot \vett E \\
\Div {\boldsymbol{\mathcal{B}}} &= N \ast \Div \vett B \ , \quad
\Rot {\boldsymbol{\mathcal{B}}} = N \ast \Rot \vett B \ ,
\end{split}
\end{equation*}
exactly as it would occur if the fields were smooth.
Thus,
multiplying the Maxwell equations by $N(\vett x-\vett y)$ and
integrating, due to the linearity of the equations
the macroscopic fields are found, as expected, to satisfy the
Maxwell equations
with charge density $\rho$ and current density
$\vett j(\vett x, t)$ which now are smooth fields rather than
distributions, and are obtained by averaging with the same procedure.
So the macroscopic fields satisfy the equations
\begin{equation*}
\begin{split}
\Div {\boldsymbol{\mathcal{E}}} &= 4\pi \rho \\
\Rot {\boldsymbol{\mathcal{E}}} &= - \frac {1}c\partial_t {\boldsymbol{\mathcal{B}}} \\
\Div {\boldsymbol{\mathcal{B}}} &= 0 \\
\Rot {\boldsymbol{\mathcal{B}}} &= \frac {4\pi}c \vett j
+ \frac {1}c\partial_t {\boldsymbol{\mathcal{E}}} \ .
\end{split}
\end{equation*}
which involve the space--averaged sources
\begin{equation}\label{eq:carica}
\rho (\vett x,t) \stackrel {\mathrm{def}}{=} \sum_{k=1}^N \sum_{j=0}^{n_k} e_j N(\vett x -
\vett x_{j,k})
\end{equation}
\begin{equation}\label{eq:corrente}
\vett j(\vett x, t) \stackrel {\mathrm{def}}{=} \sum_{k=1}^N \sum_{j=0}^{n_k} e_j
\dot{\vett x}_{j,k} N(\vett x-\vett x_{j,k})\ .
\end{equation}
\subsection*{The microscopic polarization field}
We now show how the space--averaged charge density $\rho$ can be
written as the divergence of a field, which should be interpreted as a still
microscopic form of the polarization field. This is obtained by expanding
the positions of the charges entering the function $N(\cdot)$,
about the centers of mass of
their molecules or atoms.
This makes the single microscopic dipoles come in.
Denote by $\vett{x}_k^0$ the position of the center of mass of the $k$--th
molecule or atom,
\footnote{In the case of crystals
the formulas are simplified if one even thinks of $\vett{x}_k^0$
as a fixed position of a cell, for example a given corner.}
and by $\vett{q}_{j,k} \stackrel {\mathrm{def}}{=} \vett x_{j,k}-\vett{x}_k^0$ the corresponding displacements
(which are assumed to be bounded)
of the charges. We have now to find
which expression does the space averaged charge density $\rho$ take, as a
function of the displacements $\vett{q}_{j,k} $. Here the familiar
procedure consists in introducing a multipole expansion and a
truncation, through which $\rho$ is shown to be the divergence of a
vector field.
We obtain this result, perhaps in a simpler and
more rigorous way, by
making use of the finite--increment Lagrange
formula, according to which for a smooth function $f$ one has
$$
f(\vett x+\vett h)- f(\vett x)=\int_0^1d\zeta\, \frac{d~}{d\zeta}
f(\vett x+\zeta \vett h)\ .
$$
Indeed one then has
\begin{equation*}
\begin{split}
N(\vett x & -\vett x_{j,k}) = N(\vett x-\vett{x}_k^0) + \int_0^1 \mathrm{d} \zeta
\frac{d~}{d\zeta}
\, N(\vett x-\vett{x}_k^0 - \zeta \vett{q}_{j,k} ) \\
& = N(x-\vett{x}_k^0) - \int_0^1 \mathrm{d} \zeta \; \vett{q}_{j,k} \cdot \nabla N(\vett x-\vett{x}_k^0 -
\zeta \vett{q}_{j,k} ) \\
& = N(\vett x-\vett{x}_k^0) - \Div \Big( \vett{q}_{j,k} \int_0^1 \mathrm{d}
\zeta \, N(\vett x-\vett{x}_k^0 - \zeta \vett{q}_{j,k} ) \Big)\ .
\end{split}
\end{equation*}
Thus, substituting this formula in the expression (\ref{eq:carica}) for the
space--averaged charge density $\rho$, and
recalling that the
molecules are neutral so that
$$
\sum_{j=0}^{n_k} e_j N(\vett x-\vett{x}_k^0) = 0 \ ,
$$
one finds
\begin{equation}\label{divergenza}
\rho = - 4 \pi \Div \vett P \ ,
\end{equation}
where the field $\vett P$ is given by
\begin{equation}\label{eq:defpol}
\vett P(\vett x) \stackrel {\mathrm{def}}{=} \sum_{k=1}^N \sum_{j=0}^{n_k} e_j \Big( \vett{q}_{j,k} \int_0^1 \mathrm{d}
\zeta \, N(\vett x-\vett{x}_k^0 - \zeta \vett{q}_{j,k} ) \Big)\ .
\end{equation}
Without much error this can be written
in the simplified form
\begin{equation}\label{eq:polarizzazione}
\vett P(\vett x) = \frac 1\Delta V \sum_{\vett{x}_k^0\in\Delta V } \sum_{j=0}^{n_k} e_j
\vett{q}_{j,k} \ ,
\end{equation}
i.e., as the sum of the dipole moments of the single molecules or atoms with
respect to their centers of mass, as one might have expected.
On the other hand we
know that, in a dielectric,
the macroscopic charge density is expressed as the divergence of
polarization.
So one might be tempted to altogether
identify $\vett P$ with the macroscopic polarization $ {\boldsymbol{\mathcal{P}}} $
itself. This however is not correct. The reason is that the field
$\vett P(\vett x)$ still is a dynamical variable, by which we mean a
function defined on the global ``mechanical phase space'' of the
charges, a point of which, call it $z$, is identified through
the positions and the momenta of all
charges. Now, $\vett P(\vett x)$ evidently depends on the phase point,
and thus
may be called the \emph{microscopic polarization field }.
The microscopic magnetization field could be given
along similar lines. However we don't need it for our aims, because
with good approximation in dielectrics one can put $\mu=1$, unless one
is just interested in magneto--optical phenomena.
\subsection*{Need for an ensemble average}
As usual in statistical mechanics, a macroscopic quantity is
defined as the average over phase space of a microscopic quantity
(a function of $z$), with respect to a given measure.
Denoting such an averaging in the mechanical phase space
by $\langle\cdot \rangle$,
the macroscopic polarization field will then be defined by
$$
{\boldsymbol{\mathcal{P}}} (\vett x) = \langle \vett P(\vett x)\rangle \ ,
$$
i.e., by
\begin{equation*}
{\boldsymbol{\mathcal{P}}} (\vett x) = \frac 1\Delta V \big\langle
\sum_{\vett{x}_k^0\in\Delta V } \sum_{j=0}^{n_k} e_j \vett{q}_{j,k} \big\rangle \ .
\end{equation*}
Now, the microscopic polarization, being itself a space--mean over many
molecules, should already satisfy some central limit theorem and so should not
fluctuate very much as the phase space point $z$ is varied.
In such a case the ensemble average just provides a ``typical value'',
so that the use of a further ensemble average
may appear to be redundant. This is not so, because it is
just by performing ensemble averages that analytical
manipulations can be performed which lead to significant results.
One such result, as we will see,
is the existence itself of electric susceptibility, namely, the fact that
polarization responds linearly to an external perturbation even if the
unperturbed system presents highly nonlinear motions. This is obtained
by Green-Kubo methods in phase space, just because of the
linearity of the equation of motion for the probability density.
A further result is the proof
of the $f$--sum rule.
However, it is not at all obvious
how phase space methods can be used in a microscopic model
which involves both retarded forces
and dissipative ones. How to do this, and how to use Hamiltonian
techniques in phase space will be shown in the next section.
\section{Ensemble average and Green--Kubo theorem for
polarization. Role of the Wheeler--Feynman identity and of the
Ewald resummation methods}\label{3}
\subsection*{Reduction to the mechanical phase space
(Wheeler--Feynman and Ewald--Oseen)}
The reduction of the original electrodynamic problem to a purely
mechanical one in the mechanical phase space is quite hard a task.
First of all, the original problem is different from those usually studied
in statistical mechanics because, due to the finite
propagation speed of the
electromagnetic interactions among the charges, the equations of motion for the
displacements $\vett{q}_{j,k} $ of the charges
turn out to be differential equations with delay. Notice that the delay
cannot be neglected, as it
produces qualitatively essential features. For example, in the case of
ionic crystals it is just retardation that produces the two new
branches of the
dispersion relation which correspond to polaritons (see formula (15)
of (\cite{alessio}),
thus explaining why visible light can propagate inside them.
Thus, in the original electrodynamic problem, having to deal with
equations with delay we know nothing about the properties of the
corresponding dynamical system,
not even how to correctly frame a Cauchy problem. Neither do we
know which is the phase space suited to the system,
nor can we know which measure should be used to define the averages.
Finally, the system is not a conservative one, at least not in any obvious
way, inasmuch as the charges should radiate energy away during their
necessarily accelerated motions.
From a heuristic point of view such problems can be overcome in the
following way. Due to the long range character of the field produced by any
single charge (a range much longer than the purely Coulomb one), in order to
determine the force acting on any charge and produced by all the other
ones, one necessarily has to perform a ``resummation'' of the forces.
This can be done in an exact way in the case of crystals (through the
so called Ewald method, as implemented for
example in \cite{alessio})
by suitably splitting the field into two contributions. The first one
essentially comes from the near (in a microscopic sense) charges,
and can thus be considered to all effects as being instantaneous,
while the second one is essentially due to the far charges.
In turn, the contribution of the far charges too can be divided into
two parts. One of them exactly cancels the radiation reaction
force (which necessarily is nonvanishing, because of the accelerated
motions of the charges). This indeed
is the so called Wheeler--Feynman identity, which was postulated by
those authors in their paper of the year 1945 and was proven, in the case
of ionic crystals, in \cite{alessio}, following \cite{cg} and \cite{mcg}.
The second part of the contribution of the far charges
enters in the same way as an external electromagnetic
field, which propagates inside matter with a suitable refractive
index (see the first term in the force entering formula (15)
of (\cite{alessio}), notwithstanding the fact that the microscopic far
fields entering the original problem
do propagate with the speed of light in vacuum (this is the so--called
Ewald--Oseen cancellation
property). So we have to deal
both with the Wheeler--Feynman property (or identity) and with
the Ewald--Oseen resummation properties.
In the case of ionic crystals both properties were proved to hold,
so that the original electrodynamic equations of motion
for the charges could be consistently dealt with
as a system of non dissipative differential equations
(possibly depending on time), of the form
$$
m_j\ddqjk = \sum_{\vett{x}_{k'}^0 \in U} \sum_{j'}\vett F_{j,j'}(\vett{q}_{j,k} - \vett{q}_{j',k'} ) + e_j
{\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,t)
$$
where $U$ is a microscopic (namely, much smaller than $\Delta V $) neighborhood
of $\vett{x}_k^0$, while the field $ {\boldsymbol{\mathcal{E}}^c} $ is what Ewald calls the
``exciting'' electric field (\emph{``erregende Feld''} in his words,
see \cite{ewald}, page 7). This is the field produced by the far
charges that actually enters the equations of motion
as if it were an external field, propagating with a macroscopic
refractive index.
Analogous proofs should be provided here for the disordered case of
interest for dielectrics.
For what concerns the Wheeler--Feynman
identity, we here do more than required, because we give in an Appendix
a proof which holds in any situation, and
actually shows the deep significance of the identity, as corresponding
to some general form of causality.
Instead, the Ewald--Oseen property is not proven here for the case of
disordered systems, and its validity is assumed to hold by analogy with the
case of crystals. We are confident that a proof may be provided on
another occasion.
\subsection*{The macroscopic polarization through a Green--Kubo
type theorem. General expression of the response function for an
absorption process }
So our phase space can be taken to be the usual one of
statistical mechanics, namely, the space having as coordinates
the positions $\vett{q}_{j,k} $ and the momenta $\vett{p}_{j,k} \stackrel {\mathrm{def}}{=} m_j\dqjk$
of all the charges of the system, and our aim is now to obtain an
expression for the electric susceptibility following the standard methods
of Green--Kubo type of quantum statistical mechanics.
Here however a difficulty arises.
Indeed the analogous methods transported to the classical case
amount to studying the Liouville equation
for the probability density in phase space, looking for its time
evolution under the action of a perturbation. However, in the quantum
case it is first of
all assumed that an unperturbed (or equilibrium) solution exists,
given exactly by the
Gibbs ensemble. Now, if one looks
at the procedures used in the proofs, one might have the impression
that the role of the Gibbs density is essential, and that the proof
couldn't be obtained without using it.
On the other hand we
have to deal with Coulomb attractive interactions,
which have the effect that the Gibbs measure does not even exist, in
the classical case.
We show here how any reference to the equilibrium Gibbs measure can be
avoided, and even in a rather simple way.
Indeed in this section the existence of susceptibility is proven, and a quite
general
expression for it is provided, essentially without introducing
any requirement at all on the equilibrium measure. Then in section
\ref{5} it will be shown how
susceptibility is expressed in terms of time--correlation functions,
if an assumption of a quite general character for the measure is introduced
(validity of the large deviation principle for momenta).
So we only assume that an equilibrium probability density exists, which
will be denoted by $\rho_0$ (no confusion
with the space--averaged charge density should occur), and
its form will not need be specified.
In other terms, $\rho_0$ is only assumed to be invariant under the
flow determined by the equations of motion, i,e.,
to be a stationary solution of the continuity equation
$$
\partial_t\rho + \vett v\cdot\nabla\rho=0 \ ,
$$
where $\vett v$ is the vector field defined by the equations of
motion in phase space for the isolated system.\footnote{For the sake
of
simplicity
we are admitting that the vector field $\vett v$ has vanishing
divergence. Nothing should change in the general case.}
Consider now the case in which there is an external electromagnetic
field $\vett E^{in}$
(for example a monochromatic wave of frequency $\omega$) which incides
on the body, with an intensity that starts increasing slowly and
then reaches a stationary
value (the so called case of an adiabatically switched on perturbation).
Then a change, say $\delta {\boldsymbol{\mathcal{E}}^c} (\vett x,t)$, will be induced on the
Ewald exciting field, which is
the one actually entering the
equations of motion for the charges. The change is due
both to the presence itself of the incoming external
field, and to the fact that the far charges which are responsible for
that field are now
moving in a modified way.
For the sake of consistency, the relation
between $\delta {\boldsymbol{\mathcal{E}}^c} $ and the incoming external
field $\vett E^{in}$ should be determined, and to this end
the validity of the Lorentz--Lorenz relation should be established. This is in
any case a necessary step, if macroscopic optics should be deduced at
all. This problem will not be dealt with in the present paper.
Under the perturbation induced by the external field,
the density $\rho$ will evolve according to the
equation
\begin{equation}\label{rho1}
\partial_t \rho + \vett v\cdot\nabla\rho +\sum_{k,j}
e_j \delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,t) \frac {\partial\rho}{\partial \vett{p}_{j,k} } =0 \ ,
\end{equation}
inasmuch as the equation of motion for $\vett{q}_{j,k} $ contains the further force
term $e_j \delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,t)$.
As $\delta {\boldsymbol{\mathcal{E}}^c} $ is assumed to be a small perturbation, one can look
for the solution as a series expansion
$$
\rho = \rho_0 + \rho_1 + \ldots \ ,
$$
and the first order term $\rho_1$ is immediately seen to satisfy
the equation
\begin{equation}\label{xxx}
\partial_t \rho_1 = - \vett v\cdot\nabla\rho_1 - \sum_{k,j}
e_j \delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,t) \frac {\partial\rho_0}{\partial \vett{p}_{j,k} } \ .
\end{equation}
Clearly the suited ``initial'' condition is the asymptotic one
\begin{equation}\label{incoming}
\rho_1\to 0\quad \mathrm{for}
\quad t\to-\infty\ ,
\end{equation}
and the corresponding well known solution is then
\begin{equation}
\rho_1(z,t) = - \int_{-\infty}^t\mathrm{d} s \sum_{k,j}
e_j \delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,s) \frac {\partial\rho_0}{\partial
\vett{p}_{j,k} }\Big(\Phi^{s-t} z \Big) \ ,
\end{equation}
where $\Phi^t z$ is the flow
relative to the \emph{unperturbed} equations of motion,
The macroscopic polarization $ {\boldsymbol{\mathcal{P}}} (\vett x,t)$ can now be computed
to first order, as the average of the microscopic polarization
$\vett P(\vett x,t)$ with respect to the density
$\rho_0 +\rho_1$. Assuming that $ {\boldsymbol{\mathcal{P}}} $
vanishes at equilibrium (absence of ferroelectricity),
one remains with the contribution of $\rho_1$ only, which gives
\begin{equation}\label{eq:polarizza}
\begin{split}
{\boldsymbol{\mathcal{P}}} (\vett x,t) = &
- \int \mathrm{d} z\, \vett P(\vett x,t)
\int_{-\infty}^t\mathrm{d} s\\ & \sum_{k,j} e_j \delta
{\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,s)
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} }\Big(\Phi^{s-t} z \Big)\ .
\end{split}
\end{equation}
One has now to insert the expression (\ref{eq:polarizzazione}) for
the microscopic polarization
$\vett P(\vett x,t)$. Then, first of all one performs
two elementary transformations
(namely, interchange of the integration orders of $s$ and $z$, an change of
variable
$z\to\Phi^{t-s}z$ -- taking into account that the modulus of the jacobian
determinant of $\Phi^t z$ is unitary\footnote{Because the unperturbed
vector field has vanishing divergence.}).
Moreover, one uses the fact that
$\delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,s)$, being a
macroscopic field, takes on essentially the same
value $\delta {\boldsymbol{\mathcal{E}}^c} (\vett x,s)$ at all points of the volume element $\Delta V $.
This eventually produces the result that macroscopic
polarization depends linearly on the exciting field. So the
macroscopic polarization can be written in
the familiar form of linear response theory, namely as
\begin{equation}\label{eq:rispostalineare}
{\boldsymbol{\mathcal{P}}} (\vett x,t) = \int_{-\infty}^t\mathrm{d} s \; \delta {\boldsymbol{\mathcal{E}}^c} (\vett x ,s)
\tilde\chi(t-s) \ ,
\end{equation}
in terms of a dielectric response function
$\Tilde\chi(t)$, which is given by
\begin{equation}\label{eq:suscettivita}
\Tilde\chi(t) \stackrel {\mathrm{def}}{=} - \frac 1\Delta V
\int \mathrm{d} z\, \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'=0}^{n_k}
e_j e_{j'}\, \vett{q}_{j',k'} (t)\, \frac {\partial\rho_0}{\partial\vett{p}_{j,k} } \ .
\end{equation}
Actually, in this expression for the response function
we have introduced one more simplification. This consists in the
fact that, when the expression (\ref{eq:polarizzazione})
for the microscopic polarization
$\vett P$ is introduced into formula (\ref{eq:polarizza}),
one has two sums over $k$
and $k'$, corresponding to two volume elements, whereas now
the first sum was restricted to just the molecules
that belong to the volume element entering the second sum.
This amounts to presuming that the microscopic dynamics in
two different macroscopic volume elements be totally
uncorrelated. This point will be discusses later.
We add now some comments.
The first one concerns the fact that in the deduction of the
formula for the dielectric response function
no reference at all was made to nonconservative forces. Indeed,
it was explicitly assumed that in the equation of motion for each
charge the radiation reaction force be canceled by a
part of the retarded forces due to the ``far'' charges of the
dielectric body.\footnote{Curiously enough, the radiation reaction
force is still taken into consideration in the paper of Callen and
Welton \cite{callen} which is usually considered to be the first
modern work on the fluctuation--dissipation theorem.}
The first scientist who realized the occurring of this
cancellation (already
in the year 1916) is the
Swedish physicist Oseen \cite{oseen}. However, his
result was ignored, having even been qualified as wrong
(\emph{``irrig''} (see \cite{jaffe}, page 266), as also was essentially ignored the
work of Wheeler and Feynman, in which the same
property was conjectured to hold quite in general.
So we are dealing with a time--reversible dynamical system.
An asymmetry in the proof was however introduced above through the
choice of the incoming
external field ${\vett E}^{in}$ (which was adiabatically switched on), and through the
corresponding choice (\ref{incoming}) for the ''initial'' (or rather,
asymptotic in the past) condition needed to solve the continuity equation for
the probability density (vanishing of $\rho_1$ as $t\to -\infty$).
Clearly. these are the choices which are responsible for the fact
that the formula just found corresponds to
an absorption process. How an emission process can be analogously
described in the present time--reversible frame,
will be shown in the next subsection.
The second remark is that the proof shows how
the existence of a linear response to the
external field is quite independent of the nature of the unperturbed
motions, which may have either an ordered or a disordered
character. The linearity of the response is inherited from that of the
Liouville equation, under the only assumption that the higher order
corrections (beyond the first one) to the equilibrium solution be
negligible.This fact is characteristic of linear response theory,
and so also occurs in its present classical formulation in phase
space. The situation
was quite different with the older approaches. In the oldest one,
typically described in Drude's book \cite{drude} but still somehow
surviving
in the Born--Wolf book \cite{bw}, to each observed spectral line was
associated the motion of a material oscillator, which was supposed to
perform linear oscillations, forced by the inciding
field. For example, in the words of Kronig \cite{kronig}, in that
approach one is dealing with
\emph{``an electric charge,
elastically bound to an equilibrium position, having} -- as he even
adds -- \emph{a damping proportional to its velocity''}.
A different attitude was taken by Van Vleck \cite{van24} who,
working in the spirit of Bohr's approach, thought it appropriate to
formulated a theory of susceptibility by assuming that the
unperturbed system performs quasi periodic motions.
Here, instead, essentially no property is required
for the unperturbed motions.
\subsection*{Emission process}
The proof of the existence of a linear response was given above
in a way suited to describe an absorption process. However, the proof was
formulated in the general frame of a time-reversible
dynamics, in such a way that different types of nonequilibrium processes
can be looked upon as determined by an asymmetry of the asymptotic
conditions. So an emission process
should be described by the same equations previously considered,
just choosing a suitable asymptotic condition, and external field
(see \cite{dirac}).
The suitable asymptotic condition can be inferred in the following way.
Recall how the absorption process was described.
For $t\to-\infty$ we have a stationary state described
by an equilibrium probability density $\rho_0$, in the presence of
a well defined exciting field $ {\boldsymbol{\mathcal{E}}^c} $. A perturbation is then introduced
through a ``free'' field ${\vett E}^{in}$, incoming
from infinity. During the process, one has a density
$\rho_0+\rho_1$ and a corresponding exciting field $ {\boldsymbol{\mathcal{E}}^c} +\delta {\boldsymbol{\mathcal{E}}^c} $, and one
presumes that eventually, for $t\to+\infty$,
one will have a new equilibrium (at a higher energy), with a density
$\rho_0'=\lim\big(\rho_0+\rho_1\big)$, together
with a new exciting field ${ {\boldsymbol{\mathcal{E}}^c} }'$ and a new free field $\vett
E^{out}$. Moreover, one should have $\vett E^{out}\simeq 0$, as
the whole incoming field is supposed to have been absorbed.
Let us now consider the inverse process, namely, the one which is
obtained with the interchanges $t \to -t$ and
$\vett{p}_{j,k} \to -\vett{p}_{j,k} $ (the Hamiltonian being assumed to be even in the
momenta). So one starts up with a density
$\rho_0'$ at $t=-\infty$, and asymptotically when
$t\to +\infty$ one gets a density $\rho_0$, whereas the electric field
is now the sum of the exciting field $ {\boldsymbol{\mathcal{E}}^c} $ and of the free field $\vett
E^{in}$. This means that the field $\vett E^{in}$ was emitted from the body,
in passing from the state $\rho_0'$ to the state $\rho_0$.
Mathematically, the process is still described through the perturbed
continuity equation
(\ref{rho1}),
provided the asymptotic condition
\begin{equation*}
\rho\to \rho_0\quad \mathrm{for}
\quad t\to +\infty\ ,
\end{equation*}
be assumed. If, as in the case of the absorption process, we look for
the solution in the form of a series,
the first correction $\rho_1$ has to satisfy the same equation
(\ref{xxx}) as before,
but now with the ``final'' condition
\begin{equation}\label{outcoming}
\rho_1\to 0\quad \mathrm{for}
\quad t\to+\infty\
\end{equation}
So the solution now has the form
\begin{equation}
\rho_1(\vett x,t) = \int_t^{+\infty}\mathrm{d} s \sum_{k,j}
e_j \delta {\boldsymbol{\mathcal{E}}^c} (\vett{x}_k^0,s) \frac {\partial\rho_0}{\partial
\vett{p}_{j,k} }\Big(\Phi^{s-t} z \Big) \ ,
\end{equation}
and thus, in the same hypotheses as before, the final polarization can be
written as
\begin{equation}\label{eq:polarizza2}
{\boldsymbol{\mathcal{P}}} (\vett x,t) = \int_t^{+\infty}\mathrm{d} s \; \delta {\boldsymbol{\mathcal{E}}^c} (\vett x ,s)
\tilde \chi (t-s)\ ,
\end{equation}
with $\tilde \chi$ given exactly by the expression
(\ref{eq:suscettivita})
that occurs in the absorption process.
\section{ Susceptibilities for absorption and for emission. Analyticity
properties, and the $f$--sum rule}\label{4}
\subsection*{Susceptibilities}
Susceptibilities are defined as responses to forcings of given
frequencies, and thus are obtained from the formulas
(\ref{eq:rispostalineare}) and
(\ref{eq:polarizza2})
if the latter are expressed in the form of convolutions,
namely, with integrals over the whole real axis ${\mathbb{R}}$.
Thus we introduce the functions
\begin{equation}
\chi^{abs}(t) \stackrel {\mathrm{def}}{=}
\left\{
\begin{array}{cc}
\tilde\chi(t)&\quad\mbox{for}\quad t > 0\ , \\
0&\quad\mbox{for}\quad t\le 0 \\
\end{array}
\right.
\end{equation}
\begin{equation}
\chi^{em}(t) \stackrel {\mathrm{def}}{=}
\left\{
\begin{array}{cc}
0&\quad\mbox{se}\quad t>0 \\
-\tilde\chi(t)&\quad\mbox{se}\quad t\le 0 \\
\end{array}
\right.
\end{equation}
so that
through the change of variables $s\to t-s$ formulas
(\ref{eq:rispostalineare}) and (\ref{eq:polarizza2})
for the polarizations
in an absorption or an emission process take the form
$$
{\boldsymbol{\mathcal{P}}} (\vett x,t) = \int_{\mathbb{R}}\mathrm{d} s \; \delta {\boldsymbol{\mathcal{E}}^c} (\vett x ,t-s)
\chi^{abs}(s) \ ,
$$
$$
{\boldsymbol{\mathcal{P}}} (\vett x,t) = \int_{\mathbb{R}}\mathrm{d} s \; \delta {\boldsymbol{\mathcal{E}}^c} (\vett x ,t-s)
\chi^{em}(s) \ ,
$$
namely, of convolutions between the change of exciting field and the function
$\chi^{abs}(t)$ or $\chi^{em}(t)$ respectively.
Now, as the Fourier transform of a convolution
is the product of the Fourier transforms (which we denote by a hat),
the relations between polarization and
exciting field can be written in the familiar form
\begin{equation}\label{eq:rispostalinearetrasf}
\begin{split}
\tilde{ {\boldsymbol{\mathcal{P}}} }(\vett x,\omega)= & \tilde{\chi}^{abs}(\omega) {\delta\tilde {\boldsymbol{\mathcal{E}}^c} }
(\vett x,\omega)\\
\tilde{ {\boldsymbol{\mathcal{P}}} }(\vett x,\omega)= & \tilde{\chi}^{em}(\omega) {\delta\tilde {\boldsymbol{\mathcal{E}}^c} }
(\vett x,\omega)
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\hat{\chi}^{abs}(\omega) = & - \int_{-\infty}^0\mathrm{d} t\ \tilde\chi(t)
e^{i\omega t} \\
\hat{\chi}^{em}(\omega) = & \int_0^{+\infty} \mathrm{d} t \ \tilde\chi(t)
e^{i\omega t}\ .
\end{split}
\end{equation}
As $\tilde \chi$ is odd (see below), by the change of variable $t\to -t$
in the second integral one gets that
$\hat{\chi}^{em}$ is the complex conjugate of $\hat{\chi}^{abs}$. So the
emission and the absorption spectra coincide.
To show that $\tilde\chi(t)$ is an odd function,
we notice that, from the definition, one has
$$
\tilde\chi(-t)=\int \mathrm{d} z \frac 1\Delta V \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }
\sum_{j,j'} e_j e_{j'} \vett{q}_{j',k'} (-t)
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} } \ ,
$$
so that, performing into the integral the substitution $\vett{p}_{j,k} \to -\vett{p}_{j,k} $,
one finds
$$
\tilde\chi(-t)= - \int \mathrm{d} z \frac 1\Delta V
\sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'} e_j e_{j'} \vett{q}_{j',k'} (t)
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} } = - \tilde\chi(t) \
$$
(indeed, as $\rho_0$ is even, its derivatives are odd, whereas, by changing
sign to the momenta, $\vett{q}_{j',k'} (-t)$ goes into $\vett{q}_{j',k'} (t)$).
\subsection*{Analyticity properties. The Kramers--Kronig relations }
It is well known that, as the function $\chi^{abs}(t)$ vanishes for $t<0$,
then its Fourier transform enjoys two relevant properties:
\begin{itemize}
\item It is analytic in the half plane $\IM \omega > 0$;
\item The Kramers--Kronig relations hold
\begin{align}\label{eq:KK}
\RE \hat \chi^{abs}((\omega) &= \frac 1\pi \int_{\mathbb{R}} \mathrm{d} \Omega \, \frac{\IM
\hat \chi^{abs}(\Omega)}{\Omega-\omega} \nonumber \\
\IM \hat \chi^{abs}((\omega) &= -\frac 1\pi \int_{\mathbb{R}} \mathrm{d} \Omega \, \frac{\RE
\hat \chi^{abs}(\Omega)}{\Omega-\omega} \ .
\end{align}
\end{itemize}
From a conceptual point of view the Kramers--Kronig
relations are often interpreted as expressing the causality
principle, the latter being meant in the sense that the affect (here,
polarization) cannot precede the cause (the exciting field).
On the other hand, analogous relations obviously hold also
for the function $\hat \chi^{em}((\omega)$, which clearly is not causal
in that sense, as $\chi^{em}(t)$ vanishes after the field is
applied.
A second remark concerns the poles of the two
susceptibilities. Since the original work of
Kramers, the emission was attributed to the presence of the radiation
reaction force (proportional to the time derivative of acceleration)
in the equations of motion.
In such a way, however, in the expression for the susceptibility,
calculated
by considering a single damped and forced oscillator, there
appeared a pole in the wrong half--plane, and Kramers himself had to
patch the expression in some suitable way. Instead,
with the full electrodynamic treatment considered here,
in virtue of the Wheeler--Feynman cancellation
the radiation reaction forces entering
the original equations of motion eventually disappears,
and the expressions of the
susceptibilities have poles in the correct half--plane.
\subsection*{The $f$--sum rule}
We finally come to the $f$--sum rule. The reason of the name will be
recalled in the next section.
For the sake of concreteness we here concentrate on the case of the
absorption susceptibility, because the formulas for the case of
emission are simply obtained by passing to the conjugate complex. In
order to have simpler notations, we also omit the superscript
${abs}$, i.e., we let $\hat{\chi}^{abs}\equiv \hat \chi$.
The $f$--sum rule states that
\begin{equation}\label{eq:fsumrule}
\int_{\mathbb{R}} \omega\IM\hat\chi(\omega)\mathrm{d}\omega= \frac \pi\Delta V
\sum_{\vett{x}_k^0\in\Delta V } \sum_{j} \frac {e_j^2}{m_j} \ ,
\end{equation}
so that it essentially relates the total absorption
to the electron charge density. Indeed one should
take into account that for nuclei the ratio
$e_j^2/m_j$ is negligible with respect to that of the
electrons, so that the sum at the right hand side can be restricted
to the electrons present in the considered volume. Thus, denoting by
$e$ and $m$ the charge and the mass of the electron, the r.h.s.
just reduces to $\pi e^2/m$ times the electron density (number of
electrons per unit volume).
The next part of this section is devoted to a proof of the $f$--sum
rule (\ref{eq:fsumrule}). We start noting that for a smooth
functions $f(t)$ one has
$$
\int_{\mathbb{R}} -i\omega \hat f(\omega)\mathrm{d}\omega= 2\pi\dot f(0) \ .
$$
Indeed, on the one hand the Fourier transform of $\dot f(t)$ is given by
$-i\omega \hat f(\omega)$, as one immediately checks by an integration by
parts. On the other hand the inversion theorem for the Fourier
transform gives
$$
\int_{\mathbb{R}} -i\omega \hat f(\omega)e^{-i\omega t}\mathrm{d}\omega= 2\pi \dot
f(t) \ .
$$
So the thesis should follow by simply taking $t=0$. However, in our case
$\dot \chi(t)$ presents a discontinuity of first type at
$t=0$, as it vanishes for $t>0$,
while being equal to $\dot{\tilde\chi}(t)$ for $t<0$. Now, the inversion
theorem tells us that at a discontinuity points
the integral equals the semi sum of the right and the left limits, so
that eventually one has
$$
\int_{\mathbb{R}} -i\omega\hat\chi(\omega)\mathrm{d}\omega= \pi
\dot{\tilde\chi}(0) \ .
$$
However, as is easily checked,\footnote{Indeed, one has
$$
\RE \hat \chi(\omega) = \int_{-\infty}^0 \tilde\chi(t)\cos(\omega t)\mathrm{d}
t \
$$
so that, changing $\omega$ into $-\omega$, the value of the integral
does not change.} $\RE \hat\chi(\omega)$ is an even
function of $\omega$,
so that one has
$$
\int_{\mathbb{R}} -i\omega\hat\chi(\omega)\mathrm{d}\omega= \int_{\mathbb{R}} \omega
\IM \hat\chi(\omega)\mathrm{d}\omega = \pi \dot{\tilde\chi}(0) \ .
$$
Now it turns out that $\dot{\tilde\chi}(0)$ can be evaluated
exactly and, as will be seen in a moment, one has
$$
\dot{\tilde\chi}(0)= \frac 1\Delta V
\sum_{\vett{x}_k^0\in\Delta V } \sum_{j} \frac {e_j^2}{m_j} \ ,
$$
which indeed proves the $f$--sum rule (\ref{eq:fsumrule}).
In order to show the latter relation, we differentiate the expression
(\ref{eq:suscettivita}) for $\tilde\chi(t)$.
Exchanging derivative and integral one gets
\begin{equation*}
\begin{split}
\dot{\tilde\chi}(0) &= - \int \mathrm{d}
z \frac 1\Delta V \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j} e_j e_{j'} \dqjkp(0)
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} } =\\
&= - \int \mathrm{d}
z \frac 1\Delta V \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j} \frac {e_j
e_{j'}}{m_{j'}} \vett{p}_{j',k'} (t)
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} } \ ,
\end{split}
\end{equation*}
where in the second line use was made of $\dqjkp(0)=\vett{p}_{j',k'} /m_{j'}$.
Now there just remains to integrate by parts. The boundary term
vanishes (due to the vanishing of the probability for a particle to have
an infinite momentum), so that
\begin{equation*}
\begin{split}
\dot{\tilde\chi}(0) &= \int \mathrm{d}
z \frac 1\Delta V \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'=0} \frac {e_j e_{j'}}{m_{j'}}
\frac {\partial\vett{p}_{j',k'} }{\partial\vett{p}_{j,k} } \rho_0 =\\
&= \int \mathrm{d}
z \frac 1\Delta V \sum_{\vett{x}_k^0\in\Delta V }\sum_{j} \frac {e_j^2}{m_{j}}
\rho_0 = \frac 1\Delta V \sum_{\vett{x}_k^0\in\Delta V }\sum_{j} \frac {e_j^2}{m_{j}} \ ,
\end{split}
\end{equation*}
inasmuch as $ \frac
{\partial\vett{p}_{j',k'} }{\partial\vett{p}_{j,k} }=\delta_{k,k'}\delta_{j,j'}$, whereas
the density $\rho_0$ is assumed to be normalized to $1$.
\section{Response functions and susceptibilities
in terms of correlation functions}\label{5}
After the detour on the analyticity properties of the dielectric
response functions and susceptibilities, which were based on the
general expression (\ref{eq:suscettivita}),
we show here how more transparent expressions are obtained if
a further property of a quite general character is introduced for
the equilibrium density $\rho_0$. The point is that formula
(\ref{eq:suscettivita}) involves sums of integrals of the type
\begin{equation}\label{integrale}
\mathcal{I}_{k,j,k',j'} = \int \mathrm{d} z \, \vett{q}_{j',k'} (t-s) \frac
{\partial\rho_0}{\partial\vett{p}_{j,k} } \ ,
\end{equation}
the computation of which requires to have
available a definite expression for the
derivative of $\rho_0$ with respect to $\vett{p}_{j,k} $.
Now, if we were allowed to take for
$ \rho_0$ the Gibbs form, the above quantity would be proportional to
$\vett{p}_{j,k} \, \rho_0$. On the other hand, essentially the same result is
guaranteed under much milder conditions, essentially under conditions
which allow for a large deviation principle to
hold with respect to the momenta only, irrespective of the
positions (which, through the attractive
Coulomb potential, introduce divergences in the classical form of
Gibbs' measure).
Indeed this allows one to get
\begin{equation}\label{ld}
\frac {\partial\rho_0}{\partial\vett{p}_{j,k} }= -\frac {1}{m_{j'}\, \sigma^2_p}\,
\vett{p}_{j,k} \, \, \rho_0\ ,
\end{equation}
where the constant $\sigma_p^2$
is nothing but the mean square deviation
of momentum, which would just reduce to temperature if the density
were the Gibbs one. For the large deviation argument one can see the
classical book of Khinchin \cite{khin}.
So we have
\begin{equation*}\begin{split}
\mathcal{I}_{k,j,k',j'}& = \frac {-1}{m_{j'}\sigma^2_P}\int \mathrm{d} z
\vett{q}_{j',k'} (t-s) \vett{p}_{j,k} \rho_0(z)\\
& = \frac {-1}{m_{j'}\sigma^2_p} \langle \vett{q}_{j',k'} (t-s) \vett{p}_{j,k} (0) \rangle \ ,
\end{split}
\end{equation*}
namely, the integrals (\ref{integrale}) are just equilibrium
time--corre\-la\-tions
between position and momentum of each charge.
This fact, by the way, makes reasonable a property that was assumed in
the last part of section \ref{3}, when passing from
(\ref{eq:polarizza}) to (\ref{eq:polarizza2}). Namely, the property that the
integrals (\ref{integrale}) should present a fast decay
with respect to spatial separation of the charges, i.e., that one should have
$$
\mathcal{I}_{k,j,k',j'} =0
$$
if the molecules $\vett{x}_k^0$ e $\vett{x}_{k'}^0 $ belong to different
volume elements.
In conclusion, the
expression (\ref{eq:suscettivita}) for the dielectric response
function can be rewritten in the form
\begin{equation}\label{eq:suscettivita2}
\tilde\chi(t) = \frac 1{\sigma_p^2}
\sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'} \frac {e_j e_{j'}}{m_j} \langle
\vett{q}_{j',k'} (t) \vett{p}_{j,k} (0)\rangle \ ,
\end{equation}
which involves equilibrium time--correlations of momenta and positions of
the charges.
Now there remains the problem that we have to compute
phase averages with respect to the equilibrium probability density
$\rho_0$, the
form of which is still essentially undetermined. A great step
forward is accomplished by making use of a general principle
of statistical mechanics according to which, under extremely mild
conditions, the phase space equilibrium averages can be computed as
corresponding time averages (see for example \cite{khin}, page
63).
So we estimate the required phase space integral
as time averages, i.e. as
\begin{equation}\label{eq:corrpq_bis}
\begin{split}
\langle &\vett{q}_{j',k'} (t) \vett{p}_{j,k} (0)\rangle=\\
= &\lim_{T\to+\infty}\frac 1{2T}\int_{-T}^T
\vett{q}_{j',k'} (t+s) \cdot \vett{p}_{j,k} (s) \mathrm{d} s \ .
\end{split}
\end{equation}
\section{Line spectrum and the ``virtual orchestra''}\label{6}
Here we show how it can at all happen
that a conservative Hamiltonian system (to which our original
electrodynamic system has been reduced) presents a line spectrum.
This depends of the qualitative properties of the dynamical orbits (or
motions) of the system, because it turns out that a discrete spectrum occurs
if the motion of the representative point in phase space is, informally
speaking, ``non chaotic''. Indeed in
dynamical systems theory the property of presenting a
continuous spectrum is sometimes even assumed to be the
characteristic property for an orbit to be
chaotic.
More precisely, one certainly has a pure line spectrum if
the motion is assumed to be ``almost periodic'' in the sense
introduced in the year 1924 by Harald Bohr, the brother of Niels Bohr.
\footnote{For an introduction to almost periodic functions see for example
\cite{nem}, Part II, Chapter 5, where in particular the relations
between almost periodicity and Liapunov stability of an orbit
are discussed.}
\subsection*{Pure line spectrum for almost periodic motions}
Almost periodicity can be defined in several equivalent
ways.
However, the following characteristic property (which thus can be taken as
a definition), is more significant for our purposes:
if an orbit, say the motion $\vett{q}_{j,k} (t)$ of a particle, is almost
periodic, then it can be represented by a generalized Fourier expansion
\begin{equation}\label{eq:sviluppoq}
\vett{q}_{j,k} (t) = \sum_n \big[\cjkn \cos(\omega_n t) + \djkn \sin(\omega_n t)\big]
\end{equation}
where the sequence $\{\omega_n\}$ of \emph{positive} frequencies is
determined in the following way. Having defined the
functions\footnote{For almost periodic functions these limits are
proven to exist. See for example the classical text \cite{besi}.}
$\vett c_{j,k}(\omega)$ and $\vett d_{j,k}(\omega)$ by
$$
\vett c_{j,k}(\omega)=\lim_{t\to+\infty} \frac 1{2t}\int_{-t}^t
\vett{q}_{j,k} (s) \cos(\omega s)\mathrm{d} s \ ,
$$
$$
\vett d_{j,k}(\omega)=\lim_{t\to+\infty} \frac 1{2t}\int_{-t}^t
\vett{q}_{j,k} (s) \sin(\omega s)\mathrm{d} s \ ,
$$
then these functions turn out to vanish for all frequencies
but for a discrete set of frequencies $\{\omega_n\}$. This
determines the frequencies.
Then, the coefficients of the expansion simply are the values of the
expansion
simply are the values of the the above functions at $\omega_n$,
i.e., one has
$$
\cjkn = \vett c_{j,k}(\omega_n) \ , \quad
\djkn = \vett d_{j,k}(\omega_n) \ .
$$
Corresponding to the expansion (\ref{eq:sviluppoq}) for the position
as a function of time, one also has an analogous expansion
for the momenta, namely,
\begin{equation}\label{eq:svilippoq}
\vett{p}_{j,k} (t) = m_j \sum_n -\omega_n \cjkn \sin(\omega_n t) + \omega_n \djkn
\cos(\omega_n t) \ ,
\end{equation}
which is obviously obtained by differentiating with respect to time
the expansion for $\vett{q}_{j,k} (t)$.
One thus obtains
\begin{equation}\label{eq:corrpq}
\begin{split}
& \langle \vett{q}_{j',k'} (t) \vett{p}_{j,k} (0)\rangle =\\
&= \sum_n \omega_n\Big[
\frac {\cjkn\cdot\cjknp +\djkn\cdot\djknp }2 \sin \omega_n t \\
&+
\frac {\cjkn\cdot\djknp - \djkn\cdot\cjknp}2\cos \omega_n t\Big]
\ .
\end{split}
\end{equation}
This relation is obtained by evaluating the integrals through the
familiar prosthaphaeresis formulas, recalling that the time
average of any non constant trigonometric function vanishes.
The result is the following one.
Defining
\begin{equation*}
\begin{split}
I_{sc}& \stackrel {\mathrm{def}}{=} \lim_{T\to+\infty}\frac 1{2T}\int_{-T}^T
\sin\omega s\cos\omega'(t+s)\mathrm{d} s\\
I_{ss}& \stackrel {\mathrm{def}}{=} \lim_{T\to+\infty}\frac 1{2T}\int_{-T}^T
\sin\omega s\sin\omega'(t+s)\mathrm{d} s \\
I_{cc}& \stackrel {\mathrm{def}}{=} \lim_{T\to+\infty}\frac 1{2T}\int_{-T}^T
\cos\omega s\cos\omega'(t+s)\mathrm{d} s \\
I_{cs}& \stackrel {\mathrm{def}}{=} \lim_{T\to+\infty}\frac 1{2T}\int_{-T}^T
\cos\omega s\sin\omega'(t+s)\mathrm{d} s\ ,
\end{split}
\end{equation*}
one finds that all the $I$'s vanish for $\omega\ne\omega'$, whereas for
$\omega=\omega'$
one has
$$
I_{sc}=I_{cs}= -\frac 12 \sin\omega t\ ,
\quad I_{ss}=I_{cc}= -\frac 12 \cos\omega t\quad .
$$
\subsection*{Form of susceptibility for
almost periodic motions}
Now, substitute into formula (\ref{eq:suscettivita2})
the expression (\ref{eq:corrpq}) just found for the
correlations. Remarking that,
due to the antisymmetry with respect to the interchange
$k,j \leftrightarrow k',j'$ of the terms occurring in the sum.
one has
$$
\sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'} \frac {e_j e_{j'}}{m_j}
\frac {\cjkn\cdot\djknp - \djkn\cdot\cjknp }2 = 0 \ ,
$$
one obtains
\begin{equation*}
\begin{split}
&\tilde\chi(t) = \frac 1{\sigma_p^2} \, \sum_n \omega_n \sin \omega_n
t\, \cdot\\
&\cdot\,
\sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j'} \frac {e_j e_{j'}}{m_j}\,
\frac {\cjkn\cdot\cjknp +\djkn\cdot\djknp }2 \ ,
\end{split}
\end{equation*}
In order to find the susceptibility there just remains to compute the
Fourier transform of $\tilde\chi(t)$. A not difficult computation
shows that one has
$$
\int_{-\infty}^0 \sin \omega_n t \,e^{i\omega t}\mathrm{d} t = \frac
{-\omega_n}{\omega_n^2 -\omega^2} +i\pi\Big(
\delta(\omega-\omega_n) +\delta(\omega+\omega_n)\Big) \ .
$$
Thus, defining
\begin{equation}\label{eq:forzaosc}
f_n \stackrel {\mathrm{def}}{=} \omega_n^2\left[ \sum_{\vett{x}_k^0,\vett{x}_{k'}^0 \in\Delta V }\sum_{j,j}
\frac {e_j e_{j'}}{m_j}
\frac {\cjkn\cdot\cjknp +\djkn\cdot\djknp }2 \right]\ ,
\end{equation}
for the real and the imaginary parts of susceptibility one finds the
expressions
\begin{equation}\label{orchestra}
\begin{split}
\RE \chi(\omega) & = \sum \frac {f_n}{\omega_n^2 - \omega^2} \\
\IM \chi(\omega) & = \pi \sum \frac {f_n}{2\omega_n} \Big(
\delta(\omega-\omega_n)
+\delta(\omega+\omega_n)\Big) \ .
\end{split}
\end{equation}
\subsection*{The ``virtual orchestra'' of Bohr, Kramers and Slater}
Due to the delta functions appearing in
the imaginary part of susceptibility,
formula (\ref{orchestra}) shows that the spectrum of a macroscopic
dielectric body performing almost periodic motions
presents infinitely tight absorption
lines, in correspondence of the frequencies
$\omega_n$. This is the way in which, in the spectrum of a
macroscopic dielectric body, ``lines'' show up without necessarily
having to make reference to energy levels of the single molecule or atom.
This result is exactly the property of a spectrum which, before the
advent of quantum mechanics, (starting from Lorentz \cite{lorentz} and
Drude up to Kronig \cite{kronig} and even Born and Wolf
\cite{bw}), was interpreted in microscopic terms by thinking
that each line should be attributed to the motion of a material
harmonic ``resonator'', of exactly that frequency.
Analogously the molecules were thought of as
constituted of charges with mutual elastic
bonds. So there would exist corresponding normal modes,
which could be
equivalently described as harmonic oscillators with characteristic
frequencies $\omega_n$ (which were introduced from outside, in
correspondence with the
observed ones).
However, as the lines are infinite in number, one was meeting with
the absurd situation that any atom or molecule had to be composed of
an infinite number of oscillating charges
For this reason such oscillators were denoted as ``virtual'' i.e., as
somehow non physical (see \cite{bks}), and each of them was
weighted with a suitable weight (usually called ``force'') $f_n$.
In the year 1925 the ``$f$--sum rule'' was empirically discovered,
according to which the ``forces'' of the virtual oscillators were not
arbitrary, but had to satisfy the rule
\begin{equation}\label{eq:fsumrulequan}
\sum_{n} f_n = \frac 1\Delta V \sum_{\vett{x}_k^0\in\Delta V } \sum_{j} \frac
{e_j^2}{m_j} \ .
\end{equation}
Namely, the sum of the ``forces'' of the oscillators just equals
the number of electrons per atom or per molecule,
times the factor $e^2/m_e$
(indeed, as already explained, the contribution of the nuclei is
negligible).
One of the big triumphs of quantum mechanics was to ``explain'' the
$f$--sum rule in terms of the quantum commutation rules.
On the other hand, such a rule holds in the classical case too. Indeed an
explicit computation gives
$$
\int_{\mathbb{R}} \omega\IM \chi(\omega)\mathrm{d} \omega = \pi \sum f_n \ ,
$$
which, using the general formula (\ref{eq:fsumrule}), actually gives
the $f$--sum rule (\ref{eq:fsumrulequan}).
\section{Broadening and chaoticity: the case of ionic crystals}\label{7}
So, a pure line spectrum occurs for stable (almost periodic) motions, whereas
a broadening of the lines and even a continuous spectrum
are expected to occur when chaoticity of the motions sets in.
This connection between optical properties of the system
and qualitative properties (order or chaos, or their coexistence) of
the corresponding orbits can be
illustrated in a particularly clear way in the case of ionic crystals.
If one is interested in the infrared spectrum,
in the expression (\ref{eq:suscettivita2}) for the dielectric response
function it is sufficient to limit oneself to
the motions of the ions only.
In such a case it is convenient to choose
as a reference point $\vett{x}_k^0$ (with respect to which the displacements
$\vett{q}_{j,k} $ are computed), an arbitrary fixed point inside each cell of
the lattice. In such a way the index $k$ is now labeling also the cells.
Following \cite{alessio} one can pass to the normal modes
of the lattice, which we here denote by $A_{ {\boldsymbol{\xi}} ,l}(t)$ and are defined by
$$
\vett{q}_{j,k} (t) = \sum_l \int_{\mathcal{B}} \vett u_l(j, {\boldsymbol{\xi}} )
A_{ {\boldsymbol{\xi}} ,l}(t) e^{i {\boldsymbol{\xi}} \cdot(\vett{x}_k^0+\vett \tau_j) } \mathrm{d} {\boldsymbol{\xi}} \ .
$$
Here, the integration is performed over the Brillouin zone $\mathcal{B}$,
the vectors $\vett u_l(j, {\boldsymbol{\xi}} )$ are the eigenvectors of the dynamical
matrix of the crystal, while the vector $\tau_j$ specifies the
equilibrium position of the $j$--th atom inside the cell $k$.
The index $l$ is now a label for the
different branches of the dispersion relation.\footnote{ We recall that, while
in the purely mechanical case the number of branches is $3N$
($N$ being the number of ions in the fundamental cell), instead,
when the interaction with the field of the far ions is taken into
account, the number of branches can vary, and polaritonic
branches can appear.}
So, one gets the relation
$$
\sum_{\vett{x}_k^0\in \Delta V} \vett{q}_{j,k} (t) \simeq (2\pi)^3
\sum_l \vett u_l(j,0) A_{0,l}(t) \ ,
$$
because, in summing over a volume element, one has
$$
\sum_{\vett{x}_k^0\in \Delta V} e^{i {\boldsymbol{\xi}} \cdot\vett{x}_k^0} \simeq
(2\pi)^3 \delta( {\boldsymbol{\xi}} ) \ .
$$
Thus, in the case of a ionic crystal the dielectric response function for
the ions can be written as
\begin{equation*}
\begin{split}
\tilde\chi(t) = & \frac 1{\sigma_p^2}
\sum_{l,l'} \Big( \sum_{j,j'} e_j e_{j'} \vett u_l(j,0) \cdot
\vett u_l(j',0) \Big) \\
&\langle A_{0,l}(t)(t) \dot A_{0,l'}(0) \rangle \ ,
\end{split}
\end{equation*}
so that the relevant quantities now are the time correlations of the modes
$A_{0,l}(t)$.
If the harmonic approximation, each mode
performs a periodic motion with frequency $\omega_l$, so that one has
$$
\langle A_{0,l}(t)(t) \dot A_{0,l'}(0) \rangle = C_l \delta_{ll'}
\sin(\omega_l t)\ ,
$$
being $\delta_{ll'}$ the Kronecker's delta,
and one ends up with a formula of the type (\ref{orchestra}), now however
with only a finite number of terms, each corresponding to a
branch of the dispersion relation (omitting the
``acoustic'' branches , for which it is $A_{0,l}=0$).
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\textwidth]{fig_corr}
\end{center}
\caption{\label{fig:due} The response function $\chi$ versus time. Solid
line refers to the system at low temperature, while broken line
refers to the system at high temperature.}
\end{figure}
On the other hand, if the nonlinear terms are taken into account
the motion is no more integrable, and the previous analysis
has to be changed. In the case of a ``small''
nonlinearity, the behavior of the correlations over some
(large) time--scale does not change with respect to the unperturbed
(i.e., linear) case, whereas over a larger time scale
the correlations are expected to decay to zero, so that one should have
$$
\langle A_{0,l}(t)(t) \dot A_{0,l}(0) \rangle = C_l e^{-\sigma_l t}
\sin(\omega_l t)\ ,
$$
In conclusion, passing to the Fourier transform, one can presume that
in the case of a small nonlinearity one should get
$$
\int_0^{+\infty} e^{i\omega t} \langle A_{0,l}(t)(t) \dot A_{0,l}(0)
\rangle = \frac {f_l}{(\omega^2 - \omega_l^2+\sigma_l^2) + 2i\sigma_l\omega} \ ,
$$
i.e., the classical expression of Lorentz and Drude
\cite{lorentz}\cite{drude}, that such authors interpreted in terms of
motions of material damped ``resonators''.
Thus the line broadening corresponds to a decay of the time correlations
which is induced by the nonlinearity and the
presumably associated chaoticity (or rather partial chaoticity)
of the motions.
Here no damping is active, neither the linear one which was heuristically
introduced by Lorentz and Drude, nor that of the radiation reaction,
which was always taken into consideration by Van Vleck, Planck and many
others. Indeed the radiation reaction, although being actually
present in the original full electrodynamic model, turns out to
be eliminated by the electrodynamic action of the far charges,
through the Wheeler--Feynman mechanism.
So much for the case of a small nonlinearity, i.e., for the case of
what may called the ``perturbation regime'' (with respect to the
linear one). Instead, in the case of a large nonlinearity
the motion is expected to be completely chaotic, displaying time
correlations completely different from those of the linear case.
In particular the spectrum should be now a continuous one, with no
peaks anymore.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{fig_spettro}
\end{center}
\caption{\label{fig:tre} Plot of $|\chi(\omega)|$ versus
$\omega$ for two different temperatures. Circles refer to the system
at a high temperature (no peak), while the triangles, which
exhibit a peak for $\omega\simeq\omega_0$, refer to the system
at a low temperature . In the inset, which concerns the system
at low temperature, the plot of $|\chi(\omega)|$
is reported for $\omega$ near $\omega_0$,
together with the best--fit Lorentzian curve (solid
line). Here, $\omega_0$ is the frequency of the optical branch.}
\end{figure}
On the other hand, when in statistical mechanics one makes reference to
the qualitative properties of the motions with respect to order
(stability)
or chaoticity type, it is usually presumed
that in the thermodynamic limit the motions should always be chaotic.
This has the consequence that, in our case, which is concerned with
macroscopic dielectric systems dealt with in a classical frame,
one would always meet with a
continuous spectrum. Now, in the domain of the theory of
classical dynamical systems, particularly in connection with
the so called Fermi--Pasta--Ulam problem, a long
debate is going on about this point, and the results of
numerical computations appeared to be not yet conclusive.
However, rather recently it was analytically proven \cite{andrea}
that in the perturbation regimes significant stability properties do
persist in the thermodynamic limit, and indeed in a form
suited for applications to statistical
mechanics. In particular, in the works \cite{fpu} and \cite{maiocchi}
it was proven that in the FPU and in related models the normal
mode energies remain correlated
for long times also in the thermodynamic limit (see also the numerical
work \cite{cggp}, or the work \cite{plasmi2} concerning plasmas).
Thus one can conclude that the conjecture that
macroscopic systems should perform chaotic motions
is, at least, not always appropriate,
and should be checked in any particular case.
Just in order to give an example which should exhibit
in a qualitative way the features described above, we report here the
results of a numerical computation performed on the classical
one--dimensional alternating
masses model (with 1024 paticles), introduced already in the year
1912 by Born and von K\'arman.
Through a numerical simulation of the dynamics we computed
the response function $\chi(t)$, defined
by (\ref{eq:suscettivita2}) with the sum extended over all
particles of the crystal, and then the corresponding spectrum.
We considered two cases relative to a low
temperature and to a larger one. The response functions for the two
temperatures are reported in
Fig.~\ref{fig:due}, whereas
the corresponding spectra (computed as the
discrete Fourier transforms) are reported in Fig.~\ref{fig:tre}.
In the case of low temperature the response function presents a well
distinct profile, apparently not very dissimilar from a
periodic one. However a decay occurs at much longer times, as
witnessed by the broadened form of the spectrum (shown in the inset of
Fig.~\ref{fig:tre}). Further results not
reported here show that with increasing temperature the broadening,
and a shift too, become larger and larger. Finally, at some high
temperature, the results reported in the figures show that
the response function presents a decay at a short time, and the
corresponding spectrum is essentially a continuum. For an analogous
phenomenon occurring in a model of interest for plasma
confinement, in which a transition from an ordered to a chaotic
motion is witnessed by the form of the spectrum,
see \cite{plasmi2}.
We leave for a future work the numerical study of the
spectrum for a realistic three--dimensional model of a ionic crystal
involving the microscopic electrodynamic forces,
already considered in \cite{alessio}
in connection with the dispersion curves.
\section{Final comments}\label{8}
So we have complemented the result obtained in \cite{alessio}, by
showing how electric susceptibility can be consistently pro\-ven to
exist for a
dielectric macroscopic body, in a classical microscopic theory in which
the full electrodynamic interactions
among the charges are taken into account. Preliminarily we had to make
essential use
of two global properties of the electrodynamic interactions, i.e.,
the Wheeler--Feynman identity and the Ewald--Oseen resummation
properties. The former was proved here for a general system in the
thermodynamic limit, whereas the latter were proven in \cite{alessio}
for crystals, their proof for a general system being still lacking.
Thus our result is at the moment proven only for crystals, although we
are firmly convinced that it can be extended to cover the case of a
generic dielectric body.
On the basis of such global electrodynamic properties, the dynamical system
can be dealt with as if it were a Hamiltonian one, and in particular the
radiation reaction forces are completely eliminated, so that
absorption and emission appear as completely
symmetrical phenomena of a time--reversal invariant system.
Susceptibility turns out to be
expressed in terms of the time correlation functions of
positions and velocities of the charges, calculated for
motions of the system at equilibrium, i.e.,
in the absence of an external field. Notice however that the system
still contains a trace of the electrodynamic interactions, because the
equations of motion of the charges, that have to be solved in order that
the time correlation functions may be computed, still contain
the force of the ``exciting field'', namely, the field originated by
the far charges, that propagates
in the body as an external field, having however
the correct refractive index.
Having reduced the original electrodynamic system to a Hamiltonian one,
susceptibility was proven to exist through methods of
Green--Kubo type. However, this required to overcome
the difficulties of working in the absence of a Gibbs
measure, which does not exist for systems with attractive Coulomb
interactions,
For what concerns the spectrum, which is the same
for absorption and for emission, we have illustrated how it
reflects the stability properties of the unperturbed
equilibrium motions of the system. For stable (almost periodic) motions, as
occurs with a crystal in the linear approximation, one
has a purely line spectrum. So, the susceptibility
presents the standard form that,
since the first work of Lorentz of the year 1872, was
explained by thinking of the system as if it were composed of
single linear material oscillators with proper frequencies equal to
the observed ones (see the booklet
\cite{pauli} of Pauli).
When chaoticity sets in, as occurs in a
crystal in the presence of nonlinearities, one might conjecture that the
motions be completely chaotic, so that
the lines completely disappear, and a continuous spectrum occurs.
We have however pointed out that the most recent analytical result
appear to support the conjecture that, at least in the case of
crystals, partially ordered motions
persist in the thermodynamic limit (i.e., for a macroscopic system).
Thus the time correlations in
general should decay only after a sufficiently long time,
with the consequence that the lines are
in general broadened. In such a case the spectrum has the form that would
occur if the system were composed of
single linear material oscillators with the observed frequencies,
having in addition suitable linear
dissipative forces. However, no physical dissipative force is
actually present in our system, because, in virtue of the
Wheeler--Feynman identity, the radiation reaction forces
are canceled by the
electrodynamic forces due to the far charges. So, the decay
of correlations occurs in
the familiar dynamical way which characterizes autonomous Hamiltonian
systems that are (at least partially) chaotic,
and has nothing to do with the radiation
reaction force, to which for example Planck, Van Vleck and many others
were thinking. Correspondingly, the poles of susceptibility
in the complex frequency plane quite naturally do lie
in the correct half plane.
In any case, while in the
theory of dynamical systems the presence of a continuous or partly
continuous spectrum is sometimes used as a tool to
qualify the ordered or chaotic character of motions, here the
situation is reversed, and it is the
spectrum itself, in its original physical optical connotation, that is a pure
line spectrum in the case of ordered motions, while presenting broadened lines
or a fully continuous aspect in the case of partly or fully chaotic motions.
\section{Appendix. {Proof of the Wheeler--Feynman identity}}\label{9}
\subsection*{Proof of the identity}
The Wheeler--Feynman identity deals with the classical problem of
the solutions of the inhomogeneous wave equation
\begin{equation*}
\dale A^{\nu} =j^{\nu}(t,\vett x) \ ,
\end{equation*}
for the four--potential $A^{\nu}$, with a given four--current
$j^{\nu}(t,\vett x)$, and states that, possibly under suitable
conditions,
the advanced potential coincides with the retarded one, or more
precisely, in terms of their difference which is a regular function,
that one has
$$
A_{ret}^\nu - A_{adv}^\nu=0\ .
$$
Clearly this in not true for an arbitrary current, and the authors, on
the basis of four arguments, advanced the conjecture that the identity
should hold if the problem is considered as a global one involving, as
they said, all charges ``of the universe''. A much more innocuous
setting in which the problem can be framed, is the standard one of
statistical mechanics, where reference is made to the ``thermodynamic
limit''. So we consider the microscopic current inside a domain of
volume $V$ (i.e., the ``truncated'' function $j_V$ which coincides
with $j$ inside the domain and vanishes outside), and take the limit
in which both the volume and the number of elementary charges
constituting the current tend to infinity, in such a way that the
charge density (number of charges per unit volume) remains constant.
Such a framing of the problem is immediately reflected in a deep
mathematical property of the current, because for the current density
one clearly has to give up any property of decrease at infinity, and
one should assume for example only the property $j^\nu\in
L^\infty(\mathbb{R}^3)$, i.e., that the density $j^\nu(t,\vett x)$ be only
locally integrable.
As a possible substitute for the global integrability condition there
is one that quite naturally comes to one's mind for its physical
significance. Moreover, it is somehow analogous to what is sometimes called the
locality condition of quantum field theory, although it rather appears
to express a kind of causality condition. Precisely, we start up
defining the autocorrelation of the current density $j^\nu$ by
\begin{equation}\label{eq:defcorr}
\mathcal{C}_{j^\nu}(s,t,\vett x) \stackrel {\mathrm{def}}{=} \lim_{V\to\mathbb{R}^3} \frac 1{V}
\int_V j^\nu(s,\vett y) j^\nu(s+t,\vett y- \vett x) \mathrm{d}\vett y \ ,
\end{equation}
where the symbol $V$ denotes both the space region of integration and
its (Lebesgue) measure. It is implicitly assumed that the average of
$j^\nu(t,\vett x)$ over the whole space--time vanishes.
Now our global hypothesis reads as follows.
\begin{definition}[Causality Condition]\label{hyp:1}
A source $j(t,\vett x)$ satisfies the Causality Condition, iff 1)
$j\in L^\infty(\mathbb{R}^3)$, 2) the correlation $\mathcal{C}_{j}(s,t,\vett x) $
exists for all $s$, $t$, $\vett x$, and 3) for all $s$ one has
\begin{equation}\label{eq:2}
\mathcal{C}_{j}(s,t,\vett x) = 0 \quad \mbox{for} \qquad c^2t^2 - \vett
x\cdot \vett x \le 0 \ .
\end{equation}
\end{definition}
In other terms we are assuming that there exists no correlation
between space--separated points of space--time.
This requirement is natural from the
physical point of view, because one should assume that the interactions
cannot propagate faster than light, so that it seems
natural to assume that space separated events be
independent.\footnote{We do not discuss here whether this is active or
passive locality in the sense of Nelson \cite{nelson}.}
We now show that the above ``causality condition'' is sufficient to
guarantee the validity of the identity. Indeed the
following Theorem~\ref{teo:main} holds:
\begin{theorem}\label{teo:main}
Consider the wave equation
\begin{equation}\label{eq:onde}
\dale A =j(t,\vett x) \ ,
\end{equation}
having as source a current $j(t,\vett x)$ satisfying the Causality
Condition~\ref{hyp:1}. Let $A_{ret}$ and $A_{adv}$ be the retarded and the
advanced solutions respectively. Then for all $t$ one has
\begin{equation}\label{eq:WF}
\lim_{V\to\infty} \frac 1V \int_V
\Big(A_{ret}(t,\vett x)-A_{adv}(t,\vett x)\Big)^2 \mathrm{d}\vett x = 0 \ .
\end{equation}
\end{theorem}
This theorem states that for causal currents the retarded and advanced
fields are almost equal, i.e., they differ at most on a set having zero
relative measure.
To prove the theorem, let us start defining by $j_V(t,\vett x)$,
the ``truncated'' current, i.e. the function coinciding with
$j(t,\vett x)$ inside $V$, and vanishing outside it. The wave equation
(\ref{eq:onde}) can be written in Fourier space (with respect to the
spatial coordinates) as
$$
\ddot A_{\vett x} + \omega_k^2 A_{\vett x}= \hat j_V(t,\vett k) \ ,
$$
where $\omega_k=c|\vett k|$, whereas $\hat j_V(t,\vett k)$ is the
space Fourier transform of the truncated current. The retarded and advanced
solutions are then given by
\begin{equation*}
\begin{split}
A^{ret}_{\vett k} & = \int^t_{-\infty}\frac
{\sin\omega_k(t-s)}{\omega_k} \hat j_V(t,\vett k) \mathrm{d} s
\\ A^{adv}_{\vett k} & = - \int_t^{\infty}\frac
{\sin\omega_k(t-s)}{\omega_k} \hat j_V(t,\vett k) \mathrm{d} s \ ,\\
\end{split}
\end{equation*}
so that one gets
\begin{equation*}
A^{ret}_{\vett k}-A^{adv}_{\vett k} = \frac 1{2i\omega_k}\Big(
e^{i\omega_k t} \hat j_V(-\omega_k,\vett k) - e^{-i\omega_k t} \hat
j_V(\omega_k,\vett k) \Big) \ ,
\end{equation*}
where $\hat j_V(\omega,\vett k)$ is the Fourier transform, with
respect to time, of $\hat j_V(t,\vett k)$. Now one uses the
Plancherel theorem, which states
\begin{equation}\label{yyy}
\int_{\mathbb{R}^3} \Big| A^{ret}(t,\vett x)-A^{adv}(t,\vett x)\Big|^2
\mathrm{d} \vett x = \int_{\mathbb{R}^3} \Big|A^{ret}_{\vett k}-A^{adv}_{\vett
k} \Big|^2 \mathrm{d} \vett k \ ,
\end{equation}
to get (use $2|\vett a\cdot \vett b|\le a^2+b^2$)
\begin{equation}\label{eq:diff}
\begin{split}
\int_{\mathbb{R}^3} \Big| & A^{ret}(t,\vett x)-A^{adv}(t,\vett
x)\Big|^2 \mathrm{d} \vett x \\
& \le \int_{\mathbb{R}^3} \frac 1{2\omega_k^2}
\Big(|j_V(-\omega_k,\vett k)|^2 + |j_V(\omega_k,\vett k)|^2
\Big)\mathrm{d} \vett k \\
&= \frac 1{2c^2} \int \Big(|j_V(-ck,\vett k)|^2 + |j_V(ck,\vett k)|^2
\Big)\mathrm{d} k\mathrm{d}\Omega \ ,
\end{split}
\end{equation}
where $\mathrm{d}\Omega$ is the surface element on the unit
sphere in the $\vett k$ space.
We now use the causal property of the current. In fact one has the
following theorem, which will be proven below:
\begin{theorem}\label{teo:due}
If $j(t,\vett x)$ is a causal current in the sense of
Definition~\ref{hyp:1}, then one has
\begin{equation}\label{eq:CCm}
\lim_{V\to+\infty} \frac 1V \int_{\mathcal{C}} |\hat {j}_V(\omega,
\vett k)|^2 \mathrm{d} \Omega\mathrm{d} R =0 \ ,
\end{equation}
on each circular cone $\mathcal{C} \stackrel {\mathrm{def}}{=} \{ |\omega| = \alpha|\vett k|,
\alpha\ge c \}$, where $\mathrm{d}\Omega$ is the surface element on the unit
sphere in the $\vett k$ space, while $\mathrm{d} R$ runs along the cone
generatrix.
\end{theorem}
So,
dividing relation (\ref{eq:diff}) by $V$, using (\ref{eq:CCm}) with
$\alpha=c$ and taking the limit, one gets
(\ref{eq:WF}).
As a comment, one may add that from (\ref{yyy}) it is rather easily seen
that the validity almost everywhere of the Wheeler--Feynman identity implies the
vanishing of the ``spectrum of the current'', i.e. of the limit of
$|\hat {j}_V(\omega,\vett k)|^2/V$, on almost the whole
light cone $\omega^2=c^2\vett k\cdot\vett k$.
So, the problem of proving the Wheeler--Feynman identity
is reduced to proving formula
(\ref{eq:CCm}) of theorem~\ref{teo:due}.
To this end, we start defining the function
\begin{equation}\label{eq:defK}
K_V(t,\vett x) \stackrel {\mathrm{def}}{=} \int j_V(s,\vett y)j_V(s+t,\vett y +\vett x) \mathrm{d}
s\mathrm{d} \vett y \ ,
\end{equation}
which, apart from the factor $1/V$, is nothing but the correlation
of the truncated current, integrated over $s$, as one would naturally
do in defining correlations for functions having domain in
space--time. One then immediately sees that:
\begin{itemize}
\item one has
\begin{equation}\label{eq:CCK}
\lim_{V\to+\infty}\frac 1V \, K_V(t,\vett x) = 0 \ , \quad
\mbox{if}\quad c^2t^2-\vett x\cdot\vett x \le 0 \;
\end{equation}
\item the Fourier transform $\hat
K_V(\omega,\vett k)$ of $K_V(t,\vett x)$ coincides with $|\hat
j_V(\omega,\vett k)|^2$.
\end{itemize}
Indeed the first property is just a translation of the fact that
$j_V(t,\vett x)$ is causal, i.e., that (\ref{eq:2}) holds, whereas
the second one is nothing but
the ``faltung'' theorem on the Fourier transform of a convolution.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.6\textwidth]{WF_mod}
\end{center}
\caption{\label{fig:uno} The domain $\mathcal{D}$ of integration in
formula (\ref{eq:6}).}
\end{figure}
Now, considering the spherical mean of the spectrum
$|\hat j_V(\omega,\vett k)|^2$, one gets
\begin{equation}\label{eq:4}
\begin{split}
\int_{S_2} &| \hat j_V(\omega,\vett k)|^2 \mathrm{d}\Omega = \frac 1{\pi^2}
\int \mathrm{d} t\mathrm{d}\vett x \,K_V(t,\vett x) \int_{S_2} e^{i(\omega t+\vett
k\cdot\vett x)} \mathrm{d} \Omega \\ &= \frac 2{\pi} \int \mathrm{d} t\mathrm{d}\vett x
\,K_V(t,\vett x)\int_{0}^\pi e^{i(\omega t + k r\cos \theta)} \sin
\theta \mathrm{d} \theta \\ &= \frac 2{\pi^2} \int \mathrm{d} t\mathrm{d} r \, r \frac
{e^{i(\omega t + k r)} - e^{i(\omega t - k r)}}{ik}
\int_{S_2} K_V(t,\vett x) \mathrm{d} \Omega\\ &=\frac 2{\pi} \int
\mathrm{d} t\mathrm{d} r\, r \tilde K_V(t,r) \frac { e^{i(\omega t + k r)}
- e^{i(\omega t - k r)}}{ik} \ ,
\end{split}
\end{equation}
where $\tilde K_V(t,r)$ is the spherical mean of
$K_V(t,\vett x)$. Now, if one makes use of the of parity property of
the correlation $K_V(t,\vett x)=K_V(-t,-\vett x)$, which easily follows
from the very definition \eqref{eq:defK}, one finds that the spherical
mean $\tilde K_V(t,r)$ is an even function of time, so that the
imaginary part of the integral in the last line of \eqref{eq:4}
vanishes, and one gets
\begin{equation}\label{eq:5}
\begin{split}
\int_{S_2} | \hat j_V(\omega,\vett k)|^2 \mathrm{d}\Omega & =\frac 2{\pi} \int
\mathrm{d} t\mathrm{d} r \, r \tilde K_V(t,r) \\
&\Big[ \frac { \sin(\omega t + k r)}k - \frac
{\sin(\omega t - k r)}k \big] \ .
\end{split}
\end{equation}
Consider now ``a ray'' in the momentum $(\omega,\vett k)$ space, i.e. all
vectors of the form $(R\omega_0,R{\vett k}_0)$, $R>0$, and integrate relation
\eqref{eq:5} along this ray: one gets
\begin{equation*}
\begin{split}
&\int_0^{\infty }\mathrm{d} R\, \int_{S_2} | \hat j_V(R\omega_0,R\vett
k_0)|^2 \mathrm{d}\Omega =\frac 2{\pi} \int \mathrm{d} t\mathrm{d} r \, r \tilde
K_V(t,r) \frac 1{k_0} \\ &\Big[ \int_0^{\infty }\mathrm{d} R\,\frac {
\sin\big(R(\omega_0 t + k_0 r)\big)}R - \int_0^{\infty }\mathrm{d}
R\,\frac {\sin\big(R(\omega_0 t - k_0 r)\big)}{R} \Big] \ .
\end{split}
\end{equation*}
Now using the relation
\begin{equation*}
\int_0^{\infty }\mathrm{d} R\,\,\frac { \sin \alpha R } R = \left\{
\begin{split}
\frac {\pi}2 \quad &\text{if} \quad \alpha >0 \\ 0\quad
&\text{if} \quad \alpha =0 \\ -\frac {\pi}2\quad &\text{if}
\quad \alpha <0 \ ,
\end{split}
\right.
\end{equation*}
one gets
\begin{equation} \label{eq:6}
\begin{split}
\int_0^{\infty }\mathrm{d} R\, \int_{S_2} | \hat j_V(R\omega_0,R\vett
k_0)|^2 & \mathrm{d}\Omega = \\
& 2 \int_{\mathcal{D}(\omega_0,k_0)} \mathrm{d} t\mathrm{d} r\, r
\tilde K_V(t,r)
\end{split}
\end{equation}
where the domain $\mathcal{D}(\omega_0,k_0)$ (depicted in
figure~\ref{fig:uno}) is the domain in the half--plane $r>0$, bounded
by the two half--lines $\omega_0 t \pm k_0r=0$. Now, dividing by
$V$ and taking the limit, the integral is seen to vanish if
$\omega_0^2-k_0^2\ge0$. In fact, by the causality
property (\ref{eq:CCK}), in that limit $\tilde K_V(t,r)/V$
vanishes for all points
inside the region bounded by the lines $ct\pm r=0$, i.e., in
particular, for all points of $\mathcal{D}(\omega_0,k_0)$. So
(\ref{eq:CCm}) holds and Theorem \ref{teo:main} is proven.
\subsection*{Use of the identity in canceling the radiation reaction
forces}
In their paper \cite{wf}, Wheeler and Feynman showed how the condition
$$
A_{ret}^\mu-A_{adv}^\mu=0\
$$
implies the vanishing of the radiation reaction force
(or self force) acting on each charge. One starts from the
relativistic equation of motion for the charge
$$
m\ddot q^\mu = f_{mec}^\mu + \tilde F^{\mu\nu}_{ret}\,\dot q_\nu +\frac
{2e^2}{3c^3}\Big( \dddot q^{\mu} +\ddot q^\nu \ddot q_\nu \dot q^\mu \Big) \ ,
$$
where $m$ and $e$ are the charge and the mass of the particle, dots
represent derivatives with respect to proper time, repeated
index means summation (Einstein convention), $ f_{mech}^\mu$ is a
four--force of mechanical (non electromagnetic) type, while
$\tilde F^{\mu\nu}_{ret}$ is the retarded electromagnetic field due to
all other charges,
evaluated at the four--position $q^\mu$ of the considered charge,
and finally the term $\frac
{2e^2}{3c^3}\Big( \dddot q^{\mu} +\ddot q^\nu \ddot q_\nu \dot q^\mu
\Big)$ is the relativistic expression for the radiation reaction
force.
The electromagnetic field $\tilde F^{\mu\nu}_{ret}$, or
rather the field $\tilde F_{ret,\mu\nu}$, is defined as
$$
\tilde F_{ret,\mu\nu} = \sum \Big(\partial_\mu A_{ret,\nu}^k
- \partial_\nu A_{ret,\mu}^k \Big) \ ,
$$
where $ A_{ret,\nu}^k$ is the retarded field produced by the $k$--th
charge, and the summation is extended over all charges but the
considered one. The field $\tilde F_{ret,\mu,\nu}$ can be rewritten in
a more useful form as
\begin{equation*}
\begin{split}
\tilde F_{ret,\mu\nu} &= \sum \Big(\partial_\mu \frac {A_{ret,\nu}^k
+A_{adv,\nu}^k}2
- \partial_\nu \frac {A_{ret,\mu}^k+A_{adv,\mu}^k}2\Big) \\
&+ \sum \Big(\partial_\mu \frac {A_{ret,\nu}^k - A_{adv,\nu}^k}2
- \partial_\nu \frac {A_{ret,\mu}^k- A_{adv,\mu}^k}2\Big)
\ ,
\end{split}
\end{equation*}
because, as we will show below, the Wheeler-- Feynman identity implies that
\begin{equation}\label{eq:canc}
\begin{split}
\sum &\Big(\partial_\mu \frac {A_{ret,\nu}^k - A_{adv,\nu}^k}2
- \partial_\nu \frac {A_{ret,\mu}^k- A_{adv,\mu}^k}2\Big) = \\
& - \frac{2e^2}{3c^3}\Big( \dddot q^{\mu} -\ddot q^\nu \ddot q_\nu \dot q^\mu
\Big)\ ,
\end{split}
\end{equation}
so that the equations of motion, at the end, can be written as
$$
m\ddot q^\mu = f_{mec}^\mu + \frac {\tilde F^{\mu\nu}_{ret} +
\tilde F^{\mu\nu}_{adv}}2 \dot q_\nu
$$
with
\begin{equation*}
\begin{split}
\frac {\tilde F_{ret,\mu\nu} + \tilde F_{adv,\mu\nu}}2 &= \\
\sum \Big( \partial_\mu &\frac {A_{ret,\nu}^k + A_{adv,\nu}^k}2 -
\partial_\nu \frac {A_{ret,\mu}^k+A_{adv,\mu}^k}2\Big) \ .
\end{split}
\end{equation*}
The new form of the equations of motion clearly shows that they are
indeed reversible and the radiation reaction has disappeared. So, such
a force force cannot be held responsible for the emission.
To show how relation (\ref{eq:canc})
follows from the Wheeler--Feynman identity, one first has to notice that
such an identity states that one has
$$
A_{\mu,ret}-A_{\mu,adv} = \sum_{\mbox{all}} \left({A_{ret,\mu}^k-
A_{adv,\mu}^k}\right)=0\ ,
$$
where the sum is extended to all charges. Thus, at all points
$x^\mu\ne q^\mu$ (i.e., at all points different from the
four--position of the considered charge) one has
\begin{equation}\label{eq:zerotot}
\sum_{\mbox{all}} \Big(\partial_\mu \frac {A_{ret,\nu}^k
- A_{adv,\nu}^k}2
- \partial_\nu \frac {A_{ret,\mu}^k- A_{adv,\mu}^k}2\Big) = 0 \ ,
\end{equation}
because the vanishing of the potentials implies the vanishing of their
derivatives.
Now, it was shown by Dirac (see \cite{dirac}) that for the
field $ \frac {A_{ret,\mu}^j- A_{adv,\mu}^j}2$ created by the particle
$q^\mu$ itself one has
\begin{equation*}
\begin{split}
\lim_{x^\mu\to q^\mu} \Big(\partial_\mu \frac {A_{ret,\nu}^j - A_{adv,\nu}^j}2
- &\partial_\nu \frac {A_{ret,\mu}^j- A_{adv,\mu}^j}2\Big) \dot q^{\mu} = \\
&\, \frac {2e^2}{3c^3}\Big( \dddot q^{\mu} +\ddot q^\nu \ddot q_\nu \dot q^\mu
\Big)\ ,
\end{split}
\end{equation*}
while on he other hand the remaining
fields are regular at $q^\mu$. So taking the limit of
the previous relation (\ref{eq:zerotot}) for $x^\mu\to q^\mu$, one
gets (\ref{eq:canc}).
|
\section{Introduction}\label{sect:intro}
In this paper we deal with the existence, non-existence and localization of positive solutions of the following system of parabolic equations subject to nonlinear, nonlocal initial conditions
\begin{equation}\label{eq:parabolic-intro}
\begin{cases}
u_t-\Delta u = f(t,x,u,v), & (t,x)\in (0,{t_{\text{max}}})\times\Omega,\\
v_t-\Delta v = g(t,x,u,v), & (t,x)\in (0,{t_{\text{max}}})\times\Omega,\\
u(t,x)=v(t,x)=0, & (t,x)\in(0,{t_{\text{max}}})\times\partial\Omega,\\
u(0,\cdot)=\alpha(u,v),\\
v(0,\cdot)=\beta(u,v),
\end{cases}
\end{equation}
where $\Omega \subset \mathbb{R}^{m}$ is a bounded domain that is Dirichlet regular, $f,g\colon (0,{t_{\text{max}}})\times\Omega\times\mathbb{R}_+\times\mathbb{R}_+\to\mathbb{R}_+$ are continuous functions and $\alpha$ and $\beta$ belong to a fairly general class of nonlinear operators that covers a variety of cases, a \emph{particular} example being
\begin{equation}\label{eq:ntime-intro}
\alpha(u,v)(x)=\int_0^{{t_{\text{max}}}} u(t,x)dt,\quad \beta(u,v)(x)=\int_0^{{t_{\text{max}}}} v(t,x)dt.
\end{equation}
Initial nonlocal conditions have been investigated in a variety of settings, for example in the case of multi-point \cite{b:Chab},
integral \cite{b:tc-rp-pr, b:Deng, b:olm-rob, b:pao, b:ras-kar1, b:ras-kar2} and nonlinear conditions \cite{b:b-t-v, b:Bou1, b:Bou2, b:Bou3}, see also the recent review \cite{b:Stik}.
In particular, a physical motivation for the integral form of the initial condition is given in \cite{b:olm-rob} for the one-dimensional heat equation. Furthermore a number of applications of nonlocal problems for evolution equations are illustrated in Section 10.2 of \cite{b:McKib}.
In a recent paper \cite{b:gi-mm-rp} the authors investigated the existence, localization and multiplicity of positive solutions of systems of $(p,q)$-Laplacian equations subject to Dirichlet boundary conditions. The main tool in \cite{b:gi-mm-rp} is the development of a
general abstract framework for the existence of fixed points of nonlinear operators acting on cones that satisfy an inequality of Harnack-type. Within this setting, the authors of \cite{b:gi-mm-rp} used the Granas fixed point index (see for example \cite{b:Deim,b:gd}) and, in order to compute the index, they used some estimates from above, using the norm, and from below, utilizing a seminorm. Here we \emph{extend} the theoretical results of \cite{b:gi-mm-rp} to a more general setting; this generalization is \emph{apparently}
simple but fruitful and is motivated by the application to the parabolic system \eqref{eq:parabolic-intro}.
In particular, we replace the use of the seminorm with the use of a more general positively homogeneous functional and, moreover, we relax the assumptions on the cone. The Remarks \ref{comp}, \ref{rem:choiceC0}, \ref{rem:nopsi} and \ref{rem:WhySuchHarnack} illustrate in details the differences between the two theoretical approaches and their applicability.
We point out that our new approach is quite general and covers, \emph{as a special case}, the system \eqref{eq:parabolic-intro}.
Inspired by the setting considered in \cite{b:rp-par}, where the problem of \emph{one} parabolic equation with nonlocal, \emph{linear} initial condition was studied, we apply our theoretical results to the parabolic system \eqref{eq:parabolic-intro}. In contrast with the paper \cite{b:gi-mm-rp}, where the space $L^\infty$ with an integral seminorm was used, here, in order to seek mild solutions of our problem, we use the classical space of continuous functions, with a very natural positively homogeneous functional, namely the minimum on a suitable subset. A similar idea has been used with success in the context of ordinary differential equations and integral equations, see for example
\cite{b:guolak}. In our case, a key role for our multiplicity results is played by a weak Harnack-type inequality, see Remark \ref{rem:Harnack}.
In the case of the system~\eqref{eq:parabolic-intro} we obtain existence, localization, multiplicity and non-existence of positive mild solutions.
We illustrate in an example the applicability of our results and we show that the constants that occur in our theory can be computed.
\section{Abstract existence theorems}\label{sect:abstr}
Let $(E_{i},|\cdot|_i)$ ($i=1,2$) be Banach spaces and let $\fl{\cdot}_i$ be positively homogeneous continuous functionals on $E_i$. In what follows, we omit the subscript in $\fl\cdot_i$, when confusion is unlikely.
Let also $G_i\subset E_i$ be closed convex wedges, which is understood to mean that
\[\lambda u+\mu v\in G_i\text{ for all }u,v\in G_i\text{ and }\lambda,\mu\geq 0.\] Moreover, let $K_i\subset G_i$ be closed convex cones, which means that $K_i$ are closed convex wedges such that $K_i\cap(-K_i)=\{0\}$. The wedges induce the natural semiorders $\preceq$ on $E_i$ in the following way: $$u\preceq v\ \text{if and only if}\ v-u\in G_i,\, u,v\in E_i.$$ By a semiorder we mean that the relation $\preceq$ is reflexive and transitive, but not necessarily antisymmetric.
We assume that the functionals $\fl{\cdot}_i$ are monotone on $K_i$ with respect to the semiorder $\preceq$, that is for $i\in\{1,2\}$
we have
\begin{equation}\label{eq:fl monot}
\text{if }u\preceq v\text{ then }\fl u \leq\fl v\text{ for }u,v\in K_i.
\end{equation}
In particular we have $\fl u\geq 0$ for $u\in K_i$.
We assume that there exist some elements $\psi_i\in K_i$ such that $|\psi_i|=1$ and for $i\in\{1,2\}$
\begin{equation}\label{eq: prec-max} u\preceq |u|\psi_i\text{ for all }u\in K_i.\end{equation}
Note that \eqref{eq:fl monot} and \eqref{eq: prec-max} yield
\begin{equation}\label{eq:rel-fl-norm}
\fl u_i\leq \fl{\psi} |u|\text{ for all } u\in K_i.\end{equation}
In particular, we have $\fl{\psi_i}>0$ if $\fl\cdot$ is nonzero.
In what follows by the \emph{compactness} of a continuous operator we mean the
relative compactness of its range. By the \emph{complete continuity} of a
continuous operator we mean the relative compactness of the image of every
bounded set of the domain.
We seek the fixed points of a completely continuous operator
\[N:=(N_1,N_2)\colon K_1\times K_2\to K_1\times K_2,\]
that is $(u,v)\in K_1\times K_2$ such that $N(u,v) =(u,v)$.
We shall discuss not only the existence, but also the localization and multiplicity
of the solutions of the nonlinear equation $N(u,v) =(u,v)$. In order to do this, we utilize the Granas fixed point index, $\mathrm{ind}_C(f,U)$ (for more information on the index and its applications
we refer the reader to \cite{b:Deim, b:gd}).
The next Proposition describes some of the useful properties of the index,
for details see Theorem 6.2, Chapter 12 of \cite{b:gd}.
\begin{prop}
\label{fpi-pro} Let $C$ be a closed convex subset of a Banach space, $%
U\subset C$ be open in $C$ and $f\colon \overline{U}\rightarrow C$ be a
compact map with no fixed points on the boundary $\partial U$ of $U.$ Then
the fixed point index has the following properties:
\emph{(i)} (Existence) If $\mathrm{ind}_{C}(f,U)\neq 0$ then $\mathrm{fix}%
(f)\neq \emptyset $, where $\mathrm{fix}f=%
\mbox{$\left\{x\in\bar
U:\;f(x)=x\right\}$}$.
\emph{(ii)} (Additivity) If $\mathrm{fix}f\subset U_1 \cup U_2\subset U$
with $U_1, U_2$ open in $C$ and disjoint, then
\begin{equation*}
\mathrm{ind}_{C}(f,U)=\mathrm{ind}_{C}(f,U_1 )+\mathrm{ind}_{C}(f,U_2).
\end{equation*}
\emph{(iii)} (Homotopy invariance) If $h\colon \overline{U}\times \lbrack
0,1]\rightarrow C$ is a compact homotopy such that $h(u,t)\neq u$ for $u\in
\partial U$ and $t\in \lbrack 0,1]$ then
\begin{equation*}
\mathrm{ind}_{C}(h(\cdot,0),U)=\mathrm{ind}_{C}(h(\cdot,1),U).
\end{equation*}
\emph{(iv)} (Normalization) If $f$ is a constant map, with $f(u)=u_{0}$ for
every $u\in \overline{U},$ then
\begin{equation*}
\mathrm{ind}_{C}(f,U)=%
\begin{cases}
1,\quad \text{if}\ u_{0}\in U \\
0,\quad \text{if}\ u_{0}\notin \overline{U}.%
\end{cases}%
\end{equation*}
In particular, $\mathrm{ind}_{C}(f,C)=1$ for every compact function $%
f:C\rightarrow C,$ since $f$ is homotopic to any $u_{0}\in C,\ $by the
convexity of $C$ (take $h\left( u,t\right) =tf\left( u\right) +\left(
1-t\right) u_{0}$).
\end{prop}
\subsection{Fixed point results}
We begin with two theorems on the existence and localization of one
solution of the operator equation $N(u,v)=(u,v)$. Set $K=K_1\times K_2$ and
\[C=C(R_1,R_2):=\left\{ \left( u,v\right) \in K_1\times K_2:\left| u\right| \leq R_1,\ \left| v\right| \leq R_2\right\},\]
for some fixed numbers $R_1, R_2$.
The first Theorem is a generalization of Theorem 2.17 of \cite{b:gi-mm-rp}.
\begin{thm}\label{thm:existence 1a}
Assume that there exist numbers $r_i,R_i$, $i=1,2$ with $0<r_i<\fl{\psi_i}R_i$ such that
\begin{equation}\label{eq:ind0}
\inf\limits_{\substack{ (u,v)\in C \\ \fl u =r_1,\fl v \leq
r_2}}\fl{ N_1 (u,v)} > r_1,\ \
\inf\limits_{\substack{ (u,v)\in C \\ \fl u \leq r_1,\fl v
=r_2}}\fl{ N_2(u,v)} > r_2,
\end{equation}
and
\begin{equation}
\sup\limits_{(u,v)\in C}|N_{i}(u,v)|\leq R_{i}\ \ \ \ (i=1,2). \label{eq:ind1}
\end{equation}
Then $N$ has at least one fixed point $(u,v)\in K_1\times K_2$ such that $|u|\leq R_1$,
$|v|\leq R_2$ and either $\fl u > r_1 $ or $\fl v >r_2$.
\end{thm}
\begin{proof}[Proof of Theorem \ref{thm:existence 1a}]
The assumption \eqref{eq:ind1} implies that $N(C)\subset C$. Therefore, by
Proposition~\ref{fpi-pro}, we obtain $\mathrm{ind}_{C}(N,C)=1$. Let
\begin{equation*}
U:=\left\{ \left( u,v\right) \in C:\fl u <r_1,\ \fl v <r_2\right\}.
\end{equation*}%
This is an open set, whose boundary $\partial U$ with respect to $C$ is equal to $\partial
U=A_1 \cup A_2,$ where
\begin{eqnarray*}
A_1 &=&\left\{ \left( u,v\right) \in C:\fl u =r_1,\
\fl v \leq r_2\right\},\ \\
A_2 &=&\left\{ \left( u,v\right) \in C:\fl u \leq
r_1,\ \fl v =r_2\right\} .
\end{eqnarray*}%
Observe that \eqref{eq:ind0} implies that there are no fixed points of $N$ on $\partial U$. Therefore, the indices $\mathrm{ind}_{C}(N,U)$ and $\mathrm{ind}_{C}(N,C\setminus \overline{U})$ are well defined and their sum, by the additivity property of the index, is equal to one. Therefore, it
suffices to prove that $\mathrm{ind}_{C}(N,U)=0.$ Take $h=(R_1 \psi_1,R_2\psi _2)\in C$ and consider the homotopy $H:C\times \left[ 0,1\right] \rightarrow C$,
\begin{equation*}
H\left( u,v,t\right) :=th+(1-t)N(u,v).
\end{equation*}%
We claim that $H$ is fixed point free on $\partial U$. Since
\begin{eqnarray}\label{star}
\fl {R_{i}\psi _{i}} &=&R_{i}\fl{ \psi _{i}}>r_{i}, \quad i=1,2.
\end{eqnarray}%
we have $\left( u,v\right) \neq h=H\left( u,v,1\right) $ for all $%
\left( u,v\right) \in \partial U.$ It remains to show that $H\left(
u,v,t\right) \neq \left( u,v\right) $ for $\left( u,v\right) \in \partial U$
and $t\in \left( 0,1\right) .$ Assume the contrary. Then there exists $%
(u,v)\in A_1 \cup A_2$ and $t\in (0,1)$ such that
\begin{equation}
(u,v)=th+(1-t)N(u,v). \label{eq:homotopy not good}
\end{equation}%
Suppose that $(u,v)\in A_1 .$ Then,
\[N_1(u,v)\preceq |N_1(u,v)|\psi_1\preceq R_1\psi_1\] and
exploiting the first coordinate of the equation \eqref{eq:homotopy not good}, we obtain
\begin{equation}
u = N_1 (u,v)+ t(R_1 \psi _1 -N_1 (u,v))\succeq N_1(u,v).
\end{equation}%
Using the monotonicity of $\fl\cdot $ and \eqref{eq:ind0} we obtain
$r_1=\fl u\geq \fl{N_1(u,v)}>r_1$, which is impossible. Similarly, we derive a contradiction if $\left(u,v\right) \in A_2.$
By the homotopy invariance of the index we obtain $\mathrm{ind}_C(N,U)=\mathrm{ind}_C(h,U).$ From (\ref{star}) we have $h\not\in \overline{U}$, hence $\mathrm{ind}_C(N,U)=\mathrm{ind}_C(h,U)=0,$ as we wished.
\end{proof}
\begin{rem}\label{rem:inf>=}
From the proof we can deduce that if we change the assumption \eqref{eq:ind0} into
\begin{equation}\label{eq:ind0>=}
\inf\limits_{\substack{ (u,v)\in C \\ \fl u =r_1,\fl v \leq
r_2}}\fl{ N_1 (u,v)} \geq r_1,\ \
\inf\limits_{\substack{ (u,v)\in C \\ \fl u \leq r_1,\fl v
=r_2}}\fl{ N_2(u,v)} \geq r_2,
\end{equation}
then we obtain at least one fixed point $(u,v)\in C$ with slightly weaker localization: $\fl u \geq r_1 $ or $\fl v \geq r_2$. The assumption \eqref{eq:ind0>=} permits the existence of fixed points of $N$ on $\partial U$. The assumption \eqref{eq:ind0} is more convenient when dealing with multiplicity results.
\end{rem}
\begin{rem}\label{rem:lowerbound_norm} We observe that, using the relation \eqref{eq:rel-fl-norm}, a lower bound for the solution in terms of the functional $\fl\cdot$ provides a lower bound for the norm of the solution, namely
\[\fl{u}> r_1 \implies |u|> \fl{\psi_1}^{-1} r_1,\quad \fl{v}> r_2 \implies |v|> \fl{\psi_2}^{-1} r_2.\]
\end{rem}
\begin{rem}\label{comp}
The main differences between Theorem \ref{thm:existence 1a} and Theorem 2.17 of \cite{b:gi-mm-rp} consist in:
\begin{itemize}
\item The possibility of considering a positively homogeneous functional $\fl\cdot$ instead of a seminorm.
\item The assumption on the cone; in Theorem 2.17 of \cite{b:gi-mm-rp} it is needed the existence of $\psi$ such that $u\leq |u|\psi_i$ for all $u\in E_i$, where $\leq$ is the order induced by the cone $K_i$, $i=1,2$. Here, instead, we can consider a semiorder.
\end{itemize}
Under the point of view of the applicability of our novel approach to parabolic problems, this is highlighted in the Remarks \ref{rem:choiceC0}, \ref{rem:nopsi} and \ref{rem:WhySuchHarnack}.
\end{rem}
The second Theorem is in the spirit of Theorem 2.9 and Remark 2.16 of \cite{b:gi-mm-rp}.
\begin{thm}\label{thm:existence 2a}
Assume that there exist numbers $r_i,R_i$, $i=1,2$ with $0<r_i<\fl{\psi_i}R_i$ such that
\begin{equation}\label{eq:ind0'}
\inf\limits_{\substack{ (u,v)\in C \\ \fl u =r_1,\fl v \geq
r_2}}\fl{ N_1 (u,v)} > r_1,\ \
\inf\limits_{\substack{ (u,v)\in C \\ \fl u \geq r_1,\fl v
=r_2}}\fl{ N_2(u,v)} > r_2,
\end{equation}
and
\begin{equation}\label{eq:ind1'}
\sup\limits_{(u,v)\in C}|N_{i}(u,v)|\leq R_{i}\ \ \ \ (i=1,2).
\end{equation}
Then $N$ has at least one fixed point $(u,v)\in K_1\times K_2$ such that $|u|\leq R_1,$
$|v|\leq R_2$ and $\fl u > r_1 $, $\fl v > r_2.$
\end{thm}
\begin{proof}[Sketch of the proof]
The proof is similar to the proof of Theorem \ref{eq:ind0} and \cite[Theorem 2.9]{b:gi-mm-rp}.
As before, the assumption \eqref{eq:ind1'} implies that $N(C)\subset C.$ Thus, $%
\mathrm{ind}_C(N,C)=1.$ In order to finish the proof, it is sufficient to
show that $\mathrm{ind}_C(N,V)=0,$ where
\begin{equation*}
V:=\left\{ \left( u,v\right) \in C:\fl{ u} <r_1 \ \text{or\
}\fl{ v} <r_2\right\} .
\end{equation*}%
We have $\partial V=B_1 \cup B_2,$ where
\begin{eqnarray*}
B_1 &=&\left\{ \left( u,v\right) \in C:\fl{ u} =r_1,\
\fl{ v} \geq r_2\right\}, \\
B_2 &=&\left\{ \left( u,v\right) \in C:\fl{ u} \geq
r_1,\ \fl{ v} =r_2\right\} .
\end{eqnarray*}%
By \eqref{eq:ind0'} we obtain $N$ has no fixed points on $\partial V$. Consider the same homotopy as in the proof of Theorem \ref{thm:existence 1a}, that is
\begin{equation*}
H\left( u,v,t\right) =th+(1-t)N(u,v),\ \text{ where }\ h=(R_1 \psi
_1,R_2\psi _2).
\end{equation*}%
As before we can prove that $H$ is fixed point free on $\partial V$. Therefore $\mathrm{ind}_C(N,V)=\mathrm{ind}_C(h,V)=0$, since, like in the previous proof, $h\notin \overline{V}.$
\end{proof}
The result, in the spirit of Lemma~4 of \cite{b:gipp-nonlin}, allows different types
of growth of the operators and is a modification of Theorem 2.4 of \cite{b:gi-mm-rp}.
\begin{thm}
\label{thm:existence or a} Assume that there exist numbers $r_i,R_i$ with $0<r_i<\fl{\psi_i}R_i$ such that
\begin{equation}\label{eq:ind1''}
\sup\limits_{(u,v)\in C}|N_{i}(u,v)|\leq R_{i}\quad (i=1,2),
\end{equation}%
and
\begin{equation}\label{eq:ind0 or}
\inf_{(u,v)\in A} \fl{N_1(u,v)} \geq r_1\ {\text{ or }}\ \inf_{(u,v)\in A}
\fl{N_2(u,v)} \geq r_2,
\end{equation}
where $A$ is a subset of the set
\begin{equation*}
U=\mbox{$\left\{(u,v)\in C:\;\fl u < r_1,\ \fl v< r_2\right\}$}.
\end{equation*}
Then $N$ has at least one fixed point $(u,v)\in K_1\times K_2$ such that $|u|\leq R_1,$
$|v|\leq R_2$ and $(u,v)\not\in A.$
\end{thm}
\begin{proof}
Since $N$ is a completely continuous mapping in the bounded closed convex
set $C,$ by Schauder's fixed point theorem, it possesses a fixed point $%
(u,v)\in C.$ We now show that the fixed point is not in $A.$ Suppose on the
contrary that $(u,v)=N(u,v)$ and $(u,v)\in A.$ Suppose that the first
inequality from \eqref{eq:ind0 or} is satisfied. Then
\begin{equation*}
r_1 >\fl u =\fl{ N_1 (u,v)} \geq r_1,
\end{equation*}%
which is impossible. Similarly we arrive at a contradiction, if the second
inequality from \eqref{eq:ind0 or} is satisfied.
\end{proof}
\subsection{Multiplicity results}
We present now some multiplicity results that are analogues of the results of Subsection 2.3 of \cite{b:gi-mm-rp}.
\begin{thm}\label{thm:existence 3sols a}
Assume that there exist numbers $\rho _{i}, r_{i},
R_{i}$ with
\begin{equation}
0<\fl{\psi_i}\rho _{i}<r_{i}<\fl{\psi_i} R_{i}\ \ \
(i=1,2),
\end{equation}%
such that
\begin{equation}\label{eq:ind0'i}
\inf\limits_{\substack{ (u,v)\in C \\ \fl u =r_1,\fl v \geq
r_2}}\fl{ N_1 (u,v)} > r_1,\ \
\inf\limits_{\substack{ (u,v)\in C \\ \fl u \geq r_1,\fl v
=r_2}}\fl{ N_2(u,v)} > r_2,
\end{equation}
\begin{equation}\label{eq:ind1'''}
\sup\limits_{(u,v)\in C}|N_{i}(u,v)|\leq R_{i}\ \ (i=1,2), \end{equation}%
and
\begin{equation}\label{eq:ind1-rho} N(u,v)\neq \lambda (u,v),\ \text{ for }\lambda \geq 1\text{ and }(\left| u\right| =\rho _1,\ \left| v\right| \leq \rho _2\text{ or } \left| u\right| \leq \rho _1,\ \left| v\right| =\rho _2).
\end{equation}
Then $N$ has at least three fixed points $(u_{i},v_{i})\in C$ $(
i=1,2,3) $ with
\begin{align*}
| u_1 | &<\rho _1,\ | v_1 |<\rho _2\ ( \text{possibly zero solution}); \\
\fl{u_2} &<r_1 \text{ or } \fl{v_2} <r_2;\ | u_2| >\rho _1 \ \text{or }|v_2| >\rho _2\ ( \text{possibly one solution component zero}) ; \\
\fl{ u_{3}} &>r_1,\ \fl{v_{3}}>r_2\ (\text{both solution components nonzero}).
\end{align*}
\end{thm}
\begin{proof}
Let $U,V$ be as in the proof of Theorems \ref{thm:existence 1a} and \ref{thm:existence 2a}. Strict inequalities
in (\ref{eq:ind0'i}) guarantee that $N$ is fixed point free on $\partial V.$
According to the proof of Theorem \ref{thm:existence 2a} we have $\mathrm{ind}%
_{C}(N,C)=1,$ $\mathrm{ind}_{C}(N,V)=0$ and therefore by the additivity
property, $\mathrm{ind}_{C}(N,C\setminus \overline{V})=1.$ Let%
\begin{equation*}
W:=\{ ( u,v) \in C:| u| <\rho _1,\
| v| <\rho _2\} .\text{ }
\end{equation*}%
From \eqref{eq:rel-fl-norm}, for every $( u,v) \in \overline{W},$ we have
\begin{equation*}
\fl{u}\leq \fl{\psi_1} |u| \leq \fl{\psi_1}\rho_1 <r_1
\end{equation*}%
and, similarly, $\fl v <r_2.$ Hence $(u,v)
\in U,$ which proves that $\overline{W}\subset U \subset V.$ Condition (\ref%
{eq:ind1-rho}) shows that $N$ is homotopic with zero on $W.$ Thus $\mathrm{%
ind}_{C}(N,W)=\mathrm{ind}_{C}(0,W)=1.$ Then $\mathrm{ind}_{C}(N,V\setminus
\overline{W})=0-1=-1.$ Consequently, there exist at least three fixed points
of $N,$ in $W,\ V\setminus \overline{W}$ and $C\setminus \overline{V}.$
\end{proof}
If we assume the following estimates of $\fl{N_{i}(u,v)}$:
\begin{equation}\label{eq:ind0''}
\inf\limits_{\substack{ (u,v)\in C \\ \fl u =r_1 }}\fl{ N_1 (u,v)} > r_1,\ \
\inf\limits_{\substack{ (u,v)\in C \\ \fl v =r_2}}\fl{ N_2(u,v)} > r_2,
\end{equation}
then we can obtain a more precise localization for the solution $\left(
u_2,v_2\right) $ in Theorem \ref{thm:existence 3sols a}, the Figure \ref{fig:3sols} (analogous to Figure 1 of \cite{b:gi-mm-rp}) illustrates this fact.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{3sols_functional.pdf}
\caption{Localization of the three solutions $(u_{i},v_{i})$ from Theorem
\protect\ref{thm:existence 3sols a} (on the left) and Theorem \protect\ref{thm:existence 3sols a'}
(on the right).}
\label{fig:3sols}
\end{figure}
\begin{thm}
\label{thm:existence 3sols a'} Suppose that all the assumptions of Theorem \emph{%
\ref{thm:existence 3sols a}} are satisfied with the condition \emph{(\ref{eq:ind0'i})}
replaced by \emph{(\ref{eq:ind0''})}. Then $N$ has at least three fixed points $%
(u_{i},v_{i})\in C$ $\left( i=1,2,3\right) $ with%
\begin{eqnarray*}
| u_1 | &<&\rho _1,\ \ | v_1 |
<\rho _2\ \ ( \text{possibly zero solution}); \\
\fl{u_2} &<&r_1,\ \ \fl{v_2}
<r_2;\ | u_2| >\rho _1 \ \text{or }|
v_2| >\rho _2\ \ ( \text{possibly one solution component
zero}) ; \\
\fl{ u_{3}} &>&r_1,\ \fl{v_{3}}>r_2\ \
( \text{both solution components nonzero}).
\end{eqnarray*}
\end{thm}
\begin{proof}
The assumption (\ref{eq:ind0''}) implies both (\ref{eq:ind0}) and (\ref{eq:ind0'}) and that
there are no fixed points of $N$ on $\partial U$ and $\partial V.$ Hence, as
in the proofs of Theorems \ref{thm:existence 1a} and \ref{thm:existence 2a}, the
indices $\mathrm{ind}_{C}(N,U)$ and $\mathrm{ind}_{C}(N,V)$ are well defined
and equal $0.$ An analysis similar to that in the proof of Theorem \ref%
{thm:existence 3sols a} shows that
\begin{equation*}
\mathrm{ind}_{C}(N,W)=1,\ \mathrm{ind}_{C}(N,U\setminus \overline{W})=-1,\
\mathrm{ind}_{C}(N,C\setminus \overline{V})=1,
\end{equation*}%
which completes the proof.
\end{proof}
In order to ensure that the solution $(u_1,v_1 )$ from the theorems above
is nonzero, and thereby to obtain three \emph{nonzero} solutions, we use
some additional assumptions on $N$.
\begin{thm}\label{thm:existence 3sols a''}
Assume that all the conditions of Theorem {\ref{thm:existence 3sols a}} or Theorem {\ref{thm:existence 3sols a'}} are satisfied.
Consider $0<\varrho_i<\fl{\psi_i}\rho _i,
$ $i=1,2.$
\emph{(i)} If $N_1(0,0)\neq 0$ or $N_2(0,0)\neq 0$, then the solution $(u_1,v_1 )$ from Theorem {\ref{thm:existence 3sols a}} or {\ref{thm:existence 3sols a'}} is nonzero.
\emph{(ii)} If
\begin{equation}
\inf\limits_{\substack{ (u,v)\in K,|u|\leq \rho _1,|v|\leq \rho _2 \\
\fl u =\varrho _1,\fl v \geq \varrho _2}}\fl{ N_1 (u,v)} \geq \varrho_1,\ \
\inf\limits_{\substack{ (u,v)\in K,|u|\leq \rho _1,|v|\leq \rho _2 \\
\fl u \geq \varrho _1,\fl{v} =\varrho _2}}\fl{N_2(u,v)} \geq \varrho_2
\label{6varrho}
\end{equation}%
and
\begin{equation*}
|N_i (u,v)|\leq \rho _i \ \ \text{for\ \ }|u|\leq \rho _1,\ |v|\leq \rho
_2\ \ \ \left( i=1,2\right),
\end{equation*}%
then we can assume that the solution $(u_1,v_1 )$ from Theorem {\ref{thm:existence 3sols a}} or Theorem {\ref{thm:existence 3sols a'}} satisfies $\fl{u_1} \geq \varrho _1 $ and $\fl{v_1}\geq \varrho _2$;
\emph{(iii)} if
\begin{equation}
\inf\limits_{\substack{ (u,v)\in K,|u|\leq \tilde{\rho}_1,|v|\leq \tilde{
\rho}_2 \\ \fl u < \varrho_1,\fl v < \varrho _2}}
\fl{N_1 (u,v)} \geq \varrho _1 \ {\text{ or }}\ \inf\limits
_{\substack{ (u,v)\in K,|u|\leq \tilde{\rho}_1,|v|\leq \tilde{\rho}_2
\\ \fl u < \varrho _1,\fl v < \varrho _2}}\fl{N_2(u,v)} \geq \varrho _2 \label{6 or varrho}
\end{equation}%
for some$\ \tilde{\rho}_1 \leq \rho _1,\ \tilde{\rho}_2\leq \rho _2,$
then we can assume that the solution $(u_1,v_1 )$ from Theorem {\ref{thm:existence 3sols a}} or Theorem {\ref{thm:existence 3sols a'}} satisfies $\fl{u_1} \geq \varrho_1 $ or $\fl{v_1} \geq \varrho _2$ or $|u_1 |>\tilde{\rho}_1 $ or $|v_1 |>\tilde{\rho}_2.$
\end{thm}
\begin{proof}
(i) The assumption impies that $(0,0)$ is not a fixed point.
(ii) The inequality follows from Theorem \ref{thm:existence 2a} applied in the
case of $r_i:=\varrho_i $ and $R_i :=\rho_i.$
(iii) From Theorem \ref{thm:existence or a} applied in the case of $r_i :=\varrho _i,$ $R_i :=\rho _i $ and
\[A=\set{(u,v)}{\fl u < \varrho_1, \fl v < \varrho_2, |u|\leq \tilde\rho_1, |v|\leq \tilde\rho_2},\]
we obtain there are no fixed points of $N$ in $A$, which ends the proof.
\end{proof}
The next Remark illustrates how Theorem \ref{thm:existence 2a} can be used to
prove the existence of more nontrivial solutions.
\begin{rem}\label{rem:many-sols}
If $N$ satisfies the conditions of Theorem \ref{thm:existence 2a} for all pairs
\begin{equation*}
0<r_i^j<\fl{\psi_i}R_i^j \ \text{for }\ i=1,2,\ j=1,2,\ldots,n,\end{equation*}
satisfying
\begin{equation*}
\fl{\psi_i}R_i^j<r_i^{j+1}\ \text{ for }\ i=1,2,\ j=1,2,\ldots,n-1,
\end{equation*}
then $N$ possesses at least $n$ nontrivial solutions $(u_j,v_j)$ with
\begin{equation*}
|u_j |\leq R_1 ^j,\ |v_j |\leq R_2^j,\ \fl{u_j} >
r_1^j,\ \fl{v_j} > r_2^j.
\end{equation*}%
Moreover, if \eqref{eq:ind1'} holds with the strict inequality, i.e. if
\[\sup\limits_{(u,v)\in K,|u|\leq R_1 ^j,|v|\leq R_2^j }|N_i (u,v)|<R_i ^j \quad(i=1,2),
\]
hold, then we have $n-1$ additional solutions $(\bar{u}_j,\bar{v}_j ),$ $%
j=1,\ldots,n-1$ such that
\begin{equation*}
|\bar u_j|<R_1^{j+1},\ |\bar v_j|<R_2^{j+1};\ |\bar u_j|>R_1^j\text{ or }|\bar v_j|>R_2^j;\ \fl{\bar u_j}
<r_1^{j+1}\text{ or }\fl{ \bar v_j} <r_2^{j+1}.
\end{equation*}%
The first conclusion follows from Theorem \ref{thm:existence 2a} applied $n$ times, whereas the second follows from Theorem \ref{thm:existence 3sols a} applied $
n-1$ times.
\end{rem}
\begin{rem}
We stress that the abstract results obtained in this section can be generalized to the case of systems of more than two equations. The idea is to consider the product space $E=\Pi_{i=1}^nE_i$ of the Banach spaces $E_i,$ endowed with the norms $|\cdot|_i,$ functionals $\fl\cdot_i $, and the pairs of cones and wedges $K_i \subset G_i\subset E_i $ such that \eqref{eq:fl monot}, \eqref{eq: prec-max} are satisfied for $i=1,2,...,n.$ In
this setting we are interested in the existence and localization of fixed points of a given operator $N\colon K\rightarrow K,$ where $K=\Pi _{i=1}^n K_i .$ For example, let us consider the sets
\begin{equation*}
C=\mbox{$\left\{u\in K:\;|u_1|_1\leq R_1,\ldots,|u_n|_n\leq R_n\right\}$},\
U=\mbox{$\left\{u\in C:\;\fl{u_1}_1 < r_1,\ldots,\fl{u_n}_n < r_n\right\}$}
\end{equation*}%
for given radii $r_i,R_i>0 $ with $r_i <\fl{\psi_i}_i R_i,\ $ $i=1,...,n.$ If
\begin{equation*}
\sup_{u\in C}|N_i (u)|_i \leq R_i,\ \ i=1,2,...,n
\end{equation*}%
and
\begin{equation*}
\inf_{\substack{ u\in \overline{U} \\ \fl{ u_i}_i =r_i }}%
\fl{N_i (u)}_i> r_i,\quad
i=1,2,...,n,
\end{equation*}%
then $N$ has at least one fixed point in $C\setminus \overline U.$
As a consequence, results analogous to ones obtained later in Section \ref%
{sect:appl}, can be established for systems with more than two differential
equations.
\end{rem}
\section{The system of parabolic equations}\label{sect:appl}
Let $\Omega\subset \mathbb{R}^m$ be a bounded domain that is Dirichlet regular (i.e. for all $\varphi\in C(\partial \Omega)$ there exists $u\in C(\overline\Omega)$ such that $u|\partial\Omega=\varphi$ and $\Delta u=0$ in a distributional sense, see \cite[Definition 6.1.1]{b:ABHN}).
This class of domains is rather large; for example, if the boundary of $\Omega$ is Lipschitz continuous, then $\Omega$ is Dirichlet regular (see \cite[Chapter II, Section 4, Proposition 4]{b:DL}).
Let us take
\[E=C_0(\Omega)=\set{u\in C(\overline{\Omega})}{u|\partial\Omega=0},\ E_+=\set{u\in E}{u(x)\geq 0\text{ for all }x\in\Omega}.\]
Let us also consider the space ${\mathcal E}=C(0,{t_{\text{max}}},E)$, ${t_{\text{max}}}>0$ and its cone of nonnegative functions ${\mathcal E}_+=C(0,{t_{\text{max}}},E_+)$. The spaces $E$ and ${\mathcal E}$ are endowed with the uniform norms, that is
\[|u|=\max\set{|u(x)|}{x\in \overline\Omega},\ u\in E, \] and
\[|u|=\max\set{|u(t)|}{0\leq t\leq {t_{\text{max}}}},\ u\in {\mathcal E}.\]
We shall discuss the parabolic system
\begin{equation}\label{eq:parabolic}
\begin{cases}
u_t-\Delta u = f(t,x,u,v) & (t,x)\in Q:=(0,{t_{\text{max}}})\times\Omega,\\
v_t-\Delta v = g(t,x,u,v) & (t,x)\in Q:=(0,{t_{\text{max}}})\times\Omega,\\
u(t,x)=v(t,x)=0 & (t,x)\in (0,{t_{\text{max}}})\times\partial\Omega,\\
u(0,\cdot)=\alpha(u,v),\\
v(0,\cdot)=\beta(u,v).
\end{cases}
\end{equation}
Here, $f,g\colon (0,{t_{\text{max}}})\times\Omega\times\mathbb{R}_+\times\mathbb{R}_+\to\mathbb{R}_+$ and $\alpha,\beta\colon{\mathcal E}\times{\mathcal E}\to E$ are continuous functions. In what follows we shall identify $u\colon[0,{t_{\text{max}}}]\times\overline\Omega\to\mathbb{R}$ with $u\colon [0,{t_{\text{max}}} ]\to C(\overline\Omega)$ via the formula $u(t)(x)=u(t,x)$.
We shall treat $\Delta$ as the operator defined on the domain $D(\Delta) =\set{u\in E}{\Delta u\in E}$.
\begin{lem} The operator $\Delta$ is a generator of an analytic (immediately) compact $C_0$-semigroup of contractions $\set{S(t)}{t\geq 0}$ on $E$. Moreover, the operators $S(t)$ are positive, i.e. $S(t)u\geq 0$ for $u\geq 0$.
\end{lem}
\begin{proof} From \cite[Theorem 6.1.8]{b:ABHN} we conclude that the semigroup $S$ is positive and contractive. From \cite[Theorem 6.1.9]{b:ABHN} we see that $S$ is analytic. According to Theorems 4.29 and 4.25 of \cite{b:EnNag}, to demonstrate the compactness of $S$ it remains to prove the compactness of the canonical injection $i\colon E_\Delta\hookrightarrow E$, where $E_\Delta=(D(\Delta),\|\cdot\|_\Delta)$ is equipped with the graph norm $\|\cdot\|_\Delta$.
Let us consider the Fr\'echet space $F=E\cap C^1(\Omega)$ equipped with the seminorms $p_n$,
\[p_0(u)=\max\set{|u(x)|}{x\in\Omega},\ p_n(u)=\max\set{|\nabla u(x)|}{x\in\overline{\Omega_n}},\ n\geq 1,\] where $(\Omega_n)$ is an open covering of $\Omega$ satisfying $\overline{\Omega_n}\subset \Omega$, $n\in{\mathbb N}$.
By \cite[Lemma 6.1.5]{b:ABHN} we know that $D(\Delta)\subset F$. The Closed Graph Theorem yields the embedding $E_\Delta \hookrightarrow F$ is continuous. The conclusion is implied by the compactness of $F \hookrightarrow E$.
\end{proof}
\begin{rem}\label{rem:choiceC0} It seems worth discussing the choice of the space $C_0(\Omega)$. In the recent paper \cite{b:gi-mm-rp}, where an elliptic system was discussed, the space $L^\infty(\Omega)$ was considered. Unfortunately, the Laplacian $\Delta$ fails to generate a $C_0$-semigroup on $L^\infty(\Omega)$. Moreover, although $\Delta$ generates semigroups on $L^p(\Omega)$, these spaces are somewhat inappropriate to obtain the localization of solutions with our approach. The fact that the set $\set{u\in E_+}{|u|\leq 1}$ does not possess a largest element (with respect to the natural order $\leq$) prevented us from using the abstract setting from \cite{b:gi-mm-rp}; for details, see Remark \ref{rem:nopsi}.
\end{rem}
We shall also consider the space $H=L^2(\Omega)$ and the Laplacian $\Delta_2$ on $H$ with Dirichlet boundary condition. Denote by $S_2\colon [0,\infty)\to B(H)$ the semigroup generated by $\Delta_2$ and by $i\colon E\to H$ the natural embedding.
\begin{prop}\label{prop:iS=S2i} $i(S(t)u)=S_2(t)i(u)$ for all $u\in E$.
\end{prop}
\begin{proof}
Since $D(\Delta_2)=\set{u\in H}{\Delta u\in H}$, $\Delta$ is the part of $\Delta_2$ in $E$ (see \cite[Paragraph II.2.3]{b:EnNag}) and, as a consequence, $\Delta_2\circ i=i\circ \Delta$. This implies that $R(\lambda,\Delta_2)\circ i=i\circ R(\lambda,\Delta)$ for $\lambda>0$. Using the Post-Widder inversion formula for $C_0$-semigroups (see Corollary 3.3.6 in \cite{b:ABHN}) we obtain for $x\in E$ and $t\geq 0$
\[i(S(t)x)=\lim_{n\to\infty}i\left(\left[\frac ntR\left(\frac nt,\Delta\right)\right]^nx\right)
=\lim_{n\to\infty}\left[\frac ntR\left(\frac nt,\Delta_2\right)\right]^ni(x)=S_2(t)i(x).\]
\end{proof}
Define \[F(u,v)(t)(x)=f(t,x,u(t)(x),v(t)(x)),\quad G(u,v)(t)(x)=g(t,x,u(t)(x),v(t)(x))).\] Under the following assumption:
\begin{equation}\label{eq:assumpt fg}
f(t,x,0,0)=g(t,x,0,0)=0\ \text{ for }x\in\partial\Omega,
\end{equation}
the operators $F,G\colon{\mathcal E}\times{\mathcal E}\to{\mathcal E}$ are continuous and bounded (map bounded sets into bounded ones).
\begin{defin}\label{def:mild-solution}
For $\xi\in E$, and $f\in {\mathcal E}$, we say that the function \[u(t)=p(\xi,f)(t):=S(t)\xi+\int_0^t S(t-s)f(s)ds\] is a {\em mild solution} of the problem $u'-\Delta u=f$ on $(0,{t_{\text{max}}})\times\Omega$ with $u(0)=\xi$.
\end{defin}
Therefore, the mild solution of the problem \eqref{eq:parabolic} is a fixed point of the vector valued operator $M$ defined as
\[M_1(u,v)=\bar S(\alpha(u,v))+\hat S(F(u,v)),\ M_2(u,v)=\bar S(\beta(u,v))+\hat S(G(u,v)),\]
where
\[\bar S\colon E\to {\mathcal E},\ \bar S(u_0)(t)=S(t)u_0,\ u_0\in E,\]
\[\hat S\colon {\mathcal E}\to{\mathcal E},\ \hat S(f)(t)=\int_0^tS(t-\tau)f(\tau)d\tau, f\in{\mathcal E}.\]
It is worth pointing out the following regularity of the mild solutions.
\begin{prop}\label{prop:strong sol}
Let $u=p(\xi,f)$ for $\xi\in E$, $f\in{\mathcal E}$, that is $u$ is a mild solution of the problem $u'-\Delta u=f$, $u(0)=\xi$. Then
\begin{enumerate}[\upshape (a)]
\item $u$ is a strong solution of that problem in the space $L^2(\Omega)$. Precisely, $u\colon [0,{t_{\text{max}}} ]\to L^2(\Omega)$ is absolutely continuous, $u'\in L^1(0,{t_{\text{max}}},L^2(\Omega))$, $u(t)\in D(\Delta_2)$ and $u'(t)=\Delta_2 u(t)+f(t)$ for almost all $t\in(0,{t_{\text{max}}} )$.
\item $u$ is a weak solution of the equation $u'-\Delta_2u=f$ in the sense that the weak spatial derivative $\nabla u(t,x)$ and weak time derivative $u_t(t,x)$ exist on $(0,{t_{\text{max}}})\times\Omega$ and
\begin{equation}\label{eq:weak solution}\int_0^{t_0}\int_\Omega\nabla u(t,x)\nabla \phi(t,x)-u\frac{\partial}{\partial t}\phi(t,x)dxdt=\int_0^{t_0}\int_\Omega f(t)\phi(t,x)dt\end{equation}
for all $\phi\in C^\infty_0((0,{t_{\text{max}}} )\times\Omega)$.
\end{enumerate}
\end{prop}
\begin{proof}
(a) Let $u$ be as in the statement.
\[\bar u(t)=S_2(t)\bar \xi+\int_0^t S_2(t-s)\bar f(s)ds,\text{ where }\bar u=i\circ u,\ \bar \xi=i(\xi),\ \bar f=i\circ f.\] The conclusion follows from \cite[Theorem 8.2.1]{b:Vrabie}, as $H=L^2(\Omega)$ is a Hilbert space.
(b)
We now prove the following statement: The weak spatial derivative $\nabla u(t,x)$ exists on $(0,{t_{\text{max}}})\times\Omega$ and belongs to $L^2((0,{t_{\text{max}}})\times \Omega)$.
Because $u(t)\in D(\Delta_2)\subset W^{1,2}_0(\Omega)$ for $t\in (0,{t_{\text{max}}} )$, the weak (spatial) derivative $\nabla u(t)$ exists for almost all $t\in (0,{t_{\text{max}}} )$. Therefore we have
\begin{equation}\label{eq:diff_x}\int_0^{t_0}\int_\Omega u(t,x)\nabla\phi(t,x)dxdt=-\int_0^{t_0}\int_\Omega \nabla (u(t))(x)\phi(t,x)dxdt,\end{equation}
for all $\phi\in C^\infty_0((0,{t_{\text{max}}})\times\Omega)$, as $\phi(t,\cdot)\in C^\infty_0(\Omega)$ for all $t\in (0,{t_{\text{max}}} )$. Thus, the weak spatial derivative $\nabla u(t,x)$ exists and is equal to $\nabla (u(t))(x)$.
Moreover,
\begin{equation}\label{eq:estim1}\|\nabla u\|^2_{L^2((0,{t_{\text{max}}})\times\Omega)} =\int_0^{t_0}\int_\Omega|\nabla u(t,x)|^2dxdt = \int_0^{t_0}\|\nabla (u(t))\|_H^2dt.\end{equation}
By \cite[Theorem 8.2.1]{b:Vrabie} we know that the function
\[t\mapsto -\<\Delta_2u(t),u(t)>_H=
\int_\Omega |(\nabla u(t))(x)|^2dx=\|\nabla (u(t))\|_H^2\] belongs to $L^1(0,{t_{\text{max}}} )$. This and \eqref{eq:estim1} prove that
$\|\nabla u\|^2_{L^2((0,{t_{\text{max}}})\times\Omega)}<\infty$.
We now prove that the weak time derivative $u_t(t,x)$ exists on $(0,{t_{\text{max}}})\times\Omega$.
Let us consider the function $\phi\in C^\infty_0((0,{t_{\text{max}}})\times\Omega)$. To avoid disambiguity let us introduce the function $\varphi\colon (0,{t_{\text{max}}} )\to C_0(\Omega)$ defined by the formula $\varphi(t)=\phi(\cdot,t)$. One can easily check that $\varphi\in C^1(0,{t_{\text{max}}},C_0(\Omega))$ and $\varphi'(t)=\phi_t(\cdot,t)$.
By the absolute continuity of $u,\varphi\colon(0,{t_{\text{max}}} )\to H$ we obtain $d(t)=\<u(t),\varphi(t)>_H$ is absolutely continuous and $d'(t)=\<u'(t),\varphi(t)>_H+\<u(t),\varphi'(t)>_H$. In particular,
\begin{equation}\label{eq:d-ac}0=d(t_0)-d(0)=\int_0^{t_0}d'(t)dt=\int_0^{t_0}\<u'(t),\varphi(t)>_H+\<u(t),\varphi'(t)>_Hdt.\end{equation}
Since $u'\in L^1(0,{t_{\text{max}}},H)$ and $u,\varphi,\varphi'\in L^\infty(0,{t_{\text{max}}},H)$, both functions $\<u'(t),\varphi(t)>_H$, $\<u(t),\varphi'(t)>_H$ are integrable. Thus, we can restate \eqref{eq:d-ac} in the following manner:
\begin{equation}\label{eq:diff_t} \int_0^{t_0}\int_\Omega u(t,x)\phi_t(t,x)dxdt=-\int_0^{t_0}\int_\Omega (u'(t))(x)\phi(t,x)dxdt.\end{equation}
This proves that the weak derivative $\frac{\partial}{\partial t}u(t,x)$ exists and is equal to $(u'(t))(x)$.
Since $u(t)\in D(\Delta_2)$ for a.a. $t\in(0,{t_{\text{max}}} )$ and $\nabla u(t,x)=\nabla (u(t))(x)$, we have
\begin{equation}\label{eq:diff_xx}\int_0^{t_0}\int_\Omega \Delta_2 u(t)(x)\phi(t,x)dxdt= -\int_0^{t_0}\int_\Omega \nabla u(t,x)\nabla\phi(t,x)dxdt.\end{equation}
Let $\phi\in C^\infty_0((0,{t_{\text{max}}})\times\Omega)$. Combining the equations $\eqref{eq:diff_xx}$, $\eqref{eq:diff_t}$ and $u'(t)=\Delta_2u(t)+f(t)$ for almost all $t\in (0,{t_{\text{max}}} )$ (see (a)) we obtain \eqref{eq:weak solution}.
\end{proof}
\begin{prop}\label{prop:M-compact} The operators $M_1,M_2$ are completely continuous if and only if $\alpha$ and $\beta$ are completely continuous.
\end{prop}
\begin{proof}
If $M_1$, $M_2$ are completely continuous, then $e_0\circ M_1=\alpha, e_0\circ M_2=\beta$ are completely continuous, where $e_0(f)=f(0)$ for $f\in{\mathcal E}$.
Now, assume that the operators $\alpha,\beta$ are completely continuous. Since the operator $\bar S$ is continuous, then the operator $\bar S\circ \alpha$ is completely continuous. We shall demonstrate that $\hat S$ is completely continuous. In order to do this we utilize a version of the Ascoli-Arzel\`{a} Theorem tailored for the space ${\mathcal E}$, see for example Theorem A.2.1 of \cite{b:Vrabie}.
First, denote by $\Xi$ the upper bound of the norms of $\|S(t)\|_{B(E)}$ for $t\in[0,{t_{\text{max}}} ]$. Because $S$ is immediate norm continuous, for any $\varepsilon>0$ there exists a number $\mu(\varepsilon)>0$ such that $\|S(t)-S(s)\|_{B(E)}<\varepsilon$ if $0<\varepsilon\leq t\leq s\leq t+\mu(\varepsilon)$.
Fix $R,\varepsilon>0$ and let $D=\set{f\in {\mathcal E}}{|f|<R}$. For a fixed $t>0$ and $f\in D$ let $\eta=\min\{t,\varepsilon\}$. We can present $\hat S(f)(t)$ as a sum $\hat S(t)=x+y$, where
\[x=S(\eta)\int_0^{t-\eta}S(t-s-\eta)f(s)ds,\quad y=\int_{t-\eta}^tS(t-s)f(s)ds.\]
It is straightforward to show that $x\in \hat C:=S(\eta)(B(0,t\Xi R))$ and that $y\in B(0,\varepsilon \Xi R)$. Because $\hat C$ is relatively compact and $\varepsilon$ is arbitrary, we obtain the set $\set{\hat S(f)(t)}{f\in D}$ is relatively compact for all $t\in[0,{t_{\text{max}}} ]$.
Now we shall prove the equicontinuity of the family $\set{\hat S(f)}{f\in D}$. In order to do it let us fix $\varepsilon>0$, $t,s\in [0,{t_{\text{max}}} ]$ and $f\in D$. Without loss of generality we can assume that $t\leq s$. Let us put $\eta=\min\{t,\varepsilon\}$ (the case $\eta=0$ appears if $t=0$). Then $\hat S(f)(s)-\hat S(f)(t)=x+y$, where
\[x=\int_0^{t}(S(s-\tau)f(\tau)-S(t-\tau)f(\tau))d\tau,\ y=\int_t^sS(s-\tau)f(\tau)d\tau.\]
\[ |x| \leq R\int_0^t\|S(s-t+\tau)-S(\tau)\|_{B(E)}d\tau\leq 2\Xi R\varepsilon+\int_\eta^t\|S(s-t+\tau)-S(\tau)\|_{B(E)}d\tau.\]
\[ |y | \leq |s-t|\Xi R.\]
Therefore $|\hat S(f)(s)-\hat S(f)(t)|\leq (3\Xi R+{t_{\text{max}}})\varepsilon$ if $|s-t|\leq \min\{\mu(\varepsilon),\varepsilon\}$.
This proves the uniform equicontinuity of the family $\set{\hat S(f)}{f\in D}$ and finishes the proof of the complete continuity of $\hat S$. Now, the complete continuity of $M_1=\bar S\circ \alpha+\hat S\circ F$ is clear. Similarly we can prove the complete continuity of $M_2$.
\end{proof}
In order to use fixed point index for compact operators and, at the same time, to avoid assuming the compactness of $\alpha$ and $\beta$, we consider the operator $N=(N_1,N_2)$, defined by the formula
\[N_1(u,v)=\bar S\left(\alpha\left(\bar S(u(0))+\hat S(F(u,v))\:,\: \bar S(v(0))+\hat S(G(u,v))\right)\right)+\hat S(F(u,v)),\]
\[N_2(u,v)=\bar S\left(\beta\left(\bar S(u(0))+\hat S(F(u,v))\:,\:\bar S(v(0))+\hat S(G(u,v))\right)\right)+\hat S(G(u,v)).\]
\begin{prop}\label{prop:fixed-sets} The sets of fixed points of the operators $M$ and $N$ coincide.
\end{prop}
\begin{proof} Note that
\begin{equation}\label{eq:t1}
N_1(u,v)=\bar S(u_0)+\hat S(F(u,v))\text{ and }N_2(u,v)=\bar S(v_0)+\hat S(G(u,v)),
\end{equation} where
$u_0=\alpha(\bar u,\bar v)$, $v_0=\beta(\bar u,\bar v)$ and
\begin{equation}\label{eq:t2}
\bar u=\bar S(u(0))+\hat S(F(u,v)),\quad \bar v=\bar S(v(0))+\hat S(G(u,v)).
\end{equation}
From \eqref{eq:t1} and the properties of $\bar S$ and $\hat S$ we have $N_1(u,v)(0)=u_0$ and $N_2(u,v)(0)=v_0$. Therefore, if $N(u,v)=(u,v)$, then
$\bar u=u$ and $\bar v=v$ and, consequently, $u_0=\alpha(u,v)$ and $v_0=\beta(u,v)$. By \eqref{eq:t1} we arrive at $M(u,v)=(u,v)$.
Conversely, if $M(u,v)=(u,v)$, then
\begin{equation}\label{eq:t3}u=\bar S(\alpha(u,v))+\hat S(F(u,v)),\ v=\bar S(\beta(u,v))+\hat S(G(u,v))\end{equation} and $u(0)=\alpha(u,v)$, $v(0)=\beta(u,v)$. Therefore we have
\[u=\bar S(u(0))+\hat S(F(u,v))\text{ and }v=\bar S(v(0))+\hat S(F(u,v)).\] Plugging this into \eqref{eq:t3} we obtain $N(u,v)=(u,v)$.
\end{proof}
From the proof it follows in particular, that
\begin{equation}\label{eq:N,N(0)}
\begin{array}{c}
N_1(u,v)=\bar S(N_1(u,v)(0))+\hat S(F(u,v)),\text{ and }\\ N_2(u,v)=\bar S(N_2(u,v)(0))+\hat S(G(u,v)),
\end{array}
\end{equation}
which will be used later.\par
From Proposition~\ref{prop:M-compact} we know that a necessary condition for the operator $M$ to be completely continuous is the complete continuity of $\alpha$ and $\beta$. In the case of the operator $N$ we can weaken the assumptions on $\alpha$ and $\beta$.
\begin{prop}\label{prop:N-compact} The operator $N$ is completely continuous if the images $\alpha(U_1\times U_2)$ and $\beta(U_1\times U_2)$ are relatively compact for all bounded sets $U_1,U_2\subset {\mathcal E}$ that satisfy the following property:
\begin{equation}\label{eq:almost-compact-sets}
\text{For all}\ \varepsilon>0\ \text{the set}\ \set{u(t)}{u\in U_i,\ t\in[\varepsilon,{t_{\text{max}}}]}\ \text{is relatively compact},\ i=1,2.
\end{equation}
\end{prop}
\begin{proof} We prove, without loss of generality, the complete continuity of $N_1$. Let $\bar\alpha(u,v)=\alpha(\bar u,\bar v)$, where $\bar u(u,v)$ and $\bar v(u,v)$ are defined by \eqref{eq:t2}. From \eqref{eq:t1} it follows that $N_1(u,v)=\bar S(\bar\alpha(u,v))+\hat S(F(u,v))$. By Proposition \ref{prop:M-compact}, it suffices to show that $\bar\alpha$ is completely continuous. This will be done if we demonstrate, that $\bar u(U\times U)$ and $\bar v(U\times U)$ satisfy \eqref{eq:almost-compact-sets} for a given bounded set $U\subset {\mathcal E}$.
Let $\varepsilon>0$ be given. Put $R=\sup\set{|u|}{u\in U}$. For $u,v\in U$ and $t\geq \varepsilon$ we obtain
\[\bar S(u(0))(t)=S(\varepsilon)S(t-\varepsilon)u(0)\in S(\varepsilon)D(0,R)=:{\mathcal C}\] and the set ${\mathcal C}$ is relatively compact.
Moreover, the proof of Proposition \ref{prop:M-compact} shows that the set $\hat S( F(U\times U))$ is relatively compact in $\mathcal E$. By the standard arguments, utilizing the compactness of $[0,{t_{\text{max}}}]$, one can show that the set \[\set{\hat S(F(u,v))(t)}{u,v\in U,\ t\in[0,{t_{\text{max}}}]}\] is totally bounded (and therefore relatively compact) in $E$. This shows that $\bar u(U\times U)$ satisfies the condition \eqref{eq:almost-compact-sets}. Similarly we can verify this condition for the set $\bar v(U\times U)$. \end{proof}
\begin{exa}\label{ex:alpha=u(t_0)} Let $\alpha(u,v)=u(t_0)$, $\beta(u,v)=v(t_0)$, where $0\leq t_0\leq {t_{\text{max}}}$. Then $\alpha,\beta$ satisfy the condition from Proposition \ref{prop:N-compact} if and only if $t_0>0$. Indeed, let $t_0=0$ and $U_1=U_2=\set{S(\cdot)u}{|u|\leq 1}$. Then the condition \eqref{eq:almost-compact-sets} is satisfied, but the set
\[\alpha(U_1\times U_2)=\set{u(0)}{u\in U_1}=\set{u\in E}{|u|\leq 1}\] is not compact.
Conversely, if $t_0>0$ and sets $U_1,U_2$ satisfy the condition \eqref{eq:almost-compact-sets}, then $\alpha(U_1\times U_2)=\set{u(t_0)}{u\in U_1}$ is compact from \eqref{eq:almost-compact-sets}.
Note that the reasoning above can be applied to the case of \emph{multi-point} conditions of the type
\[\alpha(u,v)=\sum_{s=1}^{k} \alpha_su(t_s),\quad \beta(u,v)=\sum_{s=1}^{r} \beta_sv(t'_s),\] where $0<t_1<\ldots<t_k$, $0<t'_1<\ldots<t'_r$ and $\alpha_s,\beta_s>0$.
\end{exa}
\begin{exa}\label{ex:alpha-integral}
Let
\[\alpha(u,v)=G_1\left(\int_0^{{t_{\text{max}}}} g_1(u(t),v(t))dt\right),\quad \beta(u,v)=G_2\left(\int_0^{{t_{\text{max}}}} g_2(u(t),v(t))dt\right),\]
where $g_1,g_2\colon\mathbb{R}_+^2\to\mathbb{R}_+$, $G_1,G_2\colon\mathbb{R}_+\to\mathbb{R}_+$ with $g_i(0,0)=G_i(0)=0$ for $i\in\{1,2\}$. Then $\alpha,\beta$ satisfy the condition from Proposition \ref{prop:N-compact}.
Indeed, let us consider the sets $U_1,U_2\subset B(0,R)$ satisfying the condition \eqref{eq:almost-compact-sets} and let $\varepsilon>0$. From the uniform continuity of $G_i$ on $[0,{t_{\text{max}}}\cdot d]$, where
\[d=\max\set{g_i(u,v)}{0\leq u,v\leq R,\ i=1,2},\] there exists $\delta>0$ such that $|G(p)-G(q)|< \varepsilon$ if $0\leq p,q\leq {t_{\text{max}}}\cdot d$, $|p-q|<\delta\cdot d$.
Put
\[U_i^\delta := \set{u(t)}{u\in U_i,\ t\in[\delta,{t_{\text{max}}}]},\ \delta>0,\ i=1,2\] and
\[\Gamma_i^\delta := \overline{\mathrm{conv}}\: g_i\!\left(U_1^\delta\times U_2^\delta\right).\] From the Mazur Theorem, which states that the closed convex hull of a compact subset in a Banach space is compact, the sets $\Gamma_i^\delta$ are compact. Therefore,
\[\int_\delta^{t_{\text{max}}} g_i(u(t),v(t))dt\in W_i^\delta:=({t_{\text{max}}}-\delta)\cdot \Gamma_i^\delta\ \text{ and }\ \left|\int_0^\delta g_i(u(t),v(t))dt\right|\leq \delta\cdot d\] for $u\in U_1$, $v\in U_2$. Thus, by the choice of $\delta$, we deduce that for all $u\in U_1$, $v\in U_2$ we have
$\alpha(u,v)=u_1+u_2$, where $u_1\in G_1\!\left(W_i^\delta\right)$ and $|u_2|<\varepsilon$. Because $G_1\!\left(W_i^\delta\right)$ is compact and $\varepsilon$ is arbitrarily small, we obtain $\alpha(U_1\times U_2)$ is relatively compact. Similarly we can proceed with $\beta$.
\end{exa}
Let $D\subset\subset \Omega$ be any open subset. Put
\[{G} = \set{u\in {\mathcal E}}{u(0,x)\geq 0\ \text{ for all }x\in D}.\]
The set $G$ is a wedge generating the semiorder $\preceq$. By $\leq$ we denote the order induced by the cone ${\mathcal E}_+$. The symbol $\leq$ will also be used to denote the order on $E$ induced by $E_+$ and the natural order on $H=L^2(\Omega)$ (that is $u\leq v$ if $u(x)\leq v(x)$ for a.a. $x\in\Omega$).
Given a function $u\in H$ we set
$$
\fl u=\einf_{x\in D}|u(x)|,
$$
(in particular
$\fl u=\inf_{x\in D} |u(x)|$ for $u\in E$) and futhermore, with abuse of notation, by the same symbol we denote the value $\fl u=\fl{u(0)}$ for $u\in{\mathcal E}$.
The following monotonicity and continuity conditions of the functional $\fl\cdot$ are satisfied:
\[\fl u\leq \fl v\text{ if }u,v\in{\mathcal E}_+\text{ and }u\preceq v\] and
\[|\fl u-\fl v|\leq |u-v|\text{ for all }u,v\in{\mathcal E} \quad\left(\text{therefore} \fl u\leq |u|\text{ for all }u\in{\mathcal E}\right).\]
Consider the function $\psi(t)=\varphi$ for all $t\in[0,{t_{\text{max}}} ]$, where $\varphi\in E$ satisfies the following conditions: $\varphi|D\equiv 1$ and $0\leq \varphi\leq 1$ in $\Omega$. Then $|\psi|=1$ and $u\preceq|u|\psi$ for every $u\in{\mathcal E}_+$.
\begin{rem}\label{rem:nopsi}
Note that in the space ${\mathcal E}=C(0,{t_{\text{max}}},C_0(\Omega))$ there is no element $\bar\psi$ that $|\bar\psi|=1$ and $u\leq|u|\bar\psi$ for every $u\in{\mathcal E}_+$. This is the main reason for considering the wedges $G_i$, $i\in\{1,2\}$, and the semiorders $\preceq$ in Section \ref{sect:abstr}.
\end{rem}
Define the cones
\[K_1=K_2=\set{u\in {\mathcal E}_+}{u(t)\geq S(t)u(0) \ \text{for all}\ t\in[0,{t_{\text{max}}}]}.\]
We see that \eqref{eq:N,N(0)} yields $N({\mathcal E}_+\times{\mathcal E}_+)\subset K_1\times K_2=K$.
\begin{prop}\label{prop:properties u}
Assume that $u_0\in E$ is a nonnegative nonzero function. Define $u(t,x)=(S(t)u_0)(x)$. Then
\begin{enumerate}[\upshape (i)]
\item $u\in C^\infty((0,\infty)\times\Omega)\cap C(\mathbb{R}_+\times \overline\Omega)$
\item $u_t=\Delta u$ on $(0,\infty)\times\Omega$
\item $u(0,x)=u_0(x)$ for $x\in\Omega$
\item $u(t,x)>0$ for $t>0$ and $x\in\Omega$.
\end{enumerate}
\end{prop}
\begin{proof}
The conclusions (i)-(iii) follow from Proposition 2.6 of \cite{b:ABHN}. To prove (iv) let us fix $x\in\Omega$ and $t>0$. Let us consider the domain $V\subset\subset \Omega$ such that $x,x_0\in V$, where $x_0\in\Omega$ is such that $u_0(x_0)>0$. From (i) it follows that $\sup_Vu(\varepsilon,\cdot)>0$ for $\varepsilon<t$ sufficiently small. From (i) and (ii) it follows that $u$ solves $u_t=\Delta u$ on $(0,\infty)\times V$. Moreover, $u\geq 0$, which follows from the positiveness of the semigroup $S$.
We shall use the \emph{parabolic Harnack inequality} (see for example Theorem 7.1.10 of \cite{b:Evans}) expressed in the following manner:
{\em Let $u\in C^2((0,\infty)\times \Omega)$ with $u_t=\Delta u$ and $u\geq 0$ in $(0,\infty)\times \Omega$. Let $V\subset\subset \Omega$ be connected. Then, for each $0<t_1<t_2$, there exists a constant $C$ such that $$\sup_Vu(t_1,\cdot)\leq C\inf_Vu(t_2,\cdot).$$}
From the parabolic Harnack inequality we have $\inf_Vu(t,\cdot)>0$. Therefore we obtain $u(t,x)>0$.
\end{proof}
\begin{lem}\label{lem:m_eta>0} Let $0\neq \eta\in E=C_0(\Omega)$.
\begin{enumerate}[\upshape (a)]
\item If $0<t_0<t_1$, then
\[m_\eta(t_0,t_1):=\min\set{(S(\tau)\eta)(x)}{x\in \overline D,\ \tau\in[t_0,t_1]}>0.\]
\item If $0\leq t_0<t_1$, then \[\fl{\int_{t_0}^{t_1}S(\tau)\eta d\tau}>0.\]
\end{enumerate}
\end{lem}
\begin{proof}
(a) By Proposition \ref{prop:properties u}(i,iv) and compactness of $[t_0,t_1]\times \overline D$ we obtain $m_\eta(t_0,t_1)>0$.
(b) We can assume that $t_0>0$. Then
\[\int_{t_0}^{t_1}(S(\tau)\eta)(x) d\tau\geq (t_1-t_0)m_\eta(t_0,t_1)>0\] for $x\in D$, which is the desired conclusion.
\end{proof}
Define
\begin{equation}\label{eq:def m} m(t_0,t_1):=\einf_{x\in D,\ t\in[t_0,t_1]}\left(S_2(t)\chi_D\right)(x).\end{equation}
\begin{cor}\label{cor:m>0}
\begin{enumerate}[\upshape (a)]
\item If $0<t_0<t_1$, then $m(t_0,t_1)>0$.
\item If $0\leq t_0\leq t_1$, then
\[\fl{\int_{t_0}^{t_1}S_2(\tau)\chi_Dd\tau}>0.\]
\end{enumerate}
\end{cor}
It is possible that $m(0,t_1)>0$ (see Example \ref{ex:[0,pi]}).
\begin{proof}
Let $\eta\in E=C_0(\Omega)$ be any nonzero function with $0\leq \eta\leq \chi_D$. The conclusion is implied by Lemma \ref{lem:m_eta>0} and the inequality
$S_2(t)\chi_D\geq S_2(t)\eta= S(t)\eta,$ which follows from the positiveness of $S_2$ and Proposition \ref{prop:iS=S2i}.
\end{proof}
Let us fix $0\leq t_0<t_1\leq{t_{\text{max}}}$ and put $m:=m(t_0,t_1)$.
\begin{rem}\label{rem:Harnack}
Take $u\in K$ and set $r:=\fl u=\inf_{x\in D}u(0,x)$. Observe that
$u(0)\geq r\chi_D$ and therefore
\[u(t)\geq S(t)u(0)=S_2(t)u(0)\geq rS_2(t)\chi_D\geq rm\chi_D\ \text{ for }t\in[t_0,t_1].\]
As a consequence, we obtain the estimate
\begin{equation}\label{eq:Harnack}
u\geq m\fl u\chi_{[t_0,t_1]\times D}\ \text{ for all }u\in K,
\end{equation}
which can be called \textit{weak Harnack-type inequality}, a counterpart of the inequality (3.4) of \cite{b:gi-mm-rp}.
\end{rem}
Let $\alpha,\beta$ be as in Example \ref{ex:alpha-integral}. Consider the assumption
\begin{equation}\label{eq:gG}
p_1u\leq g_1(u,v)\leq q_1u,\ p_2v\leq g_2(u,v)\leq q_2v,\ P_iu\leq G_i(u)\leq Q_i u,
\end{equation}
for all $u,v\geq 0$, $i\in\{1,2\}$, where $0<p_i\leq q_i$, $0<P_i\leq Q_i$, $i\in\{1,2\}$. Similar assumptions were used, for nonlinear nonlocal conditions in the context of ODEs, in \cite{b:gi-caa, b:gipp-nonlin}.
Under the assumption \eqref{eq:gG} we have the following estimates:
\begin{equation}\label{eq:alfa,beta-estimates}
\begin{split}
p_1P_1 \int_0^{{t_{\text{max}}}} u(t)dt \leq \alpha(u,v)\leq q_1Q_1\int_0^{{t_{\text{max}}}} u(t)dt,\\ p_2P_2\int_0^{{t_{\text{max}}}} v(t)dt\leq \beta(u,v)\leq q_2Q_2\int_0^{{t_{\text{max}}}} v(t)dt.
\end{split}\end{equation}
Put
\begin{equation}\label{eq:inf-fg}
f_{r,R}^0:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ mr_1\leq u\leq R_1,\ 0\leq v\leq R_2}} \frac{f(t,x,u,v)}{r_1},\quad
g_{r,R}^0:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ 0\leq u\leq R_1,\ mr_2\leq v\leq R_2}} \frac{g(t,x,u,v)}{r_2},
\end{equation}
\begin{equation}\label{eq:inf'-fg}
f_{r,R}:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ mr_1\leq u\leq R_1,\ mr_2\leq v\leq R_2}} \frac{f(t,x,u,v)}{r_1},\quad
g_{r,R}:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ mr_1\leq u\leq R_1,\ mr_2\leq v\leq R_2}} \frac{g(t,x,u,v)}{r_2}
\end{equation}
and
\begin{equation}\label{eq:sup-fg}
f^{R}:=\sup\limits_{\substack{t\in[0,{t_{\text{max}}}],\ x\in\Omega \\ 0\leq u\leq R_1,\ 0\leq v\leq R_2}}\frac{f(t,x,u,v)}{R_1},\quad
g^{R}:=\sup\limits_{\substack{t\in[0,{t_{\text{max}}}],\ x\in\Omega \\ 0\leq u\leq R_1,\ 0\leq v\leq R_2}}\frac{g(t,x,u,v)}{R_2}.
\end{equation}
Evidently, $f_{r,R}^0\leq f_{r,R}\leq f^{R}$ and $g_{r,R}^0\leq g_{r,R}\leq g^{R}$.
\begin{lem}\label{lem:estimates} Let $\alpha,\beta$ be as in Example \ref{ex:alpha-integral}. Assume that the inequalities \eqref{eq:gG} are satisfied.
\begin{enumerate}[\upshape (a)]
\item If $|u|\leq R_1$, $|v|\leq R_2$, then \[
\frac{|N_1(u,v)|}{R_1}\leq q_1Q_1C_1+ f^R (q_1Q_1C_2 + C_1),\quad
\frac{|N_2(u,v)|}{R_2}\leq q_2Q_2C_1+ g^R (q_2Q_2C_2 + C_1),
\]
where the constants
\begin{equation}\label{eq:C1C2}C_1=\left|\int_0^{t_{\text{max}}} S_2(\tau)\chi_\Omega d\tau\right|,\quad C_2=\left|\int_0^{t_{\text{max}}}\int_0^t S_2(\tau)\chi_\Omega d\tau dt\right|\end{equation} are positive.
\item The following implications hold:
\begin{align*}
\fl u\geq r_1\implies \frac{\fl{N_1(u,v)}}{r_1}\geq p_1P_1(c_1+f_{r,R}^0c_2),\\ \fl v\geq r_2\implies \frac{\fl{N_2(u,v)}}{r_2}\geq p_2P_2(c_1+g_{r,R}^0c_2),\end{align*}
where the constants
\begin{equation}\label{eq:c1c2}c_1= \fl{\int_0^{t_{\text{max}}} S_2(\tau)\chi_D d\tau},\ c_2=\fl{ \int_{t_0}^{t_{\text{max}}}\int_{t_0}^{\min\{t,t_1\}}S_2(t-\tau)\chi_D d\tau dt}\end{equation} are positive.
\item If $\fl u\geq r_1 $ and $\fl v\geq r_2$, then
\[\frac{\fl{N_1(u,v)}}{r_1}\geq p_1P_1(c_1+f_{r,R}c_2),\quad \frac{\fl{N_2(u,v)}}{r_2}\geq p_2P_2(c_1+g_{r,R}c_2),\]
\end{enumerate}
\end{lem}
\begin{proof}
Consider $u,v$ such that $|u|\leq R_1$, $|v|\leq R_2$.
(a) Using the symbols $u_0$, $\bar u$, $\bar v$ introduced in the proof of Proposition \ref{prop:fixed-sets} and having in mind the equality $u_0=\alpha(\bar u,\bar v)$ and the estimates \eqref{eq:alfa,beta-estimates} we obtain
\begin{equation}\label{eq:a1}
\frac 1{q_1Q_1}u_0 \leq \int_0^{t_{\text{max}}}\bar u(t)dt.
\end{equation}
Exploiting the equation \eqref{eq:t2} we obtain
\begin{equation}\label{eq:a2}
\int_0^{t_{\text{max}}}\bar u(t)dt = \int_0^{t_{\text{max}}} S(t)u(0)dt+\int_0^{t_{\text{max}}}\int_0^t S(t-\tau)F(u,v)(\tau)d\tau dt.
\end{equation}
From \eqref{eq:a1}, \eqref{eq:a2}, \eqref{eq:sup-fg} and the fact that $u\leq R_1\chi_\Omega$ it can be concluded that
\begin{equation}\label{eq:a3}\frac 1{q_1Q_1}u_0 \leq R_1\int_0^{t_{\text{max}}} S_2(t)\chi_\Omega dt+R_1f^R\int_0^{t_{\text{max}}}\int_0^t S_2(\tau)\chi_\Omega d\tau dt.\end{equation}
From \eqref{eq:t1}, we obtain
\begin{equation}\label{eq:a4}
N_1(u,v)(t)=S(t)u_0+\int_0^tS(t-\tau)F(u,v)(\tau)d\tau\leq
S(t)u_0+R_1f^R\int_0^{t_{\text{max}}} S_2(\tau)\chi_\Omega d\tau.
\end{equation}
Combining \eqref{eq:a3} and \eqref{eq:a4} an applying the contractiveness of $S(t)$ gives
\[
\frac{|N_1(u,v)|}{R_1}\leq q_1Q_1C_1+ f^R (q_1Q_1C_2 + C_1).
\]
Similarly, we can obtain the estimate of $|N_2(u,v)|$.
(b) Assume now that $\fl u\geq r_1$. Then $u(0)\geq r_1\chi_D$. From \eqref{eq:Harnack} and \eqref{eq:inf-fg} we obtain
\[u(t)\geq mr_1\chi_D\text{ and } F(u,v)(t)\geq r_1f_{r,R}^0\chi_D \text{ for }t\in[t_0,t_1].\]
Using the symbols $\bar u,\bar v$ introduced in the proof of Proposition \ref{prop:fixed-sets} we obtain
\begin{align*}
\frac 1{p_1P_1}N_1(u,v)(0) & =\frac 1{p_1P_1} \alpha(\bar u,\bar v)\geq \int_0^{t_{\text{max}}} \bar u(t)dt \\
& = \int_0^{t_{\text{max}}} S(t) u(0)dt+ \int_0^{t_{\text{max}}}\int_0^t S(t-\tau)F(u,v)(\tau)d\tau dt\geq \\
& \geq r_1\int_0^{t_{\text{max}}} S_2(t)\chi_Ddt+ r_1f_{r,R}^0 \int_{t_0}^{t_{\text{max}}}\int_{t_0}^{\min\{t,t_1\}}S_2(t-\tau) \chi_D d\tau dt.
\end{align*}
Using the superadditivity of $\fl\cdot$ we obtain
\[\frac{\fl{N_1(u,v)}}{r_1}\geq p_1P_1(c_1+f_{r,R}^0c_2).\]
In the same manner we can obtain the estimate of $\fl{N_2(u,v)}$.
(c) Using the fact, that $\fl v\geq r_2$ implies $v(t)\geq mr_2\chi_D$ and following the calculations analogous to those above, we obtain the conclusion.
The positiveness of the constants $c_1,c_2,C_1,C_2$ follows from Corollary \ref{cor:m>0}.
\end{proof}
\begin{rem}\label{rem:WhySuchHarnack}
As it was pointed out in Remark \ref{rem:Harnack}, the Harnack-type inequality \eqref{eq:Harnack} is an analogue of the inequality (3.4) of \cite{b:gi-mm-rp}. These two inequalities play a crucial role in obtaining the estimates from below and are utilized for the calculation of the fixed point index on some suitable subsets of a cone.
The difference between these two Harnack-type inequalities deserves a comment, as our choice to use the inequality \eqref{eq:Harnack} led us to build the new theory presented in Section~\ref{sect:abstr}. The inequality (3.4) of \cite{b:gi-mm-rp} was directly derived from a Harnack-type inequality given by Trudinger \cite{b:Trud-ell}. The natural counterpart in our context would be the \textit{parabolic} Harnack inequality by Trudinger \cite[Theorem 1.2]{b:Trud}, which is valid for all weak supersolutions $u$ of the equation $u_t-\Delta u=0$ (that is, functions $u\geq 0$ such that $u_t-\Delta u\geq 0$). This inequality could be expressed in the following manner:
\begin{equation}\label{eq:Harnack-bad}u\geq c\|u\|\chi_{Q_-},
\end{equation} where $c>0$ is a constant and \[\|u\|=\int_{Q_+}|u(t,x)|dtdx,\ Q_+=\left[t^0,t^1\right]\times D,\ Q_-=\left[t_0,t_1\right]\times D\] for
$0<t^0<t^1<t_0<t_1\leq {t_{\text{max}}}$. However, the use of this inequality is somewhat unnatural and it seems that it leads to additional complication of the argument and to worse results. For the sake of brevity, we provide here only a brief explanation.
1) The inequality \eqref{eq:Harnack-bad} gives a lower bound on the values of $u$ on $Q_-$, which depend on the values of $u$ on $Q_+$. The proof of Lemma \ref{lem:estimates}(b) shows that it is more convenient to utilize the dependence of values $u|Q_-$ on values of $u|\{0\}\times D$, due to the presence of the nonlocal boundary condition $u(0,x)=\alpha(u,v)(x)$.
2) The inequality \eqref{eq:Harnack} is actually a consequence of the inequality $u\geq m\fl u\chi_{[t_0,t_1]\times D}$ used for $u(t)=S_2(t)\chi_D$, which follows from the very definition of $m$. On the other hand, the inequality \eqref{eq:Harnack-bad} is more general in the sense that the constant $c$ is so chosen that a supersolution $u$ of the equation $u_t-\Delta u=0$ satisfies the estimate $u(t,x)\geq c$ on $Q_-$ \emph{whenever} $\|u\|\geq 1$. This universality, which in other context proves to be very important, is not exploited in our consideration, and unfortunately, it effects in a negative way the constants arising in the lower bounds of the nonlinearities (the counterparts of $c_1$ and $c_2$). And lastly, those constants are more difficult to be established. In other words, having in mind the nature of the calculations from Lemma \ref{lem:estimates}, the use of minimum $\fl u$ is more convenient, natural and effective than the use of the integral seminorm $\|u\|$.
\end{rem}
\subsection{Existence results}
We are now prepared to establish some sufficient conditions for the existence of nonnegative nontrivial solutions of the problem \eqref{eq:parabolic}. In what follows we shall assume that $\alpha,\beta$ are as in Example~\ref{ex:alpha-integral} and that the estimates \eqref{eq:gG} is satisfied.
\begin{thm}\label{thm:existence1 appl}
Assume there exist radii $0<r_i<R_i$, $i\in\{1,2\}$ such that
\begin{equation}\label{eq:ind1-appl}
q_1Q_1C_1+ f^R (q_1Q_1C_2 + C_1)\leq 1,\quad
q_2Q_2C_1+ g^R (q_2Q_2C_2 + C_1)\leq 1
\end{equation}
and
\begin{equation}\label{eq:ind0-appl}
p_1P_1(c_1+f_{r,R} c_2)>1,\quad
p_2P_2(c_1+g_{r,R} c_2)>1.
\end{equation}
Then there exists at least one nonnegative solution $(u,v)$ of the problem \eqref{eq:parabolic} such that $|u|\leq R_1,$
$|v|\leq R_2$ and $\fl u > r_1 $, $\fl v > r_2$.
\end{thm}
\begin{proof} Lemma \ref{lem:estimates} implies that the conditions \eqref{eq:ind0} and \eqref{eq:ind1} are satisfied. Moreover, $r_i<\fl{\psi_i} R_i$, since $\fl{\psi_i}=1$. The assertion follows from Theorem \ref{thm:existence 2a}.
\end{proof}
\begin{rem} It is worth mentioning that the condition $\fl u > r_1$ implies $|u|> r_1$, which follows from both the definitions of $|\cdot|$ and $\fl\cdot$ and from Remark \ref{rem:lowerbound_norm}.
\end{rem}
\begin{thm}\label{thm:existence or appl}
Let $0<r_1<R_1$, $0<r_2<R_2$ and let $\tilde R_1\leq R_1$, $\tilde R_2\leq R_2$. Define the quantities
\begin{equation*}\label{eq:inf''-fg}
f_{r,\tilde R}^{00}:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ 0\leq u\leq \tilde R_1,\ 0\leq v\leq \tilde R_2}} \frac{f(t,x,u,v)}{r_1},\quad
g_{r,\tilde R}^{00}:=\inf\limits_{\substack{t\in[t_0,t_1],\ x\in D \\ 0\leq u\leq \tilde R_1,\ 0\leq v\leq \tilde R_2}} \frac{g(t,x,u,v)}{r_2}.
\end{equation*}
Assume that the condition \eqref{eq:ind1-appl} is satisfied and that
\begin{equation}
\label{eq:ind0 or appl}
f_{r,\tilde R}^{00} \geq \left(p_1P_1c_2\right)^{-1} \ \text{ or }\ g_{r,\tilde R}^{00} \geq \left(p_2P_2c_2\right)^{-1}.
\end{equation}
Then there exists a nontrivial nonnegative solution $(u,v)$ of \eqref{eq:parabolic} such that $ | u | \leq R_1$, $|v|\leq R_2$ and
\begin{equation}
\fl u \geq r_1 \text{ or }\fl v \geq r_2 \text{ or } |u|>\tilde R_1 \text{ or }|v|>\tilde R_2.
\label{eq:conclusion ororor}
\end{equation}
In particular, if $\tilde R_1 =R_1 $ and $\tilde R _2 =R_2,$ then there exists a nontrivial nonnegative solution $(u,v)$ with either $ \fl u \geq r_1 $ or $ \fl v \geq r_2$.
\end{thm}
\begin{proof}
As in the previous proof we know that the condition \eqref{eq:ind1''} from Theorem~\ref{thm:existence or a} is satisfied. Let
\[(u,v)\in A:= \set{(u,v)}{\fl u<r_1,\ \fl v< r_2,\ |u|\leq \tilde R_1,\ |v|\leq \tilde R_2 }.\]
In the same way as in the proof of Lemma \ref{lem:estimates}(b) we can demonstrate that
\begin{equation}\label{eq:estim(or)} \frac{\fl{N_1(u,v)}}{r_1}\geq p_1P_1c_2 f_{r,\tilde R}^{00},\ \frac{\fl{N_2(u,v)}}{r_2}\geq p_2P_2c_2 g_{r,\tilde R}^{00}\ \text{ for } |u|\leq\tilde R_1\text{ and }|v|\leq\tilde R_2.
\end{equation}
Therefore, the assumption \eqref{eq:ind0 or appl} implies the condition \eqref{eq:ind0 or} is satisfied and we can apply Theorem \ref{thm:existence or a} to obtain a solution $(u,v)\not\in A$. Clearly, this is equivalent to \eqref{eq:conclusion ororor}.
\end{proof}
\begin{rem}
The importance of Theorem \ref{thm:existence or appl} consists in the fact that the
assumption \eqref{eq:ind0 or appl} involves only one component of the system
nonlinearity $(f,g).$ Therefore, it allows different kind of growth of $f$
and $g$ near the origin.
A similar remark also applies to the following theorem.
\end{rem}
\begin{thm} \label{thm:existence 3sols appl}
Assume there exist radii $0<\rho_i<r_i<R_i$, $i\in\{1,2\}$, such that
\[
q_1Q_1C_1+ f^R (q_1Q_1C_2 + C_1)\leq 1,\quad
q_2Q_2C_1+ g^R (q_2Q_2C_2 + C_1)\leq 1,
\]
\[
q_1Q_1C_1+ f^\rho (q_1Q_1C_2 + C_1)\leq 1,\quad
q_2Q_2C_1+ g^\rho (q_2Q_2C_2 + C_1)\leq 1
\]
and
\begin{equation}\label{eq:ind0-appl 3sols}
p_1P_1(c_1+f_{r,R} c_2)>1,\quad
p_2P_2(c_1+g_{r,R} c_2)>1
\end{equation}
Then there exist three nonnegative solutions $(u_i,v_i)$ ($i=1,2,3$) of the system \eqref{eq:parabolic} with
\begin{align*}
| u_1 | &<\rho _1,\ | v_1 | <\rho _2\ ( \text{possibly the zero solution}) ; \\
\fl{ u_2} &<r_1 \text{ or } \fl{ v_2} <r_2;\ | u_2| >\rho _1 \text{ or }|v_2| >\rho _2\ (\text{possibly one solution component zero}) ; \\
\fl{ u_{3}} &>r_1,\ \fl{ v_{3}} >r_2\ (\text{both solution components nonzero}) .
\end{align*}
By the following slight strengthening of the assumption~\eqref{eq:ind0-appl 3sols}:
\begin{equation}\label{eq:ind0-appl 3sols'}
p_1P_1(c_1+f_{r,R}^0 c_2)>1,\quad
p_2P_2(c_1+g_{r,R}^0 c_2)>1
\end{equation}
we obtain a slight improvement of the precision in localizing the second solution:
\[\fl{ u_2} <r_1,\ \fl{ v_2} <r_2;\ | u_2| >\rho _1 \ \text{or }|
v_2| >\rho _2.\]
Moreover, having given numbers $0<\varrho_i< \rho_i$ ($i=1,2$),
\emph{(i)} if
\[p_1P_1(c_1+f_{\varrho,\rho} c_2)>1,\quad p_2P_2(c_1+g_{\varrho,\rho} c_2)>1,\]
then $\fl{u_1} \geq \varrho_1$ and $\fl{v_1} \geq \varrho_2$;
\emph{(ii)} if
\[
f_{\varrho,\tilde\rho}^{00} \geq \left(p_1P_1c_2\right)^{-1} \ \text{ or }\ g_{\varrho,\tilde \rho}^{00} \geq \left(p_2P_2c_2\right)^{-1}
\]
for some $\tilde \rho_1\leq \rho _1,$ $\tilde{\rho}_2\leq \rho _2$, then $\fl{u_1} \geq \varrho_1$ or $\fl{ v_1} \geq \varrho_2$ or $|u_1|>\tilde \rho_1$ or $|v_1|>\tilde\rho_2$.
\end{thm}
\begin{proof} One can use Lemma \ref{lem:estimates}. The first assertion follows from Theorem \ref{thm:existence 3sols a}, while second follows from Theorem \ref{thm:existence 3sols a'}. The third part of the conclusion, i.e. assertions (i) and (ii), is a consequence of Theorem \ref{thm:existence 3sols a''} and \eqref{eq:estim(or)}.
\end{proof}
\subsection{Non-existence results} We now present some sufficient conditions for the non-existence of positive solutions of the system~\eqref{eq:parabolic}. We still assume that $\alpha,\beta$ are as in Example~\ref{ex:alpha-integral} and that \eqref{eq:gG} is satisfied.
\begin{lem}\label{lem:fl>0} Let $(u,v)$ be a solution of the system~\eqref{eq:parabolic}. If $u\neq 0$ then $\fl{u}>0$ and if $v\neq 0$ then $\fl{v}>0$.
\end{lem}
\begin{proof} Let $0\leq t_0<{t_{\text{max}}}$ be such that $u(t_0)\neq 0$. From Proposition \ref{prop:fixed-sets} we know that $(u,v)$ is a fixed point of $M$. Thus,
\[\fl{u(0)} = \fl{\alpha(u,v)}\geq p_1P_1\fl{\int_0^{t_{\text{max}}} u(t)dt}\geq p_1P_1\fl{\int_{t_0}^{t_{\text{max}}} u(t)dt}\] and $u(t)\geq S(t-t_0)u(t_0)$ for $t\geq t_0$. From Lemma \ref{lem:m_eta>0} we therefore obtain
\[\fl{u}=\fl{u(0)}\geq p_1P_1 \fl{\int_{0}^{{t_{\text{max}}}-t_0} S(t)u(t_0) dt}>0.\]
\end{proof}
\begin{thm}\label{thm:non-existence} Put
\[\overline e_i = \max\left(\frac{1-q_iQ_iC_1}{q_iQ_iC_2 + C_1},\frac{1-q_iQ_i{t_{\text{max}}}}{C_1} \right),\quad \underline e_i= \frac{(p_iP_i)^{-1}-c_1}{mc_2},\ i\in\{0,1\}.\]
Assume that $(u_0,v_0)$ is a nonnegative solution of the system~\eqref{eq:parabolic}. If one of the following conditions holds:
\begin{align}
\label{eq:nonexistence u1} f(t,x,u,v)< \overline e_1 u &\text{ for all }t\in[0,{t_{\text{max}}}],\ x\in\Omega,\ u>0,\ v\geq 0,\\
\label{eq:nonexistence u2} f(t,x,u,v)> \underline e_1 u &\text{ for all }t\in[t_0,t_1],\ x\in D,\ u>0,\ v\geq 0,
\end{align}
then $u_0=0$.
Similarly, if one of the following conditions holds:
\begin{align}
\label{eq:nonexistence v1} g(t,x,u,v)< \overline e_2 v &\text{ for all }t\in[0,{t_{\text{max}}}],\ x\in\Omega,\ u\geq 0,\ v> 0,\\
\label{eq:nonexistence v2} g(t,x,u,v)> \underline e_2 v &\text{ for all }t\in[t_0,t_1],\ x\in D,\ u\geq 0,\ v> 0,
\end{align}
then $v_0=0$.
In particular, if $p_1P_1c_1>1$ ($p_2P_2c_1>1$), then $u=0$ ($v=0$), regardless of the properties of $f$ ($g$).
\end{thm}
\begin{proof} Suppose on the contrary that $u\neq 0$. Then $u_0=N_1(u_0,v_0)=M_1(u_0,v_0)$. Put $R_1=|u_0|$, $R_2=\max\{|v_0|,1\}$ and $r_1=\fl {u_0}>0$.
Assume that the inequality \eqref{eq:nonexistence u1} holds.
Observe that $f(t,x,u,v)/R_1<\overline e_1$ for $t\in[0,{t_{\text{max}}}]$, $x\in\overline\Omega$, $0\leq u\leq R_1$ and $0\leq v\leq R_2$. Therefore, $f^R<\overline e_1$.
From Lemma \ref{lem:estimates}(a) we obtain
\begin{equation}\label{eq:estim 1} 1=\frac{|u_0|}{R_1}=\frac{|N_1(u_0,v_0)|}{R_1}\leq q_1Q_1C_1+ f^R (q_1Q_1C_2 + C_1).\end{equation}
Similarly, in the analogous manner as in the proof of Lemma \ref{lem:estimates}(a), one can show that
\begin{equation}\label{eq:estim 2} 1=\frac{|u_0|}{R_1}=\frac{|M_1(u_0,v_0)|}{R_1}\leq q_1Q_1{t_{\text{max}}}+ f^R C_1.\end{equation}
The estimates \eqref{eq:estim 1} and \eqref{eq:estim 2} give $f^R\geq \overline e_1$, a contradiction.
Assume now that \eqref{eq:nonexistence u2} holds.
Observe that $f(t,x,u,v)/r_1>m\underline e_1$ for $t\in[t_0,t_1]$, $x\in D$, $mr_1\leq u\leq R_1$ and $0\leq v\leq R_2$. Therefore, $f_{r,R}^0>m\underline e_1$.
On the other hand, from Lemma \ref{lem:estimates}(b) we obtain
\begin{equation}\label{eq:estim 3} 1=\frac{\fl{u_0}}{r_1}=\frac{\fl{N_1(u_0,v_0)}}{r_1}\geq p_1P_1(c_1+f_{r,R}^0c_2).\end{equation}
From \eqref{eq:estim 3} we conclude that $f_{r,R}^0\leq m\underline e_1$, a contradiction.
The second assertion can be proved analogously.
\end{proof}
\begin{cor}\label{cor:non-existence}
\emph{(i)} If one of the inequalities \eqref{eq:nonexistence u1}-\eqref{eq:nonexistence v2} holds, then there are no positive solutions of the system \eqref{eq:parabolic}.
\emph{(ii)} If one of the inequalities \eqref{eq:nonexistence u1}-\eqref{eq:nonexistence u2} holds and one of the inequalities \eqref{eq:nonexistence v1}-\eqref{eq:nonexistence v2} holds, then there are no
nontrivial nonnegative solutions of the system \eqref{eq:parabolic}.
\end{cor}
\subsection{An example}\label{ex:[0,pi]}
In this one-dimensional example we show that all the constants $C_1,C_2,c_1,c_2,d$ that occur in our theory can be computed.
Let $\Omega=[0,\pi]$. Let us put $D=[b,\pi-b]$ for a fixed $0< b< \pi/2$. Then
\[\chi_D=\sum_{k=0}^\infty \frac{4}{(2k+1)\pi}\cos\big((2k+1)b\big)\cdot \sin\big((2k+1)x\big) \] and
\[\chi_\Omega=\sum_{k=0}^\infty \frac{4}{(2k+1)\pi}\sin\big((2k+1)x\big),\] where the convergence is considered in the space $L^2(\Omega)$.
The evolution of $\chi_D$ and $\chi_\Omega$ is given by the formulae
\[(S_2(t)\chi_D)(x)=\sum_{k=0}^\infty \frac{4}{(2k+1)\pi}\cos\big((2k+1)b\big)\cdot e^{-(2k+1)^2t}\cdot \sin\big((2k+1)x\big),\]
\[(S_2(t)\chi_\Omega)(x)=\sum_{k=0}^\infty \frac{4}{(2k+1)\pi}\cdot e^{-(2k+1)^2t}\cdot \sin\big((2k+1)x\big).\]
Put $b=\pi/4$, $t_0=0$ and $t_1={t_{\text{max}}}=1$. Then we have
\[m=(S_2(1)\chi_D)\left(\frac\pi 4\right)=
\sum_{k=0}^\infty (-1)^k\frac2{(2k+1)\pi} e^{-(2k+1)^2}\approx 0.23.
\]
Because
\[
\int_0^{t_{\text{max}}} S_2(\tau)\chi_Dd\tau(x)=\sum_{k=0}^\infty \frac{4}{(2k+1)^3\pi}\cos(2k+1)\frac\pi 4\cdot \left(1-e^{-(2k+1)^2}\right)\cdot \sin\big((2k+1)x\big)
\]
and
\[
\int_{0}^{t_{\text{max}}}\int_{0}^{t}S_2(t-\tau) \chi_D d\tau dt=
\sum_{k=0}^\infty \frac{4}{(2k+1)^5\pi}\cos(2k+1)\frac\pi 4\cdot e^{-(2k+1)^2}\cdot \sin\big((2k+1)x\big),
\]
we obtain
\[
c_1=\sum_{k=0}^\infty \frac{2\cdot (-1)^k}{(2k+1)^3\pi}\left(1-e^{-(2k+1)^2}\right)\approx 0.38\]
and
\[
c_2=\sum_{k=0}^\infty \frac{2\cdot (-1)^k}{(2k+1)^5\pi} e^{-(2k+1)^2} \approx 0.23
\]
Moreover, because
\[\int_0^{t_{\text{max}}} S_2(\tau)\chi_\Omega d\tau =\sum_{k=0}^\infty \frac{4}{(2k+1)^3\pi}\cdot \left(1-e^{-(2k+1)^2}\right)\cdot \sin\big((2k+1)x\big)\] and
\[
\int_0^{t_{\text{max}}}\int_0^t S_2(\tau)\chi_\Omega d\tau dt=
\sum_{k=0}^\infty \frac{4}{(2k+1)^5\pi}\cdot e^{-(2k+1)^2}\cdot \sin\big((2k+1)x\big),
\]
\begin{equation}C_1=\left|\int_0^{t_{\text{max}}} S_2(\tau)\chi_\Omega d\tau\right|,\quad C_2=\left|\int_0^{t_{\text{max}}}\int_0^t S_2(\tau)\chi_\Omega d\tau dt\right|\end{equation}
we obtain
\[
C_1= \sum_{k=0}^\infty \frac{4\cdot (-1)^k}{(2k+1)^3\pi}\cdot \left(1-e^{-(2k+1)^2}\right)\approx 0.77
\]
and
\[
C_2=
\sum_{k=0}^\infty \frac{4\cdot (-1)^k}{(2k+1)^5\pi}\cdot e^{-(2k+1)^2}\approx 0.47.
\]
If we put $G_i(x)=g_i(x)=x$, then $p_i=P_i=q_i=Q_i=1$ and the conditions \eqref{eq:ind1-appl} and \eqref{eq:ind0-appl} are equivalent to the following inequalities:
\[
f^R,g^R\leq \frac{1-C_1}{C_1 + C_2}\approx 0.19,\quad
f_{r,R},g_{r,R} >\frac{1-c_1}{c_2} \approx 2.64.
\]
The numerical calculations indicate that the choice $b=\pi/4$ is optimal, i.e. the ratio $(1-c_1)/(c_2m)$ is the smallest.
\section*{Acknowledgement}
The authors are indebted to Prof. Precup for the useful discussions on the nonlocal parabolic case, in particular for providing them with the works \cite{b:tc-rp-pr,b:rp-par}.
G. Infante was partially supported by G.N.A.M.P.A. - INdAM (Italy).
This paper was written during the postdoctoral
stage of M. Maciejewski at the University of Calabria, supported by a
research fellowship within the project ``Enhancing Educational Potential of
Nicolaus Copernicus University in the Disciplines of Mathematical and
Natural Sciences'' (Project no. POKL.04.01.01-00-081/10) and
by the NCN Grant 2013/09/B/ST1/01963.
|
\section{The German Protocol Code (Chou \etal~\cite{CMP})}
\label{AppendixGermaCode}
\begin{minipage}[t]{\textwidth}\tiny
\hspace*{-0.1\textwidth}
\begin{tabular}[t]{ll}
\begin{minipage}[t]{0.4\textheight}
\begin{verbatim}
const ---- Configuration parameters ----
NODE_NUM : 4;
DATA_NUM : 2;
type ---- Type declarations ----
NODE : scalarset(NODE_NUM);
DATA : scalarset(DATA_NUM);
CACHE_STATE : enum {I, S, E};
CACHE : record State : CACHE_STATE; Data : DATA; end;
MSG_CMD : enum {Empty, ReqS, ReqE, Inv, InvAck, GntS, GntE};
MSG : record Cmd : MSG_CMD; Data : DATA; end;
var ---- State variables ----
Cache : array [NODE] of CACHE; -- Caches
Chan1 : array [NODE] of MSG; -- Channels for Req*
Chan2 : array [NODE] of MSG; -- Channels for Gnt* and Inv
Chan3 : array [NODE] of MSG; -- Channels for InvAck
InvSet : array [NODE] of boolean; -- Nodes to be invalidated
ShrSet : array [NODE] of boolean; -- Nodes having S or E copies
ExGntd : boolean; -- E copy has been granted
CurCmd : MSG_CMD; -- Current request command
CurPtr : NODE; -- Current request node
MemData : DATA; -- Memory data
AuxData : DATA; -- Latest value of cache line
---- Initial states ----
ruleset d : DATA do startstate "Init"
for i : NODE do
Chan1[i].Cmd := Empty; Chan2[i].Cmd := Empty; Chan3[i].Cmd := Empty;
Cache[i].State := I; InvSet[i] := false; ShrSet[i] := false;
end;
ExGntd := false; CurCmd := Empty; MemData := d; AuxData := d;
end end;
---- State transitions ----
ruleset i : NODE do rule "SendReqS"
Chan1[i].Cmd = Empty & Cache[i].State = I
==>
Chan1[i].Cmd := ReqS;
end end;
ruleset i : NODE do rule "SendReqE"
Chan1[i].Cmd = Empty & (Cache[i].State = I | Cache[i].State = S)
==>
Chan1[i].Cmd := ReqE;
end end;
ruleset i : NODE do rule "RecvReqS"
CurCmd = Empty & Chan1[i].Cmd = ReqS
==>
CurCmd := ReqS; CurPtr := i; Chan1[i].Cmd := Empty;
for j : NODE do InvSet[j] := ShrSet[j] end;
end end;
ruleset i : NODE do rule "RecvReqE"
CurCmd = Empty & Chan1[i].Cmd = ReqE
==>
CurCmd := ReqE; CurPtr := i; Chan1[i].Cmd := Empty;
for j : NODE do InvSet[j] := ShrSet[j] end;
end end;
\end{verbatim}
\end{minipage}
&
\begin{minipage}[t]{0.4\textheight}
\begin{verbatim}
ruleset i : NODE do rule "SendInv"
Chan2[i].Cmd = Empty & InvSet[i] = true &
( CurCmd = ReqE | CurCmd = ReqS & ExGntd = true )
==>
Chan2[i].Cmd := Inv; InvSet[i] := false;
end end;
ruleset i : NODE do rule "SendInvAck"
Chan2[i].Cmd = Inv & Chan3[i].Cmd = Empty
==>
Chan2[i].Cmd := Empty; Chan3[i].Cmd := InvAck;
if (Cache[i].State = E) then Chan3[i].Data := Cache[i].Data end;
Cache[i].State := I; undefine Cache[i].Data;
end end;
ruleset i : NODE do rule "RecvInvAck"
Chan3[i].Cmd = InvAck & CurCmd != Empty
==>
Chan3[i].Cmd := Empty; ShrSet[i] := false;
if (ExGntd = true)
then ExGntd := false; MemData := Chan3[i].Data; undefine Chan3[i].Data end;
end end;
ruleset i : NODE do rule "SendGntS"
CurCmd = ReqS & CurPtr = i & Chan2[i].Cmd = Empty & ExGntd = false
==>
Chan2[i].Cmd := GntS; Chan2[i].Data := MemData; ShrSet[i] := true;
CurCmd := Empty; undefine CurPtr;
end end;
ruleset i : NODE do rule "SendGntE"
CurCmd = ReqE & CurPtr = i & Chan2[i].Cmd = Empty & ExGntd = false &
forall j : NODE do ShrSet[j] = false end
==>
Chan2[i].Cmd := GntE; Chan2[i].Data := MemData; ShrSet[i] := true;
ExGntd := true; CurCmd := Empty; undefine CurPtr;
end end;
ruleset i : NODE do rule "RecvGntS"
Chan2[i].Cmd = GntS
==>
Cache[i].State := S; Cache[i].Data := Chan2[i].Data;
Chan2[i].Cmd := Empty; undefine Chan2[i].Data;
end end;
ruleset i : NODE do rule "RecvGntE"
Chan2[i].Cmd = GntE
==>
Cache[i].State := E; Cache[i].Data := Chan2[i].Data;
Chan2[i].Cmd := Empty; undefine Chan2[i].Data;
end end;
ruleset i : NODE; d : DATA do rule "Store"
Cache[i].State = E
==>
Cache[i].Data := d; AuxData := d;
end end;
---- Invariant properties ----
invariant "CtrlProp"
forall i : NODE do forall j : NODE do
i != j -> (Cache[i].State = E -> Cache[j].State = I) &
(Cache[i].State = S -> Cache[j].State = I | Cache[j].State = S)
end end;
invariant "DataProp"
( ExGntd = false -> MemData = AuxData ) &
forall i : NODE do Cache[i].State != I -> Cache[i].Data = AuxData end;
\end{verbatim}
\end{minipage}
\end{tabular}
\end{minipage}
\\ ~ \\ ~
\section{Proof of Soundness}
\label{sec:proof}
Before proving the theorem, we first establish the following lemma:
\newtheorem*{lem1}{Lemma}
\begin{lem1}
The disjunction of predicates of all invariants in \InvSet holds,
i.e., $\bigvee_{inv \in \InvSet} inv.pred$ holds.
\end{lem1}
\begin{proof}
We prove this by induction over the splitting step in our method.
{\em Base Case:} Our method starts with the initial invariant
$true \Rightarrow \big(\forall i \in
IndexSet: \g(\AgentRuleSet(i))\big)$ in \InvSet. Thus, it trivially
satisfies the lemma.
\revised{The proof here}
{\em Induction Step:} Next, suppose at some point during the
generation of invariants, the set of candidates is \InvSet. On model
checking, invariant $inv$ in \InvSet fails with
$\g(\AgentRuleSet(i_f))$ being \false for agent $i_f$. In case there
is an error in the protocol or flows due to a rule $rl(i_f)$ being
enabled for agent $i_f$ in the failing state, the loop exits without
modifying \InvSet and so the lemma holds trivially. In the second
case, the invariant is split into invariants $inv1$ and $inv2$ by
using conflict condition $conf$.
Now for this case, $inv1.pred = (inv.pred \land \lnot conf)$ and
$inv2.pred = (inv.pred \land conf)$. Clearly, the disjunction of
predicates $inv1$ and $inv2$ equals to $inv.pred$, the predicate of
$inv$. Thus, the disjunction of predicates of the new and old set of
invariants is the same, i.e., $\bigvee_{inv \in \InvSet} inv.pred$ =
$\bigvee_{inv \in \InvSet'} inv.pred$, where the new set of invariants
$\InvSet'$ $=$ $\InvSet$ $\setminus$ $\{inv\}$ $\cup$ $\{inv1,inv2\}$.
Hence, by induction, the above lemma holds.
\end{proof}
Now, using the above lemma, we prove the following theorem to
establish soundness of our method:
\begin{lem3}
If the set of invariants $\InvSet$ along with the set of assertions
\AssertSet hold, they collectively imply \mdeadlock freedom, i.e.,
$\big(\big(\bigwedge_{inv \in \InvSet} (\Protocol \models inv)\big)$
$\land$ $\big(\bigwedge_{asrt \in \AssertSet} (\Protocol \models
asrt)\big)\big)$ $\Rightarrow \big(\Protocol \models (\bigvee_{i}
\bigvee_{j} rl_j(i).\rho)\big)$.
\end{lem3}
\begin{proof}
Let the protocol be in some reachable state $s$. We argue that some
agent has at least one rule enabled in every such reachable state. By
the above lemma, $\bigvee_{inv \in \InvSet} inv.pred$ holds in state
$s$. Thus, there must exist some invariant $inv$ such that its
predicate holds in $s$, i.e., $\exists inv \in \InvSet: inv.pred =
true$.
Now, let $inv$ be $inv.pred \Rightarrow (\forall
i \in \ISet{inv}: \, \g(\AgentRuleSet(i)))$. Then, since the assertion
$inv.pred \Rightarrow \ISet{inv} \neq \{\}$ is in the set \AssertSet,
which holds as well, there is some agent $i_0$ such that it is in
$\ISet{inv}$ and $\g(\AgentRuleSet(i_0))$ holds, i.e., $\exists
i_o \in \ISet{inv}: \, \g(\AgentRuleSet(i_0))$. Thus, agent $i_0$ is
enabled in the state $s$, and so the state is not an \mdeadlock state.
\end{proof}
\section{Deadlock freedom from Flow Properties}
\section{Verifying Flow Properties for Unbounded Agents}
\label{sec:CMP}
We now show how to verify the invariants in $\InvSet$ for an unbounded
number of agents by leveraging the data-type reduction abstraction
along with the CMP method.
{\bf Abstraction: Data-type Reduction}
Since the invariant is of the form $inv.pred \Rightarrow \big(\forall
i \in \ISet{inv}: \g(\AgentRuleSet(i))\big)$, by symmetry, it is
sufficient to check: $inv.pred \Rightarrow \big((1 \in \ISet{inv})
\Rightarrow \big(\g(\AgentRuleSet(1))\big)\big)$. In order to verify
this invariant, just the variables of agent $1$ are required. Then,
our abstraction keeps just the agent $1$, and discards the variables
of all the other agents by replacing them with a state-less {\em
environment} agent. We refer to agent $1$ as a {\em concrete} agent
and the environment as \Other with {\em id} $o$.
In the original protocol, since all the agents other than agent $1$
interact with it by updating the global variables, the actions of
these agents on the global variables are over-approximated by the
environment agent. This environment agent does not have any local
state. The construction of this agent \Other is automatic and
accomplished syntactically: further details on the automatic
construction are available in \cite{ParamVerMurali}. The final
constructed abstraction then consists of: (1) a concrete agent $1$,
(2) an environment agent \Other with {\em id} $o$, and (3) invariants
specified on variables of agent $1$ and global variables. This
abstraction is referred to as {\em data-type reduction}. If the
original protocol is \Protocol, and invariant set $\InvSet$, we denote
this abstraction by $data\_type$ and thus the abstract model by
$data\_type(\Protocol)$ and the abstracted invariants on agent 1 by
$data\_type(\InvSet)$.
\nop{
In order to prove the invariants for an unbounded number of agents in
the protocol model, a finite model is constructed by using data-type
reduction. This abstraction works by essentially keeping a small
number of agents, as is, and discarding the state of all other
agents. The agents which are left as is are referred to as {\em
concrete} agents. Leveraging symmetry, suppose agent $1$ is kept as
is and is the concrete agent. Now,
The agents which are left as is are referred to as {\em
concrete} agents. Since all the other agents, which are discarded,
interact with the concrete agent by modifying the global state, their
actions on the global state are over-approximated and replaced with an
abstract environment agent, referred to as \Other. The actions of the
environment agent \Other create {\em interference} for the concrete
agents: if the invariants hold despite the interference, then they
hold for the unbounded model as well. The construction of this
environment agent \Other is automatic and accomplished
syntactically~\cite{ParamVerMurali}.
Next, once the abstraction (with agents $1$, $2$ and \Other) is
constructed, the invariants are proven for the concrete agent $1$ (and if true,
they hold for all agents by symmetry)
}
{\bf \emph{Abstraction for German Protocol}} We now describe how the
rule $SendGntE(i)$ gets abstracted in $data\_type(\Protocol)$. In the
abstract model, there is one concrete agent $1$, which has the rule
$SendGntE(1)$. Next, $SendGntE(o)$ is constructed as follows. (1) The
guard is abstracted by replacing all atoms consisting of local
variables (e.g. $GntChannel[i] = Empty$) with \true or \false
depending on which results in an over-abstraction and by replacing any
usage of $i$ in atoms with global variables (e.g. $CurPtr=i$) with $o$
(i.e. $CurPtr=o$). (2) The action is abstracted by discarding any
assignments to local variables. Further, assignments to global pointer
variables are abstracted as well: any usage of $i$ (e.g. $CurPtr:=i$)
is replaced by $o$ (i.e. $CurPtr:=o$). The rule for agent \Other is
shown below:
\vspace*{0.1cm}
\begin{minipage}{\textwidth}
\footnotesize
\RulesetSingle{SendGntE($o$)}{CurCmd = ReqE $\land$
CurPtr = o $\land$ true $\land$ Exgntd = false $\land$
\hspace*{.6cm} ShrSet = \{\}}{ShrSet := \{o\}; ExGntd := true; CurCmd :=
Empty;
\hspace*{.6cm}CurPtr := NULL;}
\end{minipage}
\nop{
FIX THIS: Next, the invariants to be proven on this abstract model are
modified as well: by symmetry, the invariants only need to be checked
for concrete values of $i$, i.e., $i=\{1,2\}$. Thus, \InvSet is
replaced by \InvConcSet in the above invariants to obtain the
following invariants:
\begin{itemize}
\item \((CurPtr=null) \Rightarrow \g(1)\)
\item \(\lnot(CurPtr=null) \land (1=CurPtr) \land (\lnot Inv\_Cond
\lor Sharer=null) \Rightarrow \g(1)\)
\item \(\lnot(curptr=null) \land (1 = Sharer) \land (Inv\_Cond \land \lnot
(Sharer=null)) \Rightarrow \g(1)\)
\end{itemize}
}
\nop{
\newtheorem*{lem4}{Theorem}
\begin{lem4}
Assuming the invariants in $\InvSet$ are single-index and the protocol
$\Protocol$ is symmetric, the abstraction $data\_type$ is sound.
\end{lem4}
We refer an interested reader to \cite{ParamVerMurali} for proof.
}
{\bf The Abstraction-Refinement Loop of the CMP Method}
The CMP method works as an abstraction-refinement loop, as shown in
\Fig{Fig:CMPMethod}. In the loop, the protocol and invariants are
abstracted using data-type reduction. If the proof does not succeed,
the user inspects the returned counterexample $cex$ and following
possibilities arise. (1) Counterexample $cex$ is real, in which case
an error is found and so the loop exits. (2) Counterexample $cex$ is
spurious and so the user refines the protocol by adding a {\em
non-interference lemma} $lem$. The function {\em strengthen} updates
the guard $rl_j(i).\rho$ of every rule $rl_j(i)$ of the protocol to
$rl_j(i).\rho \land lem(j)$; this way, on re-abstraction with
$data\_type$ in line 1, the new abstract protocol model is
refined. Additional details on the CMP method are available in
\cite{CMP, ParamVerifCMPKristic}.
\nop{
A useful heuristic for determining if the counterexample is spurious
is to examine the behavior of the \Other agent in the counterexample
trace. Typically, in a spurious trace, the rules of the \Other agent
fire when they are not supposed to (for example doing a memory write
when another agent, say $1$, has exclusive access).
}
\input{CMPPseudocode.tex}
\nop{
\newtheorem*{lem5}{Theorem}
\begin{lem5}
Assuming the invariants in $\InvSet$ are single-index and the protocol
$\Protocol$ is symmetric, the CMP loop is sound, i.e.,
$data\_type(\Protocol^{\#}) \models$ $data\_type(\InvSet^{\#})$
$\Rightarrow$ $\Protocol(N) \models$ $\InvSet$ for all values of $N$.
\end{lem5}
We refer an interested reader to \cite{ParamVerMurali} for proof.
}
\nop{
Since the constructed environment agent \Other is completely
unconstrained, on model checking, the property under check may get
violated by the constructed abstract model due to spurious behaviors
exhibited by the agent \Other. The CMP method works by repeatedly
abstracting refined versions of the protocol model until either the
property is proven to be correct, or a real bug is found.
\begin{figure}
\includegraphics[scale=0.5]{./Figures/CMPLoop.jpg}
\caption{The CMP Method}
\label{figCMPLoop}
\vspace{-0.75cm}
\end{figure}
\Fig{figCMPLoop} shows the flow of the CMP method. In the method, the
protocol is first abstracted. Next, model checking of the protocol is
done---in case the model checking succeeds, the property is proven to
be correct for an arbitrary number of agents. If on the other hand the
model checking fails, the user inspects the counter-example. In case,
the counterexample is a valid counterexample, a real bug has been
found and so the loop exits. On the other hand, if the counterexample
is spurious, refinement must be done. This is done in two
steps. First, the user comes up with a lemma $lem$ for strengthening
the protocol model. The strengthening is done by adding the lemmas to
the guards of all the rules in the protocol---a rule $rl$ defined as
\(\rho \Rightarrow a\) is strengthened to \(\rho \wedge lem
\Rightarrow a\). Second, re-abstraction is done for the strengthened
protocol: the abstracted version of the rule $rl$ for the thread
\Other is more constrained than the original abstracted version---thus
resulting in a refined abstraction. Note that during this refinement,
no additional state is added to the model.
This helps in keeping the model tractable for model checking. {\em
This efficient refinement along with the efficiency of data type
reduction is the key factor for the success of this method for
industrial scale protocols.}
}
\nop{as follows: suppose that the lemma $lem$ is used to refine the
model. Now consider a rule $r$ of the protocol \Protocol defined as:
\(\rho \Rightarrow a\). Then, refining \Protocol with $lem$ involves
changing this rule to \(\rho \wedge lem \Rightarrow a\) (we refer to
this as $strengthen$) and then re-abstracting the new program with
the new rule obtained.
This abstraction-refinement continues in a loop, as shown in
\Fig{Fig:CMPMethod}. In the loop, whenever the model is verified (line
2), if the proof succeeds, the property is proven (line 9). If, on
the other hand, there is a counterexample for the refined system, the
user must distinguish between the following cases by examining the
counterexample. 1) The counterexample is valid and so a real bug is
found. (line 4) 2) The counterexample is spurious and so refinement must
be done. (lines 6-8)
A useful heuristic for determining if the counterexample is spurious
is to examine the behavior of the \Other agent in the counterexample
trace. Typically, in a spurious trace, the rules of the \Other agent
fire when they are not supposed to (for example doing a memory write
when another agent, say 1, has exclusive access).
\input{CMPPseudocode.tex}
\paragraph{Note:} Non-interference lemmas were not required for verification
of the invariants we generated for the German protocol; they were only
required for the Flash protocol. Then, for the sake of brevity we skip
the explanation of refinement through an example and refer interested
reader to BLA. [COMPLETE ME]
}
\nop{
\subsection{Advantages of the CMP Method}
The key advantage of the CMP method is that no extra state is added
during refinement: only the guards of rules are strengthened.
This helps in keeping the model tractable for model checkers---this
along with the efficiency of data type reduction is the key factor for
the success of this approach for industrial scale protocols.
Another key advantage of the CMP method is that the lemmas used for
strengthening are also checked during the process and any false lemma
that is added will be detected by the model
checker~\cite{ParamVerifCMPKristic}. This guarantees that the
refinement step does not affect the soundness of the proof and greatly
enhances its convenience for the user.
}
\nop{ Formally, suppose that the lemma $L$ is used. Now consider a
rule $r$ of the program $P$ defined as: \(\rho \Rightarrow
a\). Then, refining $P$ with $L$ involves changing this rule to
\(\rho \wedge L \Rightarrow a\) and then re-abstracting the new
program with the new rule obtained. Observe that in this refinement
approach, no extra state gets added to the abstract model. This is
important for the efficiency of the CMP method based verification.
This abstraction-refinement continues in a loop, as shown in \Fig{}.
In the loop, whenever a counterexample is returned by the model checker,
the user determines if the counterexample is real or spurious.
Whenever a violation
On detecting a
violation, the user examines the counter-example.
If this refined system passes the model checker, the property is
proven. If, on the other hand, there is another counterexample for
the refined system, the user must distinguish between three possible
cases by examining the counterexample. 1) The counterexample is
valid. 2) The counterexample is not valid and the lemma is correct,
in which case further refinement is required.
3) The counterexample is not valid and a lemma is incorrect, in which
case the incorrect lemma must be removed or modified.
}
\section{Fine-Grained Synchronization Algorithm}
\begin{figure}
\centering
\begin{minipage}{.65\textwidth}
{\sc CMP}$(\Protocol(N),\InvSet)$
\begin{algorithmic}[1]
\IndState[0]{$\Protocol^{\#}=\Protocol(N)$;$\InvSet^{\#}=\InvSet$}
\IndState[0]{$while \; data\_type(\Protocol^{\#}) \not\models$
$data\_type(\InvSet^{\#}) \; do$}
\IndState[1]{examine counterexample $cex$}
\IndState[1]{if $cex$ is real, exit}
\IndState[1]{if spurious:}
\IndState[2]{find lemma $lem=\forall i.lem(i)$}
\IndState[2]{$\Protocol^{\#}=strengthen(\Protocol^{\#},lem)$}
\IndState[2]{$\InvSet^{\#} = \InvSet^{\#} \cup lem$}
\end{algorithmic}
\end{minipage}
\caption{The CMP method}
\label{Fig:CMPMethod}
\end{figure}
\section{Conclusions and Future Work}
In this paper we have presented a method to prove freedom from a
practically motivated deadlock error which spans the entire cache
coherence protocol, an \mdeadlock. Our method exploits high-level
information in the form of message sequence diagrams---these are
referred to as {\em flows} and are readily available in industrial
documents as charts and tables. Using our method, a set of invariants
can be derived which collectively imply \mdeadlock freedom. These
invariants enable the direct application of industrial scale techniques
for parameterized verification.
As part of future work, we plan to take up verification of livelock
freedom by exploiting flows. Verifying livelock requires formally
defining a notion of the protocol doing useful work. This information
is present in flows---efficiently exploiting this is part of our
ongoing research.
\section{Running Example: German protocol}
We use the German protocol as our running example.
\subsection{Flows for German Protocol}
The flows for the German protocol are shown in
\Fig{figGermanFlows}. These flows represent the order in which various
messages in the protocol are passed, in order to accomplish a high
level request. Thus, they provide high level information about the
protocol. The German protocol consists of the following flows:
\begin{itemize}
\item \Shared: The \Shared flow, shown in \Fig{figGermanFlowsShrd}
processes a request from cache $i$ for shared access. It consists of
messages \{$sendReqS$, $RecvReqS$, $SendGntS$, $RecvGntS$\} passed
between the requesting cache and directory.
\item \Exclusive: Similar to shared flow, shown in
\Fig{figGermanFlowsExcl}.
\item \Invalidate: The \Invalidate flow is a sub-flow of \Shared and
\Exclusive flows: it has rules \{$sendInv$, $sendInvAck$,
$recvInvAck$\}. It is present in both figures,
\Fig{figGermanFlowsShrd} and \Fig{figGermanFlowsExcl}.
\end{itemize}
\subsection{Protocol Code}
The code for the German protocol in Murphi language~\cite{} is
presented in \Fig{figGermanCode} (courtesy~\cite{CMP}).
\paragraph{Murphi Syntax:} Murphi language consists of {\em rulesets} which is a collection of
{\em rules}, where each rule is a guarded action. For example,
$SendReqS$ is a ruleset which represents a set of rules $1..N$, where
rule $i$ (denoted by $SendReqS[i]$) corresponds to $i^{th}$ cache
sending a request for shared access. The guard of $SendReqS[i]$ is
$Chan1[i].Cmd = Empty \land Cache[i].State = S$ and the action is
$Chan1[i].Cmd := ReqS$. {\em Observe that each message in the flows
shown above correspond to a ruleset in the code.}
\paragraph{Murphi Semantics:} Each rule is executed atomically. Thus, at any time,
Murphi non-deterministically selects a rule which has its guard true
and executes it.
\paragraph{Relevant Global State:} The directory points to the cache
whose request is currently being served (for a particular cache line)
using the pointer \curptr. Further, it stores the request type from
\curptr in \curcmd. Next, \exgntd is set to true if some cache has
exclusive access, else it is set to false. Finally, a set of
variables, \shrset, stores the caches which share a particular cache
line. Thus, \shrset[i] is true if cache $i$ has a copy of the
line. Finally, an auxiliary variable $cache\_inv$ (FIX: not shown in
code) points to some cache $c$ such that \shrset[c] is
true---$cache\_inv$ is null if and only if \shrset is false for all
caches.
\begin{figure}
\hspace*{-5cm} \includegraphics[scale=1]{./Figures/germanProtocol.pdf}
\caption{German Protocol}
\label{figGermanCode}
\end{figure}
\subsection{Specifying and Proving Progress}
We prove that at all times, progress of some flow can always happen.
Since the flows are acyclic in nature and finite in length~\cite{},
progress of some flow at all times implies that the protocol does not
deadlock.
We denote the logical or of the guards of all the rules of a flow $fl$
by $progress(fl)$. Then, $progress(\Shared(i))$ equals to:
$$guard(sendReqS(i)) \lor guard(RecvReqS(i)) \lor guard(SendGntS(i))
\lor guard(RecvGntS(i)).$$ Similarly, $progress(\Invalidate(i)) =
guard(sendInv(i)) \lor guard(sendInvAck(i)) \lor
guard(recvInvAck(i))$.
\subsubsection{Generating Progress Invariants}
The progress invariants are first generated by iteratively model
checking a model with 2-3 caches. After generation, these invariants are then
checked for an unbounded number of caches.
The generation is done using 2-3 caches because of two factors: (1) a
smaller model is easy to iterate model checking on and, (2) progress
proof from smaller model typically generalizes to unbounded number of
caches---a consequence of the fact that protocol designers think in
terms of a small number of caches.
The generation is done as follows:
(1) First, the user adds invariants for progress of flow corresponding
to each request. Thus, the invariant
\begin{equation}
\forall i:progress(i), \tag{inv-1}\label{inv1}
\end{equation}
where, $progress(i) = (progress(\Shared(i))|progress(\Exclusive(i)))$,
is added.
(2) Second, model checking is done on this. If the model checking
succeeds, progress is proven.
(3) If the model checking fails, as happens in this case, the user
inspects the counter-example. The following situations can create the
counter-example:
a) Progress of the request is stuck since another flow is making
progress at the directory. In the German protocol, in case the
directory is processing a request from a cache, it points to the cache
through $curptr$. Thus, if $curptr!=i$, $progress(i)$ will fail for
cache $i$. This can be seen from the guard of $SendGntS(i)$ and
$SendGntE(i)$ in protocol code---the guard requires $curptr=i$. Thus,
the user can strengthen the property by splitting the
invariant~\ref{inv1} into two as follows:
\begin{equation}
(curptr=NULL) \rightarrow \forall i: progress(i)\tag{inv-1.1}\label{inv11}
\end{equation}
\begin{equation}
\forall i:(curptr=i) \rightarrow progress(i) \tag{inv-1.2}\label{inv12}
\end{equation}
Note that the disjunction of antecedents of the above two invariants
is $true$. This is required to prove progress in all cases.
b) Next, even after the above strengthening, the invariants fail. This
is because an invalidate may be required before the flow can progress.
This can be seen from the guards of $SendGntS$ and
$SendGntE$---$ExGntd$ must be false and in case of $SendGntE$, the
ShrSet must be null. We assume that the one of the caches in the
ShrSet is pointed by $cache\_inv$ (implemented as an auxiliary
variable) and that $cache\_inv$ is null if and only if ShrSet is
empty. Then, $((CurCmd=ReqE)|((CurCmd=ReqS) \land ExGntd)) \land
cache\_inv!=NULL)$ implies that invalidate must happen.
Thus invariant~\ref{inv12} can be split into two invariants as
follows:
\begin{equation}
\forall i:(curptr=i) \land !(Inv\_Cond \land
cache\_inv=NULL)\rightarrow progress(i)\tag{inv-1.2.1}\label{inv121}
\end{equation}
\begin{equation}
\forall i,j: (curptr=i) \land Inv\_Cond \land (j= cache\_inv)
\rightarrow progress\_invalidate(j),\tag{inv-1.2.2}\label{inv122}
\end{equation}
where, $Inv\_Cond$ is $((CurCmd=ReqE)|((CurCmd=ReqS) \land ExGntd))$.
For the German protocol, \ref{inv11}, \ref{inv121}, \ref{inv122} are
invariants.
\nop{
The following situations can create the counter-example: a) Progress
of a request is stuck due to a conflicting flow (directory
pending). The information of this conflicting flow is typically stored
in the directory or can be added as an auxiliary variable. (b) Waiting
for another flow to complete -- for example invalidates. In this case,
the directory again stores information about the flow it is waiting
for. And thus, information is added by the user accordingly.
}
\subsubsection{Proving Progress Invariants}
Once generated, the progress invariants can be proved in accordance
with the CMP method~\cite{CMP}. This is easy.
\section{Experiments}
\label{sec:Experiments}
Using our approach, we verified Murphi (CMurphi 5.4.6) implementations
of the German and Flash protocols (available
online~\cite{ProtocolDeadlockCodeFiles}). Our experiments were done on
a 2.40 GHz Intel Core 2 Quad processor, with 3.74 GB RAM, running
Ubuntu 9.10.
{\bf German Protocol} We verified the invariants discussed in
\Sec{sec:GermanInvariantGeneration}, in order to prove \mdeadlock
freedom. We chose to use an abstraction with 2 agents and an
environment agent, so that the mutual exclusion property can also be
checked.
The proof finished in 217s with 7M states explored. No
non-interference lemmas were required to refine the model, in order to
verify the invariants presented in
\Sec{sec:GermanInvariantGeneration}. Since typically protocols are
also verified for properties like data integrity (i.e. the data stored
in the cache is consistent with what the processors intended to write)
and mutual exclusion, we model checked the above invariants along with
these properties. In this case, the abstract model was constrained
and model checking this model was faster and took 0.1 sec with 1763
states explored.
{\bf \emph{Buggy Version}} We injected a simple error in the German
protocol in order to introduce an \mdeadlock. In the bug, an agent
being invalidated drops the acknowledgement $SendInvAck$ it is
supposed to send to the directory. This results in the entire protocol
getting blocked, hence an \mdeadlock situation. This was detected by
the failing of the invariant \Inv{inv-1.2.2}, discussed in
\Sec{sec:GermanInvariantGeneration}.
{\bf Flash Protocol} Next, we verified the Flash
protocol~\cite{FlashProtocol} for deadlock freedom. The Flash protocol
implements the same high-level requests as the German protocol. It
also uses a directory which has a Boolean variable $Pending$ which is
\true if the directory is busy processing a request from an agent
pointed to by another variable $CurSrc$ (name changed from original
protocol for ease of presentation). However, the Flash protocol uses
two key optimizations over the German protocol. First, the Flash
protocol enables the cache agents to directly forward data between
each other instead of via the directory, for added speed. This is
accomplished by the directory by forwarding incoming requests from the
agent $i$ to the destination agent, $FwDst(i)$, with the relevant
data. Second, the Flash protocol uses non-blocking invalidates, i.e,
the \Exclusive flow does not have to wait for the \Invalidate flow to
complete for the sharing agents in $ShrSet$. Due to these
optimizations, the flows of the Flash protocol are significantly more
complex than those of German protocol. Further, due to forwarding,
some rules involve two agents instead of one for the German protocol:
thus the flows involve two agents as well. Each flow then is of the
form $\flow_k(i,j)$, where $i$ is the requesting agent for a flow and
$j=FwDst(i)$ is the destination agent to which the request may be
forwarded by the directory. Then, we define $\AgentRuleSet(i)$ to be
equal to $\bigcup_k\flow_k(i,FwDst(i))$.
We derived the invariants from the flows by keeping $c$ to be equal to
3, as each request encompasses a maximum of 2 agents (forwarding and
invalidation do not happen simultaneously in a flow). The final
invariants derived using our method are as follows:
{\bf \emph{Directory Not Busy:}} If the directory is not busy (i.e.,
$Pending$ is \false), any agent $i$ can send a request. Thus the
invariant \Inv{invF-1}: \(\lnot(Pending) \Rightarrow \big(\forall i
\in \IndexSet: \g(\AgentRuleSet(i))\big).\)
However, if the directory is busy (i.e., $Pending$ is \true), two
possibilities arise. (1) It may be busy since it is processing a
request from agent $CurSrc$. Or, (2) in case the request from $CurSrc$
requires an invalidate, the directory may remain busy with
invalidation even after the request from $CurSrc$ has been
served. This is because Flash allows the request from $CurSrc$ to
complete before invalidation due to non-blocking invalidates. Hence
the following invariants:
{\bf \emph{Directory Busy with Request:}} Invariant \Inv{invF-2}:
\(\big((Pending) \land (ShrSet=\{\})\big) \Rightarrow \big(\forall i
\in \ISet{\Inv{invF-2}} \g(\AgentRuleSet(i))\big),\) where
$\ISet{\Inv{invF-2}} = \{i| \, (i \in \IndexSet) \land (i=CurSrc)\}$.
{\bf \emph{Directory Busy with Invalidate:}} Invariant \Inv{invF-3}:
\(\big((Pending) \land \lnot (ShrSet=\{\})\big) \Rightarrow
\big(\forall i \in \ISet{invF-3} \g(\AgentRuleSet(i))\big),\) where
$\ISet{\Inv{invF-3}} = \{i| \, (i \in \IndexSet) \land (i \in
ShrSet)\}$.
{\bf \emph{Runtime:}} We verified the above invariants
along with the mutual exclusion and the data integrity properties for
an unbounded model abstracted by keeping 3 concrete agents (one agent
behaves as a directory) and constructing an environment agent
\Other. The verification took 5127s with about 20.5M states and 152M
rules fired. In this case we reused the lemmas used in prior work by
Chou \etal~\cite{CMP} for verifying the mutual exclusion and data
integrity properties in order to refine the agent \Other.
{\bf \emph{Verifying Flash vs German Protocol:}} \revised{This
sub-section} The flows of the Flash protocol involve two indices: we
eliminated the second index by replacing it with the variable
$FwDst(i)$ which stores information of the forwarded cache and thus
made the verification similar to the German protocol case. Next, Flash
protocol uses lazy invalidate: even if the original request has
completed, the directory may still be busy with the invalidate. As
explained above, this was in contrast to the German protocol and
resulted in an additional invariant \Inv{invF-3}.
{\bf \emph{Comparison with Other Techniques:}} \revised{This
sub-section} The only technique we are aware of which handles Flash
with a high degree of automation is by Bingham
\etal~\cite{DeadlockBingham}. While a direct comparison of the runtime
between their approach and ours is infeasible for this paper, we note
that the invariants generated using our approach only require an
over-abstraction in contrast to theirs which requires a
mixed-abstraction. This is an advantage since development of automatic
and scalable over-abstraction based parameterized safety verification
techniques is a promising area of ongoing research
(e.g. \cite{ProtocolFlashAmitGoel}) which our approach directly
benefits from.
\nop{ {\bf \emph{Verification Experience:}} The Flash protocol
presented us with unique challenges during verification. First, the
usage of forwarding made reasoning about Flash hard: as discussed
above, $\AgentRuleSet(i)$ is single index (denotes requests sent by
agent $i$). However, since, some of the transitions in
$\AgentRuleSet(i)$ are two index (second index corresponds to
$FwDst(i)$), the above invariants really are two index invariants,
i.e., of the form $\forall i,j: \phi(i,j)$. We had to take this into
account for verifying for unbounded number of agents.
Next, as discussed above, due to the non-blocking invalidate, when the
failure happens due to an invalidate in progress, the conflict
condition $conf$ is still $Pending$; we used $sharer=null$ as $conf$,
where $sharer$ is an auxiliary variable. This is because, unlike
German protocol, Flash sends the grant without waiting for invalidate
to complete. However, the directory keeps $Pending$ \true and waits
for invalidate to complete, before serving fresh requests, thus
blocking multiple flows and making the identification of $conf$
difficult.
}
\nop{
The request from the first agent, may get forwarded to the
second agent. We note that in this case, the second agent is pointed
to by an auxiliary pointer $CurDst$.
Thus, for showing \mdeadlock freedom, we start with initially: \(\true
\Rightarrow \forall i \in \IndexSet,j \in \IndexSet:
\g(\AgentRuleSet(i,j))\). Since a maximum of two agents are involved in
the flows for the Flash protocol, we kept $c$ to be equal to $3$.
On model checking, invariant inv-1 failed trivially due to the trivial
case that the agent $j$ gets randomly assigned. So we split the invariant indicating
that if CurDst is not null, $j$ must be $CurDst$.
--- \(\lnot (CurDst = null) \Rightarrow \forall i \in \IndexSet,(j = CurDst):
\g(\AgentRuleSet(i,j))\)
and
\((CurDst = null) \Rightarrow \forall i \in \IndexSet, j \in
\IndexSet: \g(\AgentRuleSet(i,j))\)
Then, the invariant failed because of pending. So, new invariant
became \((CurDst = null) \land (pending) \Rightarrow \forall i =CurSrc, j
\in \IndexSet: \g(\AgentRuleSet(CurSrc,j))\)
and
--- became \((CurDst = null) \land (!pending) \Rightarrow \forall i, j
\in \IndexSet: \g(\AgentRuleSet(i,j))\)
Next, failed because agent sharer was getting invalidated. Conflict condition
tricky as invalidate is lazy - so only pending is true. So another check that
if there is a sharer, it must be getting invalidated. So split.
--- \((CurDst = null) \land (pending) \land (sharer=null)\Rightarrow \forall i =CurSrc, j
\in \IndexSet: \g(\AgentRuleSet(CurSrc,j))\)
--- \((CurDst = null) \land (pending) \land \lnot(sharer=null)\Rightarrow \forall i = Sharer, j
\in \IndexSet: \g(\AgentRuleSet(CurSrc,j))\)
}
\nop{
and we iterated on refining
the invariant, as in \Sec{Bla}[FILL]. The counterexamples in
refinement showed: (1) that the directory points to the agent whose
flow is being processed using the pointer $CurSrc$~\footnote{Name
changed from publicly available protocol for clarity.} and (2) in
case of requests from $CurSrc$ which directly get forwarded to another
agent, the destination agent is stored in $CurDst$. Finally, the
sharer of the cache line is stored in the variable $Sharer$
Thus, the final generated invariants are:
\begin{itemize}
\item $CurSrc=null \rightarrow \forall i,j: \g(i,j)$
\item $\lnot(CurSrc=null) \land \lnot(CurDst = null) \rightarrow \g(CurSrc,CurDst)$
\item $\lnot(CurSrc=null) \land (CurDst) = null) \land (Sharer = null)$ \\
$\rightarrow \forall j: \g(CurSrc,j)$
\item $\lnot(CurSrc=null) \land (CurDst) = null) \land \lnot(Sharer = null)$ \\
$\rightarrow \forall j: \g(Sharer,j)$
\end{itemize}
}
\nop{
The first invariant states that if the forward destination
stored in $CurDst$ is not $null$, the flow with source $CurSrc$ and
destination $CurDst$ must be enabled. This is stated as the
invariant:
\[\lnot(CurSrc=null) \land \lnot(CurDst = null) \rightarrow
\g(CurSrc,CurDst)\tag{inv-1.2.1}.\]
\[\lnot(CurSrc=null) \land (CurDst) = null) \land
(Sharer = null) \\ \rightarrow \forall j:
\g(CurSrc,j)\tag{inv-1.2.2.1}.\]
\[\lnot(CurSrc=null) \land (CurDst) = null) \land \lnot(Sharer = null) \\
\rightarrow \forall j: \g(Sharer,j)\tag{inv-1.2.2.2}.\]
}
\nop{
\begin{enumerate}
\item The first invariant states that if $CurSrc$ is $null$, then some
flow for any source agent $i$ and destination agent $j$ is
enabled. This was stated as: \[CurSrc=null \rightarrow \forall i,j:
\g(i,j)\tag{inv-1.1}.\]
\item The second invariant states that if $CurSrc$ is not null, then
some flow with source $CurSrc$ and any destination $j$ is
enabled. This was stated as:
\[\lnot CurSrc=null \rightarrow \forall j: \g(CurSrc,j)\tag{inv-1.2}.\]
\end{enumerate}
On model checking, invariant inv-1.2 failed, as many flows have a
single destination agent and thus for other destination agents, no
flow may be enabled. Information about the destination node to which
a request is sent (or forwarded) to, is stored in the variable
$CurDst$. Then, we split inv-1.2 into two invariants as follows:
\begin{enumerate}
\item The first invariant states that if the forward destination
stored in $CurDst$ is not $null$, the flow with source $CurSrc$ and
destination $CurDst$ must be enabled. This is stated as the
invariant:
\[\lnot(CurSrc=null) \land \lnot(CurDst = null) \rightarrow
\g(CurSrc,CurDst)\tag{inv-1.2.1}.\]
\item The second invariant states that if the forward destination
value stored in $CurDst$ is $null$, all flows with source $CurSrc$
with any destination must be enabled. This is specified as the
invariant:
\[\lnot(CurSrc=null)\land (CurDst = null) \rightarrow \forall
j: \g(CurSrc,j)\tag{inv-1.2.2}.\]
\end{enumerate}
On model checking the above invariants, the invariant inv-1.2.2
failed: as in this case, at times an \Invalidate is happening to
another agent and thus the request is not enabled. The last added
sharer for the cache line is pointed to by the witness variable
$Sharer$. Then, we split the invariant inv-1.2.2 into two invariants:
\begin{enumerate}
\item The first invariant states that if the last sharer which needs
to be invalidated is $null$ then any flow with source $CurSrc$ must
be enabled. This is stated as the invariant:
\begin{align*}
\hspace*{-.4cm}\lnot(CurSrc=null) \land (CurDst) = null) \land
(Sharer = null) \\ \rightarrow \forall j:
\g(CurSrc,j)\tag{inv-1.2.2.1}.
\end{align*}
\item The second invariant states that if the last sharer is not
$null$, then the some flow (i.e. the \Invalidate flow) for the last
sharer must be enabled. This is stated as the invariant:
\begin{align*}
\hspace*{-.4cm} \lnot(CurSrc=null) \land (CurDst) = null) \land \lnot(Sharer = null) \\
\rightarrow \forall j: \g(Sharer,j)\tag{inv-1.2.2.2}.
\end{align*}
\end{enumerate}
On model checking with 3 agents, invariants inv-1.1, inv-1.2.1,
inv-1.2.2.1, inv-1.2.2.2 were verified to be $true$ for the Flash
protocol. Since these agents involve a small number of indices, they
were verified for an unbounded number of indices using the CMP method.
}
\nop{
Next, in order to do a sanity check and see if non-interference lemmas
are required for verifying \Enabled alone, we verified just the
\Enabled property (without mutual exclusion and data integrity
properties). In this case we had to add some non-interference lemmas
constraining the interference from the abstract agent (not detailed
for brevity). However, since properties like data integrity were not
being checked, the agent had to be constrained lesser. Thus, the
verification took 48230 sec to finish with 99M states explored and
1935M rules fired.
}
\nop{
Verification of flash protocol for unbounded agents was done using the
CMP method. [Non-interference lemmas?]
----non-interference lemmas same as with mutex.
\subsubsection{Runtime}
The running time for the verification of the flash protocol was
-- with mutual exclusion
20556798 states, 152413697 rules fired in 4830.49s.
-- without mutual exclusion
98844472 states, 1935368444 rules fired in 48230.24s.
}
\nop{
\section{Discussion: Livelock verification for German Protocol}
We now discuss how our approach can be extended for proving livelock
freedom for the German protocol example. The flows for the German
protocol have characteristics which simplify the definition and
verification of livelock--we leave a general definition and
verification of livelock freedom to future work.
The flows for German protocol have following key properties:
\begin{enumerate}
\item The flows for German protocol are acyclic. This property is also
true for flows of the protocols we have seen so far.
\item All flows for the German protocol, on completion, map to a
request successfully served. This is not $true$ for other protocols
in general: for flash for example, flows may finish due to a
transition representing failure to grant permission and so a
high-level request is not served.
\end{enumerate}
Due to the above key properties, it is straightforward to see that if,
at all times, progress can happen along a flow, i.e. some guard of
some flow is always enabled and on firing, the next message in the
flow is the one which can become enabled, then, livelock can not
exist.
First, we strengthen the definition of a flow. For simplicity, we
define a flow to be a total order of messages (in general it is a
partial order however this is enough for German flow).
Thus, $fl$: $\{rl_1,rl_2,\ldots,rl_n\}$, where $rl_{i+1}$ must happen
after $rl_i$. Further, we say that a flow $fl$ is in state $i$ if the
rule $fl_i$ is the only rule which can be enabled. Thus the flow
starts from state 1.
Further, for each agent, a maximum of $n$ instances of flow $fl$ can
exist, due to bounded state.
Then, let $aux_{fl}$ track the state of a flow and let it be
updated in every rule $rl_i$:
1) progress for a flow:
$progress(fl(i))$ = $\bigwedge_i$ $(aux_{fl}=i)$ $\rightarrow$ $rl_i.g$
Then, progress of an agent $i$ is: $progress(i) = \bigvee_{fl(i)}
progress(fl(i))$, where $fl(i)$ is an enabled flow in $i$ (note that
there can be multiple enabled flow of the same type in $i$).
Then, we can show the progress property: i.e. that at all times, there
exists an agent, such that $progress(i)$ holds.
Claim: progress implies livelock freedom.
Intuition: Due to acyclic nature of flows, this is the case.
\subsubsection{Runtime}
The verification of the progress property for the German protocol took
about 97 sec, with 14.5M states explored and 9.2M rules fired.
}
\section{Protocols, Flows and \MDeadlock Freedom: Background}
\label{sec:formalModel}
\subsection{Preliminaries}
A protocol \Protocol(N) consists of $N$ symmetric cache agents, with
ids from the set $\IndexSet = \{1,2,3,\ldots,N\}$. We follow our prior
approach~\cite{ParamVerMurali} (which was inspired by the approach of
Kristic~\cite{ParamVerifCMPKristic}) in formalizing cache coherence
protocols.
{\em Index Variables:} The protocol uses index variables from the set
\IndexVarSet quantified over the set of index values
$\IndexSet$. Thus, if $i$ is an index variable, then $i$ takes values
from the domain $\IndexSet$.
{\em State Variables:} The state of the protocol consists of local and
global variables, shared between agents. Each of these type of
variables can either hold values from Boolean domain (variables with
values from generic finite domains can be represented as a set of
Boolean variables) or pointers which can hold agent ids. We represent
the Boolean variables in the global state as $G_B$, while the pointers
as $G_P$. The Boolean local state variables of each agent $i$ is
encoded as $L_B[i]$, while the local pointer variables as $L_P[i]$.
The pointer variables, thus, have values from the domain $\IndexSet
\cup \{null\}$, where $null$ represents that the variable does not
hold any index value.
{\em Rules: } The protocol consists of a set of rules written on the
state variables. Each rule $rl(i)$\footnote{We assume that each rule
consists of one index $i$ for ease of exposition, our method applies
to rules with more than one index as well.} can be written
as: \[rl(i): rl(i).\rho \rightarrow rl(i).a,\]
where, $i$ is a variable (referred to as {\em index} variable) with
values in \IndexSet, $rl(i)$ is the rule name, $rl(i).p$, the guard,
is a propositional formula on the global variables and local variables
of agent with index value stored in $i$, and $rl(i).a$ is the action
which updates the global variables and local variables of agent with
index value stored in $i$.
{\em Protocol:} A {\em protocol} is a state transition system
$(S,\Theta,T)$, where $S$ is the set of protocol states, $\Theta$ the
initial states of the protocol, and $T \subseteq S \times S$ is the
transition relation described using a set of rules. The protocol state
$s$, then, is a valuation of the variables in $V_S$ and $V_L[i]$, for
all $i$. Thus, there is a transition $\tau(i_v)$ from state $s$ to
$s'$ if there is a rule $rl(i)$ and value of index variable $i=i_v$,
s.t. $rl(i_v).\rho$ holds in $s$, and $s'$ is obtained by applying
$rl(i_v).a$ to $s$. In state $s$, we say that the rule $rl(i)$ is {\em
enabled} (for agent $i_v$) if the guard $rl(i_v).\rho$ is $true$.
When the enabled transition $\tau(i_v)$ is executed, we say that the
rule $rl(i)$ is {\em fired} for agent $i_v$.
{\em Note:} The protocol is symmetric since the set of local variables
for each agent is identical, and the rules, $rl(i)$ can be
instantiated for each agent as well.
{\em Protocol execution:} We assume interleaving semantics for the
protocol, i.e., at any state $s$, any of the enabled rules can fire
and when a rule is fired, its action is applied to update the state
atomically. An execution trace of the protocol from a state $s_0$ to a
state $s_n$ with $n$ transitions in between looks as follows: $s_0
\rightarrow^{\tau_0} \rightarrow^{\tau_1} \rightarrow^{\tau_2} \ldots
\rightarrow^{\tau_{n-1}} s_n$.
\nop{
{\em Transitions:} The transitions of the protocol are indexed over
$i$, \(\tau_1(i),\tau_2(i),\tau_3(i) \ldots \tau_k(i)\). In this
paper, we express the state transitions of the protocol using {\em
rules} (we use the words rule and transition interchangeably). A
rule $rl(i)$ is a guard-action pair which is written as $rl(i).\rho
\Rightarrow rl(i).a$, where the guard $rl(i).\rho$ (also referred to
as $\rho(i)$ for brevity) is a Boolean expression on variables in
$V_L[i]$ and $V_S$, and the action $rl(i).a$ (also referred to as
$a(i)$ for brevity) is an update to variables in $V_L[i]$ and
$V_S$. We say that a rule is {\em enabled} in a state $s$ if its guard
$\rho$ evaluates to $true$ in $s$, else the rule is {\em disabled}. An
enabled rule is said to have fired when it applies its action $a$ to
update the state. When a rule fires, the corresponding transition
occurs.
}
\subsubsection{\MDeadlock Definition}
We define a protocol state $s$ to be a {\em \mdeadlock state} if no rule
in that state is enabled. Then a protocol is \mdeadlock free if in all
states, there exists at least one transition which is enabled. This
can be expressed as the invariant:
\[\bigvee_{i} \bigvee_{rl(i)} rl(i).\rho,\]
i.e., the disjunction of guards of all rules of all agents $i$ must be
$true$ in all reachable states.
\subsubsection{Flows}
Flows describe the basic organization of rules for implementing the
high-level requests in a protocol (for example a request for \Shared
or \Exclusive access or an \Invalidate). We model a flow $fl(i)$ as a
set of rules of the form $\{$$rl_1(i), rl_2(i), \ldots rl_n(i)$$\}$
which accomplish a high-level request\footnote{For ease of exposition
we assume that the guard and action of a rule are over the variables
of a single agent. Thus, a flow containing such rules also involves
a single agent. In general, a rule and thus a flow can involve a
larger but fixed number of interacting agents as well. Our approach
can easily be generalized to that case.}. These rules are ordered,
i.e., for rule $rl_k(i)$ to fire, all rules before that in the flow
must have already been fired. Then, given an execution trace, flows
naturally induce an ordering condition on the trace s.t. a rule
$rl_k(i)$ can fire only if all rules $rl_1(i) \ldots rl_{k-1}(i)$ have
fired. Formally, we denote this by $rl_k(i).pre^{fl}$ which holds on
a trace if the rules preceding $rl_k(i)$ have all fired.
A protocol essentially then consists of a set of flows $fl_1, fl_2,
fl_3, \ldots, fl_k$. In order to process a high-level request, a
protocol may use a combination of flows. As an example, in order to
execute a request for exclusive access, the German protocol uses the
\Exclusive and \Invalidate flows.
\subsubsection{Some definitions:}
Given a set of transitions $T$, the operator $\g(T)$ is $true$ if at
least one transition in the set of transitions $T$ must be enabled,
else it is $false$. Thus, it is a disjunction of the guards of
transitions in $T$. Then, we say an agent $i$ is enabled if at least
one of its rules is enabled, i.e., $\g(i)$ holds. Next, we say that a
flow is enabled if at least one of its rules is enabled, i.e.,
$\g(fl(i))$ holds.
Finally, in case a flow $fl(i)$ is not enabled, we say it is {\em
blocked} on rule $rl(i)$ if the precondition of the rule
$rl(i).pre^{fl}$ holds but the guard of the rule $rl(i).\rho$ is
$false$.
\subsection{German Protocol Implementation}
\nop{The German protocol implements 3 high-level requests: (1) an
\Exclusive request initiated by an agent for getting a read or a write
permission (agent state with read and write permission is denoted by
$E$), (2) a \Shared request for getting a read only permission (agent
state denoted by $S$), and (3) an \Invalidate request initiated by the
directory to take away all the permissions from an agent (agent state
with no permissions is denoted by $I$).
}
The German protocol consists of agents which can have \Exclusive
($E$), \Shared ($S$) or \Invalid ($I$) access to a cache line: the
agents request these access rights by sending messages to a shared
directory. The directory is modeled as a set of global variables. The
directory serves one agent at a time: it stores the id of the agent
being served in the variable $CurPtr$. It also stores the nature of
request in the variable $CurCmd$ with values in $\{ReqE,ReqS,Empty\}$,
where $ReqE$ represents a request for \Exclusive access, $ReqS$ for
\Shared and $Empty$ for no request. Finally, the directory tracks if
\Exclusive access is granted to some agent or not using the variable
$ExGntd$: it is $true$ if access is granted and $false$ otherwise.
The code for the implementation of the \Exclusive request is shown in
\Fig{Fig:ExclusiveCode}\footnote{For increasing clarity, compared to
the original code presented in Appendix~\ref{AppendixGermaCode},
some variable names have been changed as well as some details not
relevant to the discussion have been skipped. Note that variable
names start with capital letters for the sake of consistency with
the original code.}.
\input{flowCode.tex}
In processing the \Exclusive request, before sending the grant
$SendGntE(i)$, the directory checks if there are any sharers (by
checking $ShrSet$ = $\{\}$). If there are sharers, the \Invalidate
flow is invoked for each agent in $ShrSet$. Upon invalidation of all
agents in $ShrSet$, the $ShrSet$ becomes empty and so the
$SendGntE(i)$ rule becomes enabled for execution.
We show the code for the $SendInv(i)$ rule below. (The complete code
of the protocol is provided in the Appendix~\ref{AppendixGermaCode}
for the interested reader.) We denote the condition for invoking
\Invalidate (i.e. enabling the first rule of the \Invalidate flow
$SendInv(i)$) by $Inv\_Cond \land (ShrSet \neq null)$, where,
\(Inv\_Cond: CurCmd = ReqE \lor (CurCmd=ReqS \land ExGntd=true)\) can
be identified from the guard of the rule $SendInv(i)$.
\vspace{.3cm}
\begin{minipage}{\textwidth}
\footnotesize
\Ruleset{$\forall$ i : \IndexSet;}{SendInv(i)}{InvChannel[i].cmd =
Empty $\land$ i $\in$ ShrSet $\land$ InvSent[i] = false
\hspace*{.4cm} $\land $ ((CurCmd = ReqE) $\lor$ (CurCmd = ReqS $\land$ ExGntd =
true))}{InvChannel[i].cmd = Inv; InvSent[i] = true;}
\end{minipage}
\nop{
The German protocol models a directory as a set of global variables.
The directory stores the index value of the agent whose request is
currently being served in a global variable $CurPtr$ with values in
$\{\{null\} \cup [1,N]\}$. $CurPtr=i$ if request of agent $i$ is being
served and $CurPtr=null$ if no request is being served. Next,
information about the current request being served by the directory is
encoded in the variable $CurCmd$ with values in $\{ReqE,ReqS,Empty\}$,
where $ReqE$ represents a request for \Exclusive access, $ReqS$ for
\Shared access and $Empty$ for no request. The directory also
maintains a list of agents which share a particular cache line,
denoted by $ShrSet$. Finally, the directory tracks if \Exclusive
access is granted to some agent or not using the Boolean variable
$ExGntd$: it is $true$ if access is granted and $false$ otherwise.
Each agent $i$ stores its current status in local variable
$Cache[i].State$ with values in ${I,S,E}$, where $I$, $S$ and $E$
stand for $Invalid$, $Shared$ and $Exclusive$, respectively. Next, the
agents and directory communicate with each other through channels: the
implementation of the German protocol uses separate channels for
separate requests. Agent $i$ sends requests for \Shared or \Exclusive
access to the directory on $ReqChannel[i]$. The directory in turn
sends grants to agent $i$ using the channel $GntChannel[i]$. Finally,
the directory sends \Invalidate to agents through the channel
$InvChannel[i]$: in order to track that an \Invalidate request is sent
to a particular agent $i$, it keeps a variable $InvSent[i]$ which is
$true$ if the request has been sent and $false$ otherwise.
\input{flowCode.tex}
\paragraph{Code for \Exclusive request}
\Fig{Fig:ExclusiveCode} shows the relevant code details\footnote{For
increasing clarity, compared to the original code presented in
Appendix~\ref{AppendixGermaCode}, some variable names have been
changed as well as some details not relevant to the discussion have
been skipped. Note that variable names start with capital letters
for the sake of consistency with the original code.} for the
implementations of the $SendReqE(i)$, $RecvReqE(i)$, $SendGntE(i)$ and
$RecvGntE(i)$ rules of the \Exclusive flow (flow shown in
\Fig{figGermanFlowsExcl}). Notice that the guard of the rule
$SendGntE(i)$ checks if there are no sharers of the cache line
($ShrSet$ = $\{\}$). Rest of the guards and actions of the rules in
the figure are self-explanatory.
\paragraph{Code for \Invalidate request}
For the implementation of the \Invalidate request (flow shown in
\Fig{figGermanFlowsInv}), for the sake of brevity, we only provide the
code for the rule $SendInv(i)$ which implements an \Invalidate request
initiated by the directory for a sharer:
\vspace{.3cm}
\begin{minipage}{.3\textwidth}
\footnotesize
\Ruleset{$\forall$ i : \IndexSet;}{SendInv(i)}{InvChannel[i].cmd =
Empty $\land$ i $\in$ ShrSet $\land$ InvSent[i] = false
\hspace*{.4cm} $\land $ ((CurCmd = ReqE) $\lor$ (CurCmd = ReqS $\land$ ExGntd =
true))}{InvChannel[i].cmd = Inv; InvSent[i] = true;}
\end{minipage} \vspace{.2cm}
The guard of this rule checks the following two key conditions: (1)
first, it checks if the current request for \Shared or \Exclusive
access requires the directory to initiate an \Invalidate request to
the agents sharing the cache line. This is true when the current
request is either for an \Exclusive access or, is for a \Shared access
with another agent already having an \Exclusive access. This is stated
as the condition: \(Inv\_Cond: CurCmd = ReqE \lor (CurCmd=ReqS \land
ExGntd=true).\) (2) And second, it checks whether there is an agent
$i$ which needs to be invalidated, by checking $(i \in ShrSet)$.
}
\nop{
This rule checks if the channel for sending invalidates
($invChannel[i]$) is empty, that $i$ is in the set of sharers
($ShrSet$), that invalidate has not yet been sent (tracked using
$invSet[i] = false$) and that the current request is either exclusive
or it is for shared but some other cache has exclusive access
(i.e. $((CurCmd = ReqE)$ $\lor$ $((CurCmd = ReqS)$ $\land$ $(ExGntd =
true)))$, denoted by $invCond$).
}
\nop{
The protocol code snippets implementing the Exclusive and Invalidate
requests for the German protocol are shown in \Fig{Fig:ExclusiveFlow}
and \Fig{Fig:InvalidateFlow} respectively\footnote{The complete code
is in the appendix.}.
\subsubsection{Flows}
The German protocol consists of three flows: \Exclusive flow, \Shared
flow and \Invalidate flow. The \Exclusive flow processes a request for
write access, \Shared for read access and the \Invalidate flow for
invalidating some cache. The \Invalidate flow is a sub-flow of \Shared
and \Exclusive flows. These flows are detailed as follows:
\begin{itemize}
\item{\Exclusive:} {\bf ReqExclusive(i)}: $SendReqE(i)$, $RecvReqE(i)$,
$SendGntE(i)$, $RecvGntE(i)$.
\item{\Shared:} {\bf ReqShared(i)}: $SendReqS(i)$, $RecvReqS(i)$,
$SendGntS(i)$, $RecvGntS(i)$.
\item{\Invalidate:} {\bf SendInval(i)}: $SendInv(i)$, $SendInvAck(i)$,
$RecvInvAck(i)$.
\end{itemize}
\subsubsection{Protocol Code}
The code snippets for the Exclusive and Invalidate flows of the German
protocol is presented in Murphi language~\cite{} in
\Fig{Fig:ExclusiveFlow} and \Fig{Fig:InvalidateFlow}
respectively\footnote{The complete code is in the appendix.}. These
code snippets, along with global state are explained below.
\nop{
\paragraph{Murphi Syntax:} Murphi language consists of {\em rulesets}
which is a collection of {\em rules}, where each rule is a guarded
action. For example, $SendReqS$ is a ruleset which represents a set of
rules $1..N$, where rule $i$ (denoted by $SendReqS[i]$) corresponds to
$i^{th}$ cache sending a request for shared access. The guard of
$SendReqS[i]$ is $Chan1[i].Cmd = Empty \land Cache[i].State = S$ and
the action is $Chan1[i].Cmd := ReqS$. {\em Observe that each message
in the flows shown above correspond to a ruleset in the code.}
\paragraph{Murphi Semantics:} Each rule is executed atomically. Thus, at any time,
Murphi non-deterministically selects a rule which has its guard true
and executes it.
}
\paragraph{Relevant Global State:} The directory points to the cache
whose request is currently being served (for a particular cache line)
using the pointer \curptr. Further, it stores the request type from
\curptr in \curcmd. Next, \exgntd is set to $true$ if some cache has
exclusive access, else it is set to false. Finally, a set of
variables, \shrset, stores the caches which share a particular cache
line. Thus, \shrset[i] is true if cache $i$ has a copy of the
line. Finally, an auxiliary variable $cache\_inv$ (FIX: not shown in
code) points to some cache $c$ such that \shrset[c] is
true---$cache\_inv$ is null if and only if \shrset is false for all
caches.
\paragraph{Code for Exclusive access}
\paragraph{Code for Invalidate access}
\input{flowCode.tex}
\nop{
The flows for the German protocol are shown in
\Fig{figGermanFlows}. These flows represent the order in which various
messages in the protocol are passed, in order to accomplish a high
level request. Thus, they provide high level information about the
protocol. The German protocol consists of the following flows:
\begin{itemize}
\item \Shared: The \Shared flow, shown in \Fig{figGermanFlowsShrd}
processes a request from cache $i$ for shared access. It consists of
messages \{$sendReqS$, $RecvReqS$, $SendGntS$, $RecvGntS$\} passed
between the requesting cache and directory.
\item \Exclusive: Similar to shared flow, shown in
\Fig{figGermanFlowsExcl}.
\item \Invalidate: The \Invalidate flow is a sub-flow of \Shared and
\Exclusive flows: it has rules \{$sendInv$, $sendInvAck$,
$recvInvAck$\}. It is present in both figures,
\Fig{figGermanFlowsShrd} and \Fig{figGermanFlowsExcl}.
\end{itemize}
}
}
\section{Protocols, Flows and \MDeadlock Freedom: Background}
\label{sec:formalModel}
\subsection{Preliminaries}
A protocol \Protocol(N) consists of $N$ symmetric cache agents, with
{\it ids} from the set $\IndexSet = \{1,2,3,\ldots,N\}$. We follow our
prior approach~\cite{ParamVerMurali} (which was inspired by the
approach of Kristic~\cite{ParamVerifCMPKristic}) in formalizing cache
coherence protocols.
{\em Index Variables:} The protocol uses index variables quantified
over the set of index values $\IndexSet$. Thus, if $i$ is an index
variable, then $i$ takes values from the domain $\IndexSet$.
{\em State Variables:} The state of the protocol consists of local
variables, and global variables shared between the agents. Each of
these types of variables can either hold values from the Boolean
domain (variables with values from generic finite domains can be
represented as a set of Boolean variables) or pointers which can hold
agent {\it ids}. We represent the Boolean variables in the global
state as $G_B$, and the pointers as $G_P$. The Boolean local state
variables of each agent $i$ are encoded as $L_B[i]$, and the local
pointer variables as $L_P[i]$. The pointer variables have values from
the domain $\IndexSet \cup \{null\}$, where $null$ represents that the
variable does not hold any index value.
{\em Expressions:} An expression is a, possibly quantified,
propositional formula with atoms $G_B$, $G_P=j$, $L_B[i]$ and
$L_P[i]=j$, where, $i$ and $j$ are index variables.
{\em Assignments:} Assignments are of the form $G_B := b$, or
$G_P:=j$, $L_B[i] := b$ or $L_P[i]:=j$, where, $b$ is a variable with
Boolean value and $i$, $j$ are index variables.
{\em Rules: } Each agent $i$ consists of a set of rules
$rl_1(i),rl_2(i),rl_3(i),\ldots,rl_k(i)$. Each rule $rl_j(i)$ can be
written as: \(rl_j(i): rl_j(i).\rho \rightarrow rl_j(i).a,\) where,
$rl_j(i)$ is the rule name, $rl_j(i).\rho$, the guard, is an
expression, and $rl_j(i).a$ is a list of assignments, such that these
assignments are restricted to only update the global variables or the
local variables of agent $i$. The local variables and rules for all
agents $i$ are symmetric.
{\em Protocol:} The above defined variables and rules naturally induce
a state transition system. A {\em protocol}, then, is a state
transition system $(S,\Theta,T)$, where $S$ is the set of protocol
states, $\Theta \subseteq S$ is the set of initial states, and $T
\subseteq S \times S$ is the transition relation. Each protocol state
$s \in S$ is a valuation of the variables $G_B$, $G_P$, and $L_B[i]$,
$L_P[i]$ for each agent $i$. There exists a transition
$\tau(i_v)=(s,s'), (s,s')\in T$ from state $s$ to $s'$ if there is a
rule $rl_j(i)$ and value of index variable $i=i_v$,
s.t. $rl_j(i_v).\rho$ holds in $s$, and $s'$ is obtained by applying
$rl_j(i_v).a$ to $s$. In state $s$, we say that the rule $rl_j(i)$ is
{\em enabled} for agent with {\it id} $i_v$ if the guard
$rl_j(i_v).\rho$ is \true. When the enabled rule is executed, its
action is applied to update the state and we say that the rule
$rl_j(i)$ has {\em fired} for agent $i_v$. The action is applied
atomically to update the state, thus the transitions of the protocol
have interleaving semantics. Finally, we define an execution {\em
trace} of the protocol as a series of transitions where each
transition is a fired rule. Thus, a trace can be represented by a
series ($rl_a(i_0)$, $rl_b(i_1)$, \ldots, $rl_s(i_k)$), where the
transition $rl_m(i_n)$ is the rule $rl_m$ fired for the agent with \id
$i_n$.
\nop{
{\em Transitions:} The transitions of the protocol are indexed over
$i$, \(\tau_1(i),\tau_2(i),\tau_3(i) \ldots \tau_k(i)\). In this
paper, we express the state transitions of the protocol using {\em
rules} (we use the words rule and transition interchangeably). A
rule $rl(i)$ is a guard-action pair which is written as $rl(i).\rho
\Rightarrow rl(i).a$, where the guard $rl(i).\rho$ (also referred to
as $\rho(i)$ for brevity) is a Boolean expression on variables in
$V_L[i]$ and $V_S$, and the action $rl(i).a$ (also referred to as
$a(i)$ for brevity) is an update to variables in $V_L[i]$ and
$V_S$. We say that a rule is {\em enabled} in a state $s$ if its guard
$\rho$ evaluates to $true$ in $s$, else the rule is {\em disabled}. An
enabled rule is said to have fired when it applies its action $a$ to
update the state. When a rule fires, the corresponding transition
occurs.
}
{\bf \MDeadlock Definition}
We define a protocol state $s$ to be an {\em \mdeadlock state} if no
rule in that state is enabled. Then, a protocol is \mdeadlock free if
in all states, there exists at least one rule which is enabled. This
can be expressed as the invariant: \(\bigvee_{i} \bigvee_{j}
rl_j(i).\rho\), i.e., the protocol is \mdeadlock free if the
disjunction of the guards of all the rules of all the agents is \true
for all the reachable states.
{\bf Flows} Flows describe the basic organization of rules for
implementing the high-level requests in a protocol (for example a
request for \Exclusive access or an \Invalidate). We model a flow as a
set of rules $\flow(i)$ of the form $\{$$rl_a(i), rl_b(i), rl_c(i),
\ldots, rl_n(i)$$\}$ which accomplish a high-level request of agent
$i$.\footnote{For ease of exposition we assume that the guard and
action of a rule are over the variables of a single agent. Thus, a
flow containing such rules also involves a single agent. In general,
a rule and thus a flow can involve a larger but fixed number of
interacting agents as well. Our approach can be easily generalized
to that case.} The rules in a flow are partially ordered, with the
partial order relation denoted as \partialorder{\flow(i)}. For
example, in the \Exclusive flow in \SubFig{figGermanFlowsExcl}, the rules
(arrows) are totally ordered along the downward direction. Thus
$SendReqE(i)$ $\partialorder{\flow_E(i)}$ $RecvReqE(i)$, where
$\flow_E$ denotes the set of rules for \Exclusive flow. For every rule
$rl_k(i)$ in the flow $\flow(i)$, the partial order naturally induces
the following {\em precondition}: for the rule $rl_k(i)$ to fire, all
the rules preceding that rule in the partial order of the flow
$\flow(i)$ must have already been fired. This precondition is denoted
by $\precon{\flow(i)}{rl_k}{i}$ and, formally, can be written as:
\[\precon{\flow(i)}{rl_k}{i} = \forall j:\big(\{(rl_j(i) \in \flow(i)) \land
(rl_j(i) \partialorder{\flow(i)} rl_k(i))\} \Rightarrow (rl_j(i).fired
= true)\big),\] where $rl_j(i).fired$ is an auxiliary variable which
is initially set to \false when the flow $\flow(i)$ starts and is set
to \true when the rule $rl_j(i)$ has fired for that flow.
\nop{
Then,
given a rule $rl_k(i)$, the flow $fl(i)$ naturally induces a {\em
precondition} that for the rule to be enabled, all the rules
$rl_a(i), rl_b(i), rl_c(i) ,\ldots, rl_{k-1}(i)$ must have already
fired. Formally, we denote this by $rl_k(i).pre^{fl}$ which holds on a
state if the rules preceding $rl_k(i)$ have all fired.
}
\revised{This Para} Designs of protocols are presented in industrial
documents as a set of flows $\flow_1(i)$, $\flow_2(i)$, $\flow_3(i)$,
$\ldots$, $\flow_k(i)$. In order to process a high-level request, a
protocol may use a combination of these flows, e.g. in order to
execute a request for \Exclusive access the German protocol uses the
\Exclusive and \Invalidate flows. Each flow in a protocol represents
an execution scenario of the protocol for processing some high-level
request. Thus many of the flows of a protocol tend to exhibit a lot of
similarity as they are different execution scenarios of the same
high-level request. This makes them fairly easy to understand. In
\Sec{sec:SpecMethod}, we show how a set of invariants collectively
implying \mdeadlock freedom can be derived from these flows.
{\bf Some definitions:}
We define the union of all the flows of agent $i$ by
$\AgentRuleSet(i)$, i.e., $\AgentRuleSet(i)=\bigcup_{k}\flow_k(i)$.
Next, we define the operator $\g$ which is \true for a set of rules,
if at least one rule in the set is enabled, else it is \false. Thus,
for example, $\g(\AgentRuleSet(i))$ holds if at least one of the rules
in $\AgentRuleSet(i)$ is enabled. In this case, we say that the agent
$i$ is enabled. Similarly, we say that a flow $\flow(i)$ is enabled if
at least one of its rules is enabled, i.e., $\g(\flow(i))$ holds. In
case a flow $\flow(i)$ is not enabled, we say that it is {\em blocked}
on some rule $rl_j(i) \in \flow(i)$ if the precondition of the rule
$\precon{\flow(i)}{rl_j}{i}$ holds but the guard of the rule
$rl_j(i).\rho$ is \false.
\subsection{German Protocol Implementation}
\nop{The German protocol implements 3 high-level requests: (1) an
\Exclusive request initiated by an agent for getting a read or a write
permission (agent state with read and write permission is denoted by
$E$), (2) a \Shared request for getting a read only permission (agent
state denoted by $S$), and (3) an \Invalidate request initiated by the
directory to take away all the permissions from an agent (agent state
with no permissions is denoted by $I$).
}
The German protocol consists of agents such that each agent can have
\Exclusive ($E$), \Shared ($S$) or \Invalid ($I$) access to a cache
line, as stored in the variable $Cache[i].State$. An agent $i$
requests these access rights by sending messages on a channel
$ReqChannel[i]$ to a shared directory which sends corresponding grants
along the channel $GntChannel[i]$. The directory is modeled as a set
of global variables which serves one agent at a time: it stores the
{\it id} of the agent being served in the variable $CurPtr$. It also
stores the nature of the request in the variable $CurCmd$ with values
in $\{ReqE,ReqS,Empty\}$, where $ReqE$ represents a request for
\Exclusive access, $ReqS$ for \Shared and $Empty$ for no
request. Finally, the directory tracks if \Exclusive access is granted
to some agent or not using the variable $ExGntd$: it is \true if
access is granted and \false otherwise. \CameraReady{A
simplified version of the code for the \Exclusive request is shown
in \Fig{Fig:ExclusiveCode}.} \TechnicalVersion{A
simplified version of the code for the \Exclusive request is shown
in \Fig{Fig:ExclusiveCode}, with the original Murphi
implementation~\cite{CMP} presented in
Appendix~\ref{AppendixGermaCode}.}
\input{flowCode.tex}
In processing the \Exclusive request, before sending the grant
$SendGntE(i)$, the directory checks if there are any sharers of the
cache line (by checking $ShrSet$ = $\{\}$). If there are sharers, the
\Invalidate flow is invoked for each agent in $ShrSet$. Upon
invalidation of all the agents in $ShrSet$, the $ShrSet$ becomes empty
and so the $SendGntE(i)$ rule becomes enabled for execution. We show
the code for the $SendInv(i)$ rule below.
\Ruleset{$\forall$ i : \IndexSet;}{SendInv(i)}{InvChannel[i].cmd =
Empty $\land$ i $\in$ ShrSet $\land$
\hspace*{.4cm} ((CurCmd = ReqE) $\lor$ (CurCmd = ReqS $\land$ ExGntd =
true))}{InvChannel[i].cmd := Invalidate;}
We note a condition $Inv\_Cond$, which must be \true for invoking the
\Invalidate flow and can be identified from the guard of $SendInv(i)$;
\(Inv\_Cond: \big(\big((CurCmd = ReqE) \lor ((CurCmd=ReqS) \land
(ExGntd=true))\big) \land (ShrSet \neq \{\})\big) \).
\nop{
The German protocol models a directory as a set of global variables.
The directory stores the index value of the agent whose request is
currently being served in a global variable $CurPtr$ with values in
$\{\{null\} \cup [1,N]\}$. $CurPtr=i$ if request of agent $i$ is being
served and $CurPtr=null$ if no request is being served. Next,
information about the current request being served by the directory is
encoded in the variable $CurCmd$ with values in $\{ReqE,ReqS,Empty\}$,
where $ReqE$ represents a request for \Exclusive access, $ReqS$ for
\Shared access and $Empty$ for no request. The directory also
maintains a list of agents which share a particular cache line,
denoted by $ShrSet$. Finally, the directory tracks if \Exclusive
access is granted to some agent or not using the Boolean variable
$ExGntd$: it is $true$ if access is granted and $false$ otherwise.
Each agent $i$ stores its current status in local variable
$Cache[i].State$ with values in ${I,S,E}$, where $I$, $S$ and $E$
stand for $Invalid$, $Shared$ and $Exclusive$, respectively. Next, the
agents and directory communicate with each other through channels: the
implementation of the German protocol uses separate channels for
separate requests. Agent $i$ sends requests for \Shared or \Exclusive
access to the directory on $ReqChannel[i]$. The directory in turn
sends grants to agent $i$ using the channel $GntChannel[i]$. Finally,
the directory sends \Invalidate to agents through the channel
$InvChannel[i]$: in order to track that an \Invalidate request is sent
to a particular agent $i$, it keeps a variable $InvSent[i]$ which is
$true$ if the request has been sent and $false$ otherwise.
\input{flowCode.tex}
\paragraph{Code for \Exclusive request}
\Fig{Fig:ExclusiveCode} shows the relevant code details\footnote{For
increasing clarity, compared to the original code presented in
Appendix~\ref{AppendixGermaCode}, some variable names have been
changed as well as some details not relevant to the discussion have
been skipped. Note that variable names start with capital letters
for the sake of consistency with the original code.} for the
implementations of the $SendReqE(i)$, $RecvReqE(i)$, $SendGntE(i)$ and
$RecvGntE(i)$ rules of the \Exclusive flow (flow shown in
\Fig{figGermanFlowsExcl}). Notice that the guard of the rule
$SendGntE(i)$ checks if there are no sharers of the cache line
($ShrSet$ = $\{\}$). Rest of the guards and actions of the rules in
the figure are self-explanatory.
\paragraph{Code for \Invalidate request}
For the implementation of the \Invalidate request (flow shown in
\Fig{figGermanFlowsInv}), for the sake of brevity, we only provide the
code for the rule $SendInv(i)$ which implements an \Invalidate request
initiated by the directory for a sharer:
\vspace{.3cm}
\begin{minipage}{.3\textwidth}
\footnotesize
\Ruleset{$\forall$ i : \IndexSet;}{SendInv(i)}{InvChannel[i].cmd =
Empty $\land$ i $\in$ ShrSet $\land$ InvSent[i] = false
\hspace*{.4cm} $\land $ ((CurCmd = ReqE) $\lor$ (CurCmd = ReqS $\land$ ExGntd =
true))}{InvChannel[i].cmd = Inv; InvSent[i] = true;}
\end{minipage} \vspace{.2cm}
The guard of this rule checks the following two key conditions: (1)
first, it checks if the current request for \Shared or \Exclusive
access requires the directory to initiate an \Invalidate request to
the agents sharing the cache line. This is true when the current
request is either for an \Exclusive access or, is for a \Shared access
with another agent already having an \Exclusive access. This is stated
as the condition: \(Inv\_Cond: CurCmd = ReqE \lor (CurCmd=ReqS \land
ExGntd=true).\) (2) And second, it checks whether there is an agent
$i$ which needs to be invalidated, by checking $(i \in ShrSet)$.
}
\nop{
This rule checks if the channel for sending invalidates
($invChannel[i]$) is empty, that $i$ is in the set of sharers
($ShrSet$), that invalidate has not yet been sent (tracked using
$invSet[i] = false$) and that the current request is either exclusive
or it is for shared but some other cache has exclusive access
(i.e. $((CurCmd = ReqE)$ $\lor$ $((CurCmd = ReqS)$ $\land$ $(ExGntd =
true)))$, denoted by $invCond$).
}
\nop{
The protocol code snippets implementing the Exclusive and Invalidate
requests for the German protocol are shown in \Fig{Fig:ExclusiveFlow}
and \Fig{Fig:InvalidateFlow} respectively\footnote{The complete code
is in the appendix.}.
\subsubsection{Flows}
The German protocol consists of three flows: \Exclusive flow, \Shared
flow and \Invalidate flow. The \Exclusive flow processes a request for
write access, \Shared for read access and the \Invalidate flow for
invalidating some cache. The \Invalidate flow is a sub-flow of \Shared
and \Exclusive flows. These flows are detailed as follows:
\begin{itemize}
\item{\Exclusive:} {\bf ReqExclusive(i)}: $SendReqE(i)$, $RecvReqE(i)$,
$SendGntE(i)$, $RecvGntE(i)$.
\item{\Shared:} {\bf ReqShared(i)}: $SendReqS(i)$, $RecvReqS(i)$,
$SendGntS(i)$, $RecvGntS(i)$.
\item{\Invalidate:} {\bf SendInval(i)}: $SendInv(i)$, $SendInvAck(i)$,
$RecvInvAck(i)$.
\end{itemize}
\subsubsection{Protocol Code}
The code snippets for the Exclusive and Invalidate flows of the German
protocol is presented in Murphi language~\cite{} in
\Fig{Fig:ExclusiveFlow} and \Fig{Fig:InvalidateFlow}
respectively\footnote{The complete code is in the appendix.}. These
code snippets, along with global state are explained below.
\nop{
\paragraph{Murphi Syntax:} Murphi language consists of {\em rulesets}
which is a collection of {\em rules}, where each rule is a guarded
action. For example, $SendReqS$ is a ruleset which represents a set of
rules $1..N$, where rule $i$ (denoted by $SendReqS[i]$) corresponds to
$i^{th}$ cache sending a request for shared access. The guard of
$SendReqS[i]$ is $Chan1[i].Cmd = Empty \land Cache[i].State = S$ and
the action is $Chan1[i].Cmd := ReqS$. {\em Observe that each message
in the flows shown above correspond to a ruleset in the code.}
\paragraph{Murphi Semantics:} Each rule is executed atomically. Thus, at any time,
Murphi non-deterministically selects a rule which has its guard true
and executes it.
}
\paragraph{Relevant Global State:} The directory points to the cache
whose request is currently being served (for a particular cache line)
using the pointer \curptr. Further, it stores the request type from
\curptr in \curcmd. Next, \exgntd is set to $true$ if some cache has
\Exclusive access, else it is set to false. Finally, a set of
variables, \shrset, stores the caches which share a particular cache
line. Thus, \shrset[i] is true if cache $i$ has a copy of the
line. Finally, an auxiliary variable $cache\_inv$ (FIX: not shown in
code) points to some cache $c$ such that \shrset[c] is
true---$cache\_inv$ is null if and only if \shrset is false for all
caches.
\paragraph{Code for Exclusive access}
\paragraph{Code for Invalidate access}
\input{flowCode.tex}
\nop{
The flows for the German protocol are shown in
\Fig{figGermanFlows}. These flows represent the order in which various
messages in the protocol are passed, in order to accomplish a high
level request. Thus, they provide high level information about the
protocol. The German protocol consists of the following flows:
\begin{itemize}
\item \Shared: The \Shared flow, shown in \Fig{figGermanFlowsShrd}
processes a request from cache $i$ for shared access. It consists of
messages \{$sendReqS$, $RecvReqS$, $SendGntS$, $RecvGntS$\} passed
between the requesting cache and directory.
\item \Exclusive: Similar to shared flow, shown in
\Fig{figGermanFlowsExcl}.
\item \Invalidate: The \Invalidate flow is a sub-flow of \Shared and
\Exclusive flows: it has rules \{$sendInv$, $sendInvAck$,
$recvInvAck$\}. It is present in both figures,
\Fig{figGermanFlowsShrd} and \Fig{figGermanFlowsExcl}.
\end{itemize}
}
}
\section{Introduction}
We verify deadlock freedom for symmetric cache coherence protocols.
Consider a cache coherence protocol \Protocol(N) where the parameter
\(N\) represents an unbounded number of (cache) agents. The protocol
implements requests sent by the agents using messages exchanged in the
protocol. As a protocol designer, the holy grail for verification is
the request-response property: i.e. every request from an agent
eventually gets a response. Since this is a liveness property and hard
for existing model checking tools, designers then resort to verifying
simpler properties such as deadlock freedom to identify causes for
response property failure.
The literature is abundant with various definitions of
deadlock~\cite{}: we verify for deadlock errors in which the entire
protocol gets stuck, i.e., no agent of the protocol can make any
transition. We refer to such failures as maximal deadlock
(\mdeadlock): this kind of failures, while weaker than other broader
classes of deadlock failures, is commonly observed in industry and is
consistent with the internal definition for deadlock used by the
Murphi model checker as well~\cite{}.
Since an \mdeadlock error involves all the agents getting stuck and
unable to make any transition, verification of \mdeadlock freedom
naturally is a parameterized verification problem. Parameterized
verification techniques work by using sound abstractions to reduce the
unbounded model to a finite model that preserves the property of
interest. These abstractions typically tend to be simple
over-abstractions which keep a small number of agents ($1$ or $2$) as
is and replace all the other agents with an abstract environment: such
abstractions have had considerable success in verifying key safety
properties like mutual exclusion and coherence even for industrial
scale protocols~\cite{CMP, ParamVerifMuraliIndust, ParamVerMurali}.
While successful for properties like mutual exclusion and coherence,
the application of such abstractions for parameterized verification of
properties like \mdeadlock gets challenging. This is because an
\mdeadlock error involves all of the arbitrarily large number of
agents getting stuck and unable to make progress: simple
over-abstractions like above will not preserve this error since they
discard state of most of the agents.
In this paper we show how high-level information about these protocols
can be leveraged to construct invariants which are collectively
stronger than the \mdeadlock property. These invariants are amenable
to efficient abstractions like data type reduction which have been
used in the past in verifying industrial scale protocols.
\subsection{Deadlock in the German Protocol}
The German protocol is a popular benchmark for existing protocol
verification work in literature~\cite{DeadlockBingham, CMP}. The
protocol implements high-level requests for \Exclusive, \Shared, and
\Invalidate access to a cache line.
Architects informally describe protocols using special diagrams,
referred to as {\em flows}. These provide information about the
messages passed in the protocol implementation for processing
high-level requests. Flows are readily available in industrial
documents in the form of charts and tables.
\Fig{figGermanFlows} shows the \Exclusive and \Invalidate flows for
the German protocol. Thus, for example, the \Exclusive request from
agent (cache) $i$ to the directory is implemented as follows: the
message (transition) $SendReqE(i)$ is sent by the agent $i$ to the
directory. The directory receives this message and updates its
book-keeping state ($RecvReqE(i)$). Next, if the directory is able to
grant a right for exclusive access, it sends the message $SendGntE(i)$
to agent $i$. The agent $i$ receives this grant ($RecvGntE(i)$) and
transitions to \Exclusive state. These messages are passed (in that
order) between the agent $i$ and the directory, as shown
\Fig{figGermanFlowsExcl}.
\begin{figure}
\centering
\hspace*{- 0.75cm}
\subfloat[\Exclusive flow]{
\label{figGermanFlowsExcl}
\includegraphics[scale=.65]{./Figures/germanExcl.jpg}
}
\hspace*{1.75cm}
\subfloat[\Invalidate flow]
{
\label{figGermanFlowsInv}
\includegraphics[scale=.65]{./Figures/germanInv.jpg}
}
\caption{Flows for the German protocol.}
\label{figGermanFlows}
\end{figure}
Next, it may be noted that the message $SendGntE(i)$ is sent only if
no other agent has \Shared access to the cache line. In case another
agent, say $j$, has access to the cache line, the directory uses the
\Invalidate(j) flow to first invalidate $j$. This flow proceeds by the
directory sending the invalidate ($SendInv(j)$), the agent $j$ sending
the acknowledgment ($SendInvAck(i)$), and the directory receiving it
($RecvInvAck(i)$).
\subsubsection{Deadlock Situation:} Now consider the case when the
acknowledgment (ack) sent by the agent $j$ gets dropped. In this case,
the directory is blocked and waiting on the $ack$. The blockage of
directory leads to all the $N$ agents in the protocol getting blocked
as well. Thus no transition in the protocol is enabled. In a realistic
implementation of the protocol, one would assume the directory would
timeout if the $ack$ is not received: even in that case, this kind of
a bug would negatively impact the performance of the protocol.
In this paper, we check for freedom against a class of deadlock errors
where the entire protocol gets blocked, i.e., all the agents are
stuck. In order to express this mathematically, we model the agents of
the protocol \Protocol(N) to have transitions such that each
transition $\tau$ has a guard $\tau.g$, which is $false$ if the
transition can not happen. Thus, the protocol is stuck if the guards
of all the transitions are false, i.e., \(\bigvee_{\tau} \tau.g.\)
\subsection{Challenge in verifying deadlock}
The key challenge in verifying the above property is that it is a
disjunction of all the guards of the parametrized protocol. Thus, a
violation of the above property will be a state in which all the
guards are false, i.e., when \(\bigwedge_{\tau} \lnot \tau.g\)
holds. However, this violation will in high likelihood be masked by
any over-abstraction due to the extra transitions added in the
over-abstraction process.
One approach to address the above issue is to use a combination of
over and under abstractions (i.e. a mixed abstraction), as described
in a prior deadlock verification work~\cite{DeadlockBingham}. While
very promising in verifying an important class of deadlock properties,
while believe the use of mixed abstraction is fairly complex and
effort intensive.
\subsection{Key Insight}
Our method tries to partition the state of the protocol such that in
each partition it can point to an agent $i$ and claim that one of its
transitions must be enabled. This is done by starting with an initial
guess of state partition for which agent $i$ must have an enabled
transition. In case the agent does not have any transition enabled,
typically there is a particular flow of the agent which is 'stuck'
because the protocol is processing a flow of another agent $j$. This
information is used to refine the partition. This way, we derive
invariants of the form: \(pred \rightarrow \phi(i)\), where $pred$ is
a predicate on protocol state and $\phi(i)$ is a disjunction of
transitions for agent $i$. Invariants of this form are much easier to
check using existing parameterized verification techniques.
\subsection{Key Contributions}
Our method proves deadlock freedom for parameterized protocols
(formalized in \Sec{sec:formalModel}). It works by first generating a
set of invariants which collectively imply deadlock freedom
(\Sec{sec:SpecMethod}). These generated invariants use just one index
(i.e. of the form $\forall i: \phi(i)$) and thus are amenable to state
of the art abstraction techniques, for verification for an unbounded
number of agents (\Sec{sec:CMP}). \Fig{figExperimentalSetup} shows the
setup for our method, and highlights the key contribution of this
paper: generation of invariants collectively implying deadlock freedom
by leveraging high-level information present in $flows$. We verified
two challenging protocols, the German and Flash protocols, using our
method (\Sec{sec:Experiments}).
\begin{figure}
\centering
\hspace*{-0.5cm} \includegraphics[scale=.5]{./Figures/progressSpecGenBigPict.jpg}
\caption{Experimental Flow}
\label{figExperimentalSetup}
\end{figure}
\nop{
We make the following key contributions:
\begin{itemize}
\item We show how deadlock freedom verification can be done without
developing any special techniques by just using existing techniques
for parameterized verification.
\item We show how high-level information can be leveraged to
iteratively specify a set of small-indexed invariants that
cumulatively imply deadlock freedom.
\item Using this method we have verified deadlock freedom of German
and Flash cache coherence protocls, two well known academic
protocols.
\end{itemize}
}
\nop{Whenever
the predicate of a subgroup holds on a protocol state, our invariants
show that one of the transitions in the subgroup associated with that
predicate holds. {\em The key insight is that each subgroup has
transitions from a small number of agents (1 or 2) and this
drastically simplifies checking if some transition in that subgroup
is enabled.}
The partitioning is done by iteratively splitting an initial guess invariant
This results in invariants of the form:
\begin{tabular}{c}
\(cond_1(i) \Rightarrow en_1(i)\) \\
\(cond_2(i) \Rightarrow en_2(i)\)\\
\( \ldots\)\\
\(cond_c(i) \Rightarrow en_c(i)\)
\end{tabular}
\subsection{Experiment, Contributions, limitations}
We model a protocol as a parameterized state transition system
\Protocol(N) which consists of \(N\) symmetric agents, where $N$ is a
finite but unbounded parameter, along with some global state shared
between the agents. The agents have ids $1,2,3,\ldots,N$ and have
local state. The transitions of the protocol are modeled as
guard-action pairs with transitions for processing the requests of
agent $i$ referred to as \(\tau_1(i),\tau_2(i),\tau_3(i) \ldots
\tau_k(i)~\footnote{The indexed transitions may be transitions of
either the agent or the global state. As an example for a cache
coherence protocol, these can be the local transitions of cache $i$
or the transitions of the global directory processing requests of
cache $i$.}.\) We assume symmetry in transitions.
We define a state to be a deadlock state if no outgoing transitions
are enabled in that state. Then, in order to prove deadlock freedom,
one needs to show that in all reachable states, at least some
transition is enabled. Doing this parametrically for an arbitrary
number of agents is hard, as compared to properties like mutual
exclusion (i.e. no two caches are in Exclusive state simultaneously).
\nop{
Moreover, any attempts to abstract behaviors of agents is counter
productive as this merely adds more transitions, thus further
obfuscating the check that all transitions in original protocol are
blocked.
}
The key difference between safety properties like mutual exclusion and
deadlock freedom is that the counterexamples to the former involve a
small number of agents whereas counterexamples to the latter may
involve all the agents of the system. More precisely, for a
parametrized system \Protocol(N) with \(N\) symmetric agents, mutual
exclusion or coherence can be expressed as a two-index property
\[\forall i, j \in [1..N]. \phi(i,j).\]
A counterexample to this will involve exactly two agents, 1 and 2 (by
symmetry). This allows us to construct finite abstractions focusing on
a small number of agents (2 in this case) that can be efficiently
model checked~\cite{CMP,ParamVerMurali, ParamVerifMuraliIndust}. On
the other hand, \emph{a counterexample to deadlock freedom may involve
all the agents of the system.} Abstracting away agents will lead to
the deadlock being missed. Thus efficient abstractions like data type
reduction~\cite{ConcDSDataTypeRed}, which are crucial in successful
verification of mutual exclusion for industrial strength
protocols~\cite{CMP, ParamVerifMuraliIndust, ParamVerMurali}, are not
directly applicable to deadlock freedom.
\subsection{Our Approach}
What we need is an effective way to reduce the unbounded parametrized
problem of verifying deadlock freedom to a finite problem amenable to
model checking. Instead of directly tackling deadlock freedom, our
method enables the user to construct a series of two indexed
properties (invariants) that collectively are stronger than deadlock
freedom.
Deadlock freedom can be specified as an existential assertion: at
every state \(s\) there exists some transition \(\tau_j(i)\) which is
enabled. Formally, this can be written as \[AG \exists i,
\tau_j. (\tau_j(i) \mbox{ is enabled}).\] In order to prove this
property, the existential quantification needs to be eliminated for
the parametrized system. This can be accomplished by identifying for
every state \(s\) the exact transition or set of transitions that have
to be enabled (i.e. skolemization). However, this is hard to
accomplish in general as (1) the exact set of transitions which may be
enabled at any state may be large and involve an arbitrary number of
indices, and (2) enumerating all states is a challenge.
\nop{the system (local
and global state) is parametrized and enumerating all the states and
then identifying transitions will lead to a huge blowup.
}
Our method addresses the above challenges---typically a protocol in
any state is processing a high-level transaction and this can be
leveraged in identifying a subset of transitions which must be
enabled. Our method works by identifying a set of transitions by
leveraging this high-level information such that at least one of these
transitions must be enabled when the high-level transaction is being
processed. Next, instead of enumerating all states, our method enables
the user to derive simple predicates on the protocol, such that these
can be used to refer to a set of states which have a common set of
enabled transitions.
More specifically, message passing protocols are structured around a
series of logical high-level transactions, called
\emph{Flows}~\cite{ParamVerMurali}---flows represent a series of
transitions such that they happen in the order specified and together
they accomplish one logical transaction of the protocol. As an
example, the German cache coherence protocol consists of \{\Exclusive,
\Shared, \Invalidate\} flows as shown in \Fig{figGermanFlows}. As the
names indicate, these deal with processing the \Exclusive, \Shared,
and \Invalidate access requests. Since a flow represents transitions
for one transaction which involves one or two agents interacting with
the global state, the transitions of the flow are typically on a small
number (one or two) of agents: invariants constructed from these
involve only a small number of agents.
Our method enables the user to derive invariants which are as follows:
\nop{
Our method depends crucially on the fact that dealing with Flows
instead of transitions allows to easily specify which Flow \(fl\) must
be enabled, that is one of the transitions appearing in it is enabled,
under a given condition. Our strategy then is to come up with a
series of properties
}
\begin{tabular}{c}
\(cond_1(i) \Rightarrow en_1(i)\) \\
\(cond_2(i) \Rightarrow en_2(i)\)\\
\( \ldots\)\\
\(cond_c(i) \Rightarrow en_c(i)\)
\end{tabular}
Here, $en_p(i)$ is a subset of enabled transitions (typically a
disjunction of the guards of the transitions of all the flows
involving the agent $i$) when the condition $cond_p(i)$
holds. Further, it is trivial to identify for which agent $i$
$cond_p(i)$ holds. As an example, for cache coherence protocols,
$cond_p(i)$ checks if a global pointer at the directory points to the
cache $i$ when the directory is processing a request of that
cache. Further, $en_p(i)$ can be generated as a disjunction of the
guards of the transitions of the flows involved in processing requests
from cache $i$.
\begin{figure}
\centering
\hspace*{- 0.75cm}
\subfloat[\Exclusive flow]{
\label{figGermanFlowsExcl}
\includegraphics[scale=.65]{./Figures/germanExcl.jpg}
}
\hspace*{1.75cm}
\subfloat[\Invalidate flow]
{
\label{figGermanFlowsInv}
\includegraphics[scale=.65]{./Figures/germanInv.jpg}
}
\caption{German protocol flows}
\label{figGermanFlows}
\end{figure}
Note that the above set of \emph{Flow invariants} are all universally
quantified properties with a small number of indices. Thus, they can
be checked using existing parametrized techniques. In this work, we
use the CMP method~\cite{ParamVerMurali,CMP} to verify these
invariants.
\nop{In fact, in the style of CMP method, the \emph{Flow
invariants} can be generated on iteratively starting from an initial
guess. Once we have a set of invariants we can try to verify them. If
the proof goes through we are done. Otherwise, the counter example
will led us to split the \(cond_1,..,cond_c\) leading to more
conditions. Since parameterized verification is undecidable in general
this process doesn't have to terminate always but in practice
protocols are well-behaved and the proof either terminates or we find
a real counter example, see~\ref{sec:experiments}.
}
\begin{figure}
\centering
\hspace*{-0.5cm} \includegraphics[scale=.5]{./Figures/progressSpecGenBigPict.jpg}
\caption{Experimental Flow}
\label{figExperimentalSetup}
\end{figure}
We make the following key contributions:
\begin{itemize}
\item We show how deadlock freedom verification can be done without
developing any special techniques by just using existing techniques
for parameterized verification.
\item We show how high-level information can be leveraged to
iteratively specify a set of small-indexed invariants that
cumulatively imply deadlock freedom.
\item Using this method we have verified deadlock freedom of German
and Flash cache coherence protocls, two well known academic
protocols.
\end{itemize}
\nop{ The rest of the paper is organized as follows. In the next
section, we formalize the notions already presented here. In
Section~\ref{example}, we demonstrate the method in action by using
the German cache coherence protocol as an example. In
Section~\ref{CMP}, we describe the CMP method and present the
overall method. Section 5 presents our experimental results and
Section 6 concludes the paper. }
\nop{
\subsubsection{Key Limitation}
The key limitation of our approach is that the specification has to be
generated manually through iteratively inspecting counterexamples,
identifying witness variables, and splitting invariants. This can be
automated if the information about the witness variables is available
in the flow diagrams (and it is in certain cases). However, we believe
that adding the information iteratively on inspecting the
counterexample is easier on the user, instead of requiring this
information a-priori in the flow diagrams.
}
}
\subsection{Relevant Related Work}
While there are many techniques for verification of standard safety
property, deadlock freedom and liveness in general are not handled as often.
Of the existing techniques, the notables are as follows. Macmillan has
verified liveness properties of FLASH protocol
parameterically~\cite{DeadlockFlash, McMillanLiveness}. The proof works on
the basis of manually supplied lemmas and fairness assumptions. Next, Baukus
et al. verified liveness properties of the German
protocol~\cite{DeadlockWSIS} using a specialized logic called WSIS. However,
the complexity of WSIS is super-exponential, thus the method is not very
scalable. Fang et al. used counter abstraction to verify liveness
properties~\cite{LivenessCounterAbstraction}: however the scalability of
counter abstraction is limited as well. Another method by the same group
includes~\cite{InvisibleRanking} to parametrically verify liveness
properties.
Bingham et al.~\cite{DeadlockBingham, DeadlockBinghamCAV} formally
verify deadlock as a safety property for protocols by specifying it
using user identified Quiescent states (i.e. a state in which no
resources are held): they specify a state to be deadlock free if a
Quiescent state is reachable from it. In order to accomplish formal
proofs of deadlock freedom, they require a combination of abstractions
(also referred to as {\em mixed}
abstraction~\cite{mixedAbstractions}): (1) an over-abstraction for the
set of reachable states and (2) an under-abstraction of the path from
a reachable state to a Quiescent state. Both these abstractions have
to be manually tuned in order to prove deadlock freedom. Using their
approach, they successfully verified Flash and German protocols.
While their approach is manual like ours, we believe that flows
provide a more natural way of thinking about protocols and thus
specifying properties using them is easier than coming up with
quiescent states and tuning mixed abstractions manually.
In contrast to all the methods above, our method doesn't develop any
new model checking algorithms for handling deadlock freedom. It
instead strengthens deadlock freedom to a set of universally
quantified small-indexed safety properties which can be verified using
standard techniques. In a loose sense, our approach to strengthening
deadlock freedom verification is similar in spirit to the approach in
the Xmas framework~\cite{xmasVmcai, DeadlockSayak} used for generating
invariants for proving progress properties.
Finally, among parameterized verification techniques, apart from the
CMP (CoMPositional) method~\cite{CMP}, there are numerous other
techniques for verifying safety properties. Some examples include
Invisible Invariants for safety~\cite{InvisibleInvariants}, counter
abstraction for safety~\cite{CounterAbstraction}, Indexed
predicates~\cite{IndexedPredicates}. The CMP method is a highly
successful parameterized verification technique, numerous work has
been done in the area of parameterized verification, for
e.g.~\cite{InvisibleInvariants, CounterAbstraction,
IndexedPredicates}. While some of these classical methods are
potentially applicable to concurrent data structures as well, given
the success of the CMP method in verifying industrial cache coherence
protocols~\cite{ParamVerifMuraliIndust}, we believe that our method,
which has a similar flavor to the CMP method, provides a scalable and
effective alternative approach.
\section{Introduction}
Verification of deadlock freedom for message passing protocols, such
as cache coherence protocols, is an important practical problem. Such
protocols involve an arbitrary number of symmetric, interacting agents
(e.g. caches). Thus, they are candidates for parameterized
verification techniques. While parameterized techniques for safety
properties have seen considerable success in verifying key properties
like mutual exclusion (i.e. no two caches have write permissions to a
cache line simultaneously) for industrial scale cache coherence
protocols~\cite{DeadlockFlash, CMP, ParamVerifMuraliIndust}, their
success in verifying deadlock freedom has been limited. In this work
we focus on enabling direct application of such state of the art
parameterized techniques for proving deadlock freedom as well.
\subsection{Characterizing Deadlock for Protocols}
We model a protocol as a parameterized state transition system
\Protocol(N) which consists of \(N\) symmetric agents, where $N$ is a
finite but unbounded parameter, along with some global state shared
between the agents. The agents have ids $1,2,3,\ldots,N$: each agent
$i$ has transitions \(\tau_1(i),\tau_2(i),\tau_3(i) \ldots
\tau_k(i)\). Each transition $\tau_j(i)$ is modeled as a guard-action
pair. A transition is {\em enabled} if its guard is $true$, else it is
{\em disabled}. When an enabled transition occurs, its action is
applied to update the state.
A protocol implements high-level requests, for example a request for
read or write permissions for a cache coherence protocol. Then, the
ultimate goal of any deadlock (or livelock) verification effort is to
prove that each high-level request eventually gets a response.
However, such a property is a liveness property. While there is
existing work in parameterized verification of liveness
properties~\cite{CounterAbstraction, Baukus:02a, InvisibleRanking,
McMillanLiveness}, these approaches typically have limited
scalability due to complex abstractions required for verifying
liveness. Further, direct verification of liveness is hard due to
limited support from existing model checking
tools~\cite{DeadlockSayak}.
\nop{
A protocol implements high-level requests, for example a request for
read or write permissions for a cache coherence protocol. Then, the
a protocol can be informally understood as a state transition system
which implements high-level requests (e.g. requests for read write
permissions in a cache coherence protocol). Each transition of this
system is represented as a guard-action pair. A transition is {\em
enabled} if its guard is $true$, else it is {\em disabled}.
A protocol is a parameterized state transition system \Protocol(N),
with $N$ symmetric agents, where $N$ is a finite but unbounded
parameter. Informally, the protocol consists of states with
transitions $\tau$, where the transition $\tau$ has a guard $\tau.g$
and an action $\tau.a$ which updates the state. A transition $\tau$ is
{\em enabled} in a state $s$ if its guard $\tau.g$ is $true$ in $s$,
else it is {\em disabled}.
A protocol implements high-level requests, for example a request for
read or write permissions for a cache coherence protocol. Then, t
For such a system, the ultimate goal of any deadlock (or livelock)
verification effort is to prove that each high-level request
eventually gets a response.
However, such a property is a liveness property---while there is
existing work in parameterized verification of liveness
properties~\cite{CounterAbstraction, DeadlockWSIS,
InvisibleRanking, McMillanLiveness}, these approaches typically
either have limited scalability due to complex abstractions required
for liveness or tend to be largely manual. Further, direct
verification of liveness is hard due to limited support from existing
model checking tools~\cite{DeadlockSayak}.
}
An alternate approach towards verifying response properties is to
analyze causes for failure of the property. Two key causes for the
failure of liveness properties are: (1) an unfair scheduler, i.e. a
request never gets a chance to get served, or (2) a lack of response
to a request due to the protocol getting stuck in a deadlock or
livelock. There are multiple formal definitions of deadlock in
literature (see for example \cite{DeadlockClassic, DeadlockBingham,
xmasVmcai, DeadlockSayak}): we define a deadlock state as one where
the protocol gets completely blocked, i.e., a deadlock state is a
state in which all its transitions are disabled. Correspondingly, we
define a livelock as a loop of one or more states in which, once a
protocol gets stuck, no high-level requests get served. Assuming
fairness, deadlock typically is a more commonly observed cause for
failures of response properties for industrial protocols compared to
livelock~\cite{DeadlockBingham}.
\subsection{Challenges in Proving Deadlock Freedom}
The main complexity in deadlock verification is that a deadlock
witness can involve an arbitrary number of agents deadlocked in a
classical resource dependency cycle~\cite{DeadlockClassic}---unlike
say mutual exclusion where a witness for the violation involves the
state of the two caches with erroneous write permissions. Formally,
mutual exclusion can be specified on a protocol model as an invariant
of the form $\forall i,j:\phi(i,j)$, where $\phi$ is a propositional
logic formula with two index variables $i$ and $j$ stating that agents
$i$ and $j$ cannot have simultaneous write permissions. Invariants
of this form involve a small number (say $k$) of universally
quantified index variables and thus can be verified using efficient
abstractions like the standard data type
reduction~\cite{ConcDSDataTypeRed}. This abstraction keeps just $k$
agents and abstracts away the state of all the other agents, replacing
them with a stateless environment. The usage of data type reduction
along with state of the art parameterized techniques has been
instrumental in successful verification of industrial scale
protocols~\cite{DeadlockFlash, CMP, ParamVerifMuraliIndust}.
However, unlike mutual exclusion, specification of deadlock freedom as
an invariant with a small number of indices is challenging: since a
deadlock violation can involve an arbitrary number of agents, its
specification $\phi$ must in general be indexed over an arbitrary
number of agents. In this work we address this by leveraging
high-level information to specify invariants involving a small number
of quantified index variables. These invariants collectively imply
deadlock freedom and can be directly verified using data type
reduction and existing parameterized verification techniques.
\subsection{Proving Deadlock Freedom using Flow Properties}
\label{sec:flows}
\subsubsection{Flows}
{\em Flows} provide information about the messages passed in the
protocol implementation for processing high-level requests. These are
readily available in industrial documents in the form of charts and
tables. In this work we show how these can be used for proving
deadlock freedom.
As an example, the German protocol consists of \{\Exclusive, \Shared,
\Invalidate\} flows, for processing \Exclusive, \Shared, and
\Invalidate requests. \Fig{figGermanFlows} shows the \Exclusive and
\Invalidate flows for the German protocol.
Thus, for example, in order to implement the \Exclusive request, the
messages $SendReqE(i)$, $RecvReqE(i)$, $SendGntE(i)$ and $RecvGntE(i)$
are passed between the agent $i$ and the directory, as shown
\Fig{figGermanFlowsExcl}.
The messages passed in the protocol implementation are modeled as
transitions and thus have guards and actions. For example,
$SendGntE(i)$ has a guard
checking that the set of caches with read access, i.e. $ShrSet$, must
be empty and an action which sends a message to cache $i$ granting a
write access. We say that a flow is {\em enabled} if one of its
messages has a guard which is $true$.
\begin{figure}
\centering
\hspace*{- 0.75cm}
\subfloat[\Exclusive flow]{
\label{figGermanFlowsExcl}
\includegraphics[scale=.65]{./Figures/germanExcl.jpg}
}
\hspace*{1.75cm}
\subfloat[\Invalidate flow]
{
\label{figGermanFlowsInv}
\includegraphics[scale=.65]{./Figures/germanInv.jpg}
}
\caption{Flows for the German protocol.}
\label{figGermanFlows}
\end{figure}
\subsubsection{Specifying Flow Properties for Deadlock Freedom}
Since in a deadlock state, there are no enabled transitions, the
guards of all transitions
are $false$. Thus, a protocol is deadlock free if at all times, there
is at least some enabled transition in the protocol. We leverage flows
to prove this and verify that at all times there is at least one flow
which is enabled.
Formally, we show the flow property referred to as \Enabled: ``at all
times, there exists {\em some} flow which is enabled''. In temporal
logic, this property can be stated as $G(\exists fl(i): \g(fl(i))$,
where $G$ is the $always$ operator of temporal logic, $fl(i)$ a flow
at agent $i$, and $\g(fl(i))$ is the disjunction of guards of all the
transitions in $fl(i)$. For example for the German protocol,
the \Enabled property can be written as $G(\exists fl(i):\g(fl(i)))$,
where $fl(i)$ $\in$ $\{$\Exclusive($i$), \Invalidate($i$),
\Shared($i$)$\}$.
In order to verify the \Enabled property, the existential quantifier
needs to be eliminated. This is accomplished by the user by adding
witness variables which store the index of the
agent which must have some flow enabled, for a set of states of the
protocol (specified using a predicate on state
variables)~\footnote{This process can be understood as skolemization
for an Invariant with an existential quantifier.}. As an example,
for the German protocol, two of the invariants obtained on adding
witness variables are as follows: (1) the \Exclusive flow is enabled
for a set of states where there are no sharers of the cache line,
i.e., $(\SetOfSharers=\{\})$ $\rightarrow$ $\forall i:
\g(\Exclusive(i))$, and (2) for the set of states where there are
sharers, the \Invalidate flow must be enabled for a sharer, i.e.,
$\lnot (\SetOfSharers = \{\})$ $\rightarrow$
$\g(\Invalidate(Sharer))$, where, $Sharer$ is a witness variable
pointing to a sharer of the cache line.
\subsection{Parameterized Verification of Flow Properties}
As can be seen from the German protocol example, the invariants
generated on iteratively adding witness variables for the enabled
flows involve a small number of index variables (equal to the maximum
number of agents involved in sending or receiving messages in a
flow). Then, these can be verified for an arbitrary number of agents
by direct application of standard data type
reduction~\cite{ConcDSDataTypeRed} along with state of the art
parameterized verification techniques. We use the CMP (CoMPositional)
method~\cite{DeadlockFlash,CMP}: this method is an
abstraction-refinement based method which, along with data type
reduction, has seen considerable success in verifying safety
properties of industrial scale protocols. The final experimental flow
is shown in \Fig{figExperimentalSetup}.
\begin{figure}
\centering
\hspace*{-0.5cm} \includegraphics[scale=.6]{./Figures/progressSpecGenBigPict.jpg}
\caption{Experimental Flow}
\label{figExperimentalSetup}
\end{figure}
\subsection{Key Contributions}
We make the following key contributions:
\begin{itemize}
\item We show how high-level information can be leveraged to specify
and then iteratively generate a set of invariants by adding witness
variables pointing to enabled flows. These invariants are amenable
to industrial scale parameterized verification techniques. (Section
3)
\item We show how the \Enabled property implies deadlock
freedom. (Section 3)
\nop{
present a method which enables the user to specify if some
flow is enabled (i.e. the \Enabled property) as a set of invariants,
by leveraging high-level information (flows).
}
\item Finally, we evaluate the utility of our approach by verifying
the Flash and German protocols for an arbitrary number of
caches. (Section 5)
\end{itemize}
\subsubsection{Key Limitation: }
The key limitation of our approach is that the specification has to be
generated manually through iteratively inspecting counterexamples,
identifying witness variables, and splitting invariants. This can
potentially be automated if the information about the witness
variables is available in the flow diagrams (and it is in certain
cases). However, we believe that adding the information iteratively on
inspecting the counterexample is easier on the user, instead of
requiring this information a-priori in the flow diagrams. Further, the
number of iterations required to add this information for even complex
protocols like Flash were observed to be small (3 for the Flash
protocol).
\subsection{Relevant Related Work}
Bingham et al.~\cite{DeadlockBingham, DeadlockBinghamCAV}
parametrically verify deadlock freedom as a safety property and can
handle complex protocols such as the Flash protocol. Like ours, their
approach is based on the CMP method as well. In particular, their
method extends the CMP method to handle formal proofs for deadlocks:
their notion of deadlock is to specify it by using user identified
Quiescent states, i.e. a state in which no resources are held. They
specify a state to be deadlock free if a Quiescent state is reachable
from it. In order to accomplish formal proofs of deadlock freedom,
they extend the CMP method by using a combination of abstractions
(also referred to as {\em mixed}
abstraction~\cite{mixedAbstractions}): (1) an over-abstraction for the
set of reachable states and (2) an under-abstraction of the path from
a reachable state to a Quiescent state. Both these abstractions have
to be manually tuned in order to prove deadlock freedom. While their
approach is manual like ours, we believe that flows provide a more
natural way of thinking about protocols and specifying
properties. Further, our approach uses well understood
over-abstractions which are easier than coming up with quiescent
states and then tuning over and under abstractions manually.
Among parameterized techniques for verifying liveness (and thus
response properties), Macmillan has verified liveness properties of
FLASH protocol parametrically~\cite{DeadlockFlash,
McMillanLiveness}---their proof works on the basis of lemmas and
fairness assumptions supplied to the model checker. In contrast, our
work leverages high-level information for specifying and verifying
deadlock: this helps in reducing manual effort. Baukus et al. verified
liveness properties of the German protocol~\cite{Baukus:02a} using
their logic WS1S. However, the complexity of WS1S is
super-exponential, thus restricting scalability. Fang et al. provide
another interesting approach for verifying liveness: they use
invisible ranking~\cite{InvisibleRanking} (they used counter
abstraction~\cite{CounterAbstraction} in prior work). Their approach
involves using ranking functions for constructing abstractions to
verify liveness of algorithms like the Szymanski’s and Bakery
algorithms.
Next, the concept of detecting deadlock (and livelock) by reducing
them to a set of invariants (e.g. progress invariants) is well known
(see for example~\cite{xmasVmcai, DeadlockSayak}). Our approach is
similar to these: we also seek to reduce deadlock freedom to a set of
invariants. However, our approach leverages flows in order to reduce
deadlock verification to a set of invariants which then can readily be
used with state of the art techniques. Further, these flows along with
the CMP method are already used in industry for designing and
verifying complex protocols~\cite{ParamVerMurali}: thus they are a
promising approach for verifying deadlock freedom as well.
\nop{Further, the usage of the CMP method in our approach helps in
achieving scalability to complex protocols like Flash---which seems
unlikely for the approaches for network protocols. } \nop{ These
approaches work with user identifying a set of primitives to model a
system, or a set of progress assumptions [Explain], or bla bla bla.
This work handles verifying deadlock freedom for complex control
logic by first modeling it using a fixed set of primitives and then
reducing it to a set of invariants. Our work is similar in spirit to
these---however to the best of our knowledge there is no existing
work for verifying deadlock freedom for protocols which leverages
the existing flow documents. The usage of flows drastically
simplifies the effort required in coming up with witness
variables. Further, none of these techniques work with the CMP
method, and thus have no chance of scaling.
}
Finally, among parameterized verification techniques for protocols,
there are numerous existing techniques in literature apart from the
CMP method. Some examples include Invisible Invariants for
safety~\cite{InvisibleInvariants}, Counter Abstraction for
safety~\cite{CounterAbstraction}, Indexed
predicates~\cite{IndexedPredicates}, ordinary model
checking~\cite{Abdullah:99a}, WS1S~\cite{Baukus:02a},
learning~\cite{Grinchtein:06} and strengthening split
invariants~\cite{Cohen:07}. Further, a transaction-based
method~\cite{ParkDill:96} and a predicate abstraction-based
method~\cite{DasDill:99} were also used to verify cache coherence
protocols. However, these techniques either do not scale to larger
protocols or tend to require increased manual effort to scale to
larger protocols. The demonstrated success of the CMP method along
with the data type reduction abstraction for industrial scale
protocols motivated us to use it in our approach.
\section{Introduction}
We consider the problem of verifying deadlock freedom for symmetric
cache coherence protocols. Consider a cache coherence protocol
\Protocol(N) where the parameter \(N\) represents an unbounded number
of cache agents. The protocol implements requests sent by the agents
using messages exchanged in the protocol. For a protocol designer, the
main property of interest is the request-response property, i.e.,
every request from an agent eventually gets a response. Since this
property is a liveness property which is hard for existing model
checking tools, designers resort to identifying causes for response
property failure, such as deadlock-style failures, and verify against
them.
The literature is abundant with various definitions of
deadlock~\cite{DeadlockClassic, DeadlockBingham}. We focus on deadlock
errors in which the entire protocol gets blocked, i.e., no agent of
the protocol can make any transition. We refer to such an error as a
{\em system-wide deadlock (\mdeadlock)}. If we model each transition
$\tau$ of the protocol to have a guard $\tau.g$, which is $false$ if
the transition is not enabled, the \mdeadlock error occurs if the
guards of all the transitions are \false, i.e., \(\bigwedge_{\tau}
\lnot (\tau.g)\) is \true. This kind of failure, while weaker than
other broader classes of deadlock failures, is commonly observed in
industrial computer system designs and is consistent with the internal
definition for deadlock used by the Murphi model checker as
well~\cite{Murphi}. This class of deadlocks is well motivated for
parameterized cache coherence protocols as these use a centralized
synchronization mechanism (e.g. a directory) and thus any deadlock
results in the directory getting blocked. It is highly likely that
such a deadlock in the shared directory will end up involving all of
the agents of the protocol getting blocked, i.e., unable to make any
transition.
Since an \mdeadlock error involves all of the unbounded number of
agents getting blocked and unable to make any transition, verification
of \mdeadlock freedom naturally is a parameterized verification
problem. Parameterized verification techniques work by using sound
abstractions to reduce the unbounded model to a finite bounded model
that preserves the property of interest. These abstractions typically
tend to be simple over-abstractions such as {\em data-type
reduction}~\cite{ConcDSDataTypeRed}. This abstraction keeps a small
number of agents ($1$ or $2$) as is and replaces all the other agents
with an abstract environment. Such abstractions along with
parameterized techniques like the CMP (CoMPositional)
method~\cite{CMP} have had considerable success in verifying key
safety properties like mutual exclusion and data integrity even for
industrial scale protocols~\cite{CMP, ParamVerifMuraliIndust,
ParamVerMurali}.
\subsection{Challenge in Verifying \MDeadlock}
While parameterized techniques are successful for safety properties
such as mutual exclusion and data integrity, the application of such
abstractions for parameterized verification of properties such as
\mdeadlock is hard. The key challenge arises from the fact that an
\mdeadlock violation is a state in which all the guards are \false,
i.e., when \(\bigwedge_{\tau} \lnot (\tau.g)\) holds; simple
over-abstractions such as data-type reduction will easily mask this
violation due to the discarded state of agents other than $1$ and $2$
and the extra transitions of the environment.
One approach to address the above issue is to use a combination of
over and under abstractions (i.e., a {\em mixed} abstraction) instead
of data-type reduction, as described in a prior deadlock verification
work~\cite{DeadlockBingham}. While promising for verifying a large
class of deadlock errors, the use of mixed abstraction requires
reasoning about over and under abstraction simultaneously and easily
becomes fairly complex.
In this paper we take a different approach. We show how high-level
information about the protocols, in the form of message sequence
diagrams referred to as {\em flows}, can be leveraged to construct
invariants which are collectively stronger than the \mdeadlock freedom
property. These invariants are amenable to efficient abstractions like
data-type reduction which have been used in the past for verifying
industrial scale protocols.
\subsection{Leveraging Flows for Deadlock Freedom}
Cache coherence protocols implement high-level requests for read
(termed \Shared) or write (termed \Exclusive) access from cache
agents, or for invalidating access rights (termed \Invalidate) of some
agent from the central directory. The implementation of these requests
is done by using a set of transitions which should occur in a specific
protocol order. This ordering information is present in diagrams
referred to as {\em message flows} (or {\em flows} for brevity). These
flows are readily available in industrial documents in the form of
message sequence charts and tables~\cite{ParamVerMurali}.
\Fig{figGermanFlows} shows two of the flows for the German cache
coherence protocol describing the processing of the \Exclusive and
\Invalidate requests. Each figure has a directory $Dir$, and two
agents $i$ and $j$. The downward vertical direction indicates the
passage of time. The \Exclusive request is sent by the cache agent $i$
to the directory $Dir$ to request a write access. The \Exclusive flow
in \SubFig{figGermanFlowsExcl} describes the temporal sequence of
transitions which occur in the implementation in order to process this
request: each message is a transition of the protocol. The message
$SendReqE(i)$ is sent by the agent $i$ to $Dir$ which receives this
message by executing the transition $RecvReqE(i)$. Next, if the
directory is able to grant \Exclusive access, it sends the message
$SendGntE(i)$ to agent $i$ which receives this grant by executing
$RecvGntE(i)$. However, in case the directory is unable to send the
grant since another agent $j$ has access to the cache line, the
directory sends a request to invalidate the access rights of $j$. The
temporal sequence of transitions which occur in the implementation in
this case are shown in the \Invalidate flow in
\SubFig{figGermanFlowsInv}. This flow proceeds by the directory sending
the $SendInv(j)$ message, the agent $j$ sending the acknowledgment
message $SendInvAck(j)$, and the directory receiving it by executing
$RecvInvAck(j)$ transition.
\begin{figure}
\centering
\hspace*{- 0.75cm}
\subfloat[\Exclusive flow]{
\label{figGermanFlowsExcl}
\includegraphics[scale=.55]{./Figures/germanExcl.jpg}
}
\hspace*{1.75cm}
\subfloat[\Invalidate flow]
{
\label{figGermanFlowsInv}
\includegraphics[scale=.55]{./Figures/germanInv.jpg}
}
\caption{Flows for the German protocol.}
\label{figGermanFlows}
\end{figure}
\nop{
Next, it may be noted that the message $SendGntE(i)$ is sent only if
no other agent has \Shared access to the cache line. In case another
agent, say $j$, has access to the cache line, the directory uses the
\Invalidate(j) flow to first invalidate $j$. This flow proceeds by the
directory sending the invalidate ($SendInv(j)$), the agent $j$ sending
the acknowledgment ($SendInvAck(i)$), and the directory receiving it
($RecvInvAck(i)$).
}
\nop{
\subsubsection{Deadlock Situation:} Now consider the case when in a buggy protocol
the agent $j$ does not send the acknowledgment $SendInvAck(i)$ due to
a coding error. In this case, the directory is blocked and waiting to
receive $SendInvAck(i)$. The blockage of directory leads to all the
$N$ agents in the protocol getting blocked as well. Thus no transition
in the protocol is enabled. In an RTL implementation of such a
protocol, one would assume the directory to timeout if the
$SendInvAck(i)$ is not received: even in that case, this kind of a
high-level bug would seriously impact the performance of the protocol.
}
\subsubsection{Freedom from \MDeadlock}
At a high-level, our method tries to exploit the fact that if the
protocol is \mdeadlock free, when none of the transitions of an agent
are enabled, another agent can be identified which must have a
transition enabled. This identification leverages the key insight that
in any state of the protocol, if all the transitions of some agent,
say $a_1$, cannot occur, then, some flow of that agent must be blocked
since it depends on another flow of another agent, say $a_2$, to
finish. Then, there are two possibilities: (1) the agent $a_2$ is
enabled, in which case the state is not an \mdeadlock state, or (2)
the agent $a_2$ is blocked as well, in which case it depends on
another agent $a_3$. If this dependence chain is acyclic, with the
final agent in the chain enabled, the protocol is \mdeadlock
free. However, if the final agent is not enabled, or if the dependence
chain has a cycle, the protocol may either have an \mdeadlock error or
there may be an error in the flow diagrams used.
As an example, for the German protocol, if the \Exclusive flow of
agent $i$ is blocked since the transition $SendGntE(i)$ cannot occur,
it is waiting for $j$ to get invalidated. In the protocol, at least
some transition of the \Invalidate flow on agent $j$ can occur. This
enables proving freedom from \mdeadlock for the protocol.
Using the above insight, by analyzing the dependence between blocked
agents, our method is able to point to an agent which must have at
least one transition enabled in every reachable state of the
protocol. Specifically, our method enables the derivation of a set of
invariants $\InvSet$ which collectively partition the reachable state
of the protocol. Each invariant then points to the agents which must
have at least one transition enabled when the protocol is in a state
belonging to its partition. These invariants are derived in a loop by
iteratively model checking them on a protocol model with $c$ agents,
where $c$ is heuristically chosen as discussed in
\Sec{sec:SpecMethod}.
\subsubsection{Verifying for an Unbounded Number of Agents}
Once the invariants in $\InvSet$ are derived, they hold for a model
with $c$ agents. These invariants use just one index (i.e., they are
of the form $\forall i: \phi(i)$) and thus, they can be verified for
an unbounded number of agents by using efficient parameterized
verification techniques such as data-type reduction along with the CMP
(CoMPositional) method~\cite{CMP}. This technique has previously been
successful for verifying mutual exclusion for industrial
protocols~\cite{ParamVerifMuraliIndust}. We note that our approach is
not limited to the CMP method: the invariants derived may be verified
by using any parameterized safety verification
technique~\cite{InvisibleInvariants, CounterAbstraction,
IndexedPredicates, ProtocolFlashAmitGoel}.
\nop{
Our method takes a Murphi model of the protocol along with an initial
broad guess for $\InvSet$ as input. Then, $\InvSet$ is iteratively
refined by model checking it on a protocol model with $c$ agents in a
loop. This refinement is done by user-assisted analysis of the
returned counterexample and splitting the failing invariant in
$\InvSet$ by leveraging flow information. Once the invariants in
$\InvSet$ pass for a model with $c$ agents, they are verified for an
unbounded number of agents using data-type reduction, along with the
CMP (CoMPositional) method~\cite{CMP}, a previously known technique
highly successful for verifying mutual exclusion for industrial
protocols~\cite{ParamVerifMuraliIndust}. \Fig{figExperimentalSetup}
shows the setup for our method, and highlights the key contribution of
this paper.
It may be noted that our approach is not limited to the CMP method:
the invariants derived may be verified using any of the numerous
parameterized safety verification
techniques~\cite{InvisibleInvariants, CounterAbstraction,
IndexedPredicates}.
}
\nop{
Our method tries to partition the state of the protocol such that in
each partition it can point to an agent $i$ and claim that one of its
transitions must be enabled. This is done by starting with an initial
guess of state partition for which agent $i$ must have an enabled
transition. In case the agent does not have any transition enabled,
typically there is a particular flow of the agent which is 'stuck'
because the protocol is processing a flow of another agent $j$. This
information is used to refine the partition. This way, we derive
invariants of the form: \(pred \rightarrow \phi(i)\), where $pred$ is
a predicate on protocol state and $\phi(i)$ is a disjunction of
transitions for agent $i$. Invariants of this form are much easier to
check using existing parameterized verification techniques.
}
\subsection{Key Contributions}
Our method proves \mdeadlock freedom for parameterized protocols
(formalized in \Sec{sec:formalModel}). It takes a Murphi model of the
protocol as input. As shown in \Fig{figExperimentalSetup}, first, a
set of invariants $\InvSet$ which collectively imply \mdeadlock
freedom are derived on a model with $c$ agents
(\Sec{sec:SpecMethod}). These invariants are verified for an unbounded
number of agents by using state-of-the-art parameterized verification
techniques (\Sec{sec:CMP}). We verified Murphi implementations of two
challenging protocols, the German and Flash protocols using our method
(\Sec{sec:Experiments}).
\begin{figure}
\centering
\hspace*{-0.5cm} \includegraphics[scale=.55]{./Figures/ExCMP.jpg}
\caption{Experimental Flow}
\label{figExperimentalSetup}
\end{figure}
{\em Limitation:} The key limitation of our approach is that the
invariants have to be derived manually by inspecting
counterexamples. This can be automated if additional information about
conflicting flows is available in the flow diagram itself.
\subsection{Relevant Related Work}
\textbf{{\em Deadlock Verification:}} The work closest to ours is by
Bingham \etal~\cite{DeadlockBingham, DeadlockBinghamCAV}. They
formally verify deadlock as a safety property for protocols by
specifying it using user-identified Quiescent states (\ie a state in
which no resources are held): they specify a protocol state to be a
deadlock state if no Quiescent state is reachable from it. They prove
freedom from such a deadlock by using a combination of over and under
abstractions (also referred to as a {\em mixed}
abstraction~\cite{mixedAbstractions}).
Their approach is a promising way to verify deadlock freedom which
scales to protocols like the Flash protocol. However, the required
tuning of both under and over abstractions simultaneously can be
complex. In contrast, we take the flow-based alternative to enable
simpler abstractions like data-type reduction.
Since the ultimate goal of any deadlock verification effort is to
verify the response property (i.e. every high-level request eventually
gets a response), we contrast our work with liveness verification
efforts as well. Among techniques for parameterized verification of
liveness, McMillan has verified liveness properties of the Flash
protocol~\cite{DeadlockFlash, McMillanLiveness}. The proof is manual
and works on the basis of user supplied lemmas and fairness
assumptions. In contrast, our method reduces manual effort by
leveraging information from flows along with the CMP method.
Among automatic approaches for verifying liveness properties, Baukus
\etal verified liveness properties of the German
protocol~\cite{DeadlockWSIS} using a specialized logic called
WSIS. Fang \etal used automatically deduced ranking
functions~\cite{InvisibleRanking} and, in a prior work, counter
abstraction~\cite{CounterAbstraction} to verify liveness
properties. While fully automatic, these approaches tend to exhibit
limited scalability for larger protocols such as Flash, due to the
inherent complexity of the liveness verification problem. In contrast
to these, our approach, while requiring some user guidance, achieves
much greater scalability and enables us to verify the Flash protocol.
\textbf{{\em Parameterized Verification Techniques:}} We note that the
invariants derived using our method can be verified for an unbounded
number of caches by any parameterized safety verification technique,
it is not dependent on the CMP method which we used. Our choice of
using the CMP method was motivated by the fact that it is the only
state-of-the-art method we are aware of which has been used
successfully for verifying protocols like Flash and other industrial
scale protocols. Among other techniques, an important technique is by
Conchon \etal~\cite{ProtocolFlashAmitGoel} which uses a backward
reachability algorithm to automatically prove a simplified version of
the Flash protocol. Next, there are numerous other prior approaches in
literature for parameterized verification of safety properties. The
CMP method falls in the broad category of approaches which use {\em
compositional reasoning}~\cite{ParamLamportCompositional,
ParamLamportCompositional1} and {\em abstraction} based techniques
to verify parameterized systems; the literature is abundant with
examples of these~\cite{ParamVerifNamjoshiCompositional,
DeadlockFlash, InvisibleInvariants, CounterAbstraction,
IndexedPredicates, DasDill:99,
ParamVerifEnvironmentAbstractionMurali, ParamVerifEnvAbstrMurali}.
Next, another category of approaches work by computing a {\em cutoff}
bound $k$ and showing that if the verification succeeds for $k$
agents, then the protocol is correct for an arbitrary number of
agents~\cite{ParamAmir1, ParamKahlon, ParamVerifVineetCutoff,
ParamVerifClarkeCutoff, ParamWahlCutoff,
ParamAbdullaCutoff}. Finally, there are approaches based on {\em
regular model checking} which use automata-based algorithms to
verify parameterized systems~\cite{ParamRegularModelCheckingClassic,
PAAbdullahRegularModelCheckingSurvey,
ParamRegularModelCheckingClassic1, ParamIteratingTransducers}. To
the best of our knowledge, the CMP method is the state-of-the-art for
protocol verification in contrast to these methods and has been used
to successfully verify larger protocols such as Flash with minimal
manual effort. (Other methods which verify Flash protocol in full
complexity are by Park \etal~\cite{ParkDill:96} and Park
\etal~\cite{DasDill:99}. As described by Talupur
\etal~\cite{ParamVerMurali}, these are significantly manual and take
much more time to finish verification of the Flash protocol compared
to the CMP method.)
\nop{ Finally, the last category of approach thr bounded, some
examples include Invisible invariants~\cite{InvisibleInvariants},
Counter abstraction~\cite{CounterAbstraction}, Indexed
predicates~\cite{IndexedPredicates}, and cutoff based
techniques~\cite{KahlonManyToFew}: }
\nop{
In contrast to all the methods above, our method doesn't develop any
new model checking algorithms for handling deadlock freedom. It
instead strengthens deadlock freedom to a set of universally
quantified small-indexed safety properties which can be verified using
standard techniques. In a loose sense, our approach to strengthening
deadlock freedom verification is similar in spirit to the approach in
the Xmas framework~\cite{xmasVmcai, DeadlockSayak} used for generating
invariants for proving progress properties.
Finally, among parameterized verification techniques for safety, we chose
to use the CMP (CoMPositional) method~\cite{CMP} for
, there are numerous
other techniques for verifying safety properties. Some examples
include Invisible Invariants for safety~\cite{InvisibleInvariants},
counter abstraction for safety~\cite{CounterAbstraction}, Indexed
predicates~\cite{IndexedPredicates}. The CMP method is a highly
successful parameterized verification technique, numerous work has
been done in the area of parameterized verification, for
e.g.~\cite{InvisibleInvariants, CounterAbstraction,
IndexedPredicates}. While some of these classical methods are
potentially applicable to verifying protocols as well, given the
success of the CMP method in verifying industrial cache coherence
protocols~\cite{ParamVerifMuraliIndust}, we believe that our method,
which has a similar flavor to the CMP method, provides a scalable and
effective alternative approach.
}
\subsection{Related Work}
\label{sec:related}
Parameterized verification, mainly of cache coherence
and mutual exclusion protocols, has received considerable
attention. Recent methods have been based on
invisible invariants~\cite{AronsRuah:01},
counter abstraction~\cite{Pnuelixu},
indexed predicates~\cite{Lahiri:04a},
ordinary model checking~\cite{Abdullah:99a},
WS1S ~\cite{Baukus:02a}, learning~\cite{Grinchtein:06},
strengthening split invariants~\cite{Cohen:07},
and environment abstraction~\cite{Self:06,Self:08}.
These techniques have had varying degrees of success,
but none of them has been
applied to a large industrial-strength protocol like Flash
(although environment abstraction~\cite{Self:08} has been applied to a
simplified version of Flash).
McMillan introduced the CMP method
in a series of papers~\cite{McMillan:Toma,McMillan:Flash,CMP}
(the proof in~\cite{McMillan:Flash} does not
completely follow the CMP framework: to get BDDs to scale, the Flash model
had to be pruned by hand.)
Chou, Mannava, and Park~\cite{CMP} elaborated on the method
and showed how it can be performed in conjunction with any model
checker. Krstic~\cite{Krstic:AVIS} gave a formalization of the method.
As far as we are aware, the CMP method is one of the few methods to handle
the full complexity of the Flash protocol. A transaction-based
method~\cite{ParkDill:96} and a predicate abstraction-based
method~\cite{DasDill:99} have both verified the safety properties of Flash.
Both the methods required extensive manual guidance and took several days to
complete the verification. Recently, some nearly automatic methods based on
BDDs have been proposed~\cite{LvYi:2007,Bingham:2008}. As with the other
BDD-based methods~\cite{McMillan:Flash}, they have trouble scaling up: they
have been able to verify only a control property, but not data. For all of
these methods applied to Flash, model checking has taken roughly a couple
hours. In contrast, our work is able to check both the data and control
properties of Flash in just 120 seconds.
We note that our method of deriving
invariants from message flows is not specific to CMP, and can potentially be
used in conjunction with any other method of protocol verification.
Finally, the compositional reasoning principle at the heart of our new
analysis of CMP is similar to principles proposed by Abadi and
Lamport~\cite{AbadiLamp:93,AbadiLamp:95}, McMillan~\cite{McMillan:Toma},
Krstic~\cite{Krstic:AVIS}, Bhattacharya et al.~\cite{Ritwik} and Chen et
al.~\cite{ChenGanesh:07}.
We introduce preliminary concepts in Section~\ref{sec:prelim},
derive protocols invariants from message flows in Section~\ref{sec:flows},
present a new analysis of CMP in Section~\ref{sec:cleancmp},
show how to use flow-derived invariants with CMP in Section~\ref{sec:augcmp},
give experimental results in Section~\ref{sec:results},
and conclude with Section~\ref{sec:conclusion}.
\section{Related Work}
Parameterized verification
Deadlock verification for Caches
Reduction of liveness to safety
Other liveness verification techniques.
Lot of work has been done in deadlock verification. The reduction of
progress to a bunch of safety properties is not
new~\cite{DeadlockSayak}.
\section{Deriving Invariants for Proving \MDeadlock Freedom}
\label{sec:SpecMethod}
In this section, we show how invariants can be derived from flows
which collectively imply \mdeadlock freedom. At a high-level, our
method works by partitioning the global state of the protocol using
predicates, such that for each partition, some agent $i$ has at least
one transition enabled. Each invariant $inv$ is of the form $inv.pred
\rightarrow \forall i \in \ISet{inv} \, \g(i)$, where $inv.pred$ is a
predicate on the global variables of the protocol, $\ISet{inv}
\subseteq \IndexSet$ s.t. $\lnot(\ISet{inv}=\{\})$~\footnote{This is
discharged as a separate assertion for model checking.} and $\g(i)$
is a disjunction of the guards of agent $i$. {\em The key insight is
that since $\g(i)$ has transitions from a single agent, model
checking $inv$ for an unbounded number of agents is drastically
simpler than checking the original \mdeadlock property, as discussed
in \Sec{sec:CMP}.}
Our method works by iteratively model checking invariant $inv$ and on
failure, refining the invariant into two invariants by splitting the
predicate and updating the set $\ISet{inv}$ for each. Suppose, an
invariant of the form $inv.pred \rightarrow \forall i \in \ISet{inv} \,
\g(i)$ fails on model checking. Then, there exists some agent $i_f$
such that when $inv.pred$ holds, $i_f \in \ISet{inv}$ holds and
$\g(i_f)$ is \false. Then, this agent has a flow $fl$ which is blocked
(i.e. it has a rule which is expected to be enabled and so has
precondition \true, however has its guard \false). This blocked flow
is typically waiting for another flow $fl'$ on another agent $i_s$ to
complete. As an example, for German protocol, the \Exclusive flow may
be blocked for agent $i_f$ with the rule $SendGntE(i_f)$ having
precondition \true but guard \false. In this case, it is waiting for
an \Invalidate request to complete for another agent $i_f$, which is
in the set $Sharers$.
The user identifies a global witness variable $\ptr$ which points to
$i_s$, along with a conflict condition $conf$ on the global state
which indicates when $i_s$ is enabled and $i_f$ fails. The user
accomplishes this by using the heuristic that typically, if the rule
$rl$ of flow $fl$ of agent $i_f$ is blocked, $conf$ can be derived by
inspecting the guard of $rl$: $conf$ is the cause for falsification of
$rl.g$. Further, $\ptr$ identifies the succeeding flow $i_s$; this
can be an auxiliary variable or a state variable already being used by
the protocol. For example, for German protocol, $conf$ can be derived
from the guard of $SendGntE$ and $\ptr$ points to some sharer of the
cache line which is being invalidated.
Using $conf$ and $\ptr$, the invariant can be split into two
invariants. (1) The first invariant excludes the case when conflict
happens from the original invariant, i.e., $inv1: inv.pred \land \lnot
conf \rightarrow \forall i \in \ISet{inv1} \, \g(i)$, where
$\ISet{inv1} = \ISet{inv}$. (2) The second invariant shows that when a
conflict happens, the agent pointed to by \ptr must be enabled and so
the protocol is still \mdeadlock-free, i.e., $conf \rightarrow \forall
i \in \ISet{conf} \, \g(i)$, where $\ISet{conf} = \{i: i \in \IndexSet
\land i=\ptr\}$. For both the invariants, assertions which check that
the corresponding set of indices is not null are also added for model
checking. For example, for $inv1$, this assertion is $inv.pred \land
\lnot conf \rightarrow \ISet{inv1}$.
Our method derives these invariants by iteratively model checking on a
small number of $c$ ($3$ for German protocol) agents. (Once the
invariants are generated on $c$ agents, they are generalized to an
unbounded number of agents, as shown is \Sec{sec:CMP}.) This number
$c$ needs to be chosen such that the proof of \mdeadlock freedom is
expected to generalize to unbounded number of agents: as a heuristic
we found that $c$ should be one more than the maximum number of agents
involved in processing a high-level request. As an example, an
exclusive request may involve two agents, a requesting agent $i$ and
an agent $j$ getting invalidated, so we chose $c$ to be equal to 3.
\nop{
The method starts with a broad enough predicate for each agent and
based on iterative model checking and user analysis of conflicting
flows in the counter-example, this predicate is iteratively split into
smaller predicates. The model checking is done iteratively on a
protocol with a small number $c$ of agents ($2$-$3$). (Once the
invariants are generated on $c$ agents, they are generalized to an
unbounded number of agents, as shown is Sec().) This number $c$ needs
to be chosen such that the proof generalizes easily to unbounded
number of agents: as a heuristic we found that $c$ should be one more
than the maximum number of agents involved in processing any
high-level request to accommodate all possible conflicts among flows.
}
\Fig{fig:MethodOverview} shows the details of the method. Our method
starts with an initial broad guess invariant, $\true \rightarrow
\forall i \in \IndexSet: \g(i)$ (line 1). This states that in all
reachable states, every agent has at least one transition enabled. As
this invariant is \false, this broad guess invariant is refined into
finer invariants, using the loop (lines 2-13). On finishing, the user
is able to derive a set of invariants, $\InvSet$, which collectively
imply \mdeadlock freedom. Further, the user is also able to derive an
assertion set, $\AssertSet$, such that for each invariant $inv$ in
$\InvSet$, an assertion in $\AssertSet$ checks if the set of
indices $\ISet{inv}$ is non-empty when $inv.pred$ holds.
\nop{
The user inspects the returned counter-example (line 5). Unless a
discrepancy between the protocol and the flow is exposed (in which
case the loop exits and the flow/protocol is fixed), the user
identifies the failing agent $i_f$. This agent has a flow $fl$ which
is blocked (i.e. it has a rule with precondition \true but the guard
\false). The blocked flow is typically waiting for another flow
($fl'$) on another agent ($i_s$) to complete (e.g. an \Invalidate
request). The user identifies a global witness variable $\ptr$ which
points to $i_s$, along with a conflict condition $conf$ on the global
state which indicates when $i_s$ is enabled and $i_f$ fails. {\em This
is essential for generalizing from the counter-example to come up
with a finer predicate.} Using $conf$ and $\ptr$, the invariants are
updated (lines 11-13).
One heuristic for coming up with $\ptr$ and $conf$ is that, typically,
the flow $fl$ of agent $i_f$ is blocked due to a missing resource-like
variable (e.g. $\lnot (CurPtr=i)$ can be interpreted as $CurPtr$ not held by
$i$). Thus, if the next rule for $fl$ which should be fired is $rl$,
then we observed that $conf$ can be derived by inspecting the guard of
$rl$: $conf$ is the cause for falsification of $rl.g$. Further, $\ptr$
identifies the succeeding flow $i_s$; this can be an auxiliary
variable or at times is a state variable already being used by the
protocol (e.g. $\ptr$ is $CurPtr$).
}
\input{splitMethod.tex}
\subsection{Specifying Invariants for the German Protocol}
\label{sec:GermanInvariantGeneration}
We now describe the complete generation of the invariants for the
German protocol. We start with the initial invariant that for all
agents, some flow is enabled. This can be stated as: \[true
\rightarrow \forall i \in \IndexSet \, \g(i)\tag{inv-1}.\]
\nop{
\begin{equation}
\forall i:\g(i), \tag{inv-1}\label{inv1}
\end{equation}
}
This invariant fails on model checking with 3 caches. A
counterexample is returned in which the \Exclusive flow of cache 1 was
expected to be enabled but is not since the \Exclusive flow of cache 2
is being processed, i.e., the pointer $CurPtr$ points to 2. Thus, on
inspecting the guard of the disabled rule ($RecvReqE$) of the
\Exclusive flow, the invariant is split by selecting the witness
variable $\ptr$ to be $CurPtr$ and the conflict condition $conf$ as
$\lnot(CurCmd=Empty)$. The generated invariants are:
\begin{enumerate}
\item The first invariant states that when $CurPtr$ is $null$, all
flows of all caches are enabled. This is specified as:
\[(CurCmd=Empty) \rightarrow \forall i \in \IndexSet \, \g(i)\tag{inv-1.1}.\]
\item The second invariant states that when $CurPtr$ is not $null$,
the agent pointed to by $CurPtr$ must have some flow enabled. This
can be stated as \(\lnot(CurPtr=null) \rightarrow \g(CurPtr)\). In
order to bring it in the quantified invariant form, we write it
as: \[\lnot(CurCmd=Empty) \rightarrow \forall i \in \ISet{inv-1.2}
\, \g(i)\tag{inv-1.2},\] where $\ISet{inv-1.2} = {i:(i \in
\IndexSet)\land(i=CurPtr)}$. Also, the an assertion checking
$\lnot (CurCmd=Empty) \rightarrow \lnot (CurPtr=null)$ is added.
\end{enumerate}
\nop{
\begin{equation}
(CurPtr=NULL) \rightarrow \forall i: enb(i)\tag{inv-1.1}\label{inv11}
\end{equation}
\begin{equation}
\lnot(CurPtr=NULL) \rightarrow \g(CurPtr) \tag{inv-1.2}\label{inv12}
\end{equation}
}
Next, on model checking the invariants inv-1.1 and inv-1.2, the
invariant inv-1.2 fails. The counterexample trace shows that when the
\Exclusive flow is supposed to be enabled for agent 1 (i.e. when
$CurPtr=1$), it is not enabled. This is due to a conflict with the
\Invalidate flow of agent 2: the guard of the rule $SendGntE$ of the
\Exclusive flow requires all sharers to be invalidated, which is what
the enabled \Invalidate of agent 2 is doing . Thus, the invariant
inv-1.2 is split by selecting the witness $\ptr$ to be an auxiliary
variable $Sharer$ which points to the last sharer added to the set of
sharers, $ShrSet$. The conflict condition checks if some \Invalidate
needs to be sent and is thus derived from the guard of the
$SendInvalidate$ rule and is equal to $(Inv\_Cond$ $\land$ $\lnot
(Sharer=null))$ [CHECK ME]. The two invariants thus obtained are:
\begin{enumerate}
\item The first invariant says that when an \Invalidate is not
required, some flow of agent pointed to by $CurPtr$ is enabled:
\[\lnot(CurCmd=Empty) \land (\lnot Inv\_Cond) \rightarrow \forall i \in \ISet{inv-1.2.1} \, \g(i)\tag{inv-1.2.1},\]
where \ISet{inv-1.2.1} = \ISet{inv-1.2}.
\item The second invariant states that if the condition for
\Invalidate is $true$ (i.e. $Inv\_Cond$ $true$) and there is a
sharer (i.e. $Sharer$ is not $null$), then some flow for agent
pointed by $Sharer$ must be $true$. This is stated as:
\[\lnot(CurCmd=Empty) \land (Inv\_Cond) \rightarrow \forall i \in
\ISet{inv-1.2.2} \, \g(i)\tag{inv-1.2.2},\] where $\ISet{inv-1.2.2}
= \{i|(i \in \IndexSet)\land(i \in ShrSet)\}$.
\end{enumerate}
Next, on model checking, the invariants inv1-1, inv-1.2.1, inv-1.2.2
hold for a model with 3 caches. Then, this set of invariants form a
candidate set to verify for a protcol model with an unbounded number
of agents, in order to proves \mdeadlock freedom. The property is
checked for unbounded agents using techniques described in
\Sec{sec:CMP}.
\subsection{Proof of soundness}
We show that the invariants in \InvSet generated using our method,
collectively imply \mdeadlock freedom. We show this by first stating
the following lemma (proof in Appendix~\ref{sec:proof}):
\newtheorem*{lem1}{Lemma}
\begin{lem1}
The following condition holds for \InvSet: $\bigvee_{inv \in \InvSet}
inv.pred = true$.
\end{lem1}
Intuitively, this condition states that the predicates of all the
invariants in \InvSet add up to \true. Thus, in every state of the
protocol, $inv.pred$ holds for some $inv \in \InvSet$. This condition
enables the proof of the following Theorem (proof in appendix).
\newtheorem*{lem3}{Theorem}
\begin{lem3}
The set of invariants $\InvSet$ collectively imply \mdeadlock freedom,
i.e., $(\forall inv \in \InvSet, \Protocol \models inv) \rightarrow
\Protocol \models (\bigvee_{i} \bigvee_{rl(i)} rl(i).\rho)$.
\end{lem3}
\section{Deriving Invariants for Proving \MDeadlock Freedom}
\label{sec:SpecMethod}
In this section, we show how a set of invariants \InvSet can be
derived from flows such that the invariants in \InvSet collectively
imply \mdeadlock freedom. At a high-level, our method tries to show
\mdeadlock freedom by partitioning the global state of the protocol
using predicates, such that for each partition, some agent $i$ has at
least one transition enabled. Each invariant $inv$ is of the form
$inv.pred \Rightarrow \big(\forall i \in \ISet{inv}: \,
\g(\AgentRuleSet(i))\big)$, where $inv.pred$ is a predicate on the
global variables of the protocol, $\ISet{inv} \subseteq \IndexSet$
s.t. $\lnot(\ISet{inv}=\{\})$ (this is discharged as a separate
assertion for model checking) and $\g(\AgentRuleSet(i))$ denotes a
disjunction of the guards of the rules in $\AgentRuleSet(i)$. {\em The
key insight is that since $\g(\AgentRuleSet(i))$ has transitions
from a single agent, the abstractions required for model checking
$inv$ for an unbounded number of agents are significantly simpler
than those for checking the original \mdeadlock
property,\footnote{\revised{This Footnote is new! }In the case of
rules involving more than one agent (say $c$), the corresponding
invariants may involve transitions from $c$ agents as well. Since
$c$ is small for practical protocols, the abstraction constructed
for verifying such invariants will be simple as well.} as
discussed in \Sec{sec:CMP}.}
\revised{Mostly Redone!} Our method iteratively model checks each
invariant in \InvSet to refine it. Suppose, the invariant $inv \in
\InvSet$ fails on model checking with the state of the protocol at
failure being $s_f$. Then, there exists some agent $i_f$ such that
when $inv.pred$ holds in $s_f$, $i_f \in \ISet{inv}$ is \true and
$\g(\AgentRuleSet(i_f))$ is \false in $s_f$. This can happen due to
two reasons: first, there may be a mismatch between the flow
specification and the rule-based protocol description. This can be due
to a missing rule in some flow, a missing flow all together, or an
implementation error: the cause for the mismatch can be discovered
from the counterexample. As an example for this case, the
counterexample may show that all flows of the agent $i_f$ are not
enabled, however the agent still has some rule $rl_e(i_f)$ enabled:
this rule may be a part of a missing flow. However, typically the
invariant $inv$ fails due to the second reason: there must exist some
flow \flow of the agent $i_f$ which is blocked (i.e. it has a rule
which is expected to be enabled and so has precondition \true but has
its guard \false). This blocked flow is waiting for another flow
\flow' of another agent $i_s$ to complete. As an example, for the
German protocol, the \Exclusive flow may be blocked for agent $i_f$
with the rule $SendGntE(i_f)$ having precondition \true but guard
\false and waiting for an \Invalidate request to complete for another
agent $i_s$ in the set $Sharers$. In this case, the set \InvSet is
refined by splitting the invariant $inv$.
The invariant $inv$ is split by, (1) splitting the predicate
$inv.pred$ to further partition the global state, and (2) updating the
set $\ISet{inv}$ for each partition. To accomplish this, the user
identifies a pointer variable from $G_P$ or $L_P[i]$ (or an auxiliary
variable) $\ptr$, such that it has the value $i_s$ in the failing
state $s_f$ (and so acts as a {\em witness} variable for $i_s$). The
user also identifies a conflict condition $conf$ on the global state
which indicates when $i_s$ is enabled and $i_f$ fails. This is done by
using the heuristic that if the rule $rl_f(i_f)$ of flow \flow of
agent $i_f$ is blocked, $conf$ can be derived by inspecting the guard
of $rl_f(i_f)$; the condition $conf$ generally is the cause for
falsification of $rl_f(i_f).\rho$.
For example, for the German protocol, $conf$ is derived from the guard
of $SendGntE$ and $\ptr$ points to some sharer which is being
invalidated.
Using $conf$ and $\ptr$, the invariant can be split into two
invariants. (1) The first invariant excludes the case when conflict
happens from the original invariant, i.e., $inv1: (inv.pred \land
\lnot conf) \Rightarrow \big(\forall i \in \ISet{inv1}: \,
\g(\AgentRuleSet(i))\big)$, where $\ISet{inv1} = \ISet{inv}$. (2) The
second invariant shows that when a conflict happens, the agent pointed
to by \ptr must be enabled and so the protocol is still
\mdeadlock-free, i.e., $inv2:$ $(inv.pred \land conf) \Rightarrow
\big(\forall i \in \ISet{inv2} :\, \g(\AgentRuleSet(i))\big)$, where
$\ISet{inv2} = \{i| \, (i \in \IndexSet) \land (i=\ptr)\}$. For both
the invariants, assertions which check that the corresponding set of
indices are non-empty are also verified. For example, for $inv1$, this
assertion is $(inv.pred \land \lnot conf) \Rightarrow \ISet{inv1}$.
\nop{
some rule $rl_e(i_f)$ of the agent $i_f$ is still enabled, even when
the invariant fails. This occurs due to a mismatch between the flow
specification and the protocol implementation---the rule $rl_e(i_f)$
is not present in any flow. This may be either due to an entire flow
missing from the description, or an error in the description of some
flow which missed out $rl_e(i_f)$. Thus, the protocol or the flows
have to be fixed. In the second case,
}
Our method derives these invariants by iteratively model checking with
a small number $c$ ($3$ for German protocol) of agents. (Once the
invariants are derived for $c$ agents, they are verified for an
unbounded number of agents, as shown is \Sec{sec:CMP}.) This number
$c$ needs to be chosen to be large enough such that the proof of
\mdeadlock freedom is expected to generalize to an unbounded number of
agents. For the protocols we verified, we found that as a heuristic,
$c$ should be one more than the maximum number of agents involved in
processing a high-level request. For the German protocol, an
\Exclusive request may involve two agents, a requesting agent $i$ and
an agent $j$ getting invalidated, so we chose $c$ to be equal to 3.
\Fig{fig:MethodOverview} shows the details of the method. It starts
with an initial broad guess invariant, $\true \Rightarrow \big(\forall
i \in \IndexSet: \, \g(\AgentRuleSet(i))\big)$ (line 1). This
indicates that in all reachable states, every agent has at least one
transition enabled. As this invariant is \false, this broad guess
invariant is refined into finer invariants, using the loop. On
finishing, the user is able to derive a set of invariants, $\InvSet$,
which collectively imply \mdeadlock freedom. Further, the user is also
able to derive an assertion set, $\AssertSet$, such that for each
invariant $inv$ in $\InvSet$, an assertion in $\AssertSet$ checks if
the set of indices $\ISet{inv}$ is non-empty when $inv.pred$ holds.
\input{splitMethod.tex}
{\bf \emph{Soundness of the Method}} \TechnicalVersion{The following
theorem shows that the invariants in \InvSet along with the
assertions in \AssertSet collectively imply \mdeadlock freedom, with
proof in Appendix~\ref{sec:proof}.} \CameraReady{The following
theorem (proof in the extended version~\cite{DJDeadlockPaperArXiv})
shows that the invariants in \InvSet along with the assertions in
\AssertSet collectively imply \mdeadlock freedom.}
\nop{\newtheorem*{lem1}{Lemma}
\begin{lem1}
The disjunction of predicates of all invariants in \InvSet holds,
i.e., $\bigvee_{inv \in \InvSet} inv.pred$ holds.
\end{lem1}
This condition intuitively implies that the preconditions of all the
invariants in $\InvSet$ partition the global state of the protocol.
This enables the proof of the following Theorem (proof in
Appendix~\ref{sec:proof}):
}
\newtheorem*{lem3}{Theorem}
\begin{lem3}
If the set of invariants $\InvSet$ along with the set of assertions
\AssertSet hold, they collectively imply \mdeadlock freedom, i.e.,
$\big(\big(\bigwedge_{inv \in \InvSet} (\Protocol \models inv)\big)$
$\land$ $\big(\bigwedge_{asrt \in \AssertSet} (\Protocol \models
asrt)\big)\big)$ $\Rightarrow \big(\Protocol \models (\bigvee_{i}
\bigvee_{j} rl_j(i).\rho)\big)$.
\end{lem3}
\subsection{Specifying Invariants for the German Protocol}
\label{sec:GermanInvariantGeneration}
\revised{This subsection was rewritten} We derive the invariants for a
model of the German protocol with 3 cache agents. We start with the
initial invariant that for all agents, some flow is enabled, i.e.,
\Inv{inv-1}: \(true \Rightarrow \big(\forall i \in \IndexSet: \,
\g(\AgentRuleSet(i))\big)\).
{\bf \emph{Iteration 1:}} Model checking the invariant \Inv{inv-1}
returns a counterexample trace ($SendReqE(1)$, $RecvReqE(1)$,
$SendReqE(2)$). Since the index of the last rule in the trace is $2$,
\g(\AgentRuleSet(2)) must be \false. This is because the rule
$RecvReqE(2)$ of the \Exclusive flow of cache 2 is not fired and thus
has precondition \true but guard \false. The user identifies the
conflict condition $conf = \lnot(CurCmd=Empty)$ from the guard of the
blocked rule $RecvReqE(2)$. Since $CurPtr$ is the witness pointer in
the protocol for the variable $CurCmd$, the witness $\ptr$ is set to
$CurPtr$. Thus, the invariant is split as follows:
\begin{itemize}
\item \Inv{inv-1.1}: \((CurCmd=Empty) \Rightarrow (\forall i \in
\IndexSet: \, \g(\AgentRuleSet(i)))\).
\item \Inv{inv-1.2}: \(\lnot(CurCmd=Empty) \Rightarrow (\forall i \in
\ISet{inv-1.2}: \, \g(\AgentRuleSet(i)))\), where $\ISet{inv-1.2} =
\{i| \, (i \in \IndexSet)\land(i=CurPtr)\}$. The assertion $\lnot
(CurCmd=Empty)$ $\Rightarrow$ $\lnot (\ISet{inv-1.2}=\{\})$ is also
checked.
\end{itemize}
{\bf \emph{Iteration 2:}} Next, on model checking the invariants
\Inv{inv-1.1} and \Inv{inv-1.2}, the invariant \Inv{inv-1.2}
fails. The counterexample trace returned is ($SendReqE(1)$,
$RecvReqE(1)$, $SendGntE(1)$, $SendReqE(2)$, $RecvReqE(2)$,
$SendReqE(2)$). Since the last rule of the counterexample is from
cache 2, \g(\AgentRuleSet(2)) must be \false even when
$CurPtr=2$. Further, there are two flows for two \Exclusive requests
by cache 2 active in the counterexample, the first with $SendReqE(2)$
fired and the second with $SendReqE(2)$, $RecvReqE(2)$ fired. Since
the first flow is blocked on the rule $RecvReqE(2)$, the guard of this
rule is inspected. The guard is \false as $CurCmd$ is not
empty. However, since the corresponding witness variable for $CurCmd$
is $CurPtr$ which is already $2$ (due to the processing of the second
flow), this is not a conflict with another cache. The conflict must
then be for the second \Exclusive flow. The second flow is blocked on
the rule $SendGntE(2)$ with precondition \true but guard \false: the
user identifies the conflict condition $conf$ from the guard of
$SendGntE$ to be $Inv\_Cond$. Now, if $Inv\_Cond$ is \true, the
\Invalidate flow for some sharer cache (cache $1$ in this trace) must
be active. Thus, the user identifies $\ptr$ to point to a sharer which
must be invalidated: this is done using the auxiliary variable
$Sharer$, which points to the last sharer to be invalidated in
$ShrSet$. Thus, the invariant \Inv{inv-1.2} is split as follows:
\nop{is due to a conflict with the \Invalidate flow of agent
2: the guard of the rule $SendGntE$ of the \Exclusive flow requires
all sharers to be invalidated, which is what the enabled \Invalidate
of agent 2 is doing . Thus, the invariant inv-1.2 is split by
selecting the witness $\ptr$ to be an auxiliary variable $Sharer$
which points to the last sharer added to the set of sharers, $ShrSet$.
The conflict condition checks if some \Invalidate needs to be sent and
is thus derived from the guard of the $SendInvalidate$ rule and is
equal to $(Inv\_Cond$ $\land$ $\lnot (Sharer=null))$ [CHECK ME]. The
two invariants thus obtained are:
}
\begin{itemize}
\item \Inv{inv-1.2.1}: \(\big(\lnot(CurCmd=Empty) \land (\lnot
Inv\_Cond)\big) \Rightarrow (\forall i \in \ISet{inv-1.2.1}: \,
\g(\AgentRuleSet(i)))\), where, \ISet{inv-1.2.1} =
\ISet{inv-1.2}. An assertion that the precondition implies the index
set is non-empty is also checked.
\item \Inv{inv-1.2.2}: \(\big(\lnot(CurCmd=Empty) \land
(Inv\_Cond)\big) \Rightarrow (\forall i \in \ISet{inv-1.2.2}: \,
\g(\AgentRuleSet(i)))\), where, $\ISet{inv-1.2.2} = \{i| \, (i \in
\IndexSet)\land(i \in ShrSet)\}$. An assertion that the precondition
implies the index set is non-empty is also checked.
\end{itemize}
{\bf \emph{Iteration 3:}} Next, on model checking, the invariants
\Inv{inv-1.1}, \Inv{inv-1.2.1}, \Inv{inv-1.2.2}, along with the added
assertions hold for a model with 3 caches. Then, to prove \mdeadlock
freedom, this set of invariants form a candidate set to verify a
protocol model with an unbounded number of agents. The property is
checked for unbounded agents using techniques described in
\Sec{sec:CMP}.
|
\section{Introduction}
\label{sec:Introduction}
Magnetization reversal in a single-domain ferromagnetic nanostructure
is an important phenomenon for both fundamental physics and applications.
The conventional method for reversing magnetization
is to apply a direct magnetic field to a ferromagnet
along the reversed direction,
where the field magnitude $H$ should be larger than
the anisotropy field (or coercivity) $H_{\rm K}$
to energetically stabilize the reversed state \cite{hubert98}.
However, this method requires a large field $H$
anti-parallel to the magnetization,
as well as large power consumption, for the reversal
because ferromagnets with large $H_{\rm K}$ are used in practical applications
to keep the high thermal stability $\Delta_{0}=MH_{\rm K}V/(2k_{\rm B}T)$,
where $M$, $V$, and $T$ are the magnetization, the volume of the ferromagnet, and the temperature,
respectively.
Recently alternative methods,
such as spin-torque-induced magnetization reversal \cite{slonczewski96,berger96,katine00,zhang02,kiselev03,kubota05,deac06,taniguchi08,taniguchi09}
and microwave-assisted magnetization reversal (MAMR)
\cite{bertotti01,bertotti01a,thirion03,denisov06,sun06,zhu08,bertotti09,bertotti09book,okamoto08,okamoto10,okamoto12,okamoto13,okamoto14,barros11,barros13,cai13},
have been proposed to reduce the reversal field magnitude.
The optical magnetization reversal with circularly polarized light \cite{stanciu07,gevaux07} is
another possibility,
where the combination of the ultrafast heating and the magnetic field,
both of which are generated by the circularly polarized laser,
enables the ultrafast magnetization reversal without an external field.
In MAMR, the microwave produces a circularly rotating magnetic field
applied to the ferromagnet,
in which the field direction lies in a plane
perpendicular to the easy axis.
A rotating frame is conventionally used to understand
why the reversal field becomes smaller than $H_{\rm K}$ in MAMR \cite{bertotti09book,okamoto13}.
In the rotating frame,
the field acting on the magnetization is independent of time.
The effect of the rotating field is converted to
an additional dc magnetic field $(2\pi f/\gamma)$ pointing in the reversed direction \cite{bertotti01,denisov06,bertotti09book},
where $f$ and $\gamma$ are the frequency of the rotating field and the gyromagnetic ratio, respectively,
i.e., the total dc magnetic field pointing in the reversed direction is $H+(2\pi f/\gamma)$.
The additional field $(2\pi f/\gamma)$ energetically stabilizes the reversed state,
and reduces the reversal field magnitude.
In a low-frequency region,
the reversal field linearly decreases as the frequency increases,
which is qualitatively consistent with this conventional picture.
However, both the experiments and the numerical simulations of
the Landau-Lifshitz-Gilbert (LLG) equation have revealed that
such a conventional picture cannot explain the dependence of the reversal field on the frequency
in a high-frequency region \cite{zhu08,okamoto08,okamoto12,okamoto13}.
In the high-frequency region,
the reversal field slightly increases as the frequency increases;
see, for example, Fig. \ref{fig:fig5} below.
Moreover, the magnitude of the total dc magnetic field for the reversal, $H+(2\pi f/\gamma)$,
in the high-frequency region is much larger than $H_{\rm K}$.
This result seems like in contradiction with
the Stoner-Wohlfarth theory \cite{hubert98},
in which the magnetization reversal should occur
when $H+(2\pi f/\gamma)$ becomes slightly larger than $H_{\rm K}$
because only the reversed state is energetically stable.
Okamoto \etal studied the dependence of the reversal field on the frequency of the rotating field
for a Co/Pt nanodot with 120 nm diameter
both experimentally and numerically based on the micromagnetic model \cite{okamoto12}.
They found that the excitation of the spin wave,
which arises from the difference in the local demagnetization field
between the end and center of the dots,
leads to a reduction of the reversal field.
This result is of great advance in understanding the reversal mechanism of MAMR.
However, it is still unclear why the reversal field jumps to a high value at a certain frequency.
Moreover, the numerical simulations based on the macrospin (single domain) model also
show the jump of the reversal field \cite{zhu08,okamoto12},
indicating that not only the excitation of the spin wave
but also other mechanisms lead to this jump.
A fabrication of the ferromagnet smaller than the exchange length (typically \cite{okamoto13}, on the order of 10 nm)
is an unavoidable and indispensable challenge
for both fundamental physics and practical applications.
In such nanostructure, the magnetization dynamics is well described by the macrospin model.
Therefore, it is important to clarify the magnetization reversal mechanism,
such as the origin of the jump of the reversal field,
using the macrospin model.
The purposes of this paper are
to explain why a large field is required to reverse the magnetization
in the high-frequency region in MAMR
and to derive equations that determine the reversal field
in both the low- and high-frequency regions
without the time-dependent solution of the macrospin LLG equation.
It is pointed out that
the rotating field not only energetically stabilizes the reversed state
but also produces a torque acting on the magnetization.
We find that this torque,
whose strength is proportional to the frequency of the rotating field,
prevents the reversal,
causing the reversal field to become
relatively large in the high-frequency region.
The direction of this preventing torque is expressed
by the triple vector product,
analogous to the spin torque \cite{slonczewski96,berger96}.
This fact motivates us
to use the energy balance equation
for the estimation of the reversal field of MAMR,
which was recently pointed out to be useful
for estimating the reversal current
in the spin-torque reversal problem \cite{hillebrands06,newhall13,pinna13,taniguchi13,taniguchi13a,taniguchi13b,taniguchi14}
but has been never applied to the MAMR problem.
The equations determining the reversal field,
Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high}), are derived
for both the low- and high-frequency regions
from the energy balance equation.
These formulas show that
the reversal field in the low-frequency region
converges to Eq. (\ref{eq:condition_low_zero_damping})
as the damping constant decreases,
while the reversal field in the high-frequency region
is independent of the damping constant.
The boundary between the low- and high-frequency regions is
also estimated from the energy balance equation.
The validities of these formulas are quantitatively confirmed by comparison with
the numerical simulation of the LLG equation.
The paper is organized as follows.
In Sec. \ref{sec:Landau-Lifshitz-Gilbert equation in rotating frame},
the energy balance equation is derived from the LLG equation
in the rotating frame.
The equations determining the reversal fields
in the low- and high-frequency regions are derived
in Secs \ref{sec:Reversal in low frequency region} and \ref{sec:Reversal in high frequency region}, respectively.
The validities of these formulas over a wide frequency range are confirmed in
Sec. \ref{sec:Comparison with numerical simulation}
by comparison with the numerical simulation of the LLG equation.
In Sec. \ref{sec:Comparison with other work},
the present result is compared with the previous work in Refs. \cite{bertotti01,bertotti01a,bertotti09,bertotti09book}.
The conclusion appears in Sec. \ref{sec:Conclusion}.
\section{Landau-Lifshitz-Gilbert equation in rotating frame}
\label{sec:Landau-Lifshitz-Gilbert equation in rotating frame}
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig1.eps}}\vspace{-3.0ex}
\caption{
Schematic view of the system.
The unit vector pointing in the magnetization direction is denoted as $\mathbf{m}$.
The external magnetic field pointing in the negative $z$ direction
and the rotating field with the frequency $f$ are denoted as $H$ and $H_{\rm ac}$, respectively.
In the rotating frame, $x^{\prime}$ and $z^{\prime}$ axes are
parallel to the rotating field and the $z$ axis, respectively.
\vspace{-3ex}}
\label{fig:fig1}
\end{figure}
The system we consider is schematically shown in Fig. \ref{fig:fig1},
where the unit vector pointing in the magnetization direction is denoted as $\mathbf{m}$.
The ferromagnet has a uniaxial easy axis
with the anisotropy field $H_{\rm K}$ along the $z$ axis.
Throughout this paper,
the initial state is taken to be $\mathbf{m}=+\mathbf{e}_{z}$,
although the following formulas are applicable to
an arbitrary initial condition.
The external field $H$ is applied to the negative $z$ direction.
The rotating field with the magnitude $H_{\rm ac}$ and the frequency $f$ is applied
in the $xy$ plane,
where the $x$ axis is parallel to the rotating field
at the initial time $t=0$.
The magnetization dynamics
under the effect of the magnetic field,
$\mathbf{H}=H_{\rm ac} \cos (2\pi ft) \mathbf{e}_{x}
+H_{\rm ac} \sin (2\pi ft) \mathbf{e}_{y}
+(-H + H_{\rm K}m_{z})\mathbf{e}_{z}$
are described by
the LLG equation \cite{landau35,landau80,gilbert04},
\begin{equation}
\frac{d \mathbf{m}}{dt}
=
-\gamma
\mathbf{m}
\times
\mathbf{H}
+
\alpha
\mathbf{m}
\times
\frac{d \mathbf{m}}{dt},
\label{eq:LLG_orig}
\end{equation}
where the Gilbert damping constant is denoted as $\alpha$.
Because the LLG equation conserves the magnitude of the magnetization,
the magnetization dynamics are described by the trajectory
on the surface of the unit sphere.
It is convenient to use
the rotating frame $x^{\prime}y^{\prime}z^{\prime}$,
in which the $z^{\prime}$ axis is parallel to the $z$ axis,
and $x^{\prime}$ axis is parallel to the rotating field $H_{\rm ac}$,
as shown in Fig. \ref{fig:fig1}.
We denote $\mathbf{m}$ in the rotating frame as $\mathbf{m}^{\prime}=(m_{x^{\prime}},m_{y^{\prime}},m_{z^{\prime}})$.
It should be noted that the value of $m_{z}$ is invariant by this transformation.
Because the value of $\alpha$ in the conventional ferromagnet is small \cite{oogane06},
higher order terms of $\alpha$ are neglected in the following;
i.e., we use the approximation that $1+\alpha^{2}\simeq 1$.
Then, the LLG equation in the rotating frame is given by
\begin{equation}
\begin{split}
\frac{d \mathbf{m}^{\prime}}{dt}
=&
-\gamma
\mathbf{m}^{\prime}
\times
\bm{\mathcal{B}}
-
\alpha
\gamma
\mathbf{m}^{\prime}
\times
\left(
\mathbf{m}^{\prime}
\times
\bm{\mathcal{B}}
\right)
\\
&+
\alpha
2\pi f
\mathbf{m}^{\prime}
\times
\left(
\mathbf{e}_{z^{\prime}}
\times
\mathbf{m}^{\prime}
\right),
\label{eq:LLG}
\end{split}
\end{equation}
where $\bm{\mathcal{B}}=H_{\rm ac} \mathbf{e}_{x^{\prime}} + [-H-(2\pi f/\gamma)+H_{\rm K}m_{z^{\prime}}] \mathbf{e}_{z^{\prime}}$
can be regarded as the magnetic field in the rotating frame.
The transformation procedure from Eq. (\ref{eq:LLG_orig}) to Eq. (\ref{eq:LLG})
is shown in Appendix A.
It can be understood from Eq. (\ref{eq:LLG}) that
the rotating field plays two roles for the reversal.
First, the magnetization dynamics can be regarded as
a motion of a point particle
in the potential $\mathscr{E}= -M \int d \mathbf{m} \cdot \bm{\mathcal{B}}$,
\begin{equation}
\mathscr{E}
=
-M H_{\rm ac}
m_{x^{\prime}}
+
M
\left(
H
+
\frac{2\pi f}{\gamma}
\right)
m_{z^{\prime}}
-
\frac{MH_{\rm K}}{2}
m_{z^{\prime}}^{2}.
\label{eq:potential}
\end{equation}
The second term on the right-hand side of Eq. (\ref{eq:potential})
indicates that
the rotating field produces the dc magnetic field $(2\pi f/\gamma)$
pointing in the negative $z^{\prime}$ direction,
and energetically stabilizes the reversed state \cite{bertotti09,bertotti09book,okamoto13}.
Second, the rotating field produces a torque
proportional to the frequency $f$,
which appears in the third term on the right-hand side of Eq. (\ref{eq:LLG}).
The important point is that
this torque points to the positive $z^{\prime}$ direction,
and therefore, prevents the reversal.
It should also be emphasized that
the torque direction is expressed by the triple vector product,
as is similar to the spin torque \cite{slonczewski96,berger96}.
Therefore, in the following calculations,
let us conventionally call this torque spin torque.
It was shown in the spin-torque reversal problem \cite{hillebrands06,newhall13,pinna13,taniguchi13,taniguchi13a,taniguchi13b,taniguchi14} that
the magnetization reversal condition can be derived
from the energy balance equation
between the work done by spin torque
and the energy dissipation due to the damping.
In the following sections,
we apply this method to investigate
the reversal field in MAMR.
To this end,
the derivative of $\mathscr{E}$ with respect to time
on the constant energy curve is necessary.
From Eq. (\ref{eq:LLG}), we find that
\begin{equation}
\begin{split}
\frac{d \mathscr{E}}{dt}
=&
-\alpha
2\pi f M
\left[
\bm{\mathcal{B}}
\cdot
\mathbf{e}_{z^{\prime}}
-
\left(
\mathbf{m}^{\prime}
\cdot
\mathbf{e}_{z^{\prime}}
\right)
\left(
\mathbf{m}^{\prime}
\cdot
\bm{\mathcal{B}}
\right)
\right]
\\
&-
\alpha
\gamma
M
\left[
\bm{\mathcal{B}}^{2}
-
\left(
\mathbf{m}^{\prime}
\cdot
\bm{\mathcal{B}}
\right)^{2}
\right].
\label{eq:dEdt}
\end{split}
\end{equation}
The integral of Eq. (\ref{eq:dEdt}) over a precession period of the magnetization
on the constant energy curve of $\mathscr{E}$ is
$\oint dt (d \mathscr{E}/dt)=\mathscr{W}_{\rm s}+\mathscr{W}_{\alpha}$,
where
\begin{equation}
\mathscr{W}_{\rm s}
=
-\alpha
2\pi f M
\oint dt
\left[
\bm{\mathcal{B}}
\cdot
\mathbf{e}_{z^{\prime}}
-
\left(
\mathbf{m}^{\prime}
\cdot
\mathbf{e}_{z^{\prime}}
\right)
\left(
\mathbf{m}^{\prime}
\cdot
\bm{\mathcal{B}}
\right)
\right],
\label{eq:W_s}
\end{equation}
\begin{equation}
\mathscr{W}_{\alpha}
=
-\alpha
\gamma
M
\oint dt
\left[
\bm{\mathcal{B}}^{2}
-
\left(
\mathbf{m}^{\prime}
\cdot
\bm{\mathcal{B}}
\right)^{2}
\right],
\label{eq:W_alpha}
\end{equation}
are the work done by spin torque
and the energy dissipation due to the damping
during the precession, respectively \cite{perko91}.
While $\mathscr{W}_{\rm s}$ can be both positive and negative
depending on the field and the frequency,
$\mathscr{W}_{\alpha}$ is always negative.
The calculation procedures of Eqs. (\ref{eq:W_s}) and (\ref{eq:W_alpha})
without the time-dependent solution of Eq. (\ref{eq:LLG})
are shown in Appendix B.
The damping constant $\alpha$ is assumed to be scalar
in the above formulation.
On the other hand, Safonov studied the magnetization relaxation near equilibrium
with a tensor damping \cite{safonov02}.
The presence of the tensor damping was also suggested
in the spin torque problem \cite{zhang09}.
In Appendix C, we discuss how the formulas derived in the following sections are modified
when the scalar damping $\alpha$ is replaced by the tensor damping.
\section{Reversal in low-frequency region}
\label{sec:Reversal in low frequency region}
In this section,
we study the reversal field in the low-frequency region.
Let us first show in Fig. \ref{fig:fig2} (a)
the trajectory of typical magnetization dynamics
in the low-frequency region
obtained by numerically solving Eq. (\ref{eq:LLG}).
The time evolution of $m_{z^{\prime}}$ is shown in Fig. \ref{fig:fig2} (b).
The values of the parameters are
$M=1000$ emu/c.c.,
$H_{\rm K}=7.5$ kOe,
$H_{\rm ac}=450$ Oe,
$\gamma=1.764\times 10^{7}$ rad/(Oe$\cdot$s),
$f=2.0$ GHz,
and $\alpha=0.01$,
which are typical values used in the experiments and the numerical simulations \cite{denisov06,zhu08,okamoto08,okamoto12,okamoto13}.
We judged that the magnetization is reversed
when the condition $m_{z^{\prime}}<-0.9$ is satisfied.
The minimum field satisfying this condition is $H=4.709$ kOe.
Starting from the initial state $\mathbf{m}^{\prime}=+\mathbf{e}_{z^{\prime}}$,
the magnetization precesses around an axis
lying in the positive $x^{\prime}z^{\prime}$ plane.
After a half period of precession,
the magnetization changes the precession direction,
and falls into the reversed state.
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig2.eps}}\vspace{-3.0ex}
\caption{
(a) Schematic view of the magnetization dynamics in the rotating frame
at the reversal field $H=4.709$ kOe,
where the frequency of the rotating field is $f=2.0$ GHz.
The trajectory is described on the unit sphere.
(b) The time evolution of $m_{z^{\prime}}$.
(c) Schematic view of the potential $\mathscr{E}$ (normalized by $MH_{\rm K})$
in the $x^{\prime}z^{\prime}$ plane.
The horizontal axis represents the angle $\theta$ of the magnetization from the $z^{\prime}$ axis.
(d) The enlarged view of panel (c) around $[\theta_{\rm initial},\theta_{\rm saddle}]$.
The directions of the damping and spin torque are indicated by arrows.
\vspace{-3ex}}
\label{fig:fig2}
\end{figure}
Next, let us analytically derive the equation determining the reversal field.
Figure \ref{fig:fig2} (c) shows the map of the potential $\mathscr{E}$
in the $x^{\prime}z^{\prime}$ plane,
where the horizontal axis is the angle $\theta$ of the magnetization from the $z^{\prime}$ axis.
The potential with $H=4.709$ kOe has metastable,
saddle, and stable points at
$\theta_{\rm metastable}=14^{\circ}$,
$\theta_{\rm saddle}=36^{\circ}$,
and $\theta_{\rm stable}=178^{\circ}$, respectively.
The directions of the damping and the spin torque
between the initial state ($\theta_{\rm initial}=0^{\circ}$) and the saddle point
in the potential $\mathscr{E}$
are indicated by
the arrows in Fig. \ref{fig:fig2} (d).
The spin torque supplies the energy to the ferromagnet
when the magnetization is in $\theta_{\rm initial} \le \theta \le \theta_{\rm metastable}$
because it is anti-parallel to the damping,
while the spin torque dissipates the energy when the magnetization is in $\theta_{\rm metastable} \le \theta \le \theta_{\rm saddle}$
because it is parallel to the damping.
The function $\mathscr{W}_{\rm s}$ in $\theta_{\rm initial} \le \theta \le \theta_{\rm saddle}$ is negative
($\mathscr{W}_{\rm s}<0$)
because the spin torque magnitude ($\propto \sin\theta$) increases
as the angle $\theta(<90^{\circ})$ increases.
Therefore, the spin torque totally dissipates the energy during the precession.
The magnetization reversal occurs
when the magnitude of
the energy $\Delta \mathscr{E}=\int dt (d \mathscr{E}/dt)$ dissipated
during the dynamics from $\theta_{\rm initial}$ to $\theta_{\rm saddle}$
is smaller than the energy difference
between the initial state and the saddle point,
$\mathscr{E}_{\rm initial}-\mathscr{E}_{\rm saddle}$,
i.e., the reversal condition is
\begin{equation}
\mathscr{E}_{\rm initial}
-
\mathscr{E}_{\rm saddle}
\ge
-\Delta \mathscr{E}.
\label{eq:reversal_condition_low}
\end{equation}
Strictly speaking,
the time-dependent solution of $\mathbf{m}^{\prime}$ is necessary to calculate $\Delta \mathscr{E}$.
However, in the low-frequency region,
because the difference between $\mathscr{E}_{\rm initial}$ and $\mathscr{E}_{\rm saddle}$ is small,
$\Delta \mathscr{E}$ can be approximated to
$[\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle}) + \mathscr{W}_{\alpha}(\mathscr{E}_{\rm saddle})]/2$.
The numerical factor $1/2$ appears
because the reversal occurs after the half period of the precession.
Therefore, the reversal field $H_{\rm reversal}$ can be defined as
the field $H$ satisfying the condition
\begin{equation}
\mathscr{E}_{\rm initial}
-
\mathscr{E}_{\rm saddle}
=
-\frac{1}{2}
\left[
\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle})
+
\mathscr{W}_{\alpha}(\mathscr{E}_{\rm saddle})
\right].
\label{eq:condition_low}
\end{equation}
Equation (\ref{eq:condition_low}) is the main result in this section.
The reversal field estimated from Eq. (\ref{eq:condition_low}) is 4.708 kOe,
which is almost identical to that (4.709 kOe) obtained
from the numerical solution of the LLG equation (\ref{eq:LLG}).
It should be noted that
the approximation $\Delta \mathscr{E} \simeq [\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle}) + \mathscr{W}_{\alpha}(\mathscr{E}_{\rm saddle})]/2$
works well for small $\alpha$
because the magnetization dynamics occurs almost on the constant energy curve
when $\alpha \ll 1$.
In the zero-damping limit,
the reversal field is estimated from the condition $\mathscr{E}_{\rm initial}-\mathscr{E}_{\rm saddle}=0$,
and is given by
\begin{equation}
H_{\rm reversal}
=
\frac{H_{\rm K}}{2}
\sin^{2}\theta_{\rm saddle}
\left(
\frac{1}{\cos\theta_{\rm saddle}}
-
1
\right)^{-1}
-
\frac{2\pi f}{\gamma},
\label{eq:condition_low_zero_damping}
\end{equation}
where $\theta_{\rm saddle}$ depends on $H_{\rm reversal}$
through the condition
$H_{\rm ac}\cos\theta_{\rm saddle}+[H_{\rm reversal}+(2\pi f/\gamma)]\sin\theta_{\rm saddle}-H_{\rm K} \sin\theta_{\rm saddle}\cos\theta_{\rm saddle}=0$.
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig3.eps}}\vspace{-3.0ex}
\caption{
Schematic view of the potential $\mathscr{E}^{\prime}=\mathscr{E}(H=H^{\prime})$
with $(H^{\prime},f)=(3.671 {\rm \ kOe},6.1 {\rm \ GHz})$.
The arrows indicate the directions of the spin torque and the damping.
\vspace{-3ex}}
\label{fig:fig3}
\end{figure}
Let us also elaborate on the frequency range
in which Eq. (\ref{eq:condition_low}) obtains
good agreement with the numerical solution of the LLG equation (\ref{eq:LLG}).
At a certain field magnitude $H^{\prime}$,
which is larger than $H_{\rm reversal}$ estimated by Eq. (\ref{eq:condition_low}),
the metastable state disappears,
and the potential has only one stable point and a saddle point.
We denote the potential $\mathscr{E}$ with $H=H^{\prime}$ as $\mathscr{E}^{\prime}=\mathscr{E}(H=H^{\prime})$,
which satisfies
$\partial \mathscr{E}^{\prime}/\partial\theta=\partial^{2}\mathscr{E}^{\prime}/\partial\theta^{2}=0$
at the saddle point $\theta_{\rm saddle}^{\prime}$.
According to the conventional Stoner-Wohlfarth theory \cite{hubert98},
the magnetization reversal should occur
because the potential has only one minimum.
However, in the present case,
the work done by spin torque, $\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle}^{\prime})$,
on the constant energy curve of
$\mathscr{E}^{\prime}(\theta_{\rm saddle}^{\prime})=\mathscr{E}_{\rm saddle}^{\prime}$,
is positive
because the direction of the spin torque is always opposite
that of the damping, as shown in Fig. \ref{fig:fig3}.
Then, the condition
\begin{equation}
\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle}^{\prime})
+
\mathscr{W}_{\alpha}(\mathscr{E}_{\rm saddle}^{\prime})
\le
0,
\label{eq:condition_boundary}
\end{equation}
should also be satisfied to reverse the magnetization:
if Eq. (\ref{eq:condition_boundary}) is not satisfied,
the spin torque preventing the reversal overcomes the damping,
and thus the magnetization cannot reverse its direction.
It is found that Eq. (\ref{eq:condition_boundary}) is satisfied for $f<6.2$ GHz
for the above parameters.
Therefore, we define the low-frequency region
in which Eq. (\ref{eq:condition_low}) is valid as $f<6.2$ GHz.
Because the reversal field discontinuously becomes large above this frequency,
we call this frequency the jump frequency.
It should be emphasized that
the jump frequency is independent of $\alpha (\ll 1)$.
\section{Reversal in high-frequency region}
\label{sec:Reversal in high frequency region}
In this section,
we study the magnetization reversal in the high-frequency region.
As mentioned in the previous section,
for $f \ge 6.2$ GHz,
the spin torque preventing the reversal becomes sufficiently large.
Then, a large field is required
to reverse the magnetization.
Figure \ref{fig:fig4} (a) shows
the trajectory of typical magnetization dynamics
in the high-frequency region
obtained by numerically solving Eq. (\ref{eq:LLG}).
The time evolution of $m_{z^{\prime}}$ is shown in Fig. \ref{fig:fig4} (b).
The values of the parameters are those used in Sec. \ref{sec:Reversal in low frequency region}
except $f=8.0$ GHz.
The minimum field satisfying the condition $m_{z^{\prime}}<-0.9$ is $H=7.109$ kOe.
Starting from the initial state,
the magnetization precesses on the constant energy curves
near the $z^{\prime}$ axis many times.
The precession amplitude slightly increases with time,
and finally the magnetization reverses to $\mathbf{m}\simeq -\mathbf{e}_{z^{\prime}}$.
The reversal trajectory covers almost all of the unit sphere,
as shown in Fig. \ref{fig:fig4} (a).
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig4.eps}}\vspace{-3.0ex}
\caption{
(a) Schematic view of the magnetization dynamics in the rotating frame
at the reversal field $H=7.109$ kOe,
where the frequency of the rotating field is $f=8.0$ GHz.
The trajectory is described on the unit sphere.
(b) The time evolution of $m_{z^{\prime}}$.
(c) Schematic view of the potential $\mathscr{E}$ (normalized by $MH_{\rm K})$
in the $x^{\prime}z^{\prime}$ plane.
The horizontal axis represents the angle $\theta$ of the magnetization from the $z^{\prime}$ axis.
(d) The enlarged view of $\mathscr{E}$ around $-30^{\circ} \le \theta \le 10^{\circ}$.
The directions of the damping and spin torque are indicated by arrows.
\vspace{-3ex}}
\label{fig:fig4}
\end{figure}
Figure \ref{fig:fig4} (c) shows the potential $\mathscr{E}$ in the $x^{\prime}z^{\prime}$-plane
at the reversal field.
According to the Stoner-Wohlfarth condition \cite{hubert98}
$H_{\rm K}^{2/3}=[H+(2\pi f/\gamma)]^{2/3}+H_{\rm ac}^{2/3}$,
a field $H$ larger than $3.0$ kOe is enough to reverse the magnetization,
above which the potential $\mathscr{E}$ has only one minimum.
Nevertheless,
a large field ($\ge 7.109$ kOe) compared with the Stoner-Wohlfarth condition
is required for the reversal in the present case
because the spin torque prevents the reversal.
Figure \ref{fig:fig4} (d) shows the enlarged view of $\mathscr{E}$ near the initial state,
where the arrows indicate the directions of the damping and the spin torque.
The maximum of $\mathscr{E}$ is located at $\theta_{\rm maximum}=-10^{\circ}$
while the angle satisfying $\mathscr{E}(\theta)=\mathscr{E}(\theta_{\rm initial})$ is located at $\theta=-19^{\circ}$.
For a reason similar to that discussed in Sec. \ref{sec:Reversal in low frequency region},
the function $\mathscr{W}_{\rm s}(\mathscr{E}_{\rm initial})$ is positive.
When $\oint dt (d \mathscr{E}/dt)=\mathscr{W}_{\rm s}(\mathscr{E}_{\rm initial})+\mathscr{W}_{\alpha}(\mathscr{E}_{\rm initial})<0$,
the magnetization loses the energy, and falls into the reversed state.
On the other hand,
when $\mathscr{W}_{\rm s}(\mathscr{E}_{\rm initial})+\mathscr{W}_{\alpha}(\mathscr{E}_{\rm initial})>0$,
the magnetization climbs $\mathscr{E}$ from $\theta_{\rm initial}$ to $\theta_{\rm maximum}$.
Therefore, the reversal field can be defined as the field satisfying the condition
\begin{equation}
\mathscr{W}_{\rm s}(\mathscr{E}_{\rm initial})
+
\mathscr{W}_{\alpha}(\mathscr{E}_{\rm initial})
=
0.
\label{eq:condition_high}
\end{equation}
Equation (\ref{eq:condition_high}) is the main result in this section.
The reversal field estimated from Eq. (\ref{eq:condition_high}) is 7.109 kOe,
which is identical to that obtained from the numerical solution of the LLG equation (\ref{eq:LLG}).
Another important conclusion from Eq. (\ref{eq:condition_high}) is that
the reversal field is independent of $\alpha$
because both $\mathscr{W}_{\rm s}$ and $\mathscr{W}_{\alpha}$ are proportional to $\alpha$.
The validity of this conclusion is investigated in the next section.
\section{Comparison with numerical simulation}
\label{sec:Comparison with numerical simulation}
In this section,
we confirm the validities of Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high})
over a wide range of the frequency $f$.
The circles in Fig. \ref{fig:fig5} show
the reversal field estimated numerically solving the LLG equation (\ref{eq:LLG}),
where the frequency range is $0 < f \le 10$ GHz.
The Gilbert damping constant is $0.01$.
The reversal field magnitude linearly decreases
as the frequency increases for $f \lesssim 6$ GHz.
Above $f \gtrsim 6$ GHz,
the reversal field jumps to a high value
at which $H_{\rm reversal}+(2\pi f/\gamma)>H_{\rm K}$,
and slightly increases as the frequency increases.
The reversal fields obtained from Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high})
are also shown in Fig. \ref{fig:fig5} by the solid lines.
As mentioned in Sec. \ref{sec:Reversal in low frequency region},
Eq. (\ref{eq:condition_low}) is valid for $f<6.2$ GHz.
Therefore, we use Eq. (\ref{eq:condition_low}) for $f<6.2$ GHz
while Eq. (\ref{eq:condition_high}) is used for $f \ge 6.2$ GHz.
Equations (\ref{eq:condition_low}) and (\ref{eq:condition_high}) show good agreement
with the numerical solution of the LLG equation (\ref{eq:LLG}),
indicating the validities of these formulas.
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig5.eps}}\vspace{-3.0ex}
\caption{
The dependence of the reversal field on the frequency of the rotating field.
The circles are obtained from Eq. (\ref{eq:LLG})
while the solid lines are obtained from Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high}).
The value of $\alpha$ is $0.01$.
The dashed line is the reversal field estimated from Ref. \cite{bertotti09},
and is discussed in Sec. \ref{sec:Comparison with other work}.
\vspace{-3ex}}
\label{fig:fig5}
\end{figure}
\begin{figure
\centerline{\includegraphics[width=0.7\columnwidth]{fig6.eps}}\vspace{-3.0ex}
\caption{
(a) The dependences of the reversal fields on the frequency for $\alpha=0.001$, $0.01$, and $0.1$.
The symbols (square, circle, and triangle) are obtained from Eq. (\ref{eq:LLG})
while the lines (dotted, solid, and dashed) are obtained from Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high}).
(b) The enlarged view of panel (a) in the low-frequency region.
\vspace{-3ex}}
\label{fig:fig6}
\end{figure}
Figure \ref{fig:fig6} (a) shows
the dependences of the reversal fields on the frequency
for $\alpha=0.001$, $0.01$, and $0.1$,
where the numerical solution of the LLG equation (\ref{eq:LLG}) is represented by the symbols
(square, circle, and triangle, respectively),
while the reversal fields obtained from Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high}) are
represented by the lines (dotted, solid, and dashed, respectively).
The frequency ($\simeq 6$ GHz) at which the reversal field of the LLG equation (\ref{eq:LLG})
jumps to a high value is independent of $\alpha$,
which is consistent with the discussion in Sec. \ref{sec:Reversal in low frequency region}.
The enlarged view in the low-frequency region is shown in Fig. \ref{fig:fig6} (b).
In the low-frequency region,
the difference between the solutions of the LLG equation (\ref{eq:LLG})
and the energy balance equation (\ref{eq:condition_low}) becomes small
as $\alpha$ decreases,
because the approximation $\Delta\mathscr{E} \simeq [\mathscr{W}_{\rm s}(\mathscr{E}_{\rm saddle}) + \mathscr{W}_{\alpha}(\mathscr{E}_{\rm saddle})]/2$
used in the derivation of Eq. (\ref{eq:condition_low}) is valid
for a sufficiently small $\alpha$.
Also, the reversal field becomes independent of $\alpha$
with decreasing $\alpha$,
which is consistent with Eq. (\ref{eq:condition_low_zero_damping}).
In the high-frequency region,
the solution of the LLG equation (\ref{eq:LLG}) is also independent of $\alpha$,
which is consistent with Eq. (\ref{eq:condition_high}).
These results also imply the validity of Eqs. (\ref{eq:condition_low}) and (\ref{eq:condition_high}).
\begin{figure
\centerline{\includegraphics[width=1.0\columnwidth]{fig7.eps}}\vspace{-3.0ex}
\caption{
The dependence of the jump frequency, which is the boundary between the low- and high-frequency regions,
on the rotating field magnitude $H_{\rm ac}$.
The circles are obtained from Eq. (\ref{eq:LLG})
while the solid line is obtained from Eq. (\ref{eq:condition_boundary}).
\vspace{-3ex}}
\label{fig:fig7}
\end{figure}
As we end this section,
we study the relation between
the rotating field magnitude $H_{\rm ac}$
and the jump frequency, i.e., the frequency defining the boundary
between the low- and high-frequency regions.
Figure \ref{fig:fig7} shows the dependences of the jump frequency on $H_{\rm ac}$
obtained from Eq. (\ref{eq:condition_boundary}) (solid line)
and the numerical solution of the LLG equation (circle).
As shown, the jump frequency monotonically increases with increasing $H_{\rm ac}$.
Although the clarification of the relation between the jump frequency and the other parameters such as $H_{\rm ac}$ is desirable,
it is difficult to analytically solve Eq. (\ref{eq:condition_boundary}) with respect to the jump frequency.
We consider that the jump frequency does not necessarily relate to, for example,
the ferromagnetic resonance (FMR) frequency
because the jump frequency is determined by the energy balance of the magnetization at the saddle point of the potential $\mathscr{E}$
while the FMR frequency is the frequency of the harmonic oscillation around the stable point.
However, a further investigation on the jump frequency is beyond the scope of this paper.
\section{Comparison with other work}
\label{sec:Comparison with other work}
In this section, we compare the above result
with the previous work of Bertotti \etal \cite{bertotti09,bertotti09book}.
They expanded the LLG equation around its steady-state point, $(\theta,\phi)$,
where $\theta$ and $\phi$ are the zenith and azimuth angles characterizing the magnetization direction,
and satisfy the following equations:
\begin{equation}
\gamma
H_{\rm ac}
\sin\phi
-
\alpha
2\pi f
\sin\theta
=
0,
\label{eq:condition_Bertotti_1}
\end{equation}
\begin{equation}
\left(
\gamma
H
+
2\pi f
\right)
\sin\theta
-
\gamma
H_{\rm K}
\sin\theta
\cos\theta
+
\gamma
H_{\rm ac}
\cos\theta
\cos\phi
=
0.
\label{eq:condition_Bertotti_2}
\end{equation}
Small deviation of the magnetization, $(\Delta\theta,\Delta\phi)$, from the steady points
satisfy $d \Delta\theta/dt = \mathsf{A}_{1,1}\Delta\theta + \mathsf{A}_{1,2}\Delta\phi$
and $d \Delta\phi/dt = \mathsf{A}_{2,1}\Delta\theta + \mathsf{A}_{2,2} \Delta\phi$,
where components of a $2 \times 2$ matrix $\mathsf{A}$ is obtained from the LLG equation.
The trace and determinant of $\mathsf{A}$ are given by
\begin{equation}
{\rm Tr}[\mathsf{A}]
=
-\frac{2 \alpha}{1+\alpha^{2}}
\left[
v
-
\frac{\sin^{2}\theta}{2}
\gamma
H_{\rm K}
+
2\pi f
\cos\theta
\right],
\label{eq:trace_Bertotti}
\end{equation}
\begin{equation}
{\rm det}[\mathsf{A}]
=
\frac{v^{2}-\gamma H_{\rm K} v \sin^{2}\theta + (\alpha 2\pi f \cos\theta)^{2}}{1+\alpha^{2}},
\label{eq:determinant_Bertotti}
\end{equation}
where $v=\alpha 2\pi f \cot\phi$.
According to Ref. \cite{bertotti09},
the reversal field is estimated
from Eqs. (\ref{eq:condition_Bertotti_1}), (\ref{eq:condition_Bertotti_2})
and (\ref{eq:determinant_Bertotti}) with the condition ${\rm det}[\mathsf{A}]=0$.
The dashed line in Fig. \ref{fig:fig5} is the reversal field estimated from this method.
As shown, the method of Bertotti \etal reveals larger reversal field in our simulation.
The difference between our and their results arises from the following reason.
In our analytical and numerical calculations,
both the microwave and external field are applied from $t=0$ with the constant magnitudes.
The initial state of the magnetization, $m_{z^{\prime}}=1$, locates above the stable or saddle point,
as shown in Fig. \ref{fig:fig2} (d).
On the other hand, Bertotti \etal considered the instability of the magnetization
around the steady point corresponding to the stable or saddle point.
In this case, a relatively large energy compared with our situation is required to overcome the potential barrier and reverse the magnetization direction.
Therefore, the reversal field estimated from the method of Bertotti \etal becomes larger than our estimation.
We emphasize that both the results of Bertotti \etal and our method are useful to estimate the reversal field.
For example, numerical simulation of the LLG equation \cite{okamoto10}
showed good agreement with the theory of Bertotti \etal \cite{bertotti01,bertotti01a,bertotti09,bertotti09book}.
In Ref. \cite{okamoto10},
the magnitude of the dc magnetic field is linearly increased with time until it reaches a certain value.
In this case, the magnetization first relaxes to a stable point of the potential,
and after that the magnetization reverses its direction
when the saturated value of the dc magnetic field is larger than
the reversal field estimated by the method of Bertotti \etal
On the other hand, our approach is applicable when
the magnitude of the dc magnetic field is fixed from $t=0$, as mentioned above.
To clarify the applicability of the theory of Bertotti \etal more precisely,
an estimation of relaxation time from the initial state to the steady point,
which should be shorter than the time to saturate the dc magnetic field magnitude, will be necessary.
\section{Conclusion}
\label{sec:Conclusion}
In conclusion,
we studied the dependence of the reversal field
in microwave-assisted magnetization reversal
on the frequency of the rotating field theoretically.
The microwave produced a dc magnetic field pointing in the reversed direction,
which energetically stabilized the reversed state.
The microwave simultaneously produced a torque
proportional to the frequency of the rotating field.
Because this torque prevented the reversal,
a large field was required to reverse the magnetization
in the high-frequency region.
The equations determining the reversal fields
in both the low- and high-frequency regions were derived
from the energy balance equation.
The formulas showed that
the reversal field in the low-frequency region became
converged to Eq. (\ref{eq:condition_low_zero_damping})
as the damping constant decreased,
while the reversal field in the high-frequency region
was independent of the damping constant.
The boundary between the low- and high-frequency regions,
which was independent of the damping constant, was
also estimated from the energy balance equation.
The comparison with the numerical solution of the Landau-Lifshitz-Gilbert equation
showed quantitatively good agreement,
guaranteeing the validities of the formula.
The author would like to acknowledge
H. Imamura, T. Yorozu, H. Kubota, H. Maehara, and S. Yuasa
for their valuable discussions.
This work was supported by JSPS KAKENHI Grant-in-Aid for Young Scientists (B) 25790044.
|
\section{Introduction}
A \textbf{continued fraction expansion} $[a_0;a_1,\dots,a_k,\dots]$ is an expression of the form
\[
a_0 + \frac{1}{a_1 + \frac{1}{a_2+ \frac{1}{a_3+ \frac{1}{\ddots}}}}
\]
For a real number $a$, we say $[a_0;a_1,\dots,a_k,\dots]$ is the continued fraction expansion of $a$ if $a=[a_0;a_1,\dots,a_k,\cdots]$ and $a_0\in \Z$, $a_i\in \N_{>0}$ for $i>0$.
Let $a$ be a real number with continued fraction expansion $[a_0;a_1,\dots,a_k,\dots]$. In this note we study a numeration system due to Ostrowski \cite{Ost} based on the continued fraction expansion of $a$. Set $q_{-1} := 0$ and $q_{0} := 1$, and for $k\geq 0$,
\begin{equation}\label{equation:intro1}
q_{k+1} := a_{k+1} \cdot q_k + q_{k-1}.
\end{equation}
Then every natural number $N$ can be written uniquely as
\[
N = \sum_{k=0}^{n} b_{k+1} q_{k},
\]
where $b_{k} \in \N$ such that $b_1<a_1$, $b_k \leq a_{k}$ and, if $b_k = a_{k}$, $b_{k-1} = 0$. We say the word $b_n\dots b_1$ is the \textbf{Ostrowski representation} of $N$ based on $a$, and we write $\rho_a(N)$ for this word. For more details on Ostrowski representations, see for example Allouche and Shallit \cite[p.106]{Automatic} or Rockett and Sz\"usz \cite[Chapter II.4]{RS}. When $a$ is the golden ratio $\phi:=\frac{1+\sqrt{5}}{2}$, the continued fraction expansion of $a$ is $[1;1,\dots]$. In this special case the sequence $(q_k)_{k\in \N}$ is the sequence of Fibonacci numbers. Thus the Ostrowski representation based on the golden ratio is precisely the better known \textbf{Zeckendorf representation} \cite{Zeckendorf}.\newline
In this paper, we will study the following question: given the continued fraction expansion of $a$ and the Ostrowski representation of two natural numbers based on $a$, is there an easy way to compute the Ostrowski representation of their sum? Ahlbach, Usatine, Frougny and Pippenger \cite{EffectiveZeckendorf} give an elegant algorithm to calculate the sum of two natural numbers in Zeckendorf representations. In this paper we generalize their work and present an elementary three pass algorithm for computing the sum of two natural numbers given in Ostrowski representation. To be precise, we show that given the continued fraction expansion of $a$, addition of two $n$-digit numbers in Ostrowski representation based on $a$ can be computed by three linear passes over the input sequence and hence in time $O(n)$. If $a$ is a quadratic number\footnote{A real number $a$ is \textbf{quadratic} if it is a solution to a quadratic equation with rational coefficients}, we establish that the graph of addition in the Ostrowski numeration system based on $a$ can be recognized by a finite automaton (see Theorem B for a precise statement). When $a$ is the golden ratio, this result is due to Frougny \cite{Frougny}\footnote{In private communication Frougny proved that whenever the continued fraction expansion of a has period 1, the stronger statement that addition in the Ostrowski numeration system associated with $a$ can be obtained by three linear passes, one left-to-right, one right-to-left and one left-to-right, where each of the passes defines a finite sequential transducer.}. \newline
Ostrowski representations arose in number theory and have strong connections to the combinatorics of words (see for example Berth\'e \cite{Berthe}). However, our main motivation for studying Ostrowski representations is their application to decidability and definability questions in mathematical logic. The results in this paper (in particular Theorem B below) play a crucial role in the work of the first author \cite{H-Twosubgroups} on expansions of the real additive group. Here we will present the following application of our work on addition in the Ostrowski numeration system to the study of expansions of Presburger Arithmetic (see Theorem A). \newline
Let $a$ be quadratic. Since the continued fraction expansion of $a$ is periodic, there is a natural number $c:= \max_{k\in \N} a_k$. Let $\Sigma_a= \{0,\dots,c\}$. So $\rho_a(N)$ is a $\Sigma_a$-word. Let $V_a : \N \to \N$ be the function that maps $x \geq 1$ with Ostrowski representation $b_n\dots b_1$ to the least $q_k$ with $b_{k+1} \neq 0$, and $0$ to $1$.
\begin{thmA} Let $a$ be quadratic. A set $X\subseteq \N^n$ is definable in $(\N,+,V_a)$ if and only if $X$ is $a$-recognizable. Hence the theory of $(\N,+,V_a)$ is decidable.
\end{thmA}
We say a set $X \subseteq \N$ is \textbf{$a$-recognizable} if $0^*\rho_a(X)$ is recognizable by a finite automaton, where $0^*\rho_a(X)$ is the set of all $\Sigma_a$-words of the form $0\dots 0 \rho_a(N)$ for some $N\in X$.
The definition of $a$-recognizability for subsets of $\N^n$ is slightly more technical and we postpone it to Section 3. The decidability of the theory of $(\N,+,V_a)$ follows immediately from the first part of the statement of Theorem A and Kleene's theorem (see Khoussainov and Nerode \cite[Theorem 2.7.2]{automata}) that the emptiness problem for finite automata is decidable. Bruy\`{e}re and Hansel \cite[Theorem 16]{Bruyere} establish Theorem A when $a$ is the golden ratio. In fact, they show that Theorem A holds for linear numeration systems whose characteristic polynomial is the minimal polynomial of a Pisot number. A similar result for numeration systems based on $(p^n)_{n\in \N}$, where $p>1$ is an integer, is due to B\"uchi \cite{B60} (for a full proof see Bruy{\`e}re, Hansel, Michaux and and Villemaire \cite{BVHM}). It is known by Shallit \cite{Shallit} and Loraud \cite[Theorem 7]{Loraud} that the set $\N$ is $a$-recognizable if and only if $a$ is quadratic. So in general the conclusion of Theorem A fails when $a$ is not quadratic. \newline
A few remarks about the proof of Theorem A are in order. The proof that every definable set is $a$-recognizable, is rather straightforward, and we follow a similar argument from Villemaire \cite{Villemaire}. For the other direction, by Hodgson \cite{Hodgson} it is enough to prove that $\N$, the graph of $V_a$ and the graph of $+$ are $a$-recognizable. While it is easy to check the $a$-recognizability of the graph of $V_a$, we have to use our algorithm for addition in Ostrowski numeration systems to show that the graph of $+$ is $a$-recognizable. Thus most of the work towards proving Theorem A goes into showing the following result.
\begin{thmB} Let $a$ is a quadratic. Then $\{ (x,y,z) \in \N^3 \ : \ x+y = z \}$ is $a$-recognizable.
\end{thmB}
We end this introduction with a brief comment about possible applications of Theorem B to the theory of Sturmian words\footnote{When preparing this paper, the authors were completely unaware of the connection between Sturmian words and Ostrowski representations. We would like to thank the anonymous referee to point out this connection.}. Let $a$ be a real number in $[0,1]$. We define
\[
f_{a}(n):=\lfloor (n+1)a \rfloor - \lfloor na \rfloor,
\]
and we denote the infinite $\{0,1\}$-word $f_a(1)f_a(2)\dots$ by $\boldsymbol{f}_{a}$. This word is called the \textbf{Sturmian characteristic word} with slope $a$. If $a$ is a quadratic irrational, the set $\{n \in \N \ : f_a(n) =1\}$ is $a$-recognizable (see \cite[Theorem 9.1.15]{Automatic}). Du, Mousavi, Schaeffer and Shallit \cite{Duetal} use this connection and Theorem B in the case of the golden ratio $\phi$ to prove results about the Fibonacci word (that is the Sturmian characteristic word with slope $\phi -1$). Because of Theorem B the techniques in \cite{Duetal} can be applied to any characteristic Sturmian word whose slope is a quadratic irrational.
\subsection*{Notation} We denote the set of natural numbers by $\{0,1,2,\dots\}$ by $\N$. Definable will always mean definable without parameters. If $\Sigma$ is a finite set, we denote the set of $\Sigma$-words by $\Sigma^*$. If $a \in \Sigma$ and $X\subseteq \Sigma^*$, we denote the set $\{ a\dots aw \ : \ w \in X\}$ of $\Sigma$-words by $a^*X$. If $x \in X^m$ for some set $X$, we write $x_i$ for the $i$-th coordinate of $x$.
\section{Ostrowski addition}
Fix a real number $a$ with continued fraction expansion $[a_0;a_1,\dots,a_k,\dots]$. In this section we present an algorithm to compute the Ostrowski representations based on $a$ of the sum of two natural numbers given in Ostrowski representation based on $a$. Since we will only consider Ostrowski representation based on $a$, we will omit the reference to $a$. In the special case that $a$ is the golden ratio, our algorithm is exactly the one presented in \cite{EffectiveZeckendorf}. Although it is not strictly necessary, the reader might find it useful to read \cite[Section 2]{EffectiveZeckendorf} first.\\
\noindent Let $M,N \in \N$ and let $x_n\dots x_1, y_n\dots y_1$ be the Ostrowski representations of $M$ and $N$. We will describe an algorithm that given the continued fraction expansion of $a$ calculates the Ostrowski representation of $M+N$. Let $s$ be the word $s_{n+1}s_n\dots s_1$ given by
\[
s_i := x_i + y_i,
\]
for $i=1,\dots,n$ and $s_{n+1}:=0$. For ease of notation, we set $m:=n+1$.\\
\noindent The algorithm consists of three linear passes over $s$: one left-to-right, one right-to-left and one left-to-right. These three passes will change the word $s$ into a word that is the Ostrowski representation of $M+N$. The first pass converts $s$ into a word whose digit at position $k$ is smaller or equal to $a_k$. The idea how to achieve this, is as follows. We will argue (see Lemma \ref{lem:smallfirst}) that whenever the digit at position $k$ is larger or equal to $a_k$, then the preceding digit has to be less than $a_{k+1}$. Using \eqref{eq:1} we can then decrease the digit at position $k$ by $a_k$, without increasing the one at position $k+1$ above $a_{k+1}$, and without changing the value the word represents. The resulting word might not yet be an Ostrowski representation of $M+N$, because the digit at position $k$ may be $a_k$ and not followed by $0$. With the second and third pass we eliminate all such occurrences.\\
\noindent The first step is an algorithm that makes a left-to-right pass over the sequence $s_m\dots s_1$ starting at $m$. That means that it starts with the most significant digit, in this case $s_m$, and works its way down to the least significant digit $s_1$. The algorithm can best be described in terms of a moving window of width four. At each step, we only consider the entries in this window. After any possible changes are performed, the window moves one position to the right. When the window reaches the last four digits, the changes are carried out as usual. Afterwards, one final operation is performed on the last three digits. The precise algorithm is as follows. Given $s=s_m\dots s_1$, we will recursively define for every $k \in \N$ with $3 \leq k \leq m+1$, a word
\[
z_k:= z_{k,m}z_{k,m-1}\dots z_{k,2}z_{k,1}.
\]
\begin{algo} Let $k=m+1$. Then set
\[
z_{m+1} := s_m\dots s_1.
\]
\noindent Let $k\in \N$ with $4 \leq k < m+1$. We now define $z_k= z_{k,m}z_{k,m-1}\dots z_{k,2}z_{k_1}$:
\begin{itemize}
\item for $i\notin \{k,k-1,k-2,k-3\}$, we set $z_{k,i} = z_{k+1,i}$,
\item the subword $z_{k,k}z_{k,k-1}z_{k,k-2}z_{k,k-3}$ is determined as follows: \\
\begin{enumerate}
\item[(A1)] if $z_{k+1,k}<a_k, z_{k+1,k-1} > a_{k-1}$ and $z_{k+1,k-2}=0$,
\[
z_{k,k}z_{k,k-1}z_{k,k-2}z_{k,k-3}=(z_{k+1,k}+1)(z_{k+1,k-1}-(a_{k-1}+1))(a_{k-2}-1)(z_{k+1,k-3}+1)
\]
\item[(A2)] if $z_{k+1,k}<a_k, a_{k-1}\leq z_{k+1,k-1} \leq 2a_{k-1}$ and $z_{k+1,k-2}>0$,
\[
z_{k,k}z_{k,k-1}z_{k,k-2}z_{k,k-3}=(z_{k+1,k}+1)(z_{k+1,k-1}-a_{k-1})(z_{k+1,k-2}-1)(z_{k+1,k-3})
\]
\item[(A3)] otherwise,
\[
z_{k,k}z_{k,k-1}z_{k,k-2}z_{k,k-3}=z_{k+1,k}z_{k+1,k-1}z_{k+1,k-2}z_{k+1,k-3}.
\]
\end{enumerate}
\end{itemize}
\vspace{0.5cm}
Let $k=3$. We now define $z_3=z_{3,m}\dots z_{3,1}$:
\begin{itemize}
\item for $i\notin\{1,2,3\}$, we set $z_{3,l} = z_{4,l}$,
\item the subword $z_{3,3}z_{3,2}z_{3,1}$ is determined as follows: \\
\begin{enumerate}
\item[(B1)] if $z_{4,3} < a_3$, $z_{4,2} > a_2$ and $z_{4,1}=0$,
\[
z_{3,3}z_{3,2}z_{3,1}=(z_{4,3}+1)(z_{4,2}-(a_2+1))(a_1-1),
\]
\item[(B2)] if $z_{4,3} < a_3$, $z_{4,2} \geq a_2$ and $a_1 \geq z_{4,1}>0$,
\[
z_{3,3}z_{3,2}z_{3,1}=(z_{4,3}+1)(z_{4,2}-a_2)(z_{4,1}-1),
\]
\item[(B3)] if $z_{4,3} < a_3$, $z_{4,2} \geq a_2$ and $z_{4,1}>a_1$,
\[
z_{3,3}z_{3,2}z_{3,1}=(z_{4,3}+1)(z_{4,2}-a_2+1)(z_{4,1}-a_1-1),
\]
\item[(B4)] if $z_{4,2} < a_2$ and $z_{4,1}\geq a_1$,
\[
z_{3,3}z_{3,2}z_{3,1}= z_{4,3}(z_{4,2}+1)(z_{4,1}-a_1).
\]
\item[(B5)] otherwise,
\[
z_{3,3}z_{3,2}z_{3,1}=z_{4,3}z_{4,2}z_{4,1}.
\]
\end{enumerate}
\end{itemize}
\end{algo}
\noindent When we speak of \textbf{the entry at position $l$ after step $k$}, we mean $z_{k,l}$. When $z_{k+1,l} \neq z_{k,l}$, we say that at step $k$ the entry in position $l$ was changed. It follows immediately from the algorithm that the only entries changed at step $k$, are in position $k,k-1,k-2$ or $k-3$.\\
\noindent The goal of Algorithm 1 is to produce a word whose entry at position $k$ is smaller or equal to $a_k$, and which represents the same value as $s$. The following two Propositions make this statement precise.
\begin{prop} Algorithm 1 leaves the value represented unchanged. That is, for every $k \in \N$ with $3\leq k \leq m+1$
\[
\sum_{i=0}^{m} z_{k,i+1}q_i = \sum_{i=0}^{m} s_{i+1}q_i.
\]
\end{prop}
\begin{proof} It follows immediately from the recursive definition of the $q_i$'s (see \eqref{equation:intro1}) that each rule of Algorithm 1 leaves the value represented unchanged. Induction on $k$ gives the statement of the Proposition.
\end{proof}
\begin{prop}\label{prop:step1} For $k>1$, $z_{3,k} \leq a_k$ and $z_{3,1} \leq a_1-1$.
\end{prop}
\noindent We will prove the following two lemmas first.
\begin{lem}\label{lem:aftertwo} Let $k\in \N$ and $k\geq 3$. Then
\begin{itemize}
\item[(i)] If $z_{k+1,k-1} = 2a_{k-1}+1$, then $z_{k+1,k-2}=0$.
\item[(ii)] If $z_{k+1,k-1} = 2a_{k-1}$, then $z_{k+1,k-2}\leq a_{k-2}$.
\end{itemize}
\end{lem}
\begin{proof} For (i), let $z_{k+1,k-1} = 2a_{k-1}+1$. It follows immediately from the rules of the algorithm that $z_{k+2,k-1}=2a_{k-1}+1$ and $z_{m+1,k-1}=2a_{k-1}$. So $x_{k-1}$ and $y_{k-1}$ are both equal to $a_{k-1}$. Hence $x_{k-2}=0, y_{k-2}=0$ and $z_{m+1,k-2}=0$. The first time that the entry in position $k-2$ can be changed, is at step $k+1$, when rule (A1) is applied. However, since $z_{k+2,k-1} = 2a_{k-1}+1$, rule (A1) was not applied at step $k+1$. Thus $z_{k+1,k-2}=z_{m+1,k-2}=0$.\\
\noindent For (ii), let $z_{k+1,k-1} = 2a_{k-1}$. If $x_{k-1}=y_{k-1}=a_{k-1}$, we argue as before to get $z_{k+1,k-2}=0$. Suppose that either $x_{k-1}\neq a_{k-1}$ or $y_{k-1} \neq a_{k-1}$. Because $z_{k+1,k-1} = 2a_{k-1}$, we get that $x_{k-1} + y_{k-1}=2a_{k-1}-1$, and that the entry in position $k-1$ had to be increased by $1$ at step $k+2$. Hence either $x_{k-1} = a_{k-1}$ or $y_{k-1}=a_{k-1}$. By the definition of Ostrowski representations, $x_{k-2} + y_{k-2} \leq a_{k-2}$. Thus $z_{k+2,k-2} \leq a_{k-2}$. Since the entry in position $k-1$ was increased by $1$ at step $k+2$, $z_{k+2,k}=a_k-1$. Thus no change is made at step $k+1$. It follows that $z_{k+1,k-2} = x_{k-2} + y_{k-2} \leq a_{k-2}$.
\end{proof}
\begin{lem}\label{lem:smallfirst} Let $k \in \N$ and $3 \leq k\leq m$.
\begin{itemize}
\item[(i$)_k$] If $z_{k+1,k-1}>a_{k-1}$, then $z_{k+1,k} < a_k$.
\item[(ii$)_k$] If $z_{k+1,k-1}=a_{k-1}$ and $z_{k+1,k-2}>0$, then $z_{k+1,k} < a_{k}$.
\end{itemize}
\end{lem}
\begin{proof} We prove the statements by induction on $k$. For $k=m$, both (i$)_{m}$ and (ii$)_{m}$ hold, because $z_{m+1,m} = 0$.
For the induction step, suppose that (i$)_{k+1}$ and (ii$)_{k+1}$ hold. We need to establish (i$)_{k}$ and (ii$)_k$.\\
\noindent We first show (i$)_k$. Suppose $z_{k+1,k-1} > a_{k-1}$. Towards a contradiction, assume that $z_{k+1,k} \geq a_k$. Since $z_{k+1,k-1}>a_{k-1}$ and the algorithm does not increase the entry in position $k-1$ above $a_{k-1}$ at step $k+1$, we have $z_{k+2,k-1}>a_{k-1}$. Because $z_{k+1,k} \geq a_k$ and the algorithm either leaves the entry in position $k$ at step $k+1$ untouched or decreases it by $a_k$ or $a_k+1$, we get that either $z_{k+2,k}=z_{k+1,k}$ or $z_{k+2,k} \in \{2a_k,2a_k+1\}$. We handle these cases separately.\\
\noindent Suppose $z_{k+2,k} \in \{2a_k,2a_k+1\}$. By (i$)_{k+1}$, $z_{k+2,k+1} < a_{k+1}$. It follows from Lemma \ref{lem:aftertwo} that, if $z_{k+2,k}=2a_k$, then $z_{k+2,k-1} \leq a_{k-1}$, and if $z_{k+2,k}=2a_k+1$, then $z_{k+2,k-1}=0$. Since one of the first two rules is applied at step $k+1$, we have that $z_{k+1,k-1} < a_{k-1}$. This contradicts our assumption that $z_{k+1,k-1} > a_{k-1}$.\\
\noindent Now, we suppose that $z_{k+2,k}=z_{k+1,k}$ and $z_{k+2,k}=a_k$. Because $z_{k+2,k-1} > a_{k-1}$, we get $z_{k+2,k+1} < a_{k+1}$ by (ii$)_{k+1}$. Hence $z_{k+1,k} = z_{k+2,k} - a_k$ by rule (A2). This contradicts $z_{k+1,k} = z_{k+2,k}$. \\
\noindent Finally, assume that $z_{k+2,k}=z_{k+1,k}$ and $z_{k+2,k}>a_k$. By (i$)_{k+1}$, $z_{k+2,k+1} < a_{k+1}$. Since $z_{k+2,k-1} > a_{k-1}$, we have $z_{k+2,k+1}< 2a_{k+1}$ by Lemma \ref{lem:aftertwo}. Applying rule (A2) gives $z_{k+1,k}=z_{k+2,k}-a_{k}$. As before, this is a contradiction.\\
\noindent We now prove (ii$)_k$. Let $z_{k+1,k-1}=a_{k-1}$ and $z_{k+1,k-2}>0$. Suppose towards a contradiction that $z_{k+1,k} \geq a_{k}$. Then $z_{k+2,k} \geq a_{k}$, because the algorithm never increases the entry at position $k$ at step $k+1$. Since $z_{k+1,k-1}=a_{k-1}$, either $z_{k+2,k-1} = a_{k-1}+1$ (in this case rule (A2) was applied) or $z_{k+2,k-1} =a_{k-1}$ (in this case rule (A3) was applied). In both cases, $z_{k+2,k+1}< a_{k+1}$ by (i$)_{k+1}$ and (ii$)_{k+1}$. Since $z_{k+2,k-1} > 0$, $z_{k+2,k}\leq 2a_k$ by Lemma \ref{lem:aftertwo}(i). Hence rule (A2) was applied at step $k+1$, and $z_{k+2,k-1}=a_{k-1}+1$. By Lemma \ref{lem:aftertwo}(ii), $z_{k+2,k} < 2a_k$.
Thus $z_{k+1,k} = z_{k+2,k} - a_k < a_k$, a contradiction.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:step1}] Suppose $k\geq 3$. Because the entry at position $k$ is not changed after step $k$, it is enough to show that $z_{k,k} \leq a_k$. We have to consider four different cases depending on the value of $z_{k+2,k}$.\\
\noindent First, consider the case that $z_{k+2,k}< a_k$. Since the algorithm does not increase the entry in position $k$ at step $k+1$, $z_{k+1,k}<a_k$. Thus $z_{k,k} \leq z_{k+1,k} + 1 \leq a_k$. \\
\noindent Suppose $z_{k+2,k} = a_k$ and $z_{k+2,k-1} > 0$. By Lemma \ref{lem:smallfirst}(ii), $z_{k+2,k+1} < a_{k+1}$. By rule (A2), $z_{k+1,k} = 0$. Hence $z_{k,k}\leq 1 \leq a_k$.\\
\noindent Suppose $z_{k+2,k} = a_k$ and $z_{k+2,k-1} = 0$. Then no change is made at step $k+1$. Thus $z_{k+1,k}=a_k$ and $z_{k+1,k-1}=0$. Since no change is made at step $k$ as well, $z_{k,k} = a_k$. \\
\noindent Finally, consider $z_{k+2,k}> a_k$. By Lemma \ref{lem:smallfirst}(i), $z_{k+2,k+1} < a_{k+1}$ . Hence either rule (A1) or rule (A2) is applied. We get that $z_{k+1,k} \leq a_k$. If $z_{k+1,k}=a_k$, then $z_{k,k}=a_k$. If $z_{k+1,k}<a_k$, then $z_{k,k} \leq z_{k+1,k} + 1 \leq a_k$.\\
\noindent Now suppose that $k< 3$. We have to show that $z_{3,k} \leq a_k$. We do so by considering several different cases depending on the values of $z_{4,2}$ and $z_{4,1}$. By Lemma \ref{lem:smallfirst}, if $z_{4,2}>a_2$, or, if $z_{4,2}=a_2$ and $z_{4,1}>0$, then $z_{4,3}<a_3$. If $z_{4,2}=a_2$ and $z_{4,1}=0$, then no changes was made.\\
\noindent Suppose that $z_{4,2}=2a_2+1$. By Lemma \ref{lem:aftertwo}, $z_{4,1}=0$ . By rule (B1), $z_{3,2} = a_2$, $z_{3,1} = a_1 -1$ and $z_{3,3} =z_{4,3}+1 \leq a_3$.\\
\noindent Now suppose that $z_{4,2} = 2a_2$. We get $z_{4,1} \leq a_1$ from Lemma \ref{lem:aftertwo}. Then either rule (B1) or rule (B2) was applied. In both cases we get that $z_{3,2}=a_2$, $z_{3,1} = z_{4,1} - 1 \leq a_1 - 1$ and $z_{3,3} =z_{4,3}+1 \leq a_3$.\\
\noindent Consider that $a_2 \leq z_{4,2} < 2a_2$ and $z_{4,1}>0$. Here either rule (B2) or rule (B3) was used. Then $z_{3,2}\leq a_2$, $z_{3,1} \leq a_1-1$ and $z_{3,3} =z_{4,3}+1 \leq a_3$.\\
\noindent The last case we have to consider is $z_{4,2}<a_2$. Depending on whether $z_{4,1}\geq a_1$, we applied either rule (B4) or rule (B5). Since $z_{4,1} \leq 2a_1-1$, we get $z_{3,1} \leq a_1 -1$ and $z_{3,2} \leq z_{3,2} +1 \leq a_2$ in both cases.
\end{proof}
\noindent We will now describe the second step towards determining the Ostrowski representation of $M+N$. This second algorithm will be a right-to-left pass over $z_3$. Given the word $z_{3,m}z_{3,m-1}\dots z_{3,2}z_{3,1}$, we will recursively generate a word
\[
w_k= w_{k,m+1}w_{k,m}\dots w_{k,2}w_{k,1}
\]
for each $k\in N$ with $k \in \N$ with $2\leq k \leq m+1$. At each step only elements in a moving window of length $3$ are changed. Because the algorithm moves right to left, we will start by defining $w_2$, and then recursively define $w_k$ for $k\geq 2$.
\begin{algo}Let $k=2$. Then set
\[
w_{2} := 0z_{3,m}z_{3,m-1}\dots z_{3,2}z_{3,1}.
\]
\noindent Let $k \in \N$ with $2< k \leq m+1$. We now define $w_k=w_{k,m+1}\dots w_{k,1}$:
\begin{itemize}
\item for $i\notin \{k,k-1,k-2\}$, we set $w_{k,i} := w_{k-1,i}$.
\item if $w_{k-1,k} < a_k$, $w_{k-1,k-1} = a_{k-1}$ and $w_{k-1,k-2}>0$, set
\[
w_{k,k}w_{k,k-1}w_{k,k-2} := (w_{k-1,k}+1)0(w_{k-1,k-2}-1),
\]
otherwise
\[
w_{k,k}w_{k,k-1}w_{k,k-2} := w_{k-1,k}w_{k-1,k-1}w_{k-1,k-2}.
\]
\end{itemize}
\end{algo}
\noindent Again it follows immediately from Equation \eqref{equation:intro1} that this algorithm leaves the value represented unchanged:
\[
\sum_{k=0}^{m} w_{m+1,k+1}q_k = \sum_{k=0}^{m} z_{3,k+1}q_k.
\]
By Proposition \ref{prop:step1} and the rules of Algorithm 2, $w_{k,i} \leq a_k$ for every $k=2,\dots,m+1$ and $i=1,\dots, m+2$.
\begin{lem}\label{lem:step2} There is no $k\in \N$ such that
\begin{itemize}
\item $w_{m+1,k} = a_k$
\item $w_{m+1,k-1} < a_{k-1}$,
\item $w_{m+1,k-2} = a_{k-2}$, and
\item $w_{m+1,k-3} > 0$.
\end{itemize}
\end{lem}
\begin{proof} Towards a contradiction, suppose that there is such an $k$. We will first show that $w_{k-2,k-3}>0, w_{k-2,k-2} = a_{k-2}$ and $w_{k-2,k-1} = a_{k-1}$.\\
\noindent Suppose that $w_{k-2,k-3}=0$. Then the algorithm would not have made any changes at step $k-2$. Thus $w_{k-1,k-3}=0$. Because the entry will not be changed later than step $k-1$, $w_{m+1,k-3}=0$. However, this contradicts $w_{m+1,k-3}>0$. Thus $w_{k-2,k-3}>0$.\\
\noindent Suppose that $w_{k-2,k-2} < a_{k-2}$. Then $w_{k-1,k-2} = w_{k-2,k-2}$. This implies that $w_{k,k-2} < a_{k-2}$ and $w_{m+1,k-2} < a_k$. This a contradiction against our assumption $w_{m+1,k-2} = a_{k-2}$. Hence $w_{k-2,k-2} = a_{k-2}$. \\
\noindent Now suppose that $w_{k-2,k-1} < a_{k-1}$. Since $w_{k-2,k-2} = a_{k-2}$ and $w_{k-2,k-3}>0$, $w_{k-1,k-2}=0$. Thus $w_{m+1,k-2} = 0$, contradicting $w_{m+1,k-2} = a_{k-2}$. So $w_{k-2,k-1} = a_{k-1}$.\\
\noindent It follows that $w_{k-1,k-1} = w_{k-2,k-1}=a_{k-1}$ and $w_{k-1,k-2} = w_{k-1,k-2}=a_{k-2}$. We will now argue that $w_{k-1,k}<a_k$.\\
\noindent Suppose towards a contradiction that $w_{k-1,k} = a_k$. Then $w_{k,k} = a_k$ and $w_{k,k-1} = a_{k-1}$. Since $w_{m+1,k-1}<a_{k-1}$, we have $w_{k,k+1} < a_{k+1}$. Thus $w_{k+1,k} = 0$. Hence $w_{m+1,k} = 0$, a contradiction. So $w_{k-1,k} < a_k$.\\
\noindent We conclude that the entry at position $k-2$ is changed at step $k$. Therefore, $w_{k,k-2} = w_{k-1,k-2} -1= a_{k-2}-1$. So $w_{m+1,k-2} = a_{k-2}-1$. This contradicts our original assumption $w_{m+1,k-2}=a_{k-2}$.
\end{proof}
\noindent The third and final step of our algorithm is a left-to-right pass over $w_{m+1}$. The moving window is again of length $3$ and we use the same rule as in step 2. Given the word $w_{m+1,m+1}\dots w_{m+1,1}$, we will recursively generate a word
\[
v_k:=v_{k,m+2}\dots v_{k,1}
\]
for each $k\in N$ with $k \in \N$ with $3\leq k \leq m+3$. Because the algorithm moves left to right, we will start by defining $w_{m+3}$ and then recursively define $w_k$ for $k\leq m+3$.
\begin{algo} Let $k=m+3$. Then set
\[
v_{m+3} := 0w_{m+1,m+1}\dots w_{m+1,1}.
\]
Let $k \in \N$ with $3\leq k \leq m+2$. We now define $v_k=v_{k,m+2}\dots v_{k,1}$:
\begin{itemize}
\item for $i\notin \{k,k-1,k-2\}$, we set $v_{k,i} := v_{k+1,i}$,
\item if $v_{k+1,k} < a_k$, $v_{k+1,k-1} = a_{k-1}$ and $v_{k+1,k-2}>0$, set
\[
v_{k,k}v_{k,k-1}v_{k,k-2} := (v_{k+1,k}+1)0(v_{k+1,k-2}-1),
\]
otherwise
\[
v_{k,k}v_{k,k-1}v_{k,k-2} := v_{k+1,k}v_{k+1,k-1}v_{k+1,k-2}.
\]
\end{itemize}
\end{algo}
\noindent As before Equation \eqref{equation:intro1} implies that this algorithm leaves the value represented unchanged:
\[
\sum_{k=0}^{m} w_{m+1,k+1}q_k = \sum_{k=0}^{m} v_{3,k+1}q_k.
\]
Moveover, we have $v_{k,i} \leq a_k$ for every $k=3,...,m+3$ and $i=1,\dots, m+2$. We will now show $v_3$ is indeed the Ostrowski representation of $M+N$. It is enough to prove the following Proposition.
\begin{prop}\label{prop:step3} Let $l \geq 3$. Then there is no $k\geq l-1$ such that $v_{l,k} = a_k$ and $v_{l,k-1} > 0$.
\end{prop}
\noindent Before we give the proof of Proposition \ref{prop:step3}, we need one more Lemma.
\begin{lem}\label{lem:step3} Let $l\in \{3,\dots,m+3\}$. Then there is no $k\in \N$ such that
\begin{itemize}
\item $v_{l,k} = a_k$
\item $v_{l,k-1} < a_{k-1}$,
\item $v_{l,k-2} = a_{k-2}$, and
\item $v_{l,k-3} > 0$.
\end{itemize}
\end{lem}
\begin{proof} We prove the Lemma by induction on $l$. By Lemma \ref{lem:step2}, there is no such $k$ for $m+3$. Suppose that the statement holds for $l+1$. We want to show the statement for $l$. Towards a contradiction, suppose that there is a $k$ such that
\begin{equation}\label{eq:1}
v_{l,k} = a_k, v_{l,k-1} < a_{k-1}, v_{l,k-2} = a_{k-2} \hbox{ and } v_{l,k-3}>0.
\end{equation}
By the induction hypothesis, it is enough to check that no change was made at step $l$; that is $v_{l,i} = v_{l+1,i}$ for $i\in \{k,...,k-3\}$. Since the algorithm only modifies the entries at position $l,l+1$ or $l+2$, we can assume that $k \in \{l-2,\dots,l+3\}$. We consider each case separately. \\
\noindent First, suppose $k=l-2$. We get that $v_{l,i}=v_{l+1,i}$ for $i\in \{k-1,k-2,k-3\}$, because they are not in the moving window at step $l$. The only possible change is at position $k$. Since $v_{l,l-2} < v_{l+1,l-2}$ by induction hypothesis, and $v_{l,l-2}=a_{l-2}$, we get $v_{l,k}=v_{l+1,k}$. So no change is made. \\
\noindent Suppose that $k=l-1$. If a change is made at step $l$, then $v_{l,k}=0$. But this contradicts \eqref{eq:1}. Hence no change is made in this case.\\
\noindent Suppose that $k=l$. If a change is made at step $l$, then $v_{l,k-2}= v_{l+1,k-2} - 1 < a_{k-2}$. As before, this contradicts \eqref{eq:1}. Thus no change is made. \\
\noindent Suppose $k=l+1$. If a change is made at step $l$, then $v_{l,k-2}=0$ contradicting \eqref{eq:1}. So no change is made in this case either.\\
\noindent Suppose $k=l+2$. If a change is made at step $l$, then $v_{l,k-3}=0$. This again contradicts \eqref{eq:1}, and hence no change is made.\\
\noindent Finally suppose $k=l+3$. By induction hypothesis, $v_{l+1,k-3}=0$. Since $v_{l,k-3}>0$, we have $v_{l+1,k-4} =a_{k-4}$ and $v_{l+1,k-5}>0$. Then
\[
v_{l+1,k-2} = a_{k-2}, v_{l+1,k-3} = 0, v_{l+1,k-4} =a_{k-4} \hbox{ and } v_{l+1,k-5}>0.
\]
This contradicts the induction hypothesis.
\end{proof}
\begin{proof}[Proof of Propositon \ref{prop:step3}] We prove this statement by induction on $l$. For $l=m+3$ the statement holds trivially, because $v_{m+3,m+2}=0$. Now suppose that the statement holds for $l+1$, but fails for $l$. Hence there is $k \geq l-1$ such that $v_{l,k} = a_k$ and $v_{l,k-1} > 0$. Since $v_{l+1,i}=v_{l,i}$ for $i>l$, we have $k\leq l+1$. We now consider the three remaining cases $k=l+1$, $k=l$ and $k=l-1$ individually.\\
\noindent If $k=l+1$, then $v_{l+1,k}=a_{l+1,k}$. By the induction hypothesis, $v_{l+1,k-1}=0$. But in order for $v_{l,k-1}>0$ to hold, we must have $v_{l+1,k-2}=a_{k-2}$ and $v_{l+1,k-3}>0$. This contradicts Lemma \ref{lem:step3}.\\
\noindent If $k=l$, then either $v_{l+1,k}=a_k$ or $v_{l+1,k} = a_k-1$. Suppose that $v_{l+1,k} = a_{k}-1$. Then $v_{l+1,k-1} = a_k$ and $v_{l+1,k-2}>0$. This implies $v_{l,k-1}=0$, which contradicts $v_{l,k-1}>0$. Suppose that $v_{l+1,k} = a_k$. By induction hypothesis, $v_{l+1,k-1}=0$. But then no change is made at step $l$, and hence $v_{l,k-1}=0$. A contradiction against $v_{l,k-1}>0$.\\
\noindent If $k=l-1$, then no change is made at step $l$, since $v_{l,l-1}=a_{l-1}$. Hence $v_{l+1,l-1}=v_{l,l-1}=a_{l-1}$ and $v_{l+1,l-2} = v_{l,l-2} > 0$. Since no change was made at step $l$, we get that $v_{l+1,l}=a_{l}$. This contradicts the induction hypothesis.
\end{proof}
\begin{cor}\label{cor:ostadd} The word $v_{3,m+2}\dots v_{3,1}$ is the Ostrowski representation of $M+N$.
\end{cor}
\section{Proof of Theorem A}
In this section we will prove Theorem A. Let $a$ be a quadratic irrational number. Let $[a_0;a_1,\dots,a_n,\dots]$ be its continued fraction expansion. Since the continued fraction expansion of $a$ is periodic, it is of the form
\[
[a_0;a_1,\dots,a_{\xi-1},\overline{a_{\xi},\dots, a_{\nu}}],
\]
where $\nu-\xi$ is the length of the repeating block and the repeating block starts at $\xi$. We can choose $\xi$ and $\nu$ such that $\xi>4$ and $\nu - \xi\geq 3$.\footnote{It might be the case that neither $\xi$ nor $\nu$ are minimal, but this will be irrelevant here.} Set $\mu := \max_i a_i$. Set $m:=2\mu + 1$. Set $\Sigma_a:=\{0,\dots,m\}.$ \newline
\noindent We first remind the reader of the definitions of finite automata and recognizability. For more details, we refer the reader to \cite{automata}. Let $\Sigma$ be a finite set. We denote by $\Sigma^*$ the set of words of finite length on $\Sigma$.
\begin{defn} A \textbf{nondeterministic finite automaton} $\Cal A$ over $\Sigma$ is a quadruple $(S,I,T,F)$, where $S$ is a finite non-empty set, called the set of states of $\Cal A$, $I$ is a subset of $S$, called the set of initial states, $T\subseteq S \times \Sigma \times S$ is a non-empty set, called the transition table of $\Cal A$ and $F$ is a subset of $S$, called the set of final states of $\Cal A$. An automaton $\Cal A=(S,I,T,F)$ is \textbf{deterministic} if $I$ contains exactly one element, and for every $s\in S$ and $w \in \Sigma^*$ there is exactly one $s' \in S$ such that $(s,w,s')\in T$.
We say that an automaton $\Cal A$ on $\Sigma$ \textbf{accepts} a word $w=w_n\dots w_1 \in \Sigma^*$ if there is a sequence $s_n,\dots, s_1,s_0 \in S$ such that $s_n \in I$, $s_0 \in F$ and for $i=1,\dots,n$, $(s_i,w_i,s_{i-1})\in T$. A subset $L\subseteq \Sigma^*$ is \textbf{recognized} by $\Cal A$ if $L$ is the set of $\Sigma$-words that are accepted by $\Cal A$. We say that $L\subseteq \Sigma^*$ is \textbf{recognizable} if $L$ is recognized by some deterministic finite automaton.
\end{defn}
\noindent It is well known (see \cite[Theorem 2.3.3]{automata}) that a set is recognizable if it is recognized by some \emph{nondeterministic} finite automaton.\newline
\noindent Let $\Sigma$ be a set containing $0$. Let $z=(z_1,\dots,z_n) \in (\Sigma^*)^n$ and let $m$ be the maximal length of $z_1,\dots,z_n$. We add to each $z_i$ the necessary number of $0$'s to get a word $z_i'$ of length $m$. The \textbf{convolution}\footnote{Here we followed the presentation in \cite{Villemaire}. For a general definition of convolution see \cite{automata}.} of $z$ is defined as the word $z_1 * \dots * z_n \in (\Sigma^n)^*$ whose $i$-th letter is the element of $\Sigma^n$ consisting of the $i$-th letters of $z_1', \dots, z_n'$.
\begin{defn} A subset $X\subset (\Sigma^*)^n$ is \textbf{$\Sigma$-recognizable} if the set
\[
\{z_1*\dots *z_n \ : \ (z_1,\dots,z_n) \in X\}
\]
is $\Sigma^n$-recognizable.
\end{defn}
\noindent We remind the reader that every natural number $N$ can be written as $N = \sum_{k=0}^{n} b_{k+1} q_{k}$, where $b_{k} \in \N$ such that $b_1<a_1$, $b_k \leq a_{k}$ and, if $b_k = a_{k}$, $b_{k-1} = 0$, and that we denoted the $\Sigma_a$-word $b_n\dots b_1$ by $\rho_a(N)$.
\begin{defn} Let $X\subseteq \N^n$. We say that $X$ is \textbf{$a$-recognizable} if the set
\[
\{ (0^{l_1}\rho_a(N_1),\dots,0^{l_n}\rho_a(N_n)) \ : \ (N_1,\dots,N_n) \in X, l_1,\dots,l_n \in \N\}
\]
is $\Sigma_a$-recognizable.
\end{defn}
\noindent In this section we will prove that a subset $X\subseteq \N^n$ is $a$-recognizable if and only if $X$ is definable in $(\N,+,V_a)$.
\subsection*{Recognizability implies definability} We will first show that whenever a set $X \subseteq \N^n$ is $a$-recognizable, then $X$ is definable in $(\N,+,V_a)$. The proof here is an adjusted version of the proofs in Villemaire \cite{Villemaire} and \cite{Bruyere}.\newline
\noindent First note that $<$ is definable in $(\N,+,V_a)$ and so is $V_a(\N) = \{ q_k \ : \ k \in \N\}$. For convenience, we write $I$ for $V_a(\N)$. We denote the successor function on $I$ by $s_I$.
\begin{defn}For $j\in \{1,\dots, m\}$, let $\epsilon_{j} \subseteq I \times \N$ be the set of $(x,y) \in I \times \N$ with
\begin{align*}
\exists z &\in \N \exists t\in \N (z<x \wedge z+jx< s_I(x) \wedge V_a(t) > x \wedge V_a(x+t) = x \wedge y = z+jx+t) \\
&\vee \exists z \in \N (z<x \wedge y < s_I(x)\wedge y = z+jx).
\end{align*}
Let $\epsilon_{0} \subseteq I \times \N$ be the set of $(x,y) \in I \times \N$ with $\bigwedge_{j=1}^m \neg \epsilon_j(x,y)$.
\end{defn}
\noindent This definition is inspired by \cite[Lemma 2.3]{Villemaire}. Obviously, $\epsilon_j$ is definable in $(\N,+,V_a)$. Because of the greediness of the Ostrowski representation, $\epsilon_j(x,y)$ holds iff $x=q_k$ for some $k\in \N$ and the coefficient
of $q_k$ in the Ostrowski representation of $y$ is $j$. We directly get the following Lemma.
\begin{lem}\label{lem:rl1} Let $l,n \in \N$ and let $\sum_{k} b_{k+1} q_k$ be the Ostrowski representation of $n$. Then
$b_{l+1} = j$ iff $\epsilon_j(q_l,n)$.
\end{lem}
\begin{defn} Let $I_e$ be the set of all $y \in I$ with
\[
\exists z \in \N \ \epsilon_1(1,z) \wedge \epsilon_1(y,z) \wedge \forall x \in I \big(\epsilon_1(x,z) \leftrightarrow \neg \epsilon_1(s_I(x),z)\big),
\]
and let $I_o$ be the set of all $y \in I$ with
\[
\exists z \in \N \ (\neg \epsilon_1(1,z)) \wedge \epsilon_1(y,z) \wedge \forall x \in I \big(\epsilon_1(x,z) \leftrightarrow \neg \epsilon_1(s_I(x),z)\big).
\]
\end{defn}
\noindent Obviously both $I_e$ and $I_o$ are definable in $(\N,+,V_a)$, $I=I_e \cup I_o$, and since $q_0=1$,
\[
I_e = \{ q_k \ : \ k \hbox{ even } \} \hbox{ and } I_o = \{ q_k \ : \ k \hbox{ odd } \}.
\]
\begin{defn} Let $U_{e} \subseteq \N$ be the set of all $y \in \N$ with
\[
\forall z \in I_o \ \epsilon_0(z,y) \wedge \forall z \in I_e \ (\epsilon_0(z,y) \vee \epsilon_1(z,y)),
\]
and $U_{o} \subseteq \N$ be the set of all $y \in \N$ with
\[
\forall z \in I_e \ \epsilon_0(z,y) \wedge \forall z \in I_o \ (\epsilon_0(z,y) \vee \epsilon_1(z,y)).
\]
\end{defn}
\noindent Again it is easy to see that $U_e$ and $U_o$ are definable in $(\N,+,V_a)$. We get the following Lemma from Lemma \ref{lem:rl1}.
\begin{lem}\label{lem:rlost} Let $n \in \N$ and let $\sum_{k} b_{k+1} q_k$ be the Ostrowski representation of $n$. Then
\begin{itemize}
\item [(i)] $n \in U_e$ if and only if for all even $k$ $b_{k+1} \leq 1$, and for all odd $k$ $b_{k+1}=0$,
\item [(ii)] $n \in U_o$ if and only if for all odd $k$ $b_{k+1} \leq 1$, and for all even $k$ $b_{k+1}=0$.
\end{itemize}
\end{lem}
\begin{defn} Let $\epsilon \subseteq I \times (U_e \times U_o)$ be the set of all $(x,(y_1,y_2))$ with
\[
(x \in I_e \rightarrow \epsilon_1(x,y_1)) \wedge (x \in I_o \rightarrow \epsilon_1(x,y_2)).
\]
\end{defn}
\begin{thm}\label{thm:recidef} Let $X \subseteq \N^n$ be $a$-recognizable. Then $X$ is definable in $(\N,+,V_a)$.
\end{thm}
\begin{proof} Let $X \subseteq \N^n$ be $a$-recognizable by a finite automaton $\Cal A = (S,I,T,F)$. Without loss generality we can assume that the set of states $S$ is $\{1,\dots,t\}$ for some $t\in \N$, and $I=\{1\}$.
Let $\varphi$ be the formula defining the following subset $Z$ of $U^t$:
\[
\{ (u_1,\dots,u_t) \in U^t \ : \forall q \in I \ \bigwedge_{i=1}^{t} \big(\epsilon(q,u_i) \rightarrow \bigwedge_{j=1, j\neq i}^t \neg \epsilon(q,u_j)\big)\}.
\]
So $Z$ is the set of tuples $(u_1,\dots,u_t) \in U^t$ such that for $q \in I$ there is at most one $i\in \{1,\dots,t\}$ such that $\epsilon(q,u_i)$. Note that $x \in X$ if there is a run $s_{1}\dots s_{m}$ of $\Cal A$ on the word given by the Ostrowski representation of the coordinates of $x$ such that $s_{1} = 1$ and $s_{m} \in F$. The idea now is to code such a run as an element of $Z$. To be precise, a tuple $(u_1,\dots,u_t) \in Z$ will code a run $s_{1}\dots s_{m}$ if for each $q_i \in I$, $s_i$ is the unique element $k$ of $\{1,\dots,t\}$ such that $\epsilon(q_i,u_{k})$. Thus $x=(x_1,\dots,x_n) \in X$ if and only if $x$ satisfies the following formula in $(\N,+,V_a)$:
\begin{align*}
\exists u_1,\dots,u_t \in U \ \exists q \in I \ \varphi(u_1,\dots,u_t) \wedge \epsilon(1,u_1) \wedge \bigvee_{l\in F} \epsilon(q,u_l)\\
\wedge \bigwedge_{(l,(\rho_1,\dots,\rho_n),k)\in T} \forall z \in I \Big((z > q) \rightarrow \bigwedge_{i=1}^n \bigwedge_{j=1}^{m} \neg \epsilon_{j}(z,x_i)\Big)\\
\wedge \Big[\big(z \leq q \wedge \epsilon(z,u_l) \wedge \bigwedge_{i=1}^n \epsilon_{\rho_i}(z,x_i)\big)\rightarrow \epsilon(s_I(z),u_k)\Big].
\end{align*}
\end{proof}
\subsection*{Definability implies recognizability} We will prove that if a subset $X\subseteq \N^n$ is definable in $(\N,+,V_a)$, then it is $a$-recognizable. By \cite{Hodgson} it is suffices to show that the set $\N$ and the relations $\{(x,y) \in \N^2 \ : \ x=y\}$, $\{(x,y,z) \in \N^3 \ : \ x+y = z\}$ and $\{(x,y) \in \N^2 \ : \ V_a(x)=y\}$ are all $a$-recognizable. It is well known that $\N$ is $a$-recognizable (see for example \cite[Theorem 8]{Shallit}), and using that knowledge it is easy to check that $\{(x,y) \in \N^2 \ : \ x=y\}$ and $\{(x,y) \in \N^2 \ : \ V_a(x)=y\}$ are $a$-recognizable. We are now going to show that $\{(x,y,z) \in \N^3 \ : \ x+y = z\}$ is $a$-recognizable.\newline
\noindent By the work in the previous section, we have an algorithm to compute addition in Ostrowski representation based on $a$. This algorithm consists of four steps, and we will now show that each of the four steps can be recognized by a finite automaton. Given two words $z=z_n\dots z_1, z'=z_n'\dots z_1' \in \rho_a(\N)$, the first step is to compute the $\Sigma_a$-word $(z_n+z_n')\dots (z_1+z_1')$, which we will denote by $z+z'$. It is straightforward to verify that the set $\{ z * z' * (z+z') \ : \ z,z' \in \rho_a(\N) \}$ is recognizable by a finite automaton. For $z,z' \in \Sigma_a^*$, we will write $z \rightsquigarrow_i z'$ if Algorithm $i$ produces $z'$ on input $z$. In the following, we will prove that the set $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z\rightsquigarrow_i z'\}$ is recognizable by a finite automaton for $i=1,2,3$. From these results it is immediate that
\begin{align*}
\{ z * z' * z'' * u_0 * u_1 * u_2 \ &: \ z,z',z'' \in \rho_a(\N), u_0,u_1,u_2 \in \Sigma_a^*,\\
& \ u_0 = z+z', u_0\rightsquigarrow_1 u_1 \rightsquigarrow_2 u_2 \rightsquigarrow_3 z'' \}
\end{align*}
is recognizable by a finite automaton. Since recognizability is preserved under projections (see \cite[Theorem 2.3.9]{automata}), $\{(x,y,z) \in \N^3 \ : \ x+y = z\}$ is $a$-recognizable by Corollary \ref{cor:ostadd}. Thus every set $X \subseteq \N^n$ definable in $(\N,+,V_a)$ is $a$-recognizable.
\subsection*{An automaton for Algorithm 1} We will now construct a non-deterministic automaton $\Cal A_1$ that recognizes the set $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_1 z'\}$. Before giving the definition of $\Cal A_1$, we need to introduce some notation. Let $A \subseteq \N_{\leq m}^4\times \N_{\leq m}^4 \times \N_{\leq m}^4$ be the set of tuples $(u,v,w)$ with
\[
w = \left\{
\begin{array}{ll}
(v_1+1,v_2-(u_2+1),u_3-1,v_4+1), & \hbox{ if $v_1<u_1, v_2 > u_{2}$ and $v_3=0$,} \\
(v_1+1,v_2-u_2,v_3-1,v_4, & \hbox{ if $v_{1}<u_1, u_2\leq v_2 \leq 2u_2$ and $v_3>0$,} \\
(v_1,v_2,v_3,v_4), & \hbox{otherwise.}
\end{array}
\right.
\]
Let $B \subseteq \N_{\leq m}^3\times \N_{\leq m}^3 \times \N_{\leq m}^3$ be the set of tuples $(u,v,w)$ with
\[
w= \left\{
\begin{array}{ll}
(v_1+1,v_2-(u_2+1),u_3-1), & \hbox{$v_1 < u_1$, $v_2 > u_2$ and $v_3=0$;} \\
(v_1+1,v_2-u_2,v_3-1), & \hbox{$v_1 < u_1$, $v_2 \geq u_2$ and $u_1 \geq v_1>0$,;} \\
(v_1+1,v_2-u_2+1,v_1-u_1-1), & \hbox{$v_{1} < u_1$, $v_{2} \geq u_2$ and $v_1>u_1$;} \\
(v_1,v_2+1,v_1-u_1), & \hbox{if $v_2 < u_2$ and $v_1\geq u_1$;} \\
(v_1,v_2,v_3), & \hbox{otherwise.}
\end{array}
\right.
\]
Note that $A$ corresponds to the rules (A1),(A2) and (A3) of Algorithm 1, while $B$ corresponds to the rules (B1)-(B5) of Algorithm 1. The values of the variable $u$ represent the relevant part of the continued fraction, the values of the variable $v$ are used to code the entries in the moving window before any changes are carried out, and the values of the variable $w$ correspond to the entries in the moving window after the changes are carried out. For $i \in \{4,\dots,\nu\}$ and $l\in \{0,1\}$,
\[
P(i,l) := \left\{
\begin{array}{ll}
(a_i,a_{i-1},a_{i-2},a_{\nu}), & \hbox{$i=\xi+2$ and $l=1$;} \\
(a_i,a_{i-1},a_{\nu},a_{\nu-1}), & \hbox{$i=\xi+1$ and $l=1$;} \\
(a_i,a_{\nu},a_{\nu-1},a_{\nu-2}), & \hbox{$i=\xi$ and $l=1$;} \\
(a_i,a_{i-1},a_{i-2},a_{i-3}), & \hbox{otherwise.} \\
\end{array}
\right.
\]
\noindent We first explain informally the construction of $\Cal A_1$. Suppose we take $z=z_l\dots z_1\in \Sigma_a^*$. Now perform Algorithm 1 on $z$, and let the word $z'=z_l'\dots z_1'$ be the output. In order to carry out the operations at step $k$ in Algorithm 1, we needed to know the values of $a_k,a_{k-1},a_{k-2},a_{k-3}$. Because of the periodicity of the continued fraction expansion of $a$, there is $i\leq \nu$ such $a_k=a_i$. Let $l$ be $1$ if $k> \nu$ and $0$ otherwise. Then $P(i,l) = (a_k,a_{k-1},a_{k-2},a_{k-3})$. Hence in order to reconstruct $(a_k,a_{k-1},a_{k-2},a_{k-3}),$ it is enough to save $i$ and whether or not $k\leq \nu$. Moreover, to perform the operations at step $k$ in Algorithm 1, we also used the values of the last three entries in the moving window after the changes in the previous step are carried out, but before the window moves to the right. Let us denote the triple consisting of these entries by $v=(v_1,v_2,v_3) \in \Sigma_a^3$. So before the operations at step $k$ are performed, the values in the moving window are $(v_1,v_2,v_3,z_{k-3})$. Note that at step $k$ in the algorithm, we are reading in $z_{k-3}$, and not $z_k$. However, the value of $z_k'$ is determined at the same step. Indeed, at step $k$ with $k\geq 4$, the entries in the moving window are changed as follows:
\[
(v_1,v_2,v_3,z_{k-3})\mapsto (z_k',v_1',v_2',v_3'),
\]
for a certain triple $(v_1',v_2',v_3') \in \Sigma_a^3$ with $A(P(i,l),v_1,v_2,v_3,z_{k-3},z_k',v_1',v_2',v_3')$. The values in the moving window for step $k-1$ will be $(v'_1,v_2',v_3',z_{k-4})$. Because the value of $z_k'$ is only determined at step $k$, and thus at the same time $z_{k-3}'$ is being read, we are required to store the value of $z_k'$ for three steps. In order to save this information when moving from state to state, we introduce another triple $(w_1,w_2,w_3) \in \Sigma_a^3$. This triple will always contain the last three digits of $z'$. That means that before step $k$, $(w_1,w_2,w_3)=(z'_{k},z'_{k-1},z'_{k-2})$. We now define the set of states of $\Cal A_1$ as the set of quadruples $(i,l,v,w)$, where $i\leq \nu$, $l \in \{0,1\}$, $v,w\in \Sigma_a^3$. The idea is that in each state of the automaton the pair $(i,l)$ codes the relevant part of the continued fraction expansion, $v$ contains the entries of the moving window, and $w\in \Sigma_a^3$ the values of $z_k'$ that we needed to save. The automaton moves from one of these states to another according to the rules described in Algorithm 1.\newline
\noindent Here is the definition of the automaton $\Cal A_1 =(S_1,I_1,T_1,F_1)$.
\begin{itemize}
\item [1.] The set $S_1$ of states of $\Cal A_1$ is
\begin{align*}
\{ (i,1,v,w) \ : \ \xi \leq i & \leq \nu, v,w \in \Sigma_a^3\}\\
&\cup \{ (i,0,v,w) \ : \ 3 \leq i\leq \nu, v,w \in \Sigma_a^3\},
\end{align*}
\item[2.] the set $I_1$ of initial states is
\[
\{ (i,l,(0,0,0),(0,0,0)) \in S \ : \ i \geq 4 \},
\]
\item[3.] the transition table $T_1$ contains the tuples $(s,(x,y),t)\in S_1 \times \Sigma_a^2 \times S_1$ that satisfy $w'=(w_2,w_3,y)$ and
one of the following conditions:
\begin{itemize}
\item[a.] $i\neq \xi, (j,l')=(i-1,l), A(P(i,l),v,x,w_1,v'),$
\item[b.] $i= \xi, l=1, (j,l')=(\nu,l), A(P(i,l),v,x,w_1,v')$,
\item[c.] $i= \xi, l=0, (j,l')=(i-1,l), A(P(i,l),v,x,w_1,v')$
\item[d.] $i=4, j=3$, $A(P(4,l),v,x,w_1,v'), B(a_3,a_2,a_1,v',w_2,w_3,y)$,
\end{itemize}
where $s=(i,l,v,w)$, $w=(w_1,w_2,w_3)$ and $t=(j,k,v',w')$,
\item[4.] the set $F_1$ of final states is $\{ (i,l,w,y) \in S_1 \ : \ i=3\}$.
\end{itemize}
\noindent We leave it to the reader to check the details that $\Cal A$ indeed recognizes the set $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_1 z'\}$. The automata we constructed is non-deterministic, but as mentioned above there is deterministic finite automaton that recognizes the same set.
\subsection*{Automata for Algorithm 2 and 3} We now describe the non-deterministic automata $\Cal A_2$ and $\Cal A_3$ recognizing the sets $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_2 z'\}$ and $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_3 z'\}$. Again, we have to fix some notation first. Let $C\subseteq \N_{\leq m}^3 \times \N_{\leq m}^3 \times \N_{\leq m}^3$ be the set of triples $(u,v,w)\in C$ such that
\[
w =\left\{
\begin{array}{ll}
(v_1+1,0,v_3-1), & \hbox{if $v_1 < u_1$, $v_2 = u_2$ and $v_3>0$;} \\
(v_1,v_2,v_3), & \hbox{otherwise.}
\end{array}
\right.
\]
The relation $C$ represents the operation performed in both Algorithm 2 and 3. As for $A$ and $B$ above, the values of the variable $u$ correspond to the relevant part of the continued fraction, while the values of the variables $v$ and $w$ represent the entries in the moving window, before and after any changes are carried out. For $i \in \{3,\dots,\nu\}$ and $l\in \{0,1\}$,
\[
Q(i,l) := \left\{
\begin{array}{ll}
(a_i,a_{i-1},a_{\nu}), & \hbox{$i=\xi+1$ and $l=1$;} \\
(a_i,a_{\nu},a_{\nu-1}), & \hbox{$i=\xi$ and $l=1$;} \\
(a_i,a_{i-1},a_{i-2}), & \hbox{otherwise.} \\
\end{array}
\right.
\]
\noindent We start with an informal description of the automaton $\Cal A_2$. Let $z=z_l\dots z_1\in \Sigma_a^*$ and suppose that $z'=z_l'\dots z_1'$ is the output of Algorithm 2 on input $z$. To perform the operations at step $k$ in Algorithm 2, we again need to know a certain part of the continued fraction expansion of $a$; in this case $(a_k,a_{k-1},a_{k-2})$. As before it is enough to know the natural numbers $i\leq \nu$ with $a_k=a_i$, and whether $k <\nu$. Set $l$ to be $1$ if $k> \nu$ and $0$ otherwise.
Then $Q(i,l) = (a_k,a_{k-1},a_{k-2})$. When constructing $\Cal A_2$, we have to be careful: the Algorithm 2 runs from the right to the left, but the automaton reads the input from the left to the right. Let $(v_1',v_2') \in \Sigma_a^2$ be such that $(z_k,v_1',v_2')$ are the entries in the moving window before the changes at step $k$ are made. Then at step $k$, the entries change as follows:
\[
(z_k,v_1',v_2')\mapsto (v_1,v_2,z_{k-2}'),
\]
for some pair $(v_1,v_2) \in \Sigma_a^2$ with $C(Q(i,l),z_k,v_1',v_2',v_1,v_2,z_{k-2}')$. So when the automaton reads in $(z_{k-2},z_{k-2}')$, the value of $z_k$ is used to determine $z_{k-2}'$. Hence in contrast to $\Cal A_1$, the automaton $\Cal A_2$ has to remember the value of $z_k$, and not the value of $z_k'$. We define the states of $\Cal A_2$ to be tuples $(i,l,v,w)\in \{0,\dots,m\}\times \{0,1\} \times \Sigma_a^2 \times \Sigma_a^2$. The pair $v$ is again used to save the entries of the moving window, and $w$ is needed to remember the previously read entries of $z$. The automaton moves from one of these states to another according to the rules described in Algorithm 2. However, since the automaton reads the input backwards, the automaton will go from a state $(i,l,v,w)$ to a state $(i',l',v',w')$ if $Q(i,l)$ and $Q(i',l')$ are the correct parts of the continued fraction expansion of $a$ and the algorithm transforms $(z_k,v_1',v_2')$ to $(v_1,v_2,z_{k-2}')$.
\newline
\noindent Here is the definition of the automaton $\Cal A_2 =(S_2,I_2,T_2,F_2)$.
\begin{itemize}
\item [1.] The set $S_2$ of states of $\Cal A_2$ is
\begin{align*}
\{ (i,1,v,w) \ : \ \xi \leq i & \leq \nu, v,w \in \Sigma_a^2\}\\
&\cup \{ (i,0,v,w) \ : \ 2 \leq i\leq \xi, v,w \in \Sigma_a^2\},
\end{align*}
\item[2.] the set $I_2$ of initial states is
\[
\{ (i,l,(0,0,0),(0,0,0)) \in S \ : \ i \geq 3 \},
\]
\item[3.] the transition table $T_2$ contains the tuples $(s,(x,y),t)\in S_2 \times \Sigma_a^2 \times S_2$ that satisfy $w'=(w_2,x)$ and
one of the following conditions:
\begin{itemize}
\item[a.] $i\neq \xi, (j,l')=(i-1,l), C(Q(i,l),w_1,v',v,y),$
\item[b.] $i= \xi, l=1, (j,l')=(\nu,l), C(Q(i,l),w_1,v',v,y)$,
\item[c.] $i= \xi, l=0, (j,l')=(i-1,l), C(Q(i,l),w_1,v',v,y)$
\item[d.] $i=3, j=2$, $C(Q(i,0),w,x,v,y)$,
\end{itemize}
where $s=(i,l,v,w)$, $w=(w_1,w_2)$ and $t=(j,k,v',w')$,
\item[4.] the set $F_2$ of final states is $\{ (i,l,w,y) \in S_2 \ : \ i=3\}$.
\end{itemize}
\noindent As in the case of Algorithm 1, we leave it to the reader to verify that $\Cal A_2$ recognizes the set $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_2 z'\}$. As before, while $\Cal A_2$ is non-deterministic, there is a deterministic automata recognizing the same set as $\Cal A_2$.\\
\noindent It is left to construct the automaton for Algorithm 3. The only difference between Algorithm 2 and 3 is the direction in which the algorithm runs over the input. Hence the only adjustment we need to make to $\Cal A_2$, is to address the change in direction. Let $\Cal A_3 =(S_2,I_2,T_3,F_2)$ be the automaton that has the same states as $\Cal A_2$, but whose transition table $T_3$ contains the tuples $(s,(x,y),t)\in S_2 \times \Sigma_a^2 \times S_2$ that satisfy $w'=(w_2,y)$ and
one of the following conditions:
\begin{itemize}
\item[a.] $i\neq \xi, (j,l')=(i-1,l), C(Q(i,l),v,x,w_1,v'),$
\item[b.] $i= \xi, l=1, (j,l')=(\nu,l), C(Q(i,l),v,x,w_1,v')$,
\item[c.] $i= \xi, l=0, (j,l')=(i-1,l), C(Q(i,l),v,x,w_1,v')$
\item[d.] $i=3, j=2$, $C(Q(i,0),v,x,w,y)$,
\end{itemize}
where $s=(i,l,v,w)$, $w=(w_1,w_2)$ and $t=(j,k,v',w')$.\\
\noindent The set $\{ z * z' \ : \ z,z' \in \Sigma_a^*, z \rightsquigarrow_3 z'\}$ is recognized by $\Cal A_3$. So there is also a deterministic automaton recognizes this set. This completes the proof of Theorem A.
\bibliographystyle{plain}
|
\section{Introduction}
Dark Energy is the dominant source causing the current accelerated expansion of the universe, as has been confirmed by observations~\cite{Riess:1998cb,Schmidt:1998ys,Bennett:2012zja,Ade:2013zuv}. Although there exist some possibilities explaining Dark Energy, a tiny positive cosmological constant would be the prime candidate, in perfect agreement with recent observations~\cite{Bennett:2012zja,Ade:2013zuv}.
If one wants to understand the purely theoretically origin of this cosmological constant, we should promote Einstein gravity to be consistent with its quantum formulation. String theory is quite motivated for this purpose as it is expected to provide the quantum nature of gravity as well as particle physics. A cosmological constant could be realized in the context of flux compactifications~\cite{Giddings:2001yu,Dasgupta:1999ss} of 10D string theories, where a vacuum expectation value of the moduli potential at minima contributes to the vacuum energy in a four-dimensional space-time universe. Since there exist many possible choices of quantized fluxes and also a number of types of compactifications, the resultant moduli potential including a variety of minima forms the string theory landscape (see reviews~\cite{Douglas:2006es,Grana:2005jc,Blumenhagen:2006ci,Silverstein:2013wua,Quevedo:2014xia,Baumann:2014nda}).
Although there exist many vacua in the string theory landscape, when we naively stabilize the moduli and obtain the minima, negative cosmological constants seem likely to come by. Hence an 'uplift' mechanism from the negative vacuum energy keeping stability should be important to realize an accelerated expanding universe. Some possible ways of the uplift mechanism have been proposed in string compactifications.
\begin{itemize}
\item Explicit SUSY breaking achieved by brane anti-brane pairs contributes positively in the potential, and thus can be used for the uplift~\cite{Kachru:2002gs,Kachru:2003aw,Kachru:2003sx}. When the D3 brane anti-brane pairs are localized at the tip of a warped throat, the potential energy may be controllable due to a warping factor. As the uplift term contributes to the potential at ${\cal O}({\cal V}^{-4/3})$, which appears larger than the F-term potential for stabilization which is in general $\mathcal{O}(\ll \mathcal{V}^{-2})$, de Sitter ($dS$) vacua with tiny positive cosmological constant may be achieved as a result of tuning of warping. A caveat of this proposal is that the SUSY breaking term needs to compensate the entire Anti-de Sitter ($AdS$) energy, so it is an open question if the SUSY breaking term, originally treated in a probe approximation, can be included as a backreaction in supergravity appropriately.
\item As an alternative uplift mechanism, one may use the complex structure sector~\cite{Saltman:2004sn}. In the type IIB setup, the complex structure moduli as well as the dilaton are often stabilized at a supersymmetric point. Owing to the no-scale structure, the potential for the complex structure sector is positive definite $V_{\rm c.s.} = e^{K} |DW|_{\rm c.s.}^2 \sim {\cal O} (\mathcal{V}^{-2})$. So when we stay at the SUSY loci, the potential is given convex downward in general and thus tractable. However, if one stabilizes the complex structure sector at non-supersymmetric points, then there appears a chance to have a positive contribution in the potential without tachyons, that may be applied for the uplift with a tuning. See also recent applications of this mechanism~\cite{Danielsson:2013rza,Blaback:2013qza,Kallosh:2014oja}.
\item When we include the leading order $\alpha'$-correction coming of ${\cal O}(\alpha'^3)$ in the K\"ahler potential~\cite{Becker:2002nn} which breaks the no-scale structure, this generates a positive contribution in the effective potential if the Euler number $\chi$ of the Calabi-Yau is given by a negative value~\cite{Balasubramanian:2004uy}. This positive term can balance the non-perturbative terms in the superpotential such that stable $dS$ vacua can be achieved in this K\"ahler Uplift model~\cite{Westphal:2006tn,Rummel:2011cd,deAlwis:2011dp} (see also~\cite{Westphal:2005yz}). In the simplest version of the K\"ahler uplifting scenario, there is an upper bound on the overall volume of the Calabi-Yau such that one may worry about higher order $\alpha'$-corrections. However, this bound can be significantly relaxed when embedded in a racetrack model~\cite{Sumitomo:2013vla}.
\item It has been proposed that the negative curvature of the internal manifold may be used for $dS$ constructions as it contributes positive in the scalar potential~\cite{Silverstein:2007ac}. Motivated by this setup, there were many attempts constructing $dS$ vacua~\cite{Haque:2008jz,Flauger:2008ad,Danielsson:2009ff,Caviezel:2009tu,deCarlos:2009fq,deCarlos:2009qm,Dong:2010pm,Andriot:2010ju,Danielsson:2010bc,Danielsson:2011au,Danielsson:2012et,Danielsson:2012by,Blaback:2013ht}. Using the necessary constraint for the extrema~\cite{Hertzberg:2007wc,Wrase:2010ew} and for the stability~\cite{Shiu:2011zt}, we see that the existence of minima requires not only negative curvature, but also the presence of orientifold planes.
\item When the stabilization mechanism does not respect SUSY, D-terms can provide a positive contribution to the potential if the corresponding D7-brane is magnetically charged under an anomalous $U(1)$~\cite{Burgess:2003ic,Cremades:2007ig,Krippendorf:2009zza}. The potential of D-terms arises of order ${\cal O} ( {\cal V}^{-n})$ with $n\geq 2$, depending on the cycle that the D7-brane wraps. If we take into account the stabilization of matter fields having a non-trivial VEV, originating from fluxed D7-branes wrapping the large four-cycle, then the uplift contribution becomes ${\cal O} ({\cal V}^{-8/3})$~\cite{Cremades:2007ig}. So a relatively mild suppression, for instance by warping, is required for this volume dependence. See also recent applications to explicit scenarios in~\cite{Cicoli:2012vw,Cicoli:2013mpa} and also together with a string-loop correction in fibred Calabi-Yaus~\cite{Cicoli:2011yh}.
\item Recently, it has been proposed that a dilaton-dependent non-perturbative term can also work for the uplift mechanism toward $dS$ vacua~\cite{Cicoli:2012fh}. The non-perturbative term depends on both the dilaton and a vanishing blow-up mode which is stabilized by a D-term. Since the D-term turns out to be trivial at the minima due to a vanishing cycle, the non-trivial dilaton as well as the vanishing cycle dependence generate the uplift term within the F-term potential. In this setup, the given uplift term is proportional to $e^{-2b \langle s \rangle}/{\cal V}$. Although the volume does not suppress the uplift term so much, we may expect an exponential fine-tuning by the dilaton dependence to balance with the moduli stabilizing F-term potential.
\end{itemize}
In this paper, we introduce an uplifting term of the form $e^{-a_s \tau_s}/{\cal V}^2$, where $\tau_s$ is the volume of a small 4-cycle, which naturally balances with the stabilizing F-term potential in the Large Volume Scenario (LVS). The following ingredients are necessary for this mechanism:
\begin{itemize}
\item one non-perturbative effect on a 4-cycle $D_2$ to realize the standard LVS moduli stabilization potential,
\item another non-perturbative effect on a different cycle $D_3$,
\item a D-term constraint that enforces the volumes $\tau$ of the two 4-cycles to be proportional $\tau_s \sim \tau_2 \propto \tau_3$ via a vanishing D-term potential.
\end{itemize}
Hence, the minimal number of K\"ahler moduli for this uplifting scenario is $h^{1,1}_+=3$. At the level of the F-term potential the effective scalar potential reduces to the standard LVS moduli stabilization potential plus the mentioned uplifting term yielding metastable $dS$ vacua. The K\"ahler moduli are stabilized at large values avoiding dangerous string- and $\alpha'$-corrections. Compared to~\cite{Cicoli:2012fh}, the dilaton can take rather arbitrary values determined by fluxes as there is no tuning required to keep the uplifting term suppressed. Note that for $h^{1,1}_+=2$, a racetrack setup with two non-perturbative effects on one small cycle does not allow stable $dS$ vacua in the LVS. Hence, we have to consider at least two small cycles and a relating D-term constraint to construct $dS$ vacua.
We also have to consider a necessary condition for coexistence of the vanishing D-term constraint with the non-perturbative terms in superpotential. If the rigid divisors for the two non-perturbative effects intersect with the divisor on which the D-term constraint is generated via magnetic flux, we have to worry whether the VEV of matter fields that are generated by this magnetic flux are given accordingly such that the coefficients of non-perturbative terms remain non-zero. On the other hand, we may avoid additional zero mode contributions in a setup with minimal intersections. A general constraint is that the non-perturbative effects and D-term potential have to fulfill all known consistency condition, for instance requiring rigid divisors, avoiding Freed-Witten anomalies~\cite{Minasian:1997mm,Freed:1999vc} and saturating D3, D5, and D7 tadpole constraints. We expect these constraints to become less severe as the number of K\"ahler moduli increases beyond $h^{1,1}_+=3$ as in principle the degrees of freedom such as flux choices and rigid divisors increases.
This paper is organized as follows. We illustrate the uplift proposal generated through the multi-K\"ahler moduli dependence in the F-term potential and the required general geometric configuration in Section~\ref{genmech_sec} and give some computational details in Appendix~\ref{app_fluxconstraint}. We further discuss the applicability of the uplift mechanism in more general Swiss-Cheese type Calabi-Yau manifolds in Section~\ref{sec:real-more-moduli}.
\section{D-term generated racetrack uplift - general mechanism} \label{genmech_sec}
We illustrate the uplift mechanism by a D-term generated racetrack in Calabi-Yaus with the following properties: there are two small 4-cycles and two linear combinations of these small cycles that are rigid such that the existence of two non-perturbative terms is guaranteed in the superpotential avoiding additional fermionic zero modes from cycle deformations or Wilson lines. We show that this setup in general allows to stabilize the moduli in a $dS$ vacuum at large volume.
\subsection{Geometric setup and superpotential}
We consider an orientifolded Calabi-Yau $X_3$ with $h^{1,1}_+\geq3$ with the following general volume form of the divisors $D_i$
\begin{equation}
\mathcal{V} = \frac16 \left(\sum_{i,j,k=1}^{h^{1,1}_+}\kappa_{i,j,k} t_i t_j t_k \right)\,,\label{gen2cycles}
\end{equation}
in terms of 2-cycle volumes $t_i$ and intersection numbers
\begin{equation}
\kappa_{ijk} = \int_{X_3} D_i \wedge D_j \wedge D_k\,.
\end{equation}
The 4-cycle volumes are given as
\begin{equation}
\tau_i = \frac{\partial \mathcal{V}}{\partial t_i} = \frac12 \kappa_{ijk} t_j t_k\,.\label{tauandt}
\end{equation}
We assume that $X_3$ has a Swiss-Cheese structure with a big cycle named $D_a$ and at least two small cycles $D_b$ and $D_c$, i.e., its volume form can be brought to the form
\begin{equation}
\mathcal{V} = \gamma_a \tau_a^{3/2} - \gamma_b \tau_b^{3/2} - \gamma_c \tau_c^{3/2} - \mathcal{V}_{\text{rest}}\,,
\end{equation}
with $\mathcal{V}_{\text{rest}}$ parametrizing the dependence of the volume on the remaining $h^{1,1}_+-3$ moduli. Now let us assume there are two rigid divisors $D_2$ and $D_3$ of which a linear combination forms the small cycles $D_b$ and $D_c$.
\begin{align}
\begin{aligned}
D_2 &= d_{2b} D_b + d_{2c} D_c\,,\\
D_3 &= d_{3b} D_b + d_{3c} D_c\,.
\end{aligned}\label{tau2taua}
\end{align}
Even if there do not exist two divisors $D_2$ and $D_3$ that are rigid, one might still be able to effectively `rigidify' one or more divisors by fixing all the deformation moduli of the corresponding D7-brane stacks via a gauge flux choice~\cite{Bianchi:2011qh,Cicoli:2012vw,Louis:2012nb}. Under these assumptions, the superpotential in terms of the K\"ahler moduli $T_i = \tau_i + i\, \zeta_i$ is of the form
\begin{equation}
W = W_0 + A_2 e^{-a_2 T_2} + A_3 e^{-a_3 T_3} = W_0 + A_2 e^{-a_2 \left(d_{2b} T_b + d_{2c} T_c \right)} + A_3 e^{-a_3 \left(d_{3b} T_b + d_{3c} T_c \right)}\,,\label{Wnonpert}
\end{equation}
with non-zero $A_2$, $A_3$ and $W_0$ being the Gukov-Vafa-Witten flux superpotential~\cite{Gukov:1999ya}.
\subsection{D7-brane and gauge flux configuration} \label{d7config3moduli_sec}
The orientifold plane O7 induces a negative D7 charge of $-8[O7]$ that has to be compensated by the positive charge of D7-branes. In general the O7 charge can be cancelled by introducing a Whitney brane with charge $8[O7]$~\cite{Collinucci:2008pf}. The non-perturbative effects of~\eqref{Wnonpert} can be either generated by ED3-instantons or gaugino condensation. For the latter, we choose a configuration with $N_2$ D7-branes on $D_2$ and $N_3$ D7-branes on $D_3$. In this case, the exponential coefficients of the non-perturbative terms in~\eqref{Wnonpert} are $a_2 = 2 \pi / N_2$ and $a_3 = 2 \pi / N_3$. The corresponding gauge group is either $SO(N)$ or $Sp(N)$ (which becomes $SU(N)$ if gauge flux is introduced), depending on if the divisor lies on the orientifold plane or not. Furthermore we introduce a third stack of $N_D$ branes on a general linear combination $D_D$ of basis divisors that is not either $D_2$ or $D_3$. This stack will introduce a D-term constraint that reduces the F-term effective scalar potential by one degree of freedom/ K\"ahler modulus. In the case of $h^{1,1}_+=3$ this corresponds to a two K\"ahler moduli LVS potential plus an uplift term that allows $dS$ vacua as we will show in Section~\ref{effectiveVgen_sec}.\footnote{In the case of $D_D$ being a linear combination of only $D_2$ and $D_3$ this divisor is only meaningful if $D_2$ and $D_3$ intersect as a linear combination of non-intersecting and rigid, i.e., local, four-cycles would not make sense. This is the reason we consider non-zero intersections between $D_2$ and $D_3$ in the first place as opposed to the more simple setup $\mathcal{V} \sim \tau_1^{3/2} - \tau_2^{3/2} - \tau_3^{3/2} - \mathcal{V}_{\text{rest}}$.} Note that in general all required D7-brane stacks have to be consistent with possible factorizations of the Whitney brane that cancels the O7 charge~\cite{Collinucci:2008pf,Cicoli:2011qg}.
The D-term constraint is enforced via a Fayet-Illiopoulos (FI) term
\begin{equation}
\xi_D = \frac{1}{\mathcal{V}} \int_{X_3} D_D \wedge J \wedge \mathcal{F}_D = \frac{1}{\mathcal{V}} q_{Dj} t_j\,,\label{FID}
\end{equation}
where $J=t_i D_i$ is the K\"ahler form on $X_3$ and $q_{Dj} = \tilde f^k_D \kappa_{Djk}$ is the anomalous $U(1)$-charge of the K\"ahler modulus $T_j$ induced by the magnetic flux $\mathcal{F}_D = \tilde f^k_D D_k$ on $D_D$. We choose flux-quanta $\tilde f^k_D$ such that $\xi_D = 0$ in~\eqref{FID} implies
\begin{equation}
\tau_c = c\, \tau_b\,,\label{tauctaubprop}
\end{equation}
with a constant $c$ depending on flux quanta and triple intersection numbers. In a concrete example it is important to check that a constant $c$ in~\eqref{tauctaubprop} is realized which is consistent with stabilizing the moduli inside the K\"ahler cone of the manifold.
An important constraint arises from the requirement of two non-vanishing non-perturbative effects $A_2, A_3\neq 0$ on generally intersecting cycles $D_2$ and $D_3$. The cancellation of Freed-Witten anomalies requires the presence of fluxes $\mathcal{F}$ on the D7-branes wrapping these divisors that can potentially forbid the contribution from gaugino condensation in the superpotential. This gauge invariant magnetic flux $\mathcal{F}$ is determined by the gauge flux $F$ on the corresponding D7-brane and pull-back of the bulk $B$-field on the wrapped four-cycle via
\begin{equation}
\mathcal{F} = F - B\,.
\end{equation}
If $D_2$ and $D_3$ intersect each other, the $B$-field can in general not be used to cancel both of theses fluxes to zero. However, it is still possible that both fluxes $\mathcal{F}_2$ and $\mathcal{F}_3$ can be chosen to be effectively trivial, such that no additional zero modes and FI-terms are introduced. These zero-modes would be generated via charged matter fields arising at the intersection of D7-brane stacks or from the bulk D7 spectrum. The constraint has to be checked on a case-by-case basis. We work out a sufficient condition on the intersections $\kappa_{ijk}$ for $\mathcal{F}_2$ and $\mathcal{F}_3$ to be trivial for the case of $D_2$ and $D_3$ not intersecting any other divisors $\kappa_{2,j,k} = \kappa_{3,j,k} =0$ for $j,k\neq 2,3$ in Appendix~\ref{app_fluxconstraint}. Furthermore, it has to be checked that $\mathcal{F}_D$ does not generate any additional zero-modes at the intersections of $D_D$ with $D_2$ and $D_3$.
Finally, the chosen D7-brane and gauge flux setup has to be consistent with D3, D5 and D7 tadpole cancellation. As for every explicit construction this has to be checked on a case-by-case basis for the particular manifold under consideration. We do not expect tadpole cancellation to be in general more restrictive than in e.g., the $AdS$ LVS. In particular, we do not require a large number of D7-branes~\cite{Westphal:2006tn} and/or racetrack effect~\cite{Sumitomo:2013vla} on a particular single divisor to achieve a large volume as in the K\"ahler Uplifting scenario.
\subsection{Effective potential of the K\"ahler moduli} \label{effectiveVgen_sec}
We start with a slightly simplified model where the F-term potential
\begin{equation}
V_F = e^{K} \left( K^{\alpha\bar{\beta}} D_\alpha W \overline{D_\beta W} - 3 |W|^2 \right)\,,\label{VF}
\end{equation}
is given by
\begin{equation}
\begin{split}
K=& - 2 \ln \left({\cal V} + {\xi \over 2}\right), \quad
{\cal V} = (T_a + \bar{T}_a)^{3/2} - (T_b + \bar{T_b})^{3/2} - (T_c + \bar{T_c})^{3/2},\\
W =& W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3(T_b + T_c)},
\end{split}
\label{F-term potential-simplified}
\end{equation}
where we have used equal intersection numbers $\gamma$ and assumed stabilization of the dilaton and complex structure moduli via fluxes~\cite{Giddings:2001yu}. The values of these parameters are not essential for the uplift dynamics we illustrate in this paper. The superpotential in~\eqref{F-term potential-simplified} corresponds to a particular choice of the general linear combination in~\eqref{Wnonpert}. The model~\eqref{F-term potential-simplified} is known to include the solutions of the LVS~\cite{Balasubramanian:2005zx} that stabilizes the moduli in a non-supersymmetric way in the presence of the leading $\alpha'$-correction~\cite{Becker:2002nn} and one non-perturbative term. The $\alpha'$-correction is given by $\xi \propto - \chi g_s^{-3/2}$ where $\chi$ is the Euler number of the Calabi-Yau manifold.\footnote{Recently, it has been argued that the leading correction in both $\alpha'$ and string coupling constants on $SU(3)$ structure manifold comes with the Euler characteristic of the six-dimensional manifold as well as Calabi-Yau compactifications~\cite{Grana:2014vva}.}
The D-term potential is given through the magnetized D7-branes wrapping the Calabi-Yau divisor $D_i$~\cite{Haack:2006cy}:
\begin{equation}
V_D = {1 \over \re (f_D)} \left(\sum_j c_{Dj} \hat{K}_j \varphi_j - \xi_D\right)^2\,,
\label{D-term potential}
\end{equation}
where the gauge kinetic function
\begin{equation}
\text{Re}\,(f_D) = \frac12\, \int_{D_D} J \wedge J - \frac{1}{2g_s}\int_{D_D} \mathcal{F}_D \wedge \mathcal{F}_D \,,\label{fD}
\end{equation}
and $\varphi_j$ are matter fields associated with the diagonal $U(1)$ charges $c_{Dj}$ of a stack of D7-branes and the FI-term $\xi_D$ is defined in~\eqref{FID}.
Now we redefine the coordinates:
\begin{equation}
T_s \equiv {1\over 2 }\left(T_b + T_c \right), \quad Z \equiv {1\over 2} \left(T_b - T_c \right).
\end{equation}
When the D7-branes wrapping the divisor $D_D$ are magnetized and the matter fields are stabilized either at $\langle \varphi_i \rangle = 0$ or satisfying $ \langle \sum c_{ij} \hat{K}_j \varphi_j \rangle =0$, the D-term potential may become
\begin{equation}
V_D \propto {1\over \re(f_D)}{1 \over {\cal V}^2} \left(\sqrt{\tau_b} - \sqrt{\tau_c} \right)^2\,,
\label{D-term potential in simple model}
\end{equation}
using $\xi_D \propto \sqrt{\tau_b} - \sqrt{\tau_c}$ implied by the flux $\mathcal{F}_D$, see~\eqref{tauctaubprop} where we use $c=1$ for simplicity. In the large volume limit, the F-term potential generically scales as ${\cal O} ({\cal V}^{-3})$ in the minima given in the LVS model. Stabilizing the K\"ahler moduli at ${\cal O} ({\cal V}^{-3})$ then requires a vanishing D-term potential, i.e., $\tau_b = \tau_c$ corresponding to $z \equiv \re Z=0$.
Thanks to the topological coupling to the two-cycle supporting magnetic flux, the imaginary mode of the $Z$ modulus is eaten by a massive $U(1)$ gauge boson through the St\"uckelberg mechanism. Since the gauge boson has a mass of order of the string scale ${\cal O} ({\cal V}^{-1/2})$, the degree of freedom of $\im Z$ charged under the anomalous $U(1)$ as well as the gauge boson is integrated out at the high scale. Hence, we are left with the stabilization of the remaining moduli fields by the F-term potential.
\subsection{F-term uplift\label{sec:f-term-uplift}}
Next we will consider the stabilization by the F-term potential given in (\ref{F-term potential-simplified}). We are interested in LVS like minima ${\cal V} \sim e^{\hat{a}_i \tau_i}$ realizing an exponentially large volume. Then the leading potential of order ${\cal V}^{-3}$ is given by
\begin{equation}
\begin{split}
V & \sim {3 W_0^2 \xi \over 4 {\cal V}^3} + {2 W_0 \over {\cal V}^2} \left( a_2 A_2 {\tau_b} e^{-a_2 \tau_b/2} + a_3 A_3 (\tau_b + \tau_c) e^{-a_3 (\tau_b + \tau_c)/2}\right) \\
& + {2 \over 3 {\cal V}} \left(a_2^2 A_2^2 \sqrt{\tau_b} e^{- a_2 \tau_b} + a_3^2 A_3^2 (\sqrt{\tau_b}+\sqrt{\tau_c}) e^{-a_3 (\tau_b + \tau_c)} + 2 a_2 a_3 A_2 A_3 \sqrt{\tau_b} e^{-a_2 \tau_b/2 - a_3 (\tau_b + \tau_c)/2} \right)\,,
\end{split}
\label{effective potential at larger volume 0}
\end{equation}
where the imaginary directions are stabilized at $\im T_i= 0$, and $\im T_a$ is stabilized by non-perturbative effects that are omitted in (\ref{F-term potential-simplified}), inducing a very small mass for $\im T_a$ and with negligible influence on the stabilization of the other moduli. Although the general minima of $\im T_i$ are given by $a_i \im T_i = m_i \pi$ with $m_i \in {\mathbb Z}$, the different solutions just change the sign of $A_i$ and thus we can simply have the potential of the above form.
As the D-term stabilizes $\tau_c = \tau_b$, the resultant potential becomes
\begin{equation}
\begin{split}
V &\sim {3 W_0^2 \xi \over 4 {\cal V}^3} + {4 W_0 \over {\cal V}^2} \left(a_2 A_2 \tau_s e^{-a_2 \tau_s} + 2 a_3 A_3 \tau_s e^{-2 a_3 \tau_s}\right)\\
&+ {2 \sqrt{2} \over 3 {\cal V}} \left({a_2^2 A_2^2 \sqrt{\tau_s}} e^{-2 a_2 \tau_s} + {4 a_3^2 A_3^2 \sqrt{\tau_s}} e^{-4 a_3 \tau_s} + 2 a_2 a_3 A_2 A_3 \sqrt{\tau_s} e^{-(a_2 + 2 a_3)\tau_s} \right)\,,
\end{split}
\label{effective potential at larger volume with ts}
\end{equation}
where we have defined $\tau_s = \re T_s$. One may consider that this form of the potential looks similar to the racetrack type. Although cross terms of $A_2, A_3$ appear due to the $T_b$ dependence, the important point for the uplift mechanism demonstrated in this paper is that the cross terms between $T_b$ dependence of the $A_2$ term and $T_c$ dependence of the $A_3$ term appear at ${\cal O} ({\cal V}^{-4})$.\footnote{Note that this would be more obvious if we start from a toy setup with $W = W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3 T_c}$, although one might not obtain the D-term constraint like $\tau_c = c \tau_b$.} If the cross term appears at ${\cal O}({\cal V}^{-3})$, it disturbs uplifting to $dS$.
We further redefine the fields and parameters such that there are no redundant parameters affecting the stabilization:
\begin{equation}
\begin{split}
&x_s = a_2 \tau_s, \quad {\cal V}_x = {\cal V} a_2^{3/2},\\
&c_i = {A_i \over W_0},\quad \xi_x = \xi a_2^{3/2}, \quad \beta = {2 a_3 \over a_2}.
\end{split}
\label{redefined parameters}
\end{equation}
Then the effective potential at order ${\cal O}({\cal V}^{-3})$ becomes
\begin{equation}
\begin{split}
\hat{V} \equiv \left({a_2^{-3} W_0^{-2}}\right) V \sim {3\xi_x \over 4 {\cal V}_x^3} + {4 c_2 x_s \over {\cal V}_x^2}e^{-x_s} + {2\sqrt{2} c_2^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 x_s } + {4 \beta c_3 x_s \over {\cal V}_x^2} e^{-\beta x_s}.
\end{split}
\label{redefined potential}
\end{equation}
We have neglected the term proportional to $c_3^2$ and the cross term between $c_2, c_3$ in the expression above. In fact, these terms are not important for the uplift mechanism of our interest, and we will justify this assumption a posteriori later. Since the uplift term comes together with $e^{-\beta x_s}$, this term contributes of ${\cal O}({\cal V}_x^{-3})$ when $\beta \sim {\cal O} (1)$. Hence, it contributes at the same order as the stabilizing F-term potential and no suppression factor provided by warping or dilaton dependence is required.
Before performing the uplift, we consider the LVS solution by setting $c_3 =0$. We use a set of parameters:
\begin{equation}
c_2 = -0.01, \quad \xi_x = 5.
\label{test input parameter}
\end{equation}
The extremal equations $\partial_I \hat{V} = 0$ at $c_3=0$ can be simplified as
\begin{equation}
\xi_x = {64 \sqrt{2} (x_s-1) x_s^{5/2} \over (4 x_s -1)^2}, \quad
c_2 = - {6\sqrt{2} (x_s-1) x_s^{1/2} \over (4 x_s -1) {\cal V}_x} e^{-x_s}.
\end{equation}
Solving the equations above, we obtain
\begin{equation}
{\cal V}_x \sim 467, \quad x_s \sim 1.50.\label{VxxsAds}
\end{equation}
We can easily check that this solution gives an $AdS$ vacuum. Note that when we have just two moduli fields ${\cal V}_x, x_s$ in the LVS, the positivity of $\xi_x$ automatically guarantees the stability of the minima since the required condition $x_i > 1$ is satisfied (see e.g.~\cite{Rummel:2013yta}).
Now we consider non-zero $c_3$ for the uplift. As $c_3$ increases, the vacuum energy of the potential minimum increases and eventually crosses the Minkowski point. In Figure~\ref{fig:uplift-illustration}, we illustrate the behavior of the minimum point by changing the value of $c_3$. Interestingly, the volume increases as the vacuum energy increases, suggesting that the effective description of the theory will be more justified toward $dS$ vacua. On the other hand, the minimum value of the Hessian decreases. Destabilization occurs when the uplift term dominates the entire potential. As this happens at higher positive values of the cosmological constant, there certainly exist a range of parameters yielding stable $dS$ vacua within this setup.
As a reference, we show numerical values of parameters close to crossing the Minkowski point. When we use
\begin{equation}
\beta = {5\over 6},
\label{test beta value}
\end{equation}
the minimum reaches Minkowski at
\begin{equation}
c_3 \sim 4.28 \times 10^{-3}, \quad {\cal V}_x \sim 3240, \quad x_s \sim 3.07.
\end{equation}
So we see that the volume increases quite drastically from the $AdS$ vacuum~\eqref{VxxsAds}. Since $c_3$ remains small compared to the input value of $c_2$, we see that our approximation neglecting the term proportional to $c_3^2$ is justified.
\begin{figure}[t]
\includegraphics[width=20.5em]{lambda-c3.pdf}
\vspace{4mm}
\includegraphics[width=20em]{lambda-ddv.pdf}
\includegraphics[width=20.5em]{lambda-vol.pdf}
\caption{\footnotesize Illustration of the D-term generated racetrack uplift mechanism. We plot the cosmological constant $\hat \Lambda$ vs $c_3$, $\text{min}(\partial^2 \hat V)$ and $\mathcal{V}_x$ at the minima of the potential, especially near the Minkowski point.}
\label{fig:uplift-illustration}
\end{figure}
In fact, it is not difficult to see how these values change when the presence of $c_3^2$ terms and cross terms between $c_2,c_3$ in the potential (\ref{effective potential at larger volume with ts}) are taken into account. With the input parameters we used in (\ref{test input parameter}), the Minkowski vacuum is obtained when
\begin{equation}
c_3 \sim 5.11 \times 10^{-3}, \quad {\cal V}_x \sim 2860, \quad x_s \sim 2.61.
\label{test solution with c_3^2}
\end{equation}
Since the obtained values are not significantly different from the case where $c_3^2$ terms and cross terms between $c_2, c_3$ are neglected, we conclude a posteriori that the uplift term is dominated by the term linear in $c_3$.
Let us comment on the stabilization of the axionic partner of each modulus field. As stated, the imaginary mode of $Z$ is eaten up by a massive gauge boson and hence integrated out at the high scale. The axionic partner of the big divisor $T_a$ is stabilized by non-perturbative effects yielding a tiny mass. The remaining modulus $\im T_s$ is stabilized by the F-term potential as the D-term potential does not depend on the latter. In the approximated potential up to ${\cal O}(\mathcal{V}^{-3})$, the Hessian of $ y_s = a_2 \im T_s$ is
\begin{equation}
\partial_{y_s}^2 \hat{V}|_{\rm ext} \sim 5.14 \times 10^{-10}\,,
\end{equation}
where we have included $c_3^2$ and $c_2, c_3$ cross terms, and used the solution (\ref{test solution with c_3^2}) and $\im T_{a} = \im T_i =0$. Thus all K\"ahler moduli are stabilized.
\subsection{Analytical estimate}
It is difficult to analytically derive a generic condition for the D-term generated racetrack uplift since the formulas are still complicated enough even after using several approximations. However, some of the expressions can be simplified under an additional reasonable approximation. In this subsection, we illustrate some analytical analyses for a better understanding.
Since we checked that the uplift mechanism works even at linear approximation of the uplift parameter $c_3$, we only keep terms up to linear order in $c_3$ and neglect the higher order terms including cross terms. The extremal condition $\partial_i \hat{V} = 0$ of the potential (\ref{redefined potential}) is now simplified by
\begin{equation}
\begin{split}
&c_2 \sim -{6\sqrt{2x_s} (x_s-1) \over 4 x_s -1} {e^{x_s} \over {\cal V}_x} + c_3 {\beta (\beta x_s -1) \over x_s -1} e^{(1-\beta)x_s} ,\\
&\xi_x \sim {64\sqrt{2} x_s^{5/2} (x_s-1) \over (4x_s-1)^2} - c_3 {32 \beta x_s^2 \left(2(\beta+2)x_s + \beta-7 \right) \over 9 (x_s-1) (4 x_s-1)} e^{-\beta x_s} {\cal V}_x .
\end{split}
\label{extremal condition of effective potential}
\end{equation}
Although our interest is the uplift toward $dS$ vacua, we have to cross the Minkowski point along the way. Thus, the condition that the minimum structure holds when uplifted to Minkowski vacua is a necessary condition for the $dS$ uplift mechanism. The condition for Minkowski at the extrema $\hat{V}|_{\rm ext} = 0$ reads
\begin{equation}
\begin{split}
c_3 \sim {18 \sqrt{2} (\beta-1) x_s^{3/2} \over \beta (4 (\beta-1) x_s - 3)^2} {e^{\beta x_s} \over {\cal V}_x}.
\end{split}
\label{minkowski condition}
\end{equation}
Next, we proceed to check the stability at the Minkowski point. Although we know the conditions to check the stability, the formula of the Hessian is yet too complicated to perform an analytical analysis. So we further focus on the region satisfying $x_s \gg 1$. The region with $x_s \gg 1$ is motivated since the $AdS$ minimum points, before adding an uplift term, are guaranteed to have a positive Hessian since all eigenvalues are positive definite when satisfying $x_s > 1$ in LVS type stabilizations (see e.g.~\cite{Rummel:2013yta}). Furthermore, higher instanton corrections can be safely neglected. As shown in Figure~\ref{fig:uplift-illustration}, the minima can be uplifted keeping the positivity of the Hessian until reaching the destabilization point with a relatively high positive vacuum energy. Hence, having $x_s \gg 1$ is motivated to see the basic feature of the D-term generated racetrack uplift mechanism. Since there is no reason to take $\beta$ to be small/large, we consider $\beta \sim {\cal O} (1)$.
When we use the approximation $x_s \gg 1$, a component of the Hessian and the determinant at the extrema become (\ref{extremal condition of effective potential})
\begin{equation}
\begin{split}
\partial_{{\cal V}_x}^2 \hat{V}|_{\rm ext} \sim& {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^5} - c_3 {8 \beta x_s e^{-\beta x_s} \over {\cal V}_x^4}
\sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^5},\\
\det \left(\partial_i \partial_j \hat{V} \right)|_{\rm ext} \sim& {162 x_s^2 \over {\cal V}_x^8} + c_3 {24\sqrt{2} \beta(\beta^2 + \beta -2) x_s^{5/2} e^{-\beta x_s} \over {\cal V}_x^7}
\sim {54 (1- \beta) x_s^2 \over {\cal V}_x^8}\,,
\end{split}
\label{analytical hessian at minkowski}
\end{equation}
where in the last step of both equations, we have used the Minkowski condition (\ref{minkowski condition}). According to Sylvester's criterion, the positivity of a matrix can be checked by the positivity of the determinant of all sub-matrices. Thus it is enough to check the positivity of the quantities in (\ref{analytical hessian at minkowski}). Therefore we conclude that the stability at the Minkowski point requires $\beta < 1$. This condition is clearly satisfied in the previous numerical example following (\ref{test beta value}), which may justify the crude approximations we took in this subsection. Note that the Hessian of the imaginary mode is guaranteed to be positive under the above used approximations:
\begin{equation}
\partial_{y_s}^2 \hat{V}|_{\rm ext} \sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^3} - c_3 {4 \beta^2 (\beta +1) x_s e^{-\beta x_s} \over {\cal V}_x^2}
\sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^3}.
\end{equation}
Finally, let us check the extremal and Minkowski conditions in the limit $x_s \gg 1$. Now all conditions are simplified to be
\begin{equation}
\begin{split}
\xi_x \sim 4 \sqrt{2} x_s^{3/2}, \quad
c_2 \sim -{3 \sqrt{x_s} \over \sqrt{2}} {e^{x_s} \over {\cal V}_x}, \quad
c_3 \sim {9 \over 4 \sqrt{2x_s} \beta (1-\beta)} {e^{\beta x_s} \over {\cal V}_x}.
\end{split}
\end{equation}
We see that the minimum point needs $\xi_x >0$ and $c_2 < 0$ in agreement with the minimum requirement of the two-moduli LVS at $AdS$. The stability condition $\beta < 1$ suggests $c_3 > 0$. In fact, the extremal condition for $\xi_x, c_2$ is simply the leading order approximation of each first term in \eqref{extremal condition of effective potential} as the $c_3$ contribution appears sub-dominant. This justifies that the linear approximation for $c_3$ is compatible with $x_s \gg 1$. Hence, we can regard the last term in the potential (\ref{redefined potential}) as the uplift term.
\section{On realization in models with more moduli\label{sec:real-more-moduli}}
In this section, we show that the uplift mechanism works well in the presence of additional K\"ahler moduli in Swiss-Cheese type Calabi-Yau compactifications. We consider a simple toy model with $h^{1,1}_+=4$, which captures the essential features of the D-term generated racetrack uplift mechanism defined by
\begin{equation}
\begin{split}
K=& - 2 \ln \left({\cal V} + {\xi \over 2}\right), \quad
{\cal V} = (T_a + \bar{T}_a)^{3/2} - (T_b + \bar{T_b})^{3/2} - (T_c + \bar{T_c})^{3/2} - (T_e + \bar{T_e})^{3/2},\\
W =& W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3 (T_b+T_c)} + A_4 e^{- a_4 T_e}.
\end{split}
\end{equation}
Again we are interested in the case of a Swiss-Cheese volume for moduli stabilization of the LVS type. Note that we used the name $T_e$ to avoid confusion with $T_D$.
Taking into account the D-term potential generated by the magnetized D7-branes wrapping the divisor $Z$, we assume again that the $Z =\frac12 (T_b - T_c)$ modulus is stabilized at $Z=0$. Setting $a_4 = a_2$ for simplicity, the effective potential at ${\cal V}^{-3}$ from the F-terms is given by
\begin{equation}
\begin{split}
{\hat{V}} \equiv \left({a_2^{-3} W_0^{-2}}\right) V \sim& {3\xi_x \over 4 {\cal V}_x^3} + {4 c_2 x_s \over {\cal V}_x^2}e^{-x_s} + {2\sqrt{2} c_2^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 x_s}
+ {4 c_4 x_4 \over {\cal V}_x^2}e^{-x_4} + {2\sqrt{2} c_4^2 x_4^{1/2}\over 3 {\cal V}_x} e^{-2 x_4}\\
&+ {4 \beta c_3 x_s \over {\cal V}_x^2} e^{-\beta x_s} + {\sqrt{2} c_3^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 \beta x_s} + {2\sqrt{2}\beta c_2 c_3 x_s^{1/2} \over 3 {\cal V}_x} e^{-(1+\beta)x_s},
\end{split}
\end{equation}
where we have further defined $x_e = a_4 \tau_e, \beta=2a_3/a_2$ and $c_4 = A_4/ W_0$ in addition to~\eqref{redefined parameters}. Here we included the term proportional to $c_3^2$ as well as the cross term $c_2 c_3$ even though they are potentially subleading.
When we use a set of parameters:
\begin{equation}
c_2 = -0.0114, \quad c_4 = -3.38 \times 10^{-4}, \quad \xi_x = 19,
\label{input of 4 moduli model}
\end{equation}
then the $AdS$ LVS minimum at $c_3=0$ is located at
\begin{equation}
\begin{split}
{\cal V}_x \sim 2740, \quad x_s \sim 2.60, \quad x_e \sim 1.12.
\end{split} \label{4modminval}
\end{equation}
The stability of multi-K\"ahler moduli models of the LVS type is ensured if the constraint $x_i>1$ is satisfied~\cite{Rummel:2013yta}. Hence, the extremal point~\eqref{4modminval} is stable.
Now we add the uplift terms $c_3\neq 0$ and $\beta=5/6$. The minimum with the input parameters (\ref{input of 4 moduli model}) reaches Minkowski at
\begin{equation}
\begin{split}
c_3 \sim 4.55 \times 10^{-3}, \quad {\cal V}_x \sim 5.64 \times 10^4, \quad x_s \sim 5.45, \quad x_e \sim 2.26.
\end{split}
\end{equation}
Although the volume is drastically changed during the uplift toward $dS$ vacua, we can check the stability of the minimum by plugging the values into the Hessian, similarly to the simple three moduli model. The cosmological constant can further increase in the positive region keeping the stability until the minima exceeds the potential barrier where decompactification happens.
\section{Discussion}
We have proposed an uplift mechanism using the structure of at least two small K\"ahler moduli $T_b$ and $T_c$ in Swiss-Cheese type compactifications. The uplift contribution arises as an F-term potential when using a D-term condition which fixes $\re T_b = c \re T_c$ at a higher scale, where $c$ is determined by magnetized fluxes on D7-branes. The uplift term becomes of the form $e^{-a_s \tau_s}/{\cal V}^2$ at large volumes, and hence it can naturally balance with the stabilizing potential in the Large Volume Scenario (LVS), without requiring suppressions in the coefficient, for instance, by warping or a dilaton dependent non-perturbative effect.
In addition, we have shown that the D-term generated racetrack uplift works in the presence of additional K\"ahler moduli. Together with the fact that constraints on the uplift parameters are rather relaxed, i.e., $\beta< 1$ and $c_3>0$, this makes us optimistic that there should be many manifolds admitting the proposed uplift mechanism.
Since the proposed uplift mechanism requires certain conditions for a D-term constraint and two non-vanishing non-perturbative effects, it should be interesting if we can construct an explicit realization of this model in a particular compactification. Such an explicit construction requires to match all known consistency conditions such as cancellation of Freed-Witten anomalies and cancellation of the D3, D5, and D7 tadpole~\cite{Cicoli:2011qg,Cicoli:2012vw,Louis:2012nb,Cicoli:2013mpa,Cicoli:2013zha}. We hope to report on an explicit example in another paper.
Furthermore, the phenomenological aspect of the proposed uplift mechanism should be interesting. Even though the moduli are essentially stabilized as in the LVS, the resultant behavior of the mass spectrum and/or soft SUSY breaking terms may be different depending on which uplift mechanism we employ to realize the $dS$ vacuum.
Finally, in this paper, we concentrated on analyzing the structure of $dS$ minima. However, the structure of the potential is also changed by the uplift term in regions that might be important for including inflationary dynamics. We relegate the analysis of possible inflation scenarios as well as phenomenological consequences compared to other uplift proposals to future work.
\section*{Acknowledgments}
We would like to thank Joseph P. Conlon and Roberto Valandro for valuable discussions and important comments, and also the organizers of the workshop "String Phenomenology 2014" held at ICTP Trieste, Italy, where some of the results of this paper were presented. YS is grateful to the Rudolph Peierls Centre for Theoretical Physics, University of Oxford where part of this work was done for their hospitality and support. MR is supported by the ERC grant `Supersymmetry Breaking in String Theory'. This work is partially supported by the Grant-in-Aid for Scientific Research (No. 23244057) from the Japan Society for the Promotion of Science.
|
\section{Introduction}
Let $k$ be a local field, and let $\oink$ be the ring of integers of $k$.
Recall that the set of maximal orders in $\matrici_2(k)$ is in correspondence with the set of vertices in the Bruhat-Tits tree (or BT-tree) for $PSL_2(k)$ \cite{trees}, \cite{vigneras}. In \cite{eichler2} we proved that the subtree $S_0(\Ha)$ whose vertices are the maximal orders containing a given suborder lies in a rather restricted family. In fact, for most orders, the tree $S_0(\Ha)$, also called the branch of $\Ha$ in what follows, is the maximal subtree whose orders lie no farther than a fixed distance, the depth $p=p(\Ha)$, from a path (Figure 1b), called the stem of $\Ha$, which can have length 0 (Figure 1a), be infinite in one (Figure 1c), or both ends (Figure 1d). Such a set is called a $p$-thick line, or a thick line if the depth is irrelevant. The exceptions are the following:
\begin{itemize}
\item If $\Ha$ is the rank 2 order generated by a nilpotent element, $S_0(\Ha)$ is an infinite leaf, which can be thought as a thick line with the stem at infinity (Figure 1e),
\item if $\Ha=\oink$, identified with the ring of integral scalar matrices, then $S_0(\Ha)$ is the whole BT-tree.
\end{itemize}
In particular, for any order contained in finitely many maximal orders, as is the cases for orders of maximal rank, the branch $S_0(\Ha)$ is completely described by two invariants, the depth $p$, and the stem length $l=l(\Ha)$. In fact, the intersection of the maximal orders containing $\Ha$ is $$\Da^{[p]}=\oink\mathbb{I}+\pi^p\Da,$$ where $\Da$ is an Eichler order of level $l$, and $\pi$ is any uniformizing parameter of $k$. These intersections were originally described by Tu \cite{tu}.
\begin{figure}
\unitlength 1mm
\linethickness{0.4pt}
\ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi
\begin{picture}(120,16)(0,0)
\put(10,14){\makebox(0,0)[cc]{a}}
\put(10,6){\line(0,1){4}}
\multiput(10,6)(-0.087,-.05){40}{\line(1,0){.0625}}
\multiput(10,6)(0.087,-.05){40}{\line(1,0){.0625}}
\multiput(10,10)(-0.087,.05){20}{\line(1,0){.0625}}
\multiput(10,10)(0.087,.05){20}{\line(1,0){.0625}}
\put(6.6,4){\line(0,-1){2}}
\multiput(6.6,4)(-0.087,.05){20}{\line(1,0){.0625}}
\put(13.4,4){\line(0,-1){2}}
\multiput(13.4,4)(0.087,.05){20}{\line(1,0){.0625}}
\put(31,14){\makebox(0,0)[cc]{b}}
\put(25,6){\line(1,0){4}}
\put(29,6){\line(0,1){4}}
\multiput(29,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(29,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(25,6)(-0.05,-0.087){40}{\line(1,0){.0625}}
\multiput(25,6)(-0.05,0.087){40}{\line(1,0){.0625}}
\put(23,9.28){\line(-1,0){2}}
\put(23,2.72){\line(-1,0){2}}
\multiput(23,9.28)(0.05,0.087){20}{\line(1,0){.0625}}
\multiput(23,2.72)(0.05,-0.087){20}{\line(1,0){.0625}}
\put(29,6){\line(1,0){5}}
\put(34,6){\line(0,1){4}}
\multiput(34,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(34,10)(0.087,0.05){20}{\line(1,0){.0625}}
\put(34,6){\line(1,0){4}}
\multiput(38,6)(0.05,-0.087){40}{\line(1,0){.0625}}
\multiput(38,6)(0.05,0.087){40}{\line(1,0){.0625}}
\put(40,9.28){\line(1,0){2}}
\put(40,2.72){\line(1,0){2}}
\multiput(40,9.28)(-0.05,0.087){20}{\line(1,0){.0625}}
\multiput(40,2.72)(-0.05,-0.087){20}{\line(1,0){.0625}}
\put(54,14){\makebox(0,0)[cc]{c}}
\put(50,6){\line(1,0){4}}
\put(54,6){\line(0,1){4}}
\multiput(54,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(54,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(50,6)(-0.05,-0.087){40}{\line(1,0){.0625}}
\multiput(50,6)(-0.05,0.087){40}{\line(1,0){.0625}}
\put(48,9.28){\line(-1,0){2}}
\put(48,2.72){\line(-1,0){2}}
\multiput(48,9.28)(0.05,0.087){20}{\line(1,0){.0625}}
\multiput(48,2.72)(0.05,-0.087){20}{\line(1,0){.0625}}
\put(54,6){\line(1,0){5}}
\put(59,6){\line(0,1){4}}
\multiput(59,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(59,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(60,6)(1,0){4}{{\rule{.4pt}{.4pt}}}
\put(76,14){\makebox(0,0)[cc]{d}}
\put(71,6){\line(0,1){4}}
\multiput(71,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(71,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(71,6)(-1,0){4}{{\rule{.4pt}{.4pt}}}
\put(71,6){\line(1,0){5}}
\put(76,6){\line(0,1){4}}
\multiput(76,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(76,10)(0.087,0.05){20}{\line(1,0){.0625}}
\put(76,6){\line(1,0){5}}
\put(81,6){\line(0,1){4}}
\multiput(81,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(81,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(82,6)(1,0){4}{{\rule{.4pt}{.4pt}}}
\put(102,14){\makebox(0,0)[cc]{e}}
\put(94,6){\line(0,1){4}}
\multiput(94,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(95.8,11)(0.087,-0.05){10}{\line(1,0){.0625}}
\put(95.8,11){\line(0,1){1}}
\multiput(94,10)(0.087,0.05){20}{\line(1,0){.0625}}
\multiput(92.2,11)(-0.087,-0.05){10}{\line(1,0){.0625}}
\put(92.2,11){\line(0,1){1}}
\multiput(94,6)(-1,0){4}{{\rule{.4pt}{.4pt}}}
\put(94,6){\line(1,0){8}}
\put(102,6){\line(0,1){4}}
\multiput(102,10)(-0.087,0.05){20}{\line(1,0){.0625}}
\multiput(102,10)(0.087,0.05){20}{\line(1,0){.0625}}
\put(102,6){\line(1,0){5}}
\put(107,6){\line(0,1){4}}
\put(107,6){\line(1,0){4}}
\end{picture}
\caption{Some examples of branches of orders when the residue field is
$\mathbb{K}=\oink/\pi\oink\cong\finitum_2$.}
\end{figure}
In this work we show how these invariant can be computed in terms of a generating set for $\Ha$. As an example we compute all branches of orders of the form $\Ha=\oink[i,j]$ where $i$ and $j$ are a pair of orthogonal pure quaternions, the standard generators of a quaternion algebra. These computations are useful to study which Eichler orders represent a given suborder. A fundamental tool for this study is the representation field, whose definition is recalled in the last section. The relative spinor image, which permits the computation of representation fields, is easily expressed in terms of a third invariant, the diameter $d(\Ha)=l(\Ha)+2p(\Ha)$. It is the diameter of $S_0(\Ha)$ as a metric space with the canonical metric in a tree. Other applications of the invariants are presented in \S5. For instance, we provide a formula for the number of maximal orders containing a given order of maximal rank, or any local order with finite invariants.
If $S$ is a thick line, we also denote its invariants by $d(S)$, $l(S)$ and $p(S)$. When the stem of $S$ is a maximal path, we write $d(S)=l(S)=\infty$. When the stem of $S$ is infinite in one direction only, we write $d(S)=l(S)=\infty/2$.
In all that follows, when $\Ha=\oink[a_1,\dots,a_n]$ we write $S_0(a_1,\dots,a_n)$ instead of $S_0(\Ha)$. Note that
$S_0(a_1,\dots,a_n)=\bigcap_{i=1}^nS_0(a_i)$.
This property is used in \cite{eichler2} to characterize branches of orders among all possible subtrees of the BT-tree. The intersection of every non-empty sub-collection of the sets on the right of this identity is a thick line or an infinity leaf,
so it suffices to describe the intersection of any two of these sets in terms of their relative positions. Then we need to determine these relative positions for explicit orders.
This needs to be done by an ad-hoc argument in every specific case, but Lemma \ref{l31} is particularly useful there. The possible shapes for each branch $S=S_0(a_i)$
are fully described by the following list \cite[Cor. 4.3 and Prop. 4.4]{eichler2}:
\begin{enumerate}
\item If $L=k(a_i)$ is a field, then $S$ is a thick line with a stem of length 1 if $L/k$ is ramified and 0 otherwise.
\item If $L=k(a_i)\cong k\times k$, then $S$ is a thick line with $l(S)=\infty$.
\item If $k(a_i)$ contains a non-trivial nilpotent element, then $S$ is an infinite leaf (Fig. 1e). \end{enumerate}
Furthermore, in either of the first two cases, the depth $p=p(\oink[a_i]\big)$ is defined by the identity $\oink[a_i]=\oink_L^{[p]}=\oink+\pi^p\oink_L$, where $\oink_L$ is the maximal order of $L$.
\section{Explicit intersections}
Let $G$ be a graph and let $\delta$ be the usual distance on it. For any subgraph $S$ of $G$ we define
$S^{[r]}=\bigcup_{x\in S}B[x;r]$, where $B[x;r]=B_\delta[x;r]$ is the closed ball of radius $r$ centered at $x$.
The length of a path $\gamma$ is denoted $l(\gamma)$.
\begin{lemma}\label{lfund}
If $G$ is a tree, while $S_1$ and $S_2$ are two subtrees with non-empty intersection, we have
$(S_1\cap S_2)^{[r]}=S_1^{[r]}\cap S_2^{[r]}$.
\end{lemma}
\begin{proof}
It is clear that $(S_1\cap S_2)^{[r]}\subseteq S_1^{[r]}\cap S_2^{[r]}$, so we prove the opposite inclusion.
First observe that since the path joining two vertices in a tree is unique, if $S_1$ and $S_2$ are path-connected so is their intersection. Now assume $x\in S_1^{[r]}\cap S_2^{[r]}$. Let $\alpha_i$ be the shortest path joining $x$ to a point
$y_i\in S_i$ (Fig. 2A), so that $l(\alpha_i)=\delta(x,y_i)\leq r$. Let $\gamma_i$ be the shortest path (in $S_i$) joining $y_i$ to a point $z_i\in S_1\cap S_2$. Let $\beta$ be the path (in $S_1\cap S_2$)
joining $z_1$ and $z_2$. By the uniqueness of paths either of the following statements hold:
\begin{itemize}\item $\alpha_2$ passes through $y_1$, $z_1$, and $z_2$
(Fig. 2B), and therefore $l(\beta)=l(\gamma_2)=0$ and $z_1=z_2=y_2$ by definition of $\alpha_2$, or
\item $\alpha_1$ passes through $y_2$, $z_2$, and $z_1$ (Fig. 2C),
and therefore $l(\beta)=l(\gamma_1)=0$ and $y_1=z_1=z_2$.
\end{itemize}
In the first case, the distance from $x$ to $z_1=y_2$ is $l(\alpha_2)\leq r$. The remaining case is analogous.
\end{proof}
\begin{figure}
\[
\includegraphics[width=1\textwidth]{fig1.jpg}
\]
\caption{Paths in the proof of Lemma \ref{lfund}. In (B) $\alpha_2$ equals the juxtaposition $\alpha_1*\gamma_1*\beta*\gamma_2^{-1}$. In (C) we have
$\alpha_1=\alpha_2*\gamma_2*\beta^{-1}*\gamma_1^{-1}$.}
\end{figure}
\begin{cor}
In the notations of the previous lemma, if $S_1\cap S_2$ is a $p$-thick line, then for every
positive integer $t$, the intersection $S_1^{[t]}\cap S_2^{[t]}$
is a $(p+t)$-thick line with the same stem.
\end{cor}
\begin{lemma}\label{lfund2}
If $G$ is a tree, while $S_1$ and $S_2$ are two subtrees satisfying $S_1\cap S_2=\emptyset$ and
$S_1^{[1]}\cap S_2^{[1]}\neq\emptyset$, then $S_1^{[1]}\cap S_2^{[1]}$ is either a point or a path
of length $1$.
\end{lemma}
\begin{proof}
Let $x\in S_1^{[1]}\cap S_2^{[1]}$. Then $x\notin S_1\cap S_2$ as the latter set is empty, whence $x\notin S_1$ or $x\notin S_2$.
If $x\notin S_1$, then $x$ is an endpoint of $S_1^{[1]}$. We conclude that $x$ has a unique neighbor $y$ in $S_1$.
In particular, $y$ is not in $S_2$, whence $y$ is either an endpoint of $S_2^{[1]}$, or $y\notin S_2^{[1]}$.
In the first case, $S_1^{[1]}\cap S_2^{[1]}$ is the path joining $x$ and $y$, while in the latter case
$S_1^{[1]}\cap S_2^{[1]}=\{x\}$.
\end{proof}
In all that follows we write $[\alpha]$ and $\{\alpha\}=\alpha-[\alpha]$ for the integral part and the fractional part, respectively, of a real number $\alpha$. We also use the conventions $\mathrm{min}(\alpha,\infty/2)=\alpha$,
$\alpha+\infty/2=\infty/2$, and $\infty/2+\infty/2=\infty$.
\begin{proposition}\label{inter1}
Let $S_1$ and $S_2$ be two thick lines whose stems $T_1$ and $T_2$ lie at a distance $e>0$ from each other, and
let $a$, $b$, $c$, and $d$, be the length of the segments of the stems determined by the unique path joining them as in Fig. 3A. Let $l_1=a+b$ and $l_2=c+d$ be the stem length of $S_1$ and $S_2$, and let $p_1$ and $p_2$ be the corresponding depths. Then:
\begin{enumerate}
\item if $e>|p_1-p_2|$, then $S_3=S_1\cap S_2$ is a thick line with invariants $p_3=\left[\frac{p_1+p_2-e}2\right]$
and $l_3=2\left\{\frac{p_1+p_2-e}2\right\}\in\{0,1\}$,
\item if $0<e\leq p_1-p_2$, then $S_3=S_1\cap S_2$ is a thick line with invariants $p_3=p_2$
and $l_3=\mathrm{min}\{\tau,c\}+\mathrm{min}\{\tau,d\}$, where $\tau=p_1-p_2-e$,
\item if $0<e\leq p_2-p_1$, then $S_3=S_1\cap S_2$ is a thick line with invariants $p_3=p_1$
and $l_3=\mathrm{min}\{\mu,a\}+\mathrm{min}\{\mu,b\}$, where $\mu=p_2-p_1-e$.
'\end{enumerate}
\end{proposition}
\begin{figure}
\[ \fbox{ \xygraph{
!{<0cm,0cm>;<.8cm,0cm>:<0cm,.8cm>::} !{(1,0) }*+{\bullet}="a" !{(0.5,0) }*+{T_1}="a1"
!{(3,0)}*+{\bullet}="b" !{(3.6,0) }*+{T_2}="b1" !{(0,1) }*+{\bullet}="c"
!{(0,-1) }*+{\bullet}="d" !{(4,1)}*+{\bullet}="e"
!{(4,-1)}*+{\bullet}="f"
!{(2,-1) }*+{\mathrm{(A)}}="z"
"a"-@{.}^e"b" "a"-_a"c" "a"-^b"d" "b"-^c"e"
"b"-_d"f" } }
\fbox{ \xygraph{
!{<0cm,0cm>;<.8cm,0cm>:<0cm,.8cm>::} !{(1,0) }*+{\bullet}="a" !{(0.7,-0.3) }*+{x_1}="a1"
!{(3,0)}*+{\bullet}="b" !{(-0.3,0.7) }*+{y_1}="b1" !{(3.4,0.3) }*+{x_2}="b1" !{(0,1) }*+{\bullet}="c"
!{(4,-1)}*+{\bullet}="e" !{(4.3,-0.7) }*+{y_1}="e1" !{(2,-1) }*+{\mathrm{(B)}}="z"
"a"-@{.}^e"b" "a"-@{--}_{\Delta_1}"c" "b"-@{--}^{\Delta_2}"e"} }
\]
\caption{Disposition of the stems in Proposition \ref{inter1} (A), and the vertices used in the proof (B).}
\end{figure}
\begin{proof}
Assume first that $e>|p_1-p_2|$. Without loss of generality we can assume $p_2\leq p_1$. Now, we can write $S_i=U_i^{[p_2]}$,
for $i\in\{1,2\}$, where $U_2$ is a path, while $U_1$ is a $(p_1-p_2)$-thick line not intersecting $U_2$. We conclude that, for some positive integer $t\leq p_2$, we have $U_1^{[t-1]}\cap U_2^{[t-1]}=\emptyset$ and $U_1^{[t]}\cap U_2^{[t]}\neq\emptyset$. We conclude from Lemma \ref{lfund2} and the corollary to Lemma \ref{lfund} that $l(S_3)\in\{0,1\}$. Since the
invariants satisfy the relation $d(S_3)=l(S_3)+2p(S_3)$, it suffices to compute the diameter. We claim that the
diameter of $S_3$ is $d=p_1+p_2-e$. The result follows easily from the claim. Now let $x_1$ be the point of $T_1$ that is closest to $T_2$, and let $x_2$ be the point of $T_2$ that is closest to $T_1$. Let $y_1$ be a point in the BT-tree
such that the path from $y_1$ to $x_2$ passes through $x_1$ and set $\delta(y_1,x_1)=\Delta_1$.
Define $y_2$ and $\Delta_2$ analogously as in
Figure 3B. Then $y_1$ belongs to $S_2$ if and only if $\Delta_1\leq p_2-e$, and all points satisfying this condition are in $S_1$. Similarly, $y_2$ belongs to $S_1$ if and only if $\Delta_2\leq p_1-e$. The result follows.
Assume now that $0<e\leq p_1-p_2$. Then we can write $S_i=U_i^{[p_2]}$,
for $i\in\{1,2\}$, where $U_2=T_2$ is a path, while $U_1$ is a $(p_1-p_2)$-thick line intersecting $U_2$.
In this case $U_1\cap U_2$ is a path of length $l_3$ as in the statement, so the result follows.
The final case is analogous.
\end{proof}
\begin{proposition}\label{inter2}
Let $S_1$ and $S_2$ be two thick lines whose stems $T_1$ and $T_2$ intersect as shown in Figure 4,
so that $l_1=a+e+b$ and $l_2=c+e+d$ are their respective stem lengths.
Let $p_1$ and $p_2$ be the depths of $S_1$ and $S_2$, respectively. Then $S_3=S_1\cap S_2$ has invariants
$p_3=\mathrm{min}\{p_1,p_2\}$ and
$$l_3=\left\{\begin{array}{rcl}
e+\mathrm{min}\{a,p_2-p_1\}+\mathrm{min}\{b,p_2-p_1\}&\textnormal{ if }&p_1\leq p_2,\\
e+\mathrm{min}\{c,p_1-p_2\}+\mathrm{min}\{d,p_1-p_2\}&\textnormal{ if }&p_2\leq p_1\end{array}\right..
$$
\end{proposition}
\begin{figure}
\[ \fbox{ \xygraph{
!{<0cm,0cm>;<.8cm,0cm>:<0cm,.8cm>::} !{(1,0) }*+{\bullet}="a" !{(2,1) }*+{T_1}="a1"
!{(3,0)}*+{\bullet}="b" !{(2,-1) }*+{T_2}="b1" !{(0,1) }*+{\bullet}="c"
!{(0,-1) }*+{\bullet}="d" !{(4,1)}*+{\bullet}="e"
!{(4,-1)}*+{\bullet}="f" "a"-^e"b" "a"-_a"c" "a"-^c"d" "b"-^b"e"
"b"-_d"f" } }
\]
\caption{Disposition of the stems in Proposition \ref{inter2}.}
\end{figure}
\begin{proof}
Assume $p_1\leq p_2$. Reasoning as before, we can assume that $p_1=0$, so $S_1=T_1$ is a line. The result follows.
The other case is analogous.
\end{proof}
In all that follows, for any metric space $X$ and any subset $A$, we define a relative depth function $p(A,-):A\rightarrow\mathbb{R}$ by
$$p(A,x)=\sup\left\{r\in[0,\infty)\Big|B(x;r)\subseteq A\right\}.$$
It is apparent that for any two sets $A_1,A_2\subseteq X$
we have $p(A_1\cap A_2,x)=\mathrm{min}\{p(A_1,x),p(A_2,x)\}$, for any point $x\in A_1\cap A_2$. This concept is useful to us since for any order $\Ha$
with a finite branch, we have $$p(\Ha)=\max_{x\in S_0(\Ha)}p\big(S_0(\Ha),x\big),$$ while this maximum is attained precisely at the points in the stem.
Recall from \cite{eichler2} that the points in the branch of an order can be classified into stem and leaf points according to the depth of each vertex and its neighbors, as shown in Fig. 5. Note that, from a leaf point there is always a unique direction leading to vertices of larger depth, whence every path without backtracking in a different direction is a path going outwards through the leaves, a path that cannot be longer that the depth of the starting point. In an infinite leaf $S$, which has no stem points, every path without backtracking has at most one vertex at maximal depth, and the path on every side from such vertex goes outwards through the leaves. Such paths can only intersect the infinite leaf on a finite path. On the other hand, starting from every vertex of the infinite leaf, there is a unique path without backtracking leading always in the direction of higher depth. Such paths, which are infinite on one direction, are called long paths of the leaf. The long paths corresponding to two vertices $\Da$ and $\Da'$ always coincide from some point $\Da''$ onwards, and the segments determined by $\Da''$ on each path, as in Fig. 6A,
satisfy $p(S,\Da'')=a+p(S,\Da)=b+p(S,\Da')$, so in particular, if $\Da$ and $\Da'$ are endpoints of $S$, i.e. their relative depths are zero, we have $a=b$.
\begin{figure}
\scriptsize
\[
\includegraphics[width=1\textwidth]{fig2.jpg}
\]
\normalsize
\caption{Stem or leaf points. Arrows indicate "outwards" directions. Double lines denote stem edges.}
\end{figure}
\begin{figure}
\[ \includegraphics[width=1\textwidth]{fig3.jpg}
\]
\caption{Two long paths in an infinite leaf (A), and a stem not intersecting an infinite leaf (B).}
\end{figure}
\begin{proposition}\label{inter3}
Let $S_1$ be a thick line and let $S_2$ be an infinite leaf. Let $T_1$ be the stem of $S_1$.
Let $p_3$ and $l_3$ be the invariants of $S_3=S_1\cap S_2$,
while $p_1$ is the depth of $S_1$. Then,
exactly one of the following conditions hold:
\begin{enumerate}
\item $T_1$ has a vertex $\Da''$ at maximal depth $p_0>p_1$ in $S_2$, which divides $T_1$ in two parts of lengths
$a$ and $b$. In this case $p_3=p_1$ and $l_3=\mathrm{min}\{a,p_0-p_1\}+\mathrm{min}\{b,p_0-p_1\}$.
\item $T_1$ has a vertex $\Da''$ at maximal depth $p_0$, with $0\leq p_0\leq p_1$. In this case $p_3=[\frac{p_1+p_0}2]$, and
$l_3=2\{\frac{p_1+p_0}2\}\in\{0,1\}$.
\item $T_1$ contains a long path of $S_2$. In this case $l_3=\infty/2$, and $p_3=p_1$.
\item $T_1$ lies at a positive distance $e\leq p_1$ from the infinite leaf $S_2$ (Fig. 6B).
In this case, $p_3=[\frac{p_1-e}2]$, and
$l_3=2\{\frac{p_1-e}2\}\in\{0,1\}$.
'\end{enumerate}
\end{proposition}
\begin{proof}
Note that, if $e=\delta(T_1,S_2)>p_1$, then $S_1\cap S_2=\emptyset$.
It follows from the paragraph preceding the proposition that (1)-(4) are actually all possible cases.
Assume the hypotheses in case 1. If $p_1=0$, so that $S_1=T_1$ is a path, then so is $S_3$,
and the result follows from a quick look at Figure 6A. More generally, note that $S_2=U_t^{[t]}$
for an arbitrary positive integer $t$ and a suitable infinite leaf $U_t$. Since $p(U_{p_1},\Da'')=p(S_2,\Da'')-p_1$,
the result follows in this case.
Assume the hypotheses in case 3. Setting $S_2=U_{p_1}^{[p_1]}$ as before, we are reduced to the case $p_1=0$,
which is trivial.
In case 2, we observe that $U_{p_1}$ does not intersect $T_1$, whence there is a positive integer
$t\leq p_1$ such that $U^{[t]}_{p_1}$ intersects $T^{[t]}_1$, but $U^{[t-1]}_{p_1}$ does not intersect $T^{[t-1]}_1$.
We conclude that $l_3\in\{0,1\}$. Now the result follows by computing the diameter of $S_3$ as in Proposition
\ref{inter1}. Case 4 is similar.
\end{proof}
\begin{example}
If $\Ha$ is an order generated by an suborder $\Ha_0\cong\oink\times\oink$ and a nilpotent element $\epsilon$,
the branch $S=S_0(\Ha)$ is a path, and $l(S)\neq\infty$, since $S_0(\epsilon)$ does not contain a maximal path.
It follows that there exist one or two maximal orders $\Da$,
containing $\Ha$, for which $\epsilon\notin\pi\Da$, the endpoints of $S$.
There exists exactly one such order if and only if $l(S)=\infty/2$, i.e., $S_0(\Ha_0)$ contains
a long path of the infinite leaf $S_0(\epsilon)$. This holds if and only if $\Ha_0\big(\mathrm{Ker}(\epsilon)\big)=\mathrm{Ker}(\epsilon)$. Up to scalar multiples, there are exactly two nilpotent elements with this property.
\end{example}
\begin{proposition}\label{inter4}
Let $S_1$ and $S_2$ be two infinite leaves, and let $S_3=S_1\cap S_2$. Then $S_3$ is infinite if and only if
$S_1$ and $S_2$ have a comon long path. If this holds, then either $S_1\subseteq S_2$ or $S_2\subseteq S_1$, and in this case $S_3$ is an infinite leaf. If $S_3$ is finite, and the maximal depth $p(S_1,\Da)$ of a vertex $\Da$ of $S_3$ is $r$, then $p_3=[\frac r2]$, and $l_3=2\{\frac r2\}\in\{0,1\}$.
\end{proposition}
\begin{proof}
By inspecting the explicit list of possible branches of orders given in \S1, we conclude that every infinite branch contains an infinite path, and this can only be a long path in each of the branches $S_1$ and $S_2$. First statement follows. Assume now that $S_3$ is infinite and let $\Da$ be an arbitrary vertex of $S_3$. Assume $p(S_1,\Da)\geq p(S_2,\Da)$. Then if $U$ is the long path starting at $\Da$, and $\Da_i$ is the $i$-th vertex of that path we have
$$S_1=\bigcup_{i=0}^\infty B\Big(\Da_i;p\big(S_1,\Da\big)+i\Big)\supseteq \bigcup_{i=0}^\infty B\Big(\Da_i;p\big(S_2,\Da\big)+i\Big)=S_2.$$
Finally, assume $S_3$ is finite. Take a vertex $\Da\in S_3$ whose depth $r=p(S_1,\Da)$ is maximal. Since any path inside $S_3$ starting from $\Da$ goes outwards through the leaves in $S_1$, we have $S_3=S_2\cap B(\Da;r)$. Note also that $\Da$ is an endpoint of $S_2$, whence the result follows by setting $a=b=p_0=0$ and $p_1=r$ in case (2) of Proposition \ref{inter3}.
\end{proof}
\section{Relative position of the branches}
The following lemma, whose proof is straightforward, is as close as we can get to give a general method to determine the relative position
of the branches:
\begin{lemma}\label{l31}
If the distance between two branches $S_1$ and $S_2$ is $d$, then $S_1^{[t_1]}$ and $S_2^{[t_2]}$ intersect
if and only if $t_1+t_2\geq d$.
\end{lemma}
If we can write two orders $\Ha_1$ and $\Ha_2$ as the contractions $\Ea_1^{[s_1]}$ and $\Ea_2^{[s_2]}$
of two orders $\Ea_1$ and $\Ea_2$ whose branches are the stems of the branches
$S_1=S_0(\Ha_1)$ and $S_2=S_0(\Ha_2)$, the preceding lemma can be used to find the distance between the
corresponding stems, and thence the invariants for the spanned order $\Ha_3=\oink[\Ha_1,\Ha_2]$.
If one of the branches is an infinite leaf, the same trick can be applied to find out the maximal depth of the intersection,
as in the following example:
\begin{example}
Let $\epsilon_1$ and $\epsilon_2$ be two nilpotent elements satisfying $\epsilon_1\epsilon_2+\epsilon_2\epsilon_1=\pi^2I$,
where $\pi$ is a uniformizing parameter of $k$ and $I$ is the identity matrix.
Then there is an order containing $\epsilon_1$ and $\frac{\epsilon_2}{\pi^t}$ if and only if $t\leq2$. It follows that
the greatest depth in $S_0(\epsilon_2)$ of an element in $S_0(\epsilon_1)$ is $2$. We conclude that $S=S_0(\epsilon_1,\epsilon_2)$ has the invariants $d(S)=2$, $l(S)=0$, and $p(S)=1$.
\end{example}
It is
sometimes better to replace one of the orders by a simpler order with the same branch (or an appropriate sub-branch),
in order to perform computations. The following example illustrate this point.
\begin{example}
Assume $k=\mathbb{Q}_3$. Let $\Ha'=(\Da_1\cap\Da_2)^{[t]}$, where $\Da_1$ is the only
maximal order containing $\eta=\lbmatrix 01{-1}0$ and $\Da_2$ is the only
maximal order containing $\eta'=\lbmatrix 0{27}{-1/27}0$. Let $\epsilon=\lbmatrix 0{81}00$, so that
$\epsilon\eta+\eta\epsilon=-81I$ and $\epsilon\eta'+\eta'\epsilon=-3I$.
Then, as before, $\Da_1$ has depth
1 in $S_0(\epsilon)$, while $\Da_2$ has depth 4. On the other hand, the relation
$\eta\eta'+\eta'\eta=-\frac{730}{27}I$
shows that the distance between $S_0(\eta)=\{\Da_1\}$ and $S_0(\eta')=\{\Da_2\}$
is exactly $3$. We conclude that the branch of $\Ha'$
has a stem of length 3 whose deepest point is the endpoint $\Da_2$ and it is at depth $4$. Therefore,
the invariants for the order $\Ha=\oink[\Ha',\epsilon]$ generated by $\Ha'$ and $\epsilon$ are
(c.f. Prop. \ref{inter3}):
\begin{itemize}
\item $p(\Ha)=t$, and $l(\Ha)=\mathrm{min}\{3,4-t\}$ if $t\leq4$,
\item $p(\Ha)=2+[\frac t2]$, and $l(\Ha)=2\{\frac t2\}$ otherwise.
\end{itemize}
\end{example}
Now we compute the distance between branches of standard generators. Before we do that, we prove a formula for
the branch of a pure quaternion. This is computed in terms of the quadratic defect of its square \cite{ohm}.
Recall that the quadratic defect of
an element $a\in k$ is the fractional ideal $\mathfrak{d}(a)$ generated by all elements $\eta\in k$ satisfying $|a-u^2|_k\geq|\eta|_k$ for all $u\in k$. There is always an element $u$ satisfying $(a-u^2)=\mathfrak{d}(a)$. Furthermore, for every uniformizing parameter $\pi$ we have $\mathfrak{d}(\pi)=(\pi)$, while
the unique unramified quadratic extension of $k$ has the form $k(\sqrt\Delta)$ for any unit $\Delta$
satisfying $\mathfrak{d}(\Delta)=(4)$. The latter is called a unit of minimal quadratic defect.
The quadratic defect of any other unit $u$ has the form $\mathfrak{d}(u)=(\pi)^{2t+1}$, where $0\leq t<e=v_k(2)$.
\begin{lemma}
The depth $p=p(S)$ of the branch $S=S_0(q)$, for any pure quaternion $q$ such that $q^2=a\in \oink$, satisfies
$\mathfrak{d}(a)=(\pi)^{2p+1}$, except in the following cases:
\begin{enumerate}
\item If $q$ generates an unramified quadratic extension, then $\mathfrak{d}(a)=(\pi)^{2p}$.
\item If $a=b^2$, for some $b\in \oink$, then $p=v_k(2b)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Without loss of generality, we can assume $a$ is either a unit or a uniformizing parameter.
Assume $\mathfrak{d}(a)=(\pi)^{2t+1}$, so in particular $k(q)/k$ is a ramified extension. Then, there exists a unit
$v\in\oink^*$ and an integer $b\in\oink$ satisfying $a=b^2+v\pi^{2t+1}$. Let $\omega=\frac{q-b}{\pi^t}$. The quaternion $\omega$ is integral over $\oink$, since its norm is $\frac{b^2-q^2}{\pi^{2t}}=-\pi v$ and its trace is
$\frac{-2b}{\pi^t}$, which is an integer since $t\leq v_k(2)$.
It follows that $\oink[q]\subseteq \oink_{k(q)}^{[t]}$, and therefore $p\geq t$ by \cite[Prop. 2.4 and Prop. 4.2]{eichler2}.
On the other hand, the branch of $\oink_{k(q)}$ is a path of length 1, and therefore its depth is $0$.
If the branch of $\oink[q]$ has
depth $p$, there must exists an element $b\in\oink$ such that $\frac{q-b}{\pi^p}$ is an integer \cite[Lemma 2.5]{eichler2}, and therefore
so is the norm $\frac{b^2-q^2}{\pi^{2p}}$. We conclude that $b^2-a\in (\pi^{2p})$, and therefore $p\leq t$.
The unramified case is similar.
Assume now that $a=b^2$, so $k(q)$ can be identified with $k\times k$. Then we can assume $q=(b,-b)$, so that $\omega:=(0,1)=\frac1{2b}[(b,b)-q]$ is integral over $\oink$. Now the proof of the inequality $p\geq v_k(2b)$ goes as before. For the converse we observe that if $\frac{q-(b',b')}{\pi^p}$ is integral over $\oink$, then both $b-b'$ and
$-b-b'$ are in the ideal $(\pi^p)$, and therefore so is their difference $2b$.
\end{proof}
\begin{cor}
In the notations of the previous lemma, if $a=q^2$ is a unit, then
the depth satisfies $p\leq v_k(2)$, with equality if and only
if $k(q)$ is isomorphic to $k\times k$ or an unramified extension. If $a$ is a uniformizing parameter,
the invariants of $S_0(q)$ are $l=1$ and $p=0$.
\end{cor}
Next result give us the relative position of the branches for a cyclic order of the form $\oink[i,j]$ as in the introduction.
\begin{lemma}
Let $\Ha=\oink[i,j]$ be a cyclic order in a split quaternion algebra $\alge=K[i,j]$, for a pair of orthogonal pure quaternions
$i$ and $j$, such that $a=i^2$, and $b=j^2$, are both in $\oink^*\cup \pi\oink^*$.
\begin{enumerate}
\item Assume $a$ and $b$ are units,and let $s$ and $r$ be the depths of $S_0(i)$ and $S_0(j)$ respectively.
\begin{enumerate}
\item If $v_k(2)-s-r<0$, then the distance between the stems of $S_0(i)$ and $S_0(j)$ is $s+r-v_k(2)$.
\item If $v_k(2)-s-r=0$, then the stems of $S_0(i)$ and $S_0(j)$ intersect in a single point.
\item If $v_k(2)-s-r>0$, the stems coincide.
\end{enumerate}
\item If $a$ is prime, then the stem of $S_0(i)$ is contained in the stem of $S_0(j)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Reasoning as in the proof of the previous lemma, there are elements $c,d\in\oink$ and integral elements $\eta\in k(i)$, $\omega\in k(j)$,
such that $i=d+\pi^s\eta$ and $j=c+\pi^r\omega$. The relation $ij=-ji$ implies
\begin{equation}\label{etaom}
\omega\eta+\eta\omega=\frac2{\pi^{r+s}}cd+\frac2{\pi^r}c\eta+\frac2{\pi^s}d\omega.
\end{equation}
Assume first that $a$ and $b$ are units and the condition in (1a) is satisfied,
so that in particular $r$ and $s$ are positive and therefore $c$ and $d$ are units. Then the first term on the
right of equation (\ref{etaom}) is dominant. In particular, if $\omega$ and $\pi^t\eta$ are in a maximal order,
the element $\pi^t(\omega\eta+\eta\omega)$ is integral, and therefore $t\geq s+r-v_k(2)$. On the other hand,
if $t\geq s+r-v_k(2)$, equation (\ref{etaom}) and the fact that $\eta$ and $\omega$ satisfy monic quadratic polynomials over $\oink$, proves that the lattice with basis $\{1,\pi^t\eta,\omega,\pi^t\eta\omega\}$ is a ring.
The result follows in Case (1a).
Assume now that $v_k(2)-s-r=0$. As before, there is a maximal order $\Da$ containing $\eta$ and $\omega$. We can
assume that $\Da=\matrici_2(\oink)$. Let $\bar{\omega}$ and $\bar{\eta}$ be the images of $\omega$ and $\eta$
in $\Da/\pi\Da\cong\matrici_2(\mathbb{K})$, where $\mathbb{K}$ is the residue field of $k$.
To prove that $\Da$ is the unique order
containing $\eta$ and $\omega$, it suffices to show that $\bar{\omega}$ and $\bar{\eta}$ have no common eigenvector.
\begin{itemize}
\item Assume first that $k$ is dyadic, and $rs\neq0$. Then the equation for $\bar{\omega}$ and $\bar{\eta}$ becomes
$ \bar{\omega}\bar{\eta}+\bar{\eta}\bar{\omega}=\bar{u}\bar{c}\bar{d}$ where the bar denotes projection to the residue field, while $c$, $d$ and $u=\frac2{\pi^{r+s}}$, are all units. If $v\in\mathbb{K}^2$ is a comon eigenvector for $\bar{\omega}$ and $\bar{\eta}$ with corresponding eigenvalues $\lambda_{\omega}$ and $\lambda_{\eta}$, we have
$0=\lambda_{\omega}\lambda_{\eta}+\lambda_{\eta}\lambda_{\omega}=\bar{u}\bar{c}\bar{d}$, a contradiction.
\item If $k$ is dyadic, but $rs=0$, we can assume $r=0$ and therefore we can also assume $j=\omega$, so $c=0$.
Note that $s\neq0$ and $u=2/\pi^s$ is a unit, since $r+s=v_k(2)$. The equation for $\bar{\omega}$ and $\bar{\eta}$ becomes $ \bar{\omega}\bar{\eta}+\bar{\eta}\bar{\omega}=\bar{u}\bar{d}\bar{\omega}$,
and the eigenvalue $\lambda_\omega$
is non-zero, as $\omega=j$ is the square root of a unit. Now the result follows as before.
\item If $k$ is non-dyadic, then $s=r=0$ and we can assume as before that $j=\omega$ and $i=\eta$. The result
follows since in this case $\bar{\omega}$ and $\bar{\eta}$ are generators of a quaternion algebra, i.e., they generate
$\matrici_2(\mathbb{K})$.
\end{itemize} The result follows in Case (1b).
Assume now that $v_k(2)-s-r>0$, so that, in particular, $k$ is dyadic. This case is similar to the preceding one, except that the equation in the residual algebra is $ \bar{\omega}\bar{\eta}+\bar{\eta}\bar{\omega}=0$. The inequality $v_k(2)-s-r>0$ implies that both $i$ and $j$ generate ramified extensions and therefore the stems are paths of length 1. It suffices, as before, to prove that
$\bar{\omega}$ and $\bar{\eta}$ have a common eigenvector, so $\eta$ and $\omega$ are contained simultaneously in more than one
maximal order. Note that $\bar{\omega}$ has an eigenvector,
since $\omega$ generates a ramified extension, and $\bar{\eta}$ leaves this eigenspace invariant, since it commutes
with $\bar{\omega}$. The result follows in Case (1c).
Finally we prove Case (2). In this case, the quadratic defect of $a$ is $(\pi)$, so we can set $i=\eta$, $s=0$, and $d=0$,
in equation (\ref{etaom}). In particular, $v_K(2)-s-r\geq 0$, so that equation (1) proves that $\eta$ and $\omega$ span
an order. It follows that there exists a maximal order $\Da$ containing $\eta$ and $\omega$, and we can assume $\Da=\matrici_2(\oink_K)$. Furthermore, we can assume $i=\lbmatrix 0a10$,
and therefore $\bar{\eta}=\lbmatrix 0010$.
In this case the equation for $\bar{\omega}$ and $\bar{\eta}$ becomes
$ \bar{\omega}\bar{\eta}+\bar{\eta}\bar{\omega}=\bar{u}\bar{d}\bar{\omega}$, and by setting $\bar{\omega}=\lbmatrix xyzw$, the preceding equation gives $y=0$, so that $\lcvector01$ is a common eigenvector.
The result follows as before.
\end{proof}
\section{The tables}
\newcommand{\mathcal{O}}{\mathcal{O}}
\newcommand{\mathfrak{p}}{\mathfrak{p}}
\newcommand{\mathfrak{d}}{\mathfrak{d}}
In this section we present complete tables for the invariants of an order of the form $\Ha=\mathcal{O}[\pi^ri,\pi^sj]$ where $i^2=\alpha$ and $j^2=\beta$ are units or primes, while $r$ and $s$ are non-negative integers. All computations are straightforward from Propositions 2.1-2.2 and Propositions 3.2-3.3.
When $k$ is non-dyadic every unit is unramified.
In this case, the depth of the branches, $p_1=p(\mathcal{O}[i])$ and $p_2=p(\mathcal{O}[j])$ are zero.
The invariants of $\Ha$ for all possible values of $\alpha$ and $\beta$ can be read from Table 1,
by switching $\alpha$ and $\beta$ if needed. Note that in the table, the case $\alpha\in\Delta\mathcal{O}^{*2}$,
$\beta\in\pi\mathcal{O}^{*}$
is not shown. This is due to the
fact that $\left(\frac{\Delta,\pi}k\right)$ is always a division algebra. The intersection of the branches
are shown in Figure 7.
\begin{table}[!hbp]
\small\[
\begin{tabular}[c]{|c|c|c|c|c|c|}\hline
$\alpha \in$ & $\beta \in $ & $r-s$ & $p(\mathfrak{H})$ & $l(\mathfrak{H})$ & $d(\mathfrak{H})$\\ \hline
{$\mathcal{O}^{*2}$} & $\mathcal{O}^{*2}$ && $\min\{r,s\}$ & $2|r-s|$ &$2\max\{r,s\}$\\ \hline
{$\mathcal{O}^{*2}$} & $ \Delta\mathcal{O}^{*2}$ &$\leq0$&$r$& $2(s-r)$ & $2s$ \\ \hline
{$\mathcal{O}^{*2}$} & $ \Delta\mathcal{O}^{*2}$ &$\geq0$&$s$& $0$ & $2s$ \\ \hline
{$\mathcal{O}^{*2}$} & {$\pi\mathcal{O}^{*}$} &$\leq0$& $r$ & $2(s-r)+1$ & $2s+1$\\ \hline
{$\mathcal{O}^{*2}$} & {$\pi\mathcal{O}^{*}$} &$\geq0$& $s$ & $1$ & $2s+1$\\ \hline
{$\Delta\mathcal{O}^{*2}$} & $\Delta\mathcal{O}^{*2} $ & &$\min\{r,s\}$&$0$ & $2\min\{r,s\}$ \\ \hline
{$\pi\mathcal{O}^{*}$} & $\pi\mathcal{O}^{*}$ && {$\min\{r,s\}$} & $1$ & {$2\min\{r,s\} + 1$} \\ \hline
\end{tabular}
\]\normalsize
\caption{Invariants of the order $\Ha$ described in the text, when $k$ is non-dyadic.}
\end{table}
\begin{figure}
\[ \includegraphics[width=1\textwidth]{fig4.jpg}
\]
\caption{All possible configurations of the branches $S_1=S_0(i)$ and
$S_2=S_0(j)$ for a non-dyadic local field. Bullets and continuous lines $(\bullet\!-\!\!\!-\bullet)$
denote the intersection $S_3=S_1\cap S_2$.}
\end{figure}
In order to simpliy Table 2, which contain the values for the invariants at
general dyadic places, we define, when both $\alpha$ and $\beta$ are units, an auxiliar invariant $\epsilon$.
Let $$g=g(\alpha,\beta;r,s)=\delta(T_1,T_2)-\Big|p\big(S_0(\pi^ri)\big)-
p\big(S_0(\pi^sj)\big)\Big|,$$
where $T_1$ and $T_2$ are the stems of $S_0(\pi^ri)$ and $S_0(\pi^sj)$ respectively.
Let
$\mho=\mathcal{O}^*\backslash(\mathcal{O}^{*2}\cup\Delta\mathcal{O}^{*2})$ be the set of ramified units. When $\beta\in\mho$,
we define $t$ by $\mathfrak{d}(\beta)=(\pi)^{2t+1}$. Similarly, When $\alpha\in\mho$,
we set $\mathfrak{d}(\alpha)=(\pi)^{2u+1}$. In case $\delta(T_1,T_2)>0$, which is the only case where $\epsilon$
is needed, the number $g$ is given by
the formula (see Lemmas 3.2-3.3):
$$g=
\left\{\begin{array}{rcl}
t+u-e-|u+r-t-s|&\textnormal{ if }&\alpha,\beta\in\mho\\
t-|e+r-t-s|&\textnormal{ if }&\alpha\in \mathcal{O}^{*2}\cup\Delta\mathcal{O}^{*2},\beta\in\mho\\
e-|r-s|&\textnormal{ if }&\alpha,\beta\in \mathcal{O}^{*2}\cup\Delta\mathcal{O}^{*2}\end{array}\right..$$
In all cases we set $\epsilon=2\{g/2\}$, i.e., $\epsilon=0$ if $g$ is even and $1$ otherwise.
Note that the invariant $t$ is defined as $0$ when $\beta\in\pi\mathcal{O}^{*}$. The same observation applies to $u$.
The last two lines in Table 2 deserve some additional explanation. The computation naturally breaks into three
cases acording to whether $t+u-e$ is positive, $0$ or negative (cf. Lemma 3.3). However, in each case the results
coincide, except for the fact that the interval $[e-2u,2t-e]$ is $\{0\}$ when $t+u-e=0$, and empty when
$t+u-e<0$.
\begin{table}[!hbp]
\scriptsize\[
\begin{tabular}[c]{|c|c|c|c|c|c|}\hline
$\alpha \in $ & $\beta \in $ &$(r-s)\in$ & $p(\mathfrak{H})$ & $l(\mathfrak{H})$ & $d(\mathfrak{H})$\\ \hline
$\mathcal{O}^{*2}$ & $\mathcal{O}^{*2}$ &$[-e,e]^c$
& $\min\{r,s\}+e$ &$2w$&$2\max\{r,s\}$\\ \hline
$\mathcal{O}^{*2}$ & $\mathcal{O}^{*2}$ &$[-e,e]$& $\frac12(e+r+s-\epsilon)$
& $\epsilon$ & $e+r+s$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[-\infty,-e]$&$e+r$ & $2w$ & $2s$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[-e,e]$&$\frac12(e+r+s-\epsilon)$
& $\epsilon$ & $e+r+s$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[e,\infty]$&$e+s$ & $0$ & $2(e+s)$ \\ \hline
$\mathcal{O}^{*2}$ & {$\pi\mathcal{O}^{*}$} &$[-\infty,-e]$& $e+r$ & $2w+1$ &$2s+1$\\ \hline
$\mathcal{O}^{*2}$ &{$\pi\mathcal{O}^{*}$} &$[-e,\infty]$& $s$ & $1$ &$2s+1$\\ \hline
$\mathcal{O}^{*2}$ & {$\mho$} &$[-\infty,-e]$&$e+r$ & $2w$ & $2s$ \\ \hline
$\mathcal{O}^{*2}$ & {$\mho$} &$[-e,-e+2t]$& $\frac12(e+r+s-\epsilon)$ & $\epsilon$ &$e+r+s$\\ \hline
$\mathcal{O}^{*2}$ & {$\mho$} &$]-e+2t,\infty]$&$t+s$ & $1$ & $2(t+s)+1$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[-e,e]^c$&
$\min\{r,s\}+e$ &$0$&$2\min\{r,s\}+2e$\\ \hline
$ \Delta\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[-e,e]$
& $\frac12(e+r+s-\epsilon)$ & $\epsilon$ & $e+r+s$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & {$\mho$} &$[-\infty,-e]$&$e+r$ & $0$ & $2(e+r)$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & {$\mho$} &$[-e,-e+2t]$& $\frac12(e+r+s-\epsilon)$
& $\epsilon$ &$e+r+s$\\ \hline
$ \Delta\mathcal{O}^{*2}$ & {$\mho$} &$]-e+2t,\infty]$&$t+s$ & $1$ & $2(t+s)+1$ \\ \hline
$\pi\mathcal{O}^{*}$ & $\pi\mathcal{O}^{*}$ &$\mathbb{R}$&$\min\{r,s\}$ & $1$ & $2\min\{r,s\}+1$ \\ \hline
$\pi\mathcal{O}^{*}$ & $\mho$ &$\mathbb{R}$&$\min\{r,t+s\}$ & $1$ & $2\min\{r,t+s\}+1$ \\ \hline
{$\mho$} & {$\mho$} &$[e-2u,2t-e]^c$& $\min\{r+u,s+t\}$ &$1$&$2\min\{r+u,s+t\}+1$\\ \hline
{$\mho$} & {$\mho$} &$[e-2u,2t-e]$ & $\frac12(e+r+s-\epsilon)$ &$\epsilon$&$e+r+s$\\ \hline
\end{tabular}
\]\normalsize
\caption{Invariants of $\Ha$ when $k$ is dyadic. Here $w=|s-r|-e$.
The numbers $t$, $u$, and $\epsilon$ are defined in the text. $I^c$ is the complement of the interval $I$.
The interval in the last line can be empty.}
\end{table}
\begin{table}[!hbp]
\scriptsize\[
\begin{tabular}[c]{|c|c|c|c|c|c|}\hline
$\alpha \in $ & $\beta \in $ &$(r-s)\in$ & $p(\mathfrak{H})$ & $l(\mathfrak{H})$ & $d(\mathfrak{H})$\\ \hline
$\mathcal{O}^{*2}$ & $\mathcal{O}^{*2}$ &$\mathbb{R}-\{0\}$
& $\min\{r,s\}+1$ &$2|r-s|-2$&$2\max\{r,s\}$\\ \hline
$\mathcal{O}^{*2}$ & $\mathcal{O}^{*2}$ &$\{0\}$& $r$
& $1$ & $2r+1$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[-\infty,-1]$&$r+1$ & $2(s-r-1)$ & $2s$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$\{0\}$&$r$
& $1$ & $2r+1$ \\ \hline
$\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$[1,\infty]$&$s+1$ & $0$ & $2(s+1)$ \\ \hline
$\mathcal{O}^{*2}$ & {$\pi\mathcal{O}^{*}$} &$[-\infty,1]$& $s$ & $1$ &$2s+1$\\ \hline
$\mathcal{O}^{*2}$ &{$\pi\mathcal{O}^{*}$} &$[1,\infty]$& $r+1$ & $2(s-r)-1$ &$2s+1$\\ \hline
$\mathcal{O}^{*2}$ & {$\mho$} &$[-\infty,-1]$&$r+1$ & $2(s-r-1)$ & $2s$ \\ \hline
$\mathcal{O}^{*2}$ & {$\mho$} &$[0,\infty]$&$s$ & $1$ & $2s+1$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$\mathbb{R}-\{0\}$&
$\min\{r,s\}+1$ &$0$&$2\min\{r,s\}+2$\\ \hline
$ \Delta\mathcal{O}^{*2}$ & $ \Delta\mathcal{O}^{*2}$ &$\{0\}$
& $r$ & $1$ & $2r+1$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & {$\mho$} &$[-\infty,-1]$&$r+1$ & $0$ & $2(r+1)$ \\ \hline
$ \Delta\mathcal{O}^{*2}$ & {$\mho$} &$[0,\infty]$&$s$ & $1$ & $2s+1$ \\ \hline
$\pi\mathcal{O}^{*}$ & $\pi\mathcal{O}^{*}\cup\mho$ &$\mathbb{R}$&$\min\{r,s\}$ & $1$ & $2\min\{r,s\}+1$ \\ \hline
$\mho$ & $\mho$ &$\mathbb{R}$&$\min\{r,s\}$ & $1$ & $2\min\{r,s\}+1$ \\ \hline
\end{tabular}
\]\normalsize
\caption{Invariants of $\Ha$ when $k$ is a dyadic field that is unramified over $\mathbb{Q}_2$.}
\end{table}
\begin{figure}
\[ \includegraphics[width=1\textwidth]{fig5.jpg}
\]
\caption{All possible configurations of the branches $S_1=S_0(i)$ and
$S_2=S_0(j)$ for an un-ramified dyadic local field. Bullets and continuous lines denote the
intersection $S_3=S_1\cap S_2$. In the pictures we assume $q=|\oink/\pi\oink|=2$.}
\end{figure}
There is a significant number of possible configurations when $k$ is non-dyadic, and we desist
from the task of drawing them all, but we actually do it in one important case, when $k$ is
an unramified extension of $\mathbb{Q}_2$ (Table 3 and Fig. 8).
In the latter case, $e=1$, so every ramified unit has quadratic defect
$(\pi)$.
A quick glance to the case $r=s=0$ in Table 3, or Figure 8 shows the following important result:
\begin{proposition}\label{p44}
When $k$ is a dyadic field that is unramified over $\mathbb{Q}_2$, every order of the form
$\mathcal{O}[i,j]\subseteq\mathbb{M}_2(k)$, where $ij=-ji$ and $i^2,j^2\in \oink$ are square free, is contained
in exactly $2$ maximal orders.
\end{proposition}
\section{Examples and applications}
\begin{example}\label{ex1}
If $k=\mathbb{Q}_2(\sqrt{-1})$, then $\pi=1+\sqrt{-1}$ is a uniformizing parameter. The quadratic defect
of $u=1+2\pi=3+2\sqrt{-1}$ is $(\pi)^3$, since $1$ and $3$ are the only squares of units in $\oink/4\oink$.
In Figure 11 we see the branch $S_0(i,j)$ for
$j^2=u$, and different values of $\alpha=i^2$. As usual $S_1=S_0\big(i)$ and
$S_2=S_0\big(j)$. Recall that, as $-1$ is a square in $k$,
the quaternion algebra $\left(\frac{1+\pi,1+\pi}{k}\right)\cong
\left(\frac{1+\pi,\pi}{k}\right)$ splits.
\begin{figure}\[
\includegraphics[width=1\textwidth]{fig6.jpg}
\]
\caption{Some configurations of the branches $S_1=S_0(i)$ and
$S_2=S_0(j)$ in example \ref{ex1}.}
\end{figure}
\end{example}
\begin{example}
The tables can be used to compute the branch of an order of the form $\oink[\eta,\omega]$ for any pair of integral
elements satisfying $k[\eta]=k[i]$ and $k[\omega]=k[j]$, as long as they generate an order.
For example, if $k=\mathbb{Q}_2(\sqrt{-1})$,
to find the invariants of the order generated by $\frac{i+1}{\pi}$ and $\frac{j+1}{\pi}$ when
$i^2=j^2=u$, as in Example 4, we just set $r=s=-1$ in Table 2.
\end{example}
\begin{example}\label{ex2}
The quaternion algebra $\alge=\left(\frac{-3,-3}{\mathbb{Q}}\right)$ ramifies in exactly $2$ places, $3$ and
$\infty$. The order $\Ha=\mathbb{Z}[i,j]\subseteq\alge$, where $ij=-ji$ and $i^2=j^2=-3$
is maximal outside the set $\{2,3\}$.
The local order $\mathbb{Z}_2[i,j]$ is contained in exactly $2$ maximal orders. Since the maximal order in $\alge_3$
is unique, there are exactly $2$ global maximal orders, $\Da$ and $\Da'$, containing $\Ha$. In fact,
$\Da=\enteri[i,\omega,\nu]$ and $\Da'=\enteri[\eta,j,\nu]$, where $j=2\omega+1$, $i=2\eta+1$, and $3\nu=ij$.
It can be proved easily, using the fact that these orders are neighbors at $2$, that all maximal
orders in $\alge$ are conjugate. See \cite{nocubic} for details.
\end{example}
\subsection{Spinor image Computations and representation fields.}
Here we recall the basic facts in the theory of representation fields, see \cite{abelianos} or \cite{cyclic} for details.
The set of spinor genera in a genus $\mathbb{O}=\mathrm{gen}(\Da)$
of Eichler orders in a quaternion algebra $\alge$
over a global field $K$ equals $[\Sigma:K]$, where $\Sigma=\Sigma(\mathbb{O})$, the spinor class field, is the
class field corresponding to the class group $K^*H(\Da)$, where
$$H(\Da)=\{N(a)|a\in\alge_\ad,\ a\Da_\ad a^{-1}=\Da_\ad\},$$
if $N$ denotes the adelic reduced norm, while $\alge_\ad$ and $\Da_\ad$ are the adelizations
of the algebra $\alge$ and the order $\Da$. The field $\Sigma$ can also be described as the largest abelian
extension of $K$ ramifying only at real places that are ramified for $\alge$, and splitting at every
finite place $\wp$ satisfying either of the following conditions:
\begin{itemize}
\item $\alge$ is ramified at $\wp$, or
\item the level of $\Da$ at $\wp$ is odd.
\end{itemize}
The number of spinor genera representing a given order
$\Ha$ equals $[\Sigma:F]$, where $F$ is the representation field, i.e., the
class field corresponding to the group $K^*H(\Da|\Ha)$, where
$$H(\Da|\Ha)=\{N(a)|a\in\alge_\ad,\ a\Ha_\ad a^{-1}\subseteq \Da_\ad\}.$$
For every finite place $\wp$ it is easy to see that the local component $H_\wp(\Da|\Ha)$, which is defined analogously,
is either $K_\wp^*$ or $\oink_\wp^*K_\wp^{*2}$. In fact $H_\wp(\Da|\Ha)=\oink_\wp^*K_\wp^{*2}$
if and only if the level of $\Da$ at $\wp$ is even and equals the local diameter $d(\Ha_\wp)$.
An order $\Ha$ embeds into every spinor genera of maximal orders if and only if $F=K$. When this is not the case,
we say that $\Ha$ is selective \cite{FriedmannQ}, \cite{lino1}.
\begin{example}\label{ex2}
Let $K=\mathbb{Q}(\sqrt{-15})$, whose Hilbert class field is $\Sigma=\mathbb{Q}(\sqrt{-3},\sqrt5)$
\cite[p. 262]{cohn}. In particular, there are exactly $2$ conjugacy classes of maximal orders in $\mathbb{M}_2(K)$.
Note that $3$ and $5$ ramify on $K/\mathbb{Q}$. Let $\mathbf{3}_0$ and $\mathbf{5}_0$ denote the corresponding places of $K$, which are inert on $\Sigma/K$. On the other hand, $2$ splits on $K/\mathbb{Q}$. The two dyadic places
$\mathbf{2}_1$ and $\mathbf{2}_2$ are inert on $\Sigma/K$. Set $\Ha=\mathcal{O}[i,j]$, where $ij=-ji$, while
$i^2=-3$ and $j^2=-15$. Then, according to the tables, the invariants of the local branches of $\Ha$ are as follows:
$$\Big(d(\Ha_{\wp}),p(\Ha_{\wp}),l(\Ha_{\wp})\Big)=\left\{\begin{array}{rcl}
(2,1,0)&\textnormal{ if }&\wp=\mathbf{3}_0\\
(2,0,2)&\textnormal{ if }&\wp=\mathbf{5}_0\\
(1,0,1)&\textnormal{ if }&\wp\in\{\mathbf{2}_1,\mathbf{2}_2\} \\
(0,0,0)&&\textnormal{otherwise}\end{array}\right..$$
It follows that $\Ha$ embeds into an Eichler order of every level dividing $(30)$. Furthermore, since
the Frobenius element at each of the relevant places
$$|[\mathbf{3}_0,\Sigma/K]|=|[\mathbf{5}_0,\Sigma/K]|=|[\mathbf{2}_1,\Sigma/K]|=|[\mathbf{2}_2,\Sigma/K]|$$
is the nontrivial element in $\mathrm{Gal}(\Sigma/K)$, there is a unique conjugacy class of Eichler orders of level $(30)$,
but $2$ conjugacy classes of Eichler orders of level $(15)$ and $\Ha$ embeds in both of them. On the other hand, for the order
$\Ha'=\mathcal{O}[i,2j]$, the local invariants are:
$$\Big(d(\Ha'_{\wp}),p(\Ha'_{\wp}),l(\Ha'_{\wp})\Big)=\left\{\begin{array}{rcl}
(2,1,0)&\textnormal{ if }&\wp\in\{\mathbf{3}_0\mathbf{2}_1,\mathbf{2}_2\}\\
(2,0,2)&\textnormal{ if }&\wp=\mathbf{5}_0\\
(0,0,0)&&\textnormal{otherwise}\end{array}\right.,$$
whence $\Ha'$ is selective on the genus of Eichler orders of level $(60)$.
\end{example}
\subsection{Number of maximal orders containing a given order.} It is not hard to show by a simple inductive argument
that the number of maximal orders in a thick line $S$ with invariants $p(S)$ and $l(S)$ is
$$\aleph(S)=[l(S)+1]q^{p(s)}+2\frac{q^{p(s)}-1}{q-1},$$ where $q$ is the number of elements in the residual field.
It follows that the number of maximal orders containing an order $\Ha$ of maximal rank, in a quaternion algebra over a global field $F$, is
$\prod_{\wp}\aleph\big(S_0(\Ha_\wp)\big)$, where the product is taken over all finite places of $F$.
Note that almost every factor in this product is one.
An important case, which is straightforward from Proposition \ref{p44} is next result:
\begin{proposition}
If $F/\mathbb{Q}$ is unramified at $2$, and $a,b\in\mathcal{O}_F$ are square free, the number of maximal orders in
an $F$-algebra $\alge\cong\left(\frac{a,b}{F}\right)$ containing the order $\Ha=\mathcal{O}_F[i,j]$,
where $ij=-ji$, $i^2=a$, and $j^2=b$, is $2^T$, where $T$
is the number of un-ramified places dividing $2ab$.
\end{proposition}
\subsection{Set of fixed points for groups of Moebius transformations.}
For any non-archimedean local field $k$, the group of Moebius transformations $\mathrm{PGL}_2(k)$
acts on the vertices of
the BT-tree by conjugation. The stabilizer of a maximal order $\Da$ is the group $k^*\Da^*$. Note that an element
$u$ is a unit in some maximal order if and only if its norm is a unit and its trace is an integer, and in this case it is a unit in every order containing it. It follows that the class $\bar u\in \mathrm{PGL}_2(k)$ of a unit $u\in\mathrm{GL}_2(k)=\matrici_2(k)^*$ stabilizes every element in $S_0(u)$. On the other hand, if
$\bar u\in \mathrm{PGL}_2(k)$ is not the class of a unit, then $\bar u$ cannot leave any vertex invariant.
In particular, if a subgroup of $\mathrm{PGL}_2(k)$ has a non-empty invariant locus in the BT-tree, then this locus is a set in the family that we described in \S1. The results in this work allow us to compute this invariant locus for isomorphic copies of the Klein group inside $\mathrm{PGL}_2(k)$.
\subsection{Generating sets of orders.}
It is apparent from Propositions 2.1-2.4 that whenever the intersection of two branches has a stem of length 2 or higher,
this stem is contained in the stem of one of the intersecting branches. As a consequence, we conclude that, if $\Ha=\mathcal{O}[a_1,\dots,a_n]$ is an order whose branch has a stem of length 2 or higher, then at least one of the generators $a_i$ spans
an algebra isomorphic to $k\times k$. The converse is, however, not true since the conclusion holds for an Eichler order
$\Da$ of level $1$. In fact, $\Da/\pi\Da\cong(\mathbb{F}_k\times\mathbb{F}_k)\oplus R$, where $R$ is a radical and $\mathbb{F}_k$ is the residue field of $k$. Note that any generating set of this algebra must contain an element whose projection to $\mathbb{F}_k\times\mathbb{F}_k$ is a generator of the latter algebra.
\section{Acknowledgements}
The first author was supported by Fondecyt, Grant number 1140533. The second author was partly supported by
Fondecyt, Grant number 1100127.
|
\section{Introduction}
A characteristic feature in standard Hopf algebra theory is its self-duality, that is, the dual of a (finite-dimensional) Hopf algebra (over a field) is a Hopf algebra again. In particular, the antipode of this dual is nothing but the transpose of the original antipode; see, for example, \cite{Swe:HA}.
In the broader setup of ({\em left} or {\em full}) Hopf algebroids over possibly noncommutative rings, this peculiar property appears to be more intricate; see \cite{Boe:HA} or \S\ref{tranquilli} for the precise definitions of these objects, we only mention here that, in contrast to full Hopf algebroids, there is no notion of antipode for left Hopf algebroids:
one rather considers the inverse of a certain Hopf-Galois map and its associated {\em translation map}.
Nevertheless, left Hopf algebroids appear as the correct generalisation of Hopf algebras over noncommutative rings, whereas full Hopf algebroids generalise Hopf algebras twisted by a character, see, for example, \cite[\S4.1.2]{Kow:HAATCT}.
Recently (after the first posting of this article), Schauenburg \cite{Schau:TDATDOAHAAHA} showed that the (skew) dual of a left Hopf algebroid (under a suitable finiteness assumption) carries some Hopf structure as well without giving an explicit expression for the inverse of the respective Hopf-Galois map or the associated translation map.
However,
instead of one dual, a left bialgebroid $ U $ rather possesses {\em two}, the {\em right dual\/} $ U^{\scalast} $ and the {\em left dual\/} $ U_{\scalastd} $, which, on top, live in a different category compared to $U$ as they are both (under certain finiteness assumptions) right bialgebroids \cite{KadSzl:BAODTEAD}.
There is no reason why one should prefer one of the duals to the other. Hence, any question concerning ``the dual of $ U $'' should be converted into a question about the pair $ ( U^{\scalast} , U_{\scalastd} ) $.
Dealing with {\em full} Hopf algebroids (see \S\ref{reminders_H-ads}) does notably worsen the situation as there are actually {\em four} duals to be taken into account, two of which are left and two of which are right bialgebroids. In this case, an answer to the question of the nature of the Hopf structure on the dual(s) has only been given in certain cases, more precisely, in the presence of integrals \cite[\S5]{BoeSzl:HAWBAAIAD}.
\subsection{Aims and objectives}
\label{aims,appunto}
As mentioned a moment ago, the object one should investigate to discover the limits of self-duality in (left) Hopf algebroid theory is a {\em pair} of duals. In short, our question reads as follows: if a left bialgebroid $ U $ is, in particular, a left (or right) Hopf algebroid, what extra structure can be found on the pair
$ ( U^{\scalast} , U_{\scalastd} ) $ of duals?
\subsection{Main results}
After highlighting in \S\ref{acquaprimavera} a multitude of module structures that exist on $\Hom$-spaces and tensor products in presence of a left or right Hopf algebroid structure and that will be used in the sequel, in \S\ref{fallouts} we review (and extend) Ph\`ung's equivalence of comodule categories (see the main text for all definitions and conventions used hereafter):
{\renewcommand{\thetheorem}{{A}}
\begin{theorem}
Let $(U,A)$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item
Let $(U,A)$ be additionally a left Hopf algebroid such that $U_{\ract}$ is projective.
Then there exists a (strict) monoidal functor $ \comodu \to \ucomod $: if $M$ is a right $U$-comodule with coaction $m \mapsto m_{(0)} \otimes_\ahha m_{(1)}$, then
$$M \to U_\ract \otimes_\ahha M, \quad m \mapsto m_{(1)-} \otimes_\ahha m_{(0)} \epsilon(m_{(1)+}),$$
defines a left comodule structure on $M$ over $U$.
\item
Let $(U,A)$ be a right Hopf algebroid such that $_{\lact}U$ is projective.
Then there exists a (strict) monoidal functor $ \ucomod \to \comodu $: if $N$ is a left $U$-comodule with coaction $n \mapsto n_{(-1)} \otimes_\ahha n_{(0)}$, then
$$N \to N \otimes_\ahha \due U \lact {}, \quad n \mapsto \epsilon(n_{(-1)[+]}) n_{(0)} \otimes_\ahha n_{(-1)[-]},$$
defines a right comodule structure on $N$ over $U$.
\item
If $U$ is both a left and right Hopf algebroid and if both $U_{\ract}$ and $_{\lact}U$ are $A$-projective,
then the functors mentioned in {\it (i)} and {\it (ii)} are quasi-inverse to each other and we have an equivalence
$$
\ucomod \simeq \comodu$$
of monoidal categories.
\end{enumerate}
\end{theorem}
}
Note that this equivalence works without the help of an antipode as there are objects that are both left and right Hopf algebroids but not full Hopf algebroids (cocommutative left Hopf algebroids, for example).
Starting from this result, under suitable finiteness hypotheses on $U$, one can construct functors $\mathbf{Mod}\mbox{-}{U_{\scalastd}} \to \mathbf{Mod}\mbox{-}{U^{\scalastd}}$ resp.\ $\mathbf{Mod}\mbox{-}{U^{\scalastd}} \to \mathbf{Mod}\mbox{-}{U_{\scalastd}}$, and from this we isolate maps $U^{\scalastd} \to U_{\scalastd}$ resp.\ $U_{\scalastd} \to U^{\scalastd}$, which even make sense without any finiteness assumptions as proven in \S\ref{linking-structures},
and which are our main object of interest.
In \S\ref{linking} we can then give the following answer to the problem mentioned in \S\ref{aims,appunto}, that is, elucidate the relation between the left and the right dual:
{\renewcommand{\thetheorem}{{B}}
\begin{theorem}
\label{acquadinepi}
Let $(U,A)$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item
If $(U,A) $ is moreover a left Hopf algebroid, there is a morphism $ S^{\scalast} : U^{\scalast} \to U_{\scalastd} $
of $ A^e $-rings with augmentation; if, in addition, both $ \due U \lact {} $ and $ U_\ract $ are finitely generated $A$-projective, then $ (S^{\scalast}, \id_\ahha) $ is a morphism of right bialgebroids.
\item
If $(U,A) $ is a right Hopf algebroid instead, there is a morphism
$ S_{\scalastd} : U_{\scalastd} \to U^{\scalast} $
of $ A^e $-rings with augmentation; if, in addition, both $ \due U \lact {} $ and $ U_\ract $ are finitely generated $A$-projective, then $ (S_{\scalast}, \id_\ahha) $ is a morphism of right bialgebroids.
\item
If $(U,A)$ is simultaneously both a left and a right Hopf algebroid, then the two morphisms are inverse to each other; hence, if both $ \due U \lact {} $ and $ U_\ract $ are finitely generated $A$-projective, then
$U^\scalast\simeq U_{\scalastd}$ as right bialgebroids.
\end{enumerate}
\end{theorem}
}
Now, as said before, for a left Hopf algebroid (which is finitely generated projective with respect to both source and target map) there is no canonical choice for which dual to consider but in view of Theorem \ref{acquadinepi}, in case the left Hopf algebroid is simultaneously a right Hopf algebroid, both duals are isomorphic and hence can be seen as {\em its} dual, which carries a Hopf structure by Schauenburg's recent result \cite{Schau:TDATDOAHAAHA}. This seems to be as close as one can get to self-duality.
Theorem \ref{acquadinepi} is a straight analogue of the construction on the dual for a (finite-dimensional) Hopf algebra $H$ (over a field) with antipode $S$ in the following sense: here, one has $ H^{\scalast} = (H_{\scalastd})^\op_\coop $
and $ S^{\scalast} $ is exactly the transpose of $S$ and therefore the antipode for the dual Hopf algebra.
Observe that this last case in Theorem \ref{acquadinepi}, {\em i.e.}, the presence of both a left and right Hopf structure is given, for example, when $ U $ is a full Hopf algebroid with bijective antipode but also in weaker cases such as for the universal enveloping algebra of a Lie-Rinehart algebra. In the situation of a full Hopf algebroid, $U^{\scalast}$ and $U_{\scalastd}$ are additionally linked (in both directions) by the transposition $\qttr S t {}{}{}$ of the antipode $ S : U \to U^\op_\coop $. However, in Theorem \ref{movimentoallaforza} we show that the map $\qttr S t {}{}{}$ in general does not coincide with $S^{\scalast}$ or $S_{\scalastd}$, in contrast to the Hopf algebra case mentioned above.
Moreover, if a left Hopf algebroid $ U $ is cocommutative with both $ \due U \lact {} $ and $ U_\ract $ finitely generated $A$-projective,
then $ U^{\scalast} = (U_{\scalastd})_\coop $ is a full Hopf algebroid (with antipode precisely given by $S^{\scalast}$), though $ U $ might be not.
We shall also see in \S\ref{lautesGequatsche} that Theorem \ref{acquadinepi} actually extends to a larger setup, in particular, it applies to some interesting cases (coming from geometry), where neither $ \due U \lact {} $ nor $ U_\ract $ are finitely generated projective but $U^\scalast$ and $U_\scalast$ are still right bialgebroids in a suitable (topological) sense, such as when $ U $ is the universal enveloping of a Lie-Rinehart algebra, or a quantisation of it.
In \S\ref{lautesGequatsche},
we illustrate these results by considering some examples related to Lie-Rinehart algebras (or Lie algebroids) and their jet spaces, as well as their quantised versions.
Moreover, in \S\ref{heisz} we consider further duality phenomena related to dualising modules, which appear in Poincar\'e duality, along with their quantisations.
\bigskip
\thanks{ {\bf Acknowledgements.}
We would like to thank the referee for useful comments and suggestions, and in particular for pointing out the article \cite{Phu:TKDFHA} to us.
We also would like to thank G.\ B{\"o}hm and L.\ El Kaoutit for their valuable comments and remarks.
S.C.\ acknowledges travel support and hospitality from both INdAM and Universit\`a di Roma Tor Vergata, where part of this work was achieved.
The research of F.G.\ was partially funded by the PRIN 2012
project {\em Moduli spaces and Lie theory}.
The research of N.K.\ was funded by an INdAM-COFUND Marie Curie grant and also supported by UniNA and Compagnia di San Paolo in the framework of the program STAR 2013.
\section{Preliminaries}
\label{tranquilli}
We list here those preliminaries with respect to bialgebroids and their duals that are needed in this article; see, {\em e.g.}, \cite{Boe:HA} and references therein for an overview on this subject.
Fix an (associative, unital, commutative) ground ring $ k $. Unadorned tensor products will always be meant over $ k $.
All other algebras, modules etc.~will
have an underlying structure of a
$k$-module. Secondly, fix an associative and unital
$k$-algebra $A$, {\em i.e.}, a ring with a
ring homomorphism
$ \eta_\ahha : k \rightarrow Z(A)$ to
its centre. Denote by
$A^\mathrm{op}$ the
opposite and by
$A^\mathrm{e} := A \otimes A^\mathrm{op}$
the enveloping algebra
of $A$, and by
$\amod$ the category of
left $A$-modules.
Recall that an {\em $ A $-ring\/} is a monoid in the monoidal category $(\amoda, \otimes_\ahha, A)$ of $(A,A)$-bimodules fulfilling the usual associativity and unitality axioms, whereas dually an {\em $A$-coring\/} is a comonoid in this category that is coassociative and counital.
\subsection{Bialgebroids}
\label{h-Hopf_algbds}
For an $ A^e $-ring $ U $ given by the $k$-algebra map
$\eta: \Ae \to U $, consider the restrictions
$ s := \eta( - \otimes 1_\uhhu) $ and $t:= \eta(1_\uhhu \otimes -)$, called {\it source\/} and {\it target\/} map, respectively.
Thus an $ A^e $-ring $ U $ carries two $ A $-module structures from the left and two from the right, namely
$$
a \lact u \ract b := s(a) t(b) u, \quad \quad a \blact u \bract b := u t(a) s(b), \eqno \forall \; a, b \in A , u \in U.
$$
If we let $ U_\ract {\otimes_{\scriptscriptstyle A}} {}_\lact U $ be the corresponding tensor product of $ U $ (as an $ A^e $-module) with itself, we define the {\it (left) Takeuchi-Sweedler product\/} as
$$
U_\ract \! \times_\ahha \! {}_\lact U \; := \;
\big\{ {\textstyle \sum_i} u_i \otimes u'_i \in U_\ract \! \otimes_{\scriptscriptstyle A} \! {}_\lact U \mid {\textstyle \sum_i} (a \blact u_i) \otimes u'_i = {\textstyle \sum_i} u_i \otimes (u'_i \bract a), \ \forall a \in A \big\}.
$$
By construction, $ U_\ract \! \times_{\scriptscriptstyle A} \! {}_\lact U $ is an $ \Ae $-submodule of $ U_\ract \! \otimes_{\scriptscriptstyle A} \! {}_\lact U $; it is also an $ A^e $-ring via factorwise multiplication, with unit $ 1_\uhhu \otimes 1_\uhhu $ and $ \eta_{{}_{U_\ract \times_{\scriptscriptstyle A} {}_\lact U}}(a \otimes \tilde{a}) := s(a) \otimes t(\tilde{a})$.
Symmetrically, one can consider the tensor product
$ U_\bract \otimes_\ahha \due U \blact {} $ and define the {\em (right) Takeuchi-Sweedler product\/} as
$U_\bract \times_\ahha \due U \blact {} $, which is an $\Ae $-ring inside
$ U_\bract \otimes_\ahha \due U \blact {} $.
\begin{definition}
A {\em left bialgebroid} $(U,A)$ is a $ k $-module $ U $ with the structure of an
$ \Ae $-ring $(U, s^\ell, t^\ell)$ and an $ A $-coring $(U, \gD_\ell, \eps)$ subject to the following compatibility relations:
\begin{enumerate}
\item
the $ \Ae $-module structure on the $ A $-coring $ U $ is that of
$ \due U \lact \ract $;
\item
the coproduct $ \Delta_\ell $ is a unital $ k $-algebra morphism taking values in $ U {}_\ract \! \times_{\scriptscriptstyle A} \! {}_\lact U $;
\item
for all $ a, b \in A $, $ u, u' \in U $, one has:
\begin{equation}
\label{castelnuovo}
\epsilon( a \lact u \ract b) = a \epsilon(u) b, \quad \epsilon(uu') = \epsilon \big( u \bract \epsilon(u')\big) = \epsilon \big(\epsilon(u') \blact u\big).
\end{equation}
\end{enumerate}
A {\it morphism\/} between left bialgebroids $(U, A)$ and $(U',A')$
is a pair $(F, f)$ of maps $F: U \to U'$, $f:A \to A'$ that commute with all structure maps in an obvious way.
\end{definition}
As for any ring, we can define the categories $\umod$ and $\modu$ of left and right modules over $U$. Note that $\umod$ forms a monoidal category but $\modu$ usually does not. However, in both cases there is a forgetful functor $\umod \to \amoda$, resp.\ $\modu \to \amoda$: whereas we denote left and right action of a bialgebroid $U$ on $M \in \umod$ or $N \in \modu$ usually by juxtaposition, for the resulting $\Ae$-module structures the notation
$$
a \lact m \ract b := s^\ell(a)t^\ell(b)m, \qquad a \blact m \bract b := ns^\ell(b)t^\ell(a)
$$
for $m \in M, \ n \in N, \ a,b \in A$ is used instead. For example, the base algebra $A$ itself is a left $U$-module via the left action
$u(a) := \epsilon( u \bract a) = \epsilon( a \blact u )$ for $ u \in U $ and $ a \in A $, but in most cases there is no right $U$-action on $A$.
Dually, one can introduce the categories $\ucomod$ and $\comodu$ of left resp.\ right $U$-comodules, both of which are monoidal; here again, one has forgetful functors $\ucomod \to \amoda$ and $\comodu \to \amoda$.
More precisely (see, {\em e.g.}, \cite{Boe:HA}), a (say) right comodule
is a right comodule of the coring underlying $U$, {\em i.e.}, a right $A$-module $M$ and a right $A$-module map
$
\due {\gD} {\emme\!} {}: M \rightarrow
M \otimes_\ahha {}_\lact U, \
m \mapsto m_{(0)} \otimes_\ahha m_{(1)},
$
satisfying the usual coassociativity and counitality axioms.
On any $M \in \comodu$ there is an induced {\em left} $A$-action given by
\begin{equation}
\label{Inducedaction}
am := m_{(0)}\gve(a \blact m_{(1)}),
\end{equation}
and $\due {\gD} {\emme\!} {}$ is then an $\Ae$-module morphism
$
M \rightarrow
M \times_\ahha {}_\lact U,
$
where $ M \times_\ahha {}_\lact U $ is the $\Ae$-submodule of
$ M \otimes_\ahha {}_\lact U$ whose elements $\sum_i m_i
\otimes_\ahha u_i$ fulfil
\begin{equation}
\label{Takeuchicoaction}
\textstyle
\sum_i a m_i \otimes_\ahha u_i
= \sum_i m_i \otimes_\ahha u_i \bract a, \
\forall a \in A.
\end{equation}
\medskip
The notion of a {\em right bialgebroid\/} is obtained if one starts with the $\Ae$-module structure given by $ \blact $ and $ \bract $ instead of $\lact$ and $\ract$. We will refrain from giving the details here and refer to \cite{KadSzl:BAODTEAD} instead.
\begin{remark}
The {\em opposite\/} of a left bialgebroid $( U, A, s^\ell, t^\ell, \Delta_\ell, \epsilon ) $ yields a {\em right\/} bialgebroid
$( U^\op, A, t^\ell, s^\ell, \Delta_\ell, \epsilon )$.
The {\em coopposite\/} of a left bialgebroid is the {\em left\/} bialgebroid given by $( U, A^\op, t^\ell, s^\ell, \Delta_\ell^\coop, \epsilon) $.
\end{remark}
\subsection{Pairings of $ U $-modules and dual bialgebroids}
\label{mayitbe}
Let $ (U,A) $ be a left bialgebroid, $M, M' \in \umod$ be left $U$-modules, and $N,N' \in \modu$ be right $U$-modules. Define
\begin{equation*}
\begin{array}{rclrcl}
\Hom_{\Aopp} (M , M') \!\!\!&:=\!\!\!& \Hom_{\Aopp} ( M_\ract , M'_\ract ), &
\Hom_\ahha( M , M') \!\!\!&:=\!\!\!& \Hom_\ahha (\due M \lact {}, \due {M'} \lact {}), \\
\Hom_{\Aopp} (N , N') \!\!\!&:=\!\!\!& \Hom_{\Aopp} ( N_{\bract} , N'_{\bract}), &
\Hom_\ahha(N , N') \!\!\!&:=\!\!\!& \Hom_\ahha(\due N \blact {}, \due {N'} \blact {}).
\end{array}
\end{equation*}
In particular, for $ M': = A $ we set $ M_{\scalastd} := \Hom_\ahha( M , A ) $ and $ M^{\scalast} := \Hom_{\Aopp}( M , A) $, called, respectively, the {\em left\/} and {\it right\/} dual of $ M $.
The notion of {\em pairing\/} between $\Ae$-bimodules is also useful (see, for instance, \cite{CheGav:DFFQG}):
\begin{definition}
\label{left and right 1}
Let $ U $ and $ W $ be two $ \Ae $-bimodules.
\begin{enumerate}
\compactlist{99}
\item
A {\em left $ \Ae $-pairing\/} is a $ k $-bilinear map $\langle\ ,\ \rangle : U \times W \to A $ such that for any $ u \in U $, $ w \in W $, and $ a \in A $, one has
\begin{equation*}
\begin{array}{rclrclrcl}
\langle u , a \lact w \rangle \!\!\!&=\!\!\!&
\langle u \ract a , w \rangle, &
\langle u , w \ract a \rangle \!\!\!&=\!\!\!&
\langle a \blact u , w \rangle, &
\langle u , a \blact w \rangle \!\!\!&=\!\!\!&
\langle u \bract a , w \rangle, \\
\langle u , w \bract a \rangle \!\!\!&=\!\!\!&
\langle u , w \rangle a, &
\langle a \lact u , w \rangle \!\!\!&=\!\!\!&
a \langle u , w \rangle.
\end{array}
\end{equation*}
\item
A {\em right $ \Ae $-pairing\/} is a $ k $-bilinear map
$\langle\ ,\ \rangle : U \times W \to A $ such that for any $ u \in U $, $ w \in W $, and $ a \in A $, one has
\begin{equation*}
\begin{array}{rclrclrcl}
\langle u , w \ract a \rangle \!\!\!&=\!\!\!&
\langle a \lact u , w \rangle, &
\langle u , a \lact w \rangle \!\!\!&=\!\!\!&
\langle u \bract a , w \rangle, &
\langle u , w \bract a \rangle \!\!\!&=\!\!\!&
\langle a \blact u , w \rangle, \\
\langle u , a \blact w \rangle \!\!\!&=\!\!\!&
a \langle u , w \rangle, &
\langle u \ract a , w \rangle \!\!\!&=\!\!\!&
\langle u , w \rangle a.
\end{array}
\end{equation*}
\end{enumerate}
\end{definition}
\begin{free text}{\bf Duals of bialgebroids.}
\label{regnetswirklich?}
Let $ U_{\scalastd} $ resp.\ $ U^{\scalast} $ be the left resp.\ right dual of a left bialgebroid. If $ \due U \lact {} $ is finitely generated projective, then $ U_{\scalastd} $ is canonically endowed with a {\em right} bialgebroid structure \cite{KadSzl:BAODTEAD} such that the evaluation pairing between $ U $ and $ U_{\scalastd} $ is a (nondegenerate) {\em left\/} pairing; similarly, if $U_\ract$ is finitely generated projective, then $ U^{\scalast} $ has a canonical {\em right} bialgebroid structure for which the natural pairing between $ U $ and $ U^\scalast $ is a {\em right\/} pairing. If instead in either case the above finitely generated projective assumption is not satisfied, then both $U^\scalast$ and $ U_{\scalastd} $ are nevertheless $ \Ae $-rings endowed with a ``counit'' map, or augmentation.
\end{free text}
\subsection{Left and right Hopf algebroids}
\label{goeseveron}
\label{half-Hopf_algebroids}
For any left bialgebroid $ U $, define the {\em Hopf-Galois maps}
\begin{equation*}
\begin{array}{rclrcl}
\ga_\ell : \due U \blact {} \otimes_{\Aopp} U_\ract &\to& U_\ract \otimes_\ahha \due U \lact,
& u \otimes_\Aopp v &\mapsto& u_{(1)} \otimes_\ahha u_{(2)} v, \\
\ga_r : U_{\!\bract} \otimes^\ahha \! \due U \lact {} &\to& U_{\!\ract} \otimes_\ahha \due U \lact,
& u \otimes^\ahha v &\mapsto& u_{(1)} v \otimes_\ahha u_{(2)}.
\end{array}
\end{equation*}
With the help of these maps, we make the following definition due to Schauenburg \cite{Schau:DADOQGHA}:
\begin{definition}
\label{def Half Hopf bialgebroids}
A left bialgebroid $U$ is called a {\em left Hopf algebroid} if $ \alpha_\ell $ is a bijection. Likewise, it is called a {\em right Hopf algebroid} if $\ga_r$ is so.
In either case, we adopt for all $u \in U$ the following (Sweedler-like) notation
\begin{equation}
\label{latoconvalida}
u_+ \otimes_\Aopp u_- := \alpha_\ell^{-1}(u \otimes_\ahha 1), \qqquad
u_{[+]} \otimes^\ahha u_{[-]} := \alpha_r^{-1}(1 \otimes_\ahha u),
\end{equation}
and call both maps $ u \mapsto u_+ \otimes_\Aopp u_- $ and $ u \mapsto u_{[+]} \otimes^\ahha u_{[-]} $ {\em translation maps}.
\end{definition}
Analogous notions exist with respect to an underlying {\em right} bialgebroid structure, but we will not give the details here.
\begin{remark}
\label{left/right-Hopf-left-bialg_cocomm}
\noindent
\begin{enumerate}
\compactlist{99}
\item
In case $A=k$ is central in $U$, one can show that $\ga_\ell$ is invertible if and only if $U$ is a Hopf algebra, and the translation map reads
$ u_+ \otimes u_- := u_{(1)} \otimes S(u_{(2)}) $, where $S$ is the antipode of $U$.
On the other hand, $U$ is a Hopf algebra with invertible antipode if and only if both $\ga_\ell$ and $\ga_r$ are invertible,
and then $ u_{[+]} \otimes u_{[-]} := u_{(2)} \otimes S^{-1}(u_{(1)}) $.
\item
The underlying left bialgebroid in a {\em full} Hopf algebroid with bijective antipode is both a left and right Hopf algebroid (but not necessarily vice versa); see \cite[Prop.\ 4.2]{BoeSzl:HAWBAAIAD} for the details of this construction.
\end{enumerate}
\end{remark}
The following proposition collects some properties of the translation maps \cite{Schau:DADOQGHA}:
\begin{prop}
Let $U$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item If $ U $ is a left Hopf algebroid, the following relations hold:
\begin{eqnarray}
\label{Sch1}
u_+ \otimes_\Aopp u_- & \in
& U \times_\Aopp U, \\
\label{Sch2}
u_{+(1)} \otimes_\ahha u_{+(2)} u_- &=& u \otimes_\ahha 1 \quad \in U_{\!\ract} \! \otimes_\ahha \! {}_\lact U, \\
\label{Sch3}
u_{(1)+} \otimes_\Aopp u_{(1)-} u_{(2)} &=& u \otimes_\Aopp 1 \quad \in {}_\blact U \! \otimes_\Aopp \! U_\ract, \\
\label{Sch4}
u_{+(1)} \otimes_\ahha u_{+(2)} \otimes_\Aopp u_{-} &=& u_{(1)} \otimes_\ahha u_{(2)+} \otimes_\Aopp u_{(2)-}, \\
\label{Sch5}
u_+ \otimes_\Aopp u_{-(1)} \otimes_\ahha u_{-(2)} &=&
u_{++} \otimes_\Aopp u_- \otimes_\ahha u_{+-}, \\
\label{Sch6}
(uv)_+ \otimes_\Aopp (uv)_- &=& u_+v_+ \otimes_\Aopp v_-u_-,
\\
\label{Sch7}
u_+u_- &=& s^\ell (\varepsilon (u)), \\
\label{Sch8}
\varepsilon(u_-) \blact u_+ &=& u, \\
\label{Sch9}
(s^\ell (a) t^\ell (b))_+ \otimes_\Aopp (s^\ell (a) t^\ell (b) )_-
&=& s^\ell (a) \otimes_\Aopp s^\ell (b),
\end{eqnarray}
where in \rmref{Sch1} we mean the Takeuchi-Sweedler product
\begin{equation*}
\label{petrarca}
U \! \times_\Aopp \! U :=
\big\{ {\textstyle \sum_i} u_i \otimes v_i \in {}_\blact U \otimes_\Aopp U_{\!\ract} \mid {\textstyle \sum_i} u_i \ract a \otimes v_i = {\textstyle \sum_i} u_i \otimes a \blact v_i, \ \forall a \in A \big\}.
\end{equation*}
\item
Analogously, if $ U $ is a right Hopf algebroid, one has:
\begin{eqnarray}
\label{Tch1}
u_{[+]} \otimes^\ahha u_{[-]} & \in
& U \times^{\scriptscriptstyle A} U, \\
\label{Tch2}
u_{[+](1)} u_{[-]} \otimes_\ahha u_{[+](2)} &=& 1 \otimes_\ahha u \quad \in U_{\!\ract} \! \otimes_\ahha \! {}_\lact U, \\
\label{Tch3}
u_{(2)[-]}u_{(1)} \otimes^\ahha u_{(2)[+]} &=& 1 \otimes^\ahha u \quad \in U_{\!\bract} \!
\otimes^\ahha \! \due U \lact {}, \\
\label{Tch4}
u_{[+](1)} \otimes^\ahha u_{[-]} \otimes_\ahha u_{[+](2)} &=& u_{(1)[+]} \otimes^\ahha
u_{(1)[-]} \otimes_\ahha u_{(2)}, \\
\label{Tch5}
u_{[+][+]} \otimes^\ahha u_{[+][-]} \otimes_\ahha u_{[-]} &=&
u_{[+]} \otimes^\ahha u_{[-](1)} \otimes_\ahha u_{[-](2)}, \\
\label{Tch6}
(uv)_{[+]} \otimes^\ahha (uv)_{[-]} &=& u_{[+]}v_{[+]}
\otimes^\ahha v_{[-]}u_{[-]}, \\
\label{Tch7}
u_{[+]}u_{[-]} &=& t^\ell (\varepsilon (u)), \\
\label{Tch8}
u_{[+]} \bract \varepsilon(u_{[-]}) &=& u, \\
\label{Tch9}
(s^\ell (a) t^\ell (b))_{[+]} \otimes^\ahha (s^\ell (a) t^\ell (b) )_{[-]}
&=& t^\ell(b) \otimes^\ahha t^\ell(a),
\end{eqnarray}
where in \rmref{Tch1} we mean the Sweedler-Takeuchi product
\begin{equation*} \label{petrarca2}
U \times^{\scriptscriptstyle A} U :=
\big\{ {\textstyle \sum_i} u_i \otimes v_i \in U_{\!\bract} \otimes^\ahha \! \due U \lact {} \mid {\textstyle \sum_i} a \lact u_i \otimes v_i = {\textstyle \sum_i} u_i \otimes v_i \bract a, \ \forall a \in A \big\}.
\end{equation*}
\end{enumerate}
\end{prop}
These two structures are not entirely independent:
\begin{lemma}
The following mixed relations
hold among left and right translation maps:
\begin{eqnarray}
\label{mampf1}
u_{+[+]} \otimes_\Aopp u_{-} \otimes^\ahha u_{+[-]} &=& u_{[+]+} \otimes_\Aopp u_{[+]-} \otimes^\ahha u_{[-]}, \\
\label{mampf2}
u_+ \otimes_\Aopp u_{-[+]} \otimes^\ahha u_{-[-]} &=& u_{(1)+} \otimes_\Aopp u_{(1)-} \otimes^\ahha u_{(2)}, \\
\label{mampf3}
u_{[+]} \otimes^\ahha u_{[-]+} \otimes_\Aopp u_{[-]-} &=& u_{(2)[+]} \otimes^\ahha u_{(2)[-]} \otimes_\Aopp u_{(1)},
\end{eqnarray}
where, for example, in the first equation \rmref{mampf1}
the second tensor product relates the first component with the third, and {\em mutatis mutandis} for the other identities.
\end{lemma}
\begin{proof}
In order to prove \rmref{mampf1}, we apply $ \ga_\ell \otimes \id $ to both sides (note that this operation is well-defined on the considered tensor products); for the right hand side we obtain, by definition,
$$
(\ga_\ell \otimes \id)( u_{[+]+} \otimes_\Aopp u_{[+]-} \otimes^\ahha u_{[-]} )
= (u_{[+]} \otimes_\ahha 1) \otimes^\ahha u_{[-]},
$$
and for the left hand side we have
\begin{equation*}
\begin{split} & (\ga_\ell \otimes \id)(u_{+[+]} \otimes_\Aopp u_{-} \otimes^\ahha u_{+[-]}) = (u_{+[+](1)}
\otimes_\ahha u_{+[+](2)}u_-) \otimes^\ahha u_{+[-]} \\
= \ & (u_{+(1)[+]} \otimes_\ahha u_{+(2)}u_-) \otimes^\ahha u_{+(1)[-]}
= (u_{[+]} \otimes_\ahha 1) \otimes^\ahha u_{[-]},
\end{split}
\end{equation*}
using \rmref{Tch4} and \rmref{Sch2}.
Since $ \ga_\ell $ is assumed to be an isomorphism, this proves \rmref{mampf1}.
Let us also prove \rmref{mampf2}; the remaining identity will be left to the reader. To this end, apply $ \id \otimes \ga_r $ to both sides in \rmref{mampf2}: for the left hand side, we obtain
\begin{equation*}
\begin{split}
(\id \otimes \ga_r)(u_{+} \otimes_\Aopp u_{-[+]} \otimes^\ahha u_{-[-]}) &=
u_{+} \otimes_\Aopp ( u_{-[+](1)}u_{-[-]} \otimes_\ahha u_{-[+](2)} ) \\
& = u_+ \otimes_\Aopp (1 \otimes_\ahha u_-)
\end{split}
\end{equation*}
by \rmref{Tch2}, and where in the second equation the first tensor product relates the first component with the third. As for the right hand side, we compute:
\begin{equation*}
\begin{split}
& (\id \otimes \ga_r)(u_{(1)+} \otimes_\Aopp u_{(1)-} \otimes^\ahha u_{(2)}) = u_{(1)+} \otimes_\Aopp (u_{(1)-(1)} u_{(2)} \otimes_\ahha u_{(1)-(2)}) \\
= \ & u_{(1)++} \otimes_\Aopp (u_{(1)-}u_{(2)} \otimes_\ahha u_{(1)+-})
= u_{+} \otimes_\Aopp (1 \otimes_\ahha u_{-}),
\end{split}
\end{equation*}
using \rmref{Sch5} and \rmref{Sch3} in the last step as follows: Eq.~\rmref{Sch3} yields
$ u_{(1)+} \otimes_\Aopp u_{(1)-} u_{(2)} \otimes_\ahha 1 = u \otimes_\Aopp 1 \otimes_\ahha 1 $
and applying $ \alpha_\ell^{-1} $ to the first and the third component gives the required equality.
\end{proof}
\section{Modules over left or right Hopf algebroids}
\label{acquaprimavera}
In this section we collect some general results about modules over left and right Hopf algebroids. Some of them are known, while others seem to have passed unnoticed so far (see Note \ref{osmanstoechter} below).
\subsection{Module structures on $\Hom$-spaces and tensor products} \label{exotic_U-actions}
Similarly as for bialgebras,
the tensor product $M_\ract \otimes_\ahha \due {M'} \lact {}$ of two left $U$-modules with left $U$-module structure given by
\begin{equation}
\label{baratti&milano}
u(m \otimes_\ahha m') := u_{(1)} m \otimes_\ahha u_{(2)}m'
\end{equation}
equips the category $\umod$ for a left bialgebroid $U$ with a monoidal structure.
On the other hand, for $M \in \umod$ and $N \in \modu$, the $\Ae$-module $\Hom_\Aopp(M_\ract, N_\bract)$ is a right $U$-module via
$$
(fu)(m) := f(u_{(1)}m)u_{(2)}.
$$
The existence of a translation map if $U$ is, on top, a left or right Hopf algebroid makes it possible to endow $\Hom$-spaces and tensor products of $U$-modules with further natural $U$-module structures. The proof of the following proposition is straightforward.
\begin{proposition}
\label{structures}
Let $(U, A)$ be a left bialgebroid, $M, M' \in \umod$ and $N, N' \in \modu$ be left resp.\ right $U$-modules, denoting the respective actions by juxtaposition.
\begin{enumerate}
\compactlist{99}
\item
Let $(U,A)$ be additionally a left Hopf algebroid.
\begin{enumerate}
\compactlist{99}
\item
The $\Ae$-module $\Hom_{\Aopp}(M,M') $ carries a left $ U $-module structure given by
\begin{equation}
\label{gianduiotto1}
(uf)(m) := u_+ \big( f(u_-m) \big).
\end{equation}
In particular, $M^{\scalast} $ is endowed with a left $ U $-module structure.
\item
The $\Ae$-module $\Hom_\ahha(N,N') $ carries a left $U$-module structure via
\begin{equation}
\label{lingotto1}
(u \rightslice f)(n) := \big( f(nu_+) \big)u_-.
\end{equation}
\item
The $\Ae$-module $\due N \blact {} \otimes_\Aopp M_\ract$
carries a right $ U $-module structure via
\begin{equation}
\label{superga1}
(n \otimes_\Aopp m) \leftslice u := nu_+ \otimes_\Aopp u_-m.
\end{equation}
\end{enumerate}
\item
Let $(U,A)$ be a right Hopf algebroid instead.
\begin{enumerate}
\compactlist{99}
\item
The $\Ae$-module $ \Hom_\ahha(M,M') $ carries a left $ U $-module structure given by
\begin{equation}
\label{gianduiotto2}
(uf)(m) := u_{[+]}\big( f(u_{[-]}m) \big).
\end{equation}
In particular, $ M_{\scalastd} $ is naturally endowed with a left $ U $-module structure.
\item
The $\Ae$-module
$\Hom_{\Aopp}(N,N') $ carries a left $ U $-module structure given by
\begin{equation}
\label{lingotto2}
(u \rightslice f)(n) := \big( f(nu_{[+]}) \big)u_{[-]}.
\end{equation}
\item
The $\Ae$-module
$ N_{\bract} \otimes^\ahha \due M \lact {} $
carries a right $ U $-module structure given by
\begin{equation}
\label{superga2}
(n \otimes^\ahha m) \leftslice u := nu_{[+]} \otimes^\ahha u_{[-]}m.
\end{equation}
\end{enumerate}
\end{enumerate}
\end{proposition}
\begin{note}
\label{osmanstoechter}
These structures are well-known for $ D $-modules (that is, when $ U = D_X $, see \cite{Bor:ADM, Kas:DMAMC}) and were later extended to $ V^\ell(L) $-modules in \cite{Che:PDFKALS}, \cite{Chemla3}. The results about tensor products can be found in \cite{KowKra:DAPIACT}, whereas \rmref{gianduiotto1} serves in \cite[Thm.~3.5]{Schau:DADOQGHA} to characterise a possible (left) Hopf structure on a bialgebroid.
\end{note}
\subsection{Switching left and right modules: dualising modules}
\label{malgenauerhinsehen}
We investigate now conditions which imply an equivalence between the categories of left and of right $ U $-modules for a left bialgebroid $ U $ which is simultaneously a left and right Hopf algebroid. As in other frameworks, this is guaranteed by the existence of a suitable {\em dualising module}. This is the content of the next result, which generalises the well-known equivalence of categories between left and right $ \cD $-modules (due to Borel \cite{Bor:ADM} and Kashiwara \cite{Kas:DMAMC}). It also generalises the equivalence between left and right modules over a Lie-Rinehart algebra, {\em cf.}~\cite{Che:PDFKALS}.
\begin{proposition}
\label{left and right modules}
Let $(U,A)$ be simultaneously a left and right Hopf algebroid.
Assume that there exists a right $ U $-module $ P $, where $ P_\bract $ is finitely generated projective over $\Aop$, such that
\begin{enumerate}
\compactlist{99}
\item
the left $U$-module morphism
$$
A \to \Hom_{\Aopp}(P,P), \quad a \mapsto \{ p \mapsto a \blact p \}
$$
is an isomorphism of $k$-modules;
\item
the evaluation map
\begin{equation}
\label{lacartachenontagliaglialberi}
\due P \blact {} \otimes_\Aopp \Hom_{\Aopp}(P , N)_{\ract} \to N, \quad p \otimes_\Aopp \phi \mapsto \phi(p)
\end{equation}
is an isomorphism for any $ N \in \modu $.
\end{enumerate}
Then
$$
\umod \to \modu, \quad M \mapsto \due P \blact {} \otimes_\Aopp M_\ract
$$
is an equivalence of categories with quasi inverse given by $ N' \mapsto \Hom_{\Aopp}(P , N') $.
\end{proposition}
\begin{proof}
For $M \in \umod$ and $N, N' \in \modu $, one checks with \rmref{mampf3} that
the map
$$
M_\ract \otimes_\ahha {}_\lact \Hom_{\Aopp}(N, N') \to \Hom_{\Aopp}(N, \due {N'} \blact {} \otimes_\Aopp M_\ract), \
m \, \otimes_\ahha \chi \mapsto \{n \mapsto \chi(n) \otimes_\Aopp m \}
$$
is a morphism of left $U $-modules, where the left $U$-module structure on the left hand side is given by \rmref{baratti&milano} combined with \rmref{lingotto2}, and on the right hand side by \rmref{lingotto2} combined with \rmref{superga1}. It is even an isomorphism if $N_\bract $ is finitely generated projective over $A$.
On the other hand, using \rmref{mampf2} and \rmref{Sch7}, one easily sees that
the evaluation \rmref{lacartachenontagliaglialberi}
is a morphism of right $ U $-modules; it is then an isomorphism by hypothesis, which finishes the proof.
\end{proof}
\begin{remark}
A right $U$-module $P$ with the properties as in the above proposition appeared in various contexts in the literature: we shall call it a {\em dualising module}. We refer to \cite{Che:PDFKALS, KowKra:DAPIACT, Hue:DFLRAATMC} for applications and details, and in particular to the situation in \S\ref{heisz}.
\end{remark}
\section{Comodule equivalences and induced maps between duals}
\label{fallouts}
The aim of this section is to construct a map between the left and right dual of a left Hopf algebroid, which in some sense replaces the missing antipode on either of the duals. This can be essentially done in two ways, either by a quite straightforward generalisation of the antipode construction on the dual of a cocommutative left Hopf algebroid as in \cite{KowPos:TCTOHA}, or by considering Ph\`ung's comodule equivalence in \cite{Phu:TKDFHA} as a starting point, as suggested by the referee of the present paper. To pursue the latter approach, we will review and slightly extend the results in {\em op.~cit.}
\subsection{A categorical equivalence for comodules}
The following theorem, originally due to \cite{Phu:TKDFHA}, shows that under the given conditions every right $U$-comodule can be transformed into a left one (resp.\ vice versa in the second case). We repeat it here for future use and also slightly extend it by saying that the two given functors are quasi-inverse to each other and that they are (strict) monoidal:
\begin{theorem}
\label{U-comod <-> comod-U}
Let $(U,A)$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item
Let $(U,A)$ be additionally a left Hopf algebroid such that $U_{\ract}$ is projective.
Then there exists a (strict) monoidal functor $F: \comodu \to \ucomod $; namely, if $M$ is a right $U$-comodule with coaction
$m \mapsto m_{(0)} \otimes_\ahha m_{(1)} $, then
\begin{equation}
\label{borromini}
\lambda_\emme: M \to U_\ract \otimes_\ahha M , \quad m \mapsto m_{(1)-} \otimes_\ahha m_{(0)} \epsilon(m_{(1)+}) ,
\end{equation}
defines a left comodule structure on $M$ over $U$.
\item
Let $(U,A)$ be a right Hopf algebroid such that $_{\lact}U$ is projective.
Then there exists a (strict) monoidal functor $ G: \ucomod \to \comodu $; namely, if $N$ is a left $U$-comodule with coaction
$n \mapsto n_{(-1)} \otimes_\ahha n_{(0)} $, then
\begin{equation}
\label{bernini}
\rho_\enne: N \to N \otimes_\ahha \due U \lact {} , \quad n \mapsto \epsilon(n_{(-1)[+]}) n_{(0)} \otimes_\ahha n_{(-1)[-]} ,
\end{equation}
defines a right comodule structure on $N$ over $U$.
\item
If $U$ is both a left and right Hopf algebroid and if both $U_{\ract}$ and $_{\lact}U$ are $A$-projective,
then the functors mentioned in {\it (i)} and {\it (ii)} are quasi-inverse to each other and we have an equivalence
$$
\ucomod \simeq \comodu
$$
of monoidal categories.
\end{enumerate}
\end{theorem}
\begin{proof}
Let us first prove that (\ref{borromini}) is well defined.
For any right $U$-comodule $ M $ with coaction $\rho: M \to M \otimes_\ahha U$, there is a well-defined map $\id_\emme \otimes_\ahha \epsilon: M \otimes_\ahha U \to M$. Its restriction to the Takeuchi product $M \times_\ahha U$ is a left $A$-module map as shows the following equation: for any $\sum_i m_i \otimes u_i \in M \times_\ahha U$ and any $a \in A$, one has
$$
\sum_i m_i \eps(a \blact u_i) = \sum_i m_i \eps(u_i \bract a) = \sum_i a m_i \eps(u_i).
$$
Thus, there is a well-defined map
$$
\id_\emme \times_\ahha \eps : \,
M \times_\ahha U \to M, \quad
\sum_i m_i \otimes u_i \mapsto \sum_i m_i \eps(u_i),
$$
and hence, in particular, the map
\begin{equation}
\label{def-phi}
\phi := (\id_\emme \times_\ahha \eps) \otimes_\Aopp \id_\uhhu : \,
(M \times_\ahha U) \times_\Aopp U \to M \times_\Aopp U
\end{equation}
is well-defined, too.
On the other hand, any right coaction corestricts to a map $M \to M \times_\ahha U$;
similarly, the translation map
$\beta^{-1}(u \otimes_\ahha 1) = u_+ \otimes_\Aopp u_-$ of $ U $ corestricts to a map $U \to U \times_\Aopp U $. Combining these two maps gives a map
\begin{equation}
\label{def-psi}
\psi: M \to M \times_\ahha (U \times_\Aopp U),
\end{equation}
and it is clear that if we could combine $\phi$ in \rmref{def-phi}
with $\psi$ in \rmref{def-psi}
followed by a tensor flip, this would
yield the map \rmref{borromini}.
Now the problem is that usually $(M \times_\ahha U) \times_\Aopp U $ and $M \times_\ahha (U \times_\Aopp U)$ are different, hence the two maps might not be composable. Introducing as in \cite[Def.~1.4]{Tak:GOAOAA} the triple Takeuchi product
\begin{equation*}
\begin{split}
M \times_\ahha U \times_\Aopp U :=
\{\textstyle\sum_i & m_i \otimes u_i\otimes v_i \in M \otimes_\ahha U \otimes_\Aopp U \mid \\
& \textstyle \sum_i am_i \otimes u_i \ract b \otimes v_i
= \sum_i m_i \otimes u_i \bract a \otimes b \blact v_i, \ \forall \ a,b \in A \}.
\end{split}
\end{equation*}
It can be seen that $ \psi $
actually maps to $M \times_\ahha U \times_\Aopp U $
but
it is a priori not clear whether $ \phi $ can be directly defined on $M \times_\ahha U \times_\Aopp U $ so as to make the two maps composable.
However, in any case there are always maps
$$
\gamma: M \times_\ahha (U \times_\Aopp U) \to M \times_\ahha U \times_\Aopp U, \quad m \otimes_\ahha u \otimes_\Aopp v \mapsto m \otimes_\ahha u \otimes_\Aopp v
$$
and
$$
\ga: (M \times_\ahha U) \times_\Aopp U \to M \times_\ahha U \times_\Aopp U, \quad m \otimes_\ahha u \otimes_\Aopp v \mapsto m \otimes_\ahha u \otimes_\Aopp v.
$$
If now $U_\ract$ is projective, $\ga$ is an isomorphism \cite[Prop.~1.7]{Tak:GOAOAA}; then the composition $\tau \circ \phi
\circ \alpha^{-1} \circ \gamma \circ \psi$
of well-defined maps (where $ \tau $ is the tensor flip) yields a well-defined map again, and on an element $m \in M$ it is an easy check that this gives the map $\gl_\emme$ in \rmref{borromini}.
That the so-defined map $\gl_\emme$ is $\Ae$-linear follows from the $\Ae$-linearity of the right coaction along with \rmref{Sch9}. That $\lambda_\emme$ indeed defines a left $U$-coaction is an easy check using \rmref{Sch5} and \rmref{Sch4}, the counitality of the bialgebroid $U$, and the coassociativity with the $\Ae$-linearity of the right $U$-coaction on $M$ again: we have for $m \in M$
\begin{equation*}
\begin{split}
(\gD_\ell &\otimes \id)\lambda_\emme(m)
= m_{(1)-(1)} \otimes_\ahha m_{(1)-(2)} \otimes_\ahha m_{(0)} \epsilon(m_{(1)+}) \\
&= m_{(1)-} \otimes_\ahha m_{(1)+-} \otimes_\ahha m_{(0)} \epsilon(m_{(1)++}) \\
&= m_{(1)-} \otimes_\ahha \big( t^\ell \epsilon(m_{(1)+(2)}) m_{(1)+(1)} \big)_- \otimes_\ahha m_{(0)} \epsilon\big(\big(t^\ell \epsilon(m_{(1)+(2)}) m_{(1)+(1)}
\big)_+\big) \\
&= m_{(2)-} \otimes_\ahha \big( t^\ell \epsilon(m_{(2)+}) m_{(1)} \big)_- \otimes_\ahha m_{(0)} \epsilon\big(\big( t^\ell \epsilon(m_{(2)+}) m_{(1)} \big)_+\big)
\\
&= (\id \otimes \lambda_\emme)\lambda_\emme(m).
\end{split}
\end{equation*}
The counitality of $\lambda_\emme$ follows from \rmref{Takeuchicoaction} along with the second equation in \rmref{castelnuovo}.
As for the claim that the so-given functor $F: \comodu \to \ucomod$ is (strict) monoidal, observe first that for any two $M, M'$ in the monoidal category $\comodu$, their tensor product $ M \otimes_\ahha M' $ is a right $U$-comodule by means of the codiagonal coaction $m \otimes_\ahha m' \mapsto (m_{(0)} \otimes_\ahha m'_{(0)}) \otimes_\ahha m'_{(1)} m_{(1)}$,
that is, with a flip in the factors in $U$. On the other hand, the tensor product of two $N, N'$ in the monoidal category $\ucomod$ becomes a left $U$-comodule again
via
$n \otimes_\ahha n' \mapsto n_{(-1)} n'_{(-1)} \otimes_\ahha (n_{(0)} \otimes_\ahha n'_{(0)})$.
By the bialgebroid properties, \rmref{Sch6}, and \rmref{Takeuchicoaction}
it is then simple to see that
\begin{equation*}
\begin{split}
( m'_{(1)} m_{(1)})_- &\otimes_\ahha ( m_{(0)} \otimes_\ahha m'_{(0)}) \epsilon\big(( m'_{(1)} m_{(1)})_+\big)
\\
&=
m_{(1)-} m'_{(1)-} \otimes_\ahha \big( m_{(0)} \otimes_\ahha m'_{(0)} \epsilon\big(m'_{(1)+} s^\ell(\epsilon(m_{(1)+}))\big) \big)
\\
&=
m_{(1)-} m'_{(1)-} \otimes_\ahha \big( m_{(0)}\epsilon(m_{(1)+}) \otimes_\ahha m'_{(0)} \epsilon(m'_{(1)+}) \big),
\end{split}
\end{equation*}
that is, $F(M \otimes_\ahha M') = F(M) \otimes_\ahha F(M')$. Also, the unit object in both $\comodu$ and $\ucomod$ is given by $A$ with coaction $a \mapsto t^\ell(a)$ resp.\ $a \mapsto s^\ell(a)$, and $F(A) = A$ now follows from \rmref{Sch9}. Moreover, note that $F$ does not affect the underlying $\Ae$-module structures of the comodules in question, and hence its (strict) monoidality follows.
The proof of {\em (ii)\/} is similar, and the last claim follows by the preceding two combined with
a direct computation: applying $GF$ to a right comodule $M \in \comodu$, the resulting right coaction on $M$ reads
$$
M \to M \otimes_\ahha \due U \lact {}, \quad m \mapsto
\epsilon(m_{(1)-[+]}) m_{(0)} \epsilon(m_{(1)+}) \otimes_\ahha m_{(1)-[-]}.
$$
By using \rmref{mampf2}, the coassociativity and counitality of the original right coaction on $M$, \rmref{Takeuchicoaction}, \rmref{castelnuovo}, and \rmref{Sch8} one obtains
\begin{equation*}
\begin{split}
\epsilon(m_{(1)-[+]}) m_{(0)} \epsilon(m_{(1)+}) \otimes_\ahha m_{(1)-[-]} &= \epsilon(m_{(1)-}) m_{(0)} \epsilon(m_{(1)+}) \otimes_\ahha m_{2} \\
&= m_{(0)} \epsilon( \epsilon(m_{(1)-}) \blact m_{(1)+}) \otimes_\ahha m_{(2)} \\
&= m_{(0)} \epsilon(m_{(1)}) \otimes_\ahha m_{(2)} = m_{(0)} \otimes_\ahha m_{(1)},
\end{split}
\end{equation*}
that is, the right coaction on $M$ we started with. An analogous consideration holds for $FG$ using \rmref{mampf3}, \rmref{Tch8}, and the Takeuchi property that holds for left $U$-comodules analogous to \rmref{Takeuchicoaction}.
\end{proof}
\begin{remark}
Note that the equivalence in Theorem \ref{U-comod <-> comod-U} does {\em not\/} boil down to the usual equivalence of left and right comodules via the antipode (as there is no antipode for left or right Hopf algebroids, not even if the bialgebroid is simultaneously both). Even if we dealt with a full Hopf algebroid, this is still a different kind of equivalence (compared to
the construction in \cite[Remark 4.6]{Boe:HA}), as follows from the considerations in \S\ref{napoleon} and \S\ref{enoteca} below.
For example, if the left Hopf algebroid $U$ is considered a right comodule over itself via the coproduct, the left $U$-coaction on $U$ from \rmref{borromini} is given by
$$
U \to U_\ract \otimes_\ahha \due U \blact {}, \quad u \mapsto u_- \otimes_\ahha u_+,
$$
that is, the ``flipped'' translation map.
On the other hand,
for Hopf {\em algebras} the construction in Theorem \ref{U-comod <-> comod-U} is exactly the equivalence induced by the antipode.
\end{remark}
\subsection{Constructing maps between the duals}
\label{wiley}
We now want to construct a map between the right and the left dual of a left Hopf algebroid. To this end, we first need to recall from
\cite[Theorem 3.1.11]{Kow:HAATCT}
the following bialgebroid generalisation of the classical bialgebra module-comodule correspondence, which, however, in its first part comes somewhat unexpected at first sight:
\begin{prop}
\label{KatCM1}
Let $(U,A)$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item
There exists a functor $ \comodu \to \mathbf{Mod}\mbox{-}{U_{\scalastd}} $; namely, if $M$ is a right $U$-comodule with coaction
$m \mapsto m_{(0)} \otimes_\ahha m_{(1)} $, then
\begin{equation}
\label{sale}
M \otimes_k U_\scalast \to M , \quad m \otimes_k \psi \mapsto m_{(0)}\psi(m_{(1)}) ,
\end{equation}
defines a right module structure over the $\Ae$-ring $U_\scalast$.
If $\due U \lact {}$ is finitely generated $A$-projective
(so that $ U_\scalast $ is a right bialgebroid),
this
functor is monoidal and has a quasi-inverse $\mathbf{Mod}\mbox{-}{U_{\scalastd}} \to \comodu$ such that there is an
equivalence
$
\comodu \simeq \mathbf{Mod}\mbox{-}{U_{\scalastd}}
$
of categories.
\item
Likewise, there exists a functor $ \ucomod \to \mathbf{Mod}\mbox{-}{U^{\scalast}} $; namely, if $N$ is a left $U$-comodule with coaction
$n \mapsto n_{(-1)} \otimes_\ahha n_{(0)} $, then
\begin{equation}
\label{pepe}
N \otimes_k U^\scalast \to N , \quad n \otimes_k \phi \mapsto \phi(n_{(-1)})n_{(0)} ,
\end{equation}
defines a right module structure over the $\Ae$-ring $U^\scalast$.
If $\due U {}\ract$ is finitely generated
$A$-projective
(so that $ U^\scalast $ is a right bialgebroid),
this functor is monoidal and has a quasi-inverse $\mathbf{Mod}\mbox{-}{U^{\scalast}} \to \ucomod$ such that there
is an equivalence
$
\ucomod \simeq \mathbf{Mod}\mbox{-}{U^{\scalast}}
$
of categories.
\end{enumerate}
\end{prop}
The case $(ii)$ of the above Proposition \ref{KatCM1} can also be found in \cite[\S5]{Schau:DADOQGHA}.
An explicit proof and a description of all involved functors is given in \cite[\S3.1]{Kow:HAATCT}, along with the respective structure maps of the right bialgebroids $(U_{\scalastd}, A, s_{\scalastd}^r, t_{\scalastd}^r, \gD_{\scalastd}^r, \pl_{\scalastd})$ and $(U^{\scalast}, A, s_r^\scalast, t_r^\scalast, \gD^{\scalast}_r, \pl^{\scalast})$, in case the respective mentioned finiteness assumptions are met.
Observe that when $(U,A)$ is both a left and a right Hopf algebroid and both $U_\ract$ as well as $\due U \lact {}$ are finitely generated projective over $A$, then \rmref{eastpak} here
below
is a commutative diagram of monoidal equivalences.
We shall also need an explicit expression of the induced coaction on $M \in \mathbf{Mod}\mbox{-}{U_{\scalastd}}$ in case $\due U \lact {}$ is finitely generated projective as in {\em (i)}: let $m \otimes_k \psi \mapsto m\psi$ denote the right $U_{\scalastd}$-action on $M$ and
$\{e_i\}_{1 \leq i \leq n} \in U, \ \{ e^i\}_{1 \leq i \leq n} \in U_{\scalastd}$ a dual basis (see, for example, \cite[p.~202]{AndFul:RACOM} for the notion of dual basis of a finitely generated projective module).
Then the resulting right $U$-coaction on $M$ can be expressed as
\begin{equation}
\label{ferra}
m \mapsto \sum_i m e^i \otimes_\ahha e_i,
\end{equation}
see \cite[Eq.~(3.1.23)]{Kow:HAATCT}.
Consider now the diagram
\begin{equation}
\label{eastpak}
\begin{gathered}
\xymatrix{ \comodu \ar@{->}[r] \ar@{->}[d] & \mathbf{Mod}\mbox{-}U_{\scalast} \ar@{..>}[d] \\ \ucomod \ar@{->}[r]
& \mathbf{Mod}\mbox{-}U^{\scalast}
}
\end{gathered}
\end{equation}
of categories, where the left vertical arrow is that from Theorem \ref{U-comod <-> comod-U} {\em (i)}.
Under the finiteness assumption for $\due U \lact {}$, the upper horizontal arrow is invertible.
One therefore
obtains a functor that corresponds to the dotted arrow if $ U_\ract{}$ is $A$-projective and
$\due U \lact {}$ is finitely generated $A$-projective.
Explicitly, by using \rmref{ferra}, \rmref{borromini}, and \rmref{pepe} one obtains on a right $U_{\scalast}$-module $M$
with $U_{\scalast}$-action $m \otimes_k \psi \mapsto m\psi$ the following right $U^{\scalast}$-action:
\begin{equation}
\label{dieda}
M \otimes_k U^{\scalast} \to M, \quad m \mapsto m \Yleft \phi := \phi({e_i}_-) m e^i \epsilon({e_i}_+) = me^i \epsilon\big({e_i}_+ s^\ell(\phi({e_i}_-))\big),
\end{equation}
where the second expression follows by taking the Takeuchi property \rmref{Takeuchicoaction} of the right coaction \rmref{ferra} into account, along with \rmref{Sch9}.
Consider now the case $M = U_{\scalast}$ as right module over itself by right multiplication;
then as in \rmref{dieda} it also carries a right $ U^\scalast $-action,
which is equivariant with respect to the regular left $ U_{\scalast} $-action, that is
\begin{equation}
\label{equivariance}
(\psi' \psi'') \Yleft \phi = \psi' ( \psi'' \Yleft \phi).
\end{equation}
In particular, this implies $ \psi \Yleft \phi = \psi ( 1_{U_{\scalast}} \Yleft \phi )$, which leads us to consider
\begin{equation}
\label{1st-def_Sstarup}
S^\scalast\phi := 1_{U_{\scalast}} \Yleft \phi = \epsilon \Yleft \phi.
\end{equation}
With \rmref{dieda}, we see that
$
S^\scalast\phi = \epsilon \Yleft \phi = e^i s^r_\scalast\big(\epsilon\big({e_i}_+ s^\ell(\phi({e_i}_-))\big)\big).
$
Hence, for any $u \in U$,
\begin{equation}
\label{sososo}
\begin{split}
S^\scalast\phi(u) &= \langle \epsilon \Yleft \phi, u\rangle = \big\langle e^i s^r_\scalast\big(\epsilon\big({e_i}_+ s^\ell(\langle \phi, {e_i}_- \rangle)\big)\big), u \big\rangle \\
& = \langle e^i, u \rangle \langle \epsilon, {e_i}_+ s^\ell(\langle \phi, {e_i}_-
\rangle) \rangle =\langle \epsilon, s^\ell( \langle e^i, u \rangle ) {e_i}_+ t^\ell(\langle \phi, {e_i}_-
\rangle) \rangle,
\end{split}
\end{equation}
where we used \cite[Eq.~(3.1.3)]{Kow:HAATCT} in the third step and \rmref{castelnuovo} in the fourth.
Inserting now into \rmref{sososo} the identity
$$
u_+ \otimes_\Aopp u_- = s^\ell( \langle e^i, u \rangle ) {e_i}_+ \otimes_\Aopp {e_i}_-,
$$
which is seen by applying the bijective Hopf-Galois map $\alpha_\ell$ from \rmref{latoconvalida} to both sides (as we assumed $U$ to be a left Hopf algebroid), one further obtains
\begin{equation}
\label{S*-prov}
S^\scalast\phi(u) =\langle \epsilon, s^\ell( \langle e^i, u \rangle ) {e_i}_+ t^\ell(\langle \phi, {e_i}_- \rangle) = \epsilon\big(u_+ t^\ell(\phi(u_-))\big).
\end{equation}
As will be discussed at length in the next section, this yields a map $S^\scalast: U^{\scalast} \to U_{\scalast}$ (as is seen using \rmref{Sch9} and \rmref{castelnuovo}) of $\Ae$-rings that even makes sense
without any projectiveness or finiteness assumptions.
By means of \rmref{sale} and \rmref{S*-prov}, the action \rmref{dieda} can then be written as
\begin{equation}
\label{diqua}
m \Yleft \phi := m S^\scalast(\phi),
\end{equation}
which, without assuming any finiteness conditions on $U$, still leads to a functor $\mathbf{Mod}\mbox{-}U_{\scalast} \to \mathbf{Mod}\mbox{-}U^{\scalast}$ between the categories of modules over $\Ae$-rings.
If instead $ U $ is a {\em right\/} Hopf algebroid,
where $U_\ract$ is
finitely generated $A$-projective and
$\due U \lact {}$ is $A$-projective,
one obtains by analogous steps a map $S_\scalast : U_{\scalast} \to U^{\scalast} $ given by
$$
S_\scalast\psi(u) = \epsilon\big(u_{[+]} s^\ell(\psi(u_{[-]}))\big)
$$
for any $u \in U$, to which analogous comments apply as above.
We will discuss the properties of these maps in detail in the subsequent \S\ref{linking-structures}
\section{Linking structure for the duals of left Hopf algebroids}
\label{linking-structures}
In this section --- the core of the present work ---, we find that the map $S^\scalast$ constructed in the previous section is linking the right dual to the left dual of a left Hopf algebroid, which is apparently as close as one can get to an explicit formula of an antipode kind-of structure on the dual. Note, however, that even in the case of a full Hopf algebroid this map is not simply the transpose of the antipode, as discussed in \S\ref{napoleon}.
In some sense, this special map amounts to sort of a generalisation of (the antipode in) a full Hopf algebroid as explained in Remark \ref{schonschoen}.
As mentioned before, the definition of the map $ S^\scalast $ (and $ S_\scalast $) actually makes sense even without any
finiteness or projectiveness assumptions. Indeed, one can trace their first appearance already in \cite{KowPos:TCTOHA} in the r\^ole of the antipode in the example of the bialgebroid of jet spaces.
In what follows, we will prove the fact that $S^\scalast$ and $S_\scalast$ are morphisms of $\Ae$-rings in a direct way, whereas the fact that under suitable finiteness assumptions they are bialgebroid morphisms is shown by using the comodule equivalence discussed
in the previous section (note, however, that even the latter can be achieved by direct computation).
In particular, since the finiteness assumptions are not needed for all properties stated below, we will be able to apply $ S^\scalast $ and $ S_\scalast $ in greater generality to the examples in \S\ref{lautesGequatsche}.
\subsection{Morphisms between left and right duals}
\label{linking}
Let $(U,A) $ be a left bialgebroid. If it is additionally a left Hopf algebroid,
its right dual $ U^{\scalast} $ (see \S\ref{mayitbe}) carries a left $U$-module structure as in \rmref{gianduiotto1};
(re-)define
\begin{equation}
\label{sstarup}
S^{\scalast}(\phi)(u) := (u\phi)(1_\uhhu) = \epsilon_\uhhu \big( u_+ t^\ell(\phi(u_-)) \big), \qquad \forall \phi \in U^{\scalast} , \ u \in U.
\end{equation}
Likewise, if the left bialgebroid $(U,A) $ is a right Hopf algebroid instead, its left dual $ U_{\scalastd} $ (see \S\ref{mayitbe} again) carries a left $U$-module structure as in \rmref{gianduiotto2}, with the help of which one (re-)defines
\begin{equation}
\label{sstardown}
S_{\scalastd}(\psi)(u) := (u\psi)(1_\uhhu) = \epsilon_\uhhu \big( u_{[+]} s^\ell( \psi(u_{[-]})) \big), \qquad \forall \psi \in U_{\scalastd} , \ u \in U.
\end{equation}
The following result presents the key properties of the maps $ S^{\scalast} $ and $ S_{\scalastd} $:
\begin{theorem}
\label{S*-morphism}
Let $(U,A)$ be a left bialgebroid.
\begin{enumerate}
\compactlist{99}
\item
If $(U,A) $ is moreover a left Hopf algebroid, \rmref{sstarup} defines a morphism
$ S^{\scalast} : U^{\scalast} \to U_{\scalastd} $
of $ A^e $-rings with augmentation (the ``counit''); if in addition both $ \due U \lact {} $ and $ U_\ract $ are finitely generated projective as $ A $-modules, then $ (S^{\scalast}, \id_\ahha) $ is a morphism of right bialgebroids. In any case, $ S^{\scalast} $ is also a morphism of left $U$-modules for the action \rmref{lingotto1} on $ U^{\scalast} $ and the left action on $ U_{\scalast} $ given by right multiplication in $U$.
\item
If $(U,A) $ is a right Hopf algebroid instead, \rmref{sstardown} defines a morphism
$ S_{\scalastd} : U_{\scalastd} \to U^{\scalast} $
of $ A^e $-rings with augmentation (the ``counit''); if in addition both $ \due U \lact {} $ and $ U_\ract $ are finitely generated projective as $ A $-modules, then $ (S_{\scalast}, \id_\ahha) $ is a morphism of right bialgebroids. In any case, $ S_{\scalastd} $ is also a morphism of left $ U $-modules for the action \rmref{lingotto2} on $ U_{\scalastd} $ and the left action on $ U^{\scalast} $ given by right multiplication in $U$.
\end{enumerate}
\end{theorem}
\begin{proof}
We only prove part {\it (i)} as {\it (ii)} follows {\em mutatis mutandis}. For the explicit computations, we will again use the notation and description of the structure maps of the two right bialgebroids
$(U_{\scalastd}, A, s_{\scalastd}^r, t_{\scalastd}^r, \gD_{\scalastd}^r, \pl_{\scalastd})$ and $(U^{\scalast}, A, s_r^\scalast, t_r^\scalast, \gD^{\scalast}_r, \pl^{\scalast})$
--- where the coproduct $ \gD_{\scalastd}^r $ or $ \gD^{\scalast}_r $ only make sense if $U_\ract$ resp.\ $\due U \lact {}$ is finitely generated $A$-projective --- as given in detail in \cite[\S3.1]{Kow:HAATCT}, together with the respective properties of left and right pairings $\langle .,. \rangle$ as in Definition \ref{left and right 1}.
Direct verification shows that $S^{\scalast}$ takes values in $U_{\scalastd}$. Besides,
for $S^{\scalast}$ to be a bialgebroid morphism, we need to show the following properties:
$$ \displaylines{
\qquad
\text{\it (a)} \hfill S^{\scalast} s^{\scalast}_r = s_{\scalastd}^r ,
\quad S^{\scalast} t_r^{\scalast} = t_{\scalastd}^r, \quad
\pl_{\scalastd} S^{\scalast} = \pl^{\scalast},
\hfill \cr
\qquad
\text{\it (b)} \hfill S^{\scalast}\big(\phi \phi'\big) = S^{\scalast} (\phi) S^{\scalast} (\phi') \hfill \cr
\qquad
\text{\it (c)} \hfill \gD^r_{\scalastd} S^{\scalast} = (S^{\scalast} \otimes S^{\scalast}) \gD^{\scalast}_r , \hfill }
$$
(where, as said before, {\it (c)\/} only makes sense if $U_\ract$ and $\due U \lact {}$ are finitely generated $A$-projective).
As for {\em (a)}, we find for $u \in U$, $a \in A$ by direct computation using
\rmref{Sch8} and \rmref{Sch9}:
\begin{equation*}
\begin{split}
S^{\scalast}\big(s_r^{\scalast}(a)\big)(u) &= \epsilon\big( u_+ t^\ell\big(s_r^{\scalast}(a)(u_-) \big)\big)
= \epsilon \big( u_+ t^\ell\big(\epsilon(u_- s^\ell(a)) \big)\big)
= \epsilon(u) a = s^r_{\scalastd}(a)(u).
\end{split}
\end{equation*}
Likewise, the second identity follows from
\begin{equation*}
\begin{split}
S^{\scalast}\big(t_r^{\scalast}(a)\big)(u) &= \epsilon\big( u_+ t^\ell\big(t_r^{\scalast}(a)(u_-) \big) \big)
= \epsilon\big( u_+ t^\ell(a \epsilon (u_-)) \big)
= \epsilon(ut^\ell(a))
= t^r_{\scalastd}(a)(u).
\end{split}
\end{equation*}
The last identity in {\em (a)} regarding the respective counits is for $\phi \in U^{\scalast}$ proven by the line
$$
\pl_{\scalastd} S^{\scalast}(\phi) = S^{\scalast}(\phi)(1_\uhhu) = \phi(1_\uhhu) = \partial^{\scalast} \phi.
$$
As for {\em (b)}, let us first more generally compute an element $S^\scalast(\phi )\psi$ for $\phi \in U^\scalast$ and $\psi \in U_\scalastd$: by \cite[Eq.~(3.1.1)]{Kow:HAATCT}, Eq.~\rmref{Sch4}, and the properties of a bialgebroid counit, we have
\begin{equation*}
\begin{split}
\langle S^\scalast(\phi )\psi ,u \rangle &=
\langle \psi, t^\ell(\langle u_{(2)} , S^\scalast(\phi)\rangle)u_{(1)}\rangle
=\langle \psi, t^\ell(\langle \epsilon, u_{(2)+}t^\ell(\langle\phi, u_{(2)-}\rangle)\rangle)u_{(1)} \rangle \\
&= \langle \psi, t^\ell(\langle \epsilon, u_{+(2)}t^\ell(\langle\phi, u_{-}\rangle)\rangle)u_{+(1)}\rangle \\
&=\langle \psi, t^\ell(\langle \epsilon, u_{+(2)}s^\ell(\langle\phi, u_{-}\rangle)\rangle)u_{+(1)} \rangle \\
&= \langle \psi, t^\ell(\langle \epsilon, u_{+(2)}\rangle) u_{+(1)}t^\ell(\langle\phi, u_{-}\rangle)\rangle
=\langle\psi, u_{+}t^\ell(\langle\phi, u_{-}\rangle)\rangle.
\end{split}
\end{equation*}
With the help of this property, by \cite[Eq.~(3.1.2)]{Kow:HAATCT}
along with \rmref{Sch5}, \rmref{Sch9}, and the fact that the counit in $U$ gives the unit in $U_\scalastd$, one sees that for all $ \phi, \phi' \in U^{\scalast} $
\begin{equation*}
\begin{split}
\langle S^{\scalast}( \phi\phi') , u \rangle
&= \langle \epsilon , u_+ t^\ell(\langle \phi \phi', u_- \rangle )\rangle
= \langle \epsilon , u_+ t^\ell(\langle \phi', s^\ell\phi(u_{-(1)}) u_{-(2)} \rangle )\rangle \\
&= \langle \epsilon , u_{++} t^\ell(\langle \phi', s^\ell\phi(u_{-}) u_{+-} \rangle )\rangle \\
&= \langle \epsilon , (u_{+} t^\ell\phi(u_-))_+ t^\ell(\langle \phi', (u_{+} t^\ell\phi(u_-))_-
\rangle )\rangle \\
&= \langle S^\scalast(\phi') \epsilon , u_{+} t^\ell\phi(u_-)\rangle = \langle S^{\scalast}(\phi)S^{\scalast}( \phi' ), u \rangle.
\end{split}
\end{equation*}
Observe that if $\due U \lact {}$ is finitely generated $A$-projective, then {\em (b)} follows by the fact that \rmref{diqua} defines an action, but in general we do not want to assume this at this point.
For proving {\em (c)} --- when $U_\ract$ and $\due U \lact {}$ are finitely generated $A$-projective ---,
one could equally do this by a straightforward somewhat technical computation. A quicker way is to use the results
in \S\ref{fallouts}:
denoting the right coproduct on $U_{\scalastd}$ resp.\ $U^{\scalast}$ by Sweedler superscripts,
one has
\begin{equation*}
\begin{split}
S^\scalast(\phi)^{(1)} \otimes_\ahha S^\scalast(\phi)^{(2)}
&=
(\epsilon \otimes_\ahha \epsilon) S^\scalast(\phi)
= (\epsilon \otimes_\ahha \epsilon) \Yleft \phi \\
&= (\epsilon \Yleft \phi^{(1)}) \otimes_\ahha (\epsilon \Yleft \phi^{(2)})
= S^\scalast(\phi^{(1)}) \otimes_\ahha S^\scalast(\phi^{(2)}),
\end{split}
\end{equation*}
where in the first equation we used the monoidal structure on $\mathbf{Mod}\mbox{-}U_{\scalast}$, and in the third the fact that all functors in \rmref{eastpak} are strict monoidal.
The second part in {\em (i)} --- about the $ U $-linearity of $ S^{\scalast} $ ---, which is straightforward, is left to the reader.
\end{proof}
\begin{remark}
\label{S* as "antipode"}
When $ U $ is just a Hopf algebra over $ A = k $ with antipode $ S $, we have $ U^{\scalast} = (U_\scalastd)^\op_\coop $, and $ S^{\scalast} $ is nothing but the transpose of $ S $. If $ U^{\scalast}$ itself is in turn a Hopf algebra --- namely, if the transpose of the multiplication $ m_\uhhu $ in $ U $ takes values in the tensor square of $ U^{\scalast} $ ---, then $ S^{\scalast} $ is just the antipode of this dual Hopf algebra $ U^{\scalast} $. In this context, Theorem \ref{S*-morphism} simply expresses the fact that the antipode in a Hopf algebra is an antimorphism of algebras and of coalgebras.
\end{remark}
In particular, in case $U$ is both a left and right Hopf algebroid we have:
\begin{theorem}
\label{criptabalbi}
Let $ (U,A) $ be simultaneously a left and a right Hopf algebroid.
Then the maps $ S^{\scalast} $ and $ S_{\scalastd} $ are inverse to each other. Hence, if both $ A $-modules $ \due U \lact {} $ and $ U_\ract $ are, in addition, finitely generated projective, $ (S^{\scalast}, \id_\ahha ) $ and $ (S_{\scalastd}, \id_\ahha) $ are isomorphisms of right bialgebroids which are inverse to each other.
\end{theorem}
\begin{proof}
As for the first statement, we directly compute by means of the bialgebroid axioms along with \rmref{mampf3} and \rmref{Tch7}, for any $ \phi \in U^{\scalast} $:
\begin{equation*}
\begin{split}
(S_{\scalastd} S^{\scalast} \phi)(u) &= \epsilon\big(u_{[+]} s^\ell(S^{\scalast} \phi(u_{[-]}))\big)
= \epsilon\big(u_{[+]} s^\ell\big(\epsilon_U(u_{[-]+} t^\ell \phi(u_{[-]-}))\big)\big) \\
&= \epsilon\big(u_{[+]}u_{[-]+} t^\ell \phi(u_{[-]-})\big)
= \epsilon\big(u_{(2)[+]}u_{(2)[-]}t^\ell \phi(u_{(1)})\big) \\
&= \phi(u_{(1)})\epsilon(u_{(2)}) = \phi(u),
\end{split}
\end{equation*}
which proves that $ S_{\scalastd} \circ S^{\scalast} = \id_{\uhhu^{\scalast}} $. Likewise, one shows that $ S^{\scalast} \circ S_{\scalastd} = \id_{\uhhu_{\scalastd}} $.
\end{proof}
\subsection{The case of a full Hopf algebroid}
\label{napoleon}
If $ H $ is a full Hopf algebroid with bijective antipode $S$ in the sense of \cite{BoeSzl:HAWBAAIAD}, then it is, in particular, both a left and right bialgebroid (see the short summary below): therefore --- still assuming that $ \due H \lact {} $ and $ H_\ract $ are both finitely generated projective as $ A $-modules ---, there is a right bialgebroid analogue to the previous constructions concerning the maps $ S^{\scalast} $ and $ S_{\scalastd} $. On the other hand, the antipode $ S $ induces by transposition new maps $ S^t $, $\qttrd S t {}{}{}{} $, etc., for the dual spaces. Hereafter we discuss links between these various maps, in particular showing that, while for the Hopf algebra case one has identities like $ S^{\scalast} = \qttrd S t {}{}{} $ ({\em cf.}~Remark \ref{S* as "antipode"}), this is no longer the case for the general setup of full Hopf algebroids as illustrated in \S\ref{enoteca} below.
\begin{free text}{\bf Reminder on full Hopf algebroids.}
\label{reminders_H-ads}
Recall that a full Hopf algebroid structure (see, for example, \cite{Boe:HA}) on a $ k $-module $ H $
consists of the following data:
\begin{enumerate}
\compactlist{99}
\item
a left bialgebroid structure $ H^\ell := ( H, A, s^\ell, t^\ell, \Delta_\ell , \epsilon) $
over a $ k $-algebra $ A $;
\item
a right bialgebroid structure $ H^r := ( H, B, s^r, t^r, \Delta_r , \partial) $ over a $ k $-algebra $ B $;
\item
the assumption that the $ k $-algebra structures for $ H $ in {\em (i)\/} and in {\em (ii)\/} be the same;
\item
a $ k $-module map $ S : H \to H $;
\item
some compatibility relations between the previously listed data for which we refer to {\em op.\ cit.}
\end{enumerate}
We shall denote by lower Sweedler indices the left coproduct $ \Delta_\ell $ and by upper indices the right coproduct $ \Delta_r $, that is,
$ \Delta_\ell(h) =: h_{(1)} \otimes_\ahha h_{(2)} $ and $ \Delta_r(h) =: h^{(1)} \otimes_\behhe h^{(2)} $
for any $ h \in H $.
As said before, a full Hopf algebroid (with bijective antipode) is both a left and right Hopf algebroid but not necessarily vice versa (as illustrated in \S\ref{enoteca}).
In this case, the translation maps in \rmref{latoconvalida} are given by
\begin{equation}
\label{laterza}
h_+ \otimes_\Aopp h_- = h^{(1)} \otimes_\Aopp S(h^{(2)}) \quad \mbox{and} \quad h_{[+]} \otimes_\Bopp h_{[-]} = h^{(2)} \otimes_\Bopp S^{-1}(h^{(1)}),
\end{equation}
formally similar as for Hopf algebras.
\end{free text}
The following lemma \cite{Boe:HA, BoeSzl:HAWBAAIAD} will be needed to prove the main result in this subsection.
\begin{lemma}
\label{vitasnella}
Let $ H $ be any Hopf algebroid. Then
\begin{enumerate}
\compactlist{99}
\item
the maps $ \nu := \partial s^\ell : A \to B^\op $ and $ \mu := \epsilon s^r : B \to A^\op $ are isomorphisms of $ k $-algebras;
\item
the pair of maps $(S, \nu) : H^\ell \to {(H^r)}^\op_\coop $ gives an isomorphism of left bialgebroids;
\item
the pair of maps
$ (S, \mu) : H^r \to {(H^\ell)}^\op_\coop $ gives an isomorphism of right bialgebroids.
\end{enumerate}
\end{lemma}
The next observation might let us consider $ S^{\scalast} $ and $ S_{\scalastd} $ as sort of an analogue of the antipode on the dual:
\begin{proposition} \label{dual-cocomm_Hopf}
Let $ (U, A) $ be a cocommutative left bialgebroid (in particular, $A$ is commutative and $s^\ell = t^\ell$).
Then $(U,A) $ is a left Hopf algebroid if and only if it is a right Hopf algebroid; in this case, assuming in addition that $ \due U \lact {} $ and $ U_\ract $ are finitely generated $A$-projective,
$ (U^{\scalast},A) = ((U_{\scalastd})_\coop,A) $ is a full Hopf algebroid with involutive antipode $ \mathscr{S} := S^{\scalast} = S_{\scalastd} $.
\end{proposition}
\begin{proof}
The first claim directly holds true by the very definitions. The rest of the proof follows {\em verbatim\/} in the footsteps of the one of Theorem 3.17 in \cite{KowPos:TCTOHA}, which considers the special case for $ U = V^\ell(L)$.
\end{proof}
As mentioned before, one can also link the duals of a Hopf algebroid $(H,S)$ by transposed maps $ {}^{t\!}S $, which usually do not coincide with $S^{\scalast}$ or $S_{\scalastd}$ (see also \S\ref{enoteca}). The next result explains
a relation
between them.
\begin{theorem}
\label{movimentoallaforza}
Let $ H $ be a Hopf algebroid such that $ \due H \lact {} $ and $ H_\ract $ are finitely generated $A$-projective. Then the diagram
\begin{equation*}
\xymatrix{ {\big( {(H^r)}^\op_\coop \big)}^{\scalast} \ar@{->}^{{}^{t\!}S}[rr] \ar@{->}_{S^{\scalast}_r}[d] & & {(H^\ell)}^{\scalast} \ar@{->}^{S^{\scalast}_\ell}[d] \\
{\big( {(H^r)}^\op_\coop \big)}_{\scalastd} \ar@{->}_{{}^{t\!}S}[rr] & & {(H^\ell)}_{\scalastd} }
\end{equation*}
of right bialgebroid morphisms
is commutative.
\end{theorem}
\begin{proof}
Let us identify $ B^\op $ and $ A $ by means of the $ k $-algebra isomorphism $ \nu : A \to B^\op $ mentioned above; then the left algebroid $ {(H^r)}_\coop^\op $ is described by the sextuple
$$ \big( {(H^r)}^\op , \widehat{s^\ell} := s^r \nu , \widehat{t^\ell} := t^r \nu , \Delta_r^\coop , \widehat{\epsilon} := {\nu}^{-1} \partial \big) . $$
Moreover, the Hopf algebroid $ \big( {(H^r)}_\coop^\op , {(H^\ell)}_\coop^\op , (S, \mu) : {(H^r)}_\coop^\op \to H^\ell \big) $ is the one we have to consider to compute $ S_r^{\scalast} $. For $ \phi \in {\big( {(H_r)}^\op_\coop \big)}_{\!\scalast} $ and $ h \in H $ we have
\begin{equation*}
\begin{split}
\langle( {}^{t\!}S \circ S_r^{\scalast})(\phi) , h \rangle
&= \widehat{\epsilon} \big( {S(h)}_{(2)} \widehat{t^\ell}\big(\langle \phi , S( {S(h)}_{(1)} ) \rangle \big)
= \big( \nu^{-1} \partial S \big) \big( h^{(1)} t^\ell\big( \langle \phi , S^2( h^{(2)})\rangle \big) \\
&= \epsilon\big( h^{(1)} t^\ell\big(\langle \phi , S^2( h^{(2)})\rangle\big) \big) \\
&= \epsilon\big( h^{(1)} t^\ell\big(\langle {}^{t\!}S(\phi) , S( h^{(2)})\rangle\big) \big)
= \langle ( S_\ell^{\scalast} \circ {}^{t\!}S)(\phi) , h \rangle,
\end{split}
\end{equation*}
where we used the explicit form \rmref{laterza} of the translation map and the fact
that $S$ is an anti-coring morphism between left and right coproduct, which proves $ {}^{t\!}S \circ S_r^{\scalast} = S_\ell^{\scalast} \circ {}^{t\!}S $ as claimed.
\end{proof}
\begin{remark}
\label{schonschoen}
In general, both maps $ S^{\scalast} $ or $ S_{\scalastd} $ can be thought of as an extension of the notion of antipode for a full Hopf algebroid, in the following sense. As mentioned in Lemma \ref{vitasnella}, the antipode in a full Hopf algebroid $H$ yields a bialgebroid morphism $ S : H^\ell \to {(H^r)}^\op_\coop $. On the other hand, if $ U $ is a left Hopf algebroid, for which $ \due U \lact {} $ and $ U_\ract $ are finitely generated projective as $ A $-modules, then we have a similar situation replacing $( H^\ell , H^r , S) $ with the triple $ ((U^{\scalast})^\op , (U_{\scalastd})_\coop , S^{\scalast} )$, and one might be tempted to define a Hopf algebroid as a triple $(U,V,S)$ of a left resp.\ right bialgebroid $U$ resp.\ $V$, where the underlying ring structure is {\em not\/} the same: this way, the apparent asymmetry of a Hopf algebroid consisting of two coring structures but only one ring structure (that makes it difficult to obtain self-duality) would be somewhat attenuated.
On the other hand, in case a left Hopf algebroid is simultaneously a right Hopf algebroid, by Theorem \ref{criptabalbi} both duals are isomorphic and hence can be seen (under the stated finiteness conditions) as {\em its} dual (right) bialgebroid, which carries a Hopf structure by the results in \cite{Schau:TDATDOAHAAHA}.
\end{remark}
\section{Examples and applications}
\label{lautesGequatsche}
In this section we present some further developments and some applications
to specific examples.
\subsection{Mixed distributive law between duals}
A direct application of the existence of the bialgebroid morphism $ S^{\scalast} $ (or $ S_{\scalastd} $) is to the setup of {\it distributive laws}. Indeed, a particular kind of mixed distributive law (or entwining) in the sense of Beck \cite{Bec:DL} can be constructed via the following recipe. Combining a morphism $(\phi_1, \phi_0) : (V, B) \to (V', B') $ of right (say) bialgebroids with a Hopf-Galois map yields
$$ \chi : \due {V'} {} \bract \otimes^\behhe \due V \lact {} \to \due V {} \bract \otimes_\behhe \due {V'} \blact {}, \quad v' \otimes^\behhe v \mapsto v^{(1)} \otimes_\behhe v' \phi(v^{(2)}),
$$
which can be easily seen to define a mixed distributive law between $ V' $ (thought of as a coalgebra) and $V$ (thought of as an algebra, although its coproduct appears in $\chi$).
Applying this to the two duals of a left bialgebroid $U$ along with $S^{\scalast}$, one obtains
$$ \chi : U_{\scalastd} {}_\bract \otimes^\ahha \due {U^{\scalast}} \lact {} \to \due {U^{\scalast}} {} \bract \otimes_\ahha \due {U_{\scalastd}} \blact {}, \qquad \psi \otimes^\ahha \phi
\mapsto \phi^{(1)} \otimes_\ahha \psi S^{\scalast}(\phi^{(2)})
$$
as a mixed distributive law between $ U^{\scalast} $ and $ U_{\scalastd} $, to which
any standard construction based on mixed distributive laws could be applied.
\subsection{Lie-Rinehart algebras and their jet spaces}
\label{enoteca}
Let $(A, L) $ be a Lie-Rinehart algebra ({\em cf.}~\cite{Rin:DFOGCA}, geometrically a Lie algebroid).
Then its (left) universal enveloping algebra $ V^\ell(L) $ carries not only the structure of a left bialgebroid over the commutative algebra $A$ (see \cite{Xu:QG}) but also of a left Hopf algebroid \cite{KowKra:DAPIACT}: on generators $a \in A$ and $X \in L$, its translation map is given by
\begin{equation}
\label{transl-map_V(L)}
a_+ \otimes_\Aopp a_- = a \otimes_\Aopp 1 , \quad X_+ \otimes_\Aopp X_- = X \otimes_\Aopp 1 - 1 \otimes_\Aopp X.
\end{equation}
Moreover, as $ V^\ell(L) $ is cocommutative, it is also a right Hopf algebroid.
{\em Full\/} Hopf algebroid structures on $ V^\ell(L) $ are in bijection with right $ V^\ell(L) $-module structures on $ A $ which play the r\^ole of possible right counits, expressed by suitable maps
$ \partial : V^\ell(L) \to A $ ({\em cf.}~\cite[\S4.2]{Kow:HAATCT} or \cite{KowPos:TCTOHA} for more information). The corresponding antipode $ S : V^\ell(L) \to V^\ell(L)^\op_\coop $ is then uniquely determined by the prescriptions
\begin{equation}
\label{militeignoto}
S(a) = a , \quad S(X) = -X + \partial(X), \qquad \forall a \in A , \ \forall X \in L,
\end{equation}
on generators. For a general Lie-Rinehart algebra (which does not arise from a Lie algebroid), such a map $\pl$ and hence the antipode might or might not exist.
Let us consider the (right) {\em jet spaces} $ J^r(L) := V^\ell(L)^{\scalast} $ and
$ {}^r\!J(L) := V^\ell(L)_{\scalastd} $. If $ L $ is finitely generated projective as an $ A $-module, then $ J^r(L) $ and $ {}^r\!J(L) $ are right bialgebroids {\em in a suitable topological sense}, as their coproduct takes values in a {\em topological\/} tensor product; concerning this, we quickly recall some non-trivial key facts, referring to \cite{KowPos:TCTOHA, CalVdB:HCAC} for further details.
First, $ V^\ell(L) $ is the direct limit of an increasing bialgebroid filtration ({\em i.e.}, the strict analogue of a bialgebra filtration) of finitely generated projective modules $ V^\ell(L)_n $; it follows that $ J^r(L)$
in turn is the inverse limit of all the $ J^r(L)_n := (V^\ell(L)_n)^{\scalast} $, which are finitely generated projective as well. Similar remarks apply to $ {}^r\!J(L) $. As $ V^\ell(L)_p \cdot V^\ell(L)_q \subseteq V^\ell(L)_{p+q} $ (for all $ p, q \in \N $), the recipe used to define the coproduct in $ U^\scalast $ when $ U $ is a left bialgebroid such
that $ U_\ract $ is
finitely generated $A$-projective (see \S \ref{regnetswirklich?}) can be applied again and yields maps
$$
J^r(L)_n = (V^\ell(L)_n)^{\scalast} {\buildrel {\Delta^{J^r}_{\,n}} \over {\relbar\joinrel\longrightarrow}} \hskip-5pt {\textstyle \sum\limits_{p+q=n}} \hskip-5pt
(V^\ell(L)_p)^{\scalast} \!
{}_{\bract} \otimes_\ahha {}_{\blact}
(V^\ell(L)_q)^{\scalast}
\hskip-1pt = \hskip-6pt {\textstyle \sum\limits_{p+q=n}} \hskip-5pt J^r(L)_p {}_{\bract} \otimes_\ahha {}_{\blact} J^r(L)_q
$$
whose inverse limit $ \Delta^{J^r} \! := \lim\limits_{\leftarrow\joinrel\relbar} \Delta^{J^r}_{\,n} $ is the coproduct of $ J^r(L) $. Similarly, one constructs ``coproduct-like maps'' $ \Delta^{{}^r\!J}_{\,n} $ for the $ {}^r\!J(L)_n := (V^\ell(L)_n)_{\scalastd} $ and then takes their inverse limit $ \Delta^{{}^r\!J} := \lim\limits_{\leftarrow\joinrel\relbar} \Delta^{{}^r\!J}_{\,n} $ as a coproduct for $ {}^r\!J(L) $.
Now, because of the very definition of the $ V^\ell(L)_n $ and of the explicit form \rmref{transl-map_V(L)} of the translation map of $ V^\ell(L) $, one easily finds that the translation map itself (much like the coproduct) maps every $ V^\ell(L)_n $ into $ \sum_{p+q=n} V^\ell(L)_p \otimes_\Aopp V^\ell(L)_q $. Then formula \rmref{sstarup} makes sense again, and thus can be used to produce a well-defined map
$$
S^{\scalast}_n : J^r(L)_n = (V^\ell(L)_n)^{\scalast} \relbar\joinrel\relbar\joinrel\longrightarrow {\big(V^\ell(L)_n\big)}_{\scalast} = {}^r\!J(L)_n.
$$
Moreover, the arguments used in the proof of Theorem \ref{S*-morphism} to show that $ S^\scalast $ preserves the coproduct apply again in the present situation, and yield a commutative diagram
\begin{equation}
\label{diagr_Sstar-Delta}
\begin{gathered}
\xymatrix{
J^r(L)_n \ar@{->}[d]_{S_n^{\scalast}} \ar[rr]^{\Delta^{J^r}_{\,n} \hskip33pt} & & {\textstyle \sum\limits_{p+q=n}} \hskip-3pt J^r(L)_p \,
{}_{\bract} \otimes_\ahha {}_{\blact}
J^r(L)_q \ar@{->}[d]^{{\sum\limits_{p+q=n}} \hskip-1pt S_p^{\scalast} \otimes S_q^{\scalast}} \\
{}^r\!J(L)_n \ar[rr]_{\Delta^{{}^r\!J}_{\,n} \hskip33pt} & & \hskip-1pt {\textstyle \sum\limits_{p+q=n}} \hskip-3pt {}^r\!J(L)_p \,
{}_{\bract} \otimes_\ahha {}_{\blact}
{}^r\!J(L)_q }
\end{gathered}
\end{equation}
\noindent
Taking the inverse limit of all these $ S^{\scalast}_n $ we get a well-defined (continuous) map
$$
S^{\scalast} : J^r(L) = V^\ell(L)^{\scalast} \relbar\joinrel\relbar\joinrel\longrightarrow V^\ell(L)_{\scalast} = {}^r\!J(L).
$$
It follows by construction that this map necessarily coincides with the same name map in \S\ref{linking}, hence it respects all $ \Ae $-ring structure maps of $ J^r(L) $ and $ {}^r\!J(L) $ as well as their counits; from \rmref{diagr_Sstar-Delta} follows that this map also respects the coproduct on both sides. All in all, this means that $ S^{\scalast} $ is a morphism of (topological) bialgebroids.
As $ V^\ell(L) $ is also a right Hopf algebroid, \S\ref{linking} also provides a map
$
S_{\scalastd} : {}^r\!J(L) \to J^r(L),
$
which again turns out to be a morphism of (topological) bialgebroids, inverse to $ S^{\scalast} $. The outcome is that
\vskip1pt
\centerline{ \it Theorem \ref{S*-morphism} holds true (in full strength) for $ U = V^\ell(L) $ }
\vskip1pt
\noindent
(replacing the formulation ``morphism of right bialgebroids'' by ``morphism of topological right bialgebroids''),
although the left bialgebroid $ V^\ell(L) $ does not comply with the finiteness assumptions required (in general) for that result.
Finally, note that both $ J^r(L) $ and $ {}^r\!J(L) $ are commutative (because $ V^\ell(L) $ is cocommutative), so they are also left bialgebroids. Identifying $ J^r(L) $ as the coopposite of
$ {}^r\!J(L) $ and with the cocommutativity of $ V^\ell(L) $, one finds that $ S^{\scalast} $ and $ S_{\scalastd} $ are equal and yield an {\em antipode\/} for $ J^r(L) $, which in this way becomes a full Hopf algebroid.
In other words, Proposition \ref{dual-cocomm_Hopf} holds true for $ U = V^\ell(L) $ and $ U^{\scalast} = J^r(L) = {}^r\!J(L)_\coop = (U_{\scalastd})_\coop $, although $ V^\ell(L) $ is {\em not\/} finitely generated projective.
\subsubsection{Difference between $S^*$ and $ {}^{t\!}S$}
In this specific example, one can explicitly observe the difference between $S^{\scalast}$ and the transpose of the antipode $S$ on $V^\ell(L)$ in \rmref{militeignoto}.
Apart from the fact mentioned above that $S^{\scalast}$ always exists while ${}^{t\!}S$ does not, this is already clear on an abstract level since these are maps of different nature as pointed out in Theorem \ref{movimentoallaforza}. Nevertheless, one directly sees here that with respect to the $A$-module structures coming from left and right multiplication in $V^\ell(L)$, the map $S^\scalast(\phi)$ is left $A$-linear whereas $ {}^{t\!}S(\phi)$ is $A$-linear from the right, for $\phi \in V^\ell(L)^{\scalast}$.
Evaluating both maps on an element in $L \subset V^\ell(L)$, one obtains
\begin{equation*}
\label{tSxVL}
{}^{t\!}S(\phi)(X) = - \phi(X) + \partial(X) \phi(1) \qquad \qquad \forall \phi \in V^\ell(L)^{\scalast}, X \in L,
\end{equation*}
on one hand, and on the other hand:
$$
S^{\scalast}(\phi)(X) = - \phi(X) + X(\phi(1)) \qquad \qquad \forall \phi \in V^\ell(L)^{\scalast} , X \in L,
$$
where $L \to \Der(A,A), \ X \mapsto \{ a \mapsto X(a)\}$ denotes the anchor of the Lie-Rinehart algebra $(A,L)$. Using the property $Xa - aX = X(a)$ with respect to the product in $V^\ell(L)$ as well as the right $A$-linearity of $\partial$, one obtains $\pl(aX) = \pl(X) a - X(a)$ and therefore
$
{}^{t\!}S(\phi)(X) - S^{\scalast}(\phi)(X) = \partial(\phi(1)X),
$
which in general does not vanish.
\subsection{Examples from quantisation}
In this section, we adapt our main constructions and results to a different setup, that of quantisations of universal enveloping algebras (of Lie-Rinehart algebras) and other associated objects. In particular, this means that we deal with yet another kind of topological bialgebroids, so that we have to clarify the nature of these objects and how the analysis and results of
the preceding sections
fits to this modified context.
\begin{definition} \label{def-q-bialgd}
\label{def-QUEAd}
Let $ \big( U, A, s^\ell, t^\ell, m, \Delta , \epsilon \big) $ be a left (resp.\ right) bialgebroid.
A {\em quantisation of $U$} (or {\em quantum bialgebroid\/}) is a {\em topological\/} left (resp.\ right) bialgebroid $ \big( U_h, A_h , s^\ell_h , t^\ell_h , m_h , \Delta_h , \epsilon_h \big) $ over a topological $ k[[h]] $-algebra $ A_h $ such that:
\begin{enumerate}
\compactlist{99}
\item
$ A_h $ is isomorphic to $ A[[h]] $ as a topological $ k[[h]] $-module, and this isomorphism induces an algebra isomorphism $ A_h \big/ h A_h \cong A[[h]] \big/ hA[[h]] \cong A $;
\item
$ U_h $ is isomorphic to $ U[[h]] $ as a topological $ k[[h]] $-module;
\item
$ U_h \big/ hU_h \cong U[[h]] \big/ hU[[h]] $ is isomorphic to $U $ as a left $ A $-bialgebroid via the isomorphism $ A_h \big/ hA_h \cong A[[h]] \big/ hA[[h]] \cong A $ mentioned in (i);
\item
the coproduct $ \Delta_h $ of $ U_h $ takes values in $ U_h \widehat{\times}_{\ahha_h} U_h $, where
\begin{equation*}
\qqquad {U_h} \widehat{\times}_{\ahha_h} {U_h} :=
\big\{ {\textstyle \sum_i} u_ i \otimes u'_i \in
{U_h}_\ract \, \widehat{\otimes}_{\ahha_h} \due {U_h} \lact {}
\mid {\textstyle \sum_i} (a \blact u_i) \otimes u'_i = {\textstyle \sum_i} u_i \otimes (u'_i \blacktriangleleft a) \big\}
\end{equation*}
is the {\em Takeuchi-Sweedler product}, and where
$
{U_h}_\ract \, \widehat{\otimes}_{\ahha_h} \due {U_h} \lact {}
$
denotes the completion of
$
{U_h}_\ract \, {\otimes}_{\ahha_h} \due {U_h} \lact {}
$
with respect to the $ h $-adic topology.
\end{enumerate}
In this setting, we shall say that $ U_h $ is a {\em quantisation}, or {\em quantum deformation}, of $ U $.
\end{definition}
\begin{remark} \quad
\label{rem.s_quant-def.s}
{\it (a)}\, The notions of quantum left or right Hopf algebroid are defined replacing the ordinary tensor product by a suitable completion, just as for $ J^r(L) $ above.
{\it (b)}\, When dealing with $ k[[h]] $-modules, any morphism ({\em i.e.}, $ k[[h]] $-linear map) is automatically continuous for the $ h $-adic topology on the source and target $ k[[h]] $-module; we shall tacitly use this fact with no further mention. In particular, for a quantum bialgebroid $ U_h $ both its (full linear) duals $ (U_h)^{\scalast} $ and $ (U_h)_{\scalastd} $ are also {\em topological duals}.
{\it (c)}\,
For a left bialgebroid $ U $ with a quantisation $ U_h $, assume that $ U $ is also a left Hopf algebroid. Then $ U_h $ is automatically a left Hopf algebroid (in a topological sense) as well by a standard argument in deformation theory:
by assumption, we have $ U_h \cong U[[h]] $ as modules over $ A_h \cong A[[h]] $; from this isomorphism one deduces similar isomorphisms for modules of homomorphisms or tensor products of modules. Moreover --- because $ U_h \Big/ h U_h \cong U $ as bialgebroids ---, all bialgebroid structure maps of $ U_h $ taken modulo $ h $ reduce to the same name structure maps of $ U $. Now, for the (topological) left bialgebroid $ U_h $ we have a well-defined Hopf-Galois map
$$
(\ga_\ell)_h : \due {U_h} \blact {} \widehat{\otimes}_{{\scriptscriptstyle{A^{\rm op}_h}}} \, {U_h}_\ract \to {U_h}_\ract \, \widehat{\otimes}_{\ahha_h} \, \due {U_h} \lact {} , \quad
u \, \widehat{\otimes}_{A^{\rm op}_h} v \mapsto u_{(1)} \, \widehat{\otimes}_{\ahha_h} \, u_{(2)} v,
$$
which belongs to
$ \Hom_{k[[h]]}\big( \due {U_h} \blact {} \widehat{\otimes}_{{\scriptscriptstyle{A^{\rm op}_h}}} {U_h}_\ract , {U_h}_\ract \widehat{\otimes}_{\ahha_h} \due {U_h} \lact {} \big) $: as mentioned above, this module is isomorphic to
$ \Hom_k\big( \due U \blact {} \otimes_{\Aopp} U_\ract, U_\ract \otimes_\ahha \due U \lact {} \big)[[h]] $, so that $ {(\ga_\ell)}_h $ expands as $ (\ga_\ell)_h = \sum_{n \in \N} a_n h^n $ for some $ a_n \in \Hom_k\big( \due U \blact {} \otimes_{\Aopp} U_\ract , U_\ract \otimes_\ahha \due U \lact {} \big) $.
In addition, as all structure maps of $ U_h $ modulo $ h $ are just those of $ U $, one has $ \ga_\ell = (\ga_\ell)_h \mod h = a_0 $. But $ U $ was a left Hopf algebroid, hence $ \ga_\ell = a_0 $ is invertible, and therefore $ (\ga_\ell)_h = \sum_{n \in \N} a_n h^n $ is invertible too, so that $ U_h $ is a left Hopf algebroid as well.
\end{remark}
\begin{free text}{\bf Universal enveloping algebras and deformations.}
As in \cite{CheGav:DFFQG}, one can consider a quantum deformation $ V^\ell(L)_h $ of $ V^\ell(L) $: as the latter is both a left and right Hopf algebroid, the same holds true for $ V^\ell(L)_h $ as well, by Remark \ref{rem.s_quant-def.s} {\it (c)\/} above.
On the other hand, the dual (right) bialgebroids $ J^r(L)_h := (V^\ell(L)_h)^{\scalast} $ and
$ {}^r{\!}J(L)_h = (V^\ell(L)_h)_{\scalastd} $ are deformations of $ J^r(L) = V^\ell(L)^{\scalast} = (V^\ell(L)_{\scalastd})_\coop $. This common ``limit'' is a full Hopf algebroid (with bijective antipode) by the above, hence in particular it is a left and right Hopf algebroid with respect to the underlying right bialgebroid structure. It then follows that the same is true for the right bialgebroids $ J^r(L)_h $ and $ {}^r{\!}J(L)_h $, but usually they are not full Hopf algebroids. Nonetheless, we can apply our constructions of \S\ref{linking} to $ U_h := V^\ell(L)_h $ and find the maps $ S^{\scalast} $ and $ S_{\scalastd} $, as we now shortly explain.
By construction, the maps $ S^{\scalast} $ and $ S_{\scalastd} $ as in (\ref{sstarup}) and (\ref{sstardown}) are given in terms of structure maps and translation maps of the (non-topological) bialgebroid $ U $: when $ U $ is replaced by $ U_h $, all those maps are continuous, hence both definitions still make sense and provide maps
$ S^{\scalast} : (U_h)^{\scalast} \to (U_h)_{\scalastd} $ and $ S_{\scalastd} : (U_h)_{\scalastd} \to (U_h)^{\scalast} $ as announced. Once these maps are properly defined (for $ U_h = V^\ell(L)_h $), the proof of all their properties still works untouched (all arguments and calculations make sense and go through in the proper setup of topological bialgebroids). In particular, Theorem \ref{criptabalbi} then assures that the two deformations $ J^r(L)_h := (U_h)^{\scalast} $ and $ {}^r{\!}J(L)_h := (U_h)_{\scalastd} $ of $ V^\ell(L)^{\scalast} = (V^\ell(L)_{\scalastd})_\coop $ are isomorphic (as right bialgebroids) via $ S^{\scalast} $ and $ S_{\scalastd} $.
\end{free text}
\subsection{Cases where a dualising module exists}
\label{heisz}
In this section, we will come back to the situation of dualising modules as in \S\ref{malgenauerhinsehen} by investigating their (deformation) quantisation. To this end, we first need to introduce some extra notation, terminology, and definitions with respect to decreasing filtrations; see, for example, \cite{Che:DPFQG, Schn:AITDM} for further basic results and details.
Let $A$ be an algebra endowed with a decreasing filtration $(F_nA)_{n \in \N}$ and consider a filtered $FA$-module denoted by $FM$, whereas its underlying $A$-module will be denoted by $M$. If $FM$ and $FN$ are two filtered $FA$-modules, then
a filtered morphism $Fu : FM \to FN$ is a morphism $u :M \to N$ of the underlying $A$-modules such that $u(F_sM) \subset F_s N$. A filtered morphism $Fu : FM \to FN$ is {\em strict} if it satisfies
$u(F_{s}M)=u(M)\cap F_{s}N$. An exact sequence of $FA$-modules is a sequence
\begin{equation}
\label{effervescentenaturale}
FM \buildrel{Fu}\over \longrightarrow FN
\buildrel{Fv} \over \longrightarrow FP
\end{equation}
such that $\Ker(F_s v) = \mathrm{Im}\,(F_s u) $, where $ F_s v := v{\big|}_{F_s N} $ and $ F_s u := u{\big|}_{F_s M} $; hence $ Fu $ is strict. If moreover $ Fv $ is also strict, then \rmref{effervescentenaturale} is a called a {\em strict exact sequence}.
The filtration of a filtered module gives rise to a topology and even a metric if the filtered module is separated, that is, if $\bigcap_{n \in \N} F_nM=\{0\} $.
For any $r \in \Z$ and for any $FA$-module $FM$, we define the
{\em shifted module $FM(r)$} as the module $M$ endowed with the filtration
$( F_{s+r}M )_{s \in \Z}$.
An $FA$-module is called {\em finite free} if isomorphic to an
$FA$-module of the type
$\bigoplus_{i=1}^{p}FA(-d_{i})$, where
$d_{1},\dots , d_{p} \in\Z$. An $FA$-module $FM$ is called {\em of finite
type} if one can find
$m_{1} \in F_{d_{1}}M, \dots, m_{p}\in F_{d_{p}}M$ such that any
$m \in F_{d}M$ may be written as
$$
m={\displaystyle \sum_{i=1}^{p}a_{d-d_{i}}}m_{i},
$$
where $a_{d-d_{i}} \in F_{d-d_{i}}A$.
We will be dealing with the case where $M$ is a $k[[h]]$-module and
$F_nM = h^n M$, the so-called {\em $h$-adic filtration}.
\begin{remark}
The existence of a translation map if $U_h$ is a left or right Hopf algebroid makes it possible to endow
\begin{itemize}
\item[--]
$\Hom$-spaces with values in a $h$-adic complete space, and
\item[--]
complete tensor products of $U_h$-modules
\end{itemize}
with further natural $U_h$-module structures. Let us make this explicit for the cases we will use, {\em i.e.}, adapt Proposition \ref{structures}.
If ${\mathcal P}_h$ is a right $U_h$-module and $N_h$ is a left $U_h$-module, then
$_{\blact}{\mathcal P}_h \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} {N_h}_{\ract}$ is endowed with a right $U_h$-module structure as follows:
if $u \in U_h$, then
$u_+ \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} u_{-} \in {}_\blact U_h \, \widehat{\otimes}_{{\scriptscriptstyle{A^{\rm op}_h}}} {U_h}_\ract$
can be written as
$u_+\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} u_- = \lim\limits_{n \to \infty} u_{+,n}\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} u_{-,n}$.
For
$x\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} y \in
{}_{\blact}{\mathcal P}_h\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} {N_h}_{\ract}$,
one defines
$$
(x\otimes_{k[[h]]}y) u := \lim\limits_{n \to \infty} x u_{+,n}\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} u_{-,n} y
\in
{}_{\blact}{\mathcal P}_h \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} {N_h}_{\ract}.
$$
As
$
\lim\limits_{n \to \infty} t^\ell(a)u_{+,n}\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}}u_{-,n} = \lim\limits_{n \to \infty} u_{+,n}\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}}u_{-,n}t^\ell(a)
$,
we have thus defined a right action of $U_h$ on
$_{\blact}{\mathcal P}_h\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} {N_h}_{\ract}$.
If ${\mathcal P}_h$ and $N_h$ are right $U_h$-modules, then
$ \Hom_{{{\scriptscriptstyle{A^{\rm op}_h}}}}({\mathcal P}_h, N_h)$ is endowed with a left $U_h$-module structure as follows: if
$u_{[+]}\otimes^{\ahha_h} u_{[-]} =\lim\limits_{n \to \infty} u_{[+],n} \otimes^{\ahha_h} u_{[-],n}
\in U_{h \bract} \, \widehat{\otimes}^{A_h} {}_\lact U_h$,
one sets for
$ \phi \in \Hom_{A_h^{\text{op}}}({\mathcal P}_h, N_h) $ and $ u \in U _h , \, p \in P_h $,
$$
u_n \phi (p) :=\phi \left ( p u_{[+],n}\right )u_{[-],n},
$$
and argues similarly as above that this defines, indeed, a left $U_h$-action on $\Hom_{A_h^{\text{op}}}({\mathcal P}_h, N_h)$.
\end{remark}
\begin{lemma}
Let $(U_h,A_h)$
be a quantum left Hopf algebroid, and let ${\mathcal P}_h$ be a right $U_h$-module such that ${\mathcal P}_{h \blacktriangleleft}$
(respectively $_{\blact}{\mathcal P}_{h}$)
is a finitely generated projective $A_h^{{\rm op}}$-module (resp.~$A_h$-module). Then
\begin{enumerate}
\compactlist{99}
\item
${\mathcal P}_h$ is complete for the $h$-adic topology.
\item
For a right $U_h$-module $N_h$, any element of $\Hom_{{{\scriptscriptstyle{A^{\rm op}_h}}}}({\mathcal P}_h, N_h)$ is continuous if we endow both modules with the $h$-adic topology.
\item
If $N_h$ is a left $U_h$-module that is complete in the $h$-adic topology, then so is the right $U_h$-module
${}_{\blact}{\mathcal P}_h\otimes_{{\scriptscriptstyle{A^{\rm op}_h}}}{N_h}_{\ract}$.
\item
If $N_h$ is a right $U_h$-module that is complete in the $h$-adic topology, then so is the left $U_h$-module
$\Hom_{{{\scriptscriptstyle{A^{\rm op}_h}}}}({\mathcal P}_h, N_h)$.
\end{enumerate}
\end{lemma}
\begin{proof}
If $N_h$ is a right $U_h$-module endowed with the $h$-adic topology, then
the $h$-adic topology on $(N_h)^p$ coincides with the product topology. Thus, if $N_h$ is complete for the $h$-adic topology, then so is $(N_h)^p$.
\begin{enumerate}
\compactlist{99}
\item
As $\cP_h$ is a finitely generated projective $A_h^{{\rm op}}$-module, it is a summand of a free module, which is complete for the $h$-adic topology as $A_h$ is so. Hence $\cP_h$ is complete for the $h$-adic topology.
\item
This is obvious as such a morphism is $k[[h]]$-linear.
\item
$\cP_h$ is a direct summand of a rank $r$ free $A_h^{{\rm op}}$-module $F_h$.
Thus
${}_{\blact}{\mathcal P}_h \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} N_{\ract}$ is a summand of $(N_h)^r$, which
is complete, hence it is itself complete.
\item
The proof of this part is analogous to the proof of (iii).
\end{enumerate}
\end{proof}
In the following, denote by $\cmodu$ resp.\ $\ucmod$ the category of right resp.\ left $U_h$-modules which are
complete for the $h$-adic topology. We then have the following result, analogous to Proposition \ref{left and right modules}:
\begin{proposition}
\label{left and right modules h}
Let $(U_h,A_h)$ be simultaneously a quantum left and right Hopf algebroid.
Assume that there exists a right $ U_h $-module $ \cP_h$, where $ {\cP_h}_{\bract} $
(resp.\ ${}_\blact \cP_h$) is finitely generated projective over $A_h^{{\rm op}}$ (resp.\ $A_h$), such that
\begin{enumerate}
\compactlist{99}
\item
the left $U_h$-module morphism
$$
A_h \to \Hom_{{\scriptscriptstyle{A^{\rm op}_h}}}(\cP_h , \cP_h), \quad a \mapsto \{ p \mapsto a \blact p \}
$$
is an isomorphism of $k[[h]]$-modules;
\item
the evaluation map
\begin{equation*}
\due {\cP_h} \blact {} \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} \Hom_{{\scriptscriptstyle{A^{\rm op}_h}}}(\cP_h , N_h)_{\ract} \to N_h,
\quad p \otimes_{{\scriptscriptstyle{A^{\rm op}_h}}} \phi \mapsto \phi(p)
\end{equation*}
is an isomorphism for any $ N_h \in \cmodu $.
\end{enumerate}
Then
$$
\ucmod \to \cmodu, \quad M_h \mapsto \due {\cP_h} \blact {} \otimes_\Aopp {M_h}_\ract
$$
is an equivalence of categories with quasi inverse given by
$ N'_h \mapsto \Hom_{{\scriptscriptstyle{A^{\rm op}_h}}}(\cP_h , N'_h) $.
\end{proposition}
We will now give an example of such a situation. Consider a left bialgebroid $(U,A)$ and a quantisation $(U_h, A_h)$ of it. Observe that the natural left $U_h$-module structure on $A_h$ quantises that of $U$ on $A$.
\begin{theorem}
\label{maximal exterior power}
Let $(U,A)$ be a left bialgebroid, where $U$ is assumed to be a $k$-Noetherian algebra. Assume that there exists an integer $d$ satisfying
$$
\Ext^i_U(A, U)=
\left \{
\begin{array}{ll}
0 & \mbox{if} \ i \neq d,\\
\Lambda & \mbox{if} \ i=d.
\end{array}
\right .
$$
Then there exists an $A_h$-module $\Lambda_h$ that is a quantisation of $\Lambda$ such that
$$
\Ext^i_{U_h}(A_h, U_h)=
\left \{
\begin{array}{ll}
0 & \mbox{if} \ i \neq d,\\
\Lambda_h & \mbox{if} \ i = d,
\end{array}
\right.
$$
where the right action of $U_h$ on
$\Ext^d_{U_h}(A_h, U_h)$ is a quantisation of the right action of $U$ on $\Ext^d_{U}(A, U)$ given by right multiplication.
\end{theorem}
We remind the reader here that $\Lambda_h$ is $\Lambda[[h]]$ as a $k[[h]]$-module.
This theorem is proven in \cite{Che:DPFQG} in the case where $A_h=k[[h]]$.
For the proof of the general case, we will need the following auxiliary statement:
\begin{lemma}
\label{resolution}
There exists a resolution of the $U_{h}$-module $A_h$ by
finite rank free (filtered) $FU_{h}$-modules
$$
\ldots \buildrel {\partial_{i+1}} \over \longrightarrow FL^{i}
\buildrel{\partial_{i}}\over \longrightarrow \ldots \buildrel{\partial_{2}} \over
\longrightarrow FL^{1} \buildrel{\partial_{1}} \over
\longrightarrow FL^{0} \longrightarrow A_h \longrightarrow \{0\},
$$
where $FL^{i}$ is $(U_h)^{d_i}$ endowed with the $h$-adic filtration
such that the associated graded complex
$$
\ldots GL^{i}\buildrel {G\partial_{i}} \over \longrightarrow
\ldots \to GL^{1}\buildrel {G\partial_{1}} \over \longrightarrow
GL^{0}\longrightarrow A[h] \longrightarrow \{0\}
$$
is a resolution of the $U[h]$-module $A[h]$.
\end{lemma}
\begin{proof}
We will construct the $p$-th module $FL^p$ by induction on $p$: for $p=0$, one may take $FL^0:=U_h$ and $\partial _0:=\epsilon$, endowed with the $h$-adic topology. Assume then that $FL^0, FL^1, \ldots, FL^p$ are already constructed along with $\partial_0, \partial_1, \ldots, \partial_p$. As $FL^p$ is topologically free, the induced filtration and the $h$-adic filtration coincide on
$\Ker \partial_p$. As $\Ker \partial_p$ is closed in $ FL^p $, it is also complete.
This $k[[h]]$-module is topologically free as it is complete for the $h$-adic topology and also torsion free; set
$\Ker \partial_p :=V_p[[h]]$.
Since $GU_h=U[h]$ is Noetherian, the (filtered) algebra $U_h$ is (filtered) Noetherian \cite[Prop.~3.0.7]{Che:DPFQG} and the
$U_h$-module $\Ker \partial_p$ is finitely generated so that the $U$-module $V_p$ is finitely generated as well. Let
$(\overline{v_1}, \ldots , \overline{v_{d_{p+1}}})$ be a generating system of the $U$-module $V_p$ and let
$(v_1, \ldots , v_{d_{p+1}})\in (\Ker \partial_p)^{d_{p+1}}$
be a lift of $(\overline{v_1}, \dots , \overline{v_{d_{p+1}}})$. Moreover, introduce the $U_h$-module morphism
$$
\partial_{p+1}: (U_h)^{d_{p+1}} \to \Ker \partial_p, \quad (u_1,\ldots ,u_{p+1}) \mapsto \sum u_i v_i,
$$
which is a strict
morphism of filtered modules. The filtered exact sequence
$$
(U_h)^{p+1}\buildrel {\partial_{p+1}} \over \longrightarrow (U_h)^p
\buildrel{\partial_{p}}\over \longrightarrow
(U_h)^{p-1}
$$
is strict exact so that the sequence
$$
(GU_h)^{p+1}\buildrel {G\partial_{p+1}} \over \longrightarrow (GU_h)^p
\buildrel{G\partial_{p}} \over
\longrightarrow (GU_h)^{p-1}
$$
is exact ({\em cf.}~\cite[Prop.~3.0.2]{Che:DPFQG}).
\end{proof}
\begin{proof}[Proof of Theorem \ref{maximal exterior power}]
The $\Ext ^\bull_{U_h}(A_h, U_h)$-groups
can be computed via
the complex $M^\bull := \big( \Hom_{U_{h}}(L^{\bullet}, U_h), \partial_{\bullet} \big)$.
Its components are endowed with the natural filtration
$$
F_s\Hom_{U_{h}}(L^{i}, U_{h}) :=
\{\lambda \in \Hom_{U_{h}}(L^{i}, U_{h})
\mid \lambda ( F_{p}L^{i} )\subset
F_{s+p}U_{h}\},
$$
and the right $FA$-modules
$F\Hom_{U_{h}}(L^{i}, U_{h})$ are isomorphic to
$(U_h)^{d_i} $ endowed with the $h$-adic filtration.
On the other hand,
the filtration of the $M^i:= \Hom_{U_{h}}(L^{i}, U_{h})$ induces a filtration on
$\Ext^i_{U_h}(A_h, U_h)$ as follows:
$$
F_s\Ext^i_{U_h}(A_h, U_h) :=
\frac{\Ker \qttr \partial t {}{} i \cap F_s M^{i} + {\rm Im} \qttr \partial t {}{} {i-1}}{{\rm Im} \qttr \partial t {}{} {i-1}}
\simeq
\frac{\Ker \qttr \partial t {}{} i\cap F_s M^{i}}{{\rm Im} \qttr \partial t {}{} {i-1}\cap
F_s M^{i-1}}.
$$
The filtration on the
$\Ext^i_{U_h}(A_h, U_h)$-groups is nothing but the $h$-adic filtration.
Reproducing the proof of \cite{Che:DPFQG}, one can see that:
\begin{itemize}
\item
if $i \neq d$, then
$\Ext^{i}_{U_{h}} (A_h, U_{h})
=\{0\}$;
\item the maps $\qttr \partial t {}{} i$ are strict filtered morphisms;
\item
$\Ext^{d}_{U_{h}} (A_h, U_h )$ is complete for the $h$-adic filtration (as it is a finitely generated $U_h^\op$-module, see \cite{Che:DPFQG}). Moreover,
$\Ext^{d}_{U_{h}}(A_h, U_{h})/h\Ext^{d}_{U_{h}}(A_h,U_{h}) \simeq \Ext^{d}_{U}(A,U)$ as $U^\op$-modules.
\end{itemize}
Let us show that $\Ext^{d}_{U_{h}} (A_h, U_{h} )$ is $h$-torsion free.
Let $[\sigma_d] \in \Ext^{d}_{U_{h}} (A_h, U_{h} )$, where $\sigma_d \in \Ker \qttr \partial t {}{} d$, be an $h$-torsion element in $\Ext^{d}_{U_{h}} (A_h, U_{h})$. There exists a minimal $n \in \mathbb{N}^{\scalast}$ such that
$h^n[\sigma_d]=0$.
Let $\sigma_{d-1} \in \Hom _{U_h}(L^{d-1}, U_h)$ be such that
$h^n \sigma_d = \qttr \partial t {}{} {d-1}(\sigma_{d-1})$. Then, by reduction modulo $h$, one obtains
$\overline{ \qttr \partial t {}{} {d-1}}(\overline{\sigma_{d-1}})=0$ and there exists $\overline{\sigma_{d-2}}$ such that
$\overline{\sigma_{d-1}}=\overline{\partial_{d-2}}\left (\overline{\sigma_{d-2}} \right )$. Let $\sigma_{d-2}$ be a lift of $\overline{\sigma_{d-2}}$. Then there exists $\tau_{d-1}$ such that
$$
\sigma_{d-1}= \qttr \partial t {}{} {d-2} (\sigma_{d-2}) + h \tau_{d-1}.
$$
Hence $h^n\sigma_{d}=h \qttr \partial t {}{} {d-1} (\tau_{d-1})$, which gives
(using the fact that $\Hom_{U_{h}}(L^{d}, U_{h}) $ is topologically free)
$h^{n-1}\sigma_{d}= \qttr \partial t {}{} {d-1} (\tau_{d-1})$.
This contradicts the minimality of $n$ so that
$\Ext^{d}_{U_{h}} (A_h, U_{h})$ is $h$-torsion free.
As $\Ext^{d}_{U_{h}}(A_h, U_{h})$ is complete for the $h$-adic topology and $h$-torsion free, it is topologically free.
\end{proof}
Combining this result with the more general structure theory as in Proposition \ref{left and right modules} resp.\ Proposition \ref{left and right modules h}, one obtains:
\begin{proposition}
\label{tavernadeiquaranta}
Let $U$ satisfy the conditions of Theorem \ref{maximal exterior power}. Assume moreover that
\begin{enumerate}
\item
$A$ is noetherian;
\item
$\Ext_U(A,U)$ is a dualising module for $(U,A)$, {\em i.e.}, satisfies the hypothesis of Proposition
\ref{left and right modules};
\item
$_\blact \Ext_U(A,U)$ is a finitely generated projective $A$-module.
\end{enumerate}
Then
$\cP_h= \Ext_{U_h}^d ( A_h, U_h)$ is a dualising module for $(U_h,A_h)$ and produces
an equivalence between the categories of left resp.\ right complete $U_h$-modules.
\end{proposition}
\begin{remark}
\label{piovera?}
Let $M_h :=M[[h]]$ and $N_h := N[[h]]$ be two $A_h^{\op}$-modules which are topologically free with respect to the $h$-adic topology. Assume moreover that $M_h$ is finitely generated projective over $A^\op_h$;
then $\Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)$ is topologically free and, as said before, is isomorphic to
$\Hom_{\Aopp}(M,N)[[h]]$ as a $k[[h]]$-module:
observe that $\Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)$ is complete for the induced topology as it is a closed subset of the topologically free $k[[h]]$-module
$\Hom_{k[[h]]}(M_h, N_h)$. On the other hand, on $\Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)$, the induced topology coincides with the $h$-adic topology. Hence $\Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)$ is complete
for the $h$-adic topology and
since it is also torsion free, it is topologically free. Let us now show that
$\Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)/h \Hom_{\scriptscriptstyle{A_h^{\op}}}(M_h, N_h)$ is isomorphic to
$\Hom_{A^{\op}}(M, N)$: in fact, there exists an $A_h^{\op}$-module $M_h^\prime$ and a finitely generated free $A_h^{\op}$-module $F_h$
such that
$M_h \oplus M_h^\prime =F_h$. Any element $\phi$ of
$\Hom_\Aopp(M,N)$ can be extended to an element of
$\Hom_\Aopp({F_h/hF_h},N)$, which, in turn, can be lifted to an element of
$\Hom_{\scriptscriptstyle{A_h^{\op}}}(F_h, N_h)$ and produces (by restriction) a lift of $\phi$.
\end{remark}
\begin{proof}[Proof of Proposition \ref{tavernadeiquaranta}]
The module ${\cP_h}_\bract$ is
a finitely generated $A_h^\op$-module as $\cP_\bract := \Ext_U(A,U)_\bract$ is a finitely generated $\Aop$-module
(see Proposition 3.0.5 of the preprint version of \cite{Che:DPFQG}).
Let $N_h$ be a finitely generated $A_h^\op$-module. It can be considered as a filtered $FA_h^\op$-module as follows: one has an epimorphism
$ {\big( A_h^{\op} \big)}^n \, {\buildrel {p} \over \longrightarrow} \, N_h \longrightarrow 0 $,
and we endow $N_h$ with the filtration $p\big( F{\big( A_h^{\op} \big)}^n \big)$.
As ${\cP}_\bract$ is a projective $A^\op$-module, ${\cP}[h]_\bract$ is a projective $A[h]^\op$-module, and
Proposition 3.0.11 of the preprint version of \cite{Che:DPFQG} shows that
$\Ext^i_{A_h^\op}(\cP_h,N_h)=\{0\}$ if $i>0$.
Let now $N_h$ be any $A_h^\op$-module. We have
$N_h =\lim\limits_{\rightarrow} N'_h$, where $N'_h$ runs over all finitely
generated $ A_h^\op $-submodules of $ N_h $. Let $F^{\bullet}$ be a resolution of $\cP$ by
finitely generated free $A_h^{\op}$-modules. We have
\begin{equation*}
\begin{split}
\Ext^{j}_{A_h^\op}(\cP_h,N_h) &= \Ext^{j}_{A_h^\op}(\cP_h,\lim\limits_{\rightarrow}N'_h)
= H^{j} \big( \Hom_{A_h^\op}(F^{\bullet},\lim\limits_{\rightarrow}N'_h)\big)\\
&= H^{j}\big( \lim\limits_{\rightarrow}\Hom_{A_h^\op}(F^{\bullet}, N'_h)\big)
= \lim\limits_{\rightarrow} H^{j} \big( \Hom_{A_h^\op}(F^{\bullet}, N'_h)\big)\\
&= \lim\limits_{\rightarrow} \Ext^{j}_{A_h^\op}(\cP_h, N'_h)
= \{0\},
\end{split}
\end{equation*}
where we used the fact that
the functor $\lim\limits_{\rightarrow}$ is exact because the set of
finitely generated submodules of $M$ is a directed set, {\em cf.}~\cite[Prop.~5.33]{Rot:AITHA}.
Thus we have proven that if $N_h$ is any $A_h^\op$-module, then
$$
\Ext^{j}_{A_h^\op} (\cP_h,N_h)=\{0\} \quad {\rm if } \ j>0.
$$
Consequently, ${\cP_{h}}_\bract$ is a projective $A^\op_h$-module; similarly, ${}_\blact \Ext_{U_h}(A_h,U_h)$ is a projective $A_h^\op$-module.
The assertion with respect to the evaluation map yet is true if $N_h$ is a topologically free $U_h$-module as it is true modulo $h$, see Remark \ref{piovera?}.
Furthermore, the functor $ N_h \mapsto \cP_h \otimes_{\scriptscriptstyle{A_h}}\Hom_{\scriptscriptstyle{A^\op_h}}(\cP_h, N_h) $ is exact as
$\cP_{h \bract} $ resp.\ $_\blact \cP_h$ is a projective $A_h^\op$-module resp.\ $A_h$-module.
Let now $N_h$ be a finitely generated $U_h$-module. Using a finite free resolution of $ N_h $, one can show (by a diagram chase argument) that the evaluation map is an isomorphism (as it is an isomorphism for any component of the resolution).
If $N_h$ is any $U_h$-module instead, one can write $ N_h = \lim\limits_{\rightarrow} N'_h $, where $N'_h$ runs over all finitely
generated submodules of $ N_h $. Since ${\cP}_h$ is a finitely generated $A_h^{\op}$-module, any element
$\phi \in \Hom_{\scriptscriptstyle{A^\op_h}}(\cP_h, N_h)$ can be considered
as an element of $\Hom_{\scriptscriptstyle{A^\op_h}}(\cP_h, N'_h)$ for a well-chosen finitely generated $A_h^{\op}$-module $ N'_h $. Using the finitely generated case, one can see that
the evaluation map is an isomorphism for any $U_h$-module $ N_h $.
As $\cP_h$ is a finitely generated projective $A_h^{\op}$-module,
the natural left $U_h$-module map
$$
A_h \to \Hom_{\scriptscriptstyle{A^\op_h}}(\cP_h, \cP_h),
\quad a \mapsto (p \mapsto a \blact p)
$$
of Proposition \ref{left and right modules h}
is an isomorphism as it is an isomorphism modulo $h$. This concludes the proof.
\end{proof}
\begin{example}
For example, if $A$ is the algebra of regular functions on a smooth affine variety $X$ and $L$ is the Lie-Rinehart algebra of vector fields over $ X $, then $ U = V^\ell(L) $ satisfies the conditions of Theorem \ref{maximal exterior power}. More generally, for any Lie-Rinehart algebra $(A,L)$, where $L$ is finitely generated projective of constant rank $d$ over a Noetherian algebra $A$, the pair $\big(A, V^\ell(L)\big)$ fulfils the conditions of Theorem \ref{maximal exterior power} and one obtains
$\Ext^d_{V^\ell(L)}(A,V^\ell(L)) = \bigwedge_\ahha^d \Hom_\ahha(L,A)$ for the dualising module (see \cite{Che:PDFKALS, Hue:DFLRAATMC} for more details in this direction).
Then, for any quantisation $ V^\ell(L)_h $ of $ V^\ell(L) $,
Proposition \ref{tavernadeiquaranta} leads to an equivalence of categories between left and right
complete $V^\ell(L)_h$-modules.
Examples of quantisations of $ V^\ell(L) $ are given in \cite{CheGav:DFFQG}.
\end{example}
|
\section{Introduction}
Although the classical collaps process suggest that an enough massive star will undergo collapse until a singularity forms, the picture deeply changes when quantum gravity (QG) is taken into account \cite{CV1, HE, CV2, CV3}. It is supposed QG effects will play a crucial role in determining the outcome of gravitational collapse during its final stages. The semiclassical analysis would suggest that information is lost if the black hole (BH) evaporates completely through Hawking radiation. Therefore, the thermal evaporation mechanism leaves probably behind a stable remnant that contains all the information falling into the BH. Hawking \cite{SH} has recently expressed objections to the AMPS firewall \cite{AMPS, AMPSS} and suggested that the correct solution of the AMPS paradox is that event horizons do not form but only apparent horizons. Moreover, Mersini-Houghton \cite{LMH} shows that due to the negative energy Hawking radiation, the collapse of the star stops at a finite radius before the singularity and the BH horizon have formed. The star bounces instead of collapsing to a BH.
Vaz \cite{CV2} proposed a resolution to the AMPS paradox, considering that the collapsing matter does not undergo continuous collapse to a singularity but condenses on the apparent horizon of the BH. He constructed static solutions with no tangential pressures. Every infalling shell of dust is accompanied by the emission of a positive energy shell from the center of the collapsing stellar object and this process of energy extraction from the center continues until the collapse terminates, describing the effect of strong quantum fluctuations close to the center \cite{CV2}.
Our starting point in this paper is Vaz's quasi-classical configuration, namely a spherically-symmetric source that occupies a finite region. Instead of using a dust cloud that condenses into the apparent horizon, our fluid stress tensor possesses a radial pressure $p_{r} = -\rho$ where $\rho$ is its energy density, with no tangential pressures. The metric inside the collapsing ''plasma ball'' is simple, with a curvature singularity at $r = 0$ which, however, is not a part of the spacetime (we have taken, as Vaz \cite{CV2} did, $r \geq r_{0}$, where $r_{0}$ is the region where quantum fluctuations are expected to dominate. In addition, we introduce a surface stress tensor on the outer boundary $r_{b}$ of the star, for the junction conditions to hold. Our proposal leads to the same relation $r_{b} = 2M + r_{0}$ between the boundary radius $r_{b}$ and the Schwarzschild mass $M$ of the star but, however, we found that $r_{0} = M$ and not $r_{0} = 2M$ as the author of \cite{CV2} obtained for his dust ball.
\section{Interior metric}
Vaz \cite{CV2} already found the solutions of Einstein's equations (without $\Lambda$) corresponding to a source located in a finite region. He did not impose any equation of state between the energy density $\rho$ and the radial pressure $p_{r}$ and set the trasversal pressures to zero. However, his interior geometry depends on an extra parameter $\gamma$ and $p_{r}$ has a complicate dependence on the radial variable $r$. To make things more simple, we propose
\begin{equation}
p_{r} = -\rho
\label{2.1}
\end{equation}
as the equation of state of the fluid inside the collapsing star. It is not the case to repeat the steps used by Vaz for obtaining the expressions for $\rho$ and $p_{r}$ and the corresponding interior geometry. With the extra restriction (2.1) the following inner geometry is obtained
\begin{equation}
ds^{2} = -\frac{r_{0}}{r} dt^{2} + \frac{r}{r_{0}}dr^{2} + r^{2} d \Omega^{2}
\label{2.2}
\end{equation}
where $d\Omega^{2}$ stands for the metric on the unit two-sphere and $r_{0}$, as we shall see, is related to the domain where the quantum fluctuations from the central region are significant. As it was noticed in \cite{CV2}, the singularity at $r = 0$ does not create problems as the solution is valid for $r \geq r_{0}$ only.
The energy-momentum tensor looks now as
\begin{equation}
T^{a}_{~b} = diag\left(-\frac{1}{8\pi r^{2}}, - \frac{1}{8\pi r^{2}}, 0, 0\right)
\label{2.3}
\end{equation}
where $a, b$ run from $0$ to $3$, $\rho = -p_{r} = 1/8\pi r^{2},~p_{\theta} = p{\phi} = 0$. Let us note that $T^{a}_{~b}$ from (2.3) satisfies all the energy conditions and $\rho$ and $g_{rr}$ acquire the same expressions as those obtained by Vaz because to get their values the expression of $g_{tt}$ is nowhere used. We also observe that, although the equation of state is of de Sitter type, the metric (2.2) is not de Sitter, due to the vanishing of the tangential pressures. In the domain of interest ($r \geq r_{0}$) the spacetime (2.2) has no singularities and no horizons. For example, the Ricci scalar is $R^{a}_{~a} = 2/r^{2}$ and the Kretschmann scalar is $R_{abcd}R^{abcd} = 4(r^{2} - 2r_{0}r + 3r_{0}^{2})/r^{6}$, with a maximum value of $8/r_{0}^{4}$ at $r = r_{0}$. Moreover, the radial acceleration of a static observer is $a^{r} = -r_{0}/2r^{2} < 0$, so that the gravitational field is repulsive.
The total mass $m(r)$ by the radius $r$ is given by \cite{SW}
\begin{equation}
m(r) = \int{4\pi r^{2}\rho(r)dr} = \frac{r}{2}
\label{2.4}
\end{equation}
as if inside any sphere of radius $r$ were a BH with mass $m(r)$. As far as the Komar energy is concerned, we obtain \cite{TP, HC}
\begin{equation}
W = 2 \int(T_{ab} - \frac{1}{2} g_{ab}T^{c}_{~c})u^{a} u^{b} N\sqrt{det\gamma} d^{3}x = 0,
\label{2.5}
\end{equation}
due to the contribution of the negative radial pressure. $N = \sqrt{-g_{tt}}$ is the lapse function, $u^{a} = (\sqrt{r/r_{0}}, 0, 0, 0)$ is the velocity vector field of a static observer and $det\gamma$ represents the determinant of the spatial metric.
The Misner-Sharp mass $M_{ms}$ is obtained from \cite{CG, NY, HPFT}
\begin{equation}
1 - \frac{2M_{ms}}{r} = g^{ab} \nabla_{a}r ~\nabla_{b}r,
\label{2.6}
\end{equation}
which gives us
\begin{equation}
M_{ms} = \frac{r}{2} - \frac{r_{0}}{2}.
\label{2.7}
\end{equation}
If we denote $r_{b}$ the radius of the boundary of the star, we have $M_{ms}(r_{b}) = r_{b}/2 - r_{0}/2$ which, as we shall see, gives exactly the Schwarzschild mass $M$ measured by an outer observer. In other words, the total mass \cite{SW} $m(r) = M_{0} + M_{ms}$, so that $M_{0} = r_{0}/2$ gives the (negative) contribution of the central mass \cite{CV2}.
\section{Junction conditions}
As we mentioned above, $r_{b}$ gives the outer boundary of the collapsing star. Our next step is to match the interior spacetime (2.2) to the exterior empty space described by the Schwarzschild geometry
\begin{equation}
ds^{2} = -(1 - \frac{2M}{R}) dT^{2} + \frac{1}{1 -\frac{2M}{R}}dR^{2} + R^{2} d \Omega^{2}
\label{3.1}
\end{equation}
where $M$ is the Schwarzschild mass, measured from large distances. Equating the 1st fundamental forms of (2.2) and (3.1), one obtains
\begin{equation}
\frac{r_{0}}{r_{b}}dt^{2} = (1 - \frac{2M}{R_{b}}) dT^{2},~~~\frac{r_{b}}{r_{0}}dr^{2} = \frac{1}{1 -\frac{2M}{R_{b}}}dR^{2},~~~r_{b} = R_{b}
\label{3.2}
\end{equation}
whence
\begin{equation}
\sqrt{\frac{r_{0}}{r_{b}}}t_{b} = \sqrt{1 -\frac{2M}{R_{b}}}T_{b},~~~\sqrt{\frac{r_{b}}{r_{0}}(1 - \frac{2M}{r_{0}})} = 1.
\label{3.3}
\end{equation}
Using the 2nd relation (3.3) we conclude that $T_{b} = t_{b}$ and therefore we have
\begin{equation}
r_{b} = r_{0} + 2M
\label{3.4}
\end{equation}
i.e. the same condition as that infered by Vaz \cite{CV2} (his Eq. 24), but using a different interior geometry. Eq. (3.4) shows that the star radius $r_{b}$ is always greater than the Schwarzschild radius, even at the final stages of the collapsing process. The difference is exactly the radius of the region of strong quantum fluctuations.
For an estimation of $r_{0}$ we call on the 2nd junction condition - the relation between the jump of the extrinsic curvature when the boundary is crossed and the surface stress tensor (the Lanczos equation)
\begin{equation}
[K_{ab}] - h_{ab}[K^{c}_{~c}] = -8\pi S_{ab}
\label{3.5}
\end{equation}
where $[K_{ab}] = K^{+}_{ab} - K^{-}_{ab}$ is the jump of the extrinsic curvature of the $r = r_{b}$ boundary $\Sigma$, $h_{ab} = g_{ab} - n_{a}n_{b}$ is the induced metric on $\Sigma$ and $n_{b}$ is the normal on $\Sigma$, with $n_{b}n^{b} = 1$. The extrinsic curvature tensor is given by
\begin{equation}
K^{\pm}_{ab} = h^{c}_{~a}\nabla_{c}n^{\pm}_{b},
\label{3.6}
\end{equation}
where $+(-)$ refers to the exterior (interior) geometries. With $n^{-}_{a} = (0, \sqrt{r/r_{0}}, 0, 0)$, the inner geometry gives us
\begin{equation}
K^{-}_{tt} = \frac{r_{0}}{2r^{2}}\sqrt{\frac{r_{0}}{r}},~~~K^{-}_{\theta \theta} = \sqrt{r_{0}r},~~~K^{-,a}_{a} = \frac{3}{2r}\sqrt{\frac{r_{0}}{r}},
\label{3.7}
\end{equation}
evaluated at $r = r_{b}$. With $n^{+}_{a} = (0, 1/\sqrt{1 - \frac{2M}{r}}, 0, 0)$, the outer Schwarzschild geometry yields
\begin{equation}
K^{+}_{TT} = -\frac{M}{R^{2}}\sqrt{1 - \frac{2M}{R}},~~~K^{+}_{\theta \theta} = R \sqrt{1 - \frac{2M}{R}},~~~K^{+,a}_{a} = \frac{2M - 3R}{R^{2}\sqrt{1 - \frac{2M}{R}}},
\label{3.8}
\end{equation}
evaluated at $r = r_{b}$. The jump of the extrinsic curvature when $\Sigma$ is crossed yields
\begin{equation}
[K^{a}_{a}] = \frac{1}{2r_{b}}\sqrt{\frac{r_{b}}{r_{0}}}.
\label{3.9}
\end{equation}
In addition, we have
\begin{equation}
[K_{tt}] = -\frac{1}{2r_{b}}\sqrt{\frac{r_{0}}{r_{b}}},~~~[K_{\theta \theta}] = 0.
\label{3.10}
\end{equation}
The next step is to propose a stress tensor on the boundary $\Sigma$. Let $S_{ab}$ be given by the expression
\begin{equation}
S_{ab} = (p_{s} + \sigma )u_{a}u_{b} + p_{s}h_{ab} + \pi_{ab},
\label{3.11}
\end{equation}
where $p_{s} = -\tau$ is the surface pressure, $\sigma$ is the surface energy density, $\tau$ is the surface tension and $ \pi_{ab}$ is the anisotropic stress tensor, with $\pi^{a}_{~a} = 0,~\pi^{a}_{~b}n^{b} = 0$. We consider the surface fluid has to be anisotropic to match the interior of the ''plasma ball'', where $p_{\theta} = 0$. By means of the equations (3.5), (3.9) and (3.10), one obtains
\begin{equation}
S_{tt} = 0,~~~8\pi S_{\theta \theta} = \frac{r_{b}}{2}\sqrt{\frac{r_{b}}{r_{0}}}.
\label{3.12}
\end{equation}
Let us assume now that
\begin{equation}
p_{s} + \sigma = 0
\label{3.13}
\end{equation}
is the equation of state of the surface fluid. It is justified by the repulsive character of the inner gravitational field which led to $a^{r} < 0$, i.e. to keep a test particle at rest we must act on it with a force toward the center. Using now (3.12) and (3.13) in (3.11), we get the set of equations
\begin{equation}
8\pi (r_{b}^{2}p_{s} + \pi_{\theta \theta}) = \frac{r_{b}}{2}\sqrt{\frac{r_{b}}{r_{0}}},~~~\pi_{tt} - \frac{r_{0}}{r_{b}}p_{s} = 0,~~~\pi^{t}_{t} + 2\pi^{\theta}_{ \theta} = 0,
\label{3.14}
\end{equation}
which leads to
\begin{equation}
p_{s} = -\sigma = - \pi^{t}_{t} = 2\pi^{\theta}_{ \theta} = \frac{1}{24\pi r_{b}}\sqrt{\frac{r_{b}}{r_{0}}}
\label{3.15}
\end{equation}
Let us make use now of the Young-Laplace equation \cite{CDR, HC}
\begin{equation}
-[p_{r}] = \tau K^{a}_{~a}.
\label{3.16}
\end{equation}
With $[p_{r}] = p_{r,out} - p_{r,in} = 1/8\pi r^{2}$ from (2.3), $\tau = \sigma$ from (3.15) and $K^{a}_{~a}$ from (3.8), we obtain
\begin{equation}
\sqrt{\frac{r_{b}}{r_{0}}} (2r_{b} - 3M) = 3r_{b}\sqrt{\frac{r_{0}}{r_{b}}}
\label{3.17}
\end{equation}
whence
\begin{equation}
2r_{b} - 3M = 3r_{0}.
\label{3.18}
\end{equation}
Solving for $r_{0}$ in (3.4) and (3.18), we finally come to
\begin{equation}
r_{0} = M.
\label{3.19}
\end{equation}
In other words, $r_{0}$ is half the gravitational radius of the Schwarzschild mass $M$. In terms of the mass $M$ the surface energy density and surface pressure become
\begin{equation}
p_{s} = -\sigma = \frac{1}{24\pi \sqrt{3}M}.
\label{3.20}
\end{equation}
For the Misner-Sharp mass we get $M_{ms} = r_{b}/2 - r_{0}/2 = M$ and the total mass $m(r_{b}) = r_{b}/2 = 3M/2$. To summarize, the negative central mass $M_{0} = r_{0}/2$ is the difference between the total mass $m(r_{b})$ till the boundary and the Misner-Sharp mass. We note that our result does not coincide with $r_{0} = r_{b}/2$ obtained by Vaz \cite{CV2}. His estimation of $r_{0}$ is based on how much energy is extracted from the center during the star collapse (every collapsing shell is accompanied by an inner, outgoing wave which will extract energy from the center). However, in the calculation of the average energy of the outgoing shell, Vaz used a power series developing which seems, in our opinion, to be debatable. For example, the r.h.s. of Eq. (27) of \cite{CV2} is obtained in the limit the shell spacing $\sigma \rightarrow 0$. More precisely, the small parameter is considered to be $\sigma r_{i}$ so that, actually, $\sigma r_{i} << 1$ (private communication from the author of \cite{CV2}). But we must take in fact $\sigma r_{i}/l^{2}_{Pl} << 1$, for to use a dimensionless expansion parameter ($l_{Pl}$ is here the Planck length). But it is hard to conceive a realistic shell to obey such an inequality, with $l_{Pl}^{2}$ at the denominator. Therefore, we consider the region of strong quantum fluctuations occupies half the Schwarzschild radius of the collapsing star.
\section{Conclusions}
Recently Hawking \cite{SH} raised several objections to the firewall formation during the collapse of a black hole. The final stage of the collapsing process is not the central singularity but only an apparent horizon forms. Based on Hawking conjecture, Vaz \cite{CV2, CV3} considers dust collapse that terminates on the apparent horizon. In his model the collapse wave function indicates that there is a process by which energy extraction from the center takes place.
We proposed in this paper a similar mechanism for a no horizon black hole formation by means of a simple interior geometry where the equation of state of the fluid is $p_{r} + \rho = 0$. To satisfy the junction conditions on the boundary of the ''plasma ball'', we introduced a surface stress tensor corresponding to an anisotropic fluid. Using the Young-Laplace equation as an extra condition, we found that the radius $r_{0}$ of the region where there are strong quantum fluctuations is half the gravitational radius of the Schwarzschild BH, contrary to Vaz's result that $r_{0} = 2M$.
|
\section{Introduction}
\setcounter{equation}{0}
Consider a spatial vector field $\vec{V}$. Its transverse---i.e.~divergence free---part $\vec{V}_T$ can be determined by making use of the projection operator\,\footnote{It can be justified by applying the relation $\nabla^2_{\bf x}\,\frac{1}{|{\bf x}-{\bf x}'|}=-4\pi\,\delta({\bf x}-{\bf x}')$ that $P$, as given by (\ref{projT}), does indeed extract the transverse part of $\vec{V}$, i.e.\,$P[\vec{V}] = \vec{V}_T$ satisfies the relation $\vec{\nabla}\cdot \vec{V}_T=0$.} $P$ given as \cite{jackson}
\begin{equation}\label{projT}
P[\vec{V}]=\vec{V}+ \frac{1}{4\pi}\,\vec{\nabla}\cdot\left[ \int\frac{\vec{\nabla}_{\vec{r}{\,}'}\cdot\vec{V}(\vec{r}{\,}') }{ |\vec{r}-\vec{r}{\,}'|} \,d\vec{r}{\,}'\right] \,.
\end{equation}
In the radiation zone instead of this operator a simplified transverse one
\begin{equation}\label{projN}
{\not\hskip-0.1cm P} [\vec{V}] = \vec{V} - (\vec{n}\cdot\vec{V})\,\vec{n}
\end{equation}
is applied, where the spatial unit vector $\vec{n}$ is pointing from the source to the point of observation. Notice that by construction ${\not\hskip-0.1cm P} [\vec{V}]$ is transverse to $\vec{n}$ but, in general, it is not divergence free. Our main concern in this paper is to demonstrate that such a replacement, as opposed to the ``folklore'', may lead to erroneous result. As a measure of the discrepancy of the action of the proper and the simplified transverse projection operators on $\vec{V}$ one may use $\Delta[\vec{V}]$ defined as
\begin{equation}\label{projD}
\Delta[\vec{V}]=P[\vec{V}]-{\not\hskip-0.1cm P} [\vec{V}] = (\vec{n}\cdot\vec{V})\,\vec{n}+\frac{1}{4\pi}\,\vec{\nabla}\cdot\left[ \int\frac{\vec{\nabla}_{\vec{r}{\,}'}\cdot\vec{V}(\vec{r}{\,}') }{ |\vec{r}-\vec{r}{\,}'|}\, d\vec{r}{\,}'\right] \,.
\end{equation}
\medskip
In this paper it will be shown that the replacement of the proper projection operator $P$ by the simplified transverse one ${\not\hskip-0.1cm P}$ may be erroneous in determining the transverse (or Coulomb gauge) part of the vector potential of an electromagnetic filed in certain physically interesting situations. More concretely, explicit examples will be investigated, each with a point charge moving on a predetermined orbit (i.e.\,the back reaction will be left out from the present considerations), and it will be demonstrated that whenever the interminable spatial motion of the source is unbounded with respect to the reference frame of the observer the replacement of the proper projection operator by the simplified transverse one does indeed yield an erroneous result and the error is of the same order as the proper Coulomb gauge vector potential itself even if the comparison is made in the radiation zone.
\medskip
Before proceeding it is important to recall that in the literature two types of asymptotic limits are applied. In the conventional approach the asymptotic limit is meant to be done by picking a $t=const$ time-slice and assuming that the distance of the observation point with respect to the origin tends to infinity. Notice that the retarded time inevitably tends to $-\infty$ in such a limiting process. Thereby, while the involved observation points are getting farther and farther from the observer they report us about earlier and earlier parts of the history of the source.
In the other approach the retarded time $t_{ret}$ is kept constant and the observations are assumed to be done further and further away to the future along a null line originating at a distinguished event of the world line of the source determined by its location at the moment $t_{ret}$. This approach suits more to the investigation of radiative processes and it provides a meaningful determination of the distance dependence of the potential. The infinite limit along the chosen null line corresponds to an ideal endpoint of this line, representing a point---in the conformal setup introduced by Penrose \cite{penrose}---at future null infinity, $\scri^+$. In all of the considered particular cases we shall indicate the limiting behavior in this latter sense.
\medskip
This paper is organized as follows:
Section\,\ref{prel} is to select the class of electromagnetic systems to which our results apply and to recall some of the basic notions and variables we shall use. The justification of the key formulas is presented separately in the Appendix. The asymptotic behavior of the proper projection and its discrepancy from the simplified one for a point-like charge moving on predetermined orbits will be discussed in Section\,\ref{cases}. In particular, the constant velocity and the oscillatory motions, along with their superposition will be investigated in some details. Section\,\ref{con} contains our final remarks.
\section{Preliminaries}\label{prel}
\setcounter{equation}{0}
The electromagnetic field is represented by a $2$-form field $F_{\,\alpha\beta}$ satisfying the Maxwell equations
\begin{equation}\label{m1}
\partial^\alpha F_{\,\alpha\beta}=-4\pi\,J_{\beta}\ \ \ {\rm and}\ \ \ \partial_{[\alpha}{F}_{\,\beta\gamma]}=0\,,
\end{equation}
where $J_{\alpha}$ denotes the electric four current vector.
It is frequently advantageous to represent the electromagnetic field by a vector potential $A_\alpha$ in terms of which the Maxwell tensor is given as
$F_{\alpha\beta}= \partial_\alpha A_{\beta}-\partial_{\beta} A_{\alpha}$, while the field equations read as
\begin{equation}\label{m2}
\partial^\alpha\left(\partial_\alpha A_{\beta}-\partial_{\beta} A_{\alpha}\right) =-4\pi\,J_{\beta}\,.
\end{equation}
The choice of $A_\alpha$ is known to be non-unique and two vector potentials $A_\alpha$ and $A'_\alpha$ are physically equivalent if there exists a real function $\chi$ such that
\begin{equation}\label{vp}
A'_\alpha=A_\alpha + \partial_\alpha \chi\,.
\end{equation}
This freedom is useful in choosing a vector potential suiting to the investigated problem. Start by splitting the vector potential $A_\alpha$ with respect to an inertial reference system\,\footnote{The minus sign in front of the scalar potential is of historical origin. The vector potential $A_\alpha$ itself entered into the discussions much later than the scalar potential.}$^{,}$\footnote{The speed of light will be retained in the equations, and the Gaussian system of units (for its determination see, e.g.\,\cite{jackson}) will be applied, with $\varepsilon_0=1$ and $\mu_0=1$ throughout.}, with coordinates $(t,\vec{r})$ and derivatives $\partial_\alpha=\left(\frac1c\,\partial_t,\vec{\nabla}\right)$, as $A_\alpha=(-\Phi,\vec{A})$. Then, if the vector potential satisfies the Lorenz gauge condition,\,\footnote{ The index raising and lowering is always done by either of the fixed background metrics $\eta_{\alpha\beta}$ or $\delta_{ij}$ of the Minkowski spacetime or the Euclidean space,
respectively. Moreover, Einstein's summation convention is used only for identical upper and lower indices.} i.e. \,$\partial^\alpha A_\alpha^L=\frac1c\,\partial_t\Phi_L + \vec{\nabla}\cdot\vec{A}_L=0$ holds for $A_\alpha^L=(-\Phi_L,\vec{A}_L)$, the Maxwell equations (\ref{m2}) simplify to
\begin{equation}
\left(-\frac1{c^2}\,\partial_t^2+\nabla^2\right)\,A_{\alpha}^L= -4\pi\,J_{\alpha} \,,\label{feemL}
\end{equation}
where $\nabla^2$ stands for the Laplace-Beltrami operator.
The other frequently used gauge is the Coulomb gauge, with vector potential $A_\alpha^C=(-\Phi_C,\vec{A}_C)$. It requires the vanishing of the spatial divergence $\vec{\nabla}\cdot\vec{A}_C$, and then the field equations read as
\begin{eqnarray}
&&\hskip1.73cm \nabla^2\,\Phi_C=-4\pi\, \rho\,,\label{feemCs}\\
&&\left(-\frac1{c^2}\,\partial_t^2+\nabla^2\right)\,\vec{A}_C =-\frac{4\pi}{c}\,\left[\vec{J}
-\frac{1}{4\pi} \vec{\nabla}\left(\partial_t\Phi_C\right)\right]\,,\label{feemCv}
\end{eqnarray}
where $-\rho$ and $\vec{J}$ denote the time and spatial part of the (locally determined) electric four current vector
$J_\alpha$, respectively.
Notice that the transverse electric current
\begin{equation}\label{tc}
\vec{J}_T=\vec{J}-\frac{1}{4\pi} \vec{\nabla}\left(\partial_t\Phi_C\right)\
\end{equation}
extends over all space regardless whether the spatial part $\vec{J}$ of $J_{\alpha}$ is localized or not \cite{jackson}. This is a direct consequence of the fact that the scalar potential $\Phi_C$ is non-local as it is subject to the Poisson equation (\ref{feemCs}) which means that a change in the charge distribution, even if it happens at an astrophysical distance, leads to an instantaneous change in $\Phi_C$.
\subsection{Potentials of moving point charges}
For a point-like source with electric charge $q$ on an orbit\,\footnote{The tangent vector of this orbit can be given as $\vec{v}(t)=\dot{\vec{R}}(t)$ whereas the pertinent spatial electric current vector reads as $\vec{J}=q \,\vec{v}(t)\,\delta^{(3)}\left[\vec{r}-\vec{R}(t)\right]$.} $\vec{R}=\vec{R}(t)$ the Lorenz gauge vector potential can be given in the familiar Li\'enard-Wiechert form as
\begin{align}
\Phi_L(t,\vec{r})={}& \,\frac{q}{|\vec{r}-\vec{R}(t_{ret})|}\,\frac1{1-\vec{\beta}(t_{ret})\,\vec{n}(t_{ret})}
\label{wundt29a} \\
\vec{A}_L(t,\vec{r})={}& \frac{q}{|\vec{r}-\vec{R}(t_{ret})|}\,\frac{\vec{\beta}(t_{ret})}{1-\vec{\beta}(t_{ret})\,\vec{n}(t_{ret})}\,, \label{wundt29b}
\end{align}
where the relation between the retarded time $t_{ret}$ and $t$ is
\begin{equation}\label{tret}
F(t,t_{ret})=t-t_{ret}-\frac{|\vec{r}-\vec{R}(t_{ret})|}{c}=0\,,
\end{equation}
$\vec{\beta}(t_{ret})$ is defined as
\begin{equation}\label{enn}
\vec{\beta}(t_{ret})=\frac{\dot{\vec{R}}(t_{ret})}{c}
\end{equation}
and the unit vector
\begin{equation}\label{enn}
\vec{n}(t_{ret})=\frac{\vec{r}-{\vec{R}}(t_{ret})}{|\vec{r}-\vec{R}(t_{ret})|}
\end{equation}
points from the location of the source $\vec{R}(t_{ret})$ at $t_{ret}$ to the observation point $\vec{r}$.
\medskip
In case of the Coulomb gauge, for a point charge (\ref{feemCs}) implies that
\begin{equation}
\Phi_C(t,\vec{r})=\frac{q}{|\vec{r}-\vec{R}(t)|} \label{wundt32} \,,
\end{equation}
and as shown\,\footnote{In order to pass from the units applied in \cite{wundt} to the Gaussian system of units used in our paper $4\pi\epsilon_0$ has to be replaced by $\frac1c$ in (44b) of \cite{wundt}.} in \cite{wundt}
\begin{equation}
\vec{A}_C(t,\vec{r})=\vec{A}_L(t,\vec{r})-q\,c\, \vec{\nabla}\cdot \left[ \int_{t_{ret}}^t \frac{1}{|\vec{r}-\vec{R}(t')|}\,dt'\right] \,,\label{wundt44b1}
\end{equation}
(see Eq.(44b) of \cite{wundt}).
Note that the second term on the right hand side of (\ref{wundt44b1}) is again an ``action at a distance'' type expression.
Taking into account (\ref{wundt29b}), along with the relation (\ref{GG}) derived in the Appendix, $\vec{A}_C(t,\vec{r})$ can also be given as
\begin{equation}\label{gauss_c}
\vec{A}_C(t,\vec{r})=\frac{q}{|\vec{r}-\vec{R}(t_{ret})|}\,\frac{\vec{\beta}(t_{ret})-\vec{n}(t_{ret})}{1-\vec{\beta}(t_{ret})\,\vec{n}(t_{ret})} + q\:c \int_{t_{ret}}^t \frac{\vec{r}-{\vec{R}}(t')}{|\vec{r}-\vec{R}(t')|^3} \,dt'\,.
\end{equation}
As it is discussed in the introduction---in the case of a time independent vector field---the discrepancy arising by the replacement of the proper projection operator $P$ by the simplified transverse one ${\not\hskip-0.1cm P}$ can be given by (\ref{projD}). In the more general case of a moving point charge, as it is verified by (\ref{GG2}) of the Appendix, in classical electrodynamics the error yielded by such a replacement is
\begin{equation}\label{GG3}
\Delta[\vec{A}]= -\frac{q\,\vec{n}(t_{ret})}{|\vec{r}-{\vec{R}}(t_{ret})|} + q\:c \int_{t_{ret}}^t \frac{\vec{r}-{\vec{R}}(t')}{|\vec{r}-\vec{R}(t')|^3} \,dt'\,.
\end{equation}
\section{Special cases}\label{cases}
\setcounter{equation}{0}
In the succeeding subsections the large distance behavior of $\vec{A}_C$ and $\Delta[\vec{A}]$ will be determined for various motions of a point charge.
\medskip
However, before proceeding it is rewarding to have a glance at some of the technical difficulties related to the evaluation of the integral term in (\ref{gauss_c}) and (\ref{GG3}). As it will be clear soon even in case of simple one-dimensional motions of a point charge the pertinent integrands may be transcendent functions the integrals of which cannot be given in closed form. Nevertheless, if considerations are restricted to the radiation zone, a careful limiting process allows to determine at least the leading order terms accurately.
\medskip
It is remarkable that whenever the motion of the point particle is restricted to a straight line and the field is also evaluated along this line the unsatisfactory behavior of the simplified transverse projection ${\not\hskip-0.1cm P}$ immediately follows. To see this assume that the above mentioned line coincides with the $x$-axis. Then, in virtue of (\ref{wundt29b}), regardless of the specific form of $R_x=R_x(t)$ the simplified transverse projection ${\not\hskip-0.1cm P}$ yields identically zero result whereas the Coulomb gauge vector potential $\vec{A}_C$ is non-zero unless the point charge is at rest.
\medskip
Below the constant velocity motion, the oscillatory motion, and the superposition of these two motions of a point charge will be considered.
By allowing a generic location of the observation point an analogous failure of the simplified transversal projection will be seen to occur whenever the orbit of the point charge is unbounded with respect to the observer's reference frame.
\subsection{Motion with constant velocity}\label{cases1}
\setcounter{equation}{0}
Start with the simplest possible motion of constant velocity. Accordingly, we shall assume\,\footnote{The vanishing of $\vec{R}$ at $t=0$ makes use of the freedom we have in choosing the origin of the reference system. Note also that this freedom is not specific to the particular case considered in this subsection so it will be applied in the other two cases, as well.} that the point charge moves along the $x$-axis with speed $\vec{v}=(\varepsilon v,0,0)$, i.e.
\begin{equation}
\vec{R}=\vec{R}(t)=\vec{v}\,t = \left\{ \begin{array} {l} R_x=\varepsilon\,v \,t \,;\\ R_{\{y,z\}}=0\,, \end{array}
\right.\label{Rx1}
\end{equation}
where $\varepsilon$ is a sign taking the values $+1$ for forward and $-1$ for backward motion, respectively. Note also that hereafter the notation ${\{y,z\}}$ is used to indicate results relevant for the $y$ and $z$ components.
\medskip
Then, by using the notation $\vec{r}=r\,\vec{e}$, where $\vec{e}$ denotes the unit spatial vector pointing from the origin to the location of the observation point, the expressions $\vec{r}-{\vec{R}}(t')$ and $|\vec{r}-{\vec{R}}(t')|$ can
be given as
\begin{equation}
\vec{r}-{\vec{R}}(t')=r\,\vec{e} - \vec{v}\,t'\,,\ \ {\rm and}\ \ \ |\vec{r}-{\vec{R}}(t')|= \sqrt{(r\,e_x - \varepsilon\,v\,t')^2 + r^2\,(1 - e_x^2)}\,,
\end{equation}
where the relation $e_x^2+e_y^2+e_z^2=1$ has been applied, while for the integrals
\begin{align}
{}& \int_{t_{ret}}^t \frac{(\vec{e}-{\vec{R}}(t'))_x}{|\vec{r}-\vec{R}(t')|^3} \,dt' =\frac{1}{\varepsilon\,v}\left[\frac{1}{\sqrt{r^2 + t^2\, v^2 - 2 \,r \,\varepsilon \,e_x \,v\,t}}-\frac{1}{\sqrt{r^2 + t_{ret}^2\, v^2 - 2\, r\,\varepsilon\, e_x\, v\, t_{ret}}}\right]\,,\label{int_unix} \\
{}& \int_{t_{ret}}^t \frac{(\vec{e}-{\vec{R}}(t'))_{\{y,z\}}}{|\vec{r}-\vec{R}(t')|^3} \,dt'\nonumber \\ {}& \hskip1.5cm = -\frac{e_{\{y,z\}}}{r\,\varepsilon\,v\,(1-e_x^2)}\left[\frac{r\,e_x - \varepsilon\,v\,t}{\sqrt{r^2 + t^2\, v^2 - 2 \,r \,\varepsilon \,e_x \,v\,t}}-\frac{r\,e_x - \varepsilon\,v\,t_{ret}}{\sqrt{r^2 + t_{ret}^2\, v^2 - 2\, r\,\varepsilon\, e_x\, v\, t_{ret}}}\right]\label{int_uniyz}
\end{align}
can be seen to hold. Taking into account (\ref{tret}) the time of observation $t$ can be expressed as
\begin{equation}
t = t_{ret} + \frac{\sqrt{(r\,e_x - \varepsilon\,v\,t_{ret})^2 + r^2\,(1 - e_x^2)}}{c}\,
\end{equation}
and determining the leading $\frac1r$-terms of the components of $\vec{A}_C$, relevant for a fixed value $t_{ret}$, we get
\begin{align}
A_{C}^x = &{}-\frac{q}{{r\,\varepsilon\, \beta}}\left[\frac{1-\beta^2}{1 - \varepsilon\,\beta\, e_x} -
\frac{1}{\sqrt{1 + \beta^2 - 2\, \beta\, e_x }}\right]+\mycal{O}(r^{-2})\,,\label{ACx} \\
A_C^{\{y,z\}} = &{} \frac{q}{{r\,\varepsilon\, \beta}}\hskip-.1cm\cdot\hskip-.1cm\frac{ e_{\{y,z\}}\,\left(e_x-\varepsilon\,\beta\right)}{1-e_x^2}
\left[\frac{1}{1-\beta\,\varepsilon\,e_x} - \frac{1}{\sqrt{1 + \beta^2 - 2 \,\varepsilon\,\beta\, e_x }}\right]+\mycal{O}(r^{-2})\,,\label{ACz}
\end{align}
where $\beta=\frac vc$. Note that the specific value of $t_{ret}$ appears only in the higher order terms.
\medskip
By evaluating the generic expression (\ref{GG3}) for $\Delta[\vec{A}]$, using again the same fixed value $t_{ret}$, we get
\begin{align}
\Delta[\vec{A}]^x = {}& -\frac{q}{{r\,\varepsilon\, \beta}} \left[ 1+ \varepsilon\,\beta\, e_x - \frac{1}{\sqrt{
1 + \beta^2 - 2 \,\varepsilon\,\beta\, e_x }}\right]+\mycal{O}(r^{-2})\,,\label{DeltaAx}\\
\Delta[\vec{A}]^{\{y,z\}} = {}& \frac{q}{{r\,\varepsilon\, \beta}}\hskip-.1cm\cdot\hskip-.1cm\frac{ e_{\{y,z\}}}{1-e_x^2}
\left[e_x -\beta\,\varepsilon\left(1 - e_x^2\right) + \frac{\varepsilon\,\beta-e_x}{\sqrt{
1 + \beta^2 - 2 \,\varepsilon\,\beta\, e_x }}\right]+\mycal{O}(r^{-2})\,.\label{DeltaAz}
\end{align}
By inspecting the above relations it gets immediately transparent that $\Delta[\vec{A}]$ is of the same order as $\vec{A}_C=P[\vec{A}_L]$. Nevertheless, one could argue that this example is not representative as there is no electromagnetic radiation associated with the uniform motion of a point charge. To this end it is important to be mentioned that---as it will be demonstrated in subsection \ref{cases3}---an analogous discrepancy of the proper and the simplified transversal projections occurs when a harmonic oscillation is superimposed on the uniform motion and then radiation is also involved.
\subsection{Oscillatory motion}\label{cases2}
Consider now a harmonic oscillation of a point charge. Accordingly, it will be assumed that
\begin{equation}
\vec{R}=\vec{R}(t)= \left\{ \begin{array} {l} R_x=a\,\sin( \omega\,t) \,;\\ R_{\{y,z\}}=0\,, \end{array}
\right.\label{Rx2}
\end{equation}
i.e.\,the charge oscillates around the origin of the reference frame of the observer along the $x$-axis with amplitude $a$.
\medskip
With (\ref{Rx2}) the integral term appearing in (\ref{gauss_c}) and (\ref{GG3}) becomes
\begin{equation}\label{int_osc0}
\int_{t_{ret}}^t \frac{{r_x}-{{R_x}}(t')}{|\vec{r}-\vec{R}(t')|^3} \,dt'=\frac{1}{r^2}\int_{t_{ret}}^t \frac{{e_x} - \frac ar \,\sin( \omega\,t') }{\left[1 - \frac {2 \,\,e_x a}r \,\sin( \omega\,t') + \frac {a^2}{r^2} \,\sin^2( \omega\,t')\right]^{\frac32}}\,dt'\,.
\end{equation}
Our aim as before is to determine the asymptotic behavior of this integral along a null line characterized by a specific value of $t_{ret}$.
In doing so notice first that for any fixed value of $r$ the integrand on the right hand side
\begin{equation}\label{int_osc2}
{\mycal F}(t',r)=\frac{{e_x} - \frac ar \,\sin( \omega\,t') }{\left[1 - \frac {2 \,\,e_x a}r \,\sin( \omega\,t') + \frac {a^2}{r^2} \,\sin^2( \omega\,t')\right]^{\frac32}}
\end{equation}
is a bounded periodic function of $t'$.
Writing now $t -t_{ret}$ as
\begin{equation}\label{tret2}
t -t_{ret}= N\cdot T+\Delta t\,,
\end{equation}
where $N$ is a sufficiently large integer and $\Delta t$ is smaller than $T$, the period of the oscillation, the integral $\int_{t_{ret}}^t {\mycal F}(t',r) \,dt'$ can be given as
\begin{equation}\label{int_osc}
\int_{t_{ret}}^t {\mycal F}(t',r) \,dt'=\sum_{i=1}^{N}\int_{t_{ret}+(i-1)\cdot T}^{t_{ret}+ i\cdot T} {\mycal F}(t',r) \,dt'+ \int_{t_{ret}+N\cdot T}^{t_{ret}+ N\cdot T + \Delta t} {\mycal F}(t',r) \,dt' \,.
\end{equation}
Due to the periodicity of the integrand in $t'$ the integrals on the right hand side of (\ref{int_osc}) are defined with respect to fixed and finite intervals. Thereby, we may apply Theorem 9.42 of \cite{rudin} ensuring that whenever the integral $\int_{a}^b {\mycal F}(t',r) \,dt'$ exists for the closed interval $[a,b]$ in $\mathbb{R}$, and the integrand ${\mycal F}(t',r=\rho^{-1})$ is (at least) a $C^1$ function of $\rho$, the relation
\begin{equation}\label{rudin}
\partial_\rho\left[\int_{a}^b {\mycal F}(t',\rho^{-1}) \,dt'\right]=\int_{a}^b \partial_\rho\left[{\mycal F}(t',\rho^{-1})\right] \,dt'\,,
\end{equation}
holds. This, in particular, implies that as far as we are only interested in the asymptotic behavior of these terms in the $r\rightarrow \infty$ limit and whenever ${\mycal F}(t',r)$ is sufficiently smooth the $\frac1{r}$-series expansion of
\begin{equation}\label{rudin}
\int_{a}^b {\mycal F}(t',r) \,dt'
\end{equation}
is equal to that of the integral of the $\frac1{r}$-series expansion of ${\mycal F}(t',r)$.
By applying these observations and taking into account that the $\frac1{r}$-series expansion of ${\mycal F}(t',r)$ in (\ref{int_osc2}) reads as
\begin{equation}\label{int_osc2_ser}
e_x-\frac ar\, \sin\left(\omega\, t'\right)\, \left[1-3\,e_x^2\right]-\frac {3\,a^2}{2\,r^2}\, \sin^2\left(\omega\, t'\right)\, \left[3-5\,e_x^2\right] + \mycal{O}(r^{-3})
\end{equation}
the integral on the right hand side of (\ref{int_osc0}) can be evaluated.
\medskip
By combining this with (\ref{gauss_c}), and taking into account $N\approx \frac{t -t_{ret}}{T}\approx \frac{r}{c\,T}$, it is straightforward to verify that
\begin{eqnarray}
A_{C}^x &\hskip-.3cm=&\hskip-.3cm -\frac {q\,a\,\omega\,\left[1-e_x^2\right]\,\cos(\omega\,t_{ret})}{r\left[c+a\,\omega\,\cos(\omega\,t_{ret})\,e_x\right]
+ \mycal{O}(r^{-2})\,,\label{ACxosc} \\
A_{C}^{\{y,z\}} &\hskip-.3cm=&\hskip-.3cm \frac {q\,a\,\omega\,e_x\,e_{\{y,z\}}\,\cos(\omega\,t_{ret})}{r\left[c+a\,\omega\,\cos(\omega\,t_{ret})\,e_x\right]}
+ \mycal{O}(r^{-2})\,.\label{ACyzosc}
\end{eqnarray}
\medskip
Similarly, in virtue of (\ref{GG3}), the leading order expressions for the error term $\Delta[\vec{A}]$ can be seen to take the form
\begin{eqnarray}
\hskip-.5cm\Delta[\vec{A}]^x&\hskip-.3cm=&\hskip-.3cm-\frac{2\,q\,a\,e_x^2\sin(\omega\,t_{ret})}{r^2}
+\mycal{O}(r^{-3})\,,\label{DeltaAxosc}\\
\hskip-.5cm\Delta[\vec{A}]^{\{y,z\}}&\hskip-.3cm=&\hskip-.3cm-\frac{2\,q\,a\,e_x\,e_{\{y,z\}}\sin(\omega\,t_{ret})}{r^2}\
+\mycal{O}(r^{-3})\,.\label{DeltaAzosc}
\end{eqnarray}
The last two relations verify that for the case of an oscillatory motion of a point charge with a fixed center of oscillation with respect to the observer's reference system the error $\Delta[\vec{A}]$ is of higher order than the Coulomb gauge vector potential itself---the latter falls off as $\frac1{r}$---hence the error caused by the replacement of the proper projection by the simplified transversal one is negligible in the asymptotic region.
\subsection{Combination of uniform and oscillatory motions}\label{cases3}
In this subsection it will be shown that the discrepancy between the proper and simplified transversal projections also occurs when radiation is present, in particular when a harmonic oscillation is superimposed on the uniform motion. It is noteworthy that this superposition can also be looked upon as an oscillation around a center moving with constant speed. The relevance of this model is suggested by the dynamical character of our Universe.
\medskip
Accordingly, we shall assume that the center of oscillation moves with constant velocity $v$ with respect to the reference system of the observe
\begin{equation}
\vec{R}=\vec{R}(t)= \left\{
\begin{array} {l} R_x=\varepsilon\, v\,t+a\,\sin( \omega\,t) \,;\\ R_{\{y,z\}}=0\,,
\end{array}
\right.\label{Rx3}
\end{equation}
with $\varepsilon$ defined as in (\ref{Rx1}).
\medskip
A glance at the right hand side of
\begin{equation}\label{int_osc02}
\int_{t_{ret}}^t \frac{{r_x}-{{R_x}}(t')}{|\vec{r}-\vec{R}(t')|^3} \,dt'=\frac{1}{r^2}\int_{t_{ret}}^t \frac{{e_x} - \frac {\varepsilon\, v\,t'+ a \,\sin( \omega\,t') }r}{\left[1 - \frac {2 \,\,e_x \left(\varepsilon\, v\,t'+ a \,\sin( \omega\,t') \right)}r + \frac {\left(\varepsilon\, v\,t'+ a \,\sin( \omega\,t') \right)^2}{r^2}\right]^{\frac32}}\,dt'\,.
\end{equation}
makes immediately clear that the determination of the rate of the fall off of this integral is complicated. In this respect it turned out to be rewarding to consider first the integral of the difference
\begin{equation}\label{int_osc03}
\frac{{e_x} - \frac {\varepsilon\, v\,t+ a \:\sin( \omega\,t) }r}{\left[1 - \frac {2 \,\,e_x \left(\varepsilon\, v\,t+ a \:\sin( \omega\,t) \right)}r + \frac {\left(\varepsilon\, v\,t+ a \,\sin( \omega\,t) \right)^2}{r^2}\right]^{\frac32}}- \frac{{e_x} - \frac {\varepsilon\, v\,t }r}{\left[1 - \frac {2 \,\,e_x \,\varepsilon\, v\,t}r + \frac { v^2\,t^2}{r^2}\right]^{\frac32}}\,.
\end{equation}
Notice that the subtracted term is the integrand applied in case of the pure uniform motion. What makes the use of this difference really advantageous is that the amplitude of the oscillation of this difference (for a fixed value of $r$) falls off at least as fast as $\frac1{t^3}$.\,\footnote{To see this a Taylor expansion in powers of $\frac1t$ has been applied keeping $t$ in $\sin(\omega t)$ fixed in $\varepsilon\, v\,t+ a \:\sin( \omega\,t)$ since $|a \:\sin( \omega\,t)|\ll |\varepsilon\, v\,t|$ while $\frac1t\rightarrow 0$. The validity of the $\frac1{t^3}$ approximation has also been verified by numerical evaluations.} This, along with the fact that we have the factor $\frac{1}{r^2}$ in front of the integral on the right hand side of (\ref{int_osc02}) and $t-t_{ret}$ goes as $\frac{r}{c}$ verifies that unless one is interested in the second or higher order contributions in $\frac{1}{r}$ the integral of this difference in (\ref{int_osc03}) may be neglected. Thereby, the leading order of the fall off
rate of the integral
\begin{equation}\label{int_osc04}
\int_{t_{ret}}^t \frac{{r_x}-{{R_x}}(t')}{|\vec{r}-\vec{R}(t')|^3} \,dt'=\frac{1}{r^2}\int_{t_{ret}}^t\frac{{e_x} - \frac {\varepsilon\, v\,t }r}{\left[1 - \frac {2 \,\,e_x \,\varepsilon\, v\,t}r + \frac { v^2\,t^2}{r^2}\right]^{\frac32}} + \mycal{O}(r^{-2})
\end{equation}
is exactly the same as that of (\ref{int_unix}).
\medskip
In determining the fall off rates of $\vec{A}_C$ and $\Delta[\vec{A}]$ one also has to evaluate the first terms of the right hand sides of (\ref{gauss_c}) and (\ref{GG3}). By combining these terms with (\ref{int_osc04}), along with the leading order terms of (\ref{ACx}) and (\ref{ACz})---which will be refereed to as $A_{C, ({\tiny\ref{ACx}})}^x$ and $A_{C, ({\tiny\ref{ACz}})}^{\{y,z\}}$---the leading order terms of $\vec{A}_C$ read as
\begin{align}
A_{C}^x = &{} A_{C}^x{}_{({\tiny\ref{ACx}})} + \frac{q\,a\,\omega\,[1- e_x^2]\, \cos(\omega\,t_{ret})}{r\,(1-\beta\,\varepsilon\,e_x)\,[c-c\,\beta\,\varepsilon\,e_x - a\,\omega\,e_x \cos(\omega\,t_{ret})]}+\mycal{O}(r^{-2}) \,,\label{ACx2} \\
A_C^{\{y,z\}} = &{} A_{C, ({\tiny\ref{ACz}})}^{\{y,z\}} -\frac{q\,a\,\omega\, e_x\, e_{\{y,z\}}\, \cos(\omega\,t_{ret})}{r\,(1-\beta\,\varepsilon\,e_x)\,[c-c\,\beta\,\varepsilon\,e_x - a\,\omega\,e_x \cos(\omega\,t_{ret})]}+\mycal{O}(r^{-2})\,,\label{ACz2}
\end{align}
whereas the leading order terms of the components of $\Delta[\vec{A}]$ are exactly the same as those in (\ref{DeltaAx}) and (\ref{DeltaAz}).
The above relations implies that in the current case, where radiation is also involved, even in the asymptotic region the discrepancy $\Delta[\vec{A}]$ is of the same order as $\vec{A}_C=P[\vec{A}_L]$.
\section{Final remarks}\label{con}
\setcounter{equation}{0}
In electrodynamics one of the most frequently used gauges is the Coulomb gauge. The transformation to this gauge from any other gauge is an important chapter of the lecture courses on electrodynamics. The present note contains a warning about the use of the simplified projection operator (\ref{projN}) when carrying out the transformation.
\medskip
In this paper some properties of the proper and simplified transverse projection operators in electrodynamics were studied. As discussed in the introduction it is widely held that the discrepancy of the action of the proper and of the simplified transversal projections is asymptotically negligible. More precisely, it is usually claimed that although the simplified transverse projection may have some error in the near or intermediate zone, this error should be negligible in the radiation zone (see, e.g.~the paragraph below Eq.\,(3.16) in \cite{jackson2}). In investigating the validity of these expectations the particular cases of the constant velocity motion, the oscillatory motion, and the superposition of these two motions of a point charge were studied in some details.
\medskip
In the case of purely oscillatory motion---i.e.~when the center of oscillation of the point charge is fixed with respect to the observer's reference frame---the generic expectation was verified by our analysis in subsection \ref{cases2}.
\medskip
Note, however, that when the interminable motion of the source is spatially unbounded with respect to the observer's reference frame the situation is different. According to our main result in this case the replacement of the proper projection operator by the simplified transverse one yields, even in the radiation zone, an erroneous result with error of the same order as the proper Coulomb gauge vector potential itself.
\medskip
Having these results it may be reasonable to ask what may be responsible for the failure of our intuition when we apply the simplified transversal projection instead of the proper one. In answering this question it is worth recalling that the simplified projection refers merely to local fields such as the vector potential in Lorenz gauge (\ref{wundt29b}) and the normal vector (\ref{enn}) pointing from the source to the observer. As opposed to this both the Coulomb gauge vector potential and the proper projection operator involve an integral term---see the second terms in (\ref{gauss_c}) and (\ref{GG3})---which is of an action at a distance type. This integral is asymptotically non-negligible when the interminable relative motion of the source and the observer is unbounded. Therefore not the reported discrepancies but their irrelevance in case of bounded motions is remarkable.
\medskip
As an interesting possible implication of the above observations recall that an analogous replacement of the proper projection operator by the simplified transverse one is applied in gravitational wave (GW) physics in determining the metric perturbation in ``transverse traceless'' ($TT$) gauge (see, e.g.~\cite{racz} for more details). Based on the dynamical character of our Universe it is highly probable that the gravitational wave sources are not fixed with respect to our GW detectors. This, in virtue of our results in subsection \ref{cases3}, may necessitate a careful revision of some of the arguments based on the use of the analog of the simplified transverse projection in linearized theory of gravity.
\medskip
Finally, it is worth emphasizing that in electrodynamics all the troubles yielded by the replacement of the proper projection operator by simplified transversal one goes away once gauge independent quantities are applied. Analogously, in the linearized theory of gravity one could avoid all the technical difficulties in describing the propagation of gravitational waves and their effects on GW detectors by using the curvature tensor instead of the gauge dependent metric perturbations (see the discussions in \cite{racz, david}).
\section*{Acknowledgments}
One of us, IR was supported by the European Union and the State of Hungary, co-financed by the European Social Fund in the framework of T\'AMOP-4.2.4.A/2-11/1-2012-0001 ``National Excellence Program''.
|
\section{INTRODUCTION} \label{sec:intro}
The upcoming Hobby-Eberly Telescope Dark Energy eXperiment (HETDEX; Ref. \citenum{Hill08a}) will amass a sample of $\sim0.8$ million Ly$\alpha$\ emitting galaxies (LAEs) that will be used as tracers of large-scale structure for constraining dark energy and measuring its evolution from $1.9 < z < 3.5$. To carry out the 120 night blind spectroscopic survey covering a 420 square degree field (9 Gpc$^{3}$), a revolutionary new multiplexed instrument called VIRUS (the Visible Integral field Replicable Unit Spectrograph; Ref. \citenum{Hill14a}) is being constructed\cite{Tuttle14} for the upgraded 9.2 m Hobby-Eberly Telescope (HET\footnote{The Hobby-Eberly Telescope is operated by McDonald Observatory on behalf of the University of Texas at Austin, the Pennsylvania State University, Ludwig-Maximillians-Universit\"{a}t M\"{u}nchen, and Georg-August-Universit\"{a}t G\"{o}ttingen.}; Ref. \citenum{Hill14b}). The VIRUS array consists of at least 150 copies of a simple fiber-fed integral field spectrograph and is the first optical astronomical instrument to leverage the economies of scale associated with large-scale replication to significantly reduce overall costs. The spectrographs are mechanically built into unit pairs and are fed by dense-pack fiber bundle integral field units (IFU) with 1/3 fill factor, each consisting of 448 fiber optic elements with a core diameter of 266 $\mu$m (1.5$^{\prime\prime}$\ on the sky). Thus, each individual spectrograph images 224 fibers. At least 75 IFUs will be arrayed on the 22$^{\prime}$\ diameter focal plane of the upgraded HET, yielding $\sim33,000$ individual spectra per exposure. Each spectrograph consists of a double-Schmidt optical design with a volume phase holographic (VPH) diffraction grating at the pupil between a $f$/3.33 folded collimator and a $f$/1.25 cryogenic camera. The spectral coverage of VIRUS is $350 < \lambda \mathrm{(nm)} < 550$ at $R = \lambda / \Delta\lambda \approx 700$, which is optimal for measuring the baryonic acoustic oscillation via the Ly$\alpha$\ emission of star-forming galaxies from $1.9 < z < 3.5$. Fig. \ref{fig:VIRUS} shows a rendering of VIRUS and its optical design.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{f1.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:VIRUS}
\textit{a}) A rendering of the upgraded HET showing the large enclosures mounted on either side of the telescope structure that contain VIRUS spectrographs. The green cables extending from the prime-focus instrument package at the top of the telescope structure to the enclosures are large bundles of fiber optics. \textit{b}) Close view of two enclosures, each containing an 8$\times$3 array of VIRUS units (48 spectrographs). \textit{c}) Section view of a single VIRUS pair. \textit{d}) Ray trace for a single VIRUS spectrograph.}
\end{figure}
The VIRUS concept was proven by the Mitchell Spectrograph (formerly known as VIRUS-P; Ref. \citenum{Hill08b}), a single prototype VIRUS spectrograph that has been in use at the McDonald Observatory 2.7 m Harlan J. Smith telescope since 2007. The instrument has excellent throughput in the blue down to 350 nm ($\sim30$\%, excluding the telescope and atmosphere). For VIRUS, better throughput is required to maximize the number of LAE detections in order to achieve the goals of HETDEX. This is especially true at the lowest redshifts of the survey (i.e., at wavelengths in the near-ultraviolet) because the surveyed volume is smaller, the number density of bright LAEs is diminished (e.g., Ref. \citenum{Ciardullo12}), and the atmospheric transmission is quickly decreasing.
An optical component that can be improved in efficiency at these wavelengths is the VPH diffraction grating that is used as the instrument's dispersing element. VPH gratings have become the standard in astronomical spectroscopy as they provide higher diffraction efficiency and versatility over classic surface relief gratings\cite{Barden98}. For an overview of the physics of VPH gratings, we refer the reader to Refs. \citenum{Arns99}, \citenum{Barden00}, and \citenum{Baldry04}. Ref. \citenum{Adams08} discusses the performance of VPH gratings developed for the Mitchell Spectrograph, which at that time pushed the technology to the highest diffraction efficiency achieved at 350 nm ($\sim60$\%). More recently, as shown in Ref. \citenum{Chonis12}, we have developed prototype VPH gratings for VIRUS that have achieved diffraction efficiencies of $\gtrsim70$\% at 350 nm. For VIRUS, a key technological challenge will be to achieve consistency in this high performance standard over a large production suite of 170 gratings.
In this paper, we present the mass production of VPH gratings for VIRUS with a focus on the acceptance testing methodologies and the as-built performance of the suite of 170 gratings. We begin in $\S$\ref{sec:gratingspec} by describing the production design of the gratings. In $\S$\ref{sec:evaluation}, we discuss the performance metrics by which the gratings are judged, and present the design of a custom apparatus that ensures the standardization of our acceptance tests. In $\S$\ref{sec:production}, the grating mass production at Sygygy Optics, LLC is discussed and the resulting performance of the grating suite is presented in $\S$\ref{sec:performance}. Our final thoughts and a summary are provided in $\S$\ref{sec:conclusions}.
\section{VIRUS VPH GRATING PRODUCTION DESIGN} \label{sec:gratingspec}
\begin{figure}[t!]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.92\textwidth]{f2.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:GratingSchem}
\textit{a}) A schematic drawing of the VIRUS VPH diffraction grating, showing a face-on view of the grating and an edge-on view. The substrate diameter is 148 mm and the total thickness of the grating assembly is 16 mm. Note that the grating and epoxy layer thicknesses are exaggerated and are not to scale. The incident angle $\alpha$ and angle of diffraction $\beta$ are shown in addition to the direction of the fringe tilt, indicated in red by $\phi$. \textit{b}) A photograph of a production VIRUS VPH grating.\vspace{5mm}
}
\end{figure}
\input{t1.tex}
A schematic drawing and photograph of a VIRUS production grating can be seen in Fig. \ref{fig:GratingSchem}. The grating assembly has physical dimensions of 148 mm (diameter) $\times$ 16 mm (total thickness). The VPH layer has a 138 mm diameter clear aperture (CA) and is sandwiched between two 8 mm thick, anti-reflection (AR) coated fused silica substrates using an optical grade adhesive. The grating has a fringe density of $930\pm2$ lines mm$^{-1}$, which provides a level of dispersion that is sufficient to cover $350<\lambda\;\mathrm{(nm)}<550$ at spectral order $m=1$. The grating will operate in transmission for unpolarized light. The key properties of the gratings are high diffraction efficiency (especially for the bluer wavelengths) and repeatability of the grating properties from unit to unit. Due to the large number of units required for VIRUS (170 science-grade gratings, plus four witness samples of lesser quality to monitor environmental degradation over the lifetime of the gratings), the gratings were delivered in batch sizes of up to 50 units (but no smaller than 10) over a 12 month time period. To accommodate the large number of units and the expected variation of performance from unit to unit, the required external diffraction efficiency for unpolarized light was defined as a mean over a given delivery batch. Uniformity from unit to unit is promoted by establishing a minimum allowable efficiency for any grating. The batch mean and minimum external diffraction efficiencies are summarized in Table \ref{tab:Efficiency}.
For HETDEX, the ideal peak diffraction efficiency is between $350 < \lambda\; \mathrm{(nm)} < 400$, and we have focused on maximizing the diffraction efficiency over this wavelength range while maintaining a sufficiently wide bandwidth to retain acceptable efficiency towards 550 nm. Given the parameters listed in this section, the Bragg condition (e.g., see Ref. \citenum{Baldry04}) dictates that the grating angle of incidence $\alpha$ is $\sim10$$^{\circ}$\ to maximize the diffraction efficiency over the desired wavelength range. However, Ref. \citenum{Burgh07} has shown that the location in detector space where the wavelength satisfying the Bragg condition is imaged is also the location of the ``Littrow recombination ghost''. This optical ghost can have a wavelength integrated strength that dominates the signal in a given resolution element of the direct spectrum and can masquerade as a solitary emission line source\cite{Adams08}. For HETDEX, this could contribute significantly to sample contamination since normal LAE detections in our redshift range do not include any other bright emission lines other than Ly$\alpha$\ itself. To mitigate this issue, the fringes are tilted by $\phi = -1$$^{\circ}$\ to decouple the Bragg condition from the Littrow configuration. Note that we have adopted the sign convention of Ref. \citenum{Burgh07} for $\phi$, where negative tilts move the plane of the fringes away from the incident beam, as depicted in Fig. \ref{fig:GratingSchem}$a$. By including $\phi$, we reduce the angle of incidence on the grating substrate to 9$^{\circ}$. This allows the retention of the diffraction efficiency curve similar to a $\alpha=10$$^{\circ}$\ grating with unslanted fringes while pushing the Littrow ghost off the CCD detector as a result of the change in the physical grating angle.
\section{EVALUATING THE VPH GRATING SUITE} \label{sec:evaluation}
\subsection{Evaluation of Prototype Gratings}\label{subsec:prototypegratings}
Prior to engaging in the full-scale production of the VIRUS VPH gratings as described in $\S$\ref{sec:production}, we carried out multiple design studies with several vendors to tune the VPH grating prescription sufficiently to meet the requirements of the HETDEX survey. These efforts have been documented extensively in Ref. \citenum{Adams08} for the Mitchell Spectrograph and Ref. \citenum{Chonis12} for VIRUS. Both of these publications include tests carried out with a custom automated test facility that allowed the full characterization of a grating over a range of $\alpha$, angle of diffraction $\beta$, and $m$. This test-bench was essential in the determination of the final specifications to which the VIRUS production gratings were built.
\subsection{Acceptance Test Requirements} \label{subsec:testrequirements}
Efficiently and consistently testing 170 gratings to validate acceptance metrics requires a different approach than the detailed characterization efforts described in Refs. \citenum{Adams08} and \citenum{Chonis12}. The following is a brief list of requirements for the acceptance tests we have performed on the mass-produced gratings:
\begin{itemize} \itemsep1pt \parskip0pt \parsep0pt
\item The testing method must provide a standardized reference for comparison to specifications.
\item Characterization and testing must take no longer than 10 minutes per grating.
\item Diffraction efficiency measurements at $\alpha=9$$^{\circ}$\ and $m=1$ must be made for $\geq3$ wavelengths within $350 < \lambda \mathrm{(nm)} < 550$.
\item Tests must provide an estimation of the spatial uniformity of the diffraction efficiency across the CA.
\item Tests must provide an estimation of the scattered light in the VPH layer for at least one wavelength in the near-ultraviolet (UV).
\item The test apparatus must be transportable to the vendor's facility and be operable in an office environment.
\item The test apparatus' calibration and data reduction must be transparent to the operator.
\end{itemize}
The test-bench discussed in the previous subsection is too large to be easily transported, and a test of a grating for multiple subapertures to provide a measure of the spatial uniformity is time consuming. Additionally, the ability to test a grating for a range of $\alpha$, $\beta$, and $m$ is not necessary for the acceptance tests. More precisely, the flexibility is undesirable as it may increase the probability of user error over a large number of grating tests. As we describe in the following subsection, we have designed a new apparatus to meet these requirements and efficiently check-out a grating directly on the production line.
\subsection{Design and Operation of the VIRUS Grating Tester}\label{subsec:testerdesign}
The design, operation, and validation of the prototype of the new grating test apparatus was described in Ref. \citenum{Chonis12}. We constructed a second, more refined copy of the apparatus and sent it to Syzygy Optics, LLC to be integrated into the production line they setup for the VIRUS VPH grating contract. Hereafter, we refer to this device as the VIRUS Grating Tester (VGT). The original prototype apparatus has remained at the University of Texas at Austin and has been used to check the results from the VGT and to characterize witness samples.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.8\textwidth]{f3.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:design}
The optomechanical design of the VGT. $a$) Ray trace of the VGT optics with the major components labeled. $b$) A cross-section of the VGT mechanical model shown at the same scale and orientation as the ray trace in panel $a$ with rough dimensions indicated for scale.
}
\end{figure}
Fig. \ref{fig:design} shows the optomechanical design of the VGT. For direct comparison to the external diffraction efficiency specification in Table \ref{tab:Efficiency}, the VGT performs measurements only at 350, 450, and 550 nm. The light from the LED sources first passes through an engineered diffuser, followed by a 300 $\mu$m diameter pinhole that is placed at the focus of a 25 mm diameter $f$/3 singlet. The collimated beam is then stopped down to 12.5 mm in diameter. The 12.5 mm beam size was chosen to avoid the differential vignetting of the dispersed collimated beam after transmission through the grating since the collimator and camera lenses have the same physical diameter (see below). With an emitted FWHM of approximately 25, 25, and 60 nm for the 350, 450 and 550 nm LEDs, respectively, the light sources are far from monochromatic. To centralize the spectral emission, we limit each respective LED by using a 10 nm FWHM narrow-band filter in the collimated beam. The final, effective measurement wavelengths after filtering the LEDs' output are 353.9, 452.3, and 549.5 nm.
After the filter, the collimated beam is split into two paths. The first path is refocused with a second $f$/3 singlet onto a silicon photodiode, which is used to monitor and self-calibrate instabilities in the LED output. The active area of the photodiode is 1.6 mm in diameter, which is large enough to accommodate the chromatic aberration attributed to the singlet lenses. The second path of the collimated beam is incident on the diffraction grating at $\alpha = 9$$^{\circ}$. The diffracted light ($\beta = 15.3$$^{\circ}$\ at 450 nm) is then focused onto a 2/3''-format, 5 megapixel CCD (3.45 $\mu$m square pixels) by a stock 25 mm diameter $f$/1 aspheric lens. The use of a CCD (rather than multiple single pixel detectors for each wavelength, such as photodiodes) is highly beneficial for the VGT. Firstly, the detector alignment and the alignment of the grating in the system (see below) is simplified since the active sensing area is larger than the individual pinhole images and the angular range of the dispersion. Additionally, the fine sampling of the chosen CCD enables accurate centroid determination for verifying the dispersion of each grating, and allows for the use of custom photometric apertures.
Due to imperfections in the grating fabrication process, the diffraction efficiency of a grating may be spatially variable across the CA\cite{Chonis12}. Ideally, one would measure the diffraction efficiency with a collimated beam matched to the CA of the grating. However such an apparatus for our 138 mm diameter gratings would be too large to be portable and would significantly increase costs. To maximize the tested area and provide an empirical estimate of the spatial variability of the diffraction efficiency with our small optical system, we have designed the VGT to easily measure up to 9 subapertures of a grating. Fig. \ref{fig:gratingcell} details the design of a special mounting cell that makes these multiple measurements possible. A marking is placed on the edge of each grating by the vendor to indicate the fringe tilt and fringe direction within $\pm1$$^{\circ}$\ so that the grating can easily be placed in the VGT mounting cell in the correct orientation. A corresponding mark on the edge of the mounting cell is used for visual alignment of the grating and ensures that the placement of the focused spot is on the CCD chip. The cell contains two press-fit locating pins which mate to a series of holes and slots on the custom tester base for constraining the rotational alignment of the grating as it is moved between the series of 9 test positions.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=\textwidth]{f4.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:gratingcell}
The mechanical design of the VGT grating mounting cell. \textit{a}) A rendering of the grating cell with major components labeled. \textit{b}) The engraving on the edge of one of the prototype VIRUS gratings that indicates the fringe orientation, including the direction of the fringe tilt. This engraving is aligned with a matching feature on the grating cell's frame for rapid initial alignment of the grating. \textit{c}) A rendering of the VGT base showing the series of holes and slots used for maintaining the rotational alignment of the grating when switching between the 9 subapertures. The positions of the subapertures are indicated by the thin blue circles on the grating face.
}
\end{figure}
The optomechanical design of the tester has been simplified by utilizing off-the-shelf components for 1'' diameter optics, and modifications to these stock parts were made where necessary to meet our needs. Custom aluminum hardware seamlessly mates with these stock components to fix the collimator and camera angles in the same configuration as a VIRUS spectrograph to an accuracy of $\pm0.1$$^{\circ}$. Including its light-tight enclosure that allows accurate testing in a fully-illuminated room, the VGT is compact and can easily fit on an office desk. It has the approximate dimensions of 485 mm tall with a 350$\times$320 mm footprint and weighs $\sim12$ kg.
The LEDs and the comparison photodiode are controlled through a data acquisition unit requiring a single USB connection to a host computer. The CCD camera interfaces with the computer through a Gigabit ethernet port and is powered by an external 12 V DC source. Both the CCD and the comparison photodiode were verified for linearity over the relevant signal levels. The operation of the VGT is controlled through custom Python software that provides near ``push-button'' simplicity for the tester's operation in a command shell environment, and includes the automated reduction of the CCD images, photometry, background subtraction of the photodiode signal, and calculation of diffraction efficiencies. Before shipment to the vendor, an absolute calibration of the VGT was provided by assembling the camera lens and CCD in a ``straight-through'' configuration without the grating and performing CCD photometry on the direct pinhole images of the LEDs. This initial calibration is then refined as needed over time automatically through the continuous monitoring of the LED output by the comparison photodiode, and manually through regular checks of a standardized 930 line mm$^{-1}$ reference grating whose absolute external diffraction efficiency is well-known through measurements made with the flexible test-bench facility discussed in $\S$\ref{subsec:prototypegratings}. The statistical uncertainties in the diffraction efficiency measurements by the VGT are $\pm0.8$\%, $\pm0.8$\%, and $\pm0.1$\% at 350, 450, and 550 nm, respectively.
\subsection{Measurements}\label{subsec:measurements}
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.9\textwidth]{f5.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:constructed}
\textit{a}) A photo of the completed VGT. When in use, a curtain drops over the open face of the enclosure to make the apparatus light-tight. \textit{b}) An image constructed from the sum of three dark subtracted CCD images, each taken with one of three LEDs turned on. Each CCD image was normalized to the peak flux within each indicated circular aperture before the summation. The three circular apertures are used for the photometric measurements from which the diffraction efficiency at each wavelength is calculated. The red elliptical annulus around the 350 nm pinhole image has a maximum angular width of 0.5$^{\circ}$ and was used to measure the near-UV scattering of the gratings.
}
\end{figure}
A photo of the completed VGT can be seen in panel $a$ of Fig. \ref{fig:constructed}. As we demonstrated in Ref. \citenum{Chonis12}, a full test of a grating requires $\lesssim10$ minutes to complete. Thanks to the fixed and rugged nature of the VGT design, the grating acceptance measurements are tailored to verifying the most critical specifications and have become standardized to avoid the unnecessary confusion that results from inconsistent measurement methods. For each VPH grating on the production line, the VGT provides the following key measurements:
\begin{itemize} \itemsep1pt \parskip0pt \parsep0pt
\item \textbf{Average External Diffraction Efficiency:} Each grating is tested at 350, 450, and 550 nm in each of the 9 different subaperture positions across the grating CA. With a 12.5 mm diameter beam size, this allows 7.4\% of the total CA area to be directly tested. The reported external diffraction efficiency for each wavelength that is to be compared with the specification in Table \ref{tab:Efficiency} is calculated as the average over these 9 subapertures. Fig. \ref{fig:constructed}$b$ shows an example of the pinhole images for each wavelength on the VGT CCD and the location of the over-sized apertures used for performing the photometry.
\item \textbf{External Diffraction Efficiency Spatial Uniformity:} Using the data gathered above for the 9 subapertures, an estimate of the spatial uniformity of the external diffraction efficiency can be calculated. For example, a simple metric for estimating the spatial diffraction efficiency uniformity is the difference between the maximum and minimum measured efficiencies at each wavelength.
\item \textbf{Near-UV Scattered Light:} Scattered light within the processed grating layer can be the result of aberrations, reflections, and other imperfections in the optical system used to expose the holographic material in which the grating is formed\cite{Barden01}. Additional sources of scattering are the epoxy layer used to bond the substrates and the surface roughness of the substrates themselves. An example of a grating with particularly bad scattering properties can be found in Fig. 18 of Ref. \citenum{Barden01}. In VIRUS, a grating that scatters a high amount of the incident beam can adversely affect the image quality of the spectrograph leading to an increase in the cross-talk between imaged fibers at the focal plane. Scattering is most pronounced at short wavelengths, so we quantify the worst-case scenario by making the measurement at 350 nm. To increase the signal-to-noise ratio for measuring faint scattered light, the images taken at all 9 subaperture positions are coadded. Two custom photometric apertures are used. First, an inner elliptical aperture is centered on the 350 nm pinhole image with major and minor axes that correspond to the ideal image size as modeled with Zemax, given the dispersion resulting from the non-monochromatic LED output. A second elliptical aperture is also used with the same center and major axis as the first, but with a minor axis that extends an additional angular distance of 0.5$^{\circ}$\ on either side of the inner elliptical aperture. The total flux within the large circular aperture used in the 350 nm diffraction efficiency calculation is first measured, followed by a measurement of the flux contained within the elliptical annulus formed between the two elliptical apertures (see Fig. \ref{fig:constructed}$b$). The goal for science-grade gratings is to have $\lesssim3$\% of the total flux scattered into the elliptical annulus for the grating to be accepted\footnote{The original specification for scattered light by the gratings stated that $\lesssim3$\% of light at $\lambda=350$ nm in a point source can be scattered into a 0.5$^{\circ}$\ solid angle cone around the $m=1$ beam at the design $\alpha$. However, as stated in $\S$\ref{subsec:testerdesign}, the narrowband-filtered LED light sources in the VGT are not monochromatic. As a result, the pinhole images on the CCD are elongated in the dispersion direction, which motivates the use of an elliptical annulus as the scattering aperture.}.
\end{itemize}
The most notable optical property that the VGT does not verify is the transmitted wavefront error (TWE). TWE measurements are made on selected gratings to verify consistency from batch to batch using a Zygo interferometer with a 6'' diameter beam (stopped down to a 138 mm diameter) at $\lambda=632$ nm in a double-pass configuration using a reference mirror. The TWE for a science-grade grating should be $<2$ waves peak-to-valley at 632 nm within the CA, including any spherical wavefront error.
\section{MASS PRODUCTION OF VPH GRATINGS} \label{sec:production}
In this section, the production line process that was used to fabricate the VIRUS VPH gratings at Syzygy Optics, LLC is summarized. For an individual VPH grating, the process begins by preparing a solution of ammonium dichromate and gelatin (i.e., dichromated gelatin; DCG) that serves as the holographic medium. This solution is poured between a glass mold and the base fused silica substrate. To achieve a layer of gelatin that is initially $\sim100$ $\mu$m thick, we attached adhesives to the substrates around the perimeter to serve as shims between the substrate and mold. We then cooled the gelatin until it congealed and removed the substrate from the mold using a releasing agent. The substrate is then dried and subsequently cured in an incubator.
To form a holographic image in the gelatin layer, each cured substrate is exposed to a 457.5 nm coherent light source that has interference in a plane that forms a 1$^{\circ}$\ angle with the gelatin layer. After the exposure, the grating is submerged in photographic fixer and then dehydrated with graded alcohol. The amount of time in the fixer and in each of the alcohol baths can be varied to produce different modulations of the gelatin layer's index of refraction. Upon removal from the final alcohol bath, we dried any remaining liquid from the grating by placing it in an oven for approximately five minutes. To obtain a preliminary analysis of the optical properties of the uncapped grating\footnote{On average, the exposed and processed DCG layer has approximately the same index of refraction as the glass that is used to cap the grating\cite{Barden00}. As a result, the VGT with its fixed 9$^{\circ}$\ angle of incidence can be used to measure a grating with or without the cap substrate in place. The only difference in the measurement without the cap substrate is the lack of an AR coated incident surface, and the lack of losses due to the internal transmittance of the fused silica. Both of these effects can be accounted for to estimate the final diffraction efficiency of the capped grating.}, we took a series of rapid measurements with the VGT to determine if the grating met the minimum required efficiency at 350, 450, and 550 nm. If the grating did not meet the minimum diffraction efficiency specification, the process of fixing and dehydration is immediately modified for the following grating. If necessary, gratings can be reprocessed through the alcohol baths to further modulate the gelatin layer's refractive index. If a grating was not uniform across the CA, however, it would not be eligible for reprocessing. Using the VGT to determine the level of uniformity in-situ prevented the reprocessing of gratings that could not consistently reach the minimum required diffraction efficiency across the CA.
Acceptable gratings were stored in a dry box for 2-3 days before retesting with the VGT to confirm that the grating properties did not significantly vary from the initial measurements. If there were no significant changes after drying, a 4-5 mm ring of the gelatin around the edge of the grating is removed and the grating is capped with the second fused silica substrate using an optical grade glue. The ring around the edge of the grating that is devoid of gelatin fills with the adhesive and encapsulates the diffractive medium. This seals the grating and prevents moisture from entering and altering the material. Once the adhesive sets, the VGT is then used for the final measurements at each of the 9 subapertures before being approved as science-grade. In the following section, we present the final VGT measurements for the 170 science-grade gratings that were fabricated using this production process.
\section{PERFORMANCE OF THE VPH GRATING SUITE} \label{sec:performance}
To be accepted as a science-grade grating, each of the 170 units must meet the basic assembly specifications. These include having a total thickness of $16.0\pm0.5$ mm, a physical diameter of $148.0\pm0.5$ mm, a radial mismatch of the two fused-silica substrates of $<0.5$ mm, and a total wedge of $<30$$^{\prime}$\ and $<10$$^{\prime}$\ perpendicular and parallel to the fringes, respectively. The VPH layer must also have a CA with a diameter $>138.0$ mm, be centered on the base substrate to within $\pm1$ mm, and be free of major bubbles and point defects. Averaged over all 170 science-grade gratings, we have measured the mean defect area to be 1.13 mm$^{2}$ (standard deviation $\sigma = 0.53$ mm$^{2}$). The maximum defect area of any individual science-grade grating is 2.40 mm$^{2}$, which corresponds to only 2.0\% of the area of a single VGT subaperture. As a result, our characterization of the average external diffraction efficiency will not be significantly affected by measuring a subaperture that contains a large bubble or point defect. Each science-grade grating must have no chips within the CA on either substrate, be able to meet a surface finish specification of 60/40 scratch/dig, and have a surface roughness of $<2$ nm within the CA. Finally, each VPH grating assembly was supplied with a unique serial number that can be traced back to a specific manufacturing date and process during production.
\subsection{Average External Diffraction Efficiency}\label{subsec:efficiency}
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth]{f6.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:Results}
The average delivered external diffraction efficiency of the VIRUS VPH grating suite. The shaded gray region in both panels indicates the external diffraction efficiency specification outlined in Table \ref{tab:Efficiency}. \textit{a}) The average VGT diffraction efficiency measurements for each grating are shown as the small colored data points (teal for 350 nm, blue for 450 nm, and green for 550 nm), which have been scattered randomly about the tested wavelengths for visual clarity. The large black data points show the average external diffraction efficiency of the entire grating suite at each of the three measured wavelengths, while the error bars show the standard deviation of the distributions. The black curves show quadratic spline fits to the diffraction efficiency measurements for notable individual gratings (the heavy solid and heavy dashed curves correspond to the best and worst overall gratings, respectively, while the light dashed curve is the grating that performs best at 350 nm with a peak external diffraction efficiency of 77.4\%). \textit{b}) A comparison of the delivered grating suite to 10 gratings simulated with RCWA. The blue curves are quadratic spline fits to the individual VGT grating measurements, while the red curves are quadratic spline fits to the simulated gratings. The large black (red) data points and error bars are the mean diffraction efficiency and standard deviation for the delivered (simulated) grating suites. The dashed red curve indicates a simulated grating that would be rejected as science-grade. The vertical blue (red) lines indicate the average wavelength of peak diffraction efficiency for the delivered (simulated) grating suites. See the text of $\S$\ref{subsec:efficiency} for more details.
}
\end{figure}
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth]{f7.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:EfficiencyHist}
Histograms showing the distribution of the external diffraction efficiency for the delivered VIRUS VPH grating suite, as measured by the VGT. Each histogram represents a cross-section cut along the $y$-axis at each wavelength from Figure \ref{fig:Results}$a$. For each panel, the solid and dashed black vertical line indicates the batch mean and minimum requirements on the external diffraction efficiency, respectively, while the solid red vertical line indicates the measured mean external diffraction efficiency for the entire grating suite.\vspace{5mm}
}
\end{figure}
The external diffraction efficiency results for the delivered science-grade VPH gratings for VIRUS are shown in Fig. \ref{fig:Results}$a$. For each of the three tested wavelengths, we show the individual, spatially averaged VGT diffraction efficiency measurements along with the suite mean and standard deviation averaged over all 170 gratings. In addition, the distribution of the diffraction efficiency for each wavelength is shown in Fig. \ref{fig:EfficiencyHist}. The histograms in that figure represent a cross-section at each wavelength along the $y$-axis of Fig. \ref{fig:Results}$a$. These results are also summarized in Table \ref{tab:Results}. In general, the measured external diffraction efficiency of the VPH grating suite very closely meets or exceeds our batch mean requirement at each of the three tested wavelengths on average. Additionally, the majority of the delivered gratings greatly exceed the minimum external diffraction efficiency specification, with only a single grating dropping below the minimum requirement at 350 nm by 0.2\%.
To quickly compare the individual gratings amongst each other quantitatively, we calculate $\int\!\eta(\lambda)\:d\lambda$ for each grating between $350 < \lambda \mathrm{(nm)} < 550$, where $\eta(\lambda)$ is the function describing the external diffraction efficiency curve. Since we only have measurements at three discrete wavelengths, we interpolate the diffraction efficiency between the measurements using a quadratic spline to determine $\eta(\lambda)$ for each grating. In Fig. \ref{fig:Results}$a$, the heavy solid curve shows the quadratic spline fit to the measurements of the best overall performing grating according to this merit function while the heavy dashed curve represents the worst grating. Despite the large difference in performance of these two gratings at longer wavelengths, both have similar efficiency at 350 nm. This is consistent with the measurements in general (see Table \ref{tab:Results}), as will also be seen in the following subsection on the spatial uniformity of the diffraction efficiency. This consistency is the result of communication with the contractor that high efficiency at 350 nm was the primary goal for the grating performance due to the multiple combined effects in the near-UV that make detecting LAEs difficult (e.g., see $\S$\ref{sec:intro}). Of all the gratings in the suite, the best-performing unit at 350 nm has a spatially averaged external diffraction efficiency of 77.4\%. The quadratic spline fit to this unit's VGT data has been highlighted in Fig. \ref{fig:Results}$a$ in addition to that of the best and worst performing gratings mentioned above. Overall, the production process was very consistent with time, as there is no statistical dependence of the external diffraction efficiency at any wavelength on the date of manufacture.
The range of variation from grating to grating seen in Figs. \ref{fig:Results} and \ref{fig:EfficiencyHist} is due to small differences in the properties of the processed DCG layer in which the grating is formed. A VPH grating can be fully described and its diffraction efficiency modeled at a given $\alpha$ and $m$ (e.g., with a Rigorous Coupled Wave Analysis; RCWA\cite{Gaylord85}) given the following properties: the fringe density, the fringe tilt $\phi$, the DCG layer refractive index $n_{\mathrm{DCG}}$ and its sinusoidal modulation $\Delta n_{\mathrm{DCG}}$, and the DCG layer thickness $d$. Of these properties at a fixed $\alpha$ and $m$, those that significantly affect the diffraction efficiency of the grating are $d$, $\Delta n_{\mathrm{DCG}}$, and $\phi$. To estimate the range of variation in these parameters that match the measured unit-to-unit diffraction efficiency variation in the delivered grating suite, we ran a series of $m=1$ RCWA models based around the targeted parameters for VIRUS. As described in Ref. \citenum{Chonis12}, those parameters are 930 line mm$^{-1}$ fringe density, $\phi=-1$$^{\circ}$, $\alpha=9$$^{\circ}$, $d=5.5$ $\mu$m, and $\Delta n_{\mathrm{DCG}}=0.037$. We assume that $n_{\mathrm{DCG}} = 1.5$ (e.g., Ref. \citenum{Barden00}), and apply factors to the RCWA modeled diffraction efficiency to take into account the transmission through the AR coated fused-silica substrates, the epoxy layer, and the transmittance of the DCG layer itself for typical physical thicknesses\cite{Barden00}. The predicted diffraction efficiency was calculated at 350, 450 and 550 nm to mimic the VGT measurements. In Fig. \ref{fig:Results}$b$, we show a sample suite of 10 RCWA modeled gratings (red quadratic splines) compared to the delivered VIRUS grating suite (blue quadratic splines). The exact values of $d$, $\Delta n_{\mathrm{DCG}}$, and $\phi$ for each RCWA model were chosen from a uniform distribution with a range about the targeted value of $\pm1.0$ $\mu$m, $\pm0.01$, and $\pm0.5$$^{\circ}$, respectively. As can be seen, the range of diffraction efficiency at each measured wavelength in the RCWA model suite qualitatively matches that of the delivered grating suite relatively well. Similar to what we observe for the delivered gratings, the wavelength with the least variation from unit to unit is 350 nm. Additionally, the overall shape of the average modeled diffraction efficiency curve matches that of the delivered grating suite well, with a difference in the average diffraction efficiency peak of only 10.8 nm (see the vertical lines in Fig. \ref{fig:Results}$b$). Of the modeled gratings, one was not considered in the discussion above due to not meeting the diffraction efficiency requirements for a science-grade grating (see the dashed red curve; the fall-off at 350 nm was due to an extremely large $\Delta n_{\mathrm{DCG}}$ coupled with $d$ that was also larger than the targeted value). In general, however, the average diffraction efficiency of the delivered gratings falls systematically short of the RCWA models at each wavelength (also, see Ref. \citenum{Chonis12} and $\S$\ref{subsec:uniformity} below). The exercise outlined above gives an estimation of the precision to which the DCG layer can be processed for modern VPH gratings.
\input{t2.tex}
\subsection{Spatial Uniformity} \label{subsec:uniformity}
From the measurement of the 9 individual subapertures across a grating CA, the VGT provides a measure of the spatial uniformity of the external diffraction efficiency. As labeled in Table \ref{tab:Results}, the ``Total'' spatial variation is a simple measure of uniformity calculated by taking the difference between the maximum and the minimum measured diffraction efficiencies between the 9 subapertures. In Fig. \ref{fig:UniformityHist}, we show the distributions of the total spatial variation for the VIRUS VPH grating suite at each of the three measured wavelengths. As was the case with the average external diffraction efficiency discussed in the previous subsection, the most consistent performance is at 350 nm, where the mean total spatial variation is $\sim2\times$ smaller than at 450 or 550 nm. The 350 nm distribution's standard deviation is also significantly smaller than the other two wavelengths.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth]{f8.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:UniformityHist}
Histograms showing the distribution of different measures of the uniformity of the external diffraction efficiency for the delivered VIRUS VPH grating suite, as measured by the VGT. Each panel corresponds to one of the three measured wavelengths and contains three histograms. The shaded histograms represent the difference between the maximum and minimum diffraction efficiency measured over the 9 VGT subapertures (``Total''). The black histograms represent the difference between the maximum diffraction efficiency measured over the 9 subapertures and the average (``High''). Finally, the red histograms represent the difference between the minimum diffraction efficiency measured over the 9 subapertures and the average (``Low''). To ease the comparison of the ``High'' and ``Low'' distributions, the ``Low'' distribution (which consists entirely of negative values) is shown as positive.
}
\end{figure}
\begin{figure}[ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth]{f9.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:UniformityCorrelation}
Scatter plots of the ``High'' and ``Low'' spatial variation measures as a function of the spatially averaged diffraction efficiency for each VPH grating in the VIRUS suite at each of the three wavelengths measured by the VGT. As described in $\S$\ref{subsec:uniformity}, the ``High'' (``Low'') spatial variation is simply the difference between the maximum (minimum) diffraction efficiency measured over the 9 subapertures and the spatially averaged diffraction efficiency that is plotted on the abscissa. Thus, there are two colored data points plotted for each grating: one above the thick dashed line at $y=0$, and one below. The larger black data points in each panel correspond to the average spatial variation measure calculated for four equally spaced bins of average diffraction efficiency, while the error bars correspond to the standard error of the mean.
}
\end{figure}
In addition to the total spatial variation, we also look at the spatial variation above and below the average external diffraction efficiency. As labeled in Table \ref{tab:Results}, the ``High'' spatial variation is simply the difference between the maximum measured diffraction efficiency among the 9 subapertures and the spatially averaged diffraction efficiency for a given grating. Similarly, the ``Low'' spatial variation is the difference between the minimum measured diffraction efficiency among the 9 subapertures and the spatially averaged diffraction efficiency, and is a negative quantity. The ``High'' (``Low'') distributions for the VIRUS VPH grating suite are also shown in the panels of Fig. \ref{fig:UniformityHist} as the black (red) histograms. In Fig. \ref{fig:UniformityHist}, the ``Low'' distributions are shown as positive to facilitate visual comparison with the ``High'' distributions. By simply looking at the mean and standard deviation of the two distributions for each wavelength (see Table \ref{tab:Results}), it is clear that the distributions significantly differ. This is confirmed by running a two-sample Kolmogorov-Smirnov test on the ``High'' and ``Low'' distributions, which indicates that the null-hypotheses (i.e., the two distributions are drawn from the same parent distribution) can be rejected at each wavelength with very high certainty. At each wavelength, the ``High'' distribution is narrower and has a lower mean than the corresponding ``Low'' distribution. This is likely due to the fact that the maximal diffraction efficiency at a given wavelength for a fixed fringe density and $\alpha$ is achieved for exactly the right combination of properties that describe the processed DCG layer (i.e., primarily $\Delta n_{\mathrm{DCG}}$ and $d$)\cite{Barden00,Baldry04}. Given some distribution of achieved values for these processed DCG layer properties about the targeted values, it is more likely that a combination of non-optimal values will be drawn rather than drawing the exact correct set of values that maximizes the diffraction efficiency. As a result, non-uniformities in the DCG layer processing that cause the diffraction efficiency to vary across the grating CA (e.g., small changes in effective gelatin thickness) are more likely to reduce the diffraction efficiency than boost it about the average in a given subaperture. The result is that the ``High'' distribution is narrower and smaller on average than that for the ``Low'' distribution. Since the spatial variations in DCG layer properties tend to decrease the measured average diffraction efficiency for a given grating, it should not be surprising that the modeled RCWA predictions discussed in the previous subsection and shown in Fig. \ref{fig:Results}$b$ (which does not consider the spatial variation of grating parameters) are systematically higher than the measurements for the delivered gratings.
Fig. \ref{fig:UniformityCorrelation} also supports the aforementioned hypotheses that small changes in the DCG properties as a function of position across the CA tend to scatter individual measurements towards lower efficiency rather than higher efficiency. In this figure, we plot both the ``High'' and ``Low'' spatial variation measures for each grating as a function of the spatially averaged diffraction efficiency for each respective wavelength. In addition, we have calculated the Spearman Rank Correlation Coefficient $\rho_{\mathrm{S}}$ and associated $p$-value for the ``High'' and ``Low'' variation measures separately at each wavelength. As can be seen in Fig. \ref{fig:UniformityCorrelation}, the ``High'' and ``Low'' variation measures are correlated with high statistical significance at 350 and 450 nm such that more uniform gratings have higher spatially averaged diffraction efficiency. However, this trend is not seen at 550 nm. For 350 and 450 nm, the fact that the ``Low'' spatial variation measure increases with increasing average diffraction efficiency is not surprising. What is more interesting is that the ``High'' spatial variation measure \textit{decreases} with increasing average diffraction efficiency. This is likely a result of the highest average diffraction efficiency gratings already having close to the optimal DCG properties that maximize the diffraction efficiency. As a result, non-uniformities in the DCG layer processing are increasingly less likely to result in a better combination of layer properties. The reason these trends are not seen at 550 nm is because it is the furthest tested wavelength from the Bragg condition (from $\S$\ref{sec:gratingspec}, recall that the VIRUS gratings were designed such that the Bragg wavelength is between $350 < \lambda\; \mathrm{(nm)} < 400$; from Fig. \ref{fig:Results}$a$, the average wavelength of the peak diffraction efficiency for the suite is 412.3 nm). As a result, the diffraction efficiency at 550 nm is not maximal, yielding a more equal likelihood that a slight change in DCG properties could either scatter the diffraction efficiency positively or negatively about the average.
\subsection{Near-UV Scattering}\label{subsec:scattering}
Fig. \ref{fig:Scattering}$a$ shows the distribution of the scattering measurements made by the VGT at 350 nm. On average, the fraction of light in the 350 nm pinhole image that is scattered into the VGT scattering aperture (see $\S$\ref{subsec:measurements}) is 3.2\%, which is slightly worse than our desired $\lesssim3$\% specification. With a standard deviation of only 0.37\%, all gratings in the VIRUS VPH grating suite perform comparably, and the amount of observed scattering in the worst performing grating (4.3\%) should not significantly increase the fiber-to-fiber cross-talk on the CCD detector beyond what is acceptable for HETDEX.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.7\textwidth]{f10.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:Scattering}
The scattered light properties of the VIRUS VPH grating suite. $a$) The distribution of the scattered light at 350 nm as measured by the VGT. The teal histogram represents the original measurements, while the red histogram represents the measurements after correction for the scattering time dependence that is suspected to be due to an increasing amount of dust on the VGT optics (see panel $b$). The dashed black vertical line indicates our original specification on the fraction of the 350 nm pinhole image's total flux that could be scattered into the VGT scattering aperture ($\lesssim3$\%; see $\S$\ref{subsec:measurements}). The solid black vertical line indicates the mean of the entire VIRUS VPH grating suite from the original measurements, and the solid red vertical line indicates the mean after the correction. $b$) The measured scattered light at 350 nm as a function of manufacturing date. The manufacturing date is represented as $\Delta t$ in days, which is the time since the completion of the first science-grade grating. The gray line represents a linear fit to the data and is used as an estimation of the increased scattering effect due to the increasing dust deposited on the VGT optics with time. The slope of this linear fit is used to correct the original measurements, and the corrected data are shown as the red histogram in panel $a$.
}
\end{figure}
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.8\textwidth]{f11.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:twe}
Surface/wavefront maps showing the TWE at $\lambda=632$ nm for a typical grating in the VIRUS VPH grating suite. The map on the left shows the $m=1$ TWE while the map on the right shows the $m=0$ TWE. The typical $m=1$ TWE is $1/2$ wave peak-to-valley at $\lambda=632$ nm, which is significantly better than the $\leq2$ wave peak-to-valley requirement for VIRUS.
}
\end{figure}
Of all the properties that were measured by the VGT, the near-UV scattering is the only one that is significantly correlated with the manufacturing date ($\rho_{\mathrm{S}}=0.61$, $p=1.07\times10^{-18}$; see Fig. \ref{fig:Scattering}$b$). Recall that the VGT operates in an office environment rather than a cleanroom. Without clean air in circulation around the VGT during the year over which these tests were carried out, we suspect that the slow increase in the measured scattered light with time is a systematic effect caused by the steady increase of dust deposited on the VGT's lenses and filters, rather than an increase in the actual grating layer scattering. To estimate the average scattering in this scenario, we assume that the VGT optics were clean as of the time of the first science-grade grating's completion, and simply subtract the slope of a linear fit to the scattering data as a function of time. The linear fit can be seen in Fig. \ref{fig:Scattering}$b$ and the resulting histogram of corrected scattering measurements can be found in Fig. \ref{fig:Scattering}$a$. The mean and standard deviation of the corrected distribution is 2.84\% and 0.30\%, respectively. After the correction, 48 gratings ($\sim28$\% of the suite) have a scattering measurement of $>3$\%.
\subsection{Transmitted Wavefront Error}\label{subsec:twe}
Fig. \ref{fig:twe} shows surface/wavefront maps for a typical VIRUS VPH grating as measured at $m=0$ and $m=1$. The surface errors do not appear to be correlated among the two measured spectral orders. The typical TWE at $\lambda=632$ nm for $m=1$ ($m=0$) is $\sim360$ nm ($\sim240$ nm) peak-to-valley. For VIRUS, this $1/2$ wave peak-to-valley performance at $m=1$ is excellent and is significantly better than our $\leq2$ wave peak-to-valley requirement. This is the result of a high performance holographic exposure system in addition to the use of thick substrates to reduce the effect of warping after the cap substrate is finally glued over the processed grating layer.
\section{SUMMARY} \label{sec:conclusions}
In this paper, we have presented the design of the VPH diffraction gratings that have been mass-produced for use in the new VIRUS array of spectrographs for the HET. The grating design was optimized to have high external diffraction efficiency in the near-UV. This is required for VIRUS to maintain sufficient throughput for the HETDEX survey, which aims to constrain dark energy and measure its evolution from $1.9 < z < 3.5$ using LAEs as tracers of large scale structure. One of the principle challenges involved in the production of the suite of 170 gratings is maintaining consistency in the high performance standard required for HETDEX. With such a large number of units, we are also faced with the challenge of efficiently and consistently validating the performance of each of the 170 gratings to ensure that the best quality units are delivered. To perform these tests, we have developed an apparatus that is very effective at providing robust acceptance test results in which measurements of the average external diffraction efficiency, spatial uniformity of the diffraction efficiency, and near-UV scattered light are provided in $\lesssim10$ minutes per grating. We have tested the suite of 170 science-grade gratings and determined that they individually meet or exceed our specifications. At near-UV wavelengths, the average grating in the suite achieves an external diffraction efficiency of $\sim70$\%. As the first optical astronomical instrument to be replicated on such a large scale, the VIRUS project has provided a useful platform on which the production of large aperture VPH gratings for astronomy can be evaluated in a statistical manner.
\acknowledgments
HETDEX is run by the University of Texas at Austin McDonald Observatory and Department of Astronomy with participation from the Ludwig-Maximilians-Universit\"{a}t M\"{u}nchen, Max-Planck-Institut f\"{u}r Extraterrestriche-Physik (MPE), Leibniz-Institut f\"{u}r Astrophysik Potsdam (AIP), Texas A\&M University (TAMU), Pennsylvania State University, Institut f\"{u}r Astrophysik G\"{o}ttingen (IAG), University of Oxford, and Max-Planck-Institut f\"{u}r Astrophysik (MPA). In addition to Institutional support, HETDEX is funded by the National Science Foundation (grant AST-0926815), the State of Texas, the US Air Force (AFRL FA9451-04-2-0355), the Texas Norman Hackerman Advanced Research Program under grants 003658-0005-2006 and 003658-0295-2007, and generous support from private individuals and foundations.
We thank the staffs of McDonald Observatory, AIP, MPE, TAMU, Oxford University Department of Physics, and IAG for their contributions to the development of VIRUS. We also acknowledge Jim Arns of Kaiser Optical Systems, Inc. for useful discussions during the development phase of the VPH gratings for the Mitchell Spectrograph and VIRUS. T.S.C. acknowledges the support of a National Science Foundation Graduate Research Fellowship.
|
\section{Supplementary Materials}
\subsection{Algorithm Schematics}
\textbf{Input:} Matrix of measurements $A \in \mathbb{C}^{M \times N}$, vector of normalised probabilities $\vec{p}$, initial guess matrix $\hat{\rho}_{in}$, SVD threshold parameter $\tau$ such that $0 < \tau < 1$, sparsity parameter $\tau_{\ell}$ such that $0 < \tau_{\ell} < 1$ and stopping condition step size $\delta_s$. In our 17-dimensional state reconstruction we choose $\tau = 0.4$, $\tau_{\ell} = 0.04$ and $\delta_s = |\vec{\rho}_s| \cdot 10^{-3}$.
\textbf{Output:} Recovered matrix $\hat{\rho}_d$
\begin{enumerate}
\item Set $A' = orth(A)$ and $\vec{p}^{\,\prime} = C \vec{p}$ where $C A = A'$
\item Set $\hat{\rho}_{s,0} = \hat{\rho}_{in}$
\item \textbf{For} $k=1:k_{max}$
\item \quad \quad Set $\hat{\rho}_{0,k} = \Gamma_{\tau, \tau_{\ell}}(\hat{\rho}_{s,k-1})$
\item \quad \quad Set $\vec{\rho}_{0,k} =vec( \hat{\rho}_{0,k})$
\item \quad \quad \textbf{For} i=1:M
\item \quad \quad \quad \quad Set $A_i = $ $i_{th}$ row of $A$
\item \quad \quad \quad \quad Set $\vec{\rho}_{i} = P(\vec{\rho}_{i-1},A_i)$
\item \quad \quad \textbf{End}
\item \quad \quad Set $\hat{\rho}_{s,k} = mat(\vec{\rho}_{M})$
\item \quad \quad Set $\delta = | \hat{\rho}_{s,k} - \hat{\rho}_{s,k-1} |$
\item \quad \quad \textbf{If} $\delta \leq \delta_s$ \textbf{Break}
\item \textbf{End}
\end{enumerate}
Here $orth(\cdot)$ is the orthogonalizing operator, $\Gamma_{\tau, \tau_{\ell}}( \cdot )$ is the operator that enforces the desired characteristics described in the results section, $vec(\cdot)$ is the operator that rearranges the elements of a matrix into a vector, $P(v,V)$ denotes the projection of a vector $v$ onto a hyperplane having normal $V$, and $mat(\cdot)$ is the operator that rearranges the elements of a vector into a square matrix.
\subsection{Projection onto a hyperplane}
To project a point $\vec{\rho}_{i-1}$ onto a hyperplane $A^{\prime}_{i}\vec{\rho} = p^{\prime}_{i}$, it is necessary to find the vector $\vec{v}_i$ that has direction $\vec{n}_i$ normal to the hyperplane $A^{\prime}_{i}\vec{\rho}=p^{\prime}_{i}$ and size $k_i$, where $k_i$ is
\begin{align}\label{stcompress4}
k_i = p^{\prime}_{i}-\braket{\vec{n}_i}{\vec{\rho}_{i-1}}.
\end{align}
The desired projection $\vec{\rho}_i$ is then
\begin{align}\label{stcompress5}
\vec{\rho}_i = \vec{\rho}_{i-1} + \vec{v}_i.
\end{align}
\subsection{Finding and correcting $\Delta\vec{\rho}$}
The error vector $\Delta\vec{\rho}$ depends on the experimental error $\Delta p_i$ associated with each of the probabilities and the measurements made to perform the reconstruction. Instead of extending the search to a non-linear space, we make use of the low rank and sparsity information to estimate the error direction, that is, to find a close approximation for the vector $\Delta\vec{\rho}=\vec{\rho}_r - \vec{\rho}_d$, where $\vec{\rho}_r$ is the projection of $\vec{\rho}_d$ onto the linear space intersection of all the hyperplanes. In order to do this we divide the measurements and the corresponding probabilities in subsets $A_{s_i}$ and $\vec{p}_{s_i}$ of sufficient size for our compressive sensing technique to yield convergence to a solution (the size of the subsets depends on the purity of the state and the estimate of the error $\Delta p_i$ on each of the probabilities). We then perform the operation-projection algorithm on each subset separately to find the vectors $\vec{\rho}_{s_i}$, low rank and sparse solutions to the corresponding subsets systems $A_{s_i}\vec{\rho} = \vec{p}_{s_i}$. We hence define a new set of vectors $\Delta \vec{\rho}_i = \vec{\rho}_{s_i} - \vec{\rho}_{d_i}$ where $\vec{\rho}_{d_i}$ is the solution resulting from applying one last set of thresholding operations to $\vec{\rho}_{s_i}$. The sum of the vectors $\Delta \vec{\rho}_i$ is taken as a close approximation for the error vector $\Delta\vec{ \rho}$. We finally compute a correction vector for the probabilities $\Delta p$ that can be subtracted from the measured probabilities $p$. The correction vector is defined as
\begin{align}\label{stcompress6}
\Delta p = A \Delta \vec{\rho}.
\end{align}
Once the probabilities have been corrected as described above, the compressive sensing algorithm can be performed, making use of the set of performed measurements $A$ and the corrected probabilities $p-\Delta p $.
\end{document} |
\section{Introduction}
Some 40 years ago Stephen Hawking predicted that black holes will emit radiation and slowly evaporate due to subtle quantum physics effects~\cite{hawking74a, hawking74b}.
This prediction continues to generate heated debate, both from within the scientific community, and (sometimes) in the popular press.
See for instance Hawking's recent opinion piece regarding the necessity of making careful physical distinctions between the mathematical concepts of \emph{event} horizon and \emph{apparent} horizon~\cite{horizons}:
``The absence of event horizons means that there are no black holes --- in the sense of regimes from which light can't escape to infinity.
There are, however, apparent horizons which persist for a period of time.''
(Hawking's opinion piece was then grossly misrepresented in the popular press, by the simple expedient of mis-quoting a key phrase; and specifically by suppressing crucial subordinate clauses and sentences.)
Related ideas along these lines have been mooted before, both by Hawking himself,
(``... a true event horizon never forms, just an apparent horizon.''~\cite{dublin}),
and (to one degree or another) by many other researchers~\cite{Roman, Ashtekar, Hayward05a,Hayward05b, micro-survey, Hossenfelder:2009, Frolov:2014, Israel:2014, Bardeen:2014}.
Indeed it is well-known that there are \emph{very many} quite \emph{different} types of horizon one can define. At a minimum: the event, apparent, trapping, isolated, dynamical, evolving, causal, Killing, non-Killing, universal, Rindler, particle, cosmological, and putative horizons, \emph{etcetera}. (See for instance~\cite{Booth:2005, Andersson:2005, Killing1, Killing2, Krasinski:2003, Helfer:2011}.)
It is less well appreciated that the precise technical differences between these horizons makes a difference when one is worried about the subtleties of the ``black hole" evaporation process.
These distinctions even make a difference when precisely defining what a ``black hole'' is --- the usual definition in terms of an \emph{event} horizon is mathematically clean, leading to many lovely theorems~\cite{Hell}, but bears little to no resemblance to anything a physicist could actually measure. (Somewhat related issues also afflict cosmology~\cite{lost, Binetruy:2014}.)
So why have \emph{event} horizons dominated so much of this discussion over the last 40 years?
Certainly \emph{event} horizons are known to exist in classical general relativity~\cite{Hell, lost, Binetruy:2014, MTW, Wald, Hartle, Poisson}, but they are extremely delicate \emph{teleological} constructions, somehow implicitly defined by a ``final cause'', requiring that nature inherently tends toward definite ends~\cite{telos}.
Mathematically, one needs to know the entire history of the universe, all the way into the infinite future, and all the way down to any spacelike singularity, to decide whether or not an \emph{event} horizon exists right here and now.
This makes \emph{event} horizons unsuitable for empirical testing in either laboratories or telescopes.
In contrast, \emph{apparent} and \emph{trapping} horizons are defined using local (or at worst quasi-local) measurements, meaning that they are at least in principle suitable for testing in \emph{finite-size laboratories or telescopes}.
\section{Detectability of apparent/ trapping horizons}
Consider for definiteness a dynamic spherically symmetric spacetime. Without any significant loss of generality we can write the metric as:
\begin{eqnarray}
{\mathrm{d}} s^2 &=& - \zeta(r,t)^2 \left\{1-{2m(r,t)\over r}\right\} {\mathrm{d}} t^2 + {{\mathrm{d}} r^2\over1-2m(r,t)/r}
\nonumber\\
&&\qquad\qquad + r^2 ({\mathrm{d}}\theta^2+\sin^2\theta\;{\mathrm{d}}\phi^2).
\end{eqnarray}
Detecting event horizons requires knowledge of $m(r,t)$ throughout the entire spacetime, and in particular into the infinite future. Detecting apparent horizons, (and the closely related trapping horizons), requires one to evaluate the expansion
$\Theta \propto \{1-2m(r,t)/ r\}$
of outward pointing null geodesics, and boils down~\cite{evolving} to making a ``local'' measurement of the quantity $2m(r,t)/r$.
To proceed, it is both useful and important to clarify the meaning of the much abused word ``local'' by carefully distinguishing the concept ``ultra-local" (measurements made at a single point) from ``quasi-local'' (measurements made in a finite-size region of spacetime; that is, using a finite-size laboratory over a finite time interval).
The equivalence principle, (more specifically the local flatness of spacetime), implies that no ultra-local measurement can (even in principle) \emph{ever} detect \emph{any} type of horizon.
(Similarly, no ultra-local measurement can, even in principle, detect spacetime curvature.)
The situation is quite different for quasi-local measurements.
In particular, it is well-known that in any finite-size laboratory one can measure tidal effects, which are controlled by the Riemann tensor.
Then in spherical symmetry, regardless of how one chooses to set up an orthonormal basis in the $r$--$t$ plane, in the $\theta$--$\phi$ plane one has:
\begin{equation}
R_{\hat\theta\hat\phi\hat\theta\hat\phi} = {2m(r,t)\over r^3}.
\end{equation}
The point is that $2m(r,t)/r^3$ is certainly measurable in a finite-size laboratory. Furthermore, by measuring the extent to which the ``verticals'' converge on each other, a finite-size laboratory can also directly measure the radial coordinate $r$. Consequently in any finite-size laboratory the quantity $2m(r,t)/r$ is physically measurable.
Detection of the presence (or absence) of an \emph{apparent} horizon is therefore a well-defined quasi-local observable.
Similar comments apply to the very closely related notion of \emph{trapping} horizon~\cite{Booth:2005, Andersson:2005, Poisson, evolving}.
For historical reasons, Hawking himself prefers to phrase his discussion in terms of \emph{apparent} horizons. However, formal developments over the last two decades or so suggest that a rephrasing in terms of \emph{trapping} horizons is potentially more fruitful. In particular, some non-symmetric foliations of spherically symmetric spacetimes can lead to rather ``unexpected'' apparent horizons~\cite{Wald:1991}, and the trapping horizons are better behaved in this regard. (See eg~\cite{Booth:2005, Andersson:2005}.)
In short, detection of the presence (or absence) of an \emph{apparent} or \emph{trapping} horizon is a well-defined quasi-local observable.
\section{Event horizons in flat spacetime}
A well-known but often under-appreciated aspect of event horizons is that (contrary to common misperceptions) they are \emph{not} necessarily associated with strong gravitational fields; at least not in any simple local manner.
In fact it is possible to arrange a situation in which a classical event horizon forms in a portion of spacetime that is a zero-curvature Riemann-flat segment of Minkowski spacetime.
To do this merely consider a spherically symmetric shell of dust with a vacuum interior. (See eg~\cite{Poisson}, and figure 1.)
By the Birkhoff theorem, the spacetime inside the shell is a segment of Minkowski space, and the spacetime outside the shell is a segment of Schwarzschild spacetime, (with some fixed constant mass parameter $m$).
As the dust shell collapses, it will eventually cross its Schwarzschild radius and a black hole will form. Consider the resulting horizon.
Then, as expected, to the future of the Schwarzschild-radius-crossing event the horizon is indeed an event horizon which is located at $r=2m$.
However, there is also a portion of the event horizon that stretches backwards from the Schwarzschild-radius-crossing event to $r=0$, the centre of the spacetime.
This portion of the event horizon is located in a segment of flat Minkowski space.
Perhaps even worse, the very existence of this event horizon is predicated on both the eternal immutability of the resulting black hole, (infinitely far into the future), \emph{and} on assuming that the Schwarzschild solution continues to hold \emph{exactly}, all the way down to the central spacelike singularity.
(There is also an apparent horizon which first forms when the shell crosses its Schwarzschild radius, and then immediately splits, with one part of the apparent horizon remaining at $r=2m$ while the second part sits on top of the dust shell and follows it down to the centre at $r=0$).
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=4cm, height=5cm]{one-shell.pdf}
\caption{The event horizon (thin solid line) reaches back into the flat Minkowski region. Part of the apparent horizon (grey line) follows the collapsing dust shell (dotted line) down to the singularity (thick black line).\vspace{-10pt} }
\label{F:1}
\end{center}
\end{figure}
\section{Non-detectability of event horizons}
Consider a spherical laboratory of radius $R$ that collects data for some time duration $T$.
Assume the equipment in the laboratory to have negligible mass, so that the gravitational self-field of the laboratory can be neglected.
Surround the laboratory with a spherical dust shell; by the Birkhoff theorem all quasi-local experiments inside the laboratory will continue to detect a Minkowski spacetime.
Now choose the dust shell such that the mass parameter of the segment of Schwarzschild spacetime outside the dust shell satisfies
\begin{equation}
2m > R + T.
\end{equation}
Let the dust shell go, drop it, arranging the timing so that the dust shell crosses its Schwarzschild radius just as the laboratory stops collecting data.
Then the resulting event horizon reaches back to engulf the entire laboratory over the entire time interval that it was collecting data.
There is simply no way that denizens of the laboratory could have detected the presence of the event horizon by any quasi-local means.
Related comments can be found (for instance) in work by Bengtsson and Senovilla~\cite{Bengtsson:2010, Bengtsson:2008}, and in the Living Review of Ashtekar and Krishnan~\cite{Ashtekar:2004}, and in Hayward's article from the year 2000~\cite{Hayward:2000}. It is perhaps a little disturbing to realise that the quite serious deficiencies and limitations exhibited by \emph{event} horizons, while well appreciated within the general relativity community, are largely not understood or appreciated in the wider physics community.
\section{Destroying horizons}
It is again well-known, if not widely appreciated, that from the point of view of QFT the Hawking flux can be viewed as a \emph{negative} flux \emph{into} the black hole~\cite{B&D}.
This suggests a simple toy model, a gedanken-experiment, that can be used to capture \emph{some} of the key features of Hawking evaporation.
Let us now surround the black hole we have just considered with a second shell of \emph{negative} mass dust, carefully tuned to make the total mass of the system zero.
(So, again by the Birkhoff theorem, inside the inner shell the spacetime is a segment of Minkowski spacetime, between the two shells it is a segment of Schwarzschild spacetime with some fixed mass parameter $m$, and outside the outer shell it is again a segment of Minkowski spacetime.)
As the negative-mass outer shell drops, it will eventually cross $r=2m$, in the process destroying the horizon that had (previously and temporarily) formed due to the inner shell crossing its Schwarzschild radius. See figure 2.
(See also the related discussion in reference~\cite{Hayward05a}, and in the book by Frolov and Novikov~\cite{Frolov-Novikov}.)
But what sort of horizon is/was this?
Certainly there is an apparent horizon in this spacetime.
As previously, the apparent horizon first forms when the inner shell crosses its Schwarzschild radius, immediately splitting, with one part of the apparent horizon remaining at $r=2m$ while the second part rides the inner shell down to the centre at $r=0$ (where a curvature singularity forms).
Later on, once the outer shell drops down to $r=2m$, the apparent horizon, (suitably defined), follows the outer shell down to the centre at $r=0$. (Where the curvature singularity then unforms.)
But where is the event horizon?
Go to the event where the outer shell finally reaches the centre, $r=0$, and then in the causal past of that event use the Schwarzschild geometry to backtrack the ``outgoing'' (ie, less ingoing) null curve into the past, until it eventually crosses the inner shell.
This null curve will always strictly satisfy $r<2m$, and so will be strictly inside (though asymptotically approaching) the apparent horizon.
Once this backtracked null curve crosses the inner shell, from that event backtrack (now in the Minkowski geometry) the outgoing radial null curve until one reaches the centre $r=0$.
This segment of the event horizon always lies strictly outside the relevant segment of the apparent horizon.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=4cm, height=5cm]{two-shell.pdf}
\caption{Part of the event horizon (thin black line) reaches back into the flat Minkowski region, and the other part reaches forward to the point where the singularity (thick black line) is destroyed. Part of the apparent horizon (grey line) follows the two collapsing dust shells (dotted lines) down to the singularity, the remaining portion lies at $r=2m$ between the two shells. The event horizon and apparent horizon occur at qualitatively different locations. \vspace{-15pt} }
\label{F:2}
\end{center}
\end{figure}
So in this toy model there is still an event horizon, but it is not where one might have thought it should be, and in particular does not extend to the infinite future.
Part of the event horizon lies in a Riemann-flat section of spacetime, and the part of the event horizon that lies in the Schwarzschild segment of the spacetime does not occur at $r=2m$.
One cannot even begin to calculate the location of the event horizon until one has final definitive information that the spacetime has stabilized into its final form.
One could always posit yet a third dust shell, waiting in the wings, ready to engulf the entire region of interest with an unexpected event horizon.
In counterpoint, there are a number of models in which the event horizon never forms in the first place.
\section{Excluding horizons completely}
The distinction between the apparent and event horizons can be so non-perturbatively extreme as to completely exclude event horizons.
In particular this is the logic underlying the physics of the Bergmann--Roman~\cite{Roman}, the Hayward~\cite{Hayward05a, Hayward05b}, and the more recent Frolov~\cite{Frolov:2014} scenarios, in all of which the apparent horizon is taken to be a topologically toroidal surface, $S^1\times S^2$, which does not intersect the always time-like $r=0$ world line.
Some other models, such as that of Ashtekar--Bojowald~\cite{Ashtekar}, are based on apparent horizons that reach down to $r=0$, (and so are topologically $\mathbb{R}\times S^2$).
Something more or less along these lines is also the underlying gestalt for Hawking's recent article~\cite{horizons}, and the models of Israel~\cite{Israel:2014} and Bardeen~\cite{Bardeen:2014}.
\section{Killing horizons}
Of course in static spacetimes and stationary spacetimes the \emph{event}, \emph{apparent}, \emph{trapping}, and \emph{Killing} horizons happen to coincide.
Thereby, physical detectability of \emph{apparent}, \emph{trapping}, and \emph{Killing} horizons implicitly implies the physical detectability of the \emph{event} horizon; but only for this restricted class of spacetimes. (See for example the discussion in reference~\cite{Poisson}.)
In non-stationary spacetimes \emph{Killing} horizons no longer exist, and the \emph{apparent} and \emph{trapping} horizons need not, and typically will not, coincide with the \emph{event} horizon, even if one might exist.
Nor, as we have seen above, need any event horizon even be ``perturbatively close'' to any apparent/trapping horizon.
\section{ADAFs}
The empirical astronomical observation of ``advection dominated accretion flows''~\cite{adaf, Narayan, Narayan:2008, Narayan:1998} is often interpreted as evidence for the existence of \emph{event} horizons --- but ADAFs cannot distinguish between event horizons and apparent/trapping horizons.
The key point is that one observes the \emph{lack} of bremsstrahlung as infalling matter coalesces with the putative black hole --- indicating that the surface is not some solid/liquid/gas interface, but is instead acting as some sort of ``one way membrane'' (at least for the temporal duration of the observations).
The ADAF observations are thus suggestive of the presence of \emph{some sort} of horizon, but cannot be used to specifically argue for the presence of \emph{event} horizons.
\section{Discussion}
Perhaps surprisingly to non-relativists, there are \emph{very many quite different} types of horizon one can define.
(At the very least, one can define the event, apparent, trapping, isolated, dynamical, evolving, causal, Killing, non-Killing, universal, Rindler, particle, and putative horizons, \emph{etcetera}...~\cite{Booth:2005, Andersson:2005, Killing1, Killing2, Hell, Krasinski:2003, Helfer:2011, lost, Binetruy:2014, MTW, Wald, Hartle, Poisson, evolving}.)
Furthermore, the precise technical differences between these horizons sometimes can (and often does) make a physical difference in physically relevant and mathematically well-defined situations.
Additionally, again perhaps surprisingly to non-relativists, \emph{event} horizons, (despite their undoubted mathematical virtues in allowing one to prove quite powerful general theorems~\cite{Hell} in classical general relativity), are in any dynamical context not physically detectable by any finite-size laboratory or telescope, severely curtailing their empirical usefulness.
So let us focus on issues that can at least in principle be testable in a finite-size laboratory.
The point is that while event horizons are certainly properties of classical general relativity (Schwarzschild spacetime, Kerr spacetime, and their evolving generalizations, \emph{etcetera}), there is little to no evidence that \emph{event} horizons survive the introduction of even semi-classical quantum effects (let alone full quantum effects).
\emph{Event} horizons are extremely delicate and depend on global geometry --- including what will happen in the infinite future.
In contrast \emph{apparent} horizons or \emph{trapping} horizons seem much more robust in this regard --- they depend on local or quasi-local physics and are much more difficult to destroy.
(Some authors prefer to use the more general term ``quasi-local horizons''.)
In particular, in somewhat different contexts, it has already been demonstrated that the presence of an \emph{event} horizon~\cite{essential},
or indeed \emph{any} sort of horizon~\cite{minimal1, minimal2, trapped} is not essential for Hawking radiation and ``black hole evaporation'' to occur, and for that matter, in some non-GR analogue spacetime contexts one can even dispense with black hole entropy
(Bekenstein entropy)~\cite{analogue, acoustic, no-entropy}.
The tangle of issues related (both directly and indirectly) to these observations has significant implications for all of black hole thermodynamics, and in particular is central to the understanding of the endpoint of the Hawking evaporation process~\cite{firewall, firewall2, Braustein:2009, no-firewall, no-firewall2, no-firewall3, no-firewall4, no-firewall5, no-firewall6, no-firewall7, no-firewall8, no-firewall9}.
\acknowledgments
This research was supported by the Marsden Fund, and by a James Cook fellowship, both administered by the Royal Society of New Zealand.
|
\section{Ordered algebraic structures.}
In this paper we construct certain ordered semi- algebraic structures which have surely been described before. The author would appreciate any references to the literature. Then we use these ordered algebraic structures for various applications. In particular, we describe measures with values in the ordered semi- algebraic structures, some of which can be used for probability calculations. The (probability) measure theory ideas may also already exist in some form. Finally, we will describe transverse measures for codimension-1 laminations with values in our ordered semi- algebraic structures.
We begin by defining a very simple ordered commutative semi-ring which was used in \cite{UO:Depth} to analyze finite height (or finite depth) measured laminations in surfaces.
\begin{example} Let $\mathbb N_0=\mathbb N\cup \{0\}=\{0,1,2,3,\ldots\}$ and let $\mathbb S$ denote the set $\{0\}\cup(\mathbb N_0 \times (0,\infty])$, where $(0,\infty]\subset\bar \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$ and where $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}=[-\infty,\infty]$ denotes the extended real line. We are using the symbol $0$ to denote 0 in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$ or $\mathbb Z$, but in the definition of $\mathbb S$ we are adding a distinct $0$. We use the
lexicographical order relation on $\mathbb N_0\times (0,\infty]$, so that $(i,t)<(j,s)$ either if $i<j$ and $s,t\in (0,\infty]$ or if $i=j$ and $t<s$. The element $0\in\mathbb S$ is a least element, $0<(i,t)$ for all $(i,t)\in(\mathbb N_0\times (0,\infty])$
We make $\mathbb S$ a topological space with the order topology as shown in Figure \ref{OrderParameter}.
The elements $(i,\infty)$ are called {\it infinities} of $\mathbb S$. Define a commutative {\it addition} operation
on $\mathbb S$ by $$(i,t)+(j,u)
=\begin{cases}(i,t+u) & \mbox{if } i=j \\
( i,t) & \mbox{if } i>j\\
(j,u) & \mbox{if } j>i
\end{cases}
$$
$$ (i,t)+0=(i,t).$$
\noindent with the convention $\infty+a=a+\infty=\infty$ for any $a\in (0,\infty]$. Define a commutative {\it multiplication} on $\mathbb S$ by
$$(i,t)(j,u)=(i+j, tu),$$ $$0(i,t)=(i,t)0=0,$$ with the convention that $a\infty=\infty a=\infty$ for any $a\in (0,\infty]$.
We say $(i,t)\in (\mathbb Z\times (0,\infty])$ is {\it real} if $i=0$. When we denote an element of $\mathbb S$ by a single symbol $x=(i,t) \in \mathbb S\setminus \{0\}$, we will use $\mathfrak{L}(x)=i$ to denote the {\it level of $x$} and $\mathfrak{R}(x)=t$ to denote the {\it real part of $x$}, which lies in $(0,\infty]$. We make the convention that $\mathfrak{R}(0)=0$ and $\mathfrak{L}(0)$ is undefined.
In terms of an operation we will define later, $\mathbb S=\mathbb N_0\circledwedge [0,\infty]$.
\end{example}
It is routine to verify the following:
\begin{lemma} $\mathbb S$ is an ordered commutative semi-ring with multiplicative identity $(0,1)$ and additive identity $0$.
\end{lemma}
We observe that the subset of $\mathbb S$ which we identify with the positive extended reals, namely $\{(0,t)\in \mathbb S:t \in (0,\infty]\}=\{x\in \mathbb S:\mathfrak{L}(x)=0\}$, has the usual addition and multiplication of the positive extended reals, and also
has the usual topology.
To avoid confusion, we define the terms ``ordered abelian semi-group," ``ordered commutative semi-ring" and ``ordered semi-field" as used in this paper.
\begin{defn} An {\it ordered abelian semi-group} is a set G equipped with a binary operation +, a total ordering $<$ and an element $0$ satisfying the following axioms for any $a,b,c\in G$:
\begin{tightenum}
\item $(a + b) + c = a + (b + c)$
\item $0 + a = a + 0 = a$
\item $a + b = b + a$
\item 0 is the least element and for all $a$ and $b$, $a+b\ge b$.
\end{tightenum}
\end{defn}
\begin{defn} An {\it ordered commutative semi-ring} is a totally ordered set $R$, with order $<$, equipped with two binary operations addition, +, and multiplication and with elements 0,1, such that $(R,+)$ is an ordered abelian semigroup and such that the following are satisfied for any $a,b,c\in R$:
\begin{tightenum}
\item $ab=ba$
\item $(ab)c = a(bc)$
\item $1a = a1 = a$
\item $a(b + c) = (ab) + (ac)$
\item $0a = a0 = 0$
\end{tightenum}
An {\it ordered semi-field} is a commutative semi-ring with the additional property that every non-zero element $a$ has a multiplicative inverse $\bar a$ such that:
\begin{tightenum}
\item[(ix)] $a\bar a=\bar a a=1$
\end{tightenum}
\end{defn}
Next we define an ordered commutative semi-ring $\mathbb O$, related to $\mathbb S$, which is pictured with the order topology in Figure \ref{OrderParameter}.
\begin{defn}\label{BLDefn} Let $\mathbb O$ denote the set $\{0\}\cup(\mathbb Z\times (0,\infty])$ where $(0,\infty]\subset\bar \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$ and where $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}=[-\infty,\infty]$ denotes the extended real line. The definition of an ordering and operations on $\mathbb O$ are essentially the same as for $\mathbb S$. We use the
lexicographical order relation on $\mathbb Z\times (0,\infty]$, so that $(i,t)<(j,s)$ either if $i<j$ and $s,t\in (0,\infty]$ or if $i=j$ and $t<s$. The element $0\in\mathbb O$ is a least element, $0<(i,t)$ for all $(i,t)\in(\mathbb Z\times (0,\infty])$.
We make $\mathbb O$ a topological space with the order topology.
The elements $(i,\infty)$ are called {\it infinities} of $\mathbb O$. Define a commutative {\it addition} operation
on $\mathbb O$ by $$(i,t)+(j,u)
=\begin{cases}(i,t+u) & \mbox{if } i=j \\
( i,t) & \mbox{if } i>j
\end{cases}
$$
$$ (i,t)+0=(i,t),$$
\noindent with the convention $\infty+a=a+\infty=\infty$ for any $a\in (0,\infty]$. Define a commutative {\it multiplication} on $\mathbb O$ by
$$(i,t)(j,u)=(i+j, tu),$$ $$0(i,t)=(i,t)0=0,$$ with the convention that $a\infty=\infty a=\infty$ for any $a\in (0,\infty]$.
We say $(i,t)\in (\mathbb Z\times (0,\infty])$ is {\it real} if $i=0$. When we denote an element of $\mathbb O$ by a single symbol $x=(i,t) \in \mathbb O\setminus \{0\}$, we will use $\mathfrak{L}(x)=i$ to denote the {\it level of $x$} and $\mathfrak{R}(x)=t$ to denote the {\it real part of $x$}, which lies in $(0,\infty]$. We make the convention that $\mathfrak{R}(0)=0$ and $\mathfrak{L}(0)$ is undefined.
We define {\it vectors} in $\mathbb O^n$ as $n$-tuples of elements of $\mathbb O$, and give $\mathbb O^n$ the product topology. We allow $n$ infinite, giving an infinite product. If $\lambda\in \mathbb O$ and $w\in \mathbb O^n$, $w=(w_1,\ldots,w_n)$, then
we define {\it scalar multiplication} by $\lambda w=(\lambda w_1, \lambda w_2,\ldots, \lambda w_n)$.
The {\it lattice points} of $\mathbb O^n$ are vectors of the form $[(i_1,\infty),(i_2,\infty),\ldots ,(i_n,\infty)]$; we use $Z$ to denote the set of lattice points in $\mathbb O^n$. The {\it origin} is the vector with all entries 0.
A {\it cone} in $\mathbb O^n$ is a subset of $\mathbb O^n$ closed under
scalar multiplication by $\lambda\in \mathbb O$.
\end{defn}
We observe that the subset of $\mathbb O$ which we identify with the positive extended reals, namely $\{(0,t)\in \mathbb O:t \in (0,\infty]\}=\{x\in \mathbb O:\mathfrak{L}(x)=0\}$, has the usual addition and multiplication of the positive extended reals, and also
has the usual topology.
Scalar multiplication shifts the levels of all entries of a vector by the same integer. Scalar multiplication by a real does not shift levels of entries.
Clearly there is an action of $\mathbb Z$ on $\mathbb O^n$. Namely, for any integer $r$ the action takes $w\in \mathbb O^n$, $w=(w_1,\ldots,w_n)$ to the scalar product $(r,1)(w_1,\ldots,w_n)$, shifting the levels of all entries by $r$. The action on $\mathbb O$ fixes 0, which means that subspaces of $\mathbb O^n$ of the form $w_{i_1}=w_{i_2}=\cdots=w_{i_s}=0$ are preserved by the action. (We use notation for finite products, but everything applies for infinite products as well.)
It is routine to verify the following:
\begin{lemma} $\mathbb O$ is an ordered commutative semi-ring with multiplicative identity $(0,1)$ and additive identity $0$. \end{lemma}
Figure \ref{OrderParameter} shows $\mathbb O$ as a topological space with the order topology.
\begin{figure}[H]
\centering
\scalebox{1}{\includegraphics{OrderParameter}}
\caption{\small The ordered commutative semi-rings $\mathbb S$, $\mathbb O$, $\bar\mathbb O$, and $\mathbb P$ viewed as topological spaces.}
\label{OrderParameter}
\end{figure}
For some applications it is useful to extend $\mathbb O$ slightly.
\begin{defn} Let $\bar\mathbb O$ denote $\mathbb O\cup\{\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\}$ with operations and the order in $\mathbb O$ extended in the obvious way: $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}>x$ for all $x\in \mathbb O$. $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}+x=x+\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$ for all $x\in \mathbb O$. Finally, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\cdot x=x\cdot \hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$ unless $x=0$, and $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\cdot 0=0\cdot \hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=0$.
\end{defn}
For applications to measure theory, we will need to define countably infinite sums in $\mathbb S$, $\mathbb O$ and $\bar \mathbb O$.
\begin{defn} In $\mathbb O$ we can define countably infinite sums of elements $x_n\in \mathbb O$, provided the summands have uniformly bounded levels. In this case, assuming the maximum of $\mathfrak{L}(x_n)$ is $M$, it is natural to define
$\displaystyle \sum_{n=1}^\infty x_n:=\sum_{\mathfrak{L}(x_n)=M}x_n$, which makes sense since $\displaystyle \sum_{\mathfrak{L}(x_n)=M}x_n=(M,\sum_{\mathfrak{L}(x_n)=M}\mathfrak{R}(x_n))$, which is expressed in terms of a countable sum of positive real numbers.
If levels in $\displaystyle \sum_{n=1}^\infty x_n$ are not uniformly bounded, the countable sum does not always make sense in $\mathbb O$. However, in $\bar \mathbb O$, we can, as before, define $\displaystyle \sum_{n=1}^\infty x_n =\sum_{\mathfrak{L}(x_n)=M}x_n$ if $\{\mathfrak{L}(x_n)\}$ has maximum value $M$. In case $\{\mathfrak{L}(x_n)\}$ is unbounded above, we define $\displaystyle \sum_{n=1}^\infty x_n=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$.
In $\mathbb S$ we can define $\displaystyle \sum_{n=1}^\infty x_n$, again provided the levels of summands $x_n$ are uniformly bounded. To be able to evaluate all countable sums, we must again adjoin a new infinity, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, to obtain $\bar\mathbb S=\mathbb S\cup\{\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\}$.
\end{defn}
Next, we will describe more general constructions for combining {\it ordered algebraic structures} which can be ordered abelian semi-groups, ordered commutative semi-rings, or ordered semi-fields, to obtain new ordered algebraic structures.
\begin{defn} Suppose $A$ and $B$ are ordered abelian semi-groups. Then we define {\it $A$ s-insert $B$} as $ A \circledvee B=A\times B$ which becomes an ordered semi-group, with order relation and operations described below. In the next definition we define another ``insert" operation. To distinguish these, we can refer to the construction we describe here as a {\it semi-group insertion} or {\it s-insertion}.
We define an order relation $<$ on $A\circledvee B$ as follows:
$(g,s)<(h,t)$ if either if $g<h$ or if $g=h$ and $s<t$.
A commutative {\it addition} operation
on $A\circledvee B$ is given by $$(g,s)+(h,t)
=\begin{cases}( g,s) & \mbox{if } g>h \\
(g,s+t) & \mbox{if } g=h
\end{cases}$$
The element $(0,0)$ is the additive identity, which we also denote simply as $0$.
If we wish to extend $A\bar\circledvee B$ by including an infinity, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, then we define $A\bar\circledvee B$ as $A\circledvee B\cup \{\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\}$ and extend the operations and order relation as follows:
For all $x\in A\circledvee B$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}>x$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}+x=x+\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$.
\end{defn}
One can verify:
\begin{lemma} If $A$ and $B$ are ordered abelian semi-groups, then $A\circledvee B$ and $A\bar\circledvee B$ are ordered abelian semi-groups with the order and addition defined above.
\end{lemma}
\begin{defn} Suppose $A$ is an ordered abelian group or an ordered abelian semigroup. Suppose $B$ is an ordered abelian semi-group, an ordered commutative semi-ring, or ordered semi-field. Then we define {\it $A$ insert $B$} as $$ A \circledwedge B=A\times (B\setminus \{0\})\cup \{0\},$$ which becomes an ordered abelian semi-group or an ordered commutative semi-ring, with order relation and operations described below. The zero added to $A\circledwedge B$ is distinct from the zero removed from $B$. In case $A$ is an ordered abelian group and $B$ is an ordered semi-field, $A\circledwedge B$ becomes an ordered semi-field. In case both $A$ and $B$ are ordered abelian semigroups, there is a difference between s-insertion and insertion.
We define an order relation $<$ on $A\circledwedge B$ as follows:
\begin{tightenum}
\item $(g,s)<(h,t)$ if either if $g<h$ or if $g=h$ and $s<t$.
\item $0<(g,s)$ for all $(g,s)$.
\end{tightenum}
Now we define the addition operation on $A\circledwedge B$.
The commutative {\it addition} operation
on $A\circledwedge B$ is given by $$(g,s)+(h,t)
=\begin{cases}( g,s) & \mbox{if } g>h \\
(g,s+t) & \mbox{if } g=h
\end{cases}
$$
$$0+(g,s)=(g,s)+0=(g,s)$$
In case $B$ is an ordered abelian group or ordered abelian semi-group, this defines the ordered abelian semi-group $A\circledwedge B$.
Provided $B$ is an ordered commutative semi-ring, we define
a commutative {\it multiplication} on $A\circledwedge B$ by
$$(g,s)(h,t)=(g+h, st),$$ $$0(g,s)=(g,s)0=0,$$
If we wish to extend $A\circledwedge B$ by including an infinity, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, then we define $A\bar\circledwedge B$ as $A\circledwedge B\cup \{\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\}$ and extend the operations and order relation as follows:
For all $x\in A\circledwedge B$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}>x$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}+x=x+\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$. Finally, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\cdot x=x\cdot \hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$ unless $x=0$, and $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\cdot 0=0\cdot \hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=0$.
When we denote a non-zero element of $A\circledwedge B$ by a single symbol $x=(g,s)$, we will use $\mathfrak{L}(x)=g$ to denote the {\it level of $x$} and $\mathfrak{R}(x)=s$ to denote the {\it residue of $x$}, which lies in $B \setminus\{0\}$. We make the convention that $\mathfrak{R}(0)=0$ and $\mathfrak{L}(0)$ is undefined (and $\mathfrak{L}(\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt})$ is undefined in $A\bar\circledwedge B$).
If $A$ is an ordered abelian group and $B$ is an ordered semi-field, then $A\circledwedge B$ is an ordered semi-field. In this case, the multiplicative inverse of $(g,s)$ is $(-g,\bar s)$, where $\bar s$ is the multiplicative inverse of $s$ in $B$.
\end{defn}
\begin{lemma}\label{WellDefinedLemma} (a) If $A$ is an ordered abelian group or semi-group and $B$ is an ordered abelian semi-group
then $A\circledwedge B$ is an ordered abelian semi-group.
\noindent (b) If $A$ is an ordered abelian group or ordered abelian semi-group and $B$ is an ordered commutative semi-ring
then $A\circledwedge B$ is an ordered commutative semi-ring. In this case, $\mathfrak{L}(xy)=\mathfrak{L}(x)+\mathfrak{L}(y)$. If $x\ne 0$, $y\ne 0$, then the level function $\mathfrak{L}$ in $A\circledwedge B$ satisfies $\mathfrak{L}(x+y)=max(\mathfrak{L}(x),\mathfrak{L}(y))$.
\noindent (c) If $A$ is an ordered abelian group and $B$ is an ordered commutative semi-field
then $A\circledwedge B$ is an ordered semi-field.
\end{lemma}
\begin{remark} We describe the operation $\circledwedge$ as an ``insertion." To construct $A\circledwedge B$, we insert a copy of $B\setminus \{0\}$ at every element of $A$, with its order, then add a different $0$ to construct $A\circledwedge B$. To remember the meaning of the symbols $\circledvee$ and $\circledwedge$, we note that the ``arrow" points down in the first, which has more limited applicability, namely it operates only on ordered abelian semi-groups.\end{remark}
\begin{exs}
\begin{intenum}
\item $\mathbb N_0\circledwedge [0,\infty]$ is the commutative semi-ring described above as $\mathbb S$, constructed from the commutative semi-ring $[0,\infty]$ in the extended reals.
\item $\mathbb Z\circledwedge [0,\infty]$ is the commutative semi-ring we described above as $\mathbb O$, constructed using the commutative semi-ring $[0,\infty]\subset \bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$.
\item Since, for example $\mathbb Z \circledwedge[0,\infty]$ can be regarded as an ordered abelian semi-group, we can construct a new ordered commutative semi-ring $(\mathbb Z\circledwedge [0,\infty])\circledwedge[0,\infty]$.
\item We can construct many more examples by iterating as often as we want, with brackets to indicate the order of $\circledwedge$ operations. A random example is $((\mathbb Z\circledwedge [0,\infty])\circledwedge(\mathbb Z\circledwedge [0,\infty]))\circledwedge[0,\infty]$.
\item We can first use the semi-group insertion operation to construct a new semi-group, and then construct a semi-ring using the new semi-group. For example $(\mathbb N_0 \circledvee \mathbb N_0)\circledwedge [0,\infty]$.
\end{intenum}
\end{exs}
In view of the examples, it is natural to ask whether in some cases the $\circledwedge$ operation is associative.
\begin{example} Consider the two ordered semi-rings $\mathbb N_0\circledwedge(\mathbb N_0\circledwedge\mathbb N_0)$ and $(\mathbb N_0\circledwedge\mathbb N_0)\circledwedge\mathbb N_0$. One might hope that $\psi((i,(j,k)))=((i,j),k)$ defines an isomorphism of ordered semi-rings. That is not the case. For example in the first group $\mathbb N_0\circledwedge(\mathbb N_0\circledwedge\mathbb N_0)$ we have $(1,(1,1))\cdot (2,(1,1))=(3,(2,1))$ whereas in the second group $(\mathbb N_0\circledwedge\mathbb N_0)\circledwedge\mathbb N_0$ we have
$((1,1),1)\cdot ((2,1),1)=((2,1),1)$. In fact, the two ordered semi-rings do not even have exactly the same underlying sets.
\end{example}
However, for semi-group insertions the situation is nicer.
\begin{lemma} If $A$, $B$, $C$ are ordered abelian semi-groups, then $(A\circledvee B)\circledvee C\cong A\circledvee(B\circledvee C)$.
\end{lemma}
\begin{proof} We can easily check that in $A\circledvee (B\circledvee C)$:
$(x_1,(x_2,x_3))+(y_1,(y_2,y_3))=(x_1,(x_2,x_3))$ if $x_1>y_1$ or $x_1=y_1$ and $x_2>y_2$,
$(x_1,(x_2,x_3))+(y_1,(y_2,y_3))=(y_1,(y_2,y_3))$ if $y_1>x_1$ or $y_1=x_1$ and $y_2>x_2$,
$(x_1,(x_2,x_3))+(y_1,(y_2,y_3))=(x_1,(x_2,x_3+y_3))$ if $x_1=y_1$ and $x_2=y_2$.
Similarly, we check that in $(A\circledvee B)\circledvee C$:
$((x_1,x_2),x_3)+((y_1,y_2),y_3)=((x_1,x_2),x_3)$ if $x_1>y_1$ or $x_1=y_1$ and $x_2>y_2$,
$((x_1,x_2),x_3)+((y_1,y_2),y_3)=((y_1,y_2),y_3)$ if $y_1>x_1$ or $y_1=x_1$ and $y_2>x_2$,
$((x_1,x_2),x_3)+((y_1,y_2),y_3)=((x_1,x_2),x_3+y_3)$ if $x_1=y_1$ and $x_2=y_2$.
One can easily show that $\psi\left((a,(b,c))\right)=\left(((a,b),c)\right)$ defines an order isomorphism, so the above identities prove that $\psi$ defines an isomorphism $(A\circledvee B)\circledvee C\cong A\circledvee(B\circledvee C)$.
\end{proof}
\begin{defn} If $A_i$ are ordered abelian semi-groups, $\displaystyle \mathop{\mathlarger{\mathlarger\circledvee}}_{i=1}^nA_i$ or $A_1\circledvee A_2\circledvee \cdots \circledvee A_n$ denotes the ordered abelian semi-group $\displaystyle( (\cdots(((A_1\circledvee A_2)\circledvee A_3)\circledvee A_3)\cdots)A_n)$.
\end{defn}
For measure theoretic applications, it is useful to be able to evaluate all countable sums (of positive elements) in an ordered algebraic structure. To make this possible, one often has to use the insert operation $\bar\circledwedge$.
\begin{defn} An ordered algebraic structure $A$ has the {\it least upper bound property (lub property)} if every set $S\subset A$ which is bounded above has a least upper bound $\sup(S)$. The ordered algebraic structure has the {\it summability property} if every countable sum of positive elements can be evaluated, yielding an element in the algebraic structure.
\end{defn}
\begin{lemma} \label{LUBLemma}Suppose $A$ is an ordered abelian group or ordered abelian semi-group with the property that every bounded non-empty set has a greatest element. Suppose $B$ is an ordered abelian semi-group or ordered semi-ring. Suppose $B$ has the least upper bound property and either
\begin{tightenum}
\item $B$ has a greatest element $\infty$, or
\item $B$ has a least element $p>0$ and $A$ has the property that every non-empty set has a least element.
\end{tightenum}
\noindent Then $A\circledvee B$ and/or $A\circledwedge B$ have the least upper bound property. Further $A\bar\circledvee B$ and/or $A\bar\circledwedge B$ is summable.
\end{lemma}
\begin{proof}
Suppose $A$ and $B$ are as in the statement. We will show that $A\circledvee B$ and $A\circledwedge B$ have the lub property. Suppose $S\subset A\circledwedge B$ or $S\subset A\circledvee B$ is bounded above. Then $$T=\{a\in A: \text{ there exists } b \text{ such that } (a,b)\in S\}$$ is bounded above and has a greatest element $M$. Let $U=\{b\in B: (M,b)\in S\}\ne\emptyset$.
\noindent {\it Case (i).} If $B$ has a greatest element $\infty$, since $B$ has the lub property and $U$ is bounded above by $\infty$, $U$ has an lub $N$ say. Then $(M,N)$ is the lub for $S$.
\noindent {\it Case (ii)} In this case, if $U$ is bounded above, then it has a least upper bound $N$, and again $(M,N)$ is the least upper bound of $S$. So we now assume $U$ is not bounded above. If $A$ does not have a greatest element, let $M'$ be the least element in $A$ greater than $M$. Then $(M',p)$ is the lub of $S$. If $A$ has a greatest element and $M$ is the greatest element, then for $S$ to be bounded above, $U=\{b\in B: (M,b)\in S\}\ne\emptyset$ must also be bounded above, and $(M,N)$ is again the lub.
\end{proof}
\begin{corollary}\label{NNCor} The ordered abelian groups $\displaystyle \mathop{\mathlarger{\mathlarger\circledvee}}_{i=1}^n\mathbb N_0$ have the lub property. Also $\mathbb S=\mathbb N_0\circledwedge [0,\infty]$ has the lub property.
\end{corollary}
\begin{proof} We use induction and Lemma \ref{LUBLemma} (ii) to show $\displaystyle \mathop{\mathlarger{\mathlarger\circledvee}}_{i=1}^n\mathbb N_0$ has the lub property. First, we check that $\displaystyle\mathop{\mathlarger{\mathlarger\circledvee}}_{i=1}^{n-1}\mathbb N_0$ has a least positive element, namely $(0,0,\ldots,0,1)$. Also, clearly $\mathbb N_0$ has the property that every bounded nonempty subset has a greatest element. Then the lemma implies $\displaystyle \mathbb N_0\circledwedge\left(\mathop{\mathlarger{\mathlarger\circledwedge}}_{i=1}^{n-1}\mathbb N_0\right)=\mathop{\mathlarger{\mathlarger\circledwedge}}_{i=1}^n\mathbb N_0$ has the lub property.
To prove the second statement, observe that $[0,\infty]$ has a greatest element $\infty$, so we can apply the lemma in case (i).\end{proof}
\begin{lemma} If $[0,\infty]$ denotes the interval in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$, then
$$\bar\mathbb S_n=(\mathbb N_0\bar\circledwedge(\mathbb N_0\bar\circledwedge\cdots\bar\circledwedge(\mathbb N_0\bar\circledwedge(\mathbb N_0\bar\circledwedge [0,\infty]))\cdots)),$$ with $n$ insertions of $\mathbb N_0$, is an ordered abelian semi-ring with the lub property, is summable and has a greatest element. In particular $\bar \mathbb S$ is summable.
\vskip 0.15in
\noindent Similarly $$\bar\mathbb O_n=(\mathbb Z\bar\circledwedge(\mathbb Z\bar\circledwedge\cdots\bar\circledwedge(\mathbb Z\bar\circledwedge(\mathbb Z\bar\circledwedge [0,\infty]))\cdots)),$$ with $n$ insertions of $\mathbb Z$, is an ordered abelian semi-ring with the lub property which is also summable. In particular $\bar \mathbb O$ is summable.
\end{lemma}
\begin{proof} We can prove the first statement using induction starting with $\mathbb N_0\bar\circledwedge[0,\infty]$. We have already proved $\mathbb N_0\circledwedge[0,\infty]$ has the lub property, and we know $\mathbb N_0\bar\circledwedge[0,\infty]$ has a greatest element. Now we inductively apply $\mathbb N_0\bar\circledwedge$ to the previous result, and at every step of the induction apply Lemma \ref{LUBLemma} (i) using $A=\mathbb N_0$ to prove the lub and summability properties.
Essentially the same proof works for the second statement.
\end{proof}
We have described quite a few ordered algebraic structures with the desirable summability and/or lub properties using our operations $\circledvee$, $\bar\circledvee$, $\circledwedge$ and $\bar\circledwedge$. There are more possibilities to explore, and we will do this in later sections.
\section{Measures.}
In this section we will assume that $\mathbb L$ is some ordered abelian semigroup with at least the lub property, and also with the summability property. (When dealing with probability measures later, we can dispense with the summability property.)
\begin{defn}\label{LMeasureDef} Suppose $\mathbb L$ is an ordered abelian semigroup with the least upper bound and summability properties. Let $(X,\Sigma)$ be a measure space with $\sigma$-algebra $\Sigma$. An {\it $\mathbb L$-measure} $\nu$ assigns an element $\nu(E)$ of $\mathbb L$ to each measurable set $E\in \Sigma$ such that:
(i) $\nu(\emptyset)=0$,
(ii) If $\{E_i\}_{i\in I}$ is a countable collection of pairwise disjoint measurable sets in $X$, then
$$\nu\left(\bigcup_{i\in I} E_i\right )=\sum_{i\in I} \nu(E_i).$$
\end{defn}
Let us consider now the special case of an $\mathbb S$-measure. Recall $\mathbb S=\mathbb N_0\circledwedge [0,\infty]$. The definitions and some ideas developed for $\mathbb S$-measures apply equally to $\mathbb O$-measures and even more generally, but for specificity for now we deal with $\mathbb S$.
Associated to an $\mathbb S$-measure $\nu$, we will define a collection of ordinary extended-real valued measures, namely a measure $\nu_k$ associated to each level $k$. If $E\subset X$ is a measurable set we define
$$\nu_k(E)
=\begin{cases}\mathfrak{R}(\nu(E))& \mbox{if } \mathfrak{L}(\nu(E))=k \\
\infty & \mbox{if }\mathfrak{L}(\nu(E))>k\\
0& \mbox{if }\mathfrak{L}(\nu(E))<k\mbox{ or if }\nu(E)=0\in \mathbb S
\end{cases}
$$
Notice that the measure $\nu$ can be recovered from the sequence $\{\nu_j\}$. Namely, to find $\nu(E)$ given all $\nu_i(E)$, let $j$ be the largest $i$ such that $\nu_i(E)>0$, then $\nu(E)=(j,\nu_j(E))\in \mathbb S$.
\begin{observation} For measurable sets $E\subset X$ with $\mathbb S$-measure $\nu$, $\mathfrak{L}(\nu(E))$ is uniformly bounded by $\mathfrak{L}(\nu(X))$.
\end{observation}
The observation follows from the fact that $\nu(E)\le\nu(X)$, so $\mathfrak{L}(\nu(E))\le \mathfrak{L}(\nu(X))$. But we note that if we replace $\mathbb S$ by $\bar \mathbb S$ by adjoining an infinity $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, and $\nu(X)=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, then the observation gives no useful information.
Informally, if $\nu$ is an $\mathbb S$-measure of total height $h$ , then a set $E$ with $\mathfrak{L}(\nu(E))=h$ might be visible to the naked eye. To see a set with $\mathfrak{L}(\nu(E))=h-1$ one might need a microscope. For lower levels one would need stronger and stronger microscopes.
If one of the measures $\nu_j$, $0\le j\le h$, is trivial in the sense that there is no $E\in \Sigma$ such that $\mathfrak{L}(\nu(E))=j$ then in some sense the finite height measure is equivalent to another one, in which there are no trivial $\nu_j$ for $0\le j\le h$. In this case we could decrease by 1 the level of every $\nu_i$ for $i>j$.
\begin{defn}\label{LevelAdjustment} Suppose $\nu$ is a finite height $\mathbb S$-measure on $X$. Suppose $ \nu_j$ is trivial in the sense that there is no $E\in \Sigma$ such that $\mathfrak{L}(\nu(E))=j$. Then $\nu$ is {\it equivalent via level alignment} to the measure $\mu$, where $\mu$ is obtained from $\nu$ by shifting some levels: $ \mu_i=\nu_i$ for $i<j$; $ \mu_i=\nu_{i+1}$ for $i\ge j$. If no level alignments are possible, we say $\nu$ is a {\it proximal} measure.
We say the $\mathbb S$-measure $\nu$ has {\it total height $h$} if $\mathfrak{L}(\nu(X))=h$.
\end{defn}
More generally, suppose $\mathbb L$ is an ordered abelian group which is summable with infinity $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$ and $\mathbb K=\mathbb Z \circledwedge \mathbb L$ (or $\mathbb K=\mathbb N_0 \circledwedge \mathbb L$). Again we can define $\mathbb L$-measures $\nu_k$:
$$\nu_k(E)
=\begin{cases}\mathfrak{R}(\nu(E))& \mbox{if } \mathfrak{L}(\nu(E))=k \\
\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt} & \mbox{if }\mathfrak{L}(\nu(E))>k\\
0& \mbox{if }\mathfrak{L}(\nu(E))<k\mbox{ or if }\nu(E)=0\in \mathbb K
\end{cases}
$$
Many definitions we have made before carry through to $\mathbb K=\mathbb Z \circledwedge \mathbb L$ (or $\mathbb K=\mathbb N_0 \circledwedge \mathbb L$).
\begin{defn} Suppose $\mathbb K=\mathbb Z \circledwedge \mathbb L$ (or $\mathbb K=\mathbb N_0 \circledwedge \mathbb L$). We will make definitions with notation for $\mathbb K=\mathbb Z \circledwedge \mathbb L$, but they also apply to $\mathbb K=\mathbb N_0 \circledwedge \mathbb L$.
The {\it total height} of a measure is {\it infinite} if $\mathfrak{L}(\nu(E))$ takes infinitely many values as $E$ varies over measurable sets, otherwise it is $\max(\mathfrak{L}(\nu(E)))-\min(\mathfrak{L}(\nu(E)))+1$.
Suppose $\nu$ is a finite height $\mathbb K$-measure on $X$. Suppose $ \nu_j$ is trivial in the sense that there is no $E\in \Sigma$ such that $\mathfrak{L}(\nu(E))=j$. Then $\nu$ is {\it equivalent via level alignment} to the measure $\mu$, where $\mu$ is obtained from $\nu$ by shifting some levels: $ \mu_i=\nu_i$ for $i<j$; $ \mu_i=\nu_{i+1}$ for $i\ge j$. If no level alignments are possible, we say $\nu$ is a {\it proximal} measure.
The measure $\mu$ is equivalent via {\it level shift} to $\nu$ if for some fixed $k$, $ \mu_i=\nu_{i-k}$.
\end{defn}
If $\mathbb K$ is commutative ordered abelian ring, we we will in many cases be able to define integrals of real-valued or $\mathbb K$-valued functions with respect to $\mathbb K$-measures on measure spaces. The definition for real-valued functions will be inductive. To start the induction, we have integrals with respect to $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-measures.
\begin{defn} If $(X,\Sigma)$ is a measurable space, $\mathbb L$ is an ordered abelian semi-ring, and $f:X\to \mathbb L$ is a function, we say $f$ is {\it measurable} if for every $c\in \mathbb L$, the set $f^{-1}(\{x:x<c\})$ is measurable.
Suppose $\mathbb L$ is a summable ordered abelian semi-ring with infinity $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, and $\mathbb K=\mathbb Z \bar\circledwedge \mathbb L$ (or $\mathbb K=\mathbb N_0 \bar\circledwedge \mathbb L$). As before, we deal with $\mathbb K=\mathbb Z \bar\circledwedge \mathbb L$ and leave the obvious modifications for $\mathbb K=\mathbb N_0 \bar\circledwedge \mathbb L$ to the reader. If $X$ is a measure space with a $\mathbb K$-measure $\nu$, and integrals of $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$- valued measurable functions have been defined for $\mathbb L$-measures on $X$ then we define the integral with respect to the $\mathbb K$-measure $\nu$ as follows. Let $f$ be a real-valued measurable function on a measurable $A\subset X$. Then we define
$$\int_A f d\nu=\sum_{k\in \mathbb Z}(k,\int_A f d\nu_k),$$
where the sum is in $\mathbb K$. (For a finite total height measure, the above sum equals one of the summands.)
Using this definition inductively, starting with $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-measures on $X$ we can define the integral for $\nu$ a $\mathbb K$-measure, $$\mathbb K=\bar\mathbb S_n=(\mathbb N_0\bar\circledwedge(\mathbb N_0\bar\circledwedge\cdots\bar\circledwedge(\mathbb N_0\bar\circledwedge(\mathbb N_0\bar\circledwedge [0,\infty]))\cdots)),$$
and for $$\mathbb K=\bar\mathbb O_n=(\mathbb Z\bar\circledwedge(\mathbb Z\bar\circledwedge\cdots\bar\circledwedge(\mathbb Z\bar\circledwedge(\mathbb Z\bar\circledwedge [0,\infty]))\cdots)).$$
Now suppose $X$ is a measure space with $\mathbb K$-measure $\nu$, $\mathbb K=\bar\mathbb O_n$ or $\mathbb K=\bar\mathbb S_n$ and suppose $B$ is a measurable set. We want to define the integral of an $\bar\mathbb O_n$-valued function. We will deal with $\bar \mathbb O_n$, but the definition is the same for $\bar \mathbb S_n$. We have already defined the integral of a $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-valued measurable function with respect to $\nu$ above, so we have also defined the integral of a $\bar \mathbb O_0$-valued function, where $\bar\mathbb O_0=[0,\infty]\subset \bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$. Inductively, suppose we have defined the integral of a $\bar \mathbb O_{j-1}$-valued measurable function on $B$. Let $B_k=\{x:(k-1,\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt})<g(x)\le (k,\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt})\in \bar\mathbb O_j\}$, a measurable set. Then we define $$\int _B g d\nu=\sum_{k\in \mathbb Z}\left(k,\int_{B_k} \mathfrak{R}(g) d\nu\right).$$
Note that $\mathfrak{R}(g)$ is an $\bar \mathbb O_{j-1}$-valued measurable function on $B_k$ which we assume is defined. We have now defined $\int _B g d\nu$ for all $\bar \mathbb O_j$-valued functions $g$, $0\le j\le n$.
\end{defn}
Thus far, we have not made an issue of the ``integrability" of a measurable $\mathbb K$-valued function in the above definition, $\mathbb K=\bar\mathbb O$; our $\mathbb K$-valued functions are positive, and the integrals may have infinite values at different levels, but otherwise measurable functions are always integrable. However, it is possible to introduce negative values for $\mathbb K$-valued functions in an artificial way, and then integrability becomes an issue.
\begin{defn} Suppose $\mathbb L$ is an ordered abelian semi-structure (without negative values). Define the {\it double} of $\mathbb L$ as $\mathbb {DL}=\{0\}\cup \{+,-\}\times (\mathbb L\setminus \{0\})$. We define $y=(+,y)$, $-y=(-,y)$, $-(-,y)=y$, and $-(+,y)=(-,y)=-y$. This becomes an ordered set with the obvious ordering $(-,y)<0$ for all $y\ne 0$, $(+,y)>0$ for all $y\ne 0$, $(+,y)<(+,z)$ and $(-,y)>(-,z)$ if $y<z$. A function with values in $\mathbb {DL}$ can be written as $f=f_+-f_-$ where
$$f_+(x)
=\begin{cases}f(x) & \mbox{if } f(x)\ge 0 \\
0 & \mbox{if } f(x)< 0
\end{cases}$$
$$f_-(x)
=\begin{cases}-f(x) & \mbox{if } f(x)\le 0 \\
0 & \mbox{if } f(x)> 0
\end{cases}$$
and $f$ is {\it measurable} if $f_+$ and $f_-$ are measurable.
Suppose now that $\mathbb L=\mathbb O$ or $\mathbb L=\mathbb S$. If $f$ is a $\mathbb {DL}$-valued function on an $\mathbb L$-measure space $X$ with measure $\nu$, and $A\subset S$ is measurable, let $P=\int_Af_+d\nu$, $N=\int_Af_-d\nu$, the integrals of the positive and negative parts of $f$. We define
$$\int_Afd\nu=\begin{cases}P=\int_Af_+d\nu & \mbox{if } \mathfrak{L}(P)>\mathfrak{L}(N) \\
N=-\int_Af_-d\nu & \mbox{if } \mathfrak{L}(P)<\mathfrak{L}(N)\\
(\mathfrak{L}(P),\mathfrak{R}(P)-\mathfrak{R}(N)) & \mbox{if } \mathfrak{L}(P)=\mathfrak{L}(N).
\end{cases}$$
\end{defn}
The above definition is awkward because $\mathbb {DL}$ does not have a subtraction operation. Algebraically, $\mathbb {DL}$ is quite defective; if one attempts to give it a structure as an additive group, the associative law fails. However, we are using an artificial sum of a negative and a positive element defined in $\mathbb {DS}$ or $\mathbb {DO}$ by
$$(i,s)+(-(j,t))=\begin{cases} (i,s) & \mbox{if } i>j \\
-(j,t) & \mbox{if } i<j\\
(i,s-t) & \mbox{if } i=j \mbox{ and } s\ne t\\
0&\mbox {if } i=j \mbox{ and } s= t.
\end{cases}$$
\begin{ex} (An interpretation of Dirac-$\delta$ functions.) We define an $\mathbb O$-measure $\nu$ for $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^n$, determined by $\nu_0$ and $\nu_{-1}$. The measure $\nu_0$ is Lebesgue measure. For $\nu_{-1}$ we use a counting measure which assigns the number of points in a set if the set is finite, and otherwise assigns $\infty$. Let the Dirac-$\delta$ function $\delta$ at $y\in \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^n$ be defined by
$$\delta(x)=\begin{cases} (1,1) & \mbox{if } x=y \\
0 & \mbox{otherwise}.
\end{cases}$$
Then $\int_{\{y\}}\delta(x)d\nu=\int_{\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^n}\delta(x)d\nu=\nu(\{y\})\delta(y)=(-1,1)(1,1)=(0,1)$, and for a ``real-valued function" $f(x)$ with the property that $\mathfrak{L}(f(x))=0$, we have $\int_{\{y\}}f(x)d\nu=(-1,1)(0,\mathfrak{R}(f(x))=(-1,\mathfrak{R}(f(x))$, which is trivial viewed at level 0. Also for a real-valued $f$ and any measurable $A$, $\int_Afd\nu= \int_Afd\nu_0$, which is just the Lebesgue integral of $f$.
\end{ex}
As with real-valued measures, if $\mathbb L$ is a suitable ordered abelian semi-ring, then different $\mathbb L$-measures may be related by a Radon-Nikodym derivatives. Thus, if for all measurable $E$ in $X$ we have $\nu(E)=\int_E f(x) d\mu$, where $f$ is an $\mathbb L$ -valued function,
we can say that $f$ is a {\it Radon-Nikodym derivative $d\nu/d\mu$}.
\begin{problem} Formulate and prove a Radon-Nikodym theorem for $\mathbb O$-measures. It seems likely that a very similar statement is true, and that there is a simple proof involving the real measures $\nu_k$ associated to an $\mathbb O$-measure $\nu$.
\end{problem}
\section {Probability measures.}
The author is far from expert in probability and statistics. Therefore it is again quite likely that at least some of the ideas presented in this section already exist in some form. Advice will be appreciated.
If $\mathbb L$ is any ordered abelian group or ordered abelian semi-group with the lub and summability properties, $\mathbb L$-measures make sense. In order to calculate probabilities we need a division operation, so we must work with ordered semi-fields, which are much less common. We also want to avoid any kind of infinite measures, at any level, so our choices are even more limited. Besides $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-measures, one obvious possibility is $\mathbb P=\mathbb Z\circledwedge [0,\infty)$. $\mathbb P$ is a sub-semi-ring of $\mathbb O$, see Figure \ref{OrderParameter}.
Notice that $\mathbb P$ has neither the lub property nor the summability property. For fixed $i\in \mathbb Z$ he set $S=\{(i,t):0<t<\infty\}$ is bounded above by $(i+1,1)$, but it does not have a least upper bound. For this reason, $\mathbb P$ is also not summable. But there is also another reason that $\mathbb P$ is not summable: the sum $\sum (i,1)$ cannot be evaluated in $\mathbb P$.
\begin{defn} If $(X,\Sigma)$ is a measure space, a {\it probability $\mathbb P$-measure} assigns an element $\nu(E)$ of $\mathbb P$ to each set $E\in \Sigma$ such that the following conditions hold:
(i) $\nu(\emptyset)=0$.
(ii) For each $j$, if $\displaystyle X_j=\bigcup_{\mathfrak{L}(\nu(E))=j}E$ and $X_j\ne \emptyset$, then $\mathfrak{R}(\nu(X_j))=1$.
(iii) If $\{E_i\}_{i\in I}$ is a countable collection of pairwise disjoint measurable sets in $X$, then
$$\nu\left(\bigcup_{i\in I} E_i\right )=\sum_{i\in I} \nu(E_i).$$
Often, we assume the probability $\mathbb P$-measure is {\it finite depth}, meaning that the following condition also holds:
(iv) The set $\{\mathfrak{L}(\nu(E)):E\in \Sigma\}$ is finite.
\end{defn}
We observe first that the set $\{\mathfrak{L}(\nu(E)):E\in \Sigma\}$ is bounded by $\mathfrak{L}(\nu(X))$. Condition (ii) in the definition guarantees that the sum in (iii) can be evaluated and is equal to an element in $\mathbb P$.
Probability $\mathbb P$-measures are useful for calculating relative probabilities of ``black swan" events.
We prefer a different point of view using ``depth" instead of height of levels. There is an action of $\mathbb Z$ on $\mathbb O$ and $\mathbb P$ by multiplication. The action takes an element $x=(i,s)$ to $(k,1)x=(k,1)(i,s)=(k+i,s)$ for each $k\in \mathbb Z$, using the multiplication in $\mathbb P$. Therefore we also have an action of $\mathbb Z$ on $\mathbb P$-measures so that $k\in\mathbb Z$ acts on $\nu$ to yield $(k,1)\nu$.
\begin{defn}
Two $\mathbb P$-measures $\nu$ and $\nu'$ are {\it shift equivalent} if $\nu'=(k,1)\nu$ for some $(k,1)\in \mathbb P$, $k\in \mathbb Z$, which means that they are in the same orbit of the $\mathbb Z$ action on $\mathbb P$-measures.
Any finite depth $\mathbb P$-measure $\nu$ is equivalent by level shift to a measure $\mu$ satisfying $-d\le \mathfrak{L}(\nu(E))\le 0$ such that there exist measurable sets $A$ and $B$ such that $\mathfrak{L}(\nu(A))=-d$ and $\mathfrak{L}(\nu(B))=0$. Using level alignment, we can further modify the measure (while possibly decreasing $d$) such that for each $j$ satisfying $-d\le j\le 0$ there exists $C$ such that $\mathfrak{L}(\nu(C))=j$. When all of these conditions are satisfied, we say $\nu$ is {\it standard}. We also say that the measure has {\it total depth} $d$.
For a standard probability $\mathbb P$-measure, we say the probability $\nu(E)$ has {\it depth $j$} if $\mathfrak{L}(\nu(E))=-j$.
\end{defn}
Using the depth terminology, we can now explain probabilities of black swan events. If $\nu(E)$ has depth 0, we can imagine that $\mathfrak{R}(\nu(E))$ represents the traditional probability. If $\nu(E)$ has greater depth, it has traditional probability 0, but it has a ``higher depth" non-zero probability. In this way, we assign probabilities
to black swan events whose traditional probability is 0. Greater depths correspond to higher ``orders of improbability."
\begin{ex} \label{Dart} (Dartboard example) This is an extremely simple example of a finite depth probability measure. Suppose we are given a circular dart board $X$ with just a single cross $Y$ drawn on it, the cross consisting of vertical and horizontal diameters. We can suppose the dart board has area 1, and suppose for simplicity that if $E$ is Lebesgue measurable with measure $\mu(E)$, then the darts have probability $\mu(E)$ of hitting $E$. The traditional probability measure assigns the measure (area) of a Borel set $E$ to the event that the dart lands on a point of $E$. Obviously there are many probability 0 events. For example, the probability that the dart hits the 1-dimensional cross is 0. We may define a $\mathbb P$-measure $\nu$ as follows. For any event $E$ with positive measure in the usual sense, we define $\nu(E)=(0,e)$ where $e=\mu(E)$ is the Lebesgue measure of $E$. If $\mu(E)$ is 0, we define $\nu(E)=(-1,\ell)$ where $\ell$ is the 1-dimensional measure (length) of $E\cap Y$. We assume that the total length of $Y$ is 1, and that each ray has length 1/4. The total depth of this $\mathbb P$-measure is 1; it has two levels. Clearly we could also define a $\mathbb P$-measure with total depth 2 (having three levels) by concentrating the depth two measure at the crossing point at the center of the dart board.
\end{ex}
The practical usefulness of of $\mathbb P$-measures comes from the fact that the usual formulas for conditional probability apply and give reasonable answers.
\begin{defn} If $A$ and $B$ are events in $X$ and $\nu$ is a probability $\mathbb P$-measure on $X$, then the {\it conditional probability of the event $A$ given $B$} is $P(A|B)=\nu(A\cap B)/\nu(B)$.
\end{defn}
\begin{example}[Dart board example continued] Suppose $B$ is the event that the dart hits the closed upper half of the dartboard. and $Y$ is the event that the dart hits the cross. Then $P(Y|B)=\nu(Y\cap B)/\nu(B)=(-1,3/4)/(0,1/2)=(-1,3/4)(0,2)=(-1,3/2)$. We obtain a probability at depth 1, but this probability is greater than $P(Y)=(-1,1)$. This says that if the level -1 probability measure on $Y\cap B$ were ``proportionately distributed" with respect to the level 0 measure on $B$, the the relative probability would be $(-1,1)$, but $Y\cap B$ has 3/2 times its share of the measure, it is ``over-represented" and the relative probability is $(-1,3/2)$.
Now suppose $A$ is the event that the dart hits the vertical ray in the upper half of the dart board. Then $P(A|Y)=(-1,1/4)/(-1,1)=(0,1/4)$, in other words the conditional probability is 1/4 as one would guess. We can calculate $P(A|B\cap Y)=(-1,1/4)/(-1,3/4)=(0,1/3)$, again as one would expect. Another less obvious example is $P(A|B)=(-1,1/4)/(0,1/2)=(-1,1/2)$. If the measure of $A$ were proportionately distributed in $B$, we would have $P(A)=(-1,1/2)$ and $P(A|B)=(-1,1)$, but $A$ is under-represented in $B$.
\end{example}
Finally, Bayes' Theorem holds in our more general probability theory.
\begin{thm} [Bayes Theorem] Suppose $(X,\Sigma)$ is a measure space and $\nu$ is a probability $\mathbb P$-measure, which we use to calculate probabilities. Suppose $\{A_i\}$ is a partition of the event space $X$. Then $P(B)={\sum _{j}P(B|A_{j})P(A_{j})}$ and
$$P(A_{i}|B)={\frac {P(B|A_{i})\,P(A_{i})}{\sum \limits _{j}P(B|A_{j})\,P(A_{j})}}\cdot$$
\end{thm}
\begin{example}[Dart board example continued] We may suppose that the event space is the dartboard itself. Let $A_1$ be the closed first quadrant of the dartboard, let $A_2$ be the interior of the second quadrant, let $A_3$ be the closed third quadrant with the center removed, and $A_4$ is the interior of the fourth quadrant.
$B$ is the event that the dart hits the horizontal line of the cross. Now we can calculate the various probabilities in Bayes' Formula to calculate $P(A_1|B)$. For example, $P(B|A_1)=P(B\cap A_1)/P(A_1)=(-1,1/4)/(0,1/4)=(-1,1)$. Similarly we calculate $P(B|A_2)=0$, $P(B|A_3)=(-1,1/4)$, $P(B|A_4)=0$. Then Bayes formula gives:
$$P(A_1|B)=\frac {(-1,1)(0,1/4)}{(-1,1)(0,1/4)+(-1,1)(0,1/4)}=\frac{(-1,1/4)}{(-1,1/2)}=(0,1/2).$$
This says the event $P(A_1|B)$ has probability 1/2 in the usual sense.
\end{example}
One can use $\mathbb P$-probability distributions to produce a probability $\mathbb P$-measure starting with an arbitrary $\mathbb P$-measure on an event space $X$, possibly a Lebesgue measure. In particular, given a probability $\mathbb P$-measure on $X$, one can obtain another probability measure from a $\mathbb P$-distribution on $X$.
\begin{defn} Suppose $(X,\mu)$ is a probability $\mathbb P$-measured space. Let $f:X\to \mathbb P$ be a function with the property that for every measurable $E\subset X$, $\nu=\int_Ef d\mu\in \mathbb P$. In general, the integral has a value in $\bar\mathbb O$ and we are requiring that the value never be $(i,\infty)$ or $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$. The measure $\nu=\int_Ef d\mu$ can then be made into a probability $\mathbb P$-measure by suitably normalizing the associated $\nu_j$ such that if $\displaystyle X_j=\bigcup_{\mathfrak{L}(\nu(E))=j}E$ and $X_j\ne \emptyset$, then $\mathfrak{R}(\nu(X_j))=1$
\end{defn}
To finish this section, we ask the question whether we can define reasonable probability measures with values in other ordered algebraic structures. Some good candidates are the following:
\begin{defn} Let
$$\mathbb P_n=(\mathbb Z\circledwedge(\mathbb Z\circledwedge\cdots\circledwedge(\mathbb Z\circledwedge(\mathbb Z\circledwedge [0,\infty)))\cdots))$$
where $\mathbb Z$ appears $n$ times in the formula. We will denote elements of this semi-field as \\ $(i_1,i_2, \ldots, i_n, s)\in \mathbb P_n$, and we let $\mathfrak{L}((i_1,i_2, \ldots, i_n, s))=(i_1,i_2, \ldots, i_n)$.
If $(X,\Sigma)$ is a measure space a {\it probability $\mathbb P_n$-measure} assigns an element $\nu(E)$ of $\mathbb P_n$ to each set $E\in \Sigma$ such that:
\begin{tightenum}
\item $\nu(\emptyset)=0$,
\item $\nu(E)=(i_1,i_2, \ldots, i_n, s)$ satisfies $i_j\le0$ and $s\le 1$.
\item If $\{E_i\}_{i\in I}$ is a countable collection of pairwise disjoint measurable sets in $X$, then
$$\nu\left(\bigcup_{i\in I} E_i\right )=\sum_{i\in I} \nu(E_i).$$
\end{tightenum}
We
observe that we have already built into our definition a suitable choice of representative by shift equivalence.
We can also perform alignment operations to ensure there are no ``gaps," and $\nu$ is {\it standard}:
\noindent (iv) \ If for some measurable $E$, $\nu(E)=(i_1,i_2,\ldots, i_n,s)$ for some $s>0$, and if $(j_1,j_2,\ldots, j_n)>(i_1,i_2,\ldots, i_n)$ in the lexicographical order, then there exists a measurable set $E'$ with $\nu(E')= (j_1,j_2,\ldots, j_n,t)$ for some $t>0$. \end{defn}
We note first that $\mathbb P_n$ is a semi-field by Lemma \ref{WellDefinedLemma} (c), so the division operation makes probability calculations possible. Again, although $\mathbb P_n$ does not have the lub property, condition (ii) ensures that partial sums of the sum in (iii) do have a least upper bound, so the sum makes sense.
It is difficult to imagine that even probability $\mathbb P_2$-measures could have any practical applications, but probability $\mathbb P$-measures certainly could have applications.
\section{Borel measures.}
Given a Hausdorff topological space $X$, a Borel measure is a assigns a measure to each set in a $\sigma$-algebra generated by open sets. Usually the values of the measure lie in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$. Suppose $\mathbb L$ is any ordered abelian semigroup with the summability property.
\begin{defn}\label{AMeasureDef} A {\it Borel $\mathbb L$-measure} $\nu$ on a Hausdorff topological space $X$ assigns an element $\nu(E)$ of $\mathbb L$ to each Borel set $E$ such that:
(i) $\nu(\emptyset)=0$,
(ii) If $\{E_i\}_{i\in I}$ is a countable collection of pairwise disjoint Borel sets in $X$, then
$$\nu\left(\bigcup_{i\in I} E_i\right )=\sum_{i\in I} \nu(E_i).$$
\end{defn}
\begin{example} We modify the dartboard example of the previous section. $X$ is the unit disk in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^2$, $Y$ is the cross in $X$ consisting of a vertical diameter union a horizontal diameter. With the usual subspace topology in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^2$, and the usual measure $\mu$ on $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^2$, we will define an Borel $\bar \mathbb S$-measure $\nu$ on $X$. It assigns $\nu(E)=(1,\mu(E))\in \bar\mathbb S $ if $\mu(E)\ne 0$, and it assigns the 1-dimensional real measure of $E\cap Y$ otherwise. More precisely, let $\rho$ be a measure on $Y$ agreeing with a metric which assigns length 1/4 to each ray of $Y$, then if $\mu(E)=0$, we define $\nu(E)=(0,\rho(E\cap Y))$. One can verify that this is a Borel $\bar \mathbb S$-measure. In fact, since all measures are finite at each level we have summability for measures in $\mathbb S$, so we could call this a Borel $\mathbb S$-measure.
\end{example}
In one way, the Borel measure in the above example is not natural, although it is natural in the setting of probability theory. Namely, there is no connection between the topology of $\bar \mathbb S$ and the topology of $X$. This motivates the following definition:
\begin{defn}\label{OpenGradedMeasureDef} A Borel $\bar\mathbb S$-measure or a Borel $\bar\mathbb O$-measure $\nu$ on a Hausdorff topological space $X$ is {\it open-graded} if the union $\bigcup E$ of Borel sets $E$ satisfying $\nu(E)< (k,\infty)$ or $\nu(E)=0$ is open in $X$. \end{defn}
The open-graded property is a property we will want for transverse measures on laminations.
\begin{ex} We let $X=[0,\infty]$, a subspace of the extended reals. We define an open-graded Borel $\bar\mathbb O$-measure essentially by identifying $X$ with our picture of $\bar\mathbb O$ in Figure \ref{OrderParameter}. More specifically, choose a homeomorphism $f:X\to \bar\mathbb O$. We also have an identification $h:\mathbb O\to \{0\}\cup(\mathbb Z\times (0,\infty])$, and we let $\pi:(\mathbb Z\times (0,\infty])\to (0,\infty]$ be the projection map. We let $\mu$ be Lebesgue measure on $(0,\infty]$. Given a Borel measurable set $E$ let $\nu(E)=(j,\mu( \pi (h\circ f(E)\cap j\times (0,\infty]))$, where $j$ is the largest $i$ such that $\mu(\pi (h\circ f(E)\cap j\times (0,\infty]))$ is non-zero. Exceptionally, when every $\mu(\pi (h\circ f(E)\cap j\times (0,\infty]))$ is $0$, we define $\nu(E)=0$, and if every $\mu(\pi (h\circ f(E)\cap j\times (0,\infty]))$ is non-zero we let $\nu(E)=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$. Now it is easy to verify that $\nu$ is open-graded.
\end{ex}
\section{Laminations, metric spaces, and trees.}
Suppose $L$ is an essential codimension-1 lamination in a compact surface $S$, or a surface with cusps as in \cite {UO:Depth}, with $\chi_g(S)<0$. Up to minor modifications, essential laminations in a surface with $\chi_g(S)<0$ can be realized as geodesic laminations for any chosen hyperbolic structure on $S$. We will usually assume that we have chosen a hyperbolic structure.
The trees dual to lifts of arbitrary essential laminations are more general trees called ``order trees." There is a definition in
\cite{DGUO:EssentialLaminations}, but we give a different definition. I am not sure who first defined these; I first heard about order trees from Peter Shalen.
\begin{defn} An {\it order tree} is a set $\mathscr{T}$ together with a subset $[x,y]$, called a {\it segment}, associated to each pair of elements, together with a linear order on $[x,y]$ such that $x$ is the least element in $[x,y]$ and $y$ is the greatest element. We allow {\it trivial segments} $[x,x]$. The set of segments should satisfy the following axioms:
\begin{tightenum}
\item The segment $[y,x]$ is the segment $[x,y]$ with the opposite order.
\item The intersection of segments $[x,y]$ and $[x,z]$ is a segment $[x,w]$.
\item If two segments intersect at a single point, $[x,y]\cap[y,z]=\{y\}$ then the union is a segment $[x,z]$.
\end{tightenum}
The order tree is a topological space: $G$ is open in $\mathscr{T}$ if for every segment $[x,y]$, $G\cap [x,y]$ is open in the order topology for $[x,y]$.
\end{defn}
Clearly $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-trees and $\bar\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-trees are also order trees, as are $\Lambda$-trees, where $\Lambda$ is an ordered abelian group. $\Lambda$-trees have been studied extensively; it would be difficult to give an adequate list of references.
Suppose now that $\mathbb L$ is any ordered abelian semi-group.
\begin{defn} An {\it $\mathbb L$-metric} on a set $X$ is a function $d:X\times X\to \mathbb L$ satisfying the usual axioms for a metric. An {\it $\mathbb L$-metric space} is the set $X$ together with an $\mathbb L$-metric.
\end{defn}
Henceforth in this section we will assume that $\mathbb L$ is an ordered abelian semi-group with the summability property; we will use Borel $\mathbb L$-measures.
\begin{defn} Suppose $\mathscr{T}$ is an order tree and suppose $\nu$ is a Borel $\mathbb L$-measure $\nu$ on the disjoint union of segments of $\mathscr{T}$ with the property that if $[x,y]$ and $[z,w]$ are segments, and $[x,y]\cap [z,w]=[u,v]$, then for any measurable set $E\subset [u,v]$, $\nu(E)$ is the same no matter which segment ($[x,y]$,$ [z,w]$, or $[u,v]$) we use to evaluate the measure. (The measure agrees on intersections of segments.) We say $\nu$ is an {\it $\mathbb L$-measure on $\mathscr{T}$}, and we say $\mathscr{T}$ with the measure $\nu$ is called an {\it $\mathbb L$-tree}.
The $\mathbb L$ measure on an order tree is {\it non-atomic} if the measure of a single point in a segment is always $0$. It has {\it full support} if it has full support on the disjoint union of segments.
\end{defn}
\begin{lemma} Suppose $\mathscr{T}$ is an $\mathbb L$-tree with a non-atomic full support measure $\nu$. Then $\mathscr{T}$ is an $\mathbb L$-metric space with metric $d(x,y)=\nu([x,y])\in \mathbb L$.
\end{lemma}
\begin{proof} Because $\nu$ has no atomic measures on points, we conclude $d(x,y)=\nu([x,y])=0$ if and only if $x=y$. To verify the triangle inequality, observe that if $x,y,z$ are points in the tree, by axiom (iii) for order trees, $[x,y]\cap [x,z]=[x,w]$ for some $w$, so $[y,w]\cup [w,z]=[y,z]$ by axiom (iii). Hence $d(y,z)=\nu ([y,z])=\nu([y,w])+\nu([w,z])\le \nu([y,x])+\nu([x,z])=d(y,x)+d(x,z)$, because $[y,w]\subset [y,x]$ and $[w,z]\subset [x,z]$.
\end{proof}
\begin{defn} Suppose $L$ is an essential lamination in $S$.
A {\it transverse $\mathbb L$-measure for $L$} is an assignment of a value $\nu(T)\in \mathbb L$ to every closed transversal $T$ of the lamination. The assignment must be invariant, in the sense that an isotopy of $T$ through transversals of $L$ moving endpoints of the transversal in a leaf of $L$ or in the completion of a component of $S\setminus L$ leaves $\nu(T)$ unchanged. An {\it $\mathbb L$-measured lamination} $(L,\mu)$ is a lamination with a transverse $\mathbb L$-measure $\mu$.
The {\it order tree dual to the lift $\tilde L$ of $L$ to the universal cover $\tilde S$ of $S$} is the set of closures of complementary regions of $\tilde L$ union non-boundary leaves. A {\it segment} $[x,y]$ is the set elements of $\mathscr{T}$ intersected by closed oriented efficient transversal $T$ for $\tilde L$ with order coming from the order on the transversal. \end{defn}
If a codimension-1 lamination admits a transverse $\bar\mathbb S$-measure or a transverse $\bar\mathbb O$-measure which is open-graded, then the support of transverse measure $\nu_k$ is a lamination, and we have a ``finite height measured lamination," as described in \cite{UO:Depth}.
Evidently a transverse $\mathbb L$-measure for a codimension-1 essential lamination in a manifold $S$ yields a measure on the dual tree.
The following proposition was proved in a more restricted setting in \cite{UO:Depth}.
\begin{proposition}\label{DualMeasureProp} Suppo
se $S$ is a compact orientable surface or a surface with cusps as in \cite{UO:Depth}, with $\chi_g(S)<0$.. Given an essential lamination $L$ in $S$, the associated dual tree $\mathscr{T}$ defined above with the given segments is an order tree.
If $L$ is $\mathbb L$-measured, with measure $\mu$, then the lifted measure $\nu=\tilde\mu$ yields an $\mathbb L$-measure $\nu$ for $\mathscr{T}$. If $\mu$ has no leaves with atomic transverse measures, $\mathscr{T}$ is an $\mathbb L$-metric space with metric $d(x,y)=\nu([x,y])$ for $x,y\in \mathscr{T}$.
\end{proposition}
\begin{proof} The first task is to show $\mathscr{T}$ is an order tree. Since $\chi_g(S)<0$, we can choose a hyperbolic structure for $S$.
Suppose $x,y$ are points in $\mathscr{T}$, representing leaves or complementary components $X$ and $Y$. A geodesic $\gamma$ from a point in $X$ to a point in $Y$ gives an efficient transversal, hence a segment $[x,y]$ in $\mathscr{T}$. The uniqueness of this segment is also easy to verify: Suppose $\beta$ is another efficient transversal from $X$ to $Y$. Choose a geodesic segment $\omega$ in $Y$ joining the final point of $\gamma$ in $ Y$ to the final point of $\beta$ in $ Y$, and similarly choose a geodesic segment $\rho$ joining the initial in point $\gamma\cap X$ to the initial point of $\beta$ in $X$. Since $\gamma\omega\bar\beta\bar\rho$ is null homotopic, we obtain a map $h:R\to \tilde S$ of a square $R$ to $\tilde S$ whose sides are mapped to $\gamma,\beta, \rho,\omega$. The null-homotopy $h$ can be homotoped such that it is transverse to $\tilde L$, and can then be further homotoped such that the induced lamination on $R$ consists of leaves joining opposite sides of $R$ mapped to $\gamma$ and $\beta$. This shows that $\gamma$ and $\beta$ yield the same segment in $\mathscr{T}$.
We verify the order tree axioms: (i) is true by construction, $[y,x]$ is $[x,y]$ with the opposite order, coming from a transversal with the opposite orientation.
For (ii), consider oriented geodesic transversal segments $\gamma$ from a point in $X$ to a point in $Y$, and $\beta$ from a point in $X$ to a point in $Z$. We may choose $\gamma$ and $\beta$ so that they do not intersect. Choose a geodesic segment $\omega$ joining the final point of $\gamma$ to the final point of $\beta$, and similarly choose a geodesic segment $\rho$ joining the initial point of $\gamma$ in $ X$ to the initial point of $\beta$ in $ X$. The simple closed $\gamma\omega\bar\beta\bar\rho$ bounds a rectangular disk $R$ in $\tilde S$ and $\tilde L\cap R$ is a lamination in $R$ which is transverse to two opposite sides $\gamma$ and $\beta$ with $\rho\subset\partial R$. Consider the set of leaves of $\tilde L\cap R$ joining $\gamma$ to $\beta$. This includes at least $\rho$ and it must be closed. So there is a largest element $w$ in $[x,y]$ which is also in $[x,z]$.
For property (iii), suppose $[x,y]$ and $[y,z]$ are (non-trivial) segments in $\mathscr{T}$ with $[x,y]\cap [y,z]=\{y\}$. Representing $[x,y]$ by an oriented geodesic segment $\beta$ and $[yz]$ by an oriented geodesic segment $\gamma$ whose initial point is the final point of $\beta$, we see that $\beta\cup \gamma$ must be an embedded path. It follows that $\beta\cup \gamma$ can be regarded as a transversal, representing $[x,z]$.
Now that we know that $\mathscr{T}$ is an order tree, it is easy to show it is an $\mathbb L$-tree. The transverse $\mathbb L$-measure $\mu$ for $L$ yields a transverse measure $\tilde \mu$ for $\tilde L$, which in turn gives a measure on transversals. Since transversals are identified with segments of $\mathscr{T}$, we have a measure $\nu$ on the transversals. Invariance of the measure $\tilde \mu$ gives an $\mathbb L$-measure $\nu$ on the disjoint union of segments of $\mathscr{T}$. If there are no leaves of $L$ with atomic measure, there are no points with atomic measure in (the segments of) $\mathscr{T}$, which shows that $d(x,y)=\nu([x,y])$ defines an $\mathbb L$-metric on $\mathscr{T}$.
\end{proof}
In \cite{UO:Depth}, the author investigated laminations with transverse $\mathbb S$-measures in a given surface $S$, defining and investigating a space of such laminations, and also investigating actions of $\pi_1(S)$ on $\mathbb S$-trees. The same program can be followed using other ordered abelian semigroups with the summability property. In fact, the first attempt in \cite{UO:Depth} used $\mathbb O$, but it turned out that $\mathbb S$-measured laminations yielded nicer lamination spaces, though the transverse structures are quite similar. For the behavior of leaves, other transverse $\mathbb L$-structures probably encode more subtleties in the dual actions on $\mathbb L$-trees. In fact, we will give examples of laminations with ``local transverse $\mathbb L$-structures", see Section \ref{LocalSection}, which suggest that many or all laminations admit these structures. On the other hand, investigating lamination spaces for more complicated choices of $\mathbb L$ may not be worthwhile.
There is a version, not included here, of Proposition \ref{DualMeasureProp} which applies to any essential lamination in a 3-manifold with a transverse $\mathbb L$-measure. Whether we are working with codimension-1 laminations in surfaces or 3-manifolds (or even higher dimensional manifolds), it is useful to use branched manifolds with invariant $\mathbb L$-weight vectors to describe the lamination. As we have observed before in \cite{UO:Depth} (when $\mathbb L=\mathbb S$) an invariant $\mathbb L$- weight vector on a codimension-1 branched manifold embedded in a manifold does not necessarily determine an embedded $\mathbb L$-measured lamination. We will assume familiarity with the material in \cite{UO:Depth}, but we recall that if a lamination with a transverse $\mathbb L$ measure is carried by a branched manifold (train track or branched surface), then the transverse measure induces an $\mathbb L$ weight on each sector (segment) of the branched manifold, and these weights satisfy branch equations, just as real transverse measures induce real weights on sectors satisfying branch equations. Recall also that the points in the
branch locus of a train track are called {\it switch points}.
If $B$ is compact with finitely many sectors, finitely many branch equations suffice to determine whether a weight vector is invariant. When $\mathbb L=[0,\infty)\subset \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$, an invariant weight vector uniquely determines a measured lamination carried by $B$. This is not in general true for arbitrary $\mathbb L$. A reasonable conjecture is that if $L\hookrightarrow M$ is a lamination with a transverse $\mathbb O$-measure or $\mathbb S$-measure $\nu$, and is carried by a codimension-1 branched manifold, then the weights induced by $\nu$ on some splitting of the branched manifold $B$ determine $L$ and its transverse measure.
\section {Mixed insertions for ordered abelian semi-groups.}
We can define more general insertion operations yielding ordered abelian semi-groups. Namely, we can insert different semigroups at different levels.
\begin{defn} Suppose $A$ is an ordered abelian group or an ordered abelian semigroup. For each $a\in A$, let $B_a$ be an ordered abelian semigroup. Then we define
$$A\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}B_a=\left\{ (a,b): a\in A, b\in (B_a\setminus\{0\})\right\}\cup\{0\}$$
\noindent We make this set into an ordered abelian semi-group by defining the order relation and the addition operation as follows:
\begin{tightenum}
\item $(g,s)<(h,t)$ if either if $g<h$ or if $g=h$ and $s<t$.
\item $0<(g,s)$ for all $(g,s)$.
\end{tightenum}
The addition operation is commutative and given by
$$(g,s)+(h,t)
=\begin{cases}( g,s) & \mbox{if } g>h \\
(g,s+t) & \mbox{if } g=h
\end{cases}
$$
$$0+(g,s)=(g,s)+0=(g,s)$$
If we wish to extend $\displaystyle A\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}B_a$ by including an infinity, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$, then we define
$$A\bar{\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}}B_a=A\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}B_a\cup \{\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}\}$$
and extend the order relation and addition operation as before, so that for all $x\in A\bar\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}B_a$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}>x$, $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}+x=x+\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}=\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$.
As before, when we denote a non-zero element of $A\bar\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\in A}B_a$ by a single symbol $x=(g,s)$, we will use $\mathfrak{L}(x)=g$ to denote the {\it level of $x$} and $\mathfrak{R}(x)=s$ to denote the {\it residue of $x$}, which lies in $B_g\setminus\{0\}$. We make the convention that $\mathfrak{R}(0)=0$ and $\mathfrak{L}(0)$ is undefined (and $\mathfrak{L}(\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt})$ is undefined if $\hbox{$\infty$\kern -8.1pt\raisebox{0.2pt}{--}\kern 1pt}$ is included).
In the same way, we define $$A\mathop{\mathlarger{\mathlarger\circledwedge}}_{k\le a\le h}B_a=\left\{ (a,b): a\in A, k\le a\le h, b\in (B_a\setminus\{0\})\right\}\cup\{0\}$$
which we will also write $\displaystyle A\mathop{\mathlarger{\mathlarger\circledwedge}}_{a=k}^hB_a$. We can similarly define $\displaystyle A\mathop{\mathlarger{\mathlarger\circledwedge}}_{a\le h}B_a$ and $\displaystyle A\mathop{\mathlarger{\mathlarger\circledwedge}}_{k\le a}B_a$.
\end{defn}
\begin{example} Let $S$ be a surface and let $\mathbb L$ be the ordered abelian group $\displaystyle \mathbb N_0\mathop{\mathlarger{\mathlarger\circledwedge}}_{0\le n\le 2} B_n$ where $B_0=[0,\infty]$, $B_1=[0,\infty]$ and $B_2= \bar\mathbb N_0$. What are the essential laminations in $S$ with transverse $\mathbb L$-measures? At the highest level, they have integer coefficients, so they are curve systems. In the complement of the curve systems, at the next level, we have a measured lamination, and in the complement of the union of the curve system with a measured lamination, we have another measured lamination. (We are ignoring the possibility of locally infinite measures, which are also allowed.) In any case, we see that $\mathbb L$-measured laminations have a very particular structure, depending on $\mathbb L$. (In this example, the $\mathbb L$- measured laminations form a subspace of the space described in \cite{UO:Depth}, because $\mathbb L$ is a sub semi-group of $\mathbb S$.)
\end{example}
\section {Laminations with local transverse $\mathbb P$-structures.}\label{LocalSection}
There is a well-known method for constructing more interesting laminations with a given local transverse structure. Suppose we are given a lamination $L\hookrightarrow M$, where $M$ is a manifold, and the lift $\tilde L$ to the universal cover $\tilde M$ has a transverse real structure $\mu$. Further suppose that the action of $\pi_1(M)$ on $\tilde L$ preserves $\mu$, but only up to scalar multiplication. Thus there is a {\it stretch homomorphism} $\phi: \pi_1(M)\to \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}_+$ such that for $\gamma\in \pi_1(M)$, the translate $\gamma (\tilde L,\mu)=(\tilde L,\phi(\gamma)\mu)$. This is called a lamination with a transverse {\it affine structure}, see \cite{UO:Stretch}, \cite{HO:AffineSpaces}, \cite{OP:Affine}, \cite{OP:Broken}. There are many laminations which do not have transverse $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-measures, but do admit affine structures. (Usually we deal with codimension-1 laminations.) The homomorphism $\log( \phi)$ can be interpreted as a cohomology class $\log\phi\in H^1(M;\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex})$.
We can play the same game with transverse $\mathbb L$ structures, but we want $\mathbb L$ to be a multiplicative group so that we can again define a homomorphism $\phi:\pi_1(M)\to (\mathbb L,\cdot)$. Thus we are forced to use $\mathbb L=\mathbb P$, or $\mathbb L=\mathbb P_n$.
\begin{defn} A {\it local transverse $\mathbb P$-structure} for a lamination $L\hookrightarrow M$ is a transverse $\mathbb P$-measure $\mu$ for the lift $\tilde L$ of $L$ to the universal cover $\tilde M$ of $M$ with the property that there exists a homomorphism $\phi:\pi_1(M)\to (\mathbb P,\cdot)$ such
that $\gamma (\tilde L,\mu)=(\tilde L,\phi(\gamma)\mu)$. We define a {\it local transverse $\mathbb P_n$ structure} for a lamination in the same way, replacing $\mathbb P$ by $\mathbb P_n$ in the definition.
Regarding $\mathbb P$ as a product $\mathbb Z\times (0,\infty)$, the homomorphism $\phi$ yields two homomorphisms: $\phi_1:\pi_1(M)\to \mathbb Z$, which can be interpreted as an element of $H^1(M;\mathbb Z)$ and is called a {\it level shift homomorphism}; and $\phi_2:\pi_1(M)\to \hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}_+$, called the {\it stretch homomorphism}, where $\log \phi_2$ can be interpreted as a cohomology class in $H^1(M;\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex})$.
\end{defn}
\begin{exs} We will give three related examples here, two of them are laminations carried by the same branched surface in a 3-manifold.
In Figure \ref{Local}(a), we see a branched surface $B$ shown immersed in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^3$. Actually, it can be embedded in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^3$, and we assume it is embedded and that $M$ is a regular neighborhood of $B$. In the figure, we show weights on $B$, which are values of $\mathbb P$. We also see a transversely oriented curve $\alpha$ representing a cohomology class, with multiplier $(0,1/2)$ in the transverse direction. Moving from one side of the cohomology class to the other, the weight is multiplied by $(0,1/2)$ in $\mathbb P$. The weights and the cohomology class represent a local $\mathbb P$ transverse structure on a lamination which is determined by the weights. In fact, this is an affine lamination. Since the levels of all weights and the multiplier are $0$, we have real weights and a real multiplier. Cutting open the branched surface on the curve representing the cohomology class, the weights on the resulting branched surface $\hat B$ represent an $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}$-measured lamination. Glueing with a stretch of 1/2 yields the affine lamination.
\begin{figure}[H]
\centering
\scalebox{1}{\includegraphics{Local}}
\caption{\small Two examples of laminations with local transverse $\mathbb P$ structures.}
\label{Local}
\end{figure}
Next we present an example with only a level shift homomorphism, using the same branched surface. We hsow weights and the cohomology class at the curve $\beta$ with multiplier $(-1,1)$, but now the multiplier at the cohomology class shifts levels. The lamination represented by the data has a ``leaf spiraling on itself," also sometimes called a spring leaf. The lamination is completely different from the one in the first example.
For a final example, we change the branched surface $B$ as shown in Figure \ref{LocalMix}. This branched surface can also be embedded in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^3$ and we assume $M$ is a regular neighborhood in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^3$ of $B$. We show a transversely oriented curve $\alpha$ with multiplier $(0,1/2)$ representing the stretch homomorphism, and another transversely oriented curve $\beta$ with multiplier $(-1,1)$ representing the level shift homomorphism. Again the data determine a lamination with a local transverse $\mathbb P$ structure.
\begin{figure}[H]
\centering
\scalebox{1}{\includegraphics{LocalMix}}
\caption{\small Example of a lamination with a local transverse $\mathbb P$ structure.}
\label{LocalMix}
\end{figure}
We observe that in all of these examples, the data determine the lamination. That is not true in general.
\end{exs}
Turning to local transverse $\mathbb P_n$-structures, we have the same definitions, but with $n$ level shift homomorphisms and one stretch homomorphism.
\begin{example} Here is an example of a lamination with a transverse local $\mathbb P_2$-structure. In this example, we take the stretch homomorphism to be trivial, so we have two level shift homomorphisms. Again, we show a branched surface $B$ embedded which can be embedded in $\hbox {\rm {R \kern -2.8ex I}\kern 1.15ex}^3$, and which we assume is so embedded, and we let $M$ be a regular neighborhood of $B$. This is actually the same branched surface shown in Figure \ref{LocalMix}. The two level shift homomorphisms are represented by curves with multipliers in $\mathbb P_2$, $\alpha$ with multiplier $(-1,0,1)$ and $\beta$ with multiplier $(0,-1,1)$.
\begin{figure}[H]
\centering
\scalebox{1}{\includegraphics{OrderedP2Ex}}
\caption{\small Example of a lamination with a local transverse $\mathbb P_2$-structure.}
\label{OrderedP2Ex}
\end{figure}
\end{example}
Suppose $L$ is an essential lamination in a 3-manifold $M$. It should be easy to describe an associated action of $\pi_1(M)$ on a $\mathbb P_n$-tree. The action does not necessarily preserve $\mathbb P_n$ measures on segments; rather, it transforms measures according to stretch and level shift homomorphisms.
Is it possible that any essential 2-dimensional lamination embedded in a 3-manifold admits a full support local transverse $\mathbb P_n$-structure for some $n$ in some regular neighborhood of itself? What about codimension-1 laminations and foliations in higher dimensional manifolds?
\bibliographystyle{amsplain}
|
\section{Background}
To validate the identity of a website such as Amazon, the web browser
relies on certificate authorities (CAs), which digitally sign
certificates vouching for the identity of the web server.
Figure~\ref{fig:certificate} provides a high-level overview of a
TLS handshake where the browser initiates a TLS connection
with the server, and retrieves the server's certificate. The browser
verifies this certificate and, if it is valid, sends random
data securely to the server that is used to seed the generation of
the necessary encryption keys.
The browser and the server can now exchange encrypted traffic.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/normal_tls_illustration}
\caption{High-level view of secure session establishment via TLS}
\label{fig:certificate}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/certvalidation_substitute.pdf}
\caption{Certificate validation. Part (a) shows expected behavior when the www.google.com webserver provides its certificate to the browser for validation. The certificate's signature is validation against an intermediate certificate, Google Internet Authority G2, which is in turn signed by a certificate in the trusted root store. Part (b) shows a substitute certificate invalidated due to lack of any signatures leading back to a root certificate. Part (c) shows a validated substitute certificate issued by a rogue or compromised certificate authority (left arrow) and a validated substitute certificate signed by an authority who injected a new trusted certificate into the root store.}
\label{fig:certificatevalidation}
\end{figure}
Web browsers authenticate a site by validating a chain of digital signatures from the site's certificate back to one of a set of trusted root certificates. These certificates comprise the ``root store'' and are
typically bundled with the operating system or browser. For example,
the certificate for www.google.com is signed by the Google Internet
Authority G2, an intermediate certificate authority run by
Google. This certificate is in turn signed by GeoTrust Global CA, a
certificate authority whose certificate is located in the root store
of the browser or operating system.
This is exemplified in Figure~\ref{fig:certificatevalidation}(a). The legitimate certificate for google.com is accepted because it has a valid chain of signatures leading back to a trusted certificate in the root store.
Figure~\ref{fig:certificatevalidation}(b) shows a case where a substitute certificate's signature cannot be traced back to a root store certificate and is rejected.
This is expected behavior for substitute certificates.
This system can be attacked by a TLS proxy inserting itself as a
man-in-the-middle between the browser and the web server. As shown in
Figure~\ref{fig:proxy}, when the browser tries to open a secure
connection to the web server, this connection is instead intercepted
by the proxy. The proxy also provides a falsified, substitute
certificate to the browser, so that it can impersonate the
original website. For this to work, the proxy must somehow control a substitute certificate for the original website that validates against the root store of the user. This can be accomplished in a variety of ways, both benign and malicious.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/proxy_illustration}
\caption{``Secure'' session establishment with involving a TLS proxy}
\label{fig:proxy}
\end{figure}
Figure~\ref{fig:certificatevalidation}(c) illustrates two ways the operator of a proxy can provide valid
certificates that do not elicit a browser warning. One way (indicated by the vertical arrow) is to directly insert a certificate into the trusted root store. For example, an
organization can use enterprise software to supply its own
certificates for the root store of all computers in the organization.
Alternatively, the IT department of an organization can create a
software image using new root certificates and require that all
computers in the organization use this image. This is generally done
for benevolent reasons, such as blocking malware and viruses,
providing intrusion detection, or protecting intellectual
property. Likewise, personal firewall software often installs a new
root certificate so that it can offer users the feature of scanning
their encrypted traffic for malware. Finally, in the case of mobile
devices, a manufacturer may add certificates to the root store,
ostensibly for speeding up web browsing. Nokia came under fire
recently for doing this \cite{nokia}.
Other ways of providing seemingly-valid certificates are more nefarious. A malicious entity can also install new certificates into an operating system's root store. For example, malware typically
has permission to add new certificates when it is installed
inadvertently by the user. Sometimes this may be as simple as asking
the user for install privileges, and many users are trained to click
"OK" whenever the operating system asks them for this. It is even possible for a malicious entity to act as a proxy without adding any
new certificates to a victim's root store.
This is shown by the left-leaning arrow from the certificate in Figure~\ref{fig:certificatevalidation}(c). A rogue certificate
authority can issue any certificate it wants, since all root certificates are allowed to sign for any domain. There have also been numerous reported cases of compromised and negligent certificate authorities that allow attackers to issue fraudulent certificates \cite{durumeric2013analysis}.
In addition, some governments, such as China and the United States, control their own root certificate
authorities, and evidence suggests that a government may coerce
authorities into granting them certificates for domains they do not
own \cite{marlinspike2011ssl, eckersley2011decentralized}.
In summary, a TLS proxy can be operated by both malicious and benign parties including parental control software, corporate firewalls, malware, hackers, and government surveillance operatives.
\section{Google Adwords Campaigns}
\label{sec:deployment}
To achieve rapid and widespread deployment of our measurement tool we leveraged the Google AdWords platform.
This strategy for using an advertising campaign to conduct an end-user measurement study has previously been used to study CSRF attacks \cite{barth2008robust}, DNS rebinding attacks \cite{jackson2009protecting}, and DNSSEC deployment \cite{huston1,huston2,lian2013measuring}. Our study is the first to use this same method to measure the deployment of TLS proxies.
The results from this study shed light on the legitimate demand for TLS proxies as well as several suspicious or duplicitous practices.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/flashtool_alternate_3.pdf}
\caption{Using an Ad Server for TLS Proxy Measurement}
\label{fig:flashtool_alt_3}
\end{figure}
The deployment of the measurement tool is given in Figure~\ref{fig:flashtool_alt_3}.
Deployment responsibilities were delegated to the Google AdWords platform, while all reports from the tool were sent back to a reporting server we controlled.
To accommodate placement in advertisements, our measurement tool was modified to contain a visible canvas on which we place a simplistic advertisement for our research lab.
Figure~\ref{fig:adimage} shows the advertisement as it appeared to web users during our measurement study.
Our measurement tool was run as soon as user's browsers loaded the advertisement, and required no interaction from the users.
\begin{figure}
\centering
\includegraphics[width=100px,height=100px]{figures/adimage.png}
\caption{Appearance of tool via Google AdWords}
\label{fig:adimage}
\end{figure}
For our ad campaigns we leveraged the CPM (cost-per-impression) bidding model for our campaign, which maximizes the number of unique clients presented with our ad.
We set the Max. CPM to \$10 USD.
To help us reach a global audience we indicated that our ad should be served to all locations and languages.
Additionally, since ads are shown only on websites that match a set of designated keywords we selected our keywords based on phrases that were currently trending globally on Google Trends~\cite{googletrends}.
We set our ad to show uniformly throughout the day so as to collect data from users in a variety of locations and situations (e.g., home, commuting, work).
Along with the certificate, we also recorded the IP address of the client tested.
This IP address was then used to query the MaxMind GeoLite \cite{maxmind} database to gather geolocation information.
\subsection{Campaign Setup for First Study}
Our initial Google AdWords advertising campaign ran from January 6, 2014 to January 30, 2014.
During the first 17 days of the study we varied the amount of money allocated to the ad campaign, but for the last week we kept it at \$500/day.
In this study we only gathered certificate data for our own website, \textit{tlsresearch.byu.edu}.
We used the following keywords for the study:
``Nelson Mandela'',
``Sports'',
``Basketball'',
``NSA'',
``Internet'',
``Freedom'',
``Paul Walker'',
``Security'',
``LeBron James'',
``Haiyan'',
``Snowden'',
``PlayStation 4'',
``Miley Cyrus'',
``Xbox One'',
``iPhone 5s''.
This initial campaign generated 4,634,386 impressions and 3,897 clicks (not required to complete the measurement) at a cost of \$4,911.97.
In total we completed 2,861,244 successful measurements.
Section~\ref{sec:measurementstudy} discusses the results from this first study.
\subsection{Campaign Setup for Second Study}
To increase the number of measurements collected and to better understand the nature of TLS proxies we conducted a second set of measurements approximately eight months after the first study.
One question unanswered by the first study was whether TLS proxies were intercepting all traffic, or whether they selectively intercepted traffic according to white or blacklists.
To shed light on this subject, we decided to gather measurements for different types of sites:
\begin{itemize}
\item {\bf Popular:} Sites from the Alexa top 25,000. Six websites were included in this category.
\item {\bf Business:} Commercial sites unlikely to be blocked by places of business. Five websites were included in this category.
\item {\bf Pornographic:} Pornographic websites (expected to be blocked by parental filters and places of business). Five websites are included in this category.
\item {\bf Authors':} The single website operated by the authors and also used in the first measurement study.
\end{itemize}
The policy restrictions of the Flash runtime prohibit establishment of socket connections to arbitrary hosts.
Thus all sites used in our study had to host permissive ``socket policy files'' that allowed connections to port 443 from any domain.
We scanned for the presence of permissive socket policy files on the entirety of the Alexa top 1 million websites, and selected the highest ranked such websites for each type to use in the second measurement. Table ~\ref{tab:hosts} lists the additional hosts we probed.
\begin{table}[t]
\scriptsize
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{lll}
\toprule
\multicolumn{3}{c}{\normalsize Website}\\
{\normalsize Top 25,000} & {\normalsize Business} & {\normalsize Porn}\\
\midrule
qq.com & airdroid.com & pornclipstv.com\\
promodj.com & webhost1.ru & porno-be.com\\
idwebgame.com & restaurantesecia.com.br & pornbasetube.com\\
parsnews.com & speedtest.net.in & pornozip.net\\
idgameland.com & iprank.ir & pornorasskazov.net\\
vcp.ir\\
\bottomrule
\end{tabular}}
\caption{Second Study Websites Probed}
\label{tab:hosts}
\end{table}
At most 17 of these sites were queried by a single served instance of our Flash measurement tool.
Due to differences in Internet connectivity quality and hardware and software performance, not all clients served with our ad were able to successfully perform TLS handshakes with all hosts.
The tool was configured to first test the connection to the authors' website, before attempting to test connections to the other hosts in parallel.
In this second study, we also targeted specific countries by
creating an additional ad campaign for each country.
The ad image used in both the global and country specific ad campaigns was the same.
Some of the countries we wanted to target were unavailable in Google AdWords (e.g., Iran, Syria) and after discussion we settled on the following five countries: China, Egypt, Pakistan, Russia, and Ukraine.
The second study ran from October 8, 2014 4:00 PM MDT to October 15, 2014 4:00 PM MDT.
The budget for the global campaign was \$500/day and the country-specific campaigns were \$50/day.
We used the following keywords for the study:
``Nelson Mandela'',
``Sports'',
``Internet Security'',
``Basketball'',
``Football'',
``Freedom'',
``NCAA'',
``Paul Walker'',
``Boston Marathon'',
``Election'',
``North Korea'',
``Harlem Shake'',
``PlayStation 4'',
``Royal Baby'',
``Cory Monteith'',
``iPhone 6'',
``iPhone 5s'',
``Samsung Galaxy S4'',
``iPhone 6 Plus'',
``TLS Proxies''.
The breakdown of costs and results are given in Table~\ref{tab:campaign2}.
In total we completed 12,314,756 successful measurements.
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{lccc}
\toprule
Campaign & Impressions & Click & Cost \\
\midrule
Global & 3,285,598 & 5,424 & \$4,021.78 \\
China & 689,233 & 652 & \$401.41 \\
Egypt & 232,218 & 1,777 & \$378.17 \\
Pakistan & 183,849 & 2,536 & \$378.26 \\
Russia & 230,474 & 203 & \$401.36 \\
Ukraine & 364,868 & 294 & \$390.69 \\
\midrule
Total & 5,079,298 & 11,077 & \$6,090.19 \\
\bottomrule
\end{tabular}}
\caption{Second Study Statistics}
\label{tab:campaign2}
\end{table}
Section~\ref{sec:measurementstudy2} discusses the results from this second study.
\section{Second Measurement Study}
\label{sec:measurementstudy2}
During the duration of this second ad campaign we successfully completed 12.3 million measurements targeting five specific countries as well as the world in general.
Of those tests, 50,761 returned a different X.509 certificate than was served by the authoritative host.
\subsection{Analysis of Issuer Organization}
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{lcc}
\toprule
Proxy Type & Connections & Percent \\
\midrule
Business/Personal Firewall & 36,005 & 70.93\% \\
Business Firewall & 1,231 & 2.43\% \\
Personal Firewall & 536 & 1.06\% \\
Parental Control & 428 & 0.84\% \\
Organization & 3,531 & 6.96\%\\
School & 482 & 0.95\% \\
Malware & 2,571 & 5.06\% \\
Unknown & 5,455 & 10.75\% \\
Telecom & 447 & 0.88\% \\
Certificate Authority & 68 & 0.13\% \\
\bottomrule
\end{tabular}}
\caption{Classification of claimed issuer in 2nd study}
\label{tab:proxytypesv2}
\end{table}
Table~\ref{tab:proxytypesv2} contains the breakdown of Issuer Organization fields from our second measurement study.
As in our first measurement study, we find that the majority of TLS proxies claim to be features of firewall solutions (74.42\%).
Organization and school names are also prevalent, accounting for another 6.01\% of Issuer Organization values.
However, we see an increase in the relative popularity of the ``Unknown'' as compared to the first study (10.75\% from 7.14\%).
The Unknown category comprises certificates with null or blank issuer fields, or otherwise uncategorizable values.
In tandem with this finding we note that the Malware category has decreased in relative popularity from 8.65\% to 5.06\%.
These results were obtained largely from our five targeted countries.
The increase in the Unknown category of TLS proxies proxying these countries is alarming.
It is possible that malware using TLS proxy features in these regions is more opaque than its earlier counterparts, opting not to disclose its identity through Issuer Organization fields.
Even if this is not the case, it is alarming to note that TLS proxies may be decreasing their already-poor visibility to users in those countries.
Another distinguishing feature of our second study's Issuer Organization fields is the presence of the names of Telecom companies in the dataset.
We found 375 proxied connections from IP addresses owned by a Korean telecom company, LG UPLUS.
Another four telecom company names were reported from an additional 72 connections.
It is unclear whether LG UPLUS and the other companies are using TLS proxies within their own office buildings or using them to intercept the traffic of their own users.
The latter behavior is not without precedent; Nokia has recently come under fire for such an operation~\cite{nokia}.
\subsection{Proxy prevalence by specific country}
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{clccc}
\toprule
Rank & Country & Proxied & Total & Percent \\
\midrule
1 & China & 563 & 2,549,301 & 0.02\% \\
2 & Ukraine & 4,329 & 1,575,053 & 0.27\% \\
3 & Russia & 4,532 & 1,116,341 & 0.40\% \\
4 & Korea & 1,722 & 836,556 & 0.21\% \\
5 & Egypt & 3,720 & 660,937 & 0.56\% \\
6 & Pakistan & 1,890 & 456,792 & 0.41\% \\
7 & Turkey & 1,975 & 411,962 & 0.48\% \\
8 & US & 3,327 & 385,811 & 0.86\% \\
9 & Japan & 2,033 & 273,532 & 0.74\% \\
10 & UK & 2,056 & 266,873 & 0.77\% \\
11 & Brazil & 1,889 & 232,454 & 0.81\% \\
12 & Taiwan & 530 & 186,942 & 0.28\% \\
13 & Romania & 2,210 & 185,749 & 1.19\% \\
14 & Indonesia & 798 & 181,971 & 0.44\% \\
15 & Germany & 1,091 & 177,586 & 0.61\% \\
16 & Italy & 737 & 145,438 & 0.50\% \\
17 & Greece & 516 & 130,613 & 0.40\% \\
18 & Poland & 456 & 127,806 & 0.36\% \\
19 & Czech Rep. & 343 & 110,170 & 0.31\% \\
20 & India & 716 & 102,869 & 0.70\% \\
& Other (209) & 15,328 & 2,200,000 & 0.70\% \\
\bottomrule
& Total & 50,761 & 12,314,756 & 0.41\% \\
\bottomrule
\end{tabular}}
\caption{Connections tested by country (2nd study)}
\label{tab:connectionsbycountryv2}
\end{table}
Our second measurement study via AdWords contained six mini-campaigns.
Five of these targeted the countries of China, Ukraine, Russia, Egypt, and Pakistan.
These countries were selected for their contemporary civil struggles and their respective governments' stance on free speech.
The final mini-campaign targeted the world in general.
Table~\ref{tab:connectionsbycountryv2} shows a breakdown of the number of connections tested per country.
We immediately see that all five specific countries targeted by our campaign lie within the top six most-prevalent countries, showing the dependability of Google Adwords' country targeting feature.
We also note the relatively low percentage of TLS-proxied traffic from China.
Before this study the authors hypothesized that this country would have a high amount of TLS-proxied traffic due to its stance on civil liberties and government surveillance.
We also find that Ukraine, Russia, Egypt and Pakistan all have a lower TLS-proxy percentage than western nations such as the United States (0.86\%) and the UK (0.77\%).
This may be due to the fact that most detected firewall solutions are of western origin and western install base.
Thus it is possible that the general lower proxy rates in our target countries is due more to consumer choice and buying power.
The relative prevalence of TLS proxies by country is visualized in Figure~\ref{fig:worldheatmap}.
Low TLS-proxy rates are signified by blue and gradually transition to red with increasing proxy rate.
The map includes connection data from our second study in 228 countries and territories.
\begin{figure}
\centering
\includegraphics[width=240px]{figures/heatmap_alternate.png}
\caption{Heat-map of TLS proxy prevalence by country. Highest = 12\% proxy rate, lowest = 0\% proxy rate}
\label{fig:worldheatmap}
\end{figure}
\subsection{Proxy behavior by type of host}
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{lccc}
\toprule
Website Type & Connections & Proxied & Percent Proxied \\
\midrule
Popular & 5,132,342 & 20,965 & 0.41\% \\
Business & 1,787,875 & 7,494 & 0.42\% \\
Pornographic & 3,004,996 & 12,458 & 0.41\% \\
Authors' & 2,353,717 & 9,844 & 0.42\% \\
\bottomrule
\end{tabular}}
\caption{Proxied connection breakdown by host type}
\label{tab:proxied_by_host_type}
\end{table}
The augmented measurement tool used in our second study connected to various types of hosts: popular, business, pornographic, and our own.
The breakdown of the prevalence of TLS proxies with respect to each host type is shown in Table~\ref{tab:proxied_by_host_type} (note that the number of connections to each type of host is varies due to the variance in bandwidth and computer performance of our users).
The percentage of proxied traffic to each type of host is nearly identical.
We also find that individual TLS proxies are also indiscriminate in their behavior toward these types of hosts.
These results suggest that TLS proxies do not employ blacklists when deciding which traffic to intercept.
Given Huang et al.'s finding of a 0.20\% TLS-proxy rate for Facebook connections, there is some evidence to suggest that at least some TLS proxies are employing whitelists when determining whether to intercept a connection.
Facebook's popularity far exceeds the popularity of our chosen popular hosts (we were constrained to hosts which served permissive Flash socket policy files), so sites akin to it are in a class of their own.
It is possible that many benevolent TLS proxies are configured to ignore extremely popular websites run by reputable organizations, perhaps to preserve some privacy and reduce performance overhead.
\subsection{More malware}
Our second study revealed a continued presence of malware in the TLS proxy space.
All of our previously discovered malware was also present in our second study, with an additional five discoveries: issuer fields containing the values Objectify Media Inc (1069 connections), Superfish, Inc. (610 connections), WiredTools LTD (131 connections), Internet Widgits Pty Ltd (67 connections), and ImpressX OU (16 connections).
Web research indicated that all these are malware products and one of them, Internet Widgits Pty Ltd, has ties to a botnet.
Combined with malware previously identified, malware in the second study accounted for 2,571 of proxied connections.
One suspicious Issuer Organization field was ``kowsar''. Certificates with this identifier appeared 268 times, and
were retrieved by 266 unique IP addresses. Unlike other Issuer
Organizations we found, this identifier did not appear to be
associated with a large organization (which would indicate a corporate
firewall) or a personal firewall product. The IP addresses given this
certificate are from numerous countries around the world and from many
different ISPs. Contrast this with the Certificate Issuer ``DSP'',
which appeared 204 times, but with only one IP. In this case, ``DSP''
is being used by the Department of Social and Family Affairs (also
called the Department of Social Protection), Ireland, and thus likely
represents a corporate firewall. The pattern for ``kowsar'' is
indicative of either a popular personal firewall or an active attack
such as a botnet.
Similar oddities appear, but on a smaller scale. For example, the
Certificate Issuer field ``Information Technology'' appeared 33 times,
covering 3 IP addresses. These IPs were from a Japanese chemical
company, an ISP in Netherlands, and a chapter of the American Red
Cross. These are such disparate organizations that this looks like
suspicious behavior, though it is possible that each of these
organizations set up a corporate firewall and chose the same name for
the certificate they generated. The Issuer field ``MYInternetS''
appeared 36 times from 6 different ISPs. Five of these are in Denmark,
from a variety of ISPs and a university, yet one is from a cable
subscriber in the United States. It is difficult to determine whether
cases like this are examples of malware.
Even more worrisome are the 5,455 instances where we could not identify the issuer, and
the 1,518 where the issuer field is null or blank. Whoever set up the TLS proxy in these cases did not want to be
identified.
\section{Initial Measurement Study}
\label{sec:measurementstudy}
Our first measurement study was targeted at a general global audience.
During the duration of this ad campaign, we served 4.36 million ads and successfully completed 2.86 million measurements.
Of those tests, 11,764 returned a different X.509 certificate than was served by our secure web server, indicating the presence of a TLS proxy.
The users behind a proxied connection that were identified by our campaign originated in 142 countries and from 8,589 distinct IP addresses. Due to the targeting algorithms used by Google AdWords, our tool's exposure to these countries is not uniformly distributed.
Table~\ref{tab:connectionsbycountry} shows the countries with
the most proxied connections in our study.
For each country, the table lists the total number of proxied connections, the total number of connections, and the percentage of total connections to that country that were proxied.
Some countries have significantly higher percentages of proxied connections than the average, including France (1.09\%), Canada (0.87\%), Belgium (0.81\%), the United States (0.79\%), and Romania (0.74\%).
Together, connections from the United States and Brazil account for 36\% of detected proxies.
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{clccc}
\toprule
Rank & Country & Proxied & Total & Percent \\
\midrule
1 & US & 2,252 & 285,078 & 0.79\% \\
2 & Brazil & 2,041 & 298,618 & 0.68\% \\
3 & France & 812 & 74,789 & 1.09\% \\
4 & UK & 759 & 259,971 & 0.29\% \\
5 & Romania & 696 & 94,116 & 0.74\% \\
6 & Germany & 499 & 187,805 & 0.27\% \\
7 & Canada & 303 & 34,695 & 0.87\% \\
8 & Turkey & 303 & 65,195 & 0.46\% \\
9 & India & 302 & 51,348 & 0.59\% \\
10 & Spain & 226 & 62,569 & 0.36\% \\
11 & Russia & 224 & 58,402 & 0.38\% \\
12 & Italy & 200 & 129,358 & 0.15\% \\
13 & S.Korea & 196 & 46,660 & 0.42\% \\
14 & Portugal & 185 & 29,799 & 0.62\% \\
15 & Poland & 182 & 110,550 & 0.16\% \\
16 & Ukraine & 160 & 61,431 & 0.26\% \\
17 & Belgium & 136 & 16,816 & 0.81\% \\
18 & Japan & 111 & 31,751 & 0.35\% \\
19 & Netherlands & 104 & 31,938 & 0.33\% \\
20 & Taiwan & 101 & 61,195 & 0.17\% \\
& Other (215) & 1,972 & 869,096 & 0.23\% \\
\bottomrule
& Total & 11,764 & 2,861,180 & 0.41\% \\
\bottomrule
\end{tabular}}
\caption{Proxied connections by country (1st study)}
\label{tab:connectionsbycountry}
\end{table}
\subsection{Analysis of Issuer Organization}
We first analyze the contents of the Issuer Organization in the
substitute certificates we collected. We use OpenSSL to
decode the certificates and store them in a database, where we can run
queries. We also manually inspect the contents of the relevant fields
to identify the issuing organization and their software products,
using web searches to determine their identity. We emphasize that our
results in this section are based on the intercepting proxy
self-identifying themselves in the certificate. It is certainly
possible that malicious proxies have hidden their tracks by
masquerading as a legitimate organization in the Issuer Organization field, and we
cannot detect this.
Table~\ref{tab:topissuernames} shows the values for the Issuer Organization field of the substitute certificates.
Table~\ref{tab:proxytypes} provides a breakdown of values present in the Issuer Organization field of the substitute certificates.
The majority of certificates from proxied connections have an Issuer Organization field matching the name of a personal or enterprise firewall (69.54\%).
Another 12.66\% have the name of an organization set as the Issuer Organization (e.g., Lawrence Livermore National Laboratory, Lincoln Financial Group).
Additionally, 7\% (829) of the substitute certificates have null values for the Issuer Organization field.
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{llc}
\toprule
Rank & Issuer Organization & Connections \\
\midrule
1 & Bitdefender & 4,788 \\
2 & PSafe Tecnologia S.A. & 1,200 \\
3 & Sendori Inc & 966 \\
4 & ESET spol. s r. o. & 927 \\
5 & Null & 829 \\
6 & Kaspersky Lab ZAO & 589 \\
7 & Fortinet & 310 \\
8 & Kurupira.NET & 267 \\
9 & POSCO & 167 \\
10 & Qustodio & 109 \\
11 & WebMakerPlus Ltd & 95 \\
12 & Southern Company Services & 62 \\
13 & NordNet & 61 \\
14 & Target Corporation & 52 \\
15 & DigiCert Inc & 49 \\
16 & ContentWatch, Inc. & 42 \\
17 & NetSpark, Inc. & 42 \\
18 & Sweesh LTD & 39 \\
19 & IBRD & 26 \\
20 & Cloud Services & 23 \\
& Other (332) & 1,121 \\
\bottomrule
\end{tabular}}
\caption{Issuer Organization field values}
\label{tab:topissuernames}
\end{table}
\begin{table}[t]
\centering
{\setlength{\extrarowheight}{2pt}
\begin{tabular}{lcc}
\toprule
Proxy Type & Connections & Percent \\
\midrule
Business/Personal Firewall & 8,101 & 68.86\% \\
Business Firewall & 69 & 0.59\% \\
Personal Firewall & 11 & 0.09\% \\
Parental Control & 156 & 1.33\% \\
Organization & 1,394 & 12.66\%\\
School & 32 & 0.27\% \\
Malware & 1,112 & 8.65\% \\
Unknown & 840 & 7.14\% \\
Telecom & 0 & 0\% \\
Certificate Authority & 49 & 0.42\% \\
\bottomrule
\end{tabular}}
\caption{Classification of claimed issuer in 1st study}
\label{tab:proxytypes}
\end{table}
The most suspicious activities discovered were revealed by certificates with an Issuer Organization that matched the names of malware. ``Sendori, Inc'', ``WebMakerPlus Ltd'', and ``IopFailZeroAccessCreate'' appeared in 966, 95, and 21 Issuer Organization fields, respectively.
Sendori poses as a legitimate enterprise, however they produce software that compromises the DNS lookup of infected machines, allowing them to redirect users to improper hosts.
A TLS proxy component is used to bypass host authenticity warnings in the browser.
The substitute certificates generated by the TLS proxy are signed by a root authority that was added to the root store of the local machine at the time of infection. Substitute certificates issued by Sendori originated from 30 distinct countries.
The WebMakerPlus malware is primarily associated with inserting advertisements into Web pages.
Since modern browsers issue warnings when insecure content is queried from secure connections, we hypothesize that WebMakerPlus uses a TLS proxy to simulate that their advertisements are served from a secure connection. Substitute certificates containing markings for WebMakerPlus originated from 16 distinct countries.
Manual Internet queries revealed that malware was responsible for an Issuer Common Name field value of ``IopFailZeroAccessCreate''. The certificates containing this value originated from 14 distinct countries. Disturbingly, each certificate contained the same 512-bit public key. This malware was also reported by \cite{huang2014analyzing}.
It is somewhat surprising that these malware programs self-identify in the substitute certificates they generate, as an attacker can arbitrarily select values for the fields in a substitute certificate.
In addition to malware discoveries, we found that the names of two companies highly associated with spam were also present in numerous Issuer Organization fields.
The names ``Sweesh LTD'', and ``AtomPark Software Inc'' were found in 39 and 20 substitute certificates, respectively.
AtomPark offers tools for spammers including ``email extractors'' and ``bulk mailers''.
Sweesh offers services to spammers to overcome ``hurdles'' faced by advertisers and publishers.
Internet searches reveal that Sweesh may be responsible for the development of WebMakerPlus.
\subsection{Negligent Behavior}
Where possible, we installed and characterized personal firewall software from many of the most common companies whose names were provided in the Issuer Organization, Issuer Organizational Unit, and Issuer Common Name fields of our collected certificates.
We characterized the behavior of these solutions when running behind our own TLS proxy which issued certificates signed by an untrusted CA. While most solutions properly rejected our forged certificates, Kurupira, a parental filter that is responsible for 267 proxied connections in our dataset, did not. When visiting \texttt{google.com} and \texttt{gmail.com}, Kurupira replaced our untrusted certificate with a signed trusted one, thus allowing attackers to perform a transparent man-in-the-middle attack against Kurupira users without having to compromise root stores. In contrast, BitDefender not only blocked this forged certificate,
but also blocked a forged certificate that resolved to a new root we installed in the victim machine's root store.
Among the negligent behavior we found are TLS proxies that generate substitute certificates with weak cryptographic strength. Our original certificate has a public key size of 2048 bits. However, we find that 5,951 (50.59\%) substitute certificates have public key sizes of 1024 bits and 21 certificates have public key sizes of 512 bits.
23 (0.20\%) TLS proxies generated substitute certificates that used MD5 for signing, 21 (0.18\%) which were also 512 bit keys.
Interestingly, some TLS proxies generated certificates that have better cryptographic strength than our certificate.
Seven (0.06\%) used certificates with a key size of 2432 and five (0.04\%) used SHA-256 for signing.
In addition to problems with cryptographic strength, we discovered that 49 (0.42\%) substitute certificates claim to be signed by DigiCert, though none of them actually are. The original certificate from our secure web server is issued by DigiCert High Assurance CA-3, indicating the TLS proxy likely copied this field when creating the substitute.
It is alarming that a TLS proxy would opt to copy this field, as it signifies a masquerading as the legitimate authority.
It is possible that these proxies are operated by malicious individuals doing their best to not be detected by the user.
Finally, we note that 110 substitute certificates have modifications to the subject field. For 51 (0.43\%) certificates, the subject did not match our website's domain.
In many cases a wildcarded IP address was used that only designated the subnet of our website.
In two cases the substitute certificate is issued to the wrong domain entirely: \texttt{mail.google.com} and \texttt{urs.microsoft.com}. These
certificates appear to be legitimate for those domains and properly
validate back to GeoTrust and Cybertrust roots, respectively.
\section{Measurement Tool}
We have developed a tool to measure the prevalence of TLS
proxies using existing, widely-deployed technologies. The tool runs silently from the
perspective of the user; no user action is required, either to
install any software or to interact with the tool. This is a
significant advantage as compared to other work that requires
client-side software installation \cite{wendlandt2008perspectives,
marlinspike2011ssl, alicherry2009doublecheck,holz2012x,
amann2012extracting, amann2012revisiting}.
\subsection{Design}
To meet our objective of using existing browser technologies without requiring further client installation, we take advantage of the widespread deployment and transparency afforded by the Adobe Flash runtime.
By hosting a Flash application on a web page the server can upload it to a visiting client, which runs it without any user interaction.
Our tool works by sending a \texttt{ClientHello} message to a TLS-enabled server and recording the \texttt{ServerHello} and \texttt{Certificate} messages received in response. The retrieved certificate(s) is then forwarded to the web server for verification.
This process is handled in three steps, illustrated in Figure~\ref{fig:flashtool}:
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/flashtool_alternate_2.pdf}
\caption{TLS Proxy Measurement}
\label{fig:flashtool}
\end{figure}
\begin{enumerate}
\item {\bf Retrieve measurement tool.} The client browser connects to the web server where the Flash application is hosted.
The application need not be visible to the user and can merely be embedded in the background of an otherwise normal web page.
The web page data, along with the embedded Flash application, is then downloaded by the client.
\item {\bf Record certificate.} The Flash tool is run automatically by the browser. The tool sends a \texttt{ClientHello} message back to the server to initiate a TLS handshake.
The tool then records the \texttt{ServerHello} and \texttt{Certificate} messages received in response and terminates the handshake.
\item {\bf Report results.} The tool reports these results back to the server using an HTTP POST request. The server then compares the certificate received with the original it sent. A mismatch indicates the presence of a TLS proxy.
\end{enumerate}
To deploy our tool a website administrator need only do the following:
\begin{enumerate}
\item {\bf Host the Flash application on the web server } and invoke it from the desired pages. The tool can be deployed transparently within existing web pages with no visible changes.
\item {\bf Host a simple ``socket policy file'' on the server.} For security reasons the Flash runtime requires that applications attempting to establish a TCP connection with a remote host first obtain permission from that host via a simple policy file. The request for this file is sent automatically by the Flash runtime. The software to host such a file is extremely simple and easily deployed. This particular security feature of Flash prohibits the tool from testing client connections to arbitrary hosts; all hosts tested must first grant permission through their respective socket policy files. We serve our socket policy file on the same port used by our web server (80). This reduces the effect of captive portals, which often block traffic targeting ports other than those used by HTTP and HTTPS (e.g., airport public access WiFi). Our socket policy server implementation is provided on our website.
\end{enumerate}
\subsection{Implementation}
To implement our tool it was necessary to retrieve the certificate used during a TLS handshake.
It would have been preferable to use JavaScript or HTML5 to retrieve the certificate used as part of a current TLS connection, but unfortunately there is no API available for this. Firefox allows a plugin to request the certificate, but plugins require manual client installation.
This left us with the alternative of establishing a plain TCP connection with the target server and then initiating a TLS handshake. Unfortunately, the ability to use a plain TCP connection rules out the use of HTML5 \texttt{WebSockets}.
Due to these constraints, we opted to use the Adobe Flash platform.
Beginning with version 11.0 of the Flash runtime, Adobe made available a \texttt{SecureSockets} API that allows developers to access certificate data from a TLS connection. However, these versions of Flash were too recent to enjoy the reported 98.9\% desktop market penetration of Flash 9.0 \cite{adobe}.
Thus we implemented our tool in ActionScript using only libraries supported by the Flash 9.0 runtime.
Using the \texttt{Socket} API provided by Flash 9.0 we implemented functionality required to perform a partial TLS handshake.
After receiving the full \texttt{Certificate} message from the desired host the handshake is aborted and the connection is closed.
The Flash application records and parses all certificates received from the \texttt{Certificate} message (as some hosts offer certificate chains) and stores them locally until it parses the final one.
All certificate data, in PEM format, is concatenated and then sent as an HTTP POST request back to the host for analysis.
Code for the measurement tool and collected datasets will be available for download at \url{https://tlsresearch.byu.edu}.
\subsection{Limitations}
\label{sec:limitations}
Our tool is unable to measure TLS proxies being used against most
mobile devices. An overwhelming majority of mobile platforms do not
support Flash and Adobe has discontinued their development of Flash
for mobile devices.
It is possible that TLS proxies could be engineered to circumvent our measurements.
At the time of our study, our measurement methodology was not well known,
so it is unlikely that any attacker was evading detection or
tampering with our reports.
However, in the case that this methodology becomes well-known, it would be difficult to prevent dedicated attackers from modifying their TLS proxies to avoid our measurements.
\section{Acknowledgments}
This work is supported by a 2014 Google Faculty Research Award.
\bibliographystyle{abbrv}
\section{Mitigation}
A large body of work seeks to detect and prevent TLS proxies,
generally regarding them as MitM attacks. Clark and van
Oorschot~\cite{clark2013sok} provide an extensive survey of this area
and provide one of the few research papers that acknowledges the
existence of benevolent TLS proxies. Below we survey the various
mitigation approaches in the field.
{\em Multi-path probing} allows clients to determine whether the
certificate they have been given for a server is different from those
seen by most other clients. Representative systems include
Perspectives~\cite{wendlandt2008perspectives},
Convergence~\cite{marlinspike2011ssl}, and
DoubleCheck~\cite{alicherry2009doublecheck}.
Crossbear~\cite{holz2012x} goes a step further to use traceroute to
localize the origin of the attack. Other systems use existing
Certificate Authorities or centralized notaries to vouch for the
authenticity of a certificate \cite{mecai,amann2012extracting,
amann2012revisiting}. Each of these systems may suffer from false
alarms due to benign changes to certificates and the presence of
multiple valid certificates for a given site \cite{amann2013acsac}.
There are several proposals to leverage the {\em shared password}
between the client and server to prevent a MitM attack. Direct
Validation of Certificates (DVCert)~\cite{dacosta2012trust} permits
the server to attest to the authenticity of all the certificates used
in a session with the web application, including certificates from
other domains. SSL/TLS Session-Aware User
Authentication~\cite{oppliger2006ssl,oppliger2008ssl} thwarts TLS MitM
attacks through user authentication tokens based on client credentials
and TLS session information. The proposed TLS-SRP
protocol~\cite{rfc5054} would extend the TLS handshake to support
mutual authentication based on a shared password.
Several proposals leverage {\em DNS} to prevent MitM attacks.
ConfiDNS~\cite{poole2006confidns} utilizes the temporal and spatial
redundancy of the existing DNS system to assess agreement for IP
resolution. The DNS-Based Authentication of Named Entities
(DANE)~\cite{hoffman2011using} protocol enables administrators to bind
hostnames to their certificates. This permits public keys to be
transmitted via DNSSEC without involving a CA.
{\em Certificate pinning}~\cite{evans2011certificate} is a Google
proposal for the web server to limit all future HTTPS connections to a
limited set of server certificates. Pinning is a trust-on-first-use
technology. The Google Chrome browser comes pre-configured with some
Google certificates already pinned in advance so the user does not
have the TOFU issue with those sites. Chrome also trusts any locally
installed trusted roots, so benevolent proxies and malware can
circumvent the pinning process. Trust Assertions for Certificate Keys
(TACK)~\cite{tack} is a TLS extension for the server to pin a signing
key that must sign all other keys in the domain.
Another approach is to use an {\em audit log} of valid certificates
issued by Certificate Authorities. Representative work in this area is
Certificate Transparency (CT)~\cite{rfc6962,ryan2014enhanced} and the
EFF sovereign keys (SK) project~\cite{sovereignkeys}. The Accountable
Key Infrastructure (AKI)~\cite{kim2013accountable} is a proposal for
new infrastructure to validate public keys and reduce the reliance on
CAs. This system also includes public log servers that support public
validation of certificate integrity.
A different approach acknowledges that there is an industry need for
TLS inspection to detect malware or protect intellectual
property. Several proposals to the IETF from industry introduce
mechanisms that would make proxies visible to the other participants
in a chain of TLS connections and could include user notification and
consent \cite{mcgrew2012internetdraft, loreto2014internetdraft}.
Another proposal \cite{nir2012internetdraft} introduces an extension
to TLS that enables a client to share decryption keys with a TLS
proxy. A potential drawback is that the key-sharing mechanism
represents a new attack surface that hackers could attempt to exploit
to acquire the decryption keys for a TLS session.
\section{Related Work}
The most closely related work in this field is a recent paper by Huang
et al., which independently develops a measurement tool that is very
similar to ours and conducts a measurement study of TLS proxies that
intercept the Facebook website \cite{huang2014analyzing}. We discussed
the primary differences in methodology in
Section~\ref{sec:introduction}. Generally speaking, the advantage of
Huang's methodology is that they find proxies specifically targeting
Facebook, whereas the advantage of our methodology is that we can
target our measurements for selected countries and for selected
websites that have permissive Flash socket policy files. This enables
us to actively collect a broader measurement of proxies.
In comparing our results to Huang, the prevalence of proxies in our
study is roughly twice what was measured by Huang (0.41\% versus
0.20\%). In addition, we find a wider array of malware, deceptive
practices, and suspicious circumstances. Both of these results are
likely due to the more comprehensive measurements we make, avoiding a
site such as Facebook that is likely on the whitelist for many
proxies. Our measurements of WebMakerPlus, Objectify Media, Superfish,
WiredTools, Internet Widgits Pty, ImpressX, and kowsar all represent
malware found only in our study. Likewise, the presence of spam
infections from Sweesh and AtomPark are unique to our study, as is the
evidence of botnets using TLS proxies. We are the first to identify a
parental filter replacing an untrusted certificate with a trusted
one. Our country-specific measurement campaigns add additional data to
the field.
The only other paper to find evidence of TLS proxies is the work from
The Netalyzer project, which analyzes the root store of Android
devices \cite{vallina2014tangled}. Their primary findings include the
use of manufacturer and vendor-specific certificates, the presence of
unusual root certs, and third party apps that manipulate the root
store. In addition, they find one case of a TLS proxy, out of 15,000
assessed TLS sessions. The app whitelists several sites, including
Facebook, Twitter, and several Google sites, but intercepts mail from
Yahoo, Google, and traffic to several major banks. It is difficult
to compare the prevalence (1 in 15K) to rates found by Huang and this
paper because the sample is from users choosing to download the
Netalyzer App.
Another closely related paper is the Crossbear
system~\cite{holz2012x}, which is designed for volunteer hunters to
work together to detect and localize real-world TLS MitM attacks.
After the client establishes a TLS connection to a website, the client
sends the received certificate chain to a central Crossbear server.
The Crossbear server establishes its own secure connection with the
website and also queries Convergence for additional data about the
website's certificate. This information is recorded in a database on
the server and is also sent to the client. If the cumulative data
received by the client suggest a MitM is present, the client performs
a traceroute operation to the malicious server and sends that
information to the Crossbear server. The Crossbear server attempts to
localize the origin of the MitM attacker by using traceroute data from
many Crossbear clients. Crossbear was deployed in 150 locations on the
PlanetLab testbed and had not detected any attacks (or benevolent TLS
proxies) at the time of the report.
Finally, a number of surveys collect and analyze SSL certificates and
certificate authorities on the Internet \cite
{eckersley2011decentralized,holz2011ssl,akhawe2013here,durumeric2013analysis,amann2013acsac}. These
studies do not examine the use of substitute certificates by TLS
proxies, but focus on issues such as TLS errors, properties of
certificates and the PKI system, and poor security practices.
|
\section{Introduction}
The fate of a massive star is determined by its initial mass, composition, and mass loss over
its life (which is still poorly understood). The explosion mechanism and remnant properties
is mainly determined by the helium core mass when the star dies. \citet{kudri2002} suggests
that the mass loss rate of a star, $\dot{m}$, is $\propto\,Z^{0.5}$, where $Z$ is the metallicity
of the star relative to the solar metallicity, $\ensuremath{Z_\odot}$. Since the first (or Pop~III) stars form in
pristine H and He gas, it is generally thought that they retain most of their mass over their
lifetimes. Pop~III stars above $80\,{\ensuremath{\it{M}_{\odot}}}$ encounter the pair production instability after
central carbon burning, in which large numbers of thermal photons create $e^-/e^+$ pairs at
the expense of pressure support in the core. If Pop~III stars are over $150\,{\ensuremath{\it{M}_{\odot}}}$ but less
than $260\,{\ensuremath{\it{M}_{\odot}}}$, core contraction due to the loss of pressure support to pair production
ignites explosive oxygen and silicon burning that results in an energetic explosion that
completely unbinds the star. This thermonuclear explosion is called a pair-instability
supernova (PSN; \citealt{barkat1967, heger2002, chen2014r, chen2014psn, chen2014ppsn}).
A PSN can explode with up to $10^{53}$ erg of energy, or about 100 times that of a Type Ia SN.
Explosive silicon burning can synthesize up to 40 {\ensuremath{\it{M}_{\odot}}}\ of {\ensuremath{^{56}\mathrm{Ni}}}{} in these events which, along
with their large explosion energies, makes them extremely bright. PSNe are therefore visible
at high redshifts and could be used to probe the properties of the first stars \citep{kasen2011,
wet12a,wet12b,pan12a}. Isotopes heavier than the iron group are completely absent from the
chemical yields of PSNe because of the absence of neutron capture processes (r- and
s-processes). The important question arises : Can the most massive stars in the local Universe
also die as PSNe? If so, how do these PSNe differ from those of Pop~III progenitors?
The detection of PSN candidates SN 2007bi and SN 2213 - 1745 \citep{galyam2009,cooke12}
has increased interest in PSNe and poses a challenge to theories of galactic star formation,
because how progenitors so massive can form at metallicities of $0.1\,\ensuremath{Z_\odot}$ today is very
unclear. Stars at these metallicities can easily lose most of their mass over their lifetimes to
strong stellar winds, and so they must form at even higher masses than Pop III stars to die
as PSNe. Understanding the observational signatures of PSNe at near-solar metallicities is
key to identifying and properly interpreting these events as more are discovered. To date,
PSN models have focused on Pop III stars. In this paper, we present 2D simulations of a PSN
of a non-zero metallicity star and investigate if mixing can affect explosive burning or
observational signatures. We first describe our numerical methods and problem setup in
\S~\ref{solpsn_Methodology}. Our results in are discussed in \S~\ref{solpsn_results}, and our
conclusions are given in \S~\ref{solpsn_conclusions}.
\section{Methodology \& Problem Setup}
\label{solpsn_Methodology}
Self-consistent multidimensional stellar evolution models from the onset of hydrogen burning to
eventual core collapse and explosion remain beyond the realm of contemporary computational
power. We instead initialize our multidimensional explosion simulation with a $500\,{\ensuremath{\it{M}_{\odot}}}$ 0.1
\ensuremath{Z_\odot}\ star that is evolved from the zero-age main sequence to the onset of core collapse in the
1D stellar evolution code {\texttt{GENEVA}} \citep{hirschi2004, whalen2013}. Our model includes
mass loss and stellar rotation to determine the final structure of the star. The initial rotational
velocity is $40\,\%$ of the critical velocity, or a surface velocity of 450 km/sec at the equator.
The star is evolved up to the onset of explosive oxygen burning, about 10 s before maximum
core contraction. Initializing this profile in a 2D simulation at this time should therefore capture
instabilities seeded by both collapse and explosive burning. The explosion time scale is about
tens of seconds which is much shorter than the rotational period of star, $P\sim$ several hours.
It is reasonable to assume that the star is non-rotating during the explosion phase, so we neglect
the effect of rotation in the 2D simulation.
We map the 1D profile of the star to a 2D cylindrical coordinate grid in $r$ and $z$ in \texttt{CASTRO}{}
with the conservative mapping scheme of \citet{chen2013}. \texttt{CASTRO}{} is a multidimensional
adaptive mesh refinement (AMR) astrophysical radiation hydrodynamics code \citep{ann2010,
zhang2011}. It has a higher-order unsplit Godunov hydro scheme and block-structured AMR.
In our models we use the Helmholtz EOS \citep{timmes2000} for stellar matter, which includes
degenerate and non-degenerate, relativistic and non-relativistic electrons, electron-positron pair
production, and an ideal gas with radiation. \texttt{CASTRO}{} evolves mass fractions for each isotope
with its own advection equation. The monopole approximation is used for self-gravity, in which a
1D gravitational potential is constructed from the radial average of the density and then
differenced to construct the gravitational force vector everywhere in the AMR hierarchy. This
approximation is well-suited to the nearly spherical symmetry of the star and is efficient. We use
the 19-isotope APPROX nuclear reaction network \citep{kepler,timmes1999}, which includes
heavy-ion reactions, alpha-chain reactions, hydrogen burning cycles, photo-disintegration of
heavy nuclei, and energy loss through thermal neutrinos. Nuclear burning is self-consistently
coupled to hydrodynamics, and we also account for energy deposition due to radioactive decay
of {\ensuremath{^{56}\mathrm{Ni}}}{} $\to$ {\ensuremath{^{56}\mathrm{Co}}}{} $\to$ {\ensuremath{^{56}\mathrm{Fe}}}{}.
The grid size is $4\cdot10^{10}\times4\cdot10^{10}\,\ensuremath{\mathrm{cm}}^2$, with 256 uniform zones in both $r$
and $z$. Up to three levels of refinement, each with a factor of 4 greater resolution, are permitted
throughout the simulation. The resolution at the finest level is $2\times10^7\,\ensuremath{\mathrm{cm}}$, which is
sufficient to resolve the characteristic scales of nuclear burning ($\sim 10^8\,\ensuremath{\mathrm{cm}}$). The grid is
refined on functions of the gradients in density, velocity, and pressure. Because we only simulate
one octant of the star, we apply reflecting boundary conditions at the inner $r$ and $z$ boundaries;
we impose outflow boundary conditions on the upper boundaries. Refined grids are nested around
the core of the star to assure that it is always at the highest resolution. We evolve our simulations
until the shock breaks through the surface of the star.
\begin{figure}[h]
\vspace{-.2cm}
\begin{center}
\subfigure[Mass fraction of selected isotopes]{\label{x500}\includegraphics[width=0.46\textwidth]{kchen_fig1.eps}}
\subfigure[$\rho r^3$ plot]{\label{p500_rhor3}\includegraphics[width=0.46\textwidth]{kchen_fig2.eps}}
\caption{Left: isotope mass fractions. The high silicon mass fraction at the core shows that oxygen
burning is complete. Right: plot of $\rho r^3$. The bump at the center is due to the helium core.
The red dashed line marks the approximate site of the formation of the shock. The shock
propagates only a short distance through a region of increasing $\rho r^3$, so it is unlikely to
decelerate and create a reverse shock or therefore fluid instabilities.}
\end{center}
\end{figure}
\section{Results}
\label{solpsn_results}
The mass of the star falls from $500\,{\ensuremath{\it{M}_{\odot}}}$ to $92.5\,{\ensuremath{\it{M}_{\odot}}}$ over its lifetime because of stellar
winds and rotation. Strong winds not only remove the hydrogen envelope but also strip away the
outer layer of the helium core. The central temperature and density of the core when it has begun
to contract are $3.31\times10^9\,\ensuremath{\mathrm{K}}$ and $1.24\times10^{6}\,\ensuremath{\mathrm{g}\,\mathrm{cm}^{-3}}$. In Figure~\ref{x500} we show
mass fractions for helium, carbon, oxygen, and silicon at the end of the GENEVA simulation. The
oxygen is nearly depleted, and explosive silicon burning is about to begin because of runaway core
contraction. Explosive silicon burning begins soon after the \texttt{CASTRO}{} run is launched, after a
brief phase of further contraction. Burning releases $3.33\times 10^{52}$ ergs and synthesizes
$3.63\,{\ensuremath{\it{M}_{\odot}}}$ of {\ensuremath{^{56}\mathrm{Ni}}}. The star is completely disrupted, leaving no compact object behind. This
large {\ensuremath{^{56}\mathrm{Ni}}}{} mass can power the PSN light curve for several months. If fluid instabilities dredge {\ensuremath{^{56}\mathrm{Ni}}}\
up from greater depths, it could affect luminosities at intermediate times. We plot oxygen and {\ensuremath{^{56}\mathrm{Ni}}}\
mass fractions $10$ s after core bounce in Figure~\ref{2d_sol_burning}. Moderate fluid instabilities
driven by nuclear burning are visible at the inner boundary of the oxygen shell, but mixing is limited.
The {\ensuremath{^{56}\mathrm{Ni}}}{} essentially remains untouched by dynamical instabilities.
The explosion drives a strong shock with initial velocities of $2-3\times10^9$ cm/s. Because the
radius of the star is only $\sim 3\times 10^{10}$ cm, the shock reaches its surface in under $20$s.
Breakout happens at a much smaller radius than in Pop III PSNe (10$^{12}$ - 10$^{13}$ cm). We
show densities and oxygen abundances at shock breakout in
Figure~\ref{2d_sol_breakout}. The SN ejecta remains roughly spherical in density and O mass
fraction, indicating there is not much mixing during the explosion. There are no prominent signs
of mixing during collapse, burning
or subsequent expansion. The reason for this can be found in the plot of $\rho r^3$ in the right
panel of Figure 1. The shock decelerates in regions of increasing $\rho r^3$ as it plows up mass,
and could form a reverse shock that is prone to Rayleigh-Taylor (RT) instabilities. But Figure~\ref{p500_rhor3}
shows that when the SN shock forms it is nearly out of the region of increasing $\rho r^3$, so there
is little time for RT instabilities, or mixing, to occur.
\begin{figure}[h]
\vspace{-.2cm}
\begin{center}
\subfigure[Fluid instabilities during the explosion]{\label{2d_sol_burning}
\includegraphics[width=0.48\textwidth]{kchen_fig3.eps}}
\subfigure[Shock breakout]{\label{2d_sol_breakout}
\includegraphics[width=0.46\textwidth]{kchen_fig4.eps}}
\caption{Left: oxygen and {\ensuremath{^{56}\mathrm{Ni}}}\ mass fractions. {\ensuremath{^{56}\mathrm{Ni}}}\ mass fractions within the red arc are greater than $0.1$.
The inner region of the oxygen-burning shell shows only mild mixing during explosive burning that
does not reach the {\ensuremath{^{56}\mathrm{Ni}}}\ layer. Right: densities (gray) and carbon abundance (contours).
The SN ejecta is fairly spherical, and there is no evidence of mixing in density or mass fractions of carbon.}
\end{center}
\end{figure}
\section{Conclusions}
\label{solpsn_conclusions}
We have examined mixing in the PSN explosions of compact non-zero metallicity stars. We
find that only mild fluid instabilities form during the explosion and do not result in visible mixing
at shock breakout. This is primarily due to the fact that RT instabilities do not have time to form
before the shock breaks out of the compact star, which is essentially a bare He core. In this
sense, such explosions may be similar to Pop~III PSNe of blue progenitors. Our model has an
explosion energy and {\ensuremath{^{56}\mathrm{Ni}}}{} yield that is similar to the $200\,{\ensuremath{\it{M}_{\odot}}}$ Pop~III PSN in \citet{
chen2014psn}.
The formation of a $500\,{\ensuremath{\it{M}_{\odot}}}$ star at near-solar metallicities is a challenge to current theories
of present-day star formation. Radiation from the star normally halts accretion well before the
star reaches such masses. Mergers between stars in dense clusters in principle could build up
stars of up to several hundred solar masses. If so, dense star clusters could be promising sites
for hunting for pair-instability supernovae in the local Universe.
\acknowledgements KC thanks the members of the CCSE at LBNL for help with \texttt{CASTRO}{}.
This work was supported by the IAU-Gruber Fellowship, Stanwood Johnston Fellowship, and
KITP Graduate Fellowship. Work at UCSC has been supported by the DOE HEP Program
under contract DE-SC0010676; the National Science Foundation (AST 0909129) and the
NASA Theory Program (NNX09AK36G). AH acknowledges support by an ARC Future Fellowship (FT120100363) and a
Monash University Larkins Fellowship. DJW acknowledges support from the European Research Council
under the European Community's Seventh Framework Programme (FP7/2007-2013) via the
ERC Advanced Grant "STARLIGHT: Formation of the First Stars" (project number 339177).
All numerical simulations were done with allocations from the University of Minnesota
Supercomputing Institute and the National Energy Research Scientific Computing Center.
|
\section{Introduction}
\label{sec:intro}
The fields of astrophysics and cosmology have benefited from the recent development of a numerical technique for effectively Lagrangian integration of the equations of hydrodynamics. This moving-mesh technique, first proposed by \cite{1985LNP...238...87B} and recently implemented in the cosmology code AREPO \citep{2010MNRAS.401..791S} and later for relativistic astrophysics in the TESS code \citep{2011ApJS..197...15D}, has had considerable success in improving upon results from Lagrangian Smoothed Particle Hydrodynamics (SPH) methods and Eulerian Adaptive Mesh Refinement (AMR) techniques. Both TESS and AREPO are three-dimensional and parallel, and have implemented solvers for the magnetohydrodynamic and viscous hydro equations. A related finite-element approach has also recently been developed which employs the same moving Voronoi mesh \citep{2014MNRAS.437..397M}. The numerical method has also been adapted for use with alternate, non-Voronoi meshes tailored to disks and jets \citep{2012ApJ...755....7D, 2013ApJ...775...87D}, as the general idea does not depend sensitively on the shape of the zones.
In general, a moving mesh can be advantageous for capturing highly supersonic flows, or situations which require precise preservation of contact discontinuities. One of the most highly-prized advantages to the moving Voronoi mesh is its ability to capture fluid instabilities at relatively low resolution \citep{2011arXiv1109.2218S}. It has also been suggested that the Voronoi mesh would be ideal for capturing turbulent flow, as the scale-by-scale structure of eddies within eddies seems naturally suited to a Lagrangian treatment \citep{2012MNRAS.423.2558B}.
However, this method is not without its detractors. A major concern regarding the Voronoi method is the noise introduced at the grid scale. First, the extremely well-preserved contact discontinuities can produce jagged edges at the grid scale, which might potentially introduce numerical issues similar to those produced by contact steepeners, which were first introduced to artificially enhance contact discontinuities \citep{1984JCoPh..54..174C}.
A much larger source of noise, however, comes from the mesh motion itself. This has been directly observed in simulations of driven turbulence. When the mesh is moved, a large amount of power is introduced artificially at the grid scale, hampering the code's ability to resolve the inertial range of turbulence at a given resolution \citep{2012MNRAS.423.2558B, 2014MNRAS.437..397M}.
The cause of this noise is easily understood. It is produced whenever two neighboring mesh points are given significantly different velocities. In this case, mesh topology can change very abruptly, either producing diffusion when a face flips and overtakes a substantial fraction of a zone, or producing noise when a face rotates quickly about its center, moving fluid around artificially even when there is no flux through the face.
Such noise will always plague Voronoi codes like TESS and AREPO, as long as there is shear in the mesh at the smallest scales. In this work, we report on a very simple way to reduce this noise, while still retaining all the advantages of the moving mesh. Simply, rather than moving the mesh-generating points with the local fluid velocity $\vec v(\vec x)$, this velocity field is smoothed on some length scale, $\lambda$: $\vec w(\vec x) = S_{\lambda}\left[ \vec v( \vec x ) \right]$, and the mesh points are moved with the velocity $\vec w$, rather than the local fluid velocity. In the examples presented here, the length scale $\lambda$ is of order the grid scale. Note that in general, one always has complete freedom when choosing the mesh velocity field $\vec w$; if $\vec w = 0$ everywhere, the method reduces to an Eulerian fixed-grid scheme. By smoothing $\vec w$, the code remains Lagrangian on large scales, but behaves like a comoving Eulerian code on small scales. The idea of using a smoothed velocity field to move mesh points has previously been proposed in the context of SPH, in the form of a technique called X-SPH \citep{1989JGR....94.6449M, 1992ARA&A..30..543M}.
It is worth noting that both TESS and AREPO already employ mesh-regularization techniques with the goal of keeping the Voronoi tessellation regular (i.e. the Voronoi cells are ``rounder" and have fewer neighbors on average). This work aims to treat a different problem with a distinct solution; the mesh velocity field is smoothed with the goal of reducing noise from rapid mesh adjustments. In doing so, there is also the side-benefit that the mesh tends to remain more regular. However, the reason for the smoothed mesh-motion is a reduction in small-scale noise.
This adaptation is simple, and it can provide improvement to the original method, as we shall demonstrate with several code tests. It is recommended that every Voronoi code use a smoothed velocity field for the mesh motion, otherwise such codes are succeptible to the aforementioned numerical noise, and the code's ability to accurately capture instabilities and turbulence (which can have shear on all scales) may be in question.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{fig1_lores.eps}
\caption{ Kelvin Helmholtz Instability, identical to the test performed in
Springel (2010). This snapshot is at time $t=1.5$. The upper panel uses a fixed mesh, which diffuses the eddies. In the center panel, the mesh is moving, and the spiral shapes formed are sharper and more pronounced. However, significant noise is introduced at the grid scale which is enhanced by the unstable shear motion, producing jagged structures. The lower panel uses the smoothed mesh motion, which significantly reduces this noise, while still maintaining contact discontinuities and capturing the growth of the instability to high accuracy (See Figures \ref{fig:kh} and \ref{fig:growth}).}
\label{fig:50}
\end{figure}
\section{Smoothed Mesh Motion}
\label{sec:smooth}
The moving Voronoi mesh technique has been articulated in other works \citep{2010MNRAS.401..791S, 2011ApJS..197...15D, 2011arXiv1109.2218S}, and we shall not repeat all of the details here. Essentially, it is a finite volume method in which the finite volumes are calculated from a Voronoi tessellation of a given set of mesh-generating marker points. These marker points can be given an arbitrary mesh velocity $\vec w(\vec x)$, though typically their motion is set to the local fluid velocity, $\vec w(\vec x) = \vec v(\vec x)$. Motion of the marker points results in mesh motion, which is accounted for by adjusting the flux through each Voronoi face, based on the face's velocity.
Note that the velocity of a face does not have to mimic the velocity of nearby marker points. In particular, if two neighboring marker points have significantly different velocities, this can result in a relative face velocity in a totally different direction from the velocity of the marker points. This generally results in a face which moves counter to the flow, which is not desirable. It can result in diffusive mixing if the face overtakes a significant fraction of one of the zones during a timestep. It can also result in noise if the face rotates about its center considerably over a timestep, an action which is not accounted for by any compensating fluxes.
This problem can be mitigated by reducing the relative velocity of neighboring marker points. In other words, instead of setting $\vec w$ equal to the local velocity, this velocity field is first processed through some smoothing operation, $\vec w = S_{\lambda}\left[ \vec v \right]$, so that neighboring marker points have similar velocities, while large-scale fluid motion is still properly subtracted off by the smoothed velocity field of the mesh.
In principle, there is freedom in the choice of the operator $S_{\lambda}$. In this work, a very simple operation is employed. Other smoothing operators are certainly possible; the choice of $S_{\lambda}$ presented here is mainly a proof-of-concept. First, the marker point velocities are set equal to the local fluid velocity, $\vec w = \vec v$. Next, an averaging operation is performed: $\vec w \rightarrow \alpha \vec w + (1-\alpha) \left< \vec w_j \right>$, where $\left< \vec w_j \right>$ represents a weighted average of the velocities of neighboring marker points (weighted by the area of the face shared in the Voronoi tessellation), and $\alpha$ is an adjustable constant chosen based on how aggressively the mesh is to be smoothed. It is then possible to perform this substitution iteratively until the velocity field $\vec w$ has the desired smoothness. In the examples presented in this work, $\alpha = 0.5$ is chosen with five iterations, but other smoothing operators could be chosen which may have more desirable properties.
\section{Code Tests}
\subsection{The Kelvin-Helmholtz Instability}
\label{sec:kh}
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{fig2_lores.eps}
\caption{ Smooth 2D Kelvin-Helmholtz test; initial conditions found in McNally et al. (2012) (snapshot at $t=1.5$, to be compared with Figure 2 of McNally et al. (2012)). Both panels ran at a resolution of $512^2$. The upper panel uses the standard mesh motion, while the lower panel uses smoothed mesh motion. Naively one might incorrectly interpret the upper panel as being more accurate; the image ``looks" more realistic, as it appears to capture details that the lower panel misses. However, since this is a smooth problem, there is a highly accurate reference solution found in McNally et al. (2012), calculated using the high-order Pencil code at high resolution. Their result does not exhibit the small-scale features found in the upper panel. Rather, these are secondary instabilities, which are artificially seeded by mesh noise.}
\label{fig:kh}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.05\columnwidth]{fig3.eps}
\caption{ Mode growth of the smooth Kelvin-Helmholtz test of Figure \ref{fig:kh} (definitions and reference solution found in ). Calculations used a mesh with $512 \times 512$ zones, and the effect of smoothing is tested by varying the number of iterations in the smoothing operator. Grid noise suppresses the growth of the instability, but by smoothing the mesh motion the growth rate is more accurately captured.}
\label{fig:growth}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=1.6\columnwidth]{fig4_lores.eps}
\caption{ Simulations of driven 2D turbulence produced by a stochastic driving field. The fluid motion is made visible by the means of a passive scalar field which was initially given by a step function in the x coordinate. The two panels are identical, except that the left panel employed standard mesh motion, while the right panel used the smoothed-mesh prescription.}
\label{fig:turb}
\end{figure*}
The Kelvin-Helmholtz instability was at first cited as one of the best arguments for using a moving Voronoi mesh. \cite{2010MNRAS.401..791S} showed that the eddies formed in the instability are much more vividly captured when moving the mesh than when the mesh is kept fixed, since a fixed mesh introduces advection errors which diffuse the eddies. Figure \ref{fig:50} shows a very low-resolution ($50 \times 50$) Kelvin-Helmholtz calculation at $t=1.5$ based on the initial conditions given in \cite{2010MNRAS.401..791S} (Springel's Figure 32). The same calculation is performed on a fixed mesh (top panel), standard moving mesh with $\vec w = \vec v$ (center), and finally smoothed mesh motion with $\vec w = S_{\lambda}[\vec v]$ (lower panel). It is clear by looking at the top and center panels that moving the mesh results in a dramatic reduction in diffusion, and much sharper contact discontinuities. The fixed mesh exhibits diffusion which comes from advection errors to which the moving mesh is less susceptible.
However, upon closer inspection, the standard moving mesh exhibits a different problem. Small-scale structures appear to form in the eddies, which may be evidence for spurious noise or numerical instability (similar features can be seen in \cite{2010MNRAS.401..791S}, Figure 32). At first glance, it might seem as though these small structures could actually be real; i.e. that they might be secondary instabilities that moving-mesh codes can capture, but which are diffused away by fixed-grid codes. In this section we will demonstrate that this is not the case. That is, secondary instabilities do exist, but in these examples they are seeded artificially by the grid-scale noise described in the previous section. The unstable nature of the flow enhances the noise and makes convergence slow. However, when the mesh motion is smoothed (lower panel of Figure \ref{fig:50}), the noise is reduced and the result is more accurate. Because the smoothing operation is not significantly diffusive, one does not have to sacrifice the sharpness of the contact discontinuities.
The calculation with smoothed mesh motion can be shown to be correct, using the smooth Kelvin-Helmholtz test proposed by \cite{2012ApJS..201...18M}. These authors started with a smooth initial condition, and evolved it using the 6th-order Pencil code \citep{2010ascl.soft10060B} at a resolution of $4096^2$. This resolution study was careful and thorough, and therefore this is likely the closest one will come to having an exact solution to this problem (relative errors in their results were of the order $10^{-3}$ or smaller everywhere on the grid; far too small to see by eye).
Figure \ref{fig:kh} shows a snapshot of the TESS code performing this problem at a resolution of $512^2$. The upper panel uses standard mesh motion, and the lower panel employes smoothed mesh motion. To the careless eye, the upper panel ``looks" more correct, as it seems as if it is picking out more details than the lower panel. However, the true, converged solution (Figure 2 in \cite{2012ApJS..201...18M}) does not exhibit these detailed features. These are secondary instabilities seeded by numerical noise. Moreover, it is very difficult to eliminate these errors with high resolution; increasing the resolution introduces the noise on smaller scales, which excites faster-growing modes. The amplitude of these errors eventually trends to zero with high resolution, but slowly.
Fortunately, smoothing the velocity field of the mesh significantly reduces this problem. Figure \ref{fig:growth} shows that the growth of the instability is more accurately captured when using smoothed mesh motion.
\subsection{Driven Turbulence}
\label{sec:turb}
\begin{figure}
\centering
\includegraphics[width=0.995\columnwidth]{fig5.eps}
\caption{ Power spectra of driven 2D turbulence using a mesh of $512 \times 512$ zones. There is a clear inertial range extending over roughly an order of magnitude in k. The inertial range shows a steep scaling somewhere between $k^{-3}$ and $k^{-5}$. At small scales, there is significant power at the dissipation scale, indicating noise induced by mesh motion. This noise is reduced and the inertial range is extended when using a smoothed velocity field for the mesh.}
\label{fig:power}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.99\columnwidth]{fig6.eps}
\caption{ Statistics of zone shapes, showing the relative frequency of zones with a given shape factor $G$ (proportional to perimiter squared over area). Zone data is taken from a $256^2$ run of the smooth KH test at $t=1.5$. $G=1$ describes a circle, while $G=1.1$ for a regular hexagon. Smoothed mesh motion reduces the number of zones in the tail, around $G = 1.4$, by nearly an order of magnitude.}
\label{fig:shape}
\end{figure}
The smoothed mesh motion is important in any problem which exhibits shear on all scales. In a turbulent cascade, there is always shear on all scales, and therefore substantial grid noise will be produced when using a Voronoi mesh to follow such a flow. If the mesh is moved, but with a smoothed prescription for the mesh motion, the mesh follows the largest eddies, and this large-scale flow is effectively subtracted off. Meanwhile, the dynamics of the smallest eddies are calculated in an effectively Eulerian manner (in the smallest eddy's center-of-momentum frame). Thus, the turbulence is captured in a natural way, but without introducing the noise produced by shearing the grid on the smallest scales.
This can be illustrated in two-dimensional simulations of driven turbulence. The fluid is stirred on large scales using a driving field generated via a stochastic process. This method for large-scale driving is identical to the one employed by \cite{2012ApJ...744...32Z}. The resultant turbulence is subsonic (Mach $0.1$). A snapshot of the 2D turbulence is shown in Figure \ref{fig:turb}, where a passive scalar is included to reveal the structures generated from the turbulent fluid motions. This passive scalar was initialized as a step function. The advantage gained in smoothing the mesh motion can be seen most clearly in the power spectrum (Figure \ref{fig:power}).
In the power spectrum, there is a clear inertial range, extending from the driving scale at $k \sim 10$ to a bottleneck at small scales. In 2D, on scales smaller than the driving scale, the power spectrum is set by an enstrophy cascade from large to small scales. Simple scaling arguments give a scale dependence of $k^{-3}$ for the power spectrum in the inertial range \citep{1967PhFl...10.1417K}, but it has been suggested that the nonlocality of 2D turbulence may cause this slope to depend on the fluid Reynolds number \citep{2010JFM...646..517B}. The important question is not of slope of the spectrum but on the extent of the inertial range and the size of the bottleneck at the dissipation scale. Clearly, the smoothed mesh motion reduces the power in this bottleneck, and increases the extent of the resolved inertial range. This could potentially be further improved with a different choice of $S_{\lambda}$.
\section{Summary}
\label{sec:sum}
The noise introduced by the Voronoi mesh motion at small scales can hamper the convergence properties of codes which employ this moving-mesh technique. Calculations which are sensitive to the detailed development of instabilities and turbulence should employ some mechanism for reducing or eliminating this noise, such as the smoothed mesh motion described in this work.
More generally, when performing astrophysical or cosmological calculations using a moving Voronoi mesh, if a certain qualitative behavior depends on whether the mesh motion is turned on, it is not necessarily the case that the moving-mesh version gives the correct answer. In the case of the smooth Kelvin-Helmholtz example of Figure \ref{fig:kh}, it may be very tempting to assume that these small structures are subtle details being picked out by the mesh motion. However, the resolution study by \cite{2012ApJS..201...18M} clearly demonstrated that this was not the case. This strongly suggests that the same is true of the non-smooth version shown in Figure \ref{fig:50}.
Given this, the smoothed mesh motion presented here provides a more accurate solution for the Kelvin-Helmholtz problem, without sacrificing the advantages gained by moving the mesh; this improvement came from subtracting noise, not from adding diffusion. The turbulence calculations (Figures \ref{fig:turb} and \ref{fig:power}) provide more evidence for this; the mesh noise is clearly exhibited in the high-k end of the power spectrum. This loud bottleneck at small scales is reduced by smoothing the mesh motion.
It should also be noted that while mesh smoothing is a distinct operation from the ``mesh steering" techniques used in AREPO and TESS (including the mesh-steering method of \cite{2012MNRAS.425.3024V}), it can help to accomplish similar goals. Figure \ref{fig:shape} shows the distribution of zone shapes, defining the ``shape factor"
\begin{equation}
G = {\rm Perimiter^2 \over 4 \pi (\rm Area)}.
\end{equation}
$G$ represents how far off zones are from being perfectly round ($G = 1$ in the limiting case of a circle, and $G = 1.1$ for a regular hexagon, but $G$ can potentially attain much larger values for non-round shapes and large aspect ratios). The figure shows that with standard mesh motion, mesh steering already does a reasonable job keeping zones round. However, smoothed mesh motion improves on this a bit, reducing the number of zones with $G \sim 1.4$ by nearly an order of magnitude.
Finally, it should be noted that the smoothing operator $S_\lambda$ presented in section \ref{sec:smooth} is not unique; undoubtedly there are other operations one can perform to smooth the velocity field $\vec w$ on small scales, which may improve performance further.
This research was supported in part by NASA through grant NNX10AF62G issued through the Astrophysics Theory Program and Fermi grant NNX13A093G and by the NSF through grant AST-1009863. We are grateful to Jim Stone, Colin McNally and Jonathan Zrake for helpful comments and discussions.
\bibliographystyle{mn2e}
|
\section{Introduction}
The large-scale structure of the Universe and the physics of galaxy formation during the epoch of Reionization are now accessible to observational tests thanks to the Hubble Space Telescope WideFieldCamera3 (WFC3). The new WFC3 transformed the field by providing the first large samples of galaxies seen in the first $500-700$ Myr after the Big Bang (redshift $z\sim7-10$), identified using the Lyman-break, or dropout, technique \citep{steidel1996}, from a variety of surveys \citep{bouwens2014,windhorst2011,trenti2011, mclure2012, bradley2012, oesch2012}. Such $z\geq6.5$ observations show that there is a decreasing number density of galaxies as the epoch of reionization is approached, albeit with growing evidence of a steepening of the galaxy luminosity function (LF) faint end \citep{bouwens2014}. By assuming that brighter galaxies are hosted in more massive dark-matter halos, the evolution of the galaxy LF can be linked to the underlying dark-matter halo mass function. Empirical models \citep{trenti2010,tacchella2013}, analytic models \citep{wyithe2013suppressed} and cosmological simulations \citep{jaacks2012, lacey2011} are all successful at reproducing the star-formation rate and LF, but there is little validation of predictions beyond galaxy number counts.
Studying the clustering properties of galaxies, through their Angular Correlation Function (ACF), is an avenue to probe the connection between the observed light and host dark-matter halos. At lower redshifts, extensive studies of clustering have been carried out, leading to measurements of the galaxy bias (the extra clustering of galaxies/halos compared to that of dark-matter), and host dark-matter halo mass \citep{adelberger1998, arnouts1999, giavalisco2001, ouchi2004, lee2006, overzier2006}.
\citet{lee2006} investigated the observed distribution of the number of galaxies in a dark-matter halo at higher redshifts, known as the halo-occupation distribution, finding evidence for multiple occupation and sub-halo clustering, extending similar studies at lower redshift (e.g., \citealt{berlind2002}).
Furthermore, clustering enables characterization of the expected evolutionary history of the observed galaxy population. For example, \citet{adelberger2005} compared their analysis of clustering of $z\lesssim3.5$ LBGs with analyses of clustering at $z=0$ to infer that their LBG sample represents the progenitors of present-day ellipticals. We can now take advantage of the large sample size and characterize the clustering properties of high-redshift Lyman-Break Galaxies (LBGs).
In this paper, we measure LBG clustering at $\overline{z}=7.2$ for the first time. We investigate qualitative trends of average bias and clustering strength with luminosity, compare our results with those at lower redshift, constrain the fraction of dark-matter halos hosting LBGs (the duty cycle), and test high-$z$ theoretical modeling of halo occupation \citep{trenti2010, wyithe2013}. This paper is organized as follows: Section \ref{section:data} describes the LBG sample. Section \ref{section:tpcf} presents our measurements of the ACF, and Section \ref{section:results} our results. Section \ref{section:conclusion} summarizes and concludes. We use the latest Planck cosmology ($\Omega_{M}$,$\Omega_{\lambda}$,$h$,$\sigma_{8}$) = ($0.315$, $0.685$, $0.673$, $0.828$) \citep{collaboration2013}. Magnitudes are in the AB system \citep{oke1974}.
\section{Data}
\label{section:data}
Our analysis is based on a combined sample of $\overline{z}=7.2$ candidate galaxies, photometrically selected as $z$ and $Y$-band dropouts from the \citet{bouwens2014} [B+14] and \citet{mclure2012} analysis of Hubble's deep and ultradeep area on the XDF/UDF fields \citep{illingworth2013} and the CANDELS data over GOODS-North and South (GN and GS, respectively; \citealt{grogin2011}). The observations span from the $4.7$ arcmin$^2$ area of the XDF, which reaches $m_{H_{160}}=29.8$ ($5\sigma$), to the $\sim280$ arcmin$^2$ of CANDELS at $m_{H_{160}}\sim27.7-28.5$. The B+14 catalogs include $N=670$ objects at $z>6.5$. The \citet{mclure2012} catalogs list $N=197$ objects within XDF/GOODS-South. Galaxies in each catalog are selected using different methods; B+14 use a two-color selection to detect the Lyman-break, while \citet{mclure2012} use multi-band spectral energy distribution fitting. A combined catalog is constructed by removing double-counted objects, and by counting the 31 pairs with separation $d<0\farcs3$ ($5$ pixels, $\sim9$kpc at $z=7.2$) as single sources, with their fluxes combined since such small separations are comparable to the HST point spread function in the $H$-band. This yields a total sample of $N=743$ objects, with $\sim80\%$ overlap for objects one magnitude above the survey detection limit, and $\sim40\%$ overlap for less robust candidates.
To ensure that merging the two catalogs does not bias our ACF measurement, we carry out a control analysis of the ACF using only the B+14 catalog (Section \ref{section:results}).
We complement our new study at $\overline{z}=7.2$, with a re-analysis of clustering at lower redshift, using the latest $B$, $V$ and $i$-dropout ($z\sim3.8-6.0$ catalogs by B+14. Table~\ref{table:fields} lists our samples, which we split between ``bright'' and ``faint'' subsamples at the median magnitude of the $\overline{z}=7.2$ sample, $M_{UV}=-19.4$, to study the redshift evolution of the bias at constant UV luminosity. A comprehensive description of the data, including number counts, luminosity functions, and redshift distributions of the samples is available in B+14.
\begin{table*}
\begin{center}
\begin{tabular}{c | c | c | c | c}
\hline
\hline
Field & $\overline{z}=3.8$ & $\overline{z}=4.9$ & $\overline{z}=5.9$ & $\overline{z}=7.2$\\
\hline
XDF &\textbf{385}(\textbf{57}) &\textbf{157}(\textbf{19}) &\textbf{104}(\textbf{13}) &\textbf{149}(\textbf{11})\\
HUDF091 &--- &\textbf{93}(\textbf{24}) &\textbf{38}(\textbf{9}) &\textbf{52}(\textbf{9})\\
HUDF092 &\textbf{147}(\textbf{50}) &\textbf{83}(\textbf{17}) &\textbf{36}(\textbf{10}) &\textbf{54}(\textbf{8})\\
GS-Deep &\textbf{1621}(\textbf{756})&\textbf{537}(\textbf{251}) &\textbf{203}(\textbf{102}) &\textbf{134}(\textbf{76})\\
GS-ERS &\textbf{757}(\textbf{445}) &\textbf{209}(\textbf{150}) &\textbf{62}(\textbf{49}) &\textbf{64}(\textbf{52})\\
GS-Wide &\textbf{422}(\textbf{304}) &\textbf{122}(\textbf{98}) &\textbf{41}(\textbf{36}) &13(11)\\
GN-Deep &\textbf{1721}(\textbf{770})&\textbf{745}(\textbf{352}) &\textbf{197}(\textbf{108}) &\textbf{220}(\textbf{151})\\
GN-Wide 1 &\textbf{398}(\textbf{302}) &\textbf{138}(\textbf{116}) &25(24) &33(33)\\
GN-Wide 2 &\textbf{485}(\textbf{346}) &\textbf{173}(\textbf{130}) &\textbf{51}(\textbf{44}) &24(24)\\
\hline
TOTAL & 5936(3030) &2257(1157) &757(395) &743(375)\\
\hline
\end{tabular}
\caption{Number of LBGs (bright LBGs) in each field at each redshift. The data are divided into bright and faint subsamples based on the median luminosity of the $\overline{z}=7.2$) sample $M_{UV}=-19.4$, which is within $\sim0.1$ mag of the median of samples at lower-$z$. Only entries in \textbf{bold} are used in the ACF analysis, since we discard fields where mean galaxy separation exceeds $>100\farcs0$.}
\label{table:fields}
\end{center}
\end{table*}
\begin{table*}
\begin{center}
\begin{tabular}{c | c c c |c c c |c c c}
\hline
\hline
$\overline{z}$ & $A_{w,total}$ & $r_{0, total}$ & $b_{total}$ & $A_{w,bright}$ & $r_{0, bright}$ & $b_{bright}$ & $A_{w,faint}$ & $r_{0, faint}$ & $b_{faint}$\\
\hline
$3.8$ & $0.44^{+0.05}_{-0.05}$ & $3.5^{+0.2}_{-0.2}$ & $3.0^{+0.2}_{-0.2}$ & $0.80^{+0.10}_{-0.10}$ & $5.1^{+0.4}_{-0.4}$ & $4.0^{+0.3}_{-0.3}$ & $0.24^{+0.08}_{-0.08}$ & $2.4^{+0.5}_{-0.5}$ & $2.2^{+0.3}_{-0.4}$\\
$4.9$ & $0.60^{+0.13}_{-0.13}$ & $3.4^{+0.4}_{-0.5}$ & $3.6^{+0.4}_{-0.4}$ & $0.68^{+0.27}_{-0.26}$ & $3.7^{+0.9}_{-1.0}$ & $3.8^{+0.7}_{-0.8}$ & $0.54^{+0.22}_{-0.22}$ & $3.2^{+0.8}_{-0.9}$ & $3.4^{+0.6}_{-0.8}$\\
$5.9$ & $0.93^{+0.34}_{-0.34}$ & $3.6^{+0.8}_{-0.9}$ & $4.4^{+0.8}_{-0.9}$ & $1.79^{+0.77}_{-0.76}$ & $5.5^{+1.4}_{-1.6}$ & $6.2^{+1.2}_{-1.5}$ & $0.34^{+0.35}_{-0.23}$ & $1.9^{+1.1}_{-1.0}$ & $2.7^{+1.2}_{-1.2}$\\
$7.2$ & $1.40^{+0.31}_{-0.31}$ & $6.7^{+0.9}_{-1.0}$ & $8.6^{+0.9}_{-1.0}$ & $1.77^{+0.82}_{-0.80}$ & $7.8^{+2.1}_{-2.5}$ & $9.7^{+2.0}_{-2.5}$ & $1.24^{+0.43}_{-0.43}$ & $6.2^{+1.3}_{-1.4}$ & $8.1^{+1.3}_{-1.6}$\\
\hline
\end{tabular}
\caption{The free ACF parameter, $A_{w}$ ($\beta=0.6$ is fixed), real-space correlation length and galaxy bias for the total, bright and faint samples at each redshift. Units of $r_{0}$ are $h^{-1}$Mpc.}
\label{table:values}
\end{center}
\end{table*}
\section{Estimating the angular correlation function and bias}
\label{section:tpcf}
The ACF measures the excess probability of finding two galaxies at an angular separation $\theta$, over a random Poisson point process. We use the ACF estimator, $w(\theta)=(DD(\theta)-2DR(\theta)+RR(\theta))/RR(\theta)$ \citep{landy1993}, where $DD(\theta)$, $DR(\theta)$ and $RR(\theta)$ are the number of galaxy-galaxy pairs, galaxy-random point pairs and random-random point pairs (respectively) counted at a separation of $\theta\pm\delta\theta$. We produced random point catalogs using a spatial Poisson process, accounting for field geometry and crowding effects using segmentation maps produced by $\tt{SExtractor}$ \citep{bertin1996}.
We use linear binning with width $12\farcs5$ to construct the ACF for the full sample. For subsamples at $\overline{z}=5.9$ and $\overline{z}=7.2$ we increase bin-width to $25\farcs0$ to account for the reduced number of pairs. We assume a power law parameterization of the ACF, $w(\theta)=A_{w}\theta^{-\beta}$. Following previous investigations \citep{lee2006,overzier2006}, we fix $\beta=0.6$, but we quantify in Section \ref{section:results} the systematic uncertainty associated with this choice. Because of finite survey area, the observed ACF is underestimated by a constant known as the integral constraint (IC):
\begin{equation}
w_{\textnormal{true}}(\theta)=w_{\textnormal{obs}}(\theta)+\textnormal{IC}.
\end{equation}
IC can be estimated from:
\begin{align}
\textnormal{IC}&=\frac{1}{\Omega^{2}}\int_{1}\int_{2}w_{\textnormal{true}}(\theta)d\Omega_{1}d\Omega_{2}\\
&=\frac{\sum_{i}RR(\theta_{i}) w_{\textnormal{true}}(\theta_{i}) }{\sum_{i}RR(\theta_{i}) }=\frac{\sum_{i}RR(\theta_{i})A_{w}\theta_{i}^{-\beta}}{\Sigma_{i}RR(\theta_{i})},
\end{align}
where $w_{\textnormal{true}}(\theta)$ is the intrinsic ACF and $w_{\textnormal{obs}}(\theta)$ is the measurement within the survey area, $\Omega$. For small area fields such as the XDF, the integral constraint is $0.96A_{w}$, while in larger fields (GOODS-S Deep) the IC starts becoming a second-order correction ($IC=0.04A_{w}$). The only uncertainty in the IC derives from assuming fixed $\beta$. $A_{w}$ is obtained from the observed ACF in the fields (each with their own IC) by maximizing the likelihood given by:
\begin{equation}
\mathscr{L}=\prod_{i=\textrm{fields}}\frac{1}{\sigma(\theta)_{obs,i}\sqrt{2\pi}} e^{-\frac{1}{2}\Big(\frac{w(\theta)_{obs,i}-A_{w}(\theta^{-\beta}-\frac{\textrm{IC}_{i}}{A_{w}})}{\sigma(\theta)_{obs,i}}\Big)^{2}}
\label{eq:fit}
\end{equation}
{where $w(\theta)_{obs,i}$, IC$_{i}$ and $\sigma(\theta)_{i}$ are the ACF measurements, integral constraints and uncertainties in field $i$, respectively.
After $A_{w}$ is determined, IC values can be added to the ACF measurements in each field, and combined to determine the intrinsic ACF shown in Figure \ref{figure:tpcf-6.5+}. Errors in the ACF are estimated using bootstrap resampling \citep{ling1986}, without including systematic sample (``cosmic'') variance uncertainty, which might be comparable to the random error (e.g., \citealt{trenti08}, and see Section~\ref{section:results} for field-to-field variations in our analysis).
We approximate the real-space correlation function with a power law $\xi(r)=(r/r_{0})^{-\gamma}$, where the coefficients are related to the ACF coefficients by the Limber transform \citep{peebles1980, adelberger2005},
\begin{align}
&\beta=\gamma-1,\\
&A_{w}=\frac{r_{0}^{\gamma} B[1/2,(\gamma-1)]\int_{0}^{\infty}dzN(z)^{2}f^{1-\gamma}g(z)^{-1}}{\big[\int_{0}^{\infty} N(z)\big]^{-2}}.
\end{align}
Here $f\equiv(1+z)D_{A}(\theta)$ where $D_{A}(\theta)$ is the angular diameter distance and $N(z)$ is the redshift distribution of the dropouts, $B$ is the beta function and $g(z)\equiv c/H(z)$. We assume the clustering evolution to be fixed in comoving coordinates over each redshift window. $N(z)$ is taken from B+14 (see their Figure~1 for quantitative details), with mean values reported in Table~\ref{table:values}, and typical standard deviations $\approx0.4$. From the real-space correlation function, we define the galaxy bias $b$ as the ratio of the galaxy variance at $8h^{-1}$cMpc (comoving megaparsecs), $\sigma_{8,g}$, to the linear matter fluctuation at $8h^{-1}$cMpc, $\sigma_{8}$:
\begin{align}
b &=\frac{\sigma_{8,g}}{\sigma_{8}},\textnormal{where}\\
\sigma_{8,g}^{2}&=\frac{72(r_{0}/8h^{-1}\textnormal{cMpc})^{\gamma}}{(3-\gamma)(4-\gamma)(6-\gamma)2^{\gamma}}
\end{align}
\section{Results}\label{section:results}
\subsection{Angular Correlation Function and Bias}
The right-most upper panel of Figure~\ref{figure:tpcf-6.5+} shows the combined ACF at $\overline{z}=7.2$. There is a clear clustering signal, detected at high confidence ($\gtrsim4\sigma$), and corresponding to a clustering length $r_{0}=6.7^{+0.9}_{-1.0} h^{-1}$cMpc. For comparison, Figure~\ref{figure:tpcf-6.5+} also shows the ACF at $\overline{z}=3.9$, $\overline{z}=4.9$ and $\overline{z}=5.9$. Table \ref{table:values} summarizes our estimates of $A_{w}$, $r_{0}$ and $b$ for these ACF measurements. Figure~\ref{figure:bias} illustrates the evolution of the galaxy bias with redshift, and compares the bias of dark-matter halos at different masses computed with the \citet{sheth1999} formalism. The bias at $\overline{z}=7.2$ is $b=8.6^{+0.9}_{-1.0}$, with a clear increase compared to the lower redshift measurements. The evolution of the clustering strength with redshift broadly follows the increase expected at an approximately constant dark-matter halo mass of $M\sim10^{10.5}~$$M_{\odot}\hspace{1mm}$ out to $z\sim6$, followed by a possible increase to $M\sim10^{11}~$$M_{\odot}\hspace{1mm}$ at $z\sim7$ ($\lesssim2\sigma$ confidence). The near-constant trend between $z=5.9$ and $z=3.8$ is consistent with empirical models \citep{trenti2013, tacchella2013}, which predict no evolution or a mild decrease in halo masses at fixed UV-luminosity. In this respect, the $\overline{z}=7.2$ result is unexpected and could potentially indicate a change in star formation efficiency or clustering properties as the Universe becomes more neutral, but the uncertainty is still large.
To investigate luminosity dependence, we split galaxies in our samples into bright and faint subsamples. We cut at the median luminosity of the $\overline{z}=7.2$ sample, $M_{UV}=-19.4$. Since LF shape evolution approximately cancels out the evolution in the $k$-corrected distance modulus, this median cut is within $0.1$ mag of the median of each sample from $z=3.8$ to $z=7.8$ allowing us to explore bias variation at near-constant UV luminosity\footnote{Note that the cut is approximate at the level of $\lesssim0.1$ mag., since the distance modulus is computed at the mean redshift for each dropout sample}. At $\overline{z}=7.2$, the clustering strength in the bright sample is $r_{0}=7.8^{+2.1}_{-2.5} h^{-1}$cMpc, while the faint sample has $r_{0}=6.2^{+1.3}_{-1.4} h^{-1}$cMpc, smaller, but consistent with no luminosity evolution within the measurement uncertainty. The presence of a luminosiy trend is more apparent at $\overline{z}=5.9$, where we observe galaxies in the bright subsample ($\overline{M}_{UV}=-20.2$) residing in more massive haloes ($M=10^{11.0^{+0.4}_{-0.6}}$$M_{\odot}\hspace{1mm}$) than those in the faint sample ($\overline{M}_{UV}=-18.6$, $M=10^{9.0^{+0.9}_{-2.1}}$$M_{\odot}\hspace{1mm}$). This is consistent with a constant mass-to-light ratio with luminosity in the $\overline{z}=5.9$ sample $\frac{\textnormal{dlogM}}{\textnormal{dlogL}}\sim1.2^{+1.8}_{-0.8}$. At lower redshift we observe a similar behavior (see also \citealt{lee2006}), with a high-confidence measurement at $\overline{z}=3.8$ (see Fig.~\ref{figure:bias} for the bias difference between LBGs in the bright and faint subsamples at $\overline{z}=3.8$ and $\overline{z}=5.9$).
\subsection{Uncertainties and Analysis Validation}\label{sec:errors}
The uncertainties presented here derive from bootstrap resampling. These may underestimate the true uncertainty because they are derived at fixed $\beta$ and cosmic variance is not accounted for. To estimate cosmic variance, we use field-to-field variations in the bias, which are shown in Figure~\ref{figure:field_to_field}. Based on a $\chi^2$ analysis of the residuals between all-fields versus individual determination of the bias, we estimate that for each \emph{individual field} measurement in the combined sample (left panel), cosmic variance uncertainty is comparable to the random error (excluding $z=3.8$). The impact on the all-field determination is mitigated by analyzing independent regions in the sky, and if we conservatively assume only two independent pointings (GOODS-N and S), we get that the systematic cosmic variance uncertainty is $\Delta b=\pm1.0$ for our new measurement at $z=7.2$. The current high-$z$ subsamples have larger random errors, thus the impact of cosmic variance is negligible compared to the random errors. We also carried out our analysis assuming different values for $\beta$. At $z=7.2$, we obtain that $\Delta\beta=\pm0.1$ corresponds to $\Delta b\sim\mp0.9$. With a higher $\beta=0.8$, we would derive $b_{\overline{z}=7.2}=6.94^{+0.76}_{-0.85}$. We investigate the impact of changing the mean and width of $N(z)$ and find that shifting $N(z)$ by $\Delta z=\pm0.2$ changes our bias estimate by $\Delta b=\pm0.1$ at each redshift interval. A $10\%$ change to the width of $N(z)$ changes our bias estimate by $\Delta b=\pm0.2$.
To ensure robustness of our analysis against the combining of catalogs, we performed several checks. We first split the $\overline{z}=7.2$ sample into $z$-dropouts ($\overline{z}=6.8$) and $Y$-dropouts ($\overline{z}=7.9$), obtaining $A_{w}=1.6^{+0.4}_{-0.4}$, $r_{0}=5.9^{+0.9}_{-1.0} h^{-1}$cMpc, $b=7.4^{+0.9}_{-1.0}$ at $\overline{z}=6.8$, and $A_{w}=1.7^{+0.9}_{-0.8}$, $r_{0}=6.1^{+1.8}_{-2.1} h^{-1}$cMpc, $b=8.6^{+2.0}_{-2.5}$ at $\overline{z}=7.8$. Additionally, we restrict the analysis to the B+14 sample, finding $A_{w}=1.5^{+0.4}_{-0.4}$, $r_{0}=7.0^{+1.1}_{-1.2} h^{-1}$cMpc, $b=8.9^{+1.1}_{-1.3}$. These determinations are all consistent with results from the full sample, demonstrating absence of systematic errors introduced by combining heterogeneous datasets.
To investigate the impact of multiple halo occupancy, which we are neglecting in our analysis but may affect the ACF in the innermost $\sim10''$ (see \citealt{lee2006}), we carried out our maximum likelihood determination of the ACF by excluding datapoints in the innermost $12.5''$. At $z=7.2$ we obtain $b=9.03^{+1.12}_{-1.29}$ for the full sample, which is completely consistent with $b=8.6\pm1.0$ from the full analysis. The lack of evidence for multiple halo occupation is expected, since multiple galaxies per halo are expected only when $M\gtrsim10^{12}~\mathrm{M_{\sun}}$ (very rare at $z>6.5$).
Finally, we verified that our low-$z$ analysis is consistent with previous published investigations. The correlation lengths of $B$ and $V$ dropouts are in general agreement with those presented in \citet{lee2006}, who found $r_{0}=2.9^{+0.2}_{-0.2}$cMpc at $z\sim3.8$ and $r_{0}=4.4^{+0.5}_{-0.5}$cMpc at $z\sim4.9$, both with a fixed slope of $\beta=0.6$, for their largest sample at each redshift. Our results at $z\sim5.9$ are consistent with those presented in \citet{overzier2006}, who found $r_{0}=4.5^{+2.1}_{-3.2}$ in the GOODS fields.
\subsection{Duty Cycle}
The duty cycle of LBGs is defined as the fraction of dark-matter halos hosting UV-bright LBGs and is linked to the star formation efficiency realized in the halos. The duty cycle depends on the occupation efficiency of dark matter haloes and the timescale that galaxies remain visible at ultraviolet wavelengths. Our clustering measurements allow us to estimate duty cycles by comparing the observed bias with that predicted by ``abudance matching'', which connects luminosity functions to the dark-matter halo mass functions (e.g., \citealt{vale04}) by matching objects at the same comoving density. Abundance matching is carried out assuming only a fraction $\epsilon_{DC}\leq1$ of halos hosts a UV-bright galaxy. The number density of haloes above a mass $M_{h}$ is matched to the number density of galaxies above a luminosity $L_{g}$ using the duty cycle, $\epsilon_{DC}\leq1$ \citep{trenti2010}:
\begin{equation}
\epsilon_{DC}\int^{+\infty}_{M_{h}}n(M_{h},z)dM_{h} = \int_{L_{g}}^{+\infty}\psi(L,z)dL
\end{equation}
where $\psi(L,z)$ is the luminosity function at redshift $z$, $n(M_{h},z)$ is the halo mass function as redshift $z$ and $\epsilon_{DC}$ is the duty cycle. We construct a mass-luminosity relation at $\overline{z}=7.2$ from abundance matching between the \citeauthor{sheth1999} mass function and a Schechter LF, defined as,
\begin{equation}
n(M)=\phi^{\star}\frac{\textrm{ln}(10)}{2.5}10^{-0.4(M-M_{\star})(\alpha+1)}e^{10^{-0.4(M-M^{\star})}}
\end{equation}
with $\phi^{\star}=0.82\times10^{-3}$, $M^{\star}=-20.2$, $\alpha=-1.86$ \citep{bouwens2011}. Matching luminosity with dark-matter halo mass gives us the average host halo mass of the galaxies, and thus a halo bias from the \citeauthor{sheth1999} formalism which depends on $\epsilon_{DC}$. This is shown in the bottom panels of Figure~\ref{figure:tpcf-6.5+}. By taking the intersection between the bias inferred from abudance matching and from clustering analysis, we derive the value of $\epsilon_{DC}$.
From Figure~\ref{figure:tpcf-6.5+}, we see that at $\overline{z}=7.2$ duty cycles near $\epsilon_{\textnormal{DC}}=1.0$ are favored for the total sample, but all duty cycle values from abundance matching are in mild tension with the observed bias (which is too high). This discrepancy could be due to the systematic errors present in the analysis (cosmic variance and/or fixed $\beta$). As discussed in Section \ref{sec:errors}, the tension at $\overline{z}=7.2$ is resolved by using a value of $\beta=0.8$. Including the fixed-$\beta$ and cosmic variance uncertainties will approximately double the total uncertainty, compared to the random error $1\sigma$ interval shown in Figure \ref{figure:bias}.
At $z\sim7$ our duty cycle measurement contrasts with the theoretical expectation from \citet{wyithe2013} who noted that for $z>4$ a low duty cycle ($\epsilon_{DC}\sim0.1-0.15$) would be required to resolve the tension between the observed star-formation rate with the comparatively low observed total stellar mass. This prediction is consistent with our measurements at $4\lesssim z\lesssim6$. Another explanation for the discrepancy may be the effect of reionization on galaxy bias measurements. \citet{wyithe2007} showed that reionization, which occurs earlier in overdense regions, can enhance the observed galaxy bias in flux-limited samples. Early reionization of overdense regions leads to different star-formation histories due to heating of the IGM, and hence a different luminosity at fixed stellar mass. This results in overdense regions hosting a younger, brighter population of LBGs, which causes the observed galaxy bias to be overestimated. At $z\sim7$, this effect may contribute at the $\sim10\%$ level ($\Delta b\sim1$) \citep{wyithe2007}, which is similar to the discrepancy observed in the lower panels of Figure \ref{figure:tpcf-6.5+}. This effect also results in an overestimation of the halo mass from the bias (a factor of $\sim3$ for halos with $M\sim10^{11}$$M_{\odot}\hspace{1mm}$).
Previous studies, such as \citet{lee2009}, found lower duty cycles at lower redshifts. As starburts remain UV-bright for $\sim100$Myr, a shorter halo assembly time at high redshifts results in a higher duty cycle, as more haloes will be seen to host UV-bright galaxies (e.g. see the top-right panel of Fig.~1 in \citealt{trenti2010}). This may qualitatively explain the increase in duty cycle observed with redshift. However, the current uncertainties in the analysis make it difficult to draw any firm conclusion on the redshift evolution of $\epsilon_{\textnormal{DC}}$.
\section{Conclusion}
\label{section:conclusion}
In this paper we carry out the first clustering analysis of LBG galaxies at $z\geq6.5$, taking advantage of a combined sample of $N=743$ candidates from the XDF, GOODS-S and GOODS-N legacy fields. We detect a positive clustering signal with $\gtrsim4\sigma$ confidence, finding the real-space correlation length and bias of our $\overline{z}=7.2$ sample to be $r_{0}=6.7^{+0.9}_{-1.0} h^{-1}$cMpc and $b=8.6^{+0.9}_{-1.0}$ respectively. This result, and its interpretation, provides fundamental insight into several aspects of galaxy formation and evolution during the epoch of reionization:
\begin{itemize}
\item We show that the bias of galaxies at fixed UV-luminosity clearly increases with redshift, from $b\sim3.0$ at $z\sim4$ to $b\sim8.6$ at $z\gtrsim7$, implying that LBGs of a fixed luminosity reside in dark-matter halos with similar masses at $4\lesssim z\lesssim6.5$, followed by a moderate halo mass increase at $z\gtrsim7$. While the current uncertainty on the bias is still too large to provide stringent tests of models of galaxy formation, the tension between predicted and measured bias at $z\gtrsim7$ could be linked to the effect of reionization \citep{wyithe2007}. Overall our analysis clearly demonstrates the potential of clustering analysis in future datasets;
\item We constrain the mass of dark-matter halos hosting the LBGs observed by Hubble, finding that $z\gtrsim7$ galaxies live in halos with $M\sim10^{11}$~$M_{\odot}\hspace{1mm}$. Since dark-matter halos grow by around two orders of magnitude in mass from $z\sim7$ to $z\sim0$ based on Extended Press-Schechter modeling (e.g., Fig. 7 in \citealt{trenti12a}), this implies that the observed population of LBGs at $z>4$ will predominantly end up in a group or small-cluster environment by the present time;
\item We observe luminosity-dependent clustering within the $\overline{z}=5.9$ sample when split at $M_{UV}=-19.4$. The change in halo bias between these two samples implies a mass-to-light ratio of $\frac{\textnormal{dlogM}}{\textnormal{dlogL}}\sim1.2^{+1.8}_{-0.8}$. A similar trend is present for galaxies at $z\geq6.5$, but with larger uncertainty;
\item Finally, we provide a first constraint on the duty cycle of $z\geq6.5$ LBGs, finding values close to unity (Figure~\ref{figure:tpcf-6.5+}) and a possible evolution of the duty cycle with redshift. Such evolution is likely related to the shorter halo assembly times at increasing redshift;
\end{itemize}
These results highlight the significance of the progress in the study of galaxies during the first Gyr, made possible by the installation of WFC3 on Hubble and by the telescope time devoted to deep surveys such as XDF and CANDELS. The measure of galaxy clustering has been a fundamental tool in galaxy formation and evolution at lower redshift starting from the establishment of the cosmic web and its connection to dark-matter halos (e.g. \citealt{davis85}). Now, we have demonstrated how this measurement is possible for samples of objects during reionization: While the uncertainty on the galaxy bias at $z\geq6.5$ is still significant, upcoming programs such as the Frontier Fields Initiative will provide even stronger constraints on the connection between baryons and dark-matter in the youth of the Universe.
\smallskip
\smallskip
\acknowledgements
We thank the anonymous referee for their useful suggestions and comments that have improved the manuscript. Support for this work was partially provided by the Australian Research Council through an Australian Laureate Fellowship FL110100072 and through Discovery Project DP140103498; by the European Commission through the Marie Curie Career Integration Fellowship PCIG12-GA-2012-333749; and by NASA through Hubble Fellowship grant HF-51278.01 and grants HSTGO-12905, HSTGO-12572, and HSTGO-11563.
|
\section{Introduction}
Phylogenetic trees, or evolutionary trees, are used in evolutionary
biology to represent the evolutionary history of an extant set of species.
In a rooted phylogenetic tree, the leaves are uniquely labelled by these species, while the unlabelled
internal nodes represent their ancestors. The root represents
the universal common ancestor of all the species. \\
In a phylogenetic tree, the \emph{out-degree} of an internal node is the number
of its children. The distance between two nodes
represents an evolutionary distance such as time or the number of mutations.
This kind of representation is appropriate for many groups of species
which include the mammals. \\
It has been observed the evolutionary patterns are not the same for all groups.
Sometimes, reticulation events come into play that do not conform to a
tree-like evolutionary process. Rather, the species under reticulation
events form a composite of genes derived from different ancestors. Indeed, reticulation
involves a gamut of events that includes hybridization, horizontal gene transfer and recombination.
In this paper, we focus primarily on hybridization. \\
Research over the years into the evolutionary history of Eukaryotes has revealed the existence
of hybridization events among certain groups of plants, birds
and fish. Spontaneous hybridization events have also been reported in
the evolutionary history of some mammals and even primates. Study
on hybridization shows that at least 25\% of plant species and 10\%
of animal species, mostly the youngest species, are involved in hybridization
events \cite{key-10}. \\
Several techniques have been devised to reconstruct phylogenetic trees
for a given set of species. This got Computational Biologists interested in the
problem of determining the 'distance' between two such trees. Distance metrics such as NNI
(Nearest Neighbor Interchange), SPR (Subtree Prune and Regraft) and
TBR (tree bisection and reconnection) have been proposed in \cite{key-11}
for measuring the distance between the two phylogenetic trees. In
a pioneering paper, Allen and Steel \cite{key-1} proposed algorithms for
estimating these distances. The hybridization number and the rooted
SPR (rSPR) distance have proven to be very useful tools for estimating
the reticulation events that have occurred. Baroni et al \cite{key-2}.
showed that the rSPR distance provides a lower bound on the number of
reticulation events. \\
Computing hybridization number, rSPR and TBR distances have been shown
to be NP-hard problems. This triggered interest in designing approximation and fixed parameter
tractable algorithms for these problems.
Hein et al. \cite{key-8} introduced the idea of a Maximum
Agreement Forest(MAF) as a new tool to determine the distance between
two phylogenies. They proposed a 3-approximation ratio algorithm
exists for computing a MAF for 2 trees and a NP-hardness
proof for the rSPR distance problem computation.
Allen and Steel \cite{key-1} showed that the TBR distance between two trees is equal to the number of
components in a MAF. They also corrected an oversight in Hein et al's paper \cite{key-8} to show
that the TBR problem is NP-hard.
Rodrigues et al. \cite{key-11} reported a 3-approximation algorithm
for computing a MAF for 2 trees and generalized this to a $(d+1)$-approximation
algorithm for computing a MAF for 2 trees with degree at most $d$.
Bonet et al \cite{key-15} pointed out that 3-approximation algorithms claimed by \cite{key-11} and \cite{key-8}
are in error and proposed a 5-approximation algorithm for the rSPR problem.
Bordewich and Semple \cite{key-3} showed that the rSPR distance between two rooted trees
is equal to the number of components in a MAF of these trees, and used this result to show the
computing the rSPR distance is NP-complete.
In a subsequent paper, Bordewich et al \cite{key-4} proposed a
3-approximation algorithm for computing the rSPR distance between two trees and a
fixed-parameter tractable algorithm of time-complexity in $O(4^kk^4+ n^3)$.
Baroni et al \cite{key-2} introduced the concept of Maximum Acyclic Agreement Forest(MAAF) and
showed that the hybridization number of two trees is one less than the number of components in a MAAF.
Chataigner \cite{key-5} obtained an 8-approximation ratio algorithm for computing the MAF on \textit{k}
($\geq$2) trees. \\
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig1.pdf}
\caption{\emph{Phylogenetic Trees}}
\label{fig:PhylogenyTrees}
\end{figure}
In this paper, we propose a 3-approximation algorithm for computing
a MAF on $k (\geq 2)$ trees by a simple extension of Bordewich et al's 3-approximation algorithm
for 2 trees \cite{key-4}. We also propose a new 3-approximation algorithm for computing a MAAF on $k (\geq 2)$ trees.
\begin{figure}
\centering
\includegraphics[scale=0.2]{fig2.pdf}
\caption{\emph{Maximum Agreement Forest of the trees in Fig.1}}
\label{fig:MAF}
\end{figure}
\section{Preliminaries}
In this section, following {[}1,2,4{]}, we introduce the terminology and notation used in this paper.
\subsection{Hybridization number}
If \textit{v} is a vertex of a directed graph (digraph, for short) \textit{D}, we
denote the in-degree of $v$ by \textit{d$^{\text{-}}$(v)} and its out-degree by
and \textit{d$^{\text{+}}$(v)}. A hybrid phylogenetic network, \textit{H}, on an extant set of species \textit{X}
consists of:
\medskip{}
(i) a \textit{rooted} acyclic digraph \textit{D} in which the root has out-degree
at least two and, for all vertices\textit{ v} with \textit{d$^{\text{+}}$(v)}
= 1, we have \textit{d$^{-}$(v)} $\geq$ 2, and \\
(ii) the set of vertices of \textit{D} with out-degree zero is precisely \textit{X}. \\
\textit{X} is called the \textit{label set} of \textit{H} and is also denoted by \textit{L(H)}. Vertices with
in-degree at least two are called \textit{hybrid vertices}.
These vertices represent an exchange
of genetic information between hypothetical ancestors. For a hybrid phylogeny
\textit{H} on \textit{X} with root $\rho$, its \emph{hybridization number}, \textit{h(H)}, is:
\begin{center}
$h(H) = \sum_{v \neq \rho}(d^-(v) - 1)$
\end{center}
\medskip{}
A rooted binary phylogenetic $X$-tree is a special type of hybrid phylogeny
in which the root has degree two and all other interior vertices have degree
three, while $X$ is its leaf-set. For two rooted binary phylogenetic \textit{X}-trees
\textit{T} and $T'$, we define:
\begin{center}
\textit{h(T,$T'$) = min \{h(H) : H is a hybrid on X that displays T and $T'$\}}
\end{center}
\subsection{Agreement Forest}
Let $T$ and $T'$ be two rooted binary phylogenetic trees.
We denote the set of leaf labels of $T$ by $L(T)$, and the
set of its edges by $E(T)$. The extent of similarity between the
two trees can be quantified by computing an \emph{agreement forest}.
This useful notion, introduced by Hein et al \cite{key-8}, is formally
defined as follows: \\
An \emph{agreement forest} $F$ for
$T$ and $T'$ is a collection of rooted binary phylogenetic
trees $t_1$, $t_2$,..., $t_n$ such that:
\begin{itemize}
\item for each $i$, $L(t_i) \subseteq L(T)$ and $\bigcup L(t_i) = L(T)$;
\item for each $t_i$, $S_i$ is the minimal subtree connecting the nodes of $L(t_i)$ in $T$; it
is identical with $t_i$ when nodes of degree 2 in $S_i$ are contracted;
\item for $i \neq j$, $S_i$ and $S_j$ are node disjoint.
\end{itemize}
The size of an agreement forest is the number of trees (or components) in the forest.
An agreement forest with the \textit{smallest} number of components is a \emph{Maximum Agreement Forest} (MAF)
for $T$ and $T'$. \\
An agreement forest is obtained by cutting the same number of edges from both
\textit{T} and $T'$, followed by contraction of degree 2 nodes from the residual trees.
The deleted edges are those which do not agree in $T$ and $T'$, suggesting that they represent different
paths of genetic inheritance i.e. hybridization events (see Figure~\ref{fig:PhylogenyTrees} and Figure~\ref{fig:MAF}). \\
The notion of a Maximum Acyclic Agreement Forest (MAAF) was introduced in [1]
to exclude agreement forests in which any vertex in the associated hybrid phylogeny inherits genetic
information from its own descendants. \\
Let $F = \{t_1, t_2, \ldots, t_n \}$ be an agreement forest for $T$ and $T'$.
Let $G_F$ be the directed graph whose vertex set
is $F$ and there is an edge from $t_i$ to $t_j$ if $i \neq j$ and
\begin{itemize}
\item either the root of $T(L(t_i))$ is an ancestor of the root of $T(L(t_j))$,
\item or the root of $T'(L(t_i))$ is an ancestor of the root of $T'(L(t_j))$.
\end{itemize}
Note that since $F$ is an agreement forest,
$T(L(t_i))$ and $T(L(t_j))$ have different roots; the
same is true of $T'(L(t_i))$ and $T'(L(t_i))$.
We say that $F$ is an \emph{acyclic agreement forest} if $G_F$ is acyclic.
$F$ is a \emph{Maximum Acyclic Agreement Forest} (MAAF) if
it has the smallest number of components.
In short, an agreement forest is a MAAF if it is a MAF and and $G_F$ is acyclic. \\
The importance of a MAAF stems from the following theorem, proved in [2].
\begin{theorem}
The hybridization number of T and $T'$ is equal to the size
of a MAAF for T and $T'$ minus one.
\end{theorem}
\subsection{Rooted Subtree Prune and Regraft(rSPR)}
Due to reticulation, two phylogenies with the same set of species exhibit
inconsistencies in their parent-child relationships.
One approach to quantifying this is to compute the rooted subtree prune and regraft (rSPR) distance between
two phylogenies [4]. This can be done by a series rSPR operations as explained below. \\
Figure~\ref{fig:spr} illustrates a typical rSPR operation in which the subtree $T_u$ rooted at $u$ is pruned and grafted
to another part of the tree by connecting $u$ to a newly-created vertex $x$ (on some pre-existing edge).
The vertex $v$, now of degree 2, is removed. For more details on rSPR operations see [3].
rSPR$(T,T')$ measures the minimum number of rSPR operations required to transform $T$ into $T'$ and is equal to the size of a MAF minus 1 (shown in [3]).
\begin{figure}
\centering
\includegraphics[trim=0mm 90mm 0mm 0mm, clip, width = 0.5\columnwidth]{fig3.pdf}
\caption{\emph{An rSPR operation}}
\label{fig:spr}
\end{figure}
\begin{comment}
\textbf{Definition 1:} For a tree \textit{T} and a subset \textit{L}
of \textit{L(T)}, the subtree induced by \textit{L}, noted \textit{T|$_{\text{L}}$}
is the tree defined as the smallest connected subgraph of \textit{T}
containing \textit{L}.
\end{comment}
\subsection{Partial order and Incompatible triples}
We define a partial order ($<$) on the edges and vertices (collectively called \emph{elements})
of a forest $F$, derived from a phylogenetic tree $T$.
For two distinct elements $x$ and $y$ that belong to the same component of $F$, we have $x < y$ if
$y$ lies on the path from $x$ to the root of this component.
\begin{figure}[b!]
\centering
\includegraphics[scale=0.3]{fig4.pdf}
\caption{\emph{A triple ab$|$c}}
\label{fig:triple}
\end{figure}
A triple (see Figure~\ref{fig:triple}) is a rooted binary phylogenetic tree that has 3 leaves.
A triple with leaf set $\{a,b,c\}$ is denoted by $ab|c$
if the path from $c$ to the root and the path from $a$ to $b$ are vertex-disjoint. \\
Let $\{a,b,c\}$ be a common leaf-set of both $T$ and $T'$.
The triple $ab|c$ is an incompatible triple of $T$ with respect to $T'$ if $ab|c$ is a
triple of $T$ only. \\
The partial order defined on elements of $F$ can be extended
to its incompatible triples. Let $ab|c$ be an incompatible triple
of $T$. If $r_{abc}$ represent the most recent common ancestor of $a$ and $c$ in $T$
and $r_{ab}$ the most recent common ancestor of $a$ and $b$ in $T$, we say $ab|c < xy|z$ if:
\begin{itemize}
\item either $r_{xyz}$ lies on the path from $r_{abc}$ to the root of $T$
\item or if $r_{abc}$ and $r_{xyz}$ are equal, $r_{xy}$ is on the path from $r_{ab}$ to the root of $T$
\end{itemize}
An incompatible triple is \emph{minimal} if it is minimal with respect to this partial order.
\subsection{Inseparable Components}
Let $F$ be the forest obtained from the two rooted binary phylogenetic
trees $T$ and $T'$ after all the incompatible triples have been taken care of.
If two components $t_x$ and $t_y$ of $F$ share a common element in $T'$,
then $t_x$ and $t_y$ are said to be \textit{inseparable} with respect to $T'$.
\section{Approximation Algorithms}
Below, by phylogenetic trees~we shall mean rooted phylogenetic trees.
In this section, we discuss algorithms for approximating a MAF as well as a MAAF for $k$ phylogenetic trees.
We refer to the 3-approximation algorithm for computing an approximate MAF for 2 phylogenetic trees~by Bordewich et al \cite{key-4}
as Bordewich's algorithm. In the next subsection, we briefly review this algorithm and show how it can be simply
extended to $k$ phylogenetic trees. In the following subsection, we show how to strengthen one of the key results in \cite{key-4}
to obtain a 2-approximation algorithm for $k~(\geq 2)$ phylogenetic trees, a substantial improvement over the 8-approximation
algorithm of Chataigner \cite{key-5}. In the next subsection, we discuss a 2-approximation
algorithm for computing an approximate MAAF for $k~(\geq 2)$ phylogenetic trees.
\subsection{Extending Bordewich et al's 3-approximation algorithm to $k$ trees}
The following lemma, proved in [4], plays a central role. Consistent with the phylogeny literature,
we use the symbols $+$ and $-$ for set-union and set-difference respectively.
\begin{lemma}
$F$ is a forest of an $X$-tree $T$, while
$e$ and $f$ are two edges in the same component of F
such that $f \in E$ and $e \notin E$, where $E$ is a subset of edges of $F$. Let $v_f$
be the end-vertex of $f$ closest to $e$, and $v_e$ an end-vertex
of $e$. If \\
(i) $v_f \sim v_e$ in F - E and \\
ii) $v_f \nsim x$ in $F-(E+e)$ for all $x$ in the leaf-set $X$, \\
then $F-E$ and $F-(E-f+e)$ yield isomorphic forests.
\end{lemma}
The essential conclusion of the above theorem is that the edges
$e$ and $f$ are connected by a linear path (see Figure~\ref{fig:central}) in $F - E$. Thus we get
isomorphic forests by substituting $e$ for $f$ in $E$. \\
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{fig5.pdf}
\caption{\emph{Pictorial illustration of Lemma 1}}
\label{fig:central}
\end{figure}
The approximation algorithm we present is a simple extension
of the 3-approximation algorithm in [4] for $k=2$.
We initialize the agreement forest $F$ to $T_1$. Next, we determine in turn the
incompatible triples of each $T_i$, $i = 2, 3, \ldots, k$ with respect to $F$ and deletes edges
from $F$ to eliminate the incompatibilities. The incompatible triples are processed with respect to their
partial order, picking a minimal one from those that remain. \\
For a minimum incompatible triple $ab|c$ in $F$ with respect to any $T_i~(i \geq 2),$
the edges that are candidates for removal are determined as follows (see Figure~\ref{pic:fig5}; this figure is
based on a similar figure in [4]).
Let $r_{abc}$ be the most recent common ancestor of $a$ and $c$ in
$T$ and $r_{ab}$ the most recent common ancestor of $a$ and $b$ in $T$. The child
edge of $r_{ab}$ leading to $a$ is denoted by $e_a$ and the child edge of $r_{ab}$
leading to $b$ is denoted by $e_{b}$. We label $e_{r}$ the child edge of $r_{abc}$
leading to $r_{ab}$. Finally, we label by $e_{c}$ the first edge on the path from $r_{abc}$ to $c$
such that for elements $c'$ in the leaf-set of $T-c$ below $e_c$,
there exists triples of the form $cc'|a$ and $cc'|b$ in both $T_i~(i \geq 2)$ and this component of $F$. \\
After all incompatibilities are resolved, the algorithm determines inseparable components $t_x$ and $t_y$
of $F$ vis-a-vis the $T_i$'s ($i \geq 2$) and deletes appropriate edges to eliminate the overlaps.
For a pair of inseparable components $t_x$ and $t_y$ in any $T_i$, $i = 2, 3, \ldots, k$ with respect to
$F$, let $v_{xy}$ denote a minimal common vertex of $t_x + t_y$ in $T_i$
with respect to the partial order on the vertices in $T_i$.
Further, $e_x$ denotes the minimal edge in $F$ whose descendants in the leaf-set are also descendants of
$v_{xy}$ in $t_x$. Similarly, $e_y$ denotes the minimal edge in $F$ whose set of
descendants in the leaf-set are also the descendants of $v_{xy}$
in $t_y$ (see Figure~\ref{pic:fig6}).
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.4]{fig6.pdf}
\caption{\emph{Minimum Incompatible triple} $ab|c$}
\label{pic:fig6}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[scale=0.4]{fig7.pdf}
\caption{\emph{Components $t_x$ and $t_y$ of $ F (= T_1)$ overlap in $T_i$, for some $i \geq 2$}}
\label{pic:fig7}
\end{figure}
\vspace{0.5cm}
\hrule
\vspace{0.5cm}
{\bf Algorithm} \emph{MAF-Approx($T_1,T_2,...,T_k$)} \\
1. $F \leftarrow T_1$;
2. for $i = 2$ to $k$ do
\qquad{}2.1. while there exists an incompatible triple in $F$ with respect to $T_i$ do
\qquad{}\qquad{}2.1.1. consider the minimal incompatible triple
$ab|c$ in $F$ with respect to $T_i$
\qquad{}\qquad{}2.1.2. $E \leftarrow \{e_a, e_c, e_r\}$ in $ab|c$
\qquad{}\qquad{}2.1.3. $F \leftarrow F - E$
\enskip{}\enskip{}\qquad{}enddo;
\enskip{}\enskip{}enddo;
3. for $i = 2$ to $k$ do
\qquad{}3.1. while there exists a pair of inseparable components in any $T_i$
($i \geq 2$) with respect to $F$ do
\qquad{}\qquad{}3.1.1. consider inseparable components $t_x$
and $t_y$ in $T_i$ with respect to $F$
\qquad{}\qquad{}3.1.2. $E \leftarrow \{e_x, e_y \}$ in $t_x$ and $t_y$
\qquad{}\qquad{}3.1.3. $F \leftarrow F - E$
\enskip{}\enskip{}\qquad{}enddo;
\enskip{}\enskip{}enddo;
4. return \textit{F};
\vspace{0.5cm}
\hrule
\vspace{0.5cm}
\begin{lemma}
Let $T_1, T_2, \ldots, T_k$ be $k$ rooted binary phylogenetic
trees and F a forest of $T_1$. \\
(i) If $ab|c$ is a minimal incompatible triple of F
with respect to some of the $T_i$'s for ($i \geq 2$), then
\begin{center}
$e(F-\{e^i_{a}, e^i_{c}, e^i_{r}\}, T_2,T_3,...,T_i,...,T_k) \leq e(F, T_2, T_3,...,T_i,...,T_k) - 1$,
\end{center}
where $e(F, T_2, T_3,...,T_k)$ denotes the size of a minimum
set $E$ of edges of $F$ such that $F - E$ forms an agreement forest of $k$
trees and $e^i_{a}, e^i_{c}, e^i_{r}$ are the edges deleted from $F$ to resolve the incompatibility
due to the triple $ab|c$. \\
(ii) If there is no incompatible triple of $F$ with respect to
any other tree, but there exist two components $t_x$ and
$t_y$ of $F$ that overlap in some tree $T_i$ ($i \geq 2$), then
for some $j \in \{x, y\}$
\begin{center}
$e(F - e^i_{j}, T_2, T_3,...,T_i,...,T_k) = e(F, T_2,T_3,...,T_i,...,T_k)$ - 1.
\end{center}
\end{lemma}
{\bf Proof:} The proof in [4] for the case when $k = 2$ goes through with some minor changes. Let's
see this for (i). \\
We have $|E| = e(F, T_2, T_3, \ldots, T_k)$, where $E$ is a minimum set of
edges such that $F-E$ yields a maximum agreement forest of $F$ and all $T_i$, $i \geq 2$. It has been shown in [4]
that there exists an $f \in E$ such that $F - (E - f + \{e_a, e_c, e_r \})$ is isomorphic to a subforest of $F - E$.
This implies that $F - (E - f + \{e_a, e_c, e_r \})$ yields an agreement forest of $F - \{e_a, e_c, e_r \}$ and all $T_i,~i \geq 2$.
So, $e(F - \{e_a, e_c, e_r \}) \leq |E-f| = e(F, T_2, T_3, \ldots, T_k) - 1$. \\
We can likewise extend the proof of (ii) for the case when $k = 2$ to this case. $\hfill \Box$ \\
Assume there are $\alpha_1$ iterations of the 1st while loop and $\alpha_2$ iterations
of the 2nd while loop in processing all $k-1$ trees, $T_2, T_3, \ldots, T_k$.
Set $\alpha = \alpha_1 + \alpha_2$. Then we have the following claim.
\begin{claim}
\textit{$\alpha$ $\leq$ e(T$_{\text{1}}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ 3$\alpha$}.
\end{claim}
{\bf Proof:} Let $F_i$ be the forest obtained after $i$ iterations of the above algorithm. The following cases arise. \\
{\bf Case 1:} $i \leq \alpha_1 $ (still in 1st while loop)
From Lemma 2, we have:
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i-1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
}- 1
\end{center}
Hence by successive applications of the above inequality, we have:
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ i $\leq$ e(F$_{0}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)}
\end{center}
Since $F_0 = T_1$, we can rewrite the above inequality as:
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ i $\leq$ e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)}
\end{center}
Moreover, since \textit{F$_{i}$} has 3 fewer edges than \textit{F$_{i-1}$} we have:
\begin{center}
\textit{e(F$_{i-1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
}+ 3
\end{center}
Applying the above $i$ times, we get
\begin{center}
\textit{e(F$_{0}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ 3i}
\end{center}
and since $F_0 = T_1$,
\begin{center}
\textit{e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ 3i}
\end{center}
{\bf Case 2:} \textit{i} $>$$\alpha$$_{\text{1}}$, (in 2nd while loop)
Again, from the second part of Lemma 2,
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i-1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)}
- 1
\end{center}
Thus by $i$ applications of the above inequality we have:
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ i $\leq$ e(F$_{0}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$) }
\end{center}
Since $F_0 = T_1$, we can rewrite the above inequality as:
\begin{center}
\textit{e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ i $\leq$ e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)}
\end{center}
Again, since \textit{F$_{i}$} has 2 fewer edges than \textit{F$_{i-1}$} we have:
\begin{center}
\textit{e(F$_{i-1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
}+ 2
\end{center}
Thus,
\begin{center}
\textit{e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$ e(F$_{i}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ 3$\alpha$$_{\text{1}}$ + 2(i - $\alpha$$_{\text{1}}$)}
\end{center}
Thus on termination of both loops, we have:
\begin{center}
\textit{e(F$_\alpha$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
+ $\alpha$$_{\text{1}}$ + $\alpha$$_{\text{2}}$ $\leq$ e(T$_{1}$,
T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$) $\leq$ e(F$_\alpha$,
T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$) + 3$\alpha$$_{\text{1}}$
+ 2$\alpha$$_{\text{2}}$}
\end{center}
As no further edge-cut is necessary when an agreement forest is generated,
\textit{e(F$_\alpha$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)}
= 0. Thus:
\begin{center}
\textit{$\alpha$$_{\text{1}}$ + $\alpha$$_{\text{2}}$ $\leq$
e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$) $\leq$
3$\alpha$$_{\text{1}}$ + 2$\alpha$$_{\text{2}}$}
\end{center}
A fortiori, we have:
\begin{center}
\textit{$\alpha$$\leq$ e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$3$\alpha$$_{\text{1}}$ + 3$\alpha$$_{\text{2}}$},
\end{center}
which simplifies to:
\begin{center}
\textit{$\alpha$$\leq$ e(T$_{1}$,T$_{\text{2}}$,T$_{\text{3}}$,...,T$_{\text{k}}$)
$\leq$3$\alpha$}
\end{center}
This proves the claim. \hfill $\Box$ \\
Since the number of edges removed is
$3\alpha_1 + 2 \alpha_2 \leq 3 \alpha \leq 3 e(T_1,T_2,T_3,...,T_k)$,
our algorithm has an approximation ratio of 3. Summarizing the above results, we have:
\begin{theorem}
Algorithm MAF-approx has an approximation ratio of 3 and time-complexity in $O(kn^5)$.
\end{theorem}
\begin{comment}
\subsection{A 2-approximation algorithm for computing an approximate MAF}
Bordewich et al's \cite{key-4} 3-approximation algorithm is based on the following key lemma:
\begin{lemma}[\cite{key-4}]
Let $ab|c$ be a minimal incompatible triple of $F$ with respect to $T'$. Then there exists an edge $f \in E$
such that $F- (E-f+ \{e_a, e_c, e_r \})$ is isomorphic to a subforest of $F-E$.
\end{lemma}
The essential idea underlying the above lemma is that the incompatibility due to a triple, $ab|c$, can be resolved
by removing the edges $e_a, e_c$ and $e_r$ (see Figure~\ref{fig6}).
The main contribution of our paper is to show it suffices to remove only the edges $e_c$ and $e_r$, as
the following lemma shows.
\begin{lemma}
Let $ab|c$ be a minimal incompatible triple of $F$ with respect to $T'$ (see Figure~\ref{fig6}). Then there exists an edge $f \in E$
such that $F- (E-f+ \{e_c, e_r \})$ is isomorphic to a subforest of $F-E$.
\end{lemma}
{\bf Proof:}
Throughout this proof, we refer to Figure~\ref{fig6}.
Assume that for all $c' \in C$, we have $c' \nsim r_c$ in $F - E$. If
$f$ is the first edge in $E$ on the path from
$r_c$ to $c$ in $F$, then by Lemma 1, $F - E$ is isomorphic to $F - (E-f+e_c)$.
Thus the lemma holds in this case. \\
Otherwise, there exists a $c'$ such that $c' \sim r_c$ in $F-E$.
Under the assumption that $ab|c$ is a minimum incompatible triple of $F$, it has been shown in \cite{key-4} that for all $y \in B + D_1$ and $a' \in A$, $a'y|c'$ is an incompatible triple.
Thus there exists at least one edge in $E$ whose deletion from $F$ removes this incompatibility.
Of these edges, let $f$ be nearest to $r_c'$. \\
Let us assume that for some $y \in B + D_1$ the
relation $y \thicksim r_{ab} \sim a'$ in $F-E$ holds ($a' \in A$).
Under this assumption, it was shown in \cite{key-4} that $c' \nsim d'$ in $F-E$ for all
$c' \in C$ and $d' \in D_1 + D_2$. \\
If $f$ is between $r_{ab}$ and $r_{abc}$,
by Lemma 1, $F-E$ is isomorphic to $F - (E-f+e_r)$; if $f$ is between $r_c'$ and $r_{abc}$,
$F-E$ is isomorphic to $F - (E-f+e_c)$. In either case, the claim of the lemma holds. \\
Now, let us suppose that there is no $y \in B + D_1$
such that $y \thicksim r_{ab}$ in $F-E$.
Then in particular, $b' \nsim r_{ab}$ for all $b' \in B$.
Assume $f$ to be the first edge
in $E$ on the path from $r_{ab}$ to $b$ in $F$. \\
We show that $F-(E-f+\{e_c, e_r\})$ is isomorphic to a subforest of $F-E$ in this case.
This isomorphism implies that for all $i,j \in X$ if $i \sim j$ in $F-(E-f+\{e_c, e_r\})$ then
$i \sim j$ in $F-E$. \\
We prove this by contradiction.
Assume there is a path from $i$ to $j$ in $F-(E-f+\{e_{\textit{c}}, e_{\textit{r}}\})$ but not in $F-E$. \\
Let $i \in B$. We have the following cases to consider with respect to the location of $j$.
\begin{itemize}
\item $j \in A$ ? \\
It is possible for $j$ to be in $A$.
\item $j \in B$ ? \\
By Lemma 1, $F-E$ is isomorphic to $F-(E-f+e_b)$. Since there is no path from $i$ to $j$ in $F-E$,
there is no path from $i$ to $j$ in $F-(E-f+e_b)$. Now, by assumption there is a path from $i$ to $j$
in $F-(E-f+\{e_{\textit{c}}, e_{\textit{r}}\})$; the edges used to traverse this path are also present
in $F-(E-f+e_b)$, except possibly for $e_b$, but only the end-point other than $r_{ab}$ can lie on this path.
The only way to resolve this contradiction is to assume that $j \notin B$.
\item $j \in D_1$ ? \\
Since $e_r$ is not in the path from $i$ to $j$, this would imply that
there exists $y$ (=$j$) such that $y \thicksim r_{ab}$, contradicting our initial asssumption.
\end{itemize}
This completes the proof. $\hfill \Box$ \\
As a consequence, we have a modified version of Lemma 2.
\begin{lemma}
Let $T_1, T_2, \ldots, T_k$ be $k$ rooted binary phylogenetic
trees and F a forest of $T_1$. \\
(i) If $ab|c$ is a minimal incompatible triple of F
with respect to some $T_i$, then
\begin{center}
$e(F-\{e^i_{c}, e^i_{r}\}, T_2,T_3,...,T_i,...,T_k) \leq e(F, T_2, T_3,...,T_i,...,T_k) - 1$,
\end{center}
where $e(F, T_2, T_3,...,T_k)$ denotes the size of a minimum
set $E$ of edges of $F$ such that $F - E$ forms an agreement forest of $k$
trees and $e^i_{a}, e^i_{c}, e^i_{r}$ are the edges deleted from $F$ to resolve the incompatibility
due to the triple $ab|c$. \\
(ii) If there is no incompatible triple of $F$ with respect to
any other tree, but there exist two components $t_x$ and
$t_y$ of $F$ that overlap in some tree $T_i$ ($i \geq 2$), then
for some $j \in \{x, y\}$
\begin{center}
$e(F - e^i_{j}, T_2, T_3,...,T_i,...,T_k) = e(F, T_2,T_3,...,T_i,...,T_k)$ - 1.
\end{center}
\end{lemma}
\hfill $\Box$
With the help of this, by arguing as in the previous section, we get a 2-approximation algorithm for
computing an approximate MAF for 2 trees, which easily extends to $k$ phylogenetic trees. Summarizing the above
discussions, we have the following theorem.
\begin{theorem}
The modified version of MAF-Approx has an approximation ratio of 2 and its time-complexity is in $O(kn^5)$
\end{theorem}
\end{comment}
\section{A 3-approximation algorithm for computing an approximate MAAF}
The roots of two components (trees) in a MAF produced by the algorithm of the previous section may have
an ancestor-descendant relationship in one tree, and the opposite in another, as in Figure~\ref{fig8}.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{fig8.pdf}
\caption{\emph{Example of a Maximum Agreement Forest of trees $T_1$ and $T_2$}}
\label{fig8}
\end{figure}
If there are no cycles in the approximate MAF, $F$, produced by the algorithm \emph{MAF-Approx}
of the previous section then we have an approximate MAAF, $F_A$, as well.
Otherwise, as in
Whidden and Zeh [13], we incorporate a preprocesing step to detect cycles that exist between roots of the
trees in the approximate MAF and eliminate these cycles. This consists of assigning to each node of $T_i, i=1,2, \ldots,k$,
a preorder visit number and the interval of preorder numbers of its descendants.
We maintain the roots of all the trees in the MAF in 2 sets: $R_p$ and $R_{up}$. The former consists of the roots of all
trees with no cycle between any pair; the roots of the latter are yet to be processed. We choose a root from $R_{up}$ and determine
if it has a cycle with any root in $R_{p}$. This is done by mapping these roots to the corresponding nodes of a pair of trees
$T_i$ and $T_j$.
The preorder intervals associated with these nodes can be used to check the existence of a cycle between these roots (for more details see
\cite{key-13}).
If a cycle does not exist, we add it to the
set $R_{p}$. Otherwise, let $t_i$ and $t_j$ be 2 trees in $F$ whose respective roots $r$ and $r'$ form a cycle.
We call such a tree-pair \emph{infeasible}.
To obtain \textit{F$_{\text{A}}$}, we delete one of the edges $e_r$ incident on $r$, as well as one of the edges $e_{r'}$
incident on $r'$ to remove this cycle. We continue doing this till the set $R_{up}$ becomes empty.
The following lemma underlies the above choice of the edges we cut.
\begin{lemma}
If $e(F, T_2, \dots, T_k)$ is the minimum number of edges that must be removed from $F$ to obtain a MAAF, then
$e(F - \{e_x\}, T_2, \ldots, T_k) = e(F, T_2, \dots, T_k) - 1$, where $x \in \{r,r'\}$; moreover,
$e(F - \{e_r, e_{r'}\}, T_2, \ldots, T_k) \leq e(F, T_2, \dots, T_k) - 1$
\end{lemma}
\textbf{Proof:} We prove that there exists a set $E$ of $e(F, T_2, \dots, T_k)$ edges of $F$ such that $F-E$ yields a
MAAF of $T_1, T_2, \ldots, T_k$ and $E \cap \{e_r, e_{r'}\} \neq \emptyset$. Otherwise, let $E$ be a set of edges
that produces a MAAF but $E \cap \{e_r, e_{r'}\} = \emptyset$. \\
Let $t_i(r_1)$ and $t_i(r_2)$ be the left and right subtrees of $t_i(r)$, and $t_j(r'_{1})$ and $t_j(r'_{2})$ the
left and right subtrees of $t_j(r')$.
There exists $i$ such that $a' \nsim_{F-E} r$ for all $a' \in X_{T(r_i)}$ or
$b' \nsim_{F-E} r'$ for all $b' \in X_{T(r_i)}$. \\
Otherwise, there exists $a_1 \in t_i(r_1)$ and $a_2 \in t_i(r_2)$ such that $a_1 \sim a_2$ and
$b_1 \in t_j(r'_1)$ and $b_2 \in t_j(r'_2)$ such that $b_1 \sim b_2$. This implies that the
cycle involving the roots $r$ and $r'$ (of $t_i$ and $t_j$ respectively) have not been removed. \\
Assume that $a' \nsim_{F-E} r$ for all $a' \in T(r_1)$. Choosing an edge $f$ closest to $r$ on the
path from $r$ to a leaf $a'$, by Lemma 1, the forests $F-E$ and $F-(E-f + e_r)$ are isomorphic. The claims
of the lemma follow from this.
$\hfill \Box$ \\
A formal description of the above algorithm is given below.
\vspace{0.5cm}
\hrule
\vspace{0.5cm}
{\bf Algorithm} \emph{MAAF-Approx($F$)} \\
// $F = \{t_1, t_2, t_3, \ldots, t_m \}$
1. Set $R_{up} \leftarrow \{ root(t_1), root(t_2), \ldots, root(t_m) \}$
2. Set $R_{p} \leftarrow \emptyset $
3. while \{$R_{up} \neq \emptyset \}$
\hspace{0.3cm} do
\hspace{0.3cm} 3.1 Pick an $r$ from $R_{up}$
\hspace{0.3cm} 3.2 If ($r$ forms a cycle with an $r'$ in $R_{p}$) then
\hspace{0.7cm} 3.2.1 Delete an edge $e_r$ incident on $r$ and an edge $e_{r'}$ incident on $r'$
\hspace{0.7cm} 3.2.2 Add roots of the subtrees of $r$ and $r'$ to $R_{up}$
\hspace{0.7cm} 3.2.3 Continue
\hspace{0.3cm} 3.3 else $R_{p} \leftarrow R_{p} + r$
\hspace{0.3cm} od
4. Return the trees whose roots are in $R_p$
\vspace{0.5cm}
\hrule
\vspace{0.5cm}
Assume there are $\beta$ iterations of \emph{MAF-Approx} and \emph{MAAF-Approx} in each of which at most 3 edges are removed.
We claim that:
\begin{claim}
$\beta \leq e(T_1, T_2, \ldots, T_k) \leq 3\beta $
\end{claim}
{\bf Proof:}
Let $F_\theta$ be the forest generated after $\theta$ iterations.
By Lemmas 2 and 3, $e(F_\theta, T_2, T_3, \ldots, T_k)) \leq e(F_{\theta-1}, T_2, T_3, \ldots, T_k) - 1$.
Hence after $\beta$ iterations, $e(F_{\beta}, T_2, T_3, \ldots, T_k) + \beta \leq e(T_1, T_2, T_3, \ldots, T_k)$
(as $e(F_0, T_2, T_3, \ldots, T_k) = e(T_1, T_2, T_3, \ldots, T_k)$). \\
Conversely, the algorithms $MAF-Approx(T_1, T_2,...,T_k)$ as well as $MAAF-Approx(F)$ account for at most 3
edge-cuts in each iteration. Hence, $e(F_{\theta-1}, T_2, T_3, \ldots, T_k) \leq e(F_\theta, T_2, T_3, \ldots, T_k) + 3$.
This implies $e(T_1, T_2, T_3, \ldots, T_k) \leq e(F_{\beta}, T_2, T_3, \ldots, T_k)+ 3\beta$. \\
Now, after $\beta$ iterations an Acyclic-MAF is generated and we
do not require any further edge-cuts. So, $e(F_{\beta}, T_2, T_3, \ldots, T_k) = 0$. \\
This proves that $\beta \leq e(T_1, T_2, T_3, \ldots, T_k) \leq 3\beta$ and consequently that our algorithm
has approximation ratio 3.
\hfill $\Box$ \\
Summarizing the above discussions, we have:
\begin{theorem}
Algorithm MAAF-approx has an approximation ratio of 2 and time complexity in $O(n^2k^2)$.
\end{theorem}
\section{Conclusion}
In this paper we have proposed approximation algorithms for finding
the Maximum Agreement Forest and the Maximum Acyclic Agreement Forest
on $k$ rooted phylogenetic trees. It is straightforward to extend the fixed-parameter
tractable algorithm for an exact MAF of 2 trees \cite{key-4} to $k (\geq 2)$ trees. \\
Extending these algorithms to $k$ unrooted trees
would be interesting, as would be to extend the results to trees of degree $d$ ($d \geq 2$).
|
\section{Introduction}
\label{sc.intro}
In relativistic heavy-ion collisions the fluctuations and correlations
of conserved charges like the baryon number, electric charge
{\it etc.}\ are considered to carry promising signals for the formation
of the exotic Quark Gluon Plasma (QGP). Close to any continuous phase
transition region these are supposed to exhibit critical behavior
\cite{asakawa,jeon_koch,Stephanov_prl,hatta_prd}. Therefore the
characteristics of quark-hadron phase transition can be understood by
analyzing the fluctuations of the system. The fluctuations are often
calculated theoretically through the respective susceptibilities.
For comparison with experimental data various combinations of the
ratios of these susceptibilities constitute important phenomenological
observables \cite{Ejiri,GavGup_11,SG_science}.
Here we shall discuss the quark number susceptibility (QNS) which
provides the response of the net quark number density to the change
in quark chemical potential. Several first principle studies have been
done to calculate the QNS in various approaches. These include the
numerical simulation of QCD on a space-time lattice (LQCD)
\cite{Gottieb_prl,Gottieb_prd,Alton,Gavai_prd1,Gavai_prd2,
HOTQCD_08,Cheng_09,WB_12,HOTQCD_12} as well as Hard Thermal
Loop (HTL) calculations \cite{Blaizot,purnendu1,purnendu2,jiang,
HTLPT_Najmul,najmul,HMS1_2013,HMS2_2013,HAMSS_2014,HBAMSS_2014}.
In the present manuscript we intend to revisit the calculations of QNS
through the phenomenological models namely the Nambu$-$Jona-Lasinio (NJL)
model and its Polyakov Loop extended version, the PNJL model. We begin
with a brief introduction for the various studies already carried out
within the various QCD inspired models to understand the thermodynamic
properties of strongly interacting matter.
QCD phase transitions for vanishing and non-vanishing baryon chemical
potential have been studied in a great detail and the possible phases
that may arise in the phase diagram have been addressed
\cite{Fukushima_PLB,BCPR_05,ratti1, rajarshidaA,rajarshidaB,
ratti_diquark,pnjl_pd_SFR,rajarshidaC,Fukushima_PRD,Kahara_2peak,
pnjl_FZL,njl_pd_CRS,CEP_pnjl_KKMY,pnjl_pd_review,pnjl_para_CHRS,
pnjl_QHPT,Sarbani,njl_pd_DR,pnjl_para_IJMP}. Similar studies were
also carried out for imaginary chemical potential
\cite{pnjl_imgnmu_SKKY,pnjl_imgnmu_SKY,pnjl_imgnmu_MSFR} and the
well-known Roberge-Weiss periodicity is discussed in that context.
Although hadrons are not present as dynamical degrees of freedom
in NJL or PNJL model, mesonic modes at real as well as imaginary
chemical potential are studied as collective excitations within Random
Phase Approximation \cite{DA_95,buballa_meson,ratti_meson,paramita_meson,
pnjl_meson_reim2,pnjl_meson_Costa,njl_meson_scrn,pnjl_meson_reim1}
and formation of baryons composed of quarks and diquarks has also been
studied by solving the Dyson-Schwinger equation \cite{pnjl_particle,
pnjl_baryon}. Mesons involving heavy quarks have been studied recently
in NJL and PNJL models \cite{Dmeson_pnjl}. The average phase factor of
QCD determinant is evaluated through PNJL model in
Ref.\cite{avg_phase_pnjl}, where it is argued that since CEP lies
within the region of vanishing phase factor, location of CEP cannot
be determined by LQCD alone. NJL model is explored in the context of
CP restoring phase transition \cite{CP_pnjl}, where it is shown that
nontrivial vacuum term of NJL model can always alter the qualitative
aspects of the high temperature phase transition. The issue of color
neutrality is crucially investigated in Ref.\cite{color_pnjl1,
color_pnjl2}. Interplay between chiral and deconfinement transition
is also investigated with U(1) valued boundary condition for
fermionic fields \cite{dqc_pnjl,dqc_njl_TKM,dqc_njl_Xu}, where
quantities namely dual quark condensate or dressed Polyakov loop
seem to be very effective in those studies. Effects of theta vacuum
on QCD phase structure is investigated using PNJL model in
Ref.\cite{thetavac_pnjl}. NJL and PNJL model are also studied within
background magnetic field \cite{magfld_pnjl_BB,magfld_pnjl_GR,
magfld_pnjl_K,magfld_pnjl_AS}. Another interesting phenomenon namely
chiral magnetic effect, which could possibly explain the observed
charge separation observed in the STAR experiment at RHIC, Brookhaven,
is also investigated through PNJL model \cite{CME1_pnjl,CME2_pnjl}.
Existence of conjectured chirally symmetric but confined phase in QCD
phase diagram which is popularly named as quarkyonic phase is
discussed in terms of PNJL model\cite{AAGNR_2008,qrkync_pnjl_MLRN,
qrkync_pnjl_SKSY,qrkync_pnjl_2013}. Role of axial anomaly and
vector interaction determining the phase diagram of QCD
is studied in Ref.\cite{vecint_pnjl_CNB,axanml_pnjl_RSCS,
axanml_pnjl_PB,vecint_pnjl_estm,vecint_pnjl_BHW}. For three degenerate
or non-degenerate flavored system, it can be shown that NJL vacuum is
unstable unless one incorporates eight-quark interaction. Modification
has been done in this direction for both NJL and PNJL model
\cite{njl_8q_OHP,njl_8q_OHBP,njl_8q_OHMBP,njl_8qpd,paramita_8qPD}.
In general NJL model in its local version is applicable within
a restricted momentum range which is governed by the cutoff parameter,
that appears while regularizing the divergent momentum integrals.
To overcome this limitation a non-local version of the model was
proposed and developed recently \cite{nonlocal_pnjl_SFR,
nonlocal_pnjl_HRCW,nonlocal_pnjl_HKW,nonlocal_pnjl_KHW,
nonlocal_pnjl_PDS}. Furthermore finite volume effect which is
relevant for studying a system created in heavy-ion collisions,
has been studied very recently in Ref.\cite{finitevol_pnjl}.
Various interesting features of Polyakov loop \cite{MAS3_2006,
MAS_12,MAS_14,aminul1} have encouraged people to study within
different formalisms. Inclusion of gluon Polyakov loop is studied in
various aspects \cite{MAS1_2006,MAS2_2006,Muller}.
Interestingly NJL model has also been studied
within Monte-Carlo framework also \cite{MCNJL1,MCNJL2}. Studies of
various transport coefficients in NJL model framework have also
been reported recently\cite{njl_Marty,etaNJL_Sabyasachi}.
The quark number susceptibility has been studied extensively within
the framework of NJL and PNJL models \cite{rajarshidaA, rajarshidaC,
FLW_2010,sasaki_redlich, ratti_plb, anirban1,fluc_WF,anirban2}.
These studies revealed the order parameter like behavior
of QNS, similar to that obtained in LQCD at vanishing baryon
chemical potential. Recently it has been shown by some of us
\cite{isobreak_pnjl} that when isospin symmetry is broken explicitly,
the baryon-isospin correlations exhibit an almost linear scaling with
the scale of isospin breaking over the entire temperature-baryon chemical
potential phase plane.
Another well-known formulation to study the strongly interacting matter
in nonperturbative regime is the Quark Meson (QM) model and its Polyakov
loop extended version (PQM), which is used to explore phase transition
and phase diagram of QCD \cite{pqm_pd_SPW1,pqm_pd_SPW2,
crit_pqm_SFR,thermodynamics_pqm_SSFR,pqm_pd_HPS, sandeep_vacuum}
as well as quark number susceptibility \cite{Schaefer_2007,kurtosis_pqm,
Schaefer_2010,skokov,qns_mu_pqm,Sandeep,charge_fluc_pqm}.
The QNS is the response of the conserved number density, $\rho$,
with infinitesimal variation in the chemical potential $\mu +\delta\mu$
as an external source. It is then defined as the second order derivative
of pressure, $\cal{P}$, with respect to $\mu$. On the other hand,
according to the fluctuation-dissipation theorem (FDT), the QNS may
also be obtained from the time-time component of the current-current
correlator in the vector channel \cite{purnendu2,
calen,Kunihiro_PLB,Hatsuda_NJL,Fujii04}.
The QNS is then expressed as
\begin{eqnarray}
\chi_q &=& \frac{\partial\rho}{\partial\mu}
=\frac{\partial^2{\mathcal{P}}}{\partial\mu^2}
=\int d^4x \langle J_0(0,\vec{x})J_0(0,\vec{0})\rangle
= -\lim_{l\rightarrow 0} {\rm{Re}} \Pi_{00}(0,l)
= \lim_{l\rightarrow 0} \beta
\int_{-\infty}^{+\infty} \frac{d \omega}{2 \pi}
\frac{-2}{1-e^{-\beta\omega}} {\rm{Im}} \Pi_{00}(\omega,l),
\label{eq.fluc_diss_def}
\end{eqnarray}
where $J_{0}$ is the temporal component of the vector current
and $\Pi_{00}(\omega,l)$ is the time-time component of the vector
correlation function or self-energy with external four momenta
$L\equiv (\omega,l=|\vec l|)$. Because of the symmetry of the system
the FDT guarantees that the
thermodynamic derivative with respect to the external source, $\mu$
is related to the time-time component of static correlation function
in the vector channel. This is known as {\it thermodynamic sum rule}
\cite{purnendu2,calen,Kunihiro_PLB}.
In usual perturbation theory the loop expansion and coupling expansion are
symmetric and the thermodynamic consistency is automatic, for a given
order of coupling $\alpha_s$. So, it really does not matter which of
the equivalent definition is used in (\ref{eq.fluc_diss_def})
to compute QNS for a given order of $\alpha_s$. For resummed approach
like Hard Thermal Loop perturbation theory, the loop expansion and
coupling expansion are not symmetric because higher loops contribute
to the lower order in $\alpha_s$. Unlike usual perturbation theory,
in resummed case an appropriate measure is to be employed
\cite{purnendu2,najmul,HAMSS_2014,HBAMSS_2014} if one desires to
compute QNS in a given order in $\alpha_s$ correctly using
(\ref{eq.fluc_diss_def}). In effective approaches like NJL
or PNJL models the QNS is usually obtained \cite{rajarshidaA, rajarshidaC,
FLW_2010,sasaki_redlich, ratti_plb, anirban1,fluc_WF,anirban2}
as the second order Taylor coefficient of pressure when it is Taylor
expanded in the direction of the quark chemical potential, $\mu$
with an approximation $\mu< T$. In model calculations any response of a
thermodynamic quantity to some external parameters should also
account for the fact that the mean fields also depend implicitly
on those external parameters \cite{rajarshidaA}. Therefore, a proper
care has to be taken to relate the thermodynamic derivatives
(viz., QNS) with the fluctuation associated with the conserved density.
One of the purposes of the present work is to demonstrate whether
the effective models like NJL and PNJL explicitly obey the FDT
vis-a-vis thermodynamic sum rule.
Furthermore, the implicit dependence of the mean fields are usually
obtained numerically but this may contain numerical errors which will
increase with the order of derivatives. This is true for either a
direct numerical derivative as well as the method of using a
fitting function for a polynomial expansion of pressure in terms
of chemical potential to corroborate with the Taylor expanded pressure.
Recently a numerical technique based on
algorithmic differentiation \cite{AD_WWS,AD_CEP,AD_SW} has been
developed to solve this shortcoming. In the present work we also propose
an alternative analytical formalism to calculate the derivatives of
the mean fields with respect to the external parameters. Using these
derivatives we are going to calculate the QNS in a consistent manner.
The paper is organized as follows. In the Sec.~\ref{sc.fldsder} we
discuss the formalism for obtaining the derivatives of the mean
fields in our alternative approach in the NJL and the PNJL models.
In Sec.~\ref{sc.qns_fdt} the calculations of QNS exploiting the FDT
for a toy model and in effective approaches like NJL as well as in
PNJL models are presented. Finally we draw our conclusions in
Sec.~\ref{sc.concl}.
\section{Calculating the derivatives of the mean fields}
\label{sc.fldsder}
\subsection{NJL model}
\label{sc.fldsder_njl}
The Lagrangian for the 2 flavor NJL model at finite quark chemical
potential ($\mu_q$) is given as
\cite{Hatsuda_NJL,Klevansky:rmp,buballa},
\begin{equation}
\mathcal{L}_{\rm NJL}=\bar{\psi}(i\slashed \partial
-m_0 +\gamma_0\mu_q)\psi + \frac{G}{2}
[(\bar{\psi}\psi)^2+(\bar{\psi}i\gamma_5\vec{\tau}\psi)^2].
\end{equation}
In the mean field approximation the pion condensate
$\langle\bar{\psi}i\gamma_5\vec{\tau}\psi\rangle=0$ and the
Lagrangian can be rewritten in terms of the chiral condensate
$\sigma=\langle\bar{\psi}\psi\rangle=
\langle\bar{u}u+\bar{d}d\rangle=\sigma_u+\sigma_d$ as,
\begin{equation}
\mathcal{L}_{\rm MF}=\bar{\psi}(i\slashed \partial
-m_0+\gamma_0\mu_q+G\sigma)\psi - \frac{G}{2}\sigma^2.
\label{eq.lag_mf}
\end{equation}
In the mean field approach, the thermodynamic potential is a functional
of the mean field $\sigma(m_0,T,\mu_q)$, and is given as,
\begin{equation}
\Omega[\sigma,m_0,T,\mu_q]
=-i\mathrm {Tr}[\ln {S_1}^{-1}]+\frac{G}{2}\sigma^2.
\label{eq.omega}
\end{equation}
Here and unless stated, \textquoteleft Tr\textquoteright\ denotes the
sum over color, flavor and Dirac indices as well as the four-momentum
and any other notation involving trace operation will be clarified
accordingly. The first term on the right hand side of (\ref{eq.omega})
is the fermionic contribution related to the dressed propagator $S_1$
where,
\begin{equation}
{S_1}^{-1}=\slashed p -m_0 +\gamma_0\mu_q+G\sigma={S_0}^{-1}+G\sigma,
\label{eq.mod_prop_inv}
\end{equation}
and $S_0$ is the bare propagator with current quark mass $m_0$.
The second term in $\Omega$ may be considered as the background
contribution of the mean field $\sigma$. To utilize thermodynamic
relations, derivatives of $\Omega$ with respect to the various
parameters and the mean fields are often necessary. In this
regard care has to be taken for the explicit appearances of
the parameters as well as their presence through the mean fields.
For example, the computation of quark number susceptibility
requires derivatives of $\Omega$ with respect to $\mu_q$ that
appear explicitly in $\Omega$ as well as their implicit effects
through $\sigma$ \cite{rajarshidaA}. This is an important observation
that we want to revisit in the present manuscript. Therefore we set the
notation for the explicit derivatives by ${\partial}/{\partial x}$ and
that for the total derivatives with ${d}/{dx}$ for some parameter $x$.
Let us start with the computation of the mean field $\sigma$. One way
to obtain $\sigma$ is to use the stationarity condition
$\partial \Omega/ \partial \sigma =0$ in the mean field approximation,
which gives,
\begin{equation}
\sigma=i\mathrm {Tr}(S_1).
\label{eq.sig_def}
\end{equation}
On the other hand, we may also use the defining equation
$\sigma=\partial\Omega/\partial m_0$, which gives the same result as in
Eq.(\ref{eq.sig_def}). However as discussed above, for the derivative with
respect to $m_0$ one should also consider the implicit dependence on $m_0$
of $\Omega$ through $\sigma$. In that case we should rather use the
relation,
\begin{equation}
\sigma=\frac{d\Omega}{dm_0}=\frac{\partial\Omega}{\partial m_0}+
\frac{\partial\Omega}{\partial\sigma}\cdot\frac{d\sigma}{dm_0}.
\label{eq.sigma_fundef}
\end{equation}
Using the stationarity condition for the mean field, the second term
vanishes and therefore in this case we get back to the original defining
equation of $\sigma$. Interestingly, if we straightway calculate
$\frac{d\Omega}{dm_0}$ from Eq.(\ref{eq.omega}) and demand it to be equal
to $\sigma$, without imposing the stationarity condition (like in
Eq.(\ref{eq.sigma_fundef})), then Eq.(\ref{eq.sig_def}) will emerge as a
consistency condition.
The transcendental nature of the solutions of (\ref{eq.sig_def})
is apparent. Thus a closed form analytical expression of $\sigma$ as
a function of $m_0$, $T$ and $\mu_q$ cannot be obtained from this equation.
One has to solve the equations numerically and obtain $\sigma(T,\mu_q)$.
The implicit derivatives, in general do not disappear, as can be seen
from the chiral susceptibility, which is the second order derivative
of thermodynamic potential with respect to $m_0$,
\begin{eqnarray}
\chi_{\sigma} = \frac{d^2 \Omega}{d m_0^2}
= \frac{\partial^2 \Omega}{\partial m_0^2} +
2\frac{\partial^2 \Omega}{\partial\sigma\partial m_0}
\cdot\frac{d\sigma}{d m_0}+
\frac{\partial^2 \Omega}{\partial\sigma^2}
\cdot\Big(\frac{d\sigma}{d m_0}\Big)^2+
\frac{\partial\Omega}{\partial\sigma}\cdot
\frac{d^2\sigma}{d m_0^2}.
\label{eq.chiral_suscp}
\end{eqnarray}
The last term in Eq.(\ref{eq.chiral_suscp}) again vanishes due to
stationarity condition of the mean field, but the second and third
terms will remain and give the implicit contributions.
This implicit dependence is what makes life a little difficult in the
mean field calculations which are otherwise quite straightforward.
While the explicit derivatives can be systematically obtained up to
any desired order in a closed analytic form, the implicit contributions
are usually obtained through numerical derivatives. Normally one has to
resort to numerical evaluation of the total derivatives like
$\chi_{\sigma}$, or the implicit part like $d\sigma/d m_0$. Such
derivatives may either be done by direct difference approximations or
by a Taylor series method as proposed by us in Ref.\cite{rajarshidaA}.
Unfortunately both these methods tend to give either large errors or
become quite insensitive as the order of the derivatives are increased.
A possible alternative to these numerical techniques has been explored
in Refs.\cite{AD_WWS,AD_CEP,AD_SW} through the method of algorithmic
differentiation. Derivatives up to very high orders may be computed in
this technique. Though very efficient and less error prone even
for obtaining very high derivatives, the method is algorithmically
involved.
One of our main focuses here is to obtain the implicit contribution in
a semi-analytic approach so that numerical uncertainties are minimized.
Here we shall outline a simple algorithm for obtaining derivatives
in a straightforward semi-analytical procedure. The procedure is
completely analytic as far as obtaining the derivatives go. We shall
argue that the total derivatives at any order can be completely
expressed only in terms of explicit derivatives up to one lower order.
Therefore one can obtain the expressions for total derivatives
completely analytically. Only the values of the final expressions
so obtained are computed numerically.
For this purpose we shall discuss the derivatives with respect to
$\mu_q$, though the methodology would be identical for any other similar
derivatives. The only numerics involved will be the momentum integrals,
and to this end all the methods of differentiation would have identical
efficiency and accuracy.
The derivatives of $\Omega$ with respect to $\mu_q$ would give the quark
number and its susceptibilities. Since the quark number is exactly
conserved one may use the Ward-Takahashi identity to derive the
corresponding three-point functions for the bare and effective theories.
The identity is given as,
\begin{equation}
q^{\mu}\Gamma_{\mu}(p,p+q)={S_1}^{-1}(p+q)-{S_1}^{-1}(p).
\label{eq.WTI_exact}
\end{equation}
Here $p$ and $q$ denote the four-momentums of the fermion and the
boson respectively. Now, $q_{\mu}\to 0$ limit of
Eq.(\ref{eq.WTI_exact}) yields the Ward Identity,
\begin{equation}
\frac{d{S_1}^{-1}}{d p^{\mu}}=\Gamma_{\mu}(p,p),
\label{eq.WI_diff}
\end{equation}
which in differential form, gives the insertion factor corresponding
to the zero momentum boson line into an internal fermion line.
In the imaginary time formalism, at finite temperature and
chemical potential, the fourth component of momentum becomes
$p_0=i(2n+1)\pi T+\mu_q$ and thus Eq.(\ref{eq.WI_diff})
can be written as \cite{kapusta},
\begin{equation}
\frac{d{S_1}^{-1}}{d \mu_q}=\Gamma_{0}(p,p).
\label{eq.WI_diff_finiteT}
\end{equation}
Using Eq.(\ref{eq.mod_prop_inv}) in the above relation we obtain,
\begin{equation}
\frac{d{S_1}^{-1}}{d \mu_q}=\gamma_0+\Big(G \frac{d \sigma}
{d \mu_q}\Big) \cdot \id_D \equiv \Gamma_0,
\label{eq.vertex_eff}
\end{equation}
where $\id_D$ is the identity matrix in Dirac space. For the bare
propagator we get the expected insertion factor for non-interacting
quarks as,
\begin{equation}
\frac{d{S_0}^{-1}}{d \mu_q}=\gamma_{0},
\label{eq.vertex_bare}
\end{equation}
Let us now consider the derivative of $\sigma$ from
Eq.(\ref{eq.sig_def}) w.r.t.\ $\mu_q$ which gives,
\begin{equation}
\frac{d \sigma}{d \mu_q}=-i\mathrm {Tr}[S_1 \Gamma_0 S_1]
=-i\mathrm{Tr}(S_1 \gamma_0 S_1)-
i\mathrm{Tr}\Big(S_1 G\frac{d \sigma}{d \mu_q}S_1\Big),
\label{eq.dsigmadmu_def}
\end{equation}
where the effective three-point function from Eq.(\ref{eq.vertex_eff})
is used. For the bare propagator $S_0$ one can easily check that
\begin{equation}
\frac{d S_0}{d \mu_q}=-S_0 \gamma_0 S_0,
\label{eq.unit}
\end{equation}
which is basically another form of Ward identity for bare three-point
function. The corresponding relation for the dressed propagator $S_1$
with the effective three-point function will be,
\begin{equation}
\frac{d S_1}{d \mu_q}={-S_1 \Gamma_0 S_1}.
\end{equation}
Rearranging terms in (\ref{eq.dsigmadmu_def}) it is possible
to write $\frac{d \sigma}{d \mu_q}$ in a closed form in terms of $m_0$,
$T$, $\mu_q$ and $\sigma$ only:
\begin{equation}
\frac{d\sigma}{d\mu_q}=
\frac{-i{\rm Tr}(S_1\gamma_0 S_1)}{1+iG{\rm Tr}({S_1}^2)}.
\label{eq.dsigmadmu_final}
\end{equation}
For the second order derivative, one may start from
Eq.(\ref{eq.dsigmadmu_def}) to get,
\begin{equation}
\frac{d^2\sigma}{d\mu_q^2}=
2i\mathrm {Tr}[S_1 \Gamma_0 S_1 \Gamma_0 S_1]-
i\mathrm{Tr}\Big(S_1 G\frac{d^2 \sigma}{d \mu_q^2}S_1\Big).
\label{eq.d2sigdmu2_def}
\end{equation}
Again rearranging terms, one can write a closed form expression for
$\frac{d^2\sigma}{d\mu_q^2}$ as a function of $m_0$, $T$, $\mu_q$,
$\sigma$ and $\frac{d \sigma}{d \mu_q}$ as,
\begin{eqnarray}
\frac{d^2\sigma}{d\mu_q^2}=
\frac{2i{\rm Tr}(S_1\Gamma_0 S_1\Gamma_0 S_1)}{1+iG{\rm Tr}({S_1}^2)}
=\frac{2i[{\rm{Tr}}(S_1\gamma_0 S_1\gamma_0 S_1)
+2(G\frac{d \sigma}{d \mu_q}){\rm{Tr}}({S_1}^3\gamma_0)
+(G\frac{d \sigma}{d \mu_q})^2{\rm{Tr}}({S_1}^3)]}
{1+iG{\rm Tr}({S_1}^2)}.
\end{eqnarray}
We have plotted the first and second order derivatives of $\sigma$
w.r.t.\ $\mu_q$ as function of $T$ in Fig.\ref{fg.fldsder1_njl}
and Fig.\ref{fg.fldsder2_njl} respectively, where we have compared
the results of semi-analytical approach presented here to that of
numerical methods like Taylor expansion or finite difference.
\begin{figure} [!ht]
\subfigure[]
{\includegraphics [scale=0.6] {fldsder1_njl.eps}
\label{fg.fldsder1_njl}}
\subfigure[]
{\includegraphics [scale=0.6] {fldsder2_njl.eps}
\label{fg.fldsder2_njl}}
\caption{First (a) and second (b) derivatives of the mean field
$\sigma$ with respect to $\mu_q$ in NJL model at $\mu_q=T_c$. Here
the points represent the result from numerical differentiation and
the lines are from the semi-analytical approach described in the text.}
\label{fg.fldsder_njl}
\end{figure}
In the same way all the higher order derivatives may be obtained
systematically as a function of $m_0$, $T$, $\mu_q$ and derivatives up
to one lower order.
This method is certainly more accurate than a direct numerical
differentiation of $\sigma$ w.r.t.\ $\mu_q$, or fitting Taylor
coefficients in an expansion w.r.t.\ $\mu_q$. No numerical
approximations or uncertainties are introduced, except for
the numerical integration of the fermionic momentum integrals.
Therefore the question of insensitivity at higher orders also
does not arise.
\subsection{PNJL model}
We now discuss the Polyakov loop enhanced NJL model. The situation here
is similar to that of the NJL model except that we now have a couple of
mean fields more in the form of the expectation value of the Polyakov
loop $\Phi$ and that of its conjugate $\bar{\Phi}$. The Lagrangian for
the 2 flavor PNJL model is given by,
\begin{equation}
{\mathcal L}_{\rm PNJL} = \bar{\psi}(i\slashed D-m_0+\gamma_0\mu)\psi +
\frac{G}{2}[(\bar{\psi}\psi)^2+(\bar{\psi}i\gamma_5\vec{\tau}\psi)^2]
- {\mathcal U}(\Phi[A],\bar{\Phi}[A],T).
\end{equation}
\noindent
where $D^\mu=\partial^\mu-ig{\mathcal A}^\mu_a\lambda_a/2$,
${\mathcal A}^\mu_a$ being the $SU(3)$ background fields and $\lambda_a$
are the Gell-Mann matrices. Here the effective Polyakov loop potential is
given by,
\begin{equation}
\frac{{\mathcal U}(\Phi,\bar{\Phi},T)}{T^4} =
-\frac{b_2(T)}{2}\Phi\bar{\Phi} -
\frac{b_3}{6}(\Phi^3+{\bar{\Phi}}^3) +
\frac{b_4}{4}(\bar{\Phi}\Phi)^2,
\label{eq.potential}
\end{equation}
with
\begin{equation*}
b_2(T) = a_0 + a_1\Big(\frac{T_0}{T}\Big) + a_2\Big(\frac{T_0}{T}\Big)^2 +
a_3\Big(\frac{T_0}{T}\Big)^3.
\end{equation*}
$\Phi$ is Polyakov loop and $\bar \Phi$ is its charge conjugate
\cite{rajarshidaB}. Values of coefficients $a_0,a_1,a_2,a_3,b_3,b_4$
have been taken from Ref.\cite{ratti1}. To take into account the effect
of SU(3) Haar measure in the PNJL model, we consider the modified
thermodynamic potential defined as \cite{rajarshidaC},
\begin{equation}
\Omega^\prime=\Omega-\kappa T^4 \ln[J(\Phi,{\bar \Phi})],
\end{equation}
where $J(\Phi,{\bar \Phi})$ is the Vandermonde (VdM)
determinant given as,
\begin{equation}
J[\Phi, {\bar \Phi}]=(27/24{\pi^2})(1-6\Phi {\bar \Phi}+
4(\Phi^3+{\bar \Phi}^3)-3{(\Phi {\bar \Phi})}^2).\nonumber\\
\end{equation}
Following Ref.\cite{rajarshidaC}, the pressure in PNJL model is defined
as $\mathcal{P}=-\Omega$.
\begin{figure} [!ht]
\subfigure []
{\includegraphics [scale=0.6] {sigder1_pnjl_nonzeromu.eps}
\label{fg.sigder1_pnjl_nonzeromu}}
\subfigure []
{\includegraphics [scale=0.6] {phiphibder1_pnjl_nonzeromu.eps}
\label{fg.phiphibder1_pnjl_nonzeromu}}
\caption{
(a) $\frac{d\sigma}{d\mu_q}$ and (b) $\frac{d\Phi}{d\mu_q}$ and
$\frac{d\bar{\Phi}}{d\mu_q}$ in the PNJL model at $\mu_q=T_c$. Here
the points represent the result from numerical differentiation, and
lines are from the semi-analytical approach described in the text.
}
\label{fg.fldsder_pnjl}
\end{figure}
In NJL model we obtained $\frac{d\sigma}{d\mu_q}$ by differentiating
the gap equation (\ref{eq.sig_def}) w.r.t.\ $\mu_q$. In the PNJL
model we do not have such gap equations for the $\Phi$ and $\bar{\Phi}$
fields. We therefore differentiate the stationarity conditions
$\frac{\partial\Omega'}{\partial X}=0$ for $X=\Phi,\bar{\Phi},\sigma$
directly with the $\mu_q$ derivatives to get,
\begin{equation}
\frac{d}{d\mu_q}\Big(\frac{\partial\Omega'}{\partial X}\Big)=0.
\label{eq.firstdermuq}
\end{equation}
Note that this equation is valid only if we insert the mean field
values in $\frac{\partial\Omega'}{\partial X}$ before taking the
$\mu_q$ derivatives. This immediately gives us,
\begin{eqnarray}
\frac{\partial}{\partial\mu_q}\Big(\frac{\partial\Omega'}{\partial X}\Big)+
\frac{\partial}{\partial\Phi}\Big(\frac{\partial\Omega'}{\partial X}\Big)
\cdot\frac{d\Phi}{d\mu_q}+
\frac{\partial}{\partial\bar{\Phi}}\Big(\frac{\partial\Omega'}{\partial X}\Big)
\cdot\frac{d\bar{\Phi}}{d\mu_q}+
\frac{\partial}{\partial\sigma}\Big(\frac{\partial\Omega'}{\partial X}\Big)
\cdot\frac{d\sigma}{d\mu_q}=0.
\end{eqnarray}
So we have the matrix equation of the form $\bf {A\cdot x=B}$,
where $\bf {A}$ is the
coefficient matrix of the variables ${\bf x}=(\frac{d\Phi}{d\mu_q},
\frac{d\bar{\Phi}}{d\mu_q},\frac{d\sigma}{d\mu_q})^{\rm T}$ and
$\bf B$ matrix has the form ${\bf B}=(-\frac{\partial}{\partial\mu_q}
(\frac{\partial\Omega'}{\partial \Phi}),
-\frac{\partial}{\partial\mu_q}(\frac{\partial\Omega'}{\partial \bar{\Phi}}),
-\frac{\partial}{\partial\mu_q}(\frac{\partial\Omega'}{\partial \sigma}))^{\rm T}$.
The above matrix equation has the solutions of the form;
\begin{equation}
\frac{d\Phi}{d\mu_q}=\frac{\triangle_1}{\triangle},
~~~~
\frac{d\bar{\Phi}}{d\mu_q}=\frac{\triangle_2}{\triangle},
~~~~
\frac{d\sigma}{d\mu_q}=\frac{\triangle_3}{\triangle},
\end{equation}
where the Cramer's determinants are given by,
\begin{equation}
\triangle = \rm{det}(\bf A) =
\left|
\begin{array}{ccc}
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big)
\end{array}
\right|,
\end{equation}
\vspace{0.2in}
\begin{equation}
\triangle_1 =
-\left|
\begin{array}{ccc}
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big)
\end{array}
\right|,
\end{equation}
\vspace{0.2in}
\begin{equation}
\triangle_2 =
-\left|
\begin{array}{ccc}
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\sigma}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big)
\end{array}
\right|,
\end{equation}
\vspace{0.2in}
\begin{equation}
\triangle_3 =
-\left|
\begin{array}{cccc}
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\Phi}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\bar{\Phi}}\Big)
\vspace{0.1in} \\
\dfrac{\partial}{\partial\Phi}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\bar{\Phi}}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big) &
\dfrac{\partial}{\partial\mu_q}\Big(\dfrac{\partial\Omega'}{\partial\sigma}\Big)
\end{array}
\right|.
\end{equation}
The elements of the determinants can be obtained from the expression
of $\Omega'$. Instead of using Cramer's rule, the solutions can also
be obtained through Gaussian Elimination method.
In Fig.\ref{fg.fldsder_pnjl} we have plotted the first order
derivatives of the mean fields along $T$ and compared the
results from two different methods as in case of NJL
in Fig.\ref{fg.fldsder_njl}.
One may similarly obtain the higher derivatives of the mean fields
with respect to $\mu_q$ by sequentially increasing the order of
derivatives to act upon (\ref{eq.firstdermuq}). The derivatives at
any order will depend upon the various thermodynamic parameters as
well as the derivatives up to one lower order. For these higher
orders also we shall have to solve the similar matrix equations of
the form $\bf A \cdot x=B$. The most interesting part is that while
the column matrix $\bf B$ changes for every order, the coefficient
matrix $\bf A$ will remain same as in first order. We expect that
this semi-analytical prescription for obtaining field derivatives
will certainly give better results specifically in higher order,
compared to numerical derivatives. The detailed study in
that direction will be presented elsewhere. Here we shall use this
elegant formulation to gain some insight into the physics aspects
of susceptibilities in the context of fluctuation-dissipation theorem.
\section{QNS from FDT}
\label{sc.qns_fdt}
\subsection{A Toy Model}
\label{sc.qns_njl}
\begin{figure} [!ht]
{\includegraphics [scale=0.4] {corltr_extlegs.eps}}
\caption{Time-time component of vector correlator.}
\label{fg.corr_free}
\end{figure}
We consider a Lagrangian with an effective mass
\begin{equation}
\mathcal{L}=\bar{\psi}(i\slashed \partial
-\hat{M} +\gamma_0 \hat{\mu})\psi
\label{eq.lag_mass}.
\end{equation}
where $\hat{M}$ is effective mass matrix and $\hat{\mu}$ is the
matrix of chemical potential and both are diagonal in flavor space.
$\gamma_0$ is the three point function
\footnote{Equivalent to a massive free theory.} and
the corresponding two-point function for flavor $f$ is
\begin{equation*}
S_f(L)=\frac{1}{\slashed L-M_f+{\mu}_f\gamma_0},
\end{equation*}
where $L$ is the four momentum.
With this simple consideration, it is obvious that
Fig.\ref{fg.corr_free} is the relevant diagram in one loop that would
contribute to the time-time component of vector correlator $\Pi_{00}$,
\begin{equation}
\Pi_{00}(q_0=\omega,q=\lvert\vec{Q}\rvert)=
-i\sum_{f=u,d}\int{\frac{d^4P}{(2\pi)^4}{\mathrm {Tr}}_{D,c}
[\gamma_0 S_f(K)\gamma_0 S_f(P)]},
\label{eq.pidef}
\end{equation}
with $K=P+Q$. Here trace $\mathrm {Tr}_{D,c}$ is over Dirac and
color indices only. Replacing
\[\int \frac{dP_0}{2\pi}\longrightarrow \frac{i}{\beta} \sum_{\omega_n}~~,\]
and performing the Dirac trace, Eq.(\ref{eq.pidef}) becomes,
\begin{eqnarray}
\Pi_{00}(\omega,q)&=& \sum_{f=u,d}\sum_n \frac{4}{\beta}
\int\frac{d^3p}{(2\pi)^3}{\mathrm{Tr}}_c
\Big\{ \frac{(i\omega_n+\omega+\tilde{\mu}_f)
(i\omega_n+\tilde{\mu}_f)+\vec{P}\cdot\vec{K}+M_f^2}
{[(i\omega_n+\omega+\tilde{\mu}_f)^2-E_{fk}^2]
[(i\omega_n+\tilde{\mu}_f)^2-E_{fp}^2]} \Big\}.
\end{eqnarray}
Now the remaining trace operation is over color space only.
Breaking into partial fractions R.H.S.\ of last equation can
be written as \cite{Klevansky:rmp},
\begin{eqnarray}
\Pi_{00}(\omega,q) &=& \sum_{f=u,d}\sum_n \frac{1}{\beta}
\int\frac{d^3p}{(2\pi)^3}\frac{1}{E_{fp}E_{fk}}
{\mathrm{Tr}}_c \nonumber \\
&& \Big\{
\frac{1}{(i\omega_n+\omega+\tilde{\mu}_f)-E_{fk}}
\frac{1}{(i\omega_n+\tilde{\mu}_f)-E_{fp}}
[E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}] \nonumber \\
&+& \frac{1}{(i\omega_n+\omega+\tilde{\mu}_f)-E_{fk}}\frac{1}
{(i\omega_n+\tilde{\mu}_f)+E_{fp}}
[E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}] \nonumber \\
&+& \frac{1}{(i\omega_n+\omega+\tilde{\mu}_f)+E_{fk}}\frac{1}
{(i\omega_n+\tilde{\mu}_f)-E_{fp}}
[E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}] \nonumber \\
&+& \frac{1}{(i\omega_n+\omega+\tilde{\mu}_f)+E_{fk}}\frac{1}
{(i\omega_n+\tilde{\mu}_f)+E_{fp}}
[E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}]
\Big\}.
\end{eqnarray}
Now performing the Matsubara summation over the discrete
frequencies, $\omega_n=(2n+1)\pi T$, we are left with,
\begin{eqnarray}
\Pi_{00}(\omega,q) &=& \sum_{f=u,d}
\int\frac{d^3p}{(2\pi)^3}\frac{1}{E_{fp}E_{fk}}
{\mathrm{Tr}}_c \nonumber \\
&& \Big\{
\frac{E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}}{\omega+E_{fp}-E_{fk}}
[\mathcal{F}_1(E_{fp}-\tilde{\mu}_f)-
\mathcal{F}_1(E_{fk}-\tilde{\mu}_f-\omega)] \nonumber \\
&+&\frac{E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}}{\omega-E_{fp}-E_{fk}}
[1-\mathcal{F}_1(E_{fp}+\tilde{\mu}_f)-
\mathcal{F}_1(E_{fk}-\tilde{\mu}_f-\omega)] \nonumber \\
&+&\frac{E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}}{\omega+E_{fp}+E_{fk}}
[\mathcal{F}_1(E_{fp}-\tilde{\mu}_f)-1+
\mathcal{F}_1(E_{fk}+\tilde{\mu}_f+\omega)] \nonumber \\
&+&\frac{E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}}{\omega-E_{fp}+E_{fk}}
[-\mathcal{F}_1(E_{fp}+\tilde{\mu}_f)+
\mathcal{F}_1(E_{fk}+\tilde{\mu}_f+\omega)]
\Big\},
\label{eq.pi_after_sum}
\end{eqnarray}
where ${\cal F}$ is the Fermi-Dirac distribution function.
If we make a change of variable $\vec{P}\rightarrow -\vec{P'}-\vec{Q}$,
in the third and fourth term then dot product of 3-vectors remains
unchanged and the momentum label of quasiparticle energy just
interchanges. Moreover keeping in mind that $e^{\beta \omega}=1$,
after simplification Eq.(\ref{eq.pi_after_sum}) becomes,
\begin{eqnarray}
\Pi_{00}(\omega,q) &=& \sum_{f=u,d}
\int\frac{d^3p}{(2\pi)^3}\frac{1}{E_{fp}E_{fk}} \nonumber \\
&& {\mathrm{Tr}}_c
\Big\{
\frac{E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}}{\omega+E_{fp}-E_{fk}}
[\mathcal{F}(E_{fp}-{\mu}_f)+\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fk}-{\mu}_f)-
\mathcal{F}(E_{fk}+{\mu}_f)] \nonumber \\
&+& [E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}]
\Big[\frac{1}{\omega-E_{fp}-E_{fk}}-
\frac{1}{\omega+E_{fp}+E_{fk}}\Big][1-\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fk}-{\mu}_f)]
\Big\}.
\label{eq.total_pi} \nonumber \\
\end{eqnarray}
\subsubsection{Calculation of $\chi_q$ from real part of $\Pi_{00}$}
After taking the real part of $\Pi_{00}(\omega,q)$ when we put $\omega=0$,
we are left with;
\begin{eqnarray}
{\textrm {Re}} \Pi_{00}(\omega=0,q) &=& \sum_{f=u,d}
\int\frac{d^3p}{(2\pi)^3}\frac{1}{E_{fp}E_{fk}}{\mathrm{Tr}}_c
\{
[E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}] \nonumber \\
&& \times\frac{\mathcal{F}(E_{fp}-{\mu}_f)+
\mathcal{F}(E_{fp}
+{\mu}_f)-\mathcal{F}(E_{fk}-{\mu}_f)-
\mathcal{F}(E_{fk}+{\mu}_f)}
{E_{fp}-E_{fk}} \nonumber \\
&-& 2
\frac{E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}}{E_{fp}+E_{fk}}
[1-\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fk}-{\mu}_f)]
\}.
\label{eq.re_pi}
\end{eqnarray}
Now we are going to use the FDT as in Eq.(\ref{eq.fluc_diss_def}).
In the limit $q\rightarrow 0$ the second term of (\ref{eq.re_pi})
vanishes and for the first term
taking care of the $\frac{0}{0}$ form using L'Hospital rule we get,
\begin{eqnarray}
\chi_q = 2\beta \sum_{f=u,d} \int\frac{d^3p}{(2\pi)^3} {\mathrm{Tr}}_c
\Big\{
\frac{e^{\beta(E_{fp}-{\mu}_f)}}
{(1+e^{\beta(E_{fp}-{\mu}_f)})^2} +
\frac{e^{\beta(E_{fp}+{\mu}_f)}}
{(1+e^{\beta(E_{fp}+{\mu}_f)})^2}
\Big\}. \label{eq.re_qns}
\end{eqnarray}
\subsubsection{Calculation of $\chi_q$ from Imaginary part of $\Pi_{00}$}
The imaginary part of the retarded correlator can be calculated
from the discontinuity in the following way;
\begin{eqnarray}
{\textrm {Im}} \Pi_{00}(\omega,q) &=& \frac{1}{2i}
[\Pi_{00}(\omega\rightarrow \omega +i\eta,q)
-\Pi_{00}(\omega\rightarrow \omega -i\eta,q)] \nonumber \\
&=& -\pi \sum_{f=u,d} \int\frac{d^3p}{(2\pi)^3}
\frac{1}{E_{fp}E_{fk}} {\mathrm{Tr}}_c
\{
(E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}) \nonumber \\
&& [\mathcal{F}(E_{fp}-{\mu}_f)+
\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fk}-{\mu}_f)-
\mathcal{F}(E_{fk}+{\mu}_f)]
\delta(\omega+E_{fp}-E_{fk}) \nonumber \\
&+& [E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K}]
[1-\mathcal{F}(E_{fp}+{\mu}_f)-
\mathcal{F}(E_{fk}-{\mu}_f)] \nonumber \\
&& [\delta(\omega-E_{fp}-E_{fk})-\delta(\omega+E_{fp}+E_{fk})]
\}.
\label{eq.im_pi}
\end{eqnarray}
The delta function in the first term of R.H.S.\ of the above equation
represents the contribution from the scattering process and the first
delta function of the second term represents the pair creation process
for $\omega>0$ \cite{Gert_calc,weldon_prd28}. The prefactors containing
Fermi-Dirac distributions to both of the above-mentioned terms can be
rearranged to show that they basically account for the statistical
weights of corresponding processes. Similarly for $\omega<0$,
one can realize some processes \cite{weldon_prd28} corresponding
to the second delta function in the second term. Although
that will clearly violates energy conservation for $\omega>0$,
since quasiparticle energies are always positive and therefore
hereinafter this term will be dropped.
As an intermediate but important step we want to show that, the
first term of R.H.S.\ of the above equation can be written as
\cite{Laine_deltaomega},
\begin{eqnarray}
&& (E_{fp}(\omega+E_{fp})+M_f^2+\vec{P}\cdot\vec{K})
\delta(\omega+E_{fp}-E_{fk}) \times \nonumber \\
&& [\mathcal{F}(E_{fp}-{\mu}_f)+
\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fp}-{\mu}_f+\omega)-
\mathcal{F}(E_{fp}+{\mu}_f+\omega)] \nonumber \\
&& \stackrel{q \rightarrow 0}{\approx}
-(E_{fp}(\omega+E_{fp})+M_f^2+\vec{P}^2) ~ \omega\delta(\omega)
~ [\mathcal{F'}(E_{fp}-{\mu}_f)+
\mathcal{F'}(E_{fp}+{\mu}_f)] \nonumber \\
&=& (E_{fp}(\omega+E_{fp})+M_f^2+\vec{P}^2)
~ \omega\delta(\omega) \times \nonumber \\
&& \Big[\mathcal{F}(E_{fp}-{\mu}_f)
\Big(1-\mathcal{F}(E_{fp}-{\mu}_f)\Big)+
\mathcal{F}(E_{fp}+{\mu}_f)
\Big(1-\mathcal{F}(E_{fp}+{\mu}_f)\Big)\Big].
\end{eqnarray}
Here, for the time being we have omitted the integration and
prefactors. Proportionality of this term to $\omega\delta(\omega)$
is important because it is related to number conservation
\cite{purnendu1,purnendu2,najmul,Kunihiro_PLB,Fujii04,Gert_calc}.
Apart from that, this kind of zero mode contribution in the
spectral functions is significant and gives rise to a
constant contribution in finite temperature Euclidean correlator
\cite{Umeda,Karsh,munshi,Ding,sourendu_meson_fluc}.
From the FDT as in Eq.(\ref{eq.fluc_diss_def}) and
using (\ref{eq.im_pi}) we get,
\begin{eqnarray}
\chi_q &=& \lim_{q\rightarrow 0}\beta \sum_{f=u,d}
\int\frac{d^3p}{(2\pi)^3} \frac{1}{E_{fp}E_{fk}} {\mathrm{Tr}}_c
\Big\{
(E_{fp}E_{fk}+M_f^2+\vec{P}\cdot\vec{K}) \nonumber \\
&& \frac{\mathcal{F}(E_{fp}-{\mu}_f)+
\mathcal{F}(E_{fp}+{\mu}_f)
-\mathcal{F}(E_{fk}-{\mu}_f)-
\mathcal{F}(E_{fk}+{\mu}_f)}
{1-e^{\beta(E_{fp}-E_{fk})}} \nonumber \\
&+& (E_{fp}E_{fk}-M_f^2-\vec{P}\cdot\vec{K})
[1-\mathcal{F}(E_{fp}+{\mu}_f)-
\mathcal{F}(E_{fk}-{\mu}_f)]
\frac{1}{1-e^{-\beta(E_{fp}+E_{fk})}}
\Big\}. \nonumber
\end{eqnarray}
For the limit of vanishing external momentum second term vanishes.
Using L'Hospital rule for the first term we are left with,
\begin{eqnarray}
\chi_q = 2\beta \sum_{f=u,d} \int\frac{d^3p}{(2\pi)^3} {\mathrm{Tr}}_c
\Big\{
\frac{e^{\beta(E_{fp}-{\mu}_f)}}
{(1+e^{\beta(E_{fp}-{\mu}_f)})^2} +
\frac{e^{\beta(E_{fp}+{\mu}_f)}}
{(1+e^{\beta(E_{fp}+{\mu}_f)})^2}
\Big\}.
\label{eq.qns_before_trace}
\end{eqnarray}
\subsubsection{Calculation of $\chi_q$ from Thermodynamic derivative}
One can write the partition function corresponding to
(\ref{eq.lag_mass}) as
\begin{eqnarray}
{\cal Z}(\beta,\{{\mu}_f\})=\int {\cal D}[\bar \psi]
{\cal D}[{\psi}]
e^{-i\int d^4x\mathcal{L}(\psi,{\bar \psi};\{{\mu}_f\})},
\label{eq.i3}
\end{eqnarray}
where $\beta$ is the inverse temperature.
The pressure can be written as
\begin{equation}
{\cal P}=\frac{1}{\cal V}
\ln{\cal Z}(\beta, \{{\mu}_f\})
\ , \label{eq.i3_1}
\end{equation}
where the four-volume, ${\cal V}=\beta V$ with
$V$ is the three-volume. One can straight away compute the pressure and
show that the $\chi_q$ obtained from it through thermodynamic derivative
with respect to ${\mu}_f$ is exactly the same as those obtained in
(\ref{eq.re_qns}) and (\ref{eq.qns_before_trace}).
Nevertheless, we now demonstrate this in a very general perspective.
${\mathcal P}'$ can be obtained from (\ref{eq.i3_1}) as
\begin{eqnarray}
\frac{\partial{\mathcal P}}{\partial {\mu}_f}
&=&\frac{-i}{{\cal V}{\cal Z}}\ {\int {\cal D}[\bar \psi]
{\cal D}[\psi]
\int d^4x \ {\bar \psi}^f(x) \gamma_0 \psi^f(x)} \,
e^{ {-i\int d^4x {\cal L} (\psi,{\bar{\psi}};\{{\mu}_f\})}} \ .
\label{eq.i3_2}
\end{eqnarray}
The quark propagator in a hot and dense medium is defined as
\begin{eqnarray}
S^f_{\alpha\sigma}(x,x')&=&\frac{
\int {\cal D}[\bar \psi] {\cal D}[\psi]
\psi^f_{\alpha}(x){\bar \psi}^f_\sigma(x') \exp \left ({-i\int d^4x {\cal L}
(\psi,{\bar{\psi}}; \{{\mu}_f\})}\right)}
{ \int {\cal D}[\bar \psi] {\cal D}[\psi]
\exp \left ({-i\int d^4x {\cal L}
(\psi,{\bar{\psi}};\{{\mu}_f\})}\right)} \ .
\label{eq.i3_3}
\end{eqnarray}
From Eq.(\ref{eq.i3_2}) one can write the quark number density as
\begin{eqnarray}
n_q=\sum_f n_f=\sum_f\frac{\partial{\cal P}}{\partial {\mu}_f}
&=& -i\sum_f \int\! \frac{d^4P}{(2\pi)^4}
{\mbox{Tr}}_{D,c}\left [S_f(P) \gamma_0 \right ].
\label{eq.i3_4}
\end{eqnarray}
Likewise, one can also obtain QNS as
\begin{eqnarray}
\chi_q=\sum_f \frac{\partial^2{\cal P}}{\partial {\mu}^2_f}
&=& i\sum_f \int\! \frac{d^4P}{(2\pi)^4}
{\mbox{Tr}}_{D,c}\left [S_f(P) \gamma_0 S_f(P) \gamma_0\right ]
=-i\mathrm {Tr}[S(P) \gamma_0 S(P) \gamma_0]
\label{eq.i3_5}
\end{eqnarray}
where the relation in (\ref{eq.unit}) is used. Now one can see that the
(\ref{eq.i3_5}) corresponds to the temporal correlator in
Fig.\ref{fg.corr_free} but at the external momentum
$Q=(\omega_q,|{\vec q}|)=0$ or amputated external legs. This static
correlator has been computed as $C_{00}$ in Appendix \ref{sc.trace},
which is exactly equal to those obtained in (\ref{eq.re_qns}) and
(\ref{eq.qns_before_trace}). This shows that the FDT vis-a-vis the
thermodynamic sum rule is satisfied in a toy model.
\subsection{NJL Model}
So far our discussions are based on the naive consideration of a toy
model with a free massive propagator, but this has set the stage for
any realistic model calculations. Here, we consider the NJL
Lagrangian of Ref.(\ref{eq.lag_mf}), in which the explicit interaction
term through chiral condensate $\sigma$ is present and this would
contribute to the physical quantities one would like to compute.
The relevant diagrams that would contribute to the correlation
function are shown in Fig.\ref{fg.corr_njl}.
\begin{figure} [!ht]
\subfigure[]
{\includegraphics [scale=0.45] {extleg_connected.eps}} ~~~~~~~~
\subfigure[]
{\includegraphics [scale=0.45] {extleg_disconnected.eps}}
\caption{Time-time component of vector correlator in NJL model.
The diagram (a) is the usual self-energy with effective propagator
and three-point vertex whereas diagram (b) has the origin of
four-fermionic interaction.}
\label{fg.corr_njl}
\end{figure}
We note that the effective propagator $S_1$, three-point vertex
$\Gamma_0$ and the chiral condensates for NJL model have already
been defined in Sec.\ref{sc.fldsder_njl} and one can easily compute
these diagrams, but we purposefully avoid this and the reason for
which will be clear later.
Let us now first concentrate on the calculation of QNS from the
thermodynamic derivative of pressure with respect to the chemical
potential. As discussed earlier in Sec.\ref{sc.fldsder_njl}, the
mean fields have implicit dependences on chemical potential, thus
the thermodynamic derivatives are to be considered appropriately.
This implies that one needs to consider the total derivative of
pressure rather than the explicit one.
With these considerations, we write the pressure
${\cal P} = -\Omega$ from (\ref{eq.omega}). The quark number density is
then given as,
\begin{eqnarray}
n_q= \frac{d \cal P}{d \mu_q}
&=&\frac{\partial \cal P}{\partial \mu_q}+
\frac{\partial \cal P} {\partial \sigma}\cdot\frac{d \sigma}{d \mu_q}
= i\mathrm {Tr}(\gamma_0 S_1)+G \frac{d \sigma}{d \mu_q}
[i\mathrm {Tr}(S_1)-\sigma]
= i\mathrm {Tr}(\Gamma_0 S_1)
-G\sigma\frac{d \sigma}{d \mu_q},
\label{eq.dp_dmu}
\end{eqnarray}
where we have used
\begin{equation}
{\frac{\partial \cal P}{\partial \mu_q}}=
i\mathrm {Tr}(\gamma_0 S_1),
\label{eq.delp_delmu}
\end{equation}
and
\begin{equation}
\frac{\partial \cal P}{\partial \sigma}=G[i\mathrm {Tr}(S_1)-\sigma].
\label{eq.delp_delsigma}
\end{equation}
It is interesting to note that for number density, $n_q$, the partial
derivative alone gives the full contribution in the mean field theory
as $\frac{\partial P}{\partial \sigma}=0$ if one uses (\ref{eq.sig_def}) in
(\ref{eq.delp_delsigma}).
One can similarly obtain the second order derivative of pressure w.r.t.\
$\mu_q$ from (\ref{eq.dp_dmu}) to get the QNS as,
\begin{eqnarray}
\chi_q= \frac{d^2 \cal P}{d \mu_q^2}
&=&\dfrac{\partial^2 \cal P}
{\partial \mu_q^2}
+\Big[\frac{\partial}{\partial \mu_q}
\Big(\frac{\partial \cal P}{\partial \sigma}\Big)
+\frac{\partial}{\partial \sigma}
\Big(\frac{\partial \cal P}{\partial \mu_q}\Big)\Big]
\cdot\frac{d \sigma}{d \mu_q}
+\frac{\partial \cal P}{\partial \sigma}\cdot
\frac{d^2 \sigma}{d \mu_q^2}
+\frac{\partial^2 \cal P}{\partial \sigma^2}\cdot
\Big(\frac{d \sigma}{d \mu_q}\Big)^2 \nonumber \\
&=& -i \mathrm {Tr}[\gamma_0 S_1 \gamma_0 S_1]-
2iG\Big(\frac{d \sigma}{d \mu_q}\Big)\mathrm {Tr}[S_1 \gamma_0 S_1]+
G \frac{d^2 \sigma}{d \mu_q^2}[i \mathrm {Tr}(S_1)-\sigma]+
G\Big(\frac{d \sigma}{d \mu_q}\Big)^2[-iG \mathrm {Tr}(S_1^2)-1] \nonumber \\
&=& -i \mathrm {Tr}(\Gamma_0 S_1\Gamma_0 S_1)
-G\Big(\frac{d \sigma}{d \mu_q}\Big)^2
+G \frac{d^2 \sigma}{d \mu_q^2}[i \mathrm {Tr}(S_1)-\sigma],
\label{eq.d2pdmu2b}
\end{eqnarray}
where various second order explicit derivatives
are used in terms of respective correlators
by using (\ref{eq.delp_delmu}) and
(\ref{eq.delp_delsigma}). These relations are noted below:
\begin{eqnarray}
\frac{\partial^2 \cal P}{\partial \mu_q^2}
&=&-i \mathrm {Tr}[\gamma_0 S_1 \gamma_0 S_1]=-C_{00},
\label{eq.explmuq}\\
\frac{\partial}{\partial \mu_q}
\Big(\frac{\partial \cal P}{\partial \sigma}\Big)
&=&-iG \mathrm {Tr}[S_1 \gamma_0 S_1]=-C_{01},
\label{eq.cross_mu_sigma2} \\
\frac{\partial}{\partial \sigma}
\Big(\frac{\partial \cal P}{\partial \mu_q}\Big)
&=&-iG \mathrm {Tr}[S_1 \gamma_0 S_1]=-C_{10},
\label{eq.cross_mu_sigma1}\\
\frac{\partial^2 \cal P}{\partial \sigma^2}
&=&G [-iG \mathrm {Tr}(S_1^2)-1]=-C_{11}-G.
\label{eq.delp_delsig2}
\end{eqnarray}
The detailed calculations of the correlators $C_{00}, C_{01}$ and
$C_{11}$ are presented in Appendix \ref{sc.trace}. Here we have used the
relation $\dfrac{\partial S_1}{\partial\sigma} =-G{S_1}^2$,
which can be easily obtained from (\ref{eq.mod_prop_inv}).
Furthermore, the last term of (\ref{eq.d2pdmu2b}) vanishes due to
(\ref{eq.sig_def}) and using the first equality of
(\ref{eq.dsigmadmu_def}) we can finally write,
\begin{equation}
\chi_q= \frac{d^2 \cal P}{d \mu_q^2}=
-i \mathrm {Tr}(\Gamma_0 S_1\Gamma_0 S_1)-
G(-i\mathrm {Tr}[S_1 \Gamma_0 S_1])^2.
\label{eq.d2pdmu2_corrl}
\end{equation}
\begin{figure} [!ht]
\subfigure[]
{\includegraphics [scale=0.45] {connected.eps}
\label{fg.corltr_con}} ~~~~~~~~
\subfigure[]
{\includegraphics [scale=0.45] {disconnected.eps}
\label{fg.corltr_dis}}
\caption{Static or amputated vector correlators with modified
propagator and effective three-point function.}
\label{fg.corltr_top}
\end{figure}
The right hand side of (\ref{eq.d2pdmu2_corrl}) can be
viewed in terms of diagrammatic topology as displayed
in Fig.\ref{fg.corltr_top}. It is evident that these are
equivalent to the vector correlator in NJL model in static limit
or amputated legs as given in Fig.\ref{fg.corr_njl}.
We note that in a mean field approach, where the mean fields
are sensitive to external source, an appropriate measure
is to be taken to satisfy the FDT vis-a-vis thermodynamic sum
rule. This implies that the inclusion of
implicit $\mu_q$ dependences of the mean fields is not ad hoc,
rather it enables us, from the field theoretic point of view,
to compute the correlators associated with the conserved density
fluctuation through diagrammatic way in NJL model.
Another interesting as well as relevant point we would like to
demonstrate below. Putting the results of (\ref{eq.cross_mu_sigma2})
and (\ref{eq.delp_delsig2}) into (\ref{eq.dsigmadmu_final}),
we can rewrite it as
\begin{equation}
\dfrac{d\sigma}{d\mu_q}=-\dfrac{\dfrac{\partial}{\partial \sigma}
\Big(\dfrac{\partial {\cal P}}{\partial \mu_q}\Big)}
{\dfrac{\partial^2 {\cal P}}{\partial \sigma^2}}
\label{eq.dsigmadmu_new}
\end{equation}
Now replacing this $\dfrac{d\sigma}{d\mu_q}$ in the first line of
Eq.(\ref{eq.d2pdmu2b}) and keeping in mind that the last but one term
will vanish due to the mean field condition, one can have
\begin{eqnarray}
\frac{d^2 {\cal P}}{d \mu_q^2}&=&\frac{\partial^2 P}{\partial \mu_q^2}
-2\frac{\Big[\frac{\partial}{\partial \sigma}
\Big(\frac{\partial {\cal P}}{\partial \mu_q}\Big)\Big]^2}
{\frac{\partial^2 {\cal P}}{\partial \sigma^2}}
+\frac{\partial^2 {\cal P}}{\partial \sigma^2}\cdot
\frac{\Big[\frac{\partial}{\partial \sigma}
\Big(\frac{\partial {\cal P}}{\partial \mu_q}\Big)\Big]^2}
{[\frac{\partial^2 {\cal P}}{\partial \sigma^2}]^2} \nonumber \\
&=&\frac{\frac{\partial^2 {\cal P}}{\partial \mu_q^2}\cdot
\frac{\partial^2 {\cal P}}{\partial \sigma^2}
-\Big[\frac{\partial}{\partial \sigma}
\Big(\frac{\partial {\cal P}}{\partial \mu_q}\Big)\Big]^2}
{\frac{\partial^2 {\cal P}}{\partial \sigma^2}}
=\frac{\partial^2 {\cal P}}{\partial \mu_q^2}-
\frac{\partial^2 {\cal P}}{\partial\mu_q\partial\sigma}\cdot
\Big( \frac{\partial^2 {\cal P}}{\partial \sigma^2} \Big)^{-1}\cdot
\frac{\partial^2 {\cal P}}{\partial\sigma\partial\mu_q}
\end{eqnarray}
This mixing pattern is already established in
Refs.\cite{Kunihiro_PLB,Fujii03} and is similar to the
mixing in susceptibilities when computed from the inverse of
curvature matrix \cite{Fukushima_PRD,Fukushima_PLB}.
\begin{figure} [!ht]
{\includegraphics [scale=0.6] {qns_compare_njl.eps}}
\caption{Quark number susceptibility in 2 flavor NJL model
at nonzero $\mu_q$.}
\label{fg.chi_q_njl}
\end{figure}
The behavior of QNS in NJL model is shown in Fig.\ref{fg.chi_q_njl}.
Here the study has been done for 2 flavor NJL model at only at non-zero
$\mu_q$. This is because at $\mu_q=0$ we have
$\frac{d\sigma}{d\mu_q}=0$ due to CP symmetry, and thus all implicit
contributions vanish. The two curves in Fig.\ref{fg.chi_q_njl}
represent explicit and total contributions to the QNS respectively.
Important contributions from the implicit $\mu_q$ dependent terms
arise close to the transition region where the change in the mean fields
is most significant. It is needless to mention that the QNS obtained
here from (\ref{eq.d2pdmu2_corrl}) comes out to be the same as that
obtained from any of the numerical derivative methods.
\subsection{PNJL Model}
\label{sc.pnjl_qns}
\begin{figure} [!ht]
\subfigure[]
{\includegraphics [scale=0.6] {qns_compare_pnjl.eps}}
\subfigure[]
{\includegraphics [scale=0.6] {qns_compare_pnjl_nonzromu.eps}}
\caption{(a) Quark number susceptibility in 2 flavor PNJL model
at $\mu_q=0$. Lattice data are taken from Ref.\cite{Alton}.
(b) QNS in PNJL model for non-zero chemical potential.}
\label{fg.qns_nonzromu_pnjl}
\end{figure}
In the present form of the PNJL model, the gluon physics is contained
only effectively in a static background field that comes through the
inclusion of Polyakov loop. We have no gluon-like quasi particles in
PNJL model. We treat such bosonic fields as purely classical ones unlike
the fermionic fields. One can only obtain their mean values through
the minimization of the thermodynamic potential. Obviously the Polyakov
loop fields are also to be determined from a set of transcendental
equations. The mean fields $\Phi$ and $\bar{\Phi}$ so obtained also
depend on $T$ and $\mu_q$ in a similar way as $\sigma$. Here one can
intuitively write as in the case of the NJL model that,
\begin{equation}
\frac{d \cal P}{d \mu_q}=\frac{\partial \cal P}{\partial \mu_q}
+\frac{\partial \cal P}{\partial \sigma}\frac{d \sigma}{d \mu_q}
+\frac{\partial \cal P}{\partial \Phi}\frac{d \Phi}{d \mu_q}
+\frac{\partial \cal P}{\partial \bar{\Phi}}\frac{d \bar{\Phi}}{d \mu_q}
\equiv\frac{\partial \cal P}{\partial \mu_q}+\sum_{X=\sigma,\Phi,\bar{\Phi}}
\frac{\partial \cal P}{\partial X}\cdot\frac{dX}{d\mu_q}.
\label{eq.delp_delmu_pnjl}
\end{equation}
Like in the NJL model, here also the second term of
(\ref{eq.delp_delmu_pnjl}) will vanish due to
$\frac{\partial \cal P}{\partial \sigma}=0$. But a crucial difference lies in
the fact that for the PNJL model,
$\frac{\partial \cal P}{\partial \Phi}\neq 0$ and
$\frac{\partial \cal P}{\partial \bar{\Phi}}\neq 0$.
This is due to the fact that $P=-\Omega\neq-\Omega'$ \cite{rajarshidaC}.
So, in case of the PNJL model even for first order derivative, we shall
have a finite contribution from implicit $\mu_q$ dependences through
$\Phi$ and $\bar{\Phi}$, i.e.
$\frac{d \cal P}{d \mu_q}\neq\frac{\partial \cal P}{\partial \mu_q}$.
Differentiating (\ref{eq.delp_delmu_pnjl}) w.r.t.\ $\mu_q$ we have,
\begin{equation}
\frac{d^2 \cal P}{d\mu_q^2}=\frac{\partial^2 \cal P}{\partial\mu_q^2}+
2\sum_{X=\sigma,\Phi,\bar{\Phi}}
\frac{\partial^2 \cal P}{\partial\mu_q\partial X}\cdot
\frac{dX}{d\mu_q}+\sum_{X=\sigma,\Phi,\bar{\Phi}}
\frac{\partial \cal P}{\partial X}\cdot\frac{d^2X}{d\mu_q^2}
+\sum_{X,Y=\sigma,\Phi,\bar{\Phi}}
\frac{\partial^2 \cal P}{\partial X\partial Y}\cdot
\frac{dX}{d\mu_q}\cdot\frac{dY}{d\mu_q},
\label{eq.qns}
\end{equation}
where, the first term is from the explicit appearances of $\mu_q$
and the other three terms contains the contributions coming from
the implicit $\mu_q$ dependences of pressure through the mean fields.
In the left panel of Fig.\ref{fg.qns_nonzromu_pnjl} the plots
of $\chi_q$ at $\mu_q=0$ are presented for the PNJL model. The
contribution from the explicit $\mu_q$ dependence and the total
contribution are shown separately. The latter contribution comes
out to be same as the QNS obtained from numerical derivatives of
pressure. Again due to CP symmetry, the non-vanishing implicit
contributions come through the $\mu_q$ dependence of $\Phi$ and
$\bar{\Phi}$ only. Our result is compared to that of
Lattice QCD\cite{Alton}. The QNS in the PNJL model for non-zero
$\mu_q$ is presented in the right panel of
Fig.\ref{fg.qns_nonzromu_pnjl}. Here again the explicit contribution
is shown separately. As is clearly evident, the presence of the
implicit contributions are significant close to the transition. The
most notable feature is that the peak in the susceptibility arises
solely due to the implicit chemical potential dependence of pressure.
Location of any critical point is therefore crucially dependent on the
proper evaluation of chemical potential dependence of the mean fields.
\section{Conclusion}
\label{sc.concl}
In QCD inspired model the fluctuation-dissipation theorem is
usually assumed to be applicable, and susceptibilities which are
associated with fluctuations are calculated from the derivatives
of pressure. In the present work through an extensive exercise we
have shown the fluctuation-dissipation theorem holds true within
the framework of NJL and PNJL models. In mean field approaches,
the mean fields are sensitive to the external source like quark
chemical potential, there should be additional contributions
coming from the implicit dependence of the mean fields on the
chemical potential. On the other hand, the temporal vector
correlator associated with the fluctuations is modified due
to the effective interaction in these model Lagrangians. Here we
have given an elegant formalism and shown that the inclusion
of implicit dependent terms through the mean fields is actually
consistent with the field theoretic point of view and consolidates
the fluctuation-dissipation theorem. For the NJL model a complete
analysis through diagrammatics could be found. While such elegant
exercise did not result for the PNJL model due to the classical
nature of the Polyakov loop, the essence of the modification
required has been clearly presented.
We have also described an analytical method for calculating the
derivative of the mean fields with respect to the chemical potential,
which forms the essential part of the modifications in the
fluctuation-dissipation theorem. This approach is essential if one
intends to study higher order derivatives for which numerical
differentiation becomes unreliable. Further studies in higher order
derivatives will be presented elsewhere. We expect these studies to
play an important role in understanding the nature of the critical
region in the phase diagram of strongly interacting matter.
\begin{acknowledgments}
We would like to thank Gert Aarts, Olaf Kaczmarek, Yan Zhu and
Mikko Laine for their crucial comments and useful communications.
Useful discussions with Najmul Haque and Chowdhury Aminul Islam are
also hereby acknowledged. A.L. acknowledges Rishi Sharma for
useful discussions.
\end{acknowledgments}
|
\subsection*{Acknowledgments}
Marion Scheepers was the first to realize the connection between Ramsey theory and
selection principles, by proving the following beautiful
qualitative extension of Ramsey's Theorem~\cite{coc1}:
Let $X$ be a topological space. If $\sone(\Om(X),\Om(X))$ holds, then for each
cover $\cU\in\Om(X)$ and each finite coloring of the set $[\cU]^2$, there is in $\Om(X)$ a cover $\cV\sub\cU$
such that the graph $[\cV]^2$ is monochromatic. Scheepers proved a large number of results of
this type, including ones jointly with Ljubi\v{s}a Ko\v{c}inac and others (e.g., \cite{coc7, SchForcing, GlCovs}).
Regarding the earlier paper~\cite{suf}, Terence Tao asked me whether
superfilters, viewed as subsets of $\beta S$, are closed.
A positive answer was known, but Tao's question pointed in a fruitful direction.
David J. Fern\'andez Bret\'on, Gili Golan and Michael (Micha\l{}) Machura read
drafts of this paper and made excellent comments.
I had long and helpful correspondences with Neil Hindman and
Imre Leader on the Milliken--Taylor Theorem.
A substantial part of this research was carried out during my Sabbatical leave at
the Faculty of Mathematics and Computer Science, Weizmann Institute of Science. I
thank Gideon Schechtman and the Faculty for their kind hospitality,
and the Faculty Teaching Committee chair, Itai Benjamini,
for the opportunity to deliver a course on Ramsey theory.
This course helped me shape the theory presented here.
|
\section{Introduction}
\label{sec:intro}
In this paper we study the absence of eigenvalues for (discrete) Jacobi operators
\[(H_{a,b}u)(n) := a(n+1)u(n+1) + b(n)u(n) + a(n) u(n-1) \quad(n\in\mathbb{Z})\]
where $a,b\in\ell_\infty(\mathbb{Z})$ such that $\frac{1}{a(\cdot)}\in \ell_\infty(\mathbb{Z})$,
and analogously (continuum) Sturm-Liouville operators
\[H_{a,\mu}u := -\partial a\partial u + u\mu,\]
where $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$ and $\mu$ is locally a complex Radon measure (for precise definitions on $\mu$ see Section \ref{sec:SL} below).
Under the assumption that the coefficients of the operators can be (locally) approximated by periodic ones in a suitable sense we prove absence of eigenvalues for these operators.
While in the Jacobi case we will work with $\ell_\infty$-approximation, for the Sturm-Liouville case we
use approximation in $L_1(\mathbb{R})$ for $a$ and a weak Wasserstein-type metric for $\mu$.
Controlling the approximation rate in a quantitative way we obtain lower bounds on the modulus of eigenvalues, where the bound is determined by the (norms of the) coefficents of the operator and the approximation rate.
In this way, our result may be seen as a quantitative version of Gordon's Theorem, which first appeared in \cite{Gordon1976}, and since then was subsequently generalized \cite{Gordon1986,Damanik2000,DamanikStolz2000, Damanik2004, Seifert2011, Seifert2012, Krueger2013, SeifertVogt2014}.
However, all the previous results stick to the case of (discrete or continuum) Schr\"odinger operators, and except of \cite{SeifertVogt2014} to a qualitative statement in the self-adjoint case.
The first quantitative result appeared in \cite{SeifertVogt2014} for not necessarily self-adjoint continuum Schr\"odinger operators on $L_2(\mathbb{R})$, where the eigenvalue bound is proven to be sharp.
We will not only consider the operators in $\ell_2(\mathbb{Z})$ and $L_2(\mathbb{R})$, respectively, but in $c_0(\mathbb{Z})$ and $C_0(\mathbb{R})$, so that we in turn obtain eigenvalue bounds also for the whole $\ell_p(\mathbb{Z})$ and $L_p(\mathbb{R})$ scale for $1\leq p<\infty$.
Note that all the sequence and function spaces we will work with are complex-valued, so that the coefficients may be complex, thus obtaining (in general) non-self-adjoint operators.
The paper is organised as follows. In Section \ref{sec:Jacobi} we deal with the discrete Jacobi case, where we also provide an example for quasiperiodic coefficients.
Section \ref{sec:SL} is devoted to the continuum Sturm-Liouville case. Here, we will also comment on optimality of our eigenvalue bound. In the appendix we provide discrete and continuum versions of Gronwall's lemma and a short lemma on norms of $SL(2,\mathbb{C})$-matrices.
\section{The Jacobi case}
\label{sec:Jacobi}
For sequences $a,b\in\ell_\infty(\mathbb{Z})$ such that $\frac{1}{a(\cdot)}\in \ell_\infty(\mathbb{Z})$ we consider the Jacobi matrix $H_{a,b}\colon c_0(\mathbb{Z})\to c_0(\mathbb{Z})$, where $c_0(\mathbb{Z})$ is the space of complex sequences with index set $\mathbb{Z}$ converging to $0$ at $\pm\infty$, defined by
\[(H_{a,b}u)(n) := a(n+1)u(n+1) + b(n)u(n) + a(n) u(n-1) \quad(n\in\mathbb{Z}).\]
We are going to prove the following theorem.
\begin{theorem}
\label{thm:Jacobi}
Let $a,b\in\ell_\infty(\mathbb{Z})$ such that $\frac{1}{a(\cdot)}\in \ell_\infty(\mathbb{Z})$.
Assume there exists a sequence $(p_m)$ in $\mathbb{N}$ with $p_m\to \infty$ and $C>0$ such that
\begin{equation}
\label{eq:Jacobi_exp}
e^{Cp_m} \max_{k=-p_m+1,\ldots,p_m} \bigl(\abs{a(k+1) - a(k+1+p_m)}+ \abs{b(k) - b(k+p_m)}\bigr)\to 0.
\end{equation}
Then $H_{a,b}$ does not have any eigenvalues with modulus less than $\frac{e^{C/2}-\norm{a}_\infty}{\norm{\frac{1}{a}}_\infty} + \norm{b}_\infty-1$.
\end{theorem}
The condition \eqref{eq:Jacobi_exp} in the preceding theorem states that the difference of the three pieces
$\bigl((a(n+1),b(n))\bigr)_{n\in\set{-p_m+1,\ldots,0}}$, $\bigl((a(n+1),b(n))\bigr)_{n\in\set{1,\ldots,p_m}}$ and $\bigl((a(n+1),b(n))\bigr)_{n\in\set{p_m+1,\ldots,2p_m}}$
tends to zero (as $m\to \infty$) faster than the exponential $e^{-Cp_m}$. This may be seen as a quantitative (in the sense of the exponential rate) version of Gordon's condition for potentials of discrete Schr\"odinger operators, see e.g.\ \cite{Gordon1976}.
In case condition \eqref{eq:Jacobi_exp} holds true for all $C>0$, we obtain an analogue of Gordon's theorem \cite{Gordon1976} for Jacobi matices.
\begin{corollary}
Let $a,b\in\ell_\infty(\mathbb{Z})$ such that $\frac{1}{a(\cdot)}\in \ell_\infty(\mathbb{Z})$.
Assume that condition \eqref{eq:Jacobi_exp} holds true for all $C>0$. Then $H_{a,b}$ does not have any eigenvalues.
\end{corollary}
Note that since $\ell_p(\mathbb{Z})\subseteq c_0(\mathbb{Z})$ for all $1\leq p<\infty$, we obtain the same result for $H_{a,b}$ considered as an operator in $\ell_p(\mathbb{Z})$ with $1\leq p<\infty$.
\begin{corollary}
Let $1\leq p<\infty$, $a,b\in\ell_\infty(\mathbb{Z})$ such that $\frac{1}{a(\cdot)}\in \ell_\infty(\mathbb{Z})$.
Assume that \eqref{eq:Jacobi_exp} is satisfied. Then $H_{a,b}\colon
\ell_p(\mathbb{Z})\to\ell_p(\mathbb{Z})$ does not have any eigenvalues with modulus less
than $\frac{e^{C/2}-\norm{a}_\infty}{\norm{\frac{1}{a}}_\infty} +
\norm{b}_\infty-1$.
In case \eqref{eq:Jacobi_exp} holds true for all $C>0$, then $H_{a,b}$ does not have any eigenvalues.
\end{corollary}
\begin{remark}
Such a result cannot hold in $\ell_\infty(\mathbb{Z})$, since periodic Jacobi matrices (i.e., Jacobi matrices with periodic sequences $a$ and $b$) have periodic eigensolutions (which are therefore in $\ell_\infty(\mathbb{Z})$).
\end{remark}
We will split the proof into several lemmas.
\begin{lemma}
Let $a,b\in\ell_\infty(\mathbb{Z})$, $a(n)\neq 0$ for all $n\in\mathbb{Z}$, $z\in\mathbb{C}$ and
$u\colon \mathbb{Z}\to \mathbb{C}$. The following are equivalent:
\begin{enumerate}
\item
$u$ is a solution of the difference equation
\[a(n+1) u(n+1) + b(n) u(n) + a(n) u(n-1) - z u(n) = 0 \quad(n\in\mathbb{Z}).\]
\item
For $m,n\in\mathbb{Z}$ we have
\[\begin{pmatrix} u(m+1)\\ a(m+1)u(m)\end{pmatrix} = T_z(m,n)\begin{pmatrix} u(n+1)\\ a(n+1)u(n)\end{pmatrix},\]
where
\[T_z(m,n) = \begin{cases}
M_z(m-1)\cdots M_z(n+1)M_z(n) & m>n,\\
I & m=n,\\
M_z(m)^{-1}\cdots M_z(n-2)^{-1}M_z(n-1)^{-1} & m<n,
\end{cases}\]
and
\[M_z(n) = \begin{pmatrix}
\frac{z-b(n+1)}{a(n+2)} & -\frac{1}{a(n+2)} \\
a(n+2) & 0
\end{pmatrix}.\]
\end{enumerate}
\end{lemma}
\begin{proof}
``(a) $\Rightarrow$ (b)'':
Fix $m,n\in\mathbb{Z}$ and let
\[T_z(m,n)\colon \begin{pmatrix} u(n+1)\\ a(n+1)u(n)\end{pmatrix} \to \begin{pmatrix} u(m+1)\\ a(m+1)u(m)\end{pmatrix}.\]
Then $T_z(m,n)$ is linear and thus can be represented by a matrix (which we will also denote by $T_z(m,n)$).
Since $u$ satisfies the difference equation, we compute
\[\begin{pmatrix}
u(n+2) \\
a(n+2)u(n+1)
\end{pmatrix} =
\underbrace{\begin{pmatrix}
\frac{z-b(n+1)}{a(n+2)} & -\frac{1}{a(n+2)} \\
a(n+2) & 0
\end{pmatrix}}_{=M_z(n)}
\begin{pmatrix}
u(n+1)\\
a(n+1)u(n)
\end{pmatrix}.\]
Thus,
\[T_z(m,n) = \begin{cases}
M_z(m-1)\cdots M_z(n+1)M_z(n) & m>n,\\
I & m=n,\\
M_z(m)^{-1}\cdots M_z(n-2)^{-1}M_z(n-1)^{-1} & m<n.
\end{cases}\]
``(b) $\Rightarrow$ (a)'':
A direct computation yields
\begin{align*}
& T_z(n+1,n) \begin{pmatrix} u(n+1)\\ a(n+1)u(n)\end{pmatrix} \\
& = \begin{pmatrix} \frac{1}{a(n+2)} \bigl(zu(n+1) - b(n+1)u(n+1) - a(n+1)u(n)\bigr) \\ a(n+1)u(n) \end{pmatrix}.
\end{align*}
Thus,
\[u(n+2) = \frac{1}{a(n+2)} \bigl(zu(n+1) - b(n+1)u(n+1) - a(n+1)u(n)\bigr),\]
which means that $u$ satisfies the difference equation.
\end{proof}
Note that $\det T_z(m,n) = 1$ for all $m,n\in\mathbb{Z}$. Furthermore, $T_z(m,n)$ ``depends locally'' on $a$ and $b$, i.e.\ $T_z(m,n)$ depends only on $a(k+1)$ and $b(k)$ for $k\in\{\min\{m,n\}+1,\ldots,\max\{m,n\}\}$.
\begin{lemma}
\label{lem:Jacobi_per}
Let $a,b\in\ell_\infty(\mathbb{Z})$ be $p$-periodic, $a(n)\neq 0$ for all $n\in\mathbb{Z}$, $z\in\mathbb{C}$, $u$ a solution of the difference equation
\[a(n+1) u(n+1) + b(n) u(n) + a(n) u(n-1) - z u(n) = 0 \quad(n\in\mathbb{Z}).\]
Then
\[\max\set{\norm{\begin{pmatrix}
u(n+1) \\ a(n+1) u(n)
\end{pmatrix}};\; n\in\set{-p,p,2p}} \geq \frac{1}{2} \norm{\begin{pmatrix}
u(1) \\ a(1) u(0)
\end{pmatrix}}.\]
\end{lemma}
\begin{proof}
By periodicity of $a$ and $b$ we have $T_z(p,0) = T_z(2p,p) = T_z(0,-p) =: T$.
The Cayley-Hamilton Theorem yields (since $\det T = 1$)
\[T^2 - (\tr T)T + I = 0.\]
In case $\abs{\tr T}\leq 1$, applying this equation to $(u(1),a(1)u(1))^\top$ we observe
\[\begin{pmatrix} u(2p+1) \\ a(2p+1) u(2p) \end{pmatrix} - \tr(T) \begin{pmatrix} u(p+1) \\ a(p+1) u(p) \end{pmatrix} = - \begin{pmatrix} u(1) \\ a(1) u(0) \end{pmatrix},\]
and therefore
\[\max\set{\norm{\begin{pmatrix}
u(n+1) \\ a(n+1) u(n)
\end{pmatrix}};\; n\in\set{p,2p}} \geq \frac{1}{2} \norm{\begin{pmatrix}
u(1) \\ a(1) u(0)
\end{pmatrix}}.\]
If $\abs{\tr(T)}>1$ we apply the equation to $(u(-p+1),a(-p+1)u(-p))^\top$ to get
\[\begin{pmatrix} u(p+1) \\ a(p+1) u(p) \end{pmatrix} + \begin{pmatrix} u(-p+1) \\ a(-p+1) u(-p) \end{pmatrix} = \tr(T) \begin{pmatrix} u(1) \\ a(1) u(0) \end{pmatrix}.\]
Thus,
\[\max\set{\norm{\begin{pmatrix}
u(n+1) \\ a(n+1) u(n)
\end{pmatrix}};\; n\in\set{-p,p}} \geq \frac{1}{2} \norm{\begin{pmatrix}
u(1) \\ a(1) u(0)
\end{pmatrix}}. \qedhere\]
\end{proof}
\begin{lemma}
\label{lem:Jacobi_diff}
Let $a,\tilde{a},b,\tilde{b}\in\ell_\infty(\mathbb{Z})$, $a(n),\tilde{a}(n)\neq 0$ for all $n\in\mathbb{Z}$, $z\in\mathbb{C}$, $u,\tilde{u}$ solutions of the difference equations
\begin{align*}
a(n+1) u(n+1) + b(n) u(n) + a(n) u(n-1) - z u(n) & = 0 \quad(n\in\mathbb{Z}),\\
\tilde{a}(n+1) \tilde{u}(n+1) + \tilde{b}(n) \tilde{u}(n) + \tilde{a}(n) \tilde{u}(n-1) - z \tilde{u}(n) & = 0 \quad(n\in\mathbb{Z}),
\end{align*}
respectively, satisfying
\[\begin{pmatrix}
u(1)\\a(1)u(0)
\end{pmatrix} = \begin{pmatrix}
\tilde{u}(1)\\\tilde{a}(1)\tilde{u}(0)
\end{pmatrix}.\]
Then, for $n\in\mathbb{Z}$, we have
\begin{align*}
& \norm{\begin{pmatrix}
u(n+1) - \tilde{u}(n+1)\\a(n+1)u(n) - \tilde{a}(n+1)\tilde{u}(n)
\end{pmatrix}}\\
& \leq \prod_{k=\min\{n+1,1\}}^{\max\{0,n\}} \norm{A(k)} \sum_{k=\min\{n+1,1\}}^{\max\{0,n\}} \norm{B(k)} \norm{\begin{pmatrix} \tilde{u}(k+\frac{1-\sgn n}{2}) \\ \tilde{a}(k+\frac{1-\sgn n}{2})\tilde{u}(k-\frac{1+\sgn n}{2})\end{pmatrix}},
\end{align*}
where
\begin{align*}
A(n) & = \begin{pmatrix}
\frac{z-b(n)}{a(n+1)} & - \frac{1}{a(n+1)} \\
a(n+1) & 0
\end{pmatrix},\\
B(n) & = \begin{pmatrix}
\left(\frac{z-b(n)}{a(n+1)} - \frac{z-\tilde{b}(n)}{\tilde{a}(n+1)}\right) & \left(- \frac{1}{a(n+1)} +\frac{1}{\tilde{a}(n+1)}\right) \\
a(n+1) - \tilde{a}(n+1) & 0
\end{pmatrix} \quad(n\in\mathbb{Z}).
\end{align*}
\end{lemma}
\begin{proof}
First, assume that $n> 0$.
We have
\begin{align*}
& \begin{pmatrix}
u(n+1) - \tilde{u}(n+1)\\a(n+1)u(n) - \tilde{a}(n+1)\tilde{u}(n)
\end{pmatrix} \\
& = \begin{pmatrix}
\frac{z-b(n)}{a(n+1)} u(n) - \frac{1}{a(n+1)} a(n) u(n-1) - \frac{z-\tilde{b}(n)}{\tilde{a}(n+1)} \tilde{u}(n) + \frac{1}{\tilde{a}(n+1)} \tilde{a}(n) \tilde{u}(n-1) \\
a(n+1) u(n) - \tilde{a}(n+1) \tilde{u}(n)
\end{pmatrix} \\
& = \underbrace{\begin{pmatrix}
\frac{z-b(n)}{a(n+1)} & - \frac{1}{a(n+1)} \\
a(n+1) & 0
\end{pmatrix}}_{=A(n)}
\begin{pmatrix}
u(n) - \tilde{u}(n)\\a(n)u(n-1) - \tilde{a}(n)\tilde{u}(n-1)
\end{pmatrix} \\&\quad
+ \underbrace{\begin{pmatrix}
\left(\frac{z-b(n)}{a(n+1)} - \frac{z-\tilde{b}(n)}{\tilde{a}(n+1)}\right) & \left(- \frac{1}{a(n+1)} +\frac{1}{\tilde{a}(n+1)}\right) \\
a(n+1) - \tilde{a}(n+1) & 0
\end{pmatrix}}_{=B(n)}
\begin{pmatrix} \tilde{u}(n) \\ \tilde{a}(n)\tilde{u}(n-1)\end{pmatrix}.
\end{align*}
Thus,
\begin{align*}
& \norm{\begin{pmatrix}
u(n+1) - \tilde{u}(n+1)\\a(n+1)u(n) - \tilde{a}(n+1)\tilde{u}(n)
\end{pmatrix}} \\
& \leq \norm{A(n)} \norm{\begin{pmatrix}
u(n) - \tilde{u}(n)\\a(n)u(n-1) - \tilde{a}(n)\tilde{u}(n-1)
\end{pmatrix}} + \norm{B(n)}\norm{\begin{pmatrix} \tilde{u}(n) \\ \tilde{a}(n)\tilde{u}(n-1)\end{pmatrix}}.
\end{align*}
The Gronwall inequality in Lemma \ref{lem:Gronwall_discrete} yields the assertion for $n> 0$.
For $n\leq 0$ we similarly obtain
\begin{align*}
& \begin{pmatrix}
u(n) - \tilde{u}(n)\\a(n)u(n-1) - \tilde{a}(n)\tilde{u}(n-1)
\end{pmatrix} \\
& = \begin{pmatrix}
u(n) - \tilde{u}(n) \\
-a(n+1) u(n+1) + (z-b(n))u(n) + \tilde{a}(n+1) \tilde{u}(n+1) + (z-\tilde{b}(n))\tilde{u}(n)
\end{pmatrix} \\
& = \begin{pmatrix}
0 & \frac{1}{a(n+1)} \\
-a(n+1) & \frac{z-b(n)}{a(n+1)}
\end{pmatrix}
\begin{pmatrix}
u(n+1) - \tilde{u}(n+1)\\a(n+1)u(n) - \tilde{a}(n+1)\tilde{u}(n)
\end{pmatrix} \\&\quad
+ \begin{pmatrix}
0 & \left( \frac{1}{a(n+1)} -\frac{1}{\tilde{a}(n+1)}\right) \\
-a(n+1) + \tilde{a}(n+1) & \left(\frac{z-b(n)}{a(n+1)} - \frac{z-\tilde{b}(n)}{\tilde{a}(n+1)}\right)
\end{pmatrix}
\begin{pmatrix} \tilde{u}(n+1) \\ \tilde{a}(n+1)\tilde{u}(n)\end{pmatrix} \\
& = A(n)^{-1} \begin{pmatrix}
u(n+1) - \tilde{u}(n+1)\\a(n+1)u(n) - \tilde{a}(n+1)\tilde{u}(n)
\end{pmatrix} + B(n)^{-1} \begin{pmatrix} \tilde{u}(n+1) \\ \tilde{a}(n+1)\tilde{u}(n)\end{pmatrix}.
\end{align*}
Applying again the Gronwall inequality in Lemma \ref{lem:Gronwall_discrete} and taking into account Lemma \ref{lem:norm_unimodular} concludes the proof for $n\leq0$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:Jacobi}]
Let $z\in\mathbb{C}$ with $\abs{z}<\frac{e^{C/2}-\norm{a}_\infty}{\norm{\frac{1}{a}}_\infty}+\norm{b}_\infty-1$, and $u$ a solution of the difference equation
\[a(n+1) u(n+1) + b(n) u(n) + a(n) u(n-1) - z u(n) = 0 \quad(n\in\mathbb{Z}).\]
Assume that $u\neq 0$; then without loss of generality let
\[\norm{\begin{pmatrix}
u(1)\\a(1)u(0)
\end{pmatrix}} = 1.\]
Let $u_m$ be the solution of the corresponding difference equation, where $a$ and $b$ are replaced by the $p_m$-periodic versions $a_m$ and $b_m$ with $a_m(k+1) = a(k+1)$ and $b_m(k) = b(k)$ for $k\in\set{1,\ldots,p_m}$, respectively,
satisfying
\[\begin{pmatrix}
u_m(1)\\a_m(1)u_m(0)
\end{pmatrix} = \begin{pmatrix}
u(1)\\a(1)u(0)
\end{pmatrix} .\]
Since for $n\in\mathbb{Z}$ we can estimate
\[\norm{M_z(n)}_{1\to 1}, \norm{M_z(n)}_{\infty\to\infty} \leq \norm{\frac{1}{a}}_\infty \norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty,\]
we obtain
\begin{align*}
\norm{T_z(n,0)} & \leq \sqrt{\norm{T_z(n,0)}_{1\to1}\norm{T_z(n,0)}_{\infty\to\infty}}\\
& \leq \bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{\abs{n}}.
\end{align*}
Hence, it follows
\[\norm{\begin{pmatrix}
u_m(n+1)\\a_m(n+1)u_m(n)
\end{pmatrix}}\leq \bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{\abs{n}} \quad (n\in\mathbb{Z},m\in\mathbb{N}).\]
We now apply Lemma \ref{lem:Jacobi_diff} with $\tilde{a} = a_m$, $\tilde{b} = b_m$ and $\tilde{u} = u_m$.
With the notation from this Lemma we obtain (noting that $A(n) = M_z(n)$)
\[\norm{A(k)}\leq \norm{\frac{1}{a}}_\infty \norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty \quad(k\in\mathbb{Z}),\]
and similarly
\[\norm{B(k)} \leq \norm{\frac{1}{a}}_\infty^2\bigl(\abs{z}+\norm{b}_\infty + 1\bigr) + 1 \quad(k\in\mathbb{Z}).\]
Thus, we have
\begin{align*}
& \norm{\begin{pmatrix}
u(n+1) - u_m(n+1)\\a(n+1)u(n) - a_m(n+1)u_m(n)
\end{pmatrix}}\\
& \leq \bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{2\abs{n}}\bigg(\norm{\frac{1}{a}}_\infty^2\bigl(\abs{z}+\norm{b}_\infty + 1\bigr) + 1\bigg)\\
& \qquad \cdot \sum_{k=\min\{n+1,1\}}^{\max\{0,n\}}\bigl(\abs{a(k+1)-a_m(k+1)} + \abs{b(k)-b_m(k)}\bigr) \\
& \leq \bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{2\abs{n}}\bigg(\norm{\frac{1}{a}}_\infty^2\bigl(\abs{z}+\norm{b}_\infty + 1\bigr) + 1\bigg) \abs{n}\\
& \qquad \cdot \max_{k=\min\{n+1,1\},\ldots,\max\{0,n\}}\bigl(\abs{a(k+1)-a_m(k+1)} + \abs{b(k)-b_m(k)}\bigr).
\end{align*}
Thus, for $n\in \{-p_m,\ldots,2p_m\}$ we obtain
\begin{align}
\nonumber
& \norm{\begin{pmatrix}
u(n+1) - u_m(n+1)\\a(n+1)u(n) - a_m(n+1)u_m(n)
\end{pmatrix}}\\\nonumber
& \leq \bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{2p_m}\bigg(\norm{\frac{1}{a}}_\infty^2\bigl(\abs{z}+\norm{b}_\infty + 1\bigr) + 1\bigg)2p_m \\
\label{eq:Jacobi_est}
& \qquad \cdot \max_{k=-p_m+1,\ldots,2p_m}\bigl(\abs{a(k+1)-a_m(k+1)} + \abs{b(k)-b_m(k)}\bigr).
\end{align}
Since $a_m(k+1) = a(k+1)$ and $b_m(k) = b(k)$ for $k\in\set{1,\ldots,p_m}$, we infer
\begin{align*}
& \max_{k=-p_m+1,\ldots,2p_m}\bigl(\abs{a(k+1)-a_m(k+1)} + \abs{b(k)-b_m(k)}\bigr) \\
& = \max_{k=-p_m+1,\ldots,p_m}\bigl(\abs{a(k+1)-a(k+1+p_m)} + \abs{b(k)-b(k+p_m)}\bigr).
\end{align*}
By assumption on $\abs{z}$ we obtain
\[\bigg(\norm{\frac{1}{a}}_\infty\norm{z-b}_\infty + \norm{\frac{1}{a}}_\infty + \norm{a}_\infty\bigg)^{2p_m} < e^{Cp_m}\]
for large $m$, so
the right-hand side in \eqref{eq:Jacobi_est} tends to zero as $m\to \infty$. Hence, there exists $m_0\in\mathbb{N}$, such that for all $m\geq m_0$ we have
\[\norm{\begin{pmatrix}
u(n+1) - u_m(n+1)\\a(n+1)u(n) - a_m(n+1)u_m(n)
\end{pmatrix}} \leq \frac{1}{4} \quad(n\in\{-p_m,\ldots,2p_m\}).\]
Since
\[\max\set{\norm{\begin{pmatrix}
u_m(n+1) \\ a_m(n+1) u_m(n)
\end{pmatrix}};\; n\in\set{-p_m,p_m,2p_m}} \geq \frac{1}{2}\]
for all $m\in\mathbb{N}$ by Lemma \ref{lem:Jacobi_per}, we conclude that
\[\limsup_{\abs{n}\to\infty} \norm{\begin{pmatrix}
u(n+1)\\a(n+1)u(n)
\end{pmatrix}} \geq \frac{1}{4}.\]
Since $\inf_{n\in\mathbb{Z}}\abs{a(n)}>0$ this implies $\limsup_{\abs{n}\to\infty}
\abs{u(n)} > 0$, and therefore $u\notin c_0(\mathbb{Z})$.
\end{proof}
\begin{example}
Let $\tilde{a},\tilde{b}\colon\mathbb{T}\to\mathbb{C}$ be $\beta$-H\"older continuous and $\tilde{a}(x)\neq 0$ for all $x\in\mathbb{T}$.
Let $\alpha>0$ satisfy
\[\abs{\alpha-\frac{p_m}{q_m}} \leq B m^{-q_m} \quad(m\in\mathbb{N})\]
for suitable $B>0$ and a suitable sequence $(\frac{p_m}{q_m})$ in $\mathbb{Q}$. Note that the set of all such numbers $\alpha$ is a dense $G_\delta$ set.
Set $a(n):= \tilde{a}(\alpha\cdot n)$ and $b(n):= \tilde{b}(\alpha\cdot n)$ for all $n\in\mathbb{N}$.
Then by the H\"older continuity and the assumption on $\alpha$ we observe
\begin{align*}
\abs{a(k+1) - a(k+1+q_m)} & \leq c\cdot \mathrm{dist}(\alpha q_m,\mathbb{Z})^\beta \leq
c\bigl(Bq_m m^{-q_m})^\beta \\
& = cB^\beta q_m^\beta e^{-\beta q_m \log m}
\end{align*}
for all $k\in\mathbb{Z}$ and $m\in\mathbb{N}$, and similarly for $b$.
Thus, \eqref{eq:Jacobi_exp} is satisfied for all $C>0$, and therefore $H_{a,b}$ does not have any eigenvalues.
\end{example}
\section{The Sturm-Liouville case}
\label{sec:SL}
We say that
\[
\mu \colon \set{B\subseteq\mathbb{R};\;B \text{ is a bounded Borel set}} \to \mathbb{C}
\]
is a \emph{local measure} if $\mathbbold{1}_K\mu := \mu(\cdot\cap K)$ is a
complex Radon measure for any compact set $K\subseteq \mathbb{R}$.
Then there exist a (unique) nonnegative Radon measure $\nu$ on $\mathbb{R}$ and
a measurable function $\sigma\colon\mathbb{R}\to \mathbb{C}$ such that $\abs{\sigma} = 1$ $\nu$-a.e.\ and
$\mathbbold{1}_K\mu = \mathbbold{1}_K\sigma\nu$ for all compact sets $K\subseteq \mathbb{R}$. The \emph{total variation} of $\mu$ is defined by $\abs{\mu}:=\nu$.
Let $\mathcal{M}_{\loc}(\R)$ be the space of all local measures on $\mathbb{R}$.
A local measure $\mu\in \mathcal{M}_{\loc}(\R)$ is called \emph{uniformly locally bounded} if
\[
\norm{\mu}_\mathrm{unif} := \sup_{x\in\mathbb{R}} \abs{\mu}((x,x+1]) < \infty.
\]
Let $\mathcal{M}_{\loc,\lu}(\R)$ denote the space of all uniformly locally bounded local measures.
The space $\mathcal{M}_{\loc,\lu}(\R)$ naturally extends $L_{1,\mathrm{loc},\mathrm{unif}}(\mathbb{R})$ to measures.
Given
$a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$ and $\mu\in\mathcal{M}_{\loc,\lu}(\R)$ we
consider the
differential operator
\begin{align*}
D(H_{a,\mu}) & := \set{u\in C_0(\mathbb{R})\cap W_{1,\mathrm{loc}}^1(\mathbb{R});\; -\partial
a\partial u + u\mu \in C_0(\mathbb{R})},\\
H_{a,\mu} & := -\partial a\partial u + u\mu
\end{align*}
in $C_0(\mathbb{R})$ (the space of continuous functions on $\mathbb{R}$ converging to $0$ at $\pm\infty$), where the terms are interpreted in the sense of distributions.
\begin{remark}
We will also consider this operator in $L_p(\mathbb{R})$, where $1\leq p<\infty$.
Then $C_0(\mathbb{R})$ has to be replaced by $L_p(\mathbb{R})$.
\end{remark}
\begin{remark}
For $\mu\in\mathcal{M}_{\loc,\lu}(\R)$ and $s,t\in\mathbb{R}$ we write
\[\int_s^t\ldots\,d\mu := \begin{cases}
\int_{(s,t]} \ldots\,d\mu & s<t,\\
0 & s=t,\\
-\int_{(t,s]} \ldots\,d\mu & s>t.
\end{cases}
\]
\end{remark}
\begin{definition}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$, $z\in\mathbb{C}$. We say that $u\in L_{1,\mathrm{loc}}(\mathbb{R})$ is a solution of
\[H_{a,\mu}u = zu,\]
if $u\in W_{1,\mathrm{loc}}^1(\mathbb{R})$ and $-\partial a \partial u + u\mu = zu$ in the
sense of distributions.
\end{definition}
\begin{lemma}
\label{lem:est_solutions}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$, $z\in\mathbb{C}$, $u$ a solution of $H_{a,\mu} u = zu$, $I\subseteq\mathbb{R}$ an
interval of length $1$, $1\leq p<\infty$. Then
\[\norm{(au')|_I}_p\leq \norm{(au')|_I}_\infty\leq M\norm{u|_I}_\infty \leq
(p+1)^{1/p}M^{(p+1)/p}\norm{u|_I}_p,\]
where $M = (2\norm{a}_\infty+\norm{\mu-z\lambda}_{\mathrm{unif}})$. In particular,
$u\in L_p(\mathbb{R})$ implies $au'\in L_p(\mathbb{R})$ and $u(x)\to 0$ as $\abs{x}\to\infty$
implies $(au')(x\rlim)\to 0$ as $\abs{x}\to\infty$.
\end{lemma}
\begin{proof}
The first estimate is trivial. For the second estimate note that
\[\abs{\int_{I} u'(t\rlim)\, dt} \leq 2\norm{u|_I}_\infty,\]
so there exists $t_0\in I$ such that $\abs{u'(t_0\rlim)}\leq 2\norm{u|_I}_\infty$.
For $t\in I$ we compute
\[\abs{(au')(t\rlim)}\leq \abs{(au')(t_0\rlim)} + \abs{\int_{t_0}^t u(s)\, d(\mu-z\lambda)(s)} \leq (2\norm{a}_{\infty}+\norm{\mu-z\lambda}_\mathrm{unif})\norm{u|_I}_\infty\]
which proves the second inequality.
Now, choose $s_0\in I$ such that $\abs{u(s_0)} = \norm{u|_I}_\infty$. Then
\[\abs{u(s_0+t)} = \abs{u(s_0) + \int_{s_0}^{s_0+t} u'(s)\,ds} \geq (1 - \abs{t} (2\norm{a}_\infty+\norm{\mu-z\lambda}_\mathrm{unif}))\norm{u|_I}_\infty\]
for all $t\in\mathbb{R}$ with $s_0+t\in I$. Hence, we conclude
\[\norm{u|_I}_p^p \geq \int_0^{1/M} (Mt)^p\norm{u|_I}_\infty^p\, dt =
\frac{1}{(p+1)M}\norm{u|_I}_\infty^p. \qedhere\]
\end{proof}
\begin{remark}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$.
There is also a unique realization of $-\partial a\partial + \mu$ via
Sturm-Liouville theory, cf.\ \cite{EckhardtTeschl2011}.
To this end, for $u\in W_{1,\mathrm{loc}}^1(\mathbb{R})$ we define $A_{a,\mu} u\in
L_{1,\mathrm{loc}}(\mathbb{R})$ by
\[(A_{a,\mu}u)(t) := (au')(t\rlim) - \int_0^t u(s)\, d\mu(s)\]
for a.a.\ $t\in\mathbb{R}$.
Define
\begin{align*}
D(H_{a,\mu}^{\mathrm{SL}}) & := \set{u\in C_0(\mathbb{R})\cap
W_{1,\mathrm{loc}}^1(\mathbb{R});\;A_{a,\mu}u\in W_{1,\mathrm{loc}}^1(\mathbb{R}),\, (A_{a,\mu}u)'\in
C_0(\mathbb{R})},\\
H_{a,\mu}^{\mathrm{SL}}u & := -(A_{a,\mu}u)'.
\end{align*}
\end{remark}
\begin{proposition}
\label{prop:Def_equal}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$. Then
$H_{a,\mu} = H_{a,\mu}^{\mathrm{SL}}$.
\end{proposition}
\begin{proof}
Let $u\in D(H_{a,\mu})$. Then $u\in W_{2}^1(\mathbb{R})\subseteq W_{1,\mathrm{loc}}^1(\mathbb{R})$ and $A_{a,\mu}u \in L_{1,\mathrm{loc}}(\mathbb{R})$. Let $\varphi\in C_c^\infty(\mathbb{R})$.
By Fubini's Theorem we observe
\begin{align*}
\int (A_{a,\mu}u)\varphi' & = \int \bigg(au' - \int_0^xu(t)\,d\mu(t)\bigg)\varphi'(x)\, dx\\
& = \int au'\varphi' - \int_\mathbb{R} \int_0^xu(t)\,d\mu(t)\varphi'(x)\, dx\\
& = \int au'\varphi' + \int_{-\infty}^0 \!\int_{(-\infty,t)} \!\!\varphi'(x)\, dx u(t)\,d\mu(t) - \int_0^\infty \!\int_{[t,\infty)}\!\!\varphi'(x)\,dx u(t)\, d\mu(t)\\
& = \int au'\varphi' + \int u\varphi\,d\mu
= \int H_{a,\mu}u\varphi.
\end{align*}
Hence, $(A_{a,\mu}u)' = -H_{a,\mu}u \in C_0(\mathbb{R})$. Therefore, also
$A_{a,\mu}u\in W_{1,\mathrm{loc}}^1(\mathbb{R})$ which implies $u\in D(H_{a,\mu}^{\mathrm{SL}})$,
$H_{a,\mu} u = H_{a,\mu}^{\mathrm{SL}}u$.
Conversely, let $u\in D(H_{a,\mu}^\mathrm{SL})$.
For $\varphi\in C_c^\infty(\mathbb{R})$
we compute by the help of Fubini's Theorem
\begin{align*}
\int -(A_{a,\mu}u)' \varphi & = \int (A_{a,\mu}u) \varphi'
= \int au' \varphi' - \int \int_0^t u(s)\,d\mu(s) \varphi'(t)\,dt \\
& = \int au' \varphi' + \int u\varphi\, d\mu.
\end{align*}
Thus, $u\in D(H_{a,\mu})$, $H_{a,\mu}u = -(A_{a,\mu}u)'$.
\end{proof}
\begin{remark}
Proposition \ref{prop:Def_equal} remains true if we consider the operators
in $L_p(\mathbb{R})$ with $1\leq p<\infty$. Furthermore, considering the operators
in $L_2(\mathbb{R})$, if additionally $a$ takes values only in a sector around the
positive real axis, we obtain an equivalent characterization of
$H_{a,\mu}$ via sectorial forms; cf.\ \cite[Remark 3.5 and Theorem
3.6]{SeifertVogt2014} in case of Schr\"odinger operators.
\end{remark}
Since the operator is now defined, we will next focus on measuring distances of
elements in $\mathcal{M}_{\loc,\lu}(\R)$.
\begin{definition}
For $\mu \in \mathcal{M}_{\loc,\lu}(\R)$ and a set $I \subseteq \mathbb{R}$ (which will usually be an interval) we define
\[
\norm{\mu}_I := \sup\set{\Bigl|\int u\, d\mu\Bigr|;\;
u\in W_{\!\infty}^1(\mathbb{R}),\: \spt u \subseteq I,\: \diam\spt u \leq 2,\: \norm{u'}_\infty \le 1 }.
\]
\end{definition}
For $\mu\in\mathcal{M}_{\loc,\lu}(\R)$ we define $\phi_\mu\colon\mathbb{R}\to\mathbb{C}$ by
\[
\phi_\mu(t) := \int_0^t d\mu
= \begin{cases}
\mu\bigl((0,t]\bigr) & \text{ if } t\ge0, \\
-\mu\bigl((t,0]\bigr) & \text{ if } t<0.
\end{cases}
\]
\begin{proposition}[see {\cite[Proposition 2.7, Remark 2.8 and Lemma 2.9]{SeifertVogt2014}}]
Let $\mu\in\mathcal{M}_{\loc,\lu}(\R)$ and $x\in\mathbb{R}$. Then
\[\norm{\mu}_{[x-1,x+1]} \leq \min_{c\in\mathbb{C}} \int_{x-1}^{x+1} \abs{\phi_\mu(t)-c}\,dt \leq 2\norm{\mu}_{[x-1,x+1]}.\]
Hence, there exists $c_x=c_{\mu,x}\in\mathbb{C}$, such that
\[\int_{x-1}^{x+1} \abs{\phi_\mu(t)-c_x}\,dt \leq 2\norm{\mu}_{[x-1,x+1]}.\]
Moreover, $c_{\mu,0}$ can be chosen such that $\abs{c_{\mu,0}}\leq \norm{\mu}_\mathrm{unif}$.
Furthermore, for $\alpha,\beta\in\mathbb{Z}$, $\alpha\leq -1$, $\beta\geq1$ and $k\in\mathbb{Z}\cap[\alpha,\beta-1]$ we have
\[\int_k^{k+1} \abs{\varphi_\mu(t)-c_0}\,dt\leq 2\max\{k+1,-k\}\norm{\mu}_{[\alpha,\beta]}.\]
\end{proposition}
We can now state Gordon's Theorem for Sturm-Liouville operators.
\begin{theorem}
\label{thm:SL}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$ and $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$.
Assume there exists $(p_m)$ in $(0,\infty)$ such that $p_m\to \infty$ and $C>0$ such that
\begin{equation}
\label{eq:SL_exp}
e^{Cp_m} \left(\norm{a - a(\cdot-p_m)}_{L_1(-p_m,p_m)} +
\norm{\mu-\mu(\cdot+p_m)}_{[-p_m,p_m]}\right)\to 0.
\end{equation}
Then $H_{(a,\mu)}$ does not have any eigenvalues with modulus less than $C^2\norm{\frac{1}{a}}_\infty - \norm{\mu}_\mathrm{unif}$.
\end{theorem}
As in the discrete case condition \eqref{eq:SL_exp} states that restrictions of $a$ and $\mu$ to the three pieces $[-p_m,0]$, $[0,p_m]$ and $[p_m,2p_m]$
do not differ to much (indeed, the difference tends to zero faster than a given exponential).
\begin{corollary}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$ and $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$.
Assume that \eqref{eq:SL_exp} is satisfied for all $C>0$. Then $H_{(a,\mu)}$ does not have any eigenvalues.
\end{corollary}
Taking into account the estimate in Lemma \ref{lem:est_solutions} we see that
an $L_p(\mathbb{R})$-eigenfunction for the $L_p(\mathbb{R})$ operator is in fact an
$C_0(\mathbb{R})$-eigenfunction for the $C_0(\mathbb{R})$ operator. Hence, we obtain the
following corollary.
\begin{corollary}
Let $1\leq p <\infty$, $a\in L_\infty(R)$ such that $\frac{1}{a}\in
L_\infty(\mathbb{R})$, $\mu\in\mathcal{M}_{\loc,\lu}(\R)$. Assume that \eqref{eq:SL_exp} holds true. Then
$H_{a,\mu}$, considered as an operator in $L_p(\mathbb{R})$, does not have any
eigenvalues with modulus less than $C^2\norm{\frac{1}{a}}_\infty -
\norm{\mu}_\mathrm{unif}$. In case \eqref{eq:SL_exp} holds true for all $C>0$,
$H_{a,\mu}$ does not have any $L_p(\mathbb{R})$-eigenvalues.
\end{corollary}
Again, we provide several lemmas for the proof of the theorem.
\begin{lemma}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$, $z\in\mathbb{C}$.
The following are equivalent:
\begin{enumerate}
\item
$u$ is a solution of the equation $H_{a,\mu}u = zu$.
\item
For $s,t\in\mathbb{R}$ we have
\[\begin{pmatrix}
u(t) \\
(au')(t\rlim)
\end{pmatrix} = T_z(t,s)\begin{pmatrix}
u(s) \\
(au')(s\rlim)
\end{pmatrix},\]
where
\[T_z(t,s) = \begin{pmatrix}
\u@DN{\Neu}(t;s) & \u@DN{\Dir}(t;s)\\
(a\u@DN{\Neu}'(\cdot;s))(t\rlim) & (a\u@DN{\Dir}'(\cdot;s))(t\rlim)
\end{pmatrix}\]
and $\u@DN{\Neu}(\cdot;s)$, $\u@DN{\Dir}(\cdot;s)$ are the (Neumann and Dirichlet) solution(s) of $Hu = zu$ satisfying
\begin{align*}
\u@DN{\Neu}(s;s) & = 1 & \u@DN{\Dir}(s;s) & = 0 & \\
(a\u@DN{\Neu}'(\cdot;s))(s\rlim) & = 0 & (a\u@DN{\Dir}'(\cdot;s))(s\rlim) & = 1 & \\
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
``(a)$\Rightarrow$(b)'': Fix $s,t\in\mathbb{R}$ and let
\[T_z(t,s)\colon \begin{pmatrix}
u(s) \\
(au')(s\rlim)
\end{pmatrix}\mapsto \begin{pmatrix}
u(t) \\
(au')(t\rlim)
\end{pmatrix}.\]
Then $T_z(t,s)$ is linear and can be represented by a matrix, which we will also denote by $T_z(t,s)$.
By the initial conditions for the Neumann and Dirichlet solution we observe
\begin{align*}
T_z(t,s) & = T_z(t,s)\begin{pmatrix} 1 & 0\\0 & 1\end{pmatrix} = T_z(t,s) \begin{pmatrix}
\u@DN{\Neu}(s;s) & \u@DN{\Dir}(s;s)\\
(a\u@DN{\Neu}'(\cdot;s))(s\rlim) & (a\u@DN{\Dir}'(\cdot;s))(s\rlim)
\end{pmatrix}\\
& = \begin{pmatrix}
\u@DN{\Neu}(t;s) & \u@DN{\Dir}(t;s)\\
(a\u@DN{\Neu}'(\cdot;s))(t\rlim) & (a\u@DN{\Dir}'(\cdot;s))(t\rlim)
\end{pmatrix}.
\end{align*}
``(b)$\Rightarrow$(a)'': For $s,t\in\mathbb{R}$ we have
\begin{align*}
u(t) & = \u@DN{\Neu}(t;s)\cdot u(s) + \u@DN{\Dir}(t;s)\cdot (au')(s\rlim),\\
(au')(t\rlim) & = (a\u@DN{\Neu}'(\cdot;s))(t\rlim)\cdot u(s) + (a\u@DN{\Dir}'(\cdot;s))(t\rlim)\cdot (au')(s\rlim).
\end{align*}
Differentiating the second equality, taking into account that $\u@DN{\Neu}(\cdot;s)$ and $\u@DN{\Dir}(\cdot;s)$ are solutions and noting the first equality yields
\begin{align*}
-(au')' & = -(a\u@DN{\Neu}'(\cdot;s))' \cdot u(s) - (a\u@DN{\Dir}'(\cdot;s))'\cdot (au')(s\rlim) \\
& = z\bigl(\u@DN{\Neu}(\cdot;s)\cdot u(s) + \u@DN{\Dir}(\cdot;s)\cdot (au')(s\rlim)\bigr) \\
& \quad - \bigl(\u@DN{\Neu}(\cdot;s)\cdot u(s) + \u@DN{\Dir}(\cdot;s)\cdot (au')(s\rlim)\bigr)\mu \\
& = zu - u\mu.
\end{align*}
Hence, $u$ is a solution of $H_{a,\mu} u = zu$.
\end{proof}
\begin{lemma}
\label{lem:SL_est1}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$, $u$ a solution of $H_{a,\mu} u = 0$. Then
\[\abs{u(t)} + \abs{(au')(t\rlim)} \leq \bigl(\abs{u(0)} + \abs{(au')(0\rlim)}\bigr)e^{(\norm{\frac{1}{a}}_\infty+\norm{\mu}_\mathrm{unif})(\abs {t}+1)} \quad(t\in\mathbb{R}).\]
\end{lemma}
\begin{proof}
Writing
\begin{align*}
u(t) & = u(0) + \int_0^t (au')(s\rlim)\frac{1}{a(s)}\,ds,\\
(au')(t\rlim) & = (au')(0\rlim) + \int_0^t u(s)\, d\mu(s),
\end{align*}
we obtain for $\varphi(t):= \abs{u(t)} + \abs{(au')(t\rlim)}$ and $\nu:=\frac{1}{a}\lambda + \abs{\mu}$ the inequality
\[\varphi(t)\leq \varphi(0) + \int_{(t,0]} \varphi(s)\, d\nu(s) \quad(t\leq 0).\]
By Gronwall's inequality (see Lemma \ref{lem:Gronwall_cont}) we infer
\[\varphi(t)\leq \varphi(0)e^{\nu((t,0])} \quad(t\leq 0).\]
Since $\norm{\nu}_\mathrm{unif}\leq \norm{\frac{1}{a}}_\infty + \norm{\mu}_\mathrm{unif}$ and $\nu((t,0]) \leq \norm{\nu}_\mathrm{unif}(\abs{t}+1)$, we obtain the assertion for $t\leq 0$.
For $t>0$ we set
\[\varphi_-(s):= \abs{u(s)} + \abs{(au')(s\llim)} \leq \varphi(0) + \int_{(0,s)} \varphi_-(r)\,d\nu(r).\]
The Gronwall's inequality in Lemma \ref{lem:Gronwall_cont} yields
\[\abs{u(s)} + \abs{(au')(s\llim)} = \varphi_-(s) \leq \varphi(0)e^{\nu((0,s))} = \bigl(\abs{u(0)} + \abs{(au')(0\rlim)}\bigr)e^{\nu((0,s))}.\]
For $s\downarrow t$ we the assertion follows, since $\nu((0,t])\leq \norm{\nu}_\mathrm{unif}(\abs{t}+1)$.
\end{proof}
\begin{lemma}
\label{lem:SL-diff1}
Let $a,\tilde{a}\in L_\infty(\mathbb{R})$ such that
$\frac{1}{a},\frac{1}{\tilde{a}}\in L_\infty(\mathbb{R})$, $\mu,\tilde{\mu}\in\mathcal{M}_{\loc,\lu}(\R)$,
and $u$ and $\tilde{u}$ two solutions of
$H_{a,\mu} u = 0$ and $H_{\tilde{a},\tilde{\mu}}\tilde{u} = 0$, respectively,
satisfying
\[\begin{pmatrix} u(0)\\(au')(0\rlim)\end{pmatrix} = \begin{pmatrix} \tilde{u}(0)\\(\tilde{a}\tilde{u}')(0\rlim)\end{pmatrix}.\]
Then, for $s,t\in\mathbb{R}$ we have
\begin{align*}
& \begin{pmatrix} u(t)-\tilde{u}(t)\\(au')(t\rlim) - (\tilde{a}\tilde{u}')(t\rlim)\end{pmatrix} \\
& = T_{\mu}(t,s) \begin{pmatrix} u(s)-\tilde{u}(s)\\(au')(s\rlim) - (\tilde{a}\tilde{u}')(s\rlim)\end{pmatrix}
+ \int_s^t T_\mu(t,r)\begin{pmatrix}0 \\\tilde{u}(r)\end{pmatrix} \,d(\mu-\tilde{\mu})(r) \\
& \quad + \int_s^t \bigg(\tfrac{1}{a(r)} - \tfrac{1}{\tilde{a}(r)}\bigg) T_\mu(t,r) \begin{pmatrix} (\tilde{a}\tilde{u}')(r\rlim)\\ 0\end{pmatrix}\,dr.
\end{align*}
\end{lemma}
\begin{proof}
Without loss of generality, let $s=0$. Integrating by parts, we obtain
\begin{align*}
& \int_0^t T_\mu(r,0)^{-1}\begin{pmatrix} 0 \\ \tilde{u}(r)\end{pmatrix}\, d(\mu-\tilde{\mu})(r)
= \begin{pmatrix}
- \int_0^t \u@DN{\Dir}(r)\tilde{u}(r)\,d(\mu-\tilde{\mu})(r) \\
\int_0^t \u@DN{\Neu}(r)\tilde{u}(r)\,d(\mu-\tilde{\mu})(r)
\end{pmatrix} \\
& = \begin{pmatrix} u(0)\\(au')(0\rlim)\end{pmatrix} - T_\mu(0,t) \!\begin{pmatrix} \tilde{u}(t) \\(\tilde{a}\tilde{u}')(t\rlim)\end{pmatrix} - \int_0^t \bigl(\tfrac{1}{a(r)} \!-\! \tfrac{1}{\tilde{a}(r)}\bigr) T_\mu(0,r) \begin{pmatrix} (\tilde{a}\tilde{u}')(r\rlim)\\ 0\end{pmatrix}dr.
\end{align*}
Multiplying by $T_\mu(t,0)$ yields the assertion, since $T_\mu(t,0) T_\mu(r,0)^{-1} = T_\mu(t,r)$.
\end{proof}
\begin{lemma}
\label{lem:SL_diff2}
Let $a,\tilde{a}\in L_\infty(\mathbb{R})$ such that
$\frac{1}{a},\frac{1}{\tilde{a}}\in L_\infty(\mathbb{R})$, $\mu,\tilde{\mu}\in\mathcal{M}_{\loc,\lu}(\R)$,
$c\in\mathbb{C}$ and $u$ and $\tilde{u}$ two solutions of
$H_{a,\mu} u = 0$ and $H_{\tilde{a},\tilde{\mu}}\tilde{u} = 0$, respectively,
satisfying
\[\begin{pmatrix} u(0)\\(au')(0\rlim)\end{pmatrix} = \begin{pmatrix} \tilde{u}(0)\\(\tilde{a}\tilde{u}')(0\rlim) - c_{\mu-\tilde{\mu},0}u(0)\end{pmatrix}.\]
Let $\alpha,\beta\in\mathbb{Z}$, $\alpha\leq -1$, $\beta\geq 1$. Let $c,\omega>0$ such that
\[\abs{\u@DN{\Neu}(t,s)}, \abs{\partial_1 \u@DN{\Dir}(t\rlim,s)} \leq ce^{\omega\abs{t-s}} \quad(s,t\in\mathbb{R}).\]
Then there exists a constant $C>0$ depending only on $\omega$ and $\norm{\tilde{\mu}}_\mathrm{unif}$ such that
\[\abs{u(t) - \tilde{u}(t)} \leq Cce^{\omega \abs{t}}
\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty
\bigl(\norm{(a-\tilde{a})|_{[\alpha,\beta]}}_{L_1(\mathbb{R})}+\norm{\mu-\tilde{\mu}}_{
[\alpha,\beta]}\bigr) \quad(t\in[\alpha,\beta]).\]
\end{lemma}
\begin{proof}
By Lemma \ref{lem:SL-diff1} we obtain
\begin{align*}
u(t) - \tilde{u}(t) & = -\u@DN{\Dir}(t) c_{\mu-\tilde{\mu},0} u(0) + \int_0^t \u@DN{\Dir}(t,r)\tilde{u}(r)\,d(\mu-\tilde{\mu})(r) \\
& \quad + \int_0^t \bigg(\frac{1}{a(r)} - \frac{1}{\tilde{a}(r)}\bigg)\u@DN{\Neu}(t,r)(\tilde{a}\tilde{u}')(r\rlim)\,dr.
\end{align*}
Since $\u@DN{\Dir}(t,t) = 0$ we have
\[\u@DN{\Dir}(t,r)\tilde{u}(r) = -\int_r^t \frac{d}{ds} \bigl(\u@DN{\Dir}(t,s)\tilde{u}(s)\bigr)\, ds.\]
Fubini's Theorem then implies
\begin{align*}
u(t) - \tilde{u}(t) & = \int_0^t \bigl(c_{\mu-\tilde{\mu},0} - \varphi_{\mu-\tilde{\mu}}(s)\bigr)\frac{d}{ds} \bigl(\u@DN{\Dir}(t,s)\tilde{u}(s)\bigr)\, ds \\
& \quad + \int_0^t \bigg(\frac{1}{a(r)} - \frac{1}{\tilde{a}(r)}\bigg)\u@DN{\Neu}(t,r)(\tilde{a}\tilde{u}')(r\rlim)\,dr.
\end{align*}
We now estimate $\frac{d}{ds} \bigl(\u@DN{\Dir}(t,s)\tilde{u}(s)\bigr)$. From $\u@DN{\Dir}(s,s) = 0$ and the assumed bound on $\abs{\partial_1 \u@DN{\Dir}(t\rlim,s)}$ we obtain $\abs{\u@DN{\Dir}(t,s)} \leq \frac{c}{\omega}e^{\omega\abs{t-s}}$ for all $s,t\in\mathbb{R}$.
Furthermore, by Lemma \ref{lem:est_solutions} we have $\norm{(\tilde{a}\tilde{u}')|_{[\alpha,\beta]}}_{\infty} \leq (2\norm{a}_\infty + \norm{\tilde{\mu}}_\mathrm{unif})\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty$.
Noting that $\u@DN{\Dir}(t,s) = -\u@DN{\Dir}(s,t)$ we hence obtain
\begin{align*}
\abs{\frac{\partial^+}{\partial s} \bigl(\u@DN{\Dir}(t,s)\tilde{u}(s)\bigr)} & = \abs{-\partial_1 \u@DN{\Dir}(s\rlim,t)\tilde{u}(s) + \u@DN{\Dir}(t,s)\tilde{u}'(s\rlim)} \\
& \leq C_0ce^{\omega\abs{t-s}}\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty \quad(s,t\in[\alpha,\beta]),
\end{align*}
with $C_0=\norm{\frac{1}{\tilde{a}}}_\infty\bigl(1+\frac{1}{\omega}(2\norm{a}_\infty + \norm{\tilde{\mu}}_\mathrm{unif})\bigr)$.
For $t\in[0,\beta]$ we therefore estimate
\begin{align*}
&\abs{u(t) - \tilde{u}(t)}\\
& \leq \int_0^t \abs{c_{\mu-\tilde{\mu},0} - \varphi_{\mu-\tilde{\mu}}(s)}\abs{\frac{d}{ds} \bigl(\u@DN{\Dir}(t,s)\tilde{u}(s)\bigr)} \, ds \\
& \quad + \int_0^t \abs{\tfrac{1}{a(r)} - \tfrac{1}{\tilde{a}(r)}} \abs{\u@DN{\Neu}(t,r)(\tilde{a}\tilde{u}')(r\rlim)} \,dr \\
& \leq C_0 c \norm{\tilde{u}|_{[\alpha,\beta]}}_\infty \sum_{k=1}^\beta \int_{k-1}^k e^{\omega(t-s)} \abs{c_{\mu-\tilde{\mu},0} - \varphi_{\mu-\tilde{\mu}}(s)}\, ds \\
& \quad + \norm{\frac{1}{a}}_\infty\norm{\frac{1}{\tilde{a}}}_\infty
(2\norm{\tilde{a}}_\infty + \norm{\tilde{\mu}}_\mathrm{unif})
\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty c e^{\omega t} \int_0^t
\norm{\tilde{a}(r)-a(r)}\, dr \\
& \leq C_0 c \norm{\tilde{u}|_{[\alpha,\beta]}}_\infty \sum_{k=1}^\beta e^{\omega(t+1-k)} 2k\norm{\mu-\tilde{\mu}}_{[\alpha,\beta]} \\
& \quad + \norm{\frac{1}{a}}_\infty\norm{\frac{1}{\tilde{a}}}_\infty
(2\norm{\tilde{a}}_\infty + \norm{\tilde{\mu}}_\mathrm{unif})
\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty c e^{\omega t} \int_0^t
\norm{\tilde{a}(r)-a(r)}\, dr \\
& \leq Cc\norm{\tilde{u}|_{[\alpha,\beta]}}_\infty e^{\omega t}
\bigl(\norm{\mu-\tilde{\mu}}_{[\alpha,\beta]} +
\norm{(a-\tilde{a})|_{[\alpha,\beta]}}_{L_1(\mathbb{R})}\bigr)
\end{align*}
where $C = C_0 \sum_{k=1}^\infty 2ke^{-\omega(k+1)} +
\norm{\frac{1}{a}}_\infty\norm{\frac{1}{\tilde{a}}}_\infty
(2\norm{\tilde{a}}_\infty + \norm{\tilde{\mu}}_\mathrm{unif})c$.
The proof for the case $t\in[\alpha,0)$ is analogous.
\end{proof}
\begin{lemma}
\label{lemm:SL_est2}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in
\mathcal{M}_{\loc,\lu}(\R)$, $u$ a solution of $H_{a,\mu} u = 0$, $\omega:=
\bigl(\norm{\mu}_\mathrm{unif}\norm{\frac{1}{a}}_\infty^{-1}\bigr)^{1/2}$.
Then
\[\bigl(\omega^2 \abs{u(t)}^2 + \abs{(au')(t\rlim)}^2\bigr)^{1/2} \leq \bigl(\omega^2 \abs{u(0)}^2 + \abs{(au')(0\rlim)}^2\bigr)^{1/2} e^{\omega(\abs{t}+1/2)} \quad(t\in\mathbb{R}).\]
\end{lemma}
\begin{proof}
Without loss of generality, let $\mu\neq 0$ (the case $\mu=0$ is trivial).
(i)
We first assume that $\mu = \rho\lambda$ with a density $\rho\in C(\mathbb{R})$. Then $au'\in C^1(\mathbb{R})$ and $(au')' = \rho u$. Let $\phi(t):= \omega^2\abs{u(t)}^2 + \abs{(au')(t)}^2$. Then
\begin{align*}
\abs{\phi'(t)} & = \abs{2\Re\bigl(\bigl(\tfrac{\omega}{\overline{a(t)}} + \abs{\rho(t)}\bigr)u(t) \overline{(au')(t)}\bigr)} \leq \bigl(\tfrac{\omega}{\abs{a(t)}} + \tfrac{\abs{\rho(t)}}{\omega}\bigr)\varphi(t).
\end{align*}
Hence, $\varphi(t)\leq \varphi(s)\exp(\omega\norm{\frac{1}{a}}_\infty\abs{t-s} + \frac{1}{\omega}\int_s^t\rho(r)\,dr)$ and therefore
\[\bigl(\omega^2 \abs{u(t)}^2 + \abs{(au')(t\rlim)}^2\bigr)\leq \bigl(\omega^2 \abs{u(s)}^2 + \abs{(au')(s)}^2\bigr)e^{\omega\norm{\tfrac{1}{a}}_\infty\abs{t-s} + \frac{1}{\omega}\abs{\mu}([s,t])}\]
for all $s,t\in\mathbb{R}$, $s<t$.
(ii)
By \cite[Proposition 2.5]{SeifertVogt2014} there exists $(\mu_n)$ in $\mathcal{M}_{\loc,\lu}(\R)$ such that $\mu_n$ has a smooth density and $\norm{\mu_n}_\mathrm{unif}\leq \norm{\mu}_\mathrm{unif}$ for all $n\in\mathbb{N}$, $\norm{\mu_n-\mu}_\mathbb{R}\to 0$ and $\limsup_{n\to\infty} \abs{\mu_n}(I) \leq \abs{\mu}(I)$ for all compact intervals $I\subseteq \mathbb{R}$.
Then \cite[Lemma 2.4]{SeifertVogt2014} implies $\mathbbold{1}_[\alpha,\beta]\mu_n\to \mathbbold{1}_{[\alpha,\beta]}\mu$ weakly for all $\alpha,\beta\in\mathbb{R}$ such that $\mu(\{\alpha\}) = \mu(\{\beta\}) = 0$.
(iii)
For $n\in\mathbb{N}$ let $u_n$ be the solution of $H_{a,\mu_n}u_n = 0$ such that
$u_n(0) = u(0)$, $(au_n)'(0\rlim) = (au')(0\rlim) + c_{\mu-\mu_n,0}u(0)$.
By Lemma \ref{lem:SL_est1}, $(u_n)$ is uniformly bounded on any compact interval, so Lemma \ref{lem:SL_diff2} implies $u_n\to u$ locally uniformly. Hence, for $s,t\in\mathbb{R}$ with $\mu(\{s\}) = \mu(\{t\}) = 0$ we obtain
\[(au_n')(t) - (au_n')(s) = \int_s^t u_n(r) \,d\mu_n(r) \to \int_s^t u(r)\,d\mu(r) = (au')(t\rlim) - (au')(s\rlim).\]
By Lemma \ref{lem:est_solutions} also $(au_n')$ is uniformly bounded on $[0,1]$, so dividing by $a(s)$ and integration with respect to $s$ yields
\begin{align*}
(au_n')(t)\int_0^1 \frac{1}{a(s)}\,ds - \bigl(u_n(1) - u_n(0)\bigr) & \to (au')(t\rlim)\int_0^1 \frac{1}{a(s)}\,ds - \bigl(u(1) - u(0)\bigr),
\end{align*}
so $(au_n')(t) \to (au')(t\rlim)$.
(iv)
Let $t>s>0$ such that $\mu(\{s\}) = \mu(\{t\}) = 0$. By (i) we have
\[\bigl(\omega^2 \abs{u_n(t)}^2 + \abs{(au_n')(t)}^2\bigr)\leq \bigl(\omega^2 \abs{u_n(s)}^2 + \abs{(au_n')(s)}^2\bigr)e^{\omega\norm{\tfrac{1}{a}}_\infty\abs{t-s} + \frac{1}{\omega}\abs{\mu_n}([s,t])}.\]
Taking the limit $n\to\infty$ noting (ii) we obtain
\[\bigl(\omega^2 \abs{u(t)}^2 + \abs{(au')(t\rlim)}^2\bigr)\leq \bigl(\omega^2 \abs{u(s)}^2 + \abs{(au')(s\rlim)}^2\bigr)e^{\omega\norm{\tfrac{1}{a}}_\infty\abs{t-s} + \frac{1}{\omega}\abs{\mu}([s,t])}.\]
(v)
For $t>0$ there exist sequences $s_n \in [0,t)$ and $(t_n)$ in $[t,\infty)$
such that $s_n\to 0$, $t_n\to t$ and $\mu(\{s_n\}) = \mu(\{t_n\}) = 0$ for
all $n\in\mathbb{N}$.
Thus, from (iv) we deduce
\[\bigl(\omega^2 \abs{u(t)}^2 + \abs{(au')(t\rlim)}^2\bigr)\leq \bigl(\omega^2 \abs{u(0)}^2 + \abs{(au')(0\rlim)}^2\bigr)e^{\omega\norm{\tfrac{1}{a}}_\infty\abs{t} + \frac{1}{\omega}\abs{\mu}((0,t])}.\]
Plugging in $\omega =
\bigl(\norm{\mu}_\mathrm{unif}\norm{\frac{1}{a}}_\infty^{-1}\bigr)^{1/2}$ yields the
assertion for $t\geq 0$. The case $t<0$ is proved analogously.
\end{proof}
We can proceed with the continuum version of Lemma \ref{lem:Jacobi_per}.
The proof is analogous to the discrete case in Lemma \ref{lem:Jacobi_per}, so we omit it here.
\begin{lemma}
\label{lem:SL_per}
Let $a\in L_\infty(\mathbb{R})$ such that $\frac{1}{a}\in L_\infty(\mathbb{R})$, $\mu \in \mathcal{M}_{\loc,\lu}(\R)$
and $p>0$ such that $a$ and $\mu$ are $p$-periodic. Let $z\in\mathbb{C}$ and $u$ a
solution of $H_{a,\mu}u = zu$. Then
\[\max\set{\norm{\begin{pmatrix} u(t)\\(au')(t\rlim)\end{pmatrix}};\; t\in\set{-p,0,2p}} \geq \frac{1}{2}\norm{\begin{pmatrix} u(0)\\(au')(0\rlim)\end{pmatrix}}.\]
\end{lemma}
\begin{lemma}[see {\cite[Lemma 5.1]{SeifertVogt2014}}]
\label{lem:equiv_Gordon}
Let $\mu\in\mathcal{M}_{\loc,\lu}(\R)$, $C> 0$. Assume there exists $(p_m)$ in $(0,\infty)$ with $p_m\to \infty$ such that
\[e^{Cp_m} \norm{\mu-\mu(\cdot+p_m)}_{[-p_m,p_m]}\to 0.\]
Then there exists $(\mu_m)$ in $\mathcal{M}_{\loc,\lu}(\R)$ such that
$\mu_m$ is periodic with period $p_m$ ($m\in\mathbb{N}$), and
\[
e^{Cp_m} \norm{\mu-\mu_m}_{[-p_m,2p_m]} \to 0 \qquad (m\to\infty).
\]
Moreover, the measures $\mu_m$ can be chosen such that
\[
\mathbbold{1}_{[\alpha_m,p_m-\alpha_m]} \mu_m = \mathbbold{1}_{[\alpha_m,p_m-\alpha_m]} \mu, \quad
\norm{\mu_m}_\mathrm{unif} \le \bigl(1+\tfrac{1}{2\alpha_m}\bigr)\norm{\mu}_\mathrm{unif}
\]
for all $m\in\mathbb{N}$, with $0 < \alpha_m \le \frac{p_m}{2}$ and $\inf_{m\in\mathbb{N}} \alpha_m>0$.
\end{lemma}
\begin{proof}[Proof of Theorem \ref{thm:SL}]
Without loss of generality, let $p_m\geq 4$ for all $m\in\mathbb{N}$.
Let $(\mu_m)$ and $(\alpha_m)$ as in Lemma \ref{lem:equiv_Gordon} such that $p_m+\alpha_m\in\mathbb{N}$ for all $m\in\mathbb{N}$, $\alpha_m\to \infty$ and $\frac{\alpha_m}{p_m}\to 0$.
Assume that $z\in\mathbb{C}$ with $\abs{z}< C^2\norm{\frac{1}{a}}_\infty - \norm{\mu}_\mathrm{unif}$ is an eigenvalue of $H_{a,\mu}$.
Let $u\neq 0$ be corresponding eigenfunction. Then $u\in C_0(\mathbb{R})$.
For $m\in\mathbb{N}$ let $u_m$ be the solution of $H_{a_m,\mu_m} u_m = zu_m$ satisfying $u_m(\alpha_m) = u(\alpha_m)$, $(a_m u_m')(\alpha_m\rlim) = (au')(\alpha_m')$.
Then $u_m = u$ on $[\alpha_m,p_m-\alpha_m]$, since $\mu_m = \mu$ on this interval. Note that $c_{\mu-\tilde{\mu},\alpha_m+1} = 0$, since $\mathbbold{1}_{[\alpha_m,\alpha_m+2]}(\mu_m-\mu) = 0$.
By Lemma \ref{lem:SL_diff2}, for $t\in[-p_m,\alpha_m]$ we obtain
\[\abs{u(t) - u_m(t)} \leq C_m
e^{\omega_m\abs{t-(\alpha_m+1)}}\bigl(\norm{a-a_m}_{L_1(-p_m,\alpha_m+1)} +
\norm{\mu-\mu_m}_{[-p_m,\alpha_m+1]}\bigr)\]
where $\omega_m = \bigl(\norm{\mu_m-z\lambda}_\mathrm{unif} \norm{\frac{1}{a_m}}_\infty^{-1}\bigr)^{1/2}$ as in Lemma \ref{lemm:SL_est2}, and $C_m$ is only depending on $\omega_m$, $\norm{\frac{1}{a}}_\infty$, $\norm{\frac{1}{a_m}}_\infty$, $\norm{\mu}_\mathrm{unif}$ and $\norm{a}_\infty$, and similarly for $t\in [p_m-\alpha_m,2p_m]$.
Hence,
\begin{align}
\label{eq:SL_conv}
& \sup_{t\in[-p_m,2p_m]} \abs{u(t) - u_m(t)} \nonumber \\
& \leq C_m e^{\omega_m (p_m+\alpha_m+1)}\bigl(\norm{a-a_m}_{L_1(-p_m,2p_m)}
+ \norm{\mu-\mu_m}_{[-p_m,2p_m]}\bigr).
\end{align}
Since $\norm{\frac{1}{a_m}}_\infty^{-1} \leq \norm{\frac{1}{a}}_\infty^{-1}$,
we have
\begin{align*}
\omega_m^2 & = \norm{\frac{1}{a_m}}_\infty^{-1} \bigl(\norm{\mu_m}_\mathrm{unif} +
\abs{z}\bigr) \leq \norm{\frac{1}{a}}_\infty^{-1}
\bigl(\bigl(1+\tfrac{1}{2\alpha_m}\bigr)\norm{\mu}_\mathrm{unif} + \abs{z}\bigr) \\
& \to \norm{\frac{1}{a}}_\infty^{-1} \bigl(\norm{\mu}_\mathrm{unif} + \abs{z}\bigr) < C^2,
\end{align*}
so for large $m$ we obtain
\[\omega_m(p_m+\alpha_m+1) \leq Cp_m.\]
Thus, for $\varepsilon>0$ there exists $m_0\in\mathbb{N}$ such that such that $\abs{u(t)-u_m(t)}\leq \varepsilon$ for all $m\geq m_0$ and $t\in[-p_m,2p_m]$.
Since $u\in C_0(\mathbb{R})$, we there exists $m_1\geq m_0$ such that $\abs{u(t)}\leq \varepsilon$ for $\abs{t}\geq p_{m_1}-1=:t_1$. Then $\abs{u_m}\leq 2\varepsilon$ on $[-p_m,2p_m]\setminus (-t_1,t_1)$, for all $m\geq m_1$.
By Lemma \ref{lem:est_solutions} we obtain $\abs{a_mu_m'}\leq 2\varepsilon(2\norm{a_m}_\infty + \norm{\mu_m-z\lambda}_\mathrm{unif})$ on that set.
Hence,
\[\bigl(u_m(\pm p_m), (a_mu_m')(\pm p_m\rlim)\bigr),\bigl(u_m(2p_m), (a_mu_m')(2 p_m\rlim)\bigr) \to 0 \quad(m\to\infty).\]
Lemma \ref{lem:SL_per} yields $\bigl(u_m(0), (a_mu_m')(0\rlim)\bigr)\to 0$.
By Lemma \ref{lem:SL_est1} we now obtain $u_m\to 0$ locally uniformly. Since
$u_m\to u$ locally uniformly by \eqref{eq:SL_conv}, we obtain $u=0$, a
contradiction.
\end{proof}
\begin{remark}
As shown in \cite[Remark 5.7 and Section 6]{SeifertVogt2014} for Schr\"odinger operators (i.e.\ $a=1$), the eigenvalue bound $C^2 -
\norm{\mu}_\mathrm{unif}$ can be sharpened to the optimal bound $C_\mu^2-\inf_{r>0} \norm{\mu}_{\mathrm{unif},r}$, where
\[C_\mu = -\liminf_{p\to\infty} \frac{1}{p} \ln \norm{\mu-\mu(\cdot+p)}_{[-p,p]}\]
and
\[\norm{\mu}_{\mathrm{unif},r} := \frac{1}{r}\sup_{t\in\mathbb{R}} \abs{\mu}\bigl((a,a+r]\bigr).\]
An analogous sharpening can also be done in our case, also yielding optimal bounds.
\end{remark}
|
\section{Introduction and statement of results}
Let $w(x)$ be the following singularly perturbed Laguerre weight
\begin{equation} \label{weight-of-the-paper}
w(x)=w(x;t, \alpha)=x^{\alpha}e^{-V_t(x)},~~~x\in (0, \infty),~~t>0,~~\alpha>0
\end{equation}
with
\begin{equation} \label{vt-def}
V_t(x):= x + \frac{t}{x}, \qquad x \in (0,\infty), \quad t >0.
\end{equation}
In this paper, we consider the Hankel determinant associated with the above weight function
\begin{equation} \label{hakel-def}
D_n[w;t]:= \det (\mu_{j+k})_{j,k=0}^{n-1},
\end{equation}
where $\mu_j$ is the $j$-th moment of $w(x)$, i.e.
\begin{equation}
\mu_j := \int_0^\infty x^j w(x) dx.
\end{equation}
It is well-known that Hankel determinants are closely related to partition functions in random matrix theory. Indeed, let $Z_n(t)$ be the partition function associated with the weight function in \eqref{weight-of-the-paper}
\begin{equation} \label{par-def}
Z_n(t):=\int_{0}^{+\infty} \cdots \int_{0}^{+\infty} \prod_{1 \leq j < k \leq n} (\lambda_i - \lambda_j)^2 \lambda_j^\alpha \exp\biggl(-\sum_{j=1}^n V_t(\lambda_j) \biggr) d\lambda_1 \cdots d\lambda_n.
\end{equation}
Then, there is only a constant difference between $D_n[w;t]$ and $Z_n(t)$, that is
\begin{equation} \label{dn-zn}
D_n[w;t] = \frac{1}{n!}Z_n(t);
\end{equation}
for example, see \cite[Eq.(1.2)]{ci}. When $t=0$, $w(x)$ is reduced to the classical Laguerre weight. The corresponding partition function $Z_n(0)$ is associated with the Laguerre unitary ensemble and given explicitly below
\begin{equation} \label{zn0}
Z_n(0) = \prod_{j=1}^n j! \ \Gamma(j + \alpha);
\end{equation}
see \cite[p.321]{mehta}.
The matrix model and Hankel determinants $D_n[w;t]$ associated with the weight function $w(x)$ in \eqref{weight-of-the-paper} were first considered by Osipov and Kanzieper \cite{Osi:Kan} in bosonic replica field theories. Later, Chen and Its \cite{ci} consider this problem again from another point of view, where their motivation partially originates from an integrable quantum field theory at finite temperature. When $n$ is fixed, they showed that the Hankel determinant $D_n[w;t]$ is the isomonodromy $\tau$-function of a Painlev\'e III equation multiplied by two factors, that is,
\begin{equation}
D_n[w;t] = C \cdot \tau_n(t) \, e^{\frac{t}{2}} \; t^{\frac{n(n+\alpha)}{2}},
\end{equation}
where $C$ is a certain constant; see Eq. (6.14) in \cite{ci}.
In this paper, we focus on the asymptotics of Hankel determinants $D_n[w;t]$ as the matrix size $n$ tends to infinity.
In recent years, there has been a considerable amount of interest in the study of asymptotics of Hankel determinants and Toeplitz determinants, due to their important applications in various branches of applied mathematics and mathematical physics, such as graphical enumeration, one-dimensional gas of impenetrable bosons, two-dimensional Ising model, etc. For more information, one may refer to \cite{bleher2005,em2003,Zhao:Cao:Dai} and \cite{Dei:Its:Kra survey}, and references therein.
Note that the asymptotic analysis of $D_n[w;t]$ is totally nontrivial, and it is difficult to derive asymptotic properties about $D_n[w;t]$ from results in Chen and Its \cite{ci}. The main difficulty in the asymptotic study comes from the fact that $w(x)$ has an essential singularity at the origin. For regular $w(x)$, there are quite a lot of asymptotic results about the corresponding matrix models and Hankel determinants in the literature. For example, if $V(x)$ is real analytic and satisfies appropriate boundary conditions, universality results for the limiting kernels have been obtained by Deift {\it{et al.}} \cite{dkmv1}. A decade later, Lubinsky \cite{Lubinsky} improved one of the main results in \cite{dkmv1} by relaxing the condition and requiring $w(x)$ to be continuous only.
Cases when the weight function has singularities, such as jump discontinuities and algebraic singularities, are also studied; cf., e.g., \cite{Its:Kra2008,xz2011} and \cite{ik1,xz2013b}.
It is worth mentioning several recent work on Fisher-Hartwig singularities
\cite{Cla:Its:Kra,Dei:Its:Kra,Krasovsky}.
In \cite{Cla:Its:Kra,Dei:Its:Kra},
emergence of a Fisher-Hartwig singularity and a transition between the two different types of asymptotic behavior for Toeplitz determinants are investigated.
More recently, in \cite{Cla:Kra}, the asymptotics of Toeplitz determinants with merging singularities are considered, and Painlev\'{e} transcendents are applied to described the transition asymptotics. The reader is also referred to the comprehensive survey paper \cite{Dei:Its:Kra survey} for historic background, updated results and interesting applications in physics models in this regard.
The asymptotic study of matrix models with an essential singularity was first done by Mezzadri and Mo \cite{Mez:Mo} and Brightmore {\it{et al.}} \cite{Bri:Mez:Mo} when they are considering asymptotic properties of the partition function associated with the following weight
\begin{equation} \label{Mo-weight}
w(x;z,s) = \exp\left( -\frac{z^2}{2x^2} + \frac{s}{x} - \frac{x^2}{2} \right), \quad z \in \mathbb{R} \setminus \{0\}, \ 0\leq s < \infty, \ x \in \mathbb{R}.
\end{equation}
Here $w(x):=w(x;z,s)$ has an essential singularity at $x=0$. As pointed out in \cite{Bri:Mez:Mo,Mez:Mo},
although the system of polynomials orthogonal with respect to \eqref{weight-of-the-paper} and \eqref{Mo-weight} can be mapped to each other when $\alpha=\pm\frac{1}{2}$ and $s=0$, the relation between the respective partition functions is still unclear.
In \cite{Bri:Mez:Mo}, Brightmore {\it{et al.}} showed that a phase transition emerges as the matrix size $n \to \infty$ and $s,z =O(1/\sqrt{n})$,
which is characterized by a Painlev\'e III transcendent.
Recently, in several different areas of mathematics and physics, matrix models arise whose weight function has an essential singularity like \eqref{weight-of-the-paper}; see, e.g., Berry and Shukla \cite{Berry:Shukla} in the study of statistics for zeros of the Riemann zeta function, Lukyanov \cite{Lukyanov} in a calculation of finite temperature expectation values in integrable quantum field theory, and \cite{Bro:Fra:Bee,Mez:Sim,Tex:Maj} in the study of the Wigner time delay,
which stands for the average time that an electron spends when scattered by an open cavity.
The Wigner delay time plays a very important role in the
theory of mesoscopic quantum dots.
It has been shown that the distribution of the inverse delay times is given by the Laguerre ensemble \cite{Bro:Fra:Bee, Tex:Maj}, and the partition function serves as the moment generating function \cite{Bri:Mez:Mo,Mez:Sim}. However, the distribution of
the Wigner delay time is far from being understood and several interesting questions remain
open \cite{Mez:Sim,Tex:Maj}.
In our previous paper \cite{Xu:Dai:Zhao}, we have studied the eigenvalue correlation kernel $K_n(x,y;t)$ for the unitary matrix ensemble associated with the weight function $w(x)$ in \eqref{weight-of-the-paper}. As mentioned in \cite{Xu:Dai:Zhao}, when $t>0$, the exponent $\frac{t}{x}$ in \eqref{weight-of-the-paper} induces an infinitely strong zero at the origin and changes the eigenvalue distribution near the hard edge $x=0$. From an orthogonal polynomial point of view, for $t>0$, the model provides a family of non-Szeg\"{o} polynomial; see \cite{cd}, and \cite{Xu:Dai:Zhao,zz2008}. The asymptotics are of the Airy type at the soft edge adjacent to the origin. The reader is also refer to a relevant discussion in \cite[p. 274]{ci}, and Corollary \ref{cor1} below, to see that the polynomials with respect to the weight \eqref{weight-of-the-paper} show a singular behavior, as compared with the classic polynomials.
In \cite{Xu:Dai:Zhao}, we consider the case $t\to 0$, when a phase transition emerges. By applying Deift-Zhou steepest descent method for Riemann-Hilbert (RH) problems and using a double scaling argument, we obtain a new limit for
the eigenvalue correlation kernel, which is related to a third-order nonlinear differential equation. This equation is integrable and its Lax pair is given explicitly. Moreover, we have further showed that this third-order equation is equivalent to a particular Painlev\'e III equation by a simple transformation. The transition of the limiting kernel to the Bessel and Airy kernels is also discussed when the parameter $t$ varies in a finite interval $(0, d]$. For more details about the limiting kernel at the hard edge $0$, interested reader may refer to Section 1.2 in \cite{Xu:Dai:Zhao}.
In the present paper, based on the analysis in \cite{Xu:Dai:Zhao},
we derive large-$n$ asymptotic formulas for the Hankel determinants, the leading coefficients and recurrence coefficients, uniformly for $t\in (0,d]$, and their transitions as the parameters $t$ tends to $0^+$ and to a fixed $d>0$. In a sense, the large degree asymptotic expansion for $H_n$ (cf. \eqref{hn-def} and \eqref{hn-asy}) can be interpreted as an asymptotic study of a particular Painlev\'{e} III transcendent (cf. \eqref{hn-eqn} below).
\subsection{Statement of results}
To state our results, we need to use a particular solution $r(s)$ to the following third-order nonlinear differential equation
\begin{equation} \label{r-eqn-introduction}
2s^2 r' r''' - s^2 {r''}^2 + 2s r' r'' - 4 s {r'}^3 + \left(2r - \frac{1}{4} \right) {r'}^2 + 1=0.
\end{equation}
As pointed out in \cite{Xu:Dai:Zhao}, the above equation is integrable and its Lax pair is given in \cite[Proposition 1.]{Xu:Dai:Zhao}.
It is also shown in \cite[Proposition 2.]{Xu:Dai:Zhao} that by a change of unknown function
$v(s) =sr'(s)$,
the third-order equation for $r(s)$ is reduced to a particular Painlev\'{e} III equation for $v(s)$, namely,
\begin{equation} \label{v-eqn-introduction}
v''=\frac {{v'}^2}{v}-\frac {v'}{s}+\frac {v^2}{s^2} +\frac \alpha s-\frac 1 v.
\end{equation}
In the following theorem, we obtain the asymptotic expansion for the Hankel determinant $D_n[w;t]$ associated with the weight in \eqref{weight-of-the-paper} as $n \to \infty$. This expansion is uniformly valid for $0<t\leq d$, where $d$ is a fixed positive constant.
\begin{thm}\label{thm1}
Let $\alpha>0,t>0$, $w(x)$ be defined in \eqref{weight-of-the-paper} and $D_n[w;0]$ be
\begin{equation}
D_n[w;0] = \frac{1}{n!} \prod_{j=1}^n j! \ \Gamma(j + \alpha).
\end{equation}
Then as $n\to\infty$, we have the following asymptotic expansion
\begin{equation} \label{dn-asy}
D_n[w;t] = D_n[w;0] \exp\biggl\{ \biggl[1+O\biggl (\frac 1{n^{1/3}}\biggr ) \biggr] \, \int_0^t \frac{1-4\alpha^2 - 8r(2n\xi)}{16 \xi} d\xi\biggr\}
\end{equation}
where the error term is uniformly valid for $t\in(0,d],d>0$ and the error term can be improved to $O(1/n)$ if $s=2nt=O(1)$. Here $r(s)$ is a particular solution to the equation \eqref{r-eqn-introduction} which is analytic for $s\in (0, +\infty)$ and satisfies the following boundary conditions
\begin{equation} \label{r-boundary-behave}
r(0)=\frac 1 8\left (1-4\alpha^2\right )~~\mbox{and}~~r(s)= \frac 32 s^{\frac 23}-\alpha s^{\frac 13}+O(1)~~\mbox{as}~~s\to +\infty.
\end{equation}
\end{thm}
\begin{rmk}
Because \eqref{r-eqn-introduction} is an integrable equation, one may construct a RH problem for $r(s)$ explicitly based on its Lax pair; see Section \ref{sec-parametrix-0} below.
\end{rmk}
Let $\pi_n(x):=\pi_n(x;t)= x^n + \textsf{p}_1(n) x^{n-1} + \cdots$ be monic polynomials orthogonal with respect to the weight function in \eqref{weight-of-the-paper}, that is
\begin{equation} \label{op-pin}
\int_0^\infty \pi_j(x) \pi_k(x) w(x) dx = h_j \delta_{j,k}.
\end{equation}
Then the Hankel determinant can be given in terms of the constants $h_j$'s in the above formula
\begin{equation}
D_n[w;t] = \prod_{j=0}^{n-1} h_j.
\end{equation}
It is well-known that orthogonal polynomials satisfy a three-term recurrence relation as follows
\begin{equation} \label{pin-recurrence}
x\pi_n(x) = \pi_{n+1} (x) + \alpha_n(t) \pi_n(x) + \beta_n(t) \pi_{n-1}(x).
\end{equation}
In \cite{ci}, for fixed $n$, Chen and Its have showed that the recurrence coefficient $\alpha_n(t)$ satisfies a particular Painlev\'e equation in terms of the parameter $t$. Moreover, based on orthogonal polynomial techniques, they proved that $D_n[w;t]$ is closely related to the isomonodromy $\tau$-function of a Painlev\'e III equation. We summarize their results in the following theorem.
\begin{thm} \label{ci-thm}{\rm{(Chen and Its \cite{ci}).}}
Let $\alpha_n(t)$ be the recurrence coefficient given in \eqref{pin-recurrence},
\begin{equation} \label{an-def}
a_n(t) := \alpha_n(t)-(2n+ 1 + \alpha ),
\end{equation}
and
\begin{equation} \label{hn-def}
H_n(t) := t \frac{d}{dt} \ln D_n[w;t].
\end{equation}
Then $a_n(t)$ satisfies the following differential equation
\begin{equation} \label{an-PIII}
a_n''(t)=\frac {[a_n'(t)]^2}{a_n}-\frac {a_n'(t)}{t}+(2n+1+\alpha)\frac{a_n(t)^2}{t^2}+\frac{a_n(t)^3}{t^2}+\frac{\alpha}{t}-\frac {1}{a_n(t)},
\end{equation}
with the initial conditions
\begin{equation}\label{an-initial}
a_n(0) = 0, \qquad a_n'(0) = \frac{1}{\alpha}.
\end{equation}
The function $H_n(t)$ also satisfies a second-order nonlinear differential equation
\begin{equation} \label{hn-eqn}
[t H_n''(t)]^2 = [n - (2n+\alpha) H_n'(t)]^2 - 4[n(n+\alpha) + t H_n'(t) - H_n(t)]H_n'(t) [H_n'(t) - 1],
\end{equation}
with initial conditions
\begin{equation} \label{hn-ic}
H_n(0) = 0, \qquad H_n'(0) = -\frac{n}{\alpha}.
\end{equation}
\end{thm}
\begin{proof}
See proofs of Theorems 1 and 3 in \cite{ci}. Although the initial conditions \eqref{hn-ic} for $H_n(t)$ are not explicitly given in \cite{ci}, one may derive them from the formulas (3.10)-(3.11) and (3.21)-(3.22) in that paper, bearing in mind that $b_n(0)=0$ and $\beta_n(0)=n(n+\alpha)$.
\end{proof}
\begin{rmk}
It has been pointed out in Chen and Its \cite{ci}, that \eqref{an-PIII} is a Painlev\'e III equation, and \eqref{hn-eqn} is the corresponding Jimbo-Miwa-Okamoto $\sigma$-form.
\end{rmk}
In deriving the asymptotics of the Hankel determinant $D_n[w;t]$, we obtain asymptotics for several quantities associated with the orthogonal polynomials $\pi_n(x)$ in \eqref{op-pin}.
Consequently, we get a pair of approximating equations of Painlev\'{e} type, as follows:
\begin{thm} \label{thm3}Let $\alpha>0$, $t>0$, and the orthogonal weight $w(x)$ be defined in \eqref{weight-of-the-paper}. Then
we have the following uniform asymptotic approximations for the recurrence coefficients $\alpha_n(t)$ and $\beta_n(t)$ in \eqref{pin-recurrence}, and for the leading coefficient $\gamma_n$ of the orthonormal polynomial $p_n(z)$
\begin{eqnarray}
a_n(t) &=& \frac{s r'(s)}{2n}\left (1 + O\left ( {n^{-1/3}}\right )\right ),\label{an-asy} \\
H_n(t) &=& - \frac{8 r(s)+4\alpha^2-1}{16}\left (1 + O\left ({n^{-1/3}}\right )\right ),\label{hn-asy} \\
\alpha_n(t) &=& 2n + \alpha+1 + \frac{s r'(s)}{2n} \left (1+ O\left ( {n^{-1/3}}\right )\right ), \label{alpha-n-asy} \\
\beta_n(t) &=& n^2 + \alpha n + \frac{4 \alpha^2-1+ 8 r(s) - 8sr'(s) }{16}\left (1 + O\left ( {n^{-1/3}}\right ) \right ), \label{beta-n-asy}\\
\gamma_n(t) &=&
\frac 1 {\sqrt{n!\, \Gamma(n+1+\alpha)}} \left (1+ \frac { 8r(s) +4\alpha^2 -1} {32n} \left (1 + O\left ({n^{-1/3}}\right ) \right )\right ), \label{gamma-n-asy}\
\end{eqnarray} where $s=2nt$, and the uniformity is for $t\in (0, d]$, $d>0$ fixed.
Moreover, substituting the above into equations \eqref{an-PIII} and \eqref{hn-eqn}, and picking up the leading order terms as $n \to \infty$, we obtain
\begin{equation} \label{an-PIII-limit}
s^2 r'(s) r'''(s) - s^2 r''(s)^2 + sr'(s) r''(s) - sr'(s)^3 -\alpha r'(s) + 1 =0
\end{equation}
and
\begin{equation} \label{hn-eqn-limit}
{s^2 {r''}(s)^2}- 2 s r'(s)^3 + \frac{8r(s)- 1}{4}{{r'}(s)^2} + {2\alpha}{r'(s)}- 1 = 0,
\end{equation}
respectively.
\end{thm}
\begin{rmk}
It is shown in \cite[Section 2.2]{Xu:Dai:Zhao} that \eqref{hn-eqn-limit} is obtained by integrating \eqref{an-PIII-limit} once. Therefore, it is of interest to see that the Painlev\'e III equation \eqref{an-PIII} and the Jimbo-Miwa-Okamoto $\sigma$-form \eqref{hn-eqn} are equivalent asymptotically, with an appropriate rescaling of variables, and with appropriate initial values. Also, it is worth mentioning that the pair of equations,
\eqref{an-PIII-limit} and \eqref{hn-eqn-limit}, have played a key role in \cite{Xu:Dai:Zhao} to justify the equivalence of
\eqref{r-eqn-introduction} with a particular Painlev\'e III equation. Indeed, it is readily observed that a combination of \eqref{r-eqn-introduction} and \eqref{an-PIII-limit} gives \eqref{hn-eqn-limit}, and that a change of unknown function $v=sr'$ turns \eqref{an-PIII-limit} into the Painlev\'e III equation \eqref{v-eqn-introduction}.
\end{rmk}
Since all the asymptotic formulas in Theorems \ref{thm1} and \ref{thm3} have uniform error terms for $t\in (0, d]$, it is of interest to consider the transition of the quantities within such a framework, as $t$ varies in the interval. We take as examples the recurrence coefficients, and the logarithmic derivative of the Hankel determinant.
\begin{cor} \label{cor1}With the same conditions in Theorem \ref{thm3}, It holds
\begin{equation*} \alpha_n(t)= 2n + \alpha+1 + \frac s {2\alpha n} + O\biggl(\frac{s^2}n \biggr) + O\biggl(\frac{1}{n^{4/3}}\biggr),~~ \beta_n(t) = n^2 + \alpha n +O\left (s^2\right ) + O\biggl(\frac{1}{n^{1/3}}\biggr), \end{equation*}
and
\begin{equation*}H_n(t)=-\frac s {2\alpha} +O\left (s^2\right ) + O\biggl(\frac{1}{n^{4/3}}\biggr), \end{equation*}
as $n\to\infty$ and $s=2nt\to 0^+$. On the other hand, as $s=2nt\to \infty$, or, more specifically, as $t\to d^-$ and $n\to\infty$, we have
\begin{equation*}\alpha_n(t)= 2n+\alpha+1+\frac {t^{2/3}}{2^{1/3}} \frac 1 {n^{1/3}}
+O\left (\frac 1 {n^{2/3}} \right ),\end{equation*}
\begin{equation*}\beta_n(t)=n^2+\alpha n+ 2^{-4/3} t^{2/3} n^{2/3}+O\left ( n^{1/3} \right ),\end{equation*}
and\begin{equation*}H_n(t)=-\frac 3 4 (2t)^{2/3} n^{2/3} +O\left ( n^{1/3} \right ) . \end{equation*}
\end{cor}
\begin{proof}
The results are justified by a combination of \eqref{hn-asy}-\eqref{beta-n-asy} with the boundary conditions \eqref{r-boundary-behave}. Here use has been made of the fact that $r'(0)=\frac 1 \alpha$, as can be derived from \eqref{an-initial} and \eqref{an-asy}.
\end{proof}
To prove our main results in Theorem \ref{thm1} and \ref{thm3}, we adopt the Deift-Zhou steepest descent method for RH problems. This method has achieved great success in the study of asymptotics for orthogonal polynomials and the corresponding random matrix models, for example, see \cite{Dei:Its:Kra,dkmv1,zxz2011,zz2008}. Since most of the RH analysis has been done in our previous paper \cite{Xu:Dai:Zhao}, we will often refer to that paper. However, to make the current paper self-contained, we will sketch the RH analysis briefly and list some formulas in Section \ref{sec-RH-analysis} below. The interested reader may find more details in \cite{Xu:Dai:Zhao}.
The rest of the paper is arranged as follows. In Section \ref{sec-original-rhp}, we provide a RH problem for the corresponding orthogonal polynomials, as well as two formulas relating $a_n(t)$ and $H_n(t)$ with the RH problem. Such relations are the starting points of our analysis. In Section \ref{sec-RH-analysis}, we list some main steps and key formulas in the RH analysis. Then the proofs of Theorems \ref{thm1} and \ref{thm3} are given in Section \ref{sec-th1-proof} and Section \ref{sec-thm3-proof}, respectively.
\section{RH problem for orthogonal polynomials} \label{sec-original-rhp}
Consider a $2\times2$ Riemann-Hilbert (RH) problem as follows:
\begin{itemize}
\item[(Y1)] $Y(z)$ is analytic in
$\mathbb{C}\backslash [0,\infty)$;
\item[(Y2)] $Y(z)$ satisfies the jump condition
\begin{equation}\label{Y-jump}
Y_+(x)=Y_-(x) \left(
\begin{array}{cc}
1 & w(x) \\
0 & 1 \\
\end{array}
\right),
\qquad x\in (0,\infty),\end{equation} where $w(x)=w(x;t)=x^\alpha e^{-x-t/x}$ is the
weight function defined in (\ref{weight-of-the-paper});
\item[(Y3)] The asymptotic behavior of $Y(z)$ at infinity is
\begin{equation}\label{Y-infty}
Y(z)=\left (I+O\left ( 1 /z\right )\right )\left(
\begin{array}{cc}
z^n & 0 \\
0 & z^{-n} \\
\end{array}
\right),\quad \mbox{as}\quad z\rightarrow
\infty ;\end{equation}
\item[(Y4)] The asymptotic behavior of $Y(z)$ at the end points $z=0$ are
\begin{equation}\label{Y-origin}Y(z)=\left(
\begin{array}{cc}
O( 1) & O( 1) \\[0.2cm]
O( 1) & O( 1)
\\
\end{array}
\right),\quad \mbox{as}\quad z\rightarrow
0 .\end{equation}
\end{itemize}
According to the significant observation of Fokas, Its and Kitaev \cite{fik}, the
solution of the above RH problem is given in terms of the monic polynomials,
orthogonal with respect to $w(x)$. This establishes an
important relation between orthogonal polynomials and
RH problems.
\begin{lem} \label{fik-lemma}
{\rm{
(Fokas, Its and Kitaev \cite{fik}).}}
The unique solution to the above RH problem for $Y$ is given by
\begin{equation}\label{Y-solution}
Y(z)= \left (\begin{array}{cc}
\pi_n(z)& \frac 1 {2\pi i}
\int_{0} ^{\infty}\frac {\pi_n(s) w(s) }{s-z} ds\\[0.2cm]
-2\pi i \gamma_{n-1}^2 \;\pi_{n-1}(z)& - \gamma_{n-1}^2\;
\int_{0} ^{\infty}\frac {\pi_{n-1}(s) w(s) }{s-z} ds \end{array} \right ),
\end{equation}
where $\pi_n(z)$ is the monic polynomial in \eqref{op-pin}, and $\gamma_{n}$ is the leading coefficient of the orthonormal polynomial $p_n(z):=\gamma_{n}\pi_n(z)$
with respect to the weight $w(x)=w(x;t)$; cf., e.g., \cite{deift} and \cite{fik}.
\end{lem}
\begin{proof}
The proof is based on the Plemelj formula and Liouville's theorem.
\end{proof}
It is worth mentioning that the important quantities $a_n(t)$ and $H_n(t)$ in Theorem \ref{ci-thm} can be expressed in terms of the entries of the RH solution $Y$. These relations are very helpful in our future calculations.
\begin{lem} \label{lem-fixn}
The following identities hold for $n \in \mathbb{N}$:
\begin{equation} \label{a-n definition from chen}
a_n(t)=2\pi i t\gamma_n^2 Y_{11}(0)Y_{12}(0)
\end{equation}
and
\begin{equation} \label{hn definition from chen}
H_n'(t) = - Y_{12}(0)Y_{21}(0),
\end{equation}
where $Y_{ij}(z)$ denotes the (i,j)-entry of the matrix-valued function $Y(z)$.
\end{lem}
\begin{proof}
From Lemma 2 in Chen and Its \cite{ci}, we have
\begin{equation} \label{an integral from chen}
a_n(t)=\frac{t}{h_n} \int_0^{\infty}\frac{\pi_n^2(y)w(y)}{y}dy = t \gamma_n^2 \int_0^{\infty}\frac{\pi_n^2(y)w(y)}{y}dy,
\end{equation}
where $h_n$ is the constant in \eqref{op-pin} and $\gamma_n$ is the leading coefficient of the orthonormal polynomial. Note that $\displaystyle\frac{\pi_n(y)}{y} = q_{n-1}(y) + \frac{\pi_n(0)}{y}$, where $q_{n-1}(y)$ is a polynomial of degree $n-1$. By the orthogonality of $\pi_n(x)$ in \eqref{op-pin}, the above formula gives us
\begin{equation}
a_n(t)=t \gamma_n^2 \int_0^{\infty}\frac{\pi_n(y) \pi_n(0)w(y)}{y}dy.
\end{equation}
Comparing the above formula with \eqref{Y-solution} gives us \eqref{a-n definition from chen}.
To get \eqref{hn definition from chen}, we obtain the following formula from Lemma 2 and Eq. (3.21) in \cite{ci}
\begin{equation} \label{hn prime integral from chen}
H_n'(t) = \frac{1}{h_{n-1}} \int_0^{\infty} \frac{\pi_n(y)\pi_{n-1}(y)w(y)}{y}dy = \gamma_{n-1}^2\int_0^{\infty}\frac{\pi_n(y)\pi_{n-1}(y)w(y)}{y}dy.
\end{equation}
Using an argument similar to the derivation of \eqref{a-n definition from chen}, we obtain \eqref{hn definition from chen}.
\end{proof}
\begin{rmk} \label{alternative derivation}
In Chen and Its \cite{ci}, the integral representation \eqref{an integral from chen} is obtained appealing to a ladder operator technique. We note that \eqref{an integral from chen}, \eqref{hn prime integral from chen} and \eqref{beta-n hn relation from chen} below can be derived alternatively, using the recurrence relation \eqref{pin-recurrence} and the specific perturbed Laguerre weight \eqref{weight-of-the-paper}. For example, using integration by parts, \eqref{an integral from chen} can be derived as
\begin{equation*}\alpha_n=\gamma_n^2 \int^\infty_0 x\pi_n^2 w(x) dx=-\gamma_n^2 \int^\infty_0 x\pi_n^2 x^\alpha e^{-\frac t x} de^{-x}= 2n +1+\alpha + t \gamma_n^2\int^\infty_0
\frac {\pi_n^2 w(x) dx} x. \end{equation*}
\end{rmk}
\section{Nonlinear steepest descent analysis} \label{sec-RH-analysis}
In this section, we give a sketch of the nonlinear steepest descent analysis for the RH problem.
In the standard Deift-Zhou analysis, one introduces a sequence of
transformations:
\[ Y \mapsto T \mapsto S \mapsto R,\]
and reduces the original RH problem for $Y$ to a new RH problem for $R$, whose jumps are close to the identity
matrix when $n$ is large. Then $R(z)$ can then be expanded into a Neumann series on the whole complex plane. Since the above transformations are all revertible, the uniform asymptotics of the orthogonal
polynomials in the complex plane are obtained for large polynomial degree $n$ when we trace back. Technique difficulties lie in the construction of the local parametrix in a neighborhood of the origin $z=0$. The parametrix possesses irregular singularity both at infinity and at the origin.
\bigskip
\noindent\textbf{Normalization: $Y \mapsto T$.} The first transformation in the Deift-Zhou steepest descent analysis
is to normalize the large-$z$ behavior of $Y(z)$ in \eqref{Y-infty} to make $T(z)\sim I$ as $z\to\infty$. To this end, we introduce the following $g$-function
\begin{equation}\label{g-function}
g(z):=\frac 2{\pi}\int_0^1 \log(z-x) \sqrt{\frac{1-x}{x}} dx, \qquad z \in \mathbb{C} \setminus (-\infty, 1],
\end{equation}
where the branch is chosen such that $\arg(z-x)\in(-\pi,\pi)$. Then the first transformation $Y \mapsto T$ is defined as
\begin{equation}\label{TrsnaformationY-T}
T(z)= (4n)^{-(n+\frac \alpha 2) \sigma_3}e^{-\frac{n l}{2}\sigma_3}Y(4nz) e^{-n(g(z)-\frac{l}{2})\sigma_3}e^{-\frac {t_n}{8nz}\sigma_3} (4n)^{\frac \alpha 2\sigma_3}
\end{equation}
for $z\in\mathbb{C}\backslash[0,\infty)$, where $l=-2(1+\ln 4)$ is the Lagrange multiplier
and $\sigma_3$ is the Pauli matrix $\begin{pmatrix}
1 &0 \\ 0 & -1
\end{pmatrix}$; see \eqref{Pauli-matrix} below. Here $t=t_n$ indicates the dependence of the parameter $t$ on the polynomial degree $n$.
It is easily verified that the large-$z$ behavior of $T(z)$ is normalized such that $T(z) = I + O(z^{-1})$ as $z \to \infty$. Note that, because there is a factor $e^{-\frac{t}{x}}$ in the weight function $w(x)$ in \eqref{weight-of-the-paper}, there is a corresponding $e^{-\frac {t_n}{8nz}\sigma_3}$ term in \eqref{TrsnaformationY-T}. As a consequence, one can easily see from \eqref{Y-origin} and \eqref{TrsnaformationY-T} that $T(z)$
possesses an essential singularity at the origin. This kind of singularity is new in the Riemann-Hilbert analysis and requires a new class of parametrix near the origin; see Section \ref{sec-parametrix-0} below.
\bigskip
\noindent\textbf{Opening of the lens: $T \mapsto S$.} Although the first transformation successfully normalizes the large-$z$ behavior of $T$, the original jump matrix \eqref{Y-jump} for $Y$ becomes more complicated. More precisely, because the function $g(z)$ in \eqref{g-function} is not analytic on the interval $(0,1)$, the jump matrix for $T$ is highly oscillatory on $(0,1)$ when the polynomial degree $n$ is large. To overcome this difficulty, Deift {\it{et al.}} \cite{dkmv1,dkm} borrowed the ideas from the classical steepest descent method for integrals and introduced the second transformation $T \mapsto S$. In this transformation, the original interval $[0,\infty)$ is deformed by opening lens. Then, the rapidly oscillatory jump matrices for $T$ will be reduced to jump matrices tending to the identity matrix exponentially, except in a neighborhood of $(0,1)$.
Before we introduce the transformation, we need one more auxiliary $\phi$-function as follows
\begin{equation}\label{phi-function}
\phi(z):=2 \int_0^z\sqrt{\frac{s-1}{s}}ds,~~z\in
\mathbb{C}\backslash[0,\infty),
\end{equation}
where $\arg z\in(0,2\pi)$, such that the Maclaurin expansion $\phi(z)= 4i\sqrt{z}\left\{ 1-\frac z 6+\cdots \right \}$ holds for $|z|<1$. From its definition, one can see that $\phi(x)>0$ for $x>1$, $\re \phi(z)<0$ in the lens-shaped domains; cf. Figure \ref{contour-for-S}, and that $\phi_\pm (x)= \pm 2i \int^x_0\sqrt{\frac{1-s} s }ds$, purely imaginary, for $x\in (0,1)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{contour-for-S.pdf}
\end{center}
\caption{\small{Contour $\Sigma_{S}$ for the RH problem for $S(z)$ in the $z$-plane.}}
\label{contour-for-S}
\end{figure}
The second transformation $T \mapsto S$ is defined as
\begin{equation}\label{transformationT-S}
S(z)=\left \{
\begin{array}{ll}
T(z), & \mbox{for $z$ outside the lens shaped region;}
\\ [.4cm]
T(z) \left( \begin{array}{cc}
1 & 0 \\
- z^{-\alpha}e^{2n\phi(z)} & 1 \\
\end{array}
\right) , & \mbox{for $z$ in the upper lens
region;}\\[.4cm]
T(z) \left( \begin{array}{cc}
1 & 0 \\
z^{-\alpha} e^{2n\phi(z)} & 1 \\
\end{array}
\right) , & \mbox{for $z$ in the lower
lens
region, }
\end{array}\right .\end{equation}
where $\arg z\in (-\pi, \pi)$. Then, $S$ satisfies a RH problem with the following jump conditions $J_S(z)$
\begin{equation} \label{S-jump}
J_S(z)= \begin{cases}
\left( \begin{array}{cc}
1 & 0 \\
z^{-\alpha} e^{2n\phi(z)} & 1 \\
\end{array}
\right), & z\in \gamma_1\cup \gamma_3, \\
\left( \begin{array}{cc}
0& x^{\alpha} \\
-x^{-\alpha} & 0 \\
\end{array}
\right), & z=x\in \gamma_2, \\
\left( \begin{array}{cc}
1 & z^{\alpha}e^{-2n \phi(z)} \\
0 & 1 \\
\end{array}
\right), & z\in (1,+\infty).
\end{cases}
\end{equation}
Of course $S(z)$ is still normalized at $\infty$ and possesses an essential singularity at 0.
\subsection{Outside parametrix}
From (\ref{S-jump}), we see that the jump matrix for $S$ is the identity matrix plus an exponentially small term for
fixed $z\in \gamma_1\cup\gamma_3\cup (1, \infty)$. Neglecting the exponentially small terms, we arrive at an approximating RH problem for $N(z)$, as follows:
\begin{description}
\item(N1)~~ $N(z)$ is analytic in $\mathbb{C}\backslash [0,1]$;
\item(N2)~~ \begin{equation}\label{N-jump} N_{+}(x)=N_{-}(x)\left(
\begin{array}{cc}
0 & x^{\alpha} \\
-x^{-\alpha} & 0 \\
\end{array}
\right)~~~\mbox{for}~~x\in (0,1);\end{equation}
\item(N3)~~ \begin{equation}\label{N-infty} N(z)= I+O( 1/ z) ,~~~\mbox{as}~~z\rightarrow\infty .\end{equation}
\end{description}
A solution to the above RH problem can be constructed explicitly,
\begin{equation}\label{N-solution} N(z)= D_{\infty}^{\sigma_3}M^{-1} a (z)^{-\sigma_3} MD(z)^{-\sigma_3}, \end{equation}
where $M=(I+i\sigma_1) /{\sqrt{2}}$, $a
(z)=\left(\frac{z-1} {z}\right)^{1/4}$ with $\arg z\in (-\pi, \pi)$ and $\arg (z-1)\in (-\pi, \pi)$, and the Szeg\"{o} function
\begin{equation*} D(z)=\left(\frac{z}{\varphi(2z-1)}\right)^{\alpha/2},~~\varphi(z)=z+\sqrt{z^2-1}, \end{equation*}
the branches are chosen such that $\varphi(z) \sim 2z$ as $z\rightarrow\infty$, and $D_{\infty}=2^{-\alpha}$.
\subsection{Local parametrix $P^{(1)}(z)$ at $z=1$ }
The jump matrices of $SN^{-1}$ are not uniformly close to the unit matrix near the end-points $0$ and $1$, thus local parametrices have to be constructed in
neighborhoods of the end-points. Near the right end-point $z=1$, we consider a small disk
$U(1,\delta)=\{z~|\;|z-1|<\delta\}$, $\delta$ being a fixed positive number.
The parametrix $P^{(1)}(z)$ in $U(1,\delta)$ can be constructed in terms of the Airy function
and its derivative as in \cite[(3.74)]{MV}; see also \cite{deift,dkm}.
\subsection{Local parametrix $P^{(0)}(z)$ at the origin} \label{sec-parametrix-0}
The parametrix, to be constructed in the neighborhood $U(0,\delta)=\{z~|\;|z|<\delta\}$ for sufficiently small $\delta$, solves a RH problem as
follows:
\begin{itemize}
\item[(a)] $P^{(0)}(z)$ is analytic in $U(0,\delta) \backslash \Sigma_{S}$;
\item[(b)] In $U(0,\delta)$, $P^{(0)}(z)$ satisfies the same jump conditions as $S(z)$ does; cf. (\ref{S-jump});
\item[(c)] $P^{(0)}(z)$ fulfils the following matching condition
on $\partial U(0,\delta)=\{~z\; |\; |z|=\delta\}$:
\begin{equation}\label{matchingP0-N}
P^{(0)}(z)N^{-1}(z)=I+ O\left ( n^{-1/3}\right )~~\mbox{as}~n \to \infty;
\end{equation}
\item[(d)] The behavior at the center $z=0$ is the same as that of $S(z)$, which possesses an essential singularity at 0.
\end{itemize}
The construction of the local parametrix is one of the main contribution in our previous paper \cite{Xu:Dai:Zhao}. It involves a model RH problem for $\Psi(\zeta,s)$ which is related to the third-order integrable ODE in \eqref{r-eqn-introduction}. The exact formula for $P^{(0)}(z)$ is given explicitly as follows:
\begin{equation} \label{p0-para}
P^{(0)}(z)= E(z)\Psi(n^2\phi^2, 2nt_n)e^{-\frac \pi 2 i\sigma_3} (-z)^{-\frac \alpha2\sigma_3}e^{n\phi(z)\sigma_3}, ~~z\in U(0,\delta) \backslash \Sigma_{S},
\end{equation}
where $E(z)$ is an analytic function in $U(0,\delta)$
\begin{equation}\label{Ez-def}
E(z)=N(z)e^{\frac \pi 2 i\sigma_3} (-z)^{\frac \alpha 2\sigma_3}\frac{I-i\sigma_1}{\sqrt{2}}
\left\{n^2\phi^2(z)\right\}^{\frac 1 4\sigma_3}.
\end{equation}
In the above two formulas, we choose $\arg(-z)\in (-\pi, \pi)$ and $\arg \left\{ n^2\phi^2(z)\right\} \in (-\pi, \pi)$.
The important function $\Psi(\zeta)=\Psi(\zeta,s)$ satisfies the following model RH problem
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{contour-for-model.pdf} \end{center}
\caption{\small{Contours and regions for the model RH problem for $\Psi$ in the $\zeta$-plane, where both sectors $\Omega_2$ and $\Omega_3$ have an opening angle $\pi/3$.}}
\label{contour-for-model}
\end{figure}
\begin{itemize}
\item[(a)] $\Psi(\zeta)$ is analytic in
$\mathbb{C}\backslash\cup^3_{j=1}\Sigma_j$, where $\Sigma_j$ are illustrated in Figure \ref{contour-for-model};
\item[(b)] $\Psi(\zeta)$ satisfies the jump condition
\begin{equation}\label{Psi-jump}
\Psi_+(\zeta)=\Psi_-(\zeta)
\left\{
\begin{array}{ll}
\left(
\begin{array}{cc}
1 & 0 \\
e^{\pi i\alpha}& 1 \\
\end{array}
\right), & \zeta \in \Sigma_1, \\[.4cm]
\left(
\begin{array}{cc}
0 &1\\
-1&0 \\
\end{array}
\right), & \zeta \in \Sigma_2, \\[.4cm]
\left(
\begin{array}{cc}
1 &0 \\
e^{-\pi i\alpha} &1 \\
\end{array}
\right), & \zeta \in \Sigma_3;
\end{array} \right .
\end{equation}
\item[(c)] The asymptotic behavior of $\Psi(\zeta)$ at infinity is
\begin{equation}\label{Psi-infty}
\Psi(\zeta, s)= \left[ I + \frac{C_{1}(s)}{\zeta} +
O\left(\frac{1}{\zeta^2}\right) \right] \; \zeta^{-\frac{1}{4} \sigma_3} \frac{I + i \sigma_1}{\sqrt{2}} e^{\sqrt{\zeta} \sigma_3},~~\arg \zeta\in (-\pi, \pi),~~\zeta\rightarrow \infty,
\end{equation}
where $C_1(s)$ is a matrix independent of $\zeta$;
\item[(d)] The asymptotic behavior of $\Psi(\zeta)$ at $\zeta=0$ is
\begin{equation}\label{Psi-origin}
\Psi(\zeta,s)=Q(s)\left\{I+O(\zeta)\right\} e^{\frac s \zeta\sigma_3}\zeta^{\frac \alpha2\sigma_3}\left\{
\begin{array}{ll}
I, & \zeta \in \Omega_1\cup\Omega_4, \\[.4cm]
\left(
\begin{array}{cc}
1 & 0 \\
-e^{\pi i\alpha}& 1 \\
\end{array}
\right), & \zeta \in \Omega_2, \\[.4cm]
\left(
\begin{array}{cc}
1 &0 \\
e^{-\pi i\alpha} &1 \\
\end{array}
\right), & \zeta \in \Omega_3
\end{array} \right .
\end{equation}
for $\arg \zeta\in (-\pi, \pi)$, as $\zeta\rightarrow 0$, where $\Omega_1-\Omega_4$ are depicted in Figure \ref{contour-for-model}, $Q(s)$ is a matrix independent of $\zeta$, such that $\det Q(s)=1$, and $\sigma_j$ are the Pauli matrices, namely,
\begin{equation}\label{Pauli-matrix}
\sigma_1=\left(
\begin{array}{cc}
0 &1 \\
1 & 0 \\
\end{array}
\right), ~~\sigma_2=\left(
\begin{array}{cc}
0 & -i \\
i & 0 \\
\end{array}
\right)~~\mbox{and}~\sigma_3=\left(
\begin{array}{cc}
1 & 0 \\
0 & -1 \\
\end{array}
\right).
\end{equation}
\end{itemize}
\begin{rmk}
The relation between the above model RH problem for $\Psi$ and the equation \eqref{r-eqn-introduction} is that, let $r(s) = i (C_{1}(s))_{12}$ with $C_1(s)$ given in \eqref{Psi-infty}, then $r(s)$ satisfies the equation \eqref{r-eqn-introduction}. For a detailed derivation, see \cite[Section 2.1]{Xu:Dai:Zhao}.
\end{rmk}
\begin{rmk}
Readers who are familiar with Riemann-Hilbert analysis may be a little surprised to see the error term $O(n^{-1/3})$ instead of $O(n^{-1})$ in \eqref{matchingP0-N}. In fact, if $s$ is fixed, the function $r(s)$ is bounded and we do get the $O(n^{-1})$ estimation. As $s= 2nt$, to achieve the uniform results in Theorem \ref{thm3} for $t\in(0,d]$, $d>0$ fixed, an asymptotic study for $r(s)$ as $s\to 0$ and $s\to +\infty$ is needed. In our previous paper \cite{Xu:Dai:Zhao}, by performing an asymptotic study of the model RH problem for $\Psi(x,s)$, we get the desired results in the following proposition. The reason why $O(n^{-1/3})$ appears can be seen from the large-$s$ asymptotic behavior of $r(s)$.
\begin{prop}(Xu, Dai and Zhao \cite{Xu:Dai:Zhao}) There exists a solution $r(s)$ of \eqref{r-eqn-introduction}, analytic for $s\in(0,\infty)$, with the following boundary conditions
\begin{equation} \label{r boundary conditions}
r(0)=\frac 18(1-4\alpha^2) \quad \mbox{and} \quad r(s)=\frac 32s^{\frac 23}-\alpha s^{\frac 13}+O(1)~ \textrm{as}~s\to +\infty.
\end{equation}
\end{prop}
\end{rmk}
\subsection{The final transformation $S\mapsto R$}
Now we bring in the final transformation by defining
\begin{equation}\label{transformationS-R}
R(z)=\left\{ \begin{array}{ll}
S(z)N^{-1}(z), & z\in \mathbb{C}\backslash \left \{ U(0,\delta)\cup U(1,\delta)\cup \Sigma_S \right \};\\[.1cm]
S(z) (P^{(0)})^{-1}(z), & z\in U(0,\delta)\backslash \Sigma_{S} ; \\[.1cm]
S(z) (P^{(1)})^{-1}(z), & z\in U(1,\delta)\backslash
\Sigma_{S} .
\end{array}\right .
\end{equation}
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{contour-for-R.pdf} \end{center}
\caption{\small{Contour $\Sigma_R$ for the RH problem for $R(z)$ in the $z$-plane.}}
\label{contour-for-R}
\end{figure}
Then, $R(z)$ solves the following RH problem:
\begin{description}
\item(R1)~~ $R(z)$ is analytic in $\mathbb{C} \backslash \Sigma_R$ (see Figure \ref{contour-for-R} for the contours);
\item(R2)~~ $R(z)$ satisfies the jump conditions
\begin{equation}\label{R-jump}
R_+(z)=R_-(z)J_{R}(z), ~~z\in\Sigma_R,
\end{equation}
where
$$
J_R(z)=\left\{ \begin{array}{ll}
P^{(0)}(z)N^{-1}(z), ~ & z\in\partial U(0,\delta),\\[.1cm]
P^{(1)}(z)N^{-1}(z),& z\in\partial
U(1,\delta),\\[.1cm]
N(z)J_S(z)N^{-1}(z), ~& \Sigma_R\setminus \partial
( U(0,\delta)\cup U(1,\delta));
\end{array}\right .$$
\item(R3)~~ $R(z)$ satisfies the following behavior at infinity:
\begin{equation}\label{R-infty}
R(z)= I+ O\left({1}/{z} \right),~~\mbox{as}~ z\rightarrow\infty .
\end{equation}
\end{description}
It follows from the matching condition (\ref{matchingP0-N}) of the local parametrices and the definition of $\phi(z)$ in \eqref{phi-function} that
\begin{equation}\label{R-jump-approx}J_R(z)=\left\{ \begin{array}{ll}
I+O\left (n^{-1/3}\right ),& z\in\partial
U(0,\delta)\cup U(1,\delta),\\[.1cm]
I+O(e^{-cn}), ~& z\in \Sigma_R\setminus \partial
( U(0,\delta)\cup U(1,\delta)),
\end{array}\right .
\end{equation}
where $c$ is a positive constant, and the error term is uniform
for $z$ on the corresponding contours. Hence we have
\begin{equation}\label{R-jump-estimate}\|J_R(z)-I\|_{L^2\cap L^{\infty}(\Sigma_R)}=O(n^{-1/3}).
\end{equation}
Then, applying the now standard procedure of norm
estimation of Cauchy operator and using the technique of deformation of contours (cf. \cite{deift,dkm}), it
follows from (\ref{R-jump-estimate}) that
\begin{equation}\label{R-approx}R(z)=I+O(n^{-1/3}),
\end{equation} uniformly for $z$ in the whole complex plane, and for $t$ in the interval $(0, d]$, $d>0$ fixed.
This completes the nonlinear steepest descent analysis.
\section{Proof of Theorem \ref{thm1}} \label{sec-th1-proof}
Since we only need to derive one-term asymptotic approximation, it is very helpful to make use of the nice relations between $a_n(t)$, $H_n(t)$ and $Y(0)$ in Lemma \ref{lem-fixn}. These relations can simplify our computations a lot. If one wishes to obtain more terms in the asymptotic expansions, we need to follow the original ideas in Deift et. al. \cite{dkm}; see also \cite[Sec. 4]{MV}.
\bigskip
\noindent\emph{Proof of Theorem \ref{thm1}.} Tracing back the transformations $R\mapsto S \mapsto T \mapsto Y$, and combining \eqref{TrsnaformationY-T}, \eqref{transformationT-S} and \eqref{transformationS-R} with \eqref{p0-para}, we have
\begin{equation}\label{Y-trace-back}
Y_+(4nx)=c_n^{\sigma_3} R(x) E(x) \Psi_- (n^2\phi^2(x), s) e^{\frac {i\pi} 2 (\alpha-1 )\sigma_3}\left(
\begin{array}{cc} 1 & 0 \\ 1 & 1 \end{array} \right)\left [w(4nx)\right ]^{-\frac 1 2\sigma_3},
\end{equation}
for $0<x<\delta$, where $ c_n= (-1)^n (4n)^{n+\frac \alpha 2} e^{\frac 1 2 {nl}}. $
To evaluate $Y(0)$, we need to find out the values of $E(x)$ and $\Psi_- (n^2\phi^2(x), s)$ as $x \to 0$. From the definitions of $\phi(z)$ and $E(z)$ in \eqref{phi-function} and \eqref{Ez-def}, one can see that
\begin{equation} \label{E-0}
\phi^2(x) = -16 x + O(x^2) \ \textrm{ and } \ E(x)= 2^{-\alpha\sigma_3} M^{-1} \begin{pmatrix}
0 & 1 \\ -1 & 0
\end{pmatrix} (2\sqrt{n})^{\sigma_3} + O(x)
\end{equation}
as $x \to 0$, where $M=(I+i\sigma_1) /{\sqrt{2}}$. Moreover, from the the behavior of $\Psi$ at zero in $\Omega_3$, see \eqref{Psi-origin}, we have
\begin{equation}
\Psi_- (f_n(x), 2nt) e^{\frac {i\pi} 2 (\alpha-1 )\sigma_3}\left( \begin{array}{cc}
1 & 0 \\ 1 & 1 \end{array} \right)\left [w(4nx)\right ]^{-\frac 1 2\sigma_3}
=Q(2nt) (I+O(f_n(x)))d^{\sigma_3},
\end{equation}
where $d=e^{2nt/f_n(x)}e^{\frac {\pi i}2(\alpha-1)}(f_n)^{\alpha/2}w(4nx)^{-\frac 12}$, and $f_n(z)=n^2 \phi^2(z) $ furnishes a conformal mapping in a neighborhood of $z=0$. Note that the explicit formula for $Q(s)$ can be obtained explicitly as follows
\begin{equation}\label{Q}
Q(s)=\sqrt{\frac {1+q'(s)}2}
\left( \begin{array}{cc}
1 & \frac{ir'(s)}{1+q'(s)} \\ i\frac{1-q'(s)}{r'(s)} & 1 \end{array}
\right)\chi ^{\sigma_3};
\end{equation} see notations and computations in \cite[Sec. 2.1]{Xu:Dai:Zhao}, with $q'(s) =-sr''(s)-\frac 1 2 r'(s)+r(s)r'(s)$,
where $\chi$ is an arbitrary non-zero factor. Using the fact that $R(x) = I + O(1/n^{1/3})$ and the above formulas, we have
\begin{eqnarray}
Y_{11}(0)Y_{12}(0)& = & (4n)^{2(n+\frac \alpha2)} e^{nl}\biggl[E(0)Q(2nt)\biggr]_{11}\biggl[E(0)Q(2nt) \biggr]_{12} (1+O ( {n^{-1/3}} )) \nonumber \\
& = & -i 2^{-2\alpha}(4n)^{2(n+\frac \alpha2)} e^{nl} nr'(2nt)(1+O( {n^{-1/3}})) \label{Y-11(0)Y-12(0)}
\end{eqnarray}
and
\begin{eqnarray}
Y_{12}(0)Y_{21}(0)&=& \biggl[E(0)Q(2nt)\biggr]_{12}\biggl[E(0)Q(2nt) \biggr]_{21}(1+O( {n^{-1/3}})) \nonumber \\
& =& n r'(2nt)(1+O( {n^{-1/3}})), \label{Y-12(0)Y-21(0)}
\end{eqnarray}
where the error term is uniform for $ t\in(0,d]$, $d>0$.
Combining \eqref{hn definition from chen} and \eqref{Y-12(0)Y-21(0)} gives us
\begin{equation}
H_n'(t) = - nr'(2nt)(1+O( {n^{-1/3}})).
\end{equation}
Recalling \eqref{r-boundary-behave} and \eqref{hn-ic}, we know that $r(0) = \frac{1-4\alpha^2}{8}$ and $H_n(0) =0$. Therefore, integrating the above formula from $0$ to $t$ gives us
\begin{equation} \label{asymptotic of h-n}
H_n(t)=\frac{1-4\alpha^2 - 8r(2nt)}{16}(1+O( {n^{-1/3}})).
\end{equation}
In view of \eqref{zn0} and \eqref{hn-def}, integrating one more time of the above formulas gives us \eqref{dn-asy}.
This completes the proof of Theorem \ref{thm1}. \hfill\qed
\section{Proof of Theorem \ref{thm3}} \label{sec-thm3-proof}
The proof the Theorem \ref{thm3} is similar to that in the previous section.
\bigskip
\noindent\emph{Proof of Theorem \ref{thm3}.} We first derive the formula for $\alpha_n(t)$ in \eqref{alpha-n-asy}. Owing to the relations in \eqref{an-def} and \eqref{a-n definition from chen}, one needs only to derive the asymptotics for $\gamma_n$ and $Y_{11}(0)Y_{12}(0)$. Note that the asymptotics of $Y_{11}(0)Y_{12}(0)$ have been given in \eqref{Y-11(0)Y-12(0)}.
To get the asymptotics for $\gamma_n$, we need the large-$z$ behavior of $Y(z)$. To achieve it, we again trace back the transformations $R\mapsto S \mapsto T \mapsto Y$. Then, from \eqref{TrsnaformationY-T}, \eqref{transformationT-S} and \eqref{transformationS-R}, we get
\begin{equation} \label{asymptotic of Y}
Y(4nz)= (4n)^{(n +\frac {\alpha}{2})\sigma_3} e^{\frac{n}{2} l\sigma_3}R(z)N(z)e^{n(g(z)-\frac 1 2l)\sigma_3}e^{\frac {t}{8nz}\sigma_3}(4n)^{-\frac {\alpha}{2}\sigma_3}
\end{equation}
for $z$ is far away from the interval $[0,\infty)$. By the uniform approximation of $R$ in \eqref{R-approx}, we have the following estimate for large $n$ and $|z|$
$$
R(z)=I+O\left (\frac{1}{n^{1/3}z}\right ),
$$
where the error term is uniform for $ t\in(0,d],d>0$. From the explicit formulas of $g(z)$ and $N(z)$ in \eqref{g-function} and \eqref{N-solution}, it is easily seen that
$$
e^{ng(z)\sigma_3}z^{-n\sigma_3}=I-\frac n {4z}\sigma_3 +O\left (\frac{1}{z^2}\right )
$$
and
$$
N(z)=I+2^{-\alpha\sigma_3}\left (-\frac \alpha 4\sigma_3-\frac 1 4 \sigma_2\right )2^{\alpha\sigma_3}\frac 1 z+O\left (\frac{1}{z^2}\right ).
$$
If we expand $Y(4nz)$ in \eqref{asymptotic of Y} as $ Y(4nz) (4nz)^{-n\sigma_3} = I + \frac{\tilde Y_1}{z} + O\left (\frac{1}{z^2}\right ) $, then according to Eq. (3.11) in Deift
{\it{et al.}} \cite{dkm}, we have
\begin{equation}
\gamma_n^2 = \frac{1}{-2\pi i \cdot (4n) \cdot (\tilde Y_1)_{12}}.
\end{equation}
Collecting the above five formulas together, we obtain
\begin{equation} \label{asymptotic of gamma-n}
\gamma_n^2=\frac 1{\pi}2^{2\alpha+1}(4n)^{-(2n+\alpha+1)}e^{-nl}(1+O(n^{-1/3})),
\end{equation}
where the error term is uniform for $ t\in(0,d],~d>0$. It is readily seen that \eqref{asymptotic of gamma-n} agrees with \eqref{gamma-n-asy}. Now combining the above formula and \eqref{Y-11(0)Y-12(0)}, we have
\begin{equation}\label{an asymptotics}
a_n(t) = 2\pi i t\gamma_n^2 Y_{11}(0)Y_{12}(0) = t r'(2nt) \left [ 1 + O( {n^{-1/3}}) \right ].
\end{equation}
Thus, \eqref{alpha-n-asy} follows the above formula and \eqref{an-def}.
A full proof of \eqref{gamma-n-asy} can be obtained as follows. Note that $h_n = \gamma^{-2}_n(t)$ and $\frac{d}{dt} w(x;t) = -\frac{1}{x} w(x)$. Then, taking derivative with respective to $t$ on both sides of the orthogonality condition \eqref{op-pin}, we get
\begin{equation}\label{log derivative gamma}
\frac {\gamma_n'(t)}{\gamma^3_n(t)}=\frac 12\int_0^{\infty} \frac{\pi_n^2(y)w(y)}{y}dy.
\end{equation}
This, together with \eqref{an integral from chen}, gives us
\begin{equation}\label{an and gamma}
2t \frac {d}{dt}\ln(\gamma_n(t))=a_n(t).
\end{equation}
Recalling \eqref{an asymptotics} and integrating the above formula, we have
\begin{equation}\label{log an asy}
\ln (\gamma_n(t))-\ln(\gamma_n(0))=\frac{r(s)-r(0)}{4n} (1+O(\frac 1{n^{1/3}})).
\end{equation}
Note that $\gamma_n(0)$ is the leading coefficients of the orthonormal Laguerre polynomials, which can be computed explicitly as follows
\begin{equation}\label{leading coeff laguerre}
\gamma_n(0)=\sqrt{\frac{D_n[w,0]}{D_{n+1}[w,0]}}=\frac {1}{\sqrt{n!\Gamma(n+1+\alpha)}};
\end{equation}
see \eqref{zn0}. Finally, substituting \eqref{leading coeff laguerre} into \eqref{log an asy} yields \eqref{gamma-n-asy}, where the initial value of $r(0)$ in \eqref{r boundary conditions} is also used.
The asymptotic formula \eqref{beta-n-asy} for $\beta_n(t)$ follows from the asymptotic formula for $H_n(t)$ in \eqref{asymptotic of h-n} and Eq. (3.22) in \cite{ci} given below
\begin{equation} \label{beta-n hn relation from chen}
\beta_n(t) = n(n+\alpha) + t H_n'(t) - H_n(t).
\end{equation}
Finally, we substitute the asymptotic formulas \eqref{alpha-n-asy} and \eqref{asymptotic of h-n} into the equations \eqref{an-PIII} and \eqref{hn-eqn}, and then collect the leading terms. The equations \eqref{an-PIII-limit} and \eqref{hn-eqn-limit} follow immediately.
This completes the proof of Theorem \ref{thm3}. \hfill\qed
\section*{Acknowledgements}
The work of Shuai-Xia Xu was supported in part by the National
Natural Science Foundation of China under grant number
11201493, GuangDong Natural Science Foundation under grant number S2012040007824, Postdoctoral Science Foundation of China under Grant No.2012M521638, and the Fundamental Research Funds for the Central Universities under grand number 13lgpy41.
Dan Dai was partially supported by a grant from City University of Hong Kong (Project No. 7004065) and a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11300814).
Yu-Qiu Zhao was supported in part by the National
Natural Science Foundation of China under grant number
10871212.
|
\section{Introduction}
The efficient computation of matrix exponentials has been extensively considered in the literature and the \defn{scaling and squaring method} is perhaps the most widely used method for matrices of dimension $n\times n$ with $n$ as large as a few hundred (see \cite{higham09tsa,moler03ndw,sidje98eas} and references therein).
For example, Matlab and Mathematica compute numerically the exponential of matrices using this method where highly efficient algorithms for general matrices exist \cite{almohy09ans,higham05tsa,higham09tsa,higham10cma}.
Given $A \in\C^{n\times n}$, the method is based on the property
\begin{equation}
\e^A = \left( \e^{A/2^s} \right)^{2^s} =
\underbrace{\left( \cdots \left.\left( \e^{A/2^s} \right)^{2}\right.^{\Ddots}\right)^{2} }_{s-\text{times}},
\end{equation}
where typically $\e^{A/2^s}$ is replaced by a polynomial approximation (e.g. a $m$th-order Taylor method, $T_m(A/2^s)$) or a rational approximation (e.g. an $2m$th-order diagonal Pad\'e method, $r_{2m}(A/2^s)$) \cite{higham09tsa,higham10cma,sastre14aae}.
The optimal choice of both $s$ and the algorithms to compute $\e^{A/2^s}$ usually depend on the value of $\|A\|$ and the desired tolerance, and
have been deeply analyzed.
The computational cost, $c(\cdot)$, is usually measured by the number of matrix--matrix products, so $c(\e^A) = s+c(\e^{A/2^s})$, where $c(\e^{A/2^s})$ has to be replaced by the cost of its numerical approximation, e.g. $c(T_{m}(A/2^s))$ or $c(r_{2m}(A/2^s))$. Given a tolerance, one has to look for the scheme which provides such accuracy with the minimum number of products
(see \cite{higham09tsa,higham10cma} and references therein).
In some cases, if the matrix $A$ has a given structure, more efficient methods can be obtained \cite{celledoni00ate,celledoni01mft} . For example, to compute the exponential of upper or lower triangular matrices, in \cite{almohy09ans} the authors show that it is advantageous to exploit the fact that the diagonal
elements of the exponential are exactly known.
It is then more efficient to replace the diagonal elements obtained using e.g. Taylor or Pad\'e approximations by the exact solution before squaring the matrix (this technique can also be extended to the first super (or sub-)diagonal elements).
On the other hand, in many cases the matrix $A$ can be considered as a small perturbation of a sparse
matrix $D$, i.e., $A=D+B$ with $\|B\|<\|D\|$ (and frequently $\|B\|\ll\|D\|$) where $\e^D$ is sparse and exactly solvable (or can be accurately and cheaply approximated numerically), and $B$ is a dense matrix.
This is the case, for example, if $D$ is diagonal (or block diagonal with small matrices along the diagonal), or if it is diagonalizable using only a few elementary transforms. This is also the case, for example, if $n=2k$ and
\[
D=\left(\begin{array}{cc} 0 & I \\ -\Omega^2 & 0 \end{array} \right)
\]
where $I$ is the $k\times k$ identity matrix and $\Omega$ is a diagonal matrix where $\e^D$ is also an sparse and trivial to compute matrix. This problem can be originated from a semidiscretization of a hyperbolic PDE or from a set of $k$ linearly coupled oscillators.
As a motivational example, let us consider the linear time-dependent system of differential equations
\[
\frac{d}{dt} X = M(\eps t) X, \qquad X(t_0)=X_0 \in\C^{n\times n}
\]
with $M \in\C^{n\times n}$ and $|\eps|\ll 1$, i.e., $M(\eps t)$ evolves adiabatically with the variable $t$. Suppose that $M(\eps t)$ is instantaneously diagonalizable, i.e., $M(\eps t)=Q(\eps t)D(\eps t) Q^{-1}(\eps t)$ with $D$ a diagonal matrix.
Then, we can consider what it is usually called the adiabatic picture in quantum mechanics (if $M$ is a skew-Hermitian matrix), i.e., the change of variables, $X=Q(\eps t)Y$ where $Y$ is the solution of the differential equation
\[
\frac{d}{dt} Y = \left(D - Q^{-1}\frac{d}{dt}Q\right) Y, \qquad
Y(t_0)=Q^{-1}(\eps t_0)X_0.
\]
A second order method in the time step $h$ which advances the solution from $t_i$ to $t_i+h$, where $Y_i\approx Y(t_i)$, is given by
\begin{equation}\label{eq.midpoint}
Y_{i+1} = e^{h\left(D_{1/2} + \eps B_{1/2}\right)} Y_{i},
\end{equation}
where
\[
D_{1/2}= D\left(\eps\left(t_{i+1/2}\right)\right) , \qquad
\eps B_{1/2}= - Q^{-1}\left(\eps \left(t_{i+1/2}\right)\right)
\frac{d}{dt}Q\left(\eps \left(t_{i+1/2}\right)\right),
\]
with $t_{i+1/2}=t_i+\frac{h}2$. Notice that $\eps B_{1/2}$ is, in general, a dense matrix with a small norm (proportional to $\eps$) due to the term $\frac{d}{dt}Q(\eps t)$.
It is then natural to look for methods that approximate the exponential \eqref{eq.midpoint} at a low computational cost while providing sufficient accuracy.
Notice that in most cases in practice it is not necessary to approximate the exponential up to round-off accuracy since the model/method itself does not reproduce the exact solution within round-off precision. However, the preservation of qualitative properties (e.g. orthogonality, symplecticity, unitarity, etc.) is in some cases of great interest \cite{iserles00lgm}.
The aim of this work is the exploration of new and more efficient algorithms which take advantage of the fact that $\e^D$ is sparse and known at a cheap computational cost and that $B$ has a small norm.
The schemes we analyze in continuation are based on splitting and composition techniques tailored for this particular problem.
For clarity in the presentation, we take the partition $s=s_1+s_2$, we set $h=2^{-s_2}, N=1/h=2^{s_2}$ and replace $B$ by $\eps B$ with $\|B\|\sim\|D\|$, and we propose a new recursive procedure that we refer as Modified Squaring
\begin{equation} \label{eq.Recursive}
X_0 = e^{b h \eps B}, \qquad
X_k = X_{k-1}e^{a_k h D}X_{k-1}, \qquad k=1,\ldots, s_1
\end{equation}
and $Y_{s_1} = e^{a_{s_1+1}hD}X_{s_1}e^{a_{s_1+1}hD}$ where $b=1/2^{s_1}$ and the parameters $a_k$ will be chosen properly to improve accuracy. The total cost is
\[
c(Y_{s_1}^{s_2}) = s_1 + s_2 + c(e^{b h \eps B})
\]
where $c(e^{b h \eps B})=c(e^{\eps B/2^s})$ is the cost to approximate this exponential. Since $\|h\eps B\|$ is very small, a low-order diagonal Pad\'e approximation can provide sufficient accuracy (for most problems it will suffice just to consider $r_2$ or $r_4$ which only require one inversion or one inversion and one product, or even a low-order Taylor approximation can also be used).
The choice $s_1=0$ corresponds to the Leapfrog or Strang method,
\begin{equation} \label{eq.leapfrog}
\e^{h(D+\eps B)} \approx \e^{hD/2} \e^{h\eps B} \e^{hD/2},
\end{equation}
where, as already mentioned, $\e^{hD/2}$ can be accurately and cheaply computed.
More accurate methods can be obtained using a general composition
\begin{equation}\label{eq.splitting}
S^{[m]}_p = \prod_{i=1}^m\e^{ha_iD} \e^{hb_i \eps B} \approx \e^{h(D+\eps B)},
\end{equation}
where the coefficients $a_i,b_i$ are chosen such that $S_p^{[m]}$ is an approximation to the exact solution up to a given order, $p$, in the parameter $h$, i.e. $S_p^{[m]}=\e^{h(D+\eps B)}+\cO(h^{p+1})$.
However, to get efficient methods it is crucial to reduce the computational cost.
Since the cost is dominated by the exponentials $\e^{hb_i \eps B}$, it is advisable to reuse as many exponentials as possible, e.g., letting $b_i=1/m$, only one exponentiation is necessary.
However, this class of methods has some limitations since for orders greater than 2, at least one of the coefficients $a_i$ and one of the $b_i$ must be negative and thus might jeopardize the re-utilization of the exponentials.
However, for small perturbations, very accurate results can still be obtained with positive coefficients.
In the particular situation when $A\in \C^{n\times n}$, complex coefficients, $a_i\in \C$, can be used without increasing the computational cost, and then fourth-order methods with all $b_i$ real and equal are achievable. The proposed recursive algorithm (\ref{eq.Recursive}) corresponds to a particular case of an splitting method where the cost has been reduced while still leaving some free parameters for optimisation.
In this work, we assume that the product $B^2$ requires $\cO\left({{n}^3}\right)$ operations but $DB$ requires only $\cO\left({k{n}^2}\right)$ with $k\ll {n}$ (e.g. $c(B^2)=1, \ c(DB)=\delta$, with $\delta\ll 1$).
Then, the commutator $\eps[D,B]=\eps(DB-BD)$ can be computed at considerably smaller cost than the product of two dense matrices while retaining a small norm due to the factor $\eps$.
It then makes sense to consider the recursive algorithm (\ref{eq.Recursive}) where the exponential $e^{b h \eps B}$ is replaced by
\begin{equation} \label{eq.leapfrogMod}
\e^{bh\eps B+ \alpha h^3\eps [A,[A,B]]}
\end{equation}
whose computational cost is similar,
but more accurate results can be obtained if the scalar parameter $\alpha$ is properly chosen.
Further exploiting this approach leads to the inclusion of the term $\beta h^5\eps [A,[A,[A,[A,B]]]]$ in the central exponential, which again, for an appropriate choice of the parameter $\beta$, decreases the error at a similar computational cost. The analysis presented in this work is also extended to the case in which not all parameters $b_i$ are taken equal.
This paper is organized as follows:
Section~2 considers the computational cost of Pad\'e and Taylor methods as well as the cost of all operations involved in the splitting schemes analyzed in this work in order to develop new algorithms which minimize the whole cost.
In Section~3 we analyze the algebraic structure of the different families of methods considered to obtain the order conditions to be satisfied by the coefficients.
In Section~4 we propose a recursive algorithms to minimize the cost of the methods and we build new methods. An error analysis is carried in Section~5 and Section~6 illustrates the performance of the methods on several numerical examples. Finally, Section~7 presents the conclusions and the appendix collects, for completeness, several new families of splitting methods which have also been analyzed.
\section{Computational cost of matrix exponentiation}
\subsection{Computational cost of Taylor and Pad\'e methods}
We first review the computational cost of the optimized Taylor and Pad\'e methods which are used in the literature and that are used as reference in the numerical examples.
\paragraph{Taylor methods}
We use the Paterson-Stockmeyer scheme (see \cite{higham08fom,higham10cma,paterson73otn}) to evaluate $T_m=\sum_{k=0}^mA^n/n!$ which minimize the required number of products.
From the Horner-scheme-like computation,
given a number of matrix products $2k$, the maximal attainable order is $m=(k+1)^2$.
In \cite{higham10cma}, it is indicated that the optimal choice for most cases corresponds to $k=3$, i.e. order $m=16$
with just 6 products given by: $A^2=AA, \ A^3=A^2A, \ A^4=A^2A^2$ and
\[
T_{16}(A) = g_0+(g_1+(g_2+(g_3+g_4 A^4) A^4) A^4) A^4,
\]
where $g_i$ are linear combinations of already computed matrices,
$g_i=\sum_{k=0}^4 c_{i,k} A^k$, with $c_{i,k}=1/(4i+k)!$ for $i=0,1,2,3$ and $g_4=\id/16$ proportional to the identity (matrix).
\paragraph{Diagonal Pad\'e methods}
Diagonal Pad\'e methods are given by the rational approximant
\begin{equation}\label{eq.24}
r_{2m}(A )=\frac{p_{m}(A )}{p_{m}(-A )},
\end{equation}
provided the polynomials $p_{m}$ are generated by the recurrence
\begin{align}
p_{0}(A ) &= \id, \ \qquad p_{1}(A )=2\id+A \nonumber \\
p_{m}(A ) &=2(2m-1)p_{m-1}(A )+A^{2}p_{m-2}(A ). \label{eq.RecPade}
\end{align}
Moreover,
$r_{2m}(A )= \e^{A }+\mathcal{O}(A^{2m+1})$, whereas
for $m=1,2$ we have
\begin{equation}\label{eq.Pade24}
r_{2}(A ) = \frac{\id+A/2}{ \id-A/2}, \qquad \qquad
r_{4}(A ) = \frac{\id+A/2+A ^{2}/12}{ \id-A/2+A ^{2}/12}.
\end{equation}
The recursive algorithm (\ref{eq.RecPade}) is, however, not an efficient way to compute $r_{2m}(A )$.
For example, the method $r_{26}(A)$ is considered among the optimal choices (with respect to accuracy and computational cost) of diagonal Pad\'e methods when round off accuracy is desired and $\|A\|$ takes relatively large values.
The algorithm to compute it is given by
\begin{equation}
(-u_{13}+v_{13}) r_{26}(A) = (u_{13}+v_{13}),
\end{equation}
with
\begin{align*}
u_{13} & = A[A_6(b_{13}A_6 + b_{11}A_4 + b_9A_2) + b_7A_6 + b_5A_4 + b_3A_2 + b_1 \id ], \\
v_{13} & = A_6(b_{12}A_6 + b_{10}A_4 + b_8A_2) + b_6A_6 + b_4A_4 + b_2A_2 + b_0\id,
\end{align*}
where $A_2=A^2, A_4=A_2^2, A_6=A_2A_4$.
Written in this form, it is evident that only six matrix multiplications and one inversion are required.
In a similar way, the method $r_{10}(A)$, which will be used in this work, only requires 3 products and one inversion.
\subsection{Computational cost of splitting methods}
Recall that we are considering a sparse and sparsely exponentiable matrix $D$, while $B$ is a dense matrix and responsible for the numerical complexity.
In order to build competitive algorithms, it is important to analyze - under these assumptions - the computational cost of all operations involved in the different classes of splitting and composition methods.
Let $X,Y$ be two dense $n\times n$ matrices and denote by $c(\cdot)$ the cost of the operations in brackets as the number of matrix--matrix products of dense matrices, e.g., $c(X Y)=1$ and $c(X+ Y)=\delta$, with $\delta\ll 1$, thereby neglecting operations with a lower complexity in the number of operations. According to this criterion, we derive Table~\ref{tab.cost}, where the dominant terms are highlighted in boldface (the cost for the inverse of a matrix is taken as 4/3 the cost of a matrix-matrix product).
\begin{table}[!ht]
\begin{center}\footnotesize
\begin{tabular}{lll}
\toprule
& Operation& Effort\\[1mm]
Sum & $c(D+ D)\approx 0$ &${\cal O}(k\, n)$, with $k\ll n$\\
& $c(X+ Y)=\delta$&${\cal O}(n^2)$\\[.5mm]
Product & $c(XY)={\bf 1}$&${\cal O}(n^3)$\\
& $c(D D)=0$&${\cal O}(k^2\, n)$\\
&$c(D X)=k\delta$&${\cal O}(k\, n^2)$\\
Inversion& $c(X^{-1} Y)={\bf 1+\frac13}$&$c(X^{-1} Y)=\frac43c(X Y)$\\[.5mm
Commutation
&$c([D,X]) = c(D X-X D)=2k\delta$ & ${\cal O}(k\, n^2)$\\
&$c([D,[D,\ldots,[D,X]\cdots]])=2rk\delta$&${\cal O}(k\, n^2)$\\[.5mm]
Exponentiation & $c(\e^D)=wk\delta$&${\cal O}(k^2\, n)$\\%\label{cost8}
& $c(r_{2}(X))={\bf 1+\frac13}$ & ${\cal O}(n^3)$\\
& $c(r_{4}(X))={\bf 2+\frac13}$ & ${\cal O}(n^3)$\\% \label{cost10}
\bottomrule
\end{tabular}
\end{center}
\caption{\label{tab.cost}Computational cost of matrix operations for the sparse and sparsely exponentiable matrix $D$ and arbitrary dense matrices $X,Y\in\mathbb{C}^{n\times n}$. The factor $w$ in $c(e^D)$ is assumed to be small, $w\ll1$.}
\end{table}
Based on this analysis, we examine the splitting method \eqref{eq.splitting} to identify the computationally relevant aspects.
In this work we assume $\delta\ll 1$ and in our computations we will take $\delta=0$ for simplicity.
First, we have to choose how to approximate the exponentials $\e^{h\eps b_i B}$ taking into account that
\begin{eqnarray}
r_2(h\eps b_i B)&=& \e^{h\eps b_i B} +{\cal O}(h^3 \eps^3) \label{eq.Pade2} , \\
r_4(h\eps b_i B)&=& \e^{h\eps b_i B} +{\cal O}(h^5 \eps^5) \label{eq.Pade4} .
\end{eqnarray}
A rough estimate for the composition \eqref{eq.splitting}, assuming all coefficients $b_i$ different, and taking into account the cost shown in Table~\ref{tab.cost}, we have
\[
c(S^{[m]}_p,r_2)=m\frac43+m-1=\frac73 m -1, \qquad
c(S^{[m]}_p,r_4)=m\frac73+m-1 = \frac{10}{3}m-1,
\]
where $c(S^{[m]}_p,r_i)$ denotes the cost of the method $S^{[m]}_p$ when the exponentials $\e^{\eps B}$ are approximated by $r_i(\eps B)$.
Repeating the coefficients $b_i$, i.e., $b_i=1/m, \ i=1,\ldots,m$, the computational cost can be reduced considerably, in this case, one gets
\[
c(S^{[m]}_p,r_2)=\frac43+(m-1)=m+\frac13, \qquad
c(S^{[m]}_p,r_4)=m+\frac43.
\]
Further simplifications are applicable and will be discussed in Sect.~\ref{sec:splitting}.
\section{The Lie algebra of perturbed systems: $(p_1,p_2)$ methods}
Following the terminology of \cite{mclachlan95cmi}, we introduce a modified error concept which is suitable for the near-integrable structure of the matrix $A$ at hand.
Letting $S^{[m]}_p$ be a $p$th-order $m$-stage consistent ($\sum_i a_i=\sum_i b_i=1$) splitting method \eqref{eq.splitting}, we expand its error as
$$
S^{[m]}_p - e^{hA} = \sum_{i=p+1} \sum_{j=1} e_{i,j} \eps^j h^{i} C_{i,j},
$$
where $e_{i,j}$ is a polynomial in the splitting coefficients $a_k,b_k$ and $C_{i,j}$ is a sum of matrix products consisting of all combinations containing $(i-j)$ sparse elements $D$ and $j$ times $B$.
Notice that in addition to the scaling $h$, we also expand in powers of the small parameter $\eps$.
The method is said to be of order $p=(p_1, p_2, \ldots)$ if $e_{i_1,1}=e_{i_2,2}=\ldots=0$ for all $i_k\leq p_k$ and $p_1\geq p_2\geq \cdots$.
Designing a method now consists of identifying the dominant error terms $e_{i,j} \eps^j h^{i}$
and finding coefficients $a_j, b_j$ to zero the polynomial $e_{i,j}$.
The main tool in this endeavor is the Baker-Campbell-Hausdorff formula which provides a series expansion of the single exponential that has been actually computed when multiplying two matrix exponentials,
$$
e^{hA}e^{hB} = e^{\bch(hA,hB)}, \quad \bch(hA,hB) = h(A+B) + \frac{h^2}{2}[A,B] + \mathcal{O}(h^3).
$$
Recursive application of this formula to a symmetric splitting \eqref{eq.splitting} establishes the concept of a modified matrix $h\tilde{{A}}$, along the lines of backward-error-analysis,
\begin{multline}\label{eq.backward}
\log(S^{[m]}_p) = h\tilde{{A}} =
hA +\tilde e_{3,1}\varepsilon h^3[D,[D,B]] +\tilde e_{3,2}\varepsilon^2 h^3[B,[D,B]]\\
+\tilde e_{5,1}\varepsilon h^5[D,[D,[D,[D,B]]]] +\tilde e_{5,2}\varepsilon^2 h^5[[D,[D,B]],[D,B]]\\ +\tilde e_{5,3}\varepsilon^2 h^5[B,[D,[D,[D,B]]]]
+\tilde e_{7,1}\varepsilon h^7[D,[D,[D,[D,[D,[D,B]]]]]] + \mathcal{O}\left(\varepsilon^3 h^5 + \varepsilon^2 h^7\right),
\end{multline}
where the $\tilde e_{i,j}$ are also polynomials in the splitting coefficients $a_k,b_k$ which multiply elements of the Lie algebra and are different from the coefficients $e_{i,j}$.
Higher-order terms can be computed by efficient algorithms \cite{casas09afa}.
\subsection{Error propagation by squaring}
The splitting method (\ref{eq.backward}) can also formally be written as
\begin{equation} \label{eq.Sp1p2}
S^{[m]}_{(p_1,p_2)} = \exp\left( h(D+\varepsilon B) +
\varepsilon \sum_{k>p_1} c_k h^k[D^k,B]+ \mathcal{O}\left(\varepsilon^2 h^{p_2+1}\right) \right)
\end{equation}
where $[D^k,B]=[D,[D,[\ldots,[D,B]\ldots]]]$ and there is only one term proportional to $\varepsilon$ at each power of $h$.
We can then define a \textit{processor}, a close to the identity map
\begin{equation} \label{eq.Proc1}
P = \exp\left( -\varepsilon \sum_{k>p_1} c_k h^{k-1}[D^{k-1},B] \right),
\end{equation}
such that the method can be written as
\begin{equation} \label{eq.Sp1p2Proc}
S^{[m]}_{(p_1,p_2)} = PKP^{-1},
\end{equation}
with
\begin{equation} \label{eq.Sp1p2K}
K= \exp\left( h(D+\varepsilon B) + \mathcal{O}\left(h^{p_2+1}\varepsilon^2\right) \right).
\end{equation}
Suppose now that the matrix $A$ can be diagonalized, $A=QD_AQ^{-1}$, then clearly
\begin{equation*}
\e^{A}=Q\e^{D_A}Q^{-1}.
\end{equation*}
The \textit{kernel} $K$ of the numerical method, on the other hand, can be diagonalized for sufficiently small $h=1/n$ and $\varepsilon$ using
\[
\hat Q = Q + \mathcal{O}\left(h^{p_2+1}\varepsilon^2\right), \qquad
\hat D_A = hD_A + \mathcal{O}\left(h^{p_2+1}\varepsilon^2\right),
\]
such that, after $n$ integration steps, we obtain
\begin{equation}\label{eq.Kn}
K^n=\hat Q\e^{\tilde D_A}\hat Q^{-1}.
\end{equation}
with $ \tilde D_A = D_A + \mathcal{O}\left(nh^{p_2+1}\varepsilon^2\right)$.
The size estimates of the above considerations lead to a favorable error propagation result which is stated in the following theorem.
\begin{theorem}\label{theorem1}
Let $A=D+\varepsilon B$ a diagonalizable matrix such that $\e^A$ is bounded and let $S^{[m]}_{(p_1,p_2)}$ be a splitting method that approximates the scaled exponential $\e^{hA}$ with $h=1/n$. Then,
for sufficiently small values of $h$ and $\varepsilon$ we have that
\begin{equation}\label{eq.thm1}
\left\|\e^A - \left( S^{[m]}_{(p_1,p_2)} \right)^n \right\| \leq C_1 h^{p_1+1}\varepsilon + n C_2 h^{p_2+1} \varepsilon^2.
\end{equation}
where $C_1,C_2$ are constants which do not depend on $h$ and $\varepsilon$.
\end{theorem}
\begin{proof}
From (\ref{eq.Sp1p2Proc}) and (\ref{eq.Kn}) we have that
\begin{equation} \label{eq.Sp1p2Proc2}
\left(S^{[m]}_{(p_1,p_2)}\right)^n = P\hat Q\e^{\tilde D_A}\hat Q^{-1}P^{-1}
= \tilde Q\e^{\tilde D_A}\tilde Q^{-1}
\end{equation}
where now $\tilde Q = P\hat{Q} = Q + \mathcal{O}\left(h^{p_1+1}\varepsilon\right)$. Then
\begin{align}\label{eq.proof1}
\left\|\e^A - \left( S^{[m]}_{(p_1,p_2)} \right)^n \right\| & =
\left\|Q\e^{D_A}Q^{-1} - \tilde Q\e^{\tilde D_A}\tilde Q^{-1} \right\| \nonumber \\
& =
\left\|Q\e^{D_A}Q^{-1} - \tilde Q\e^{D_A}Q^{-1} + \tilde Q\e^{D_A}Q^{-1} - \tilde Q\e^{\tilde D_A}\tilde Q^{-1} \right\| \nonumber \\
& \leq
\|Q - \tilde Q\| \ \|\e^{D_A}Q^{-1}\| + \|\tilde Q\| \ \|\e^{D_A}Q^{-1} - \e^{\tilde D_A}\tilde Q^{-1} \| \nonumber.
\end{align}
The right summand is expanded in a similar way to
\begin{align
\|\e^{D_A}Q^{-1} - \e^{\tilde D_A}\tilde Q^{-1} \| & =
\|\e^{D_A}Q^{-1} - \e^{\tilde D_A}Q^{-1} + \e^{\tilde D_A}Q^{-1} - \e^{\tilde D_A}\tilde Q^{-1} \| \\
& \leq
\|\e^{D_A} - \e^{\tilde D_A}\| \ \|Q^{-1}\| + \|\e^{\tilde D_A}\| \ \|Q^{-1} - \tilde Q^{-1} \|. \nonumber
\end{align}
Taking into account that $\tilde D_A = D_A + \mathcal{O}\left(nh^{p_2+1}\varepsilon^2\right)$, $\tilde Q = Q + \mathcal{O}\left(h^{p_1+1}\varepsilon\right) $, and that $\e^A$ is bounded we obtained the desired result for sufficiently small values of $h$ and $\varepsilon$.
\end{proof}
This result indicates that the error is the sum of a local error of order $\mathcal{O}\left(\varepsilon\right)$ plus a global error of order $\mathcal{O}\left(\varepsilon^2\right)$.
For problems which require a relatively large number of squaring (a large value of $n=2^s$) the dominant error of the splitting methods is proportional to $\varepsilon^2$. Then, to build methods which are accurate for different values of $s$ it seems convenient to look for methods of effective order $(p_1,p_2)$ with $p_1>p_2$
The following numerical example illustrates the results obtained.
{\bf Example} Let
\beq\label{eq.example}
A= \left( \begin{array}{cc} \eps & 1+\eps \\ -1+\eps & - \eps
\end{array}\right), \qquad \qquad
D= \left( \begin{array}{cc} 0 & 1 \\ -1 & 0
\end{array}\right)
\eeq
with $\eps=10^{-1},10^{-3}$, and approximate $\e^{2^sA}=\left(\cdots \left(\e^A\right)^2\cdots\right)^2$ to a relatively low accuracy. To approximate $\e^{A}$, we consider a fourth-order Taylor method, $T_4(A)$ (that only requires 2 products) and a fourth-order Pad\'e approximation, $r_4(A)$ (with a cost of one product and one inversion, equivalent to $1+4/3$ products).
We compare the obtained results with the second-order splitting method \eqref{eq.leapfrog}, which we denote by $S_2^{[2,a]}$ or, since in this case $p_1=p_2=2$, $S_{(2,2)}^{[2,a]}$, where the exponential $\e^D$ is computed exactly and $\eps B$ is approximated with the second order diagonal Pad\'e method, $r_2(\eps B)$.
The exact solution is given by
\[
\e^{2^sA}= \left( \begin{array}{cc}
\cos(2^s\mu) + \frac{\eps}{\mu} \sin(2^s\mu) & \frac{1+\eps}{\mu}\sin(2^s\mu) \\
- \frac{1-\eps}{\mu} \sin(2^s\mu) & \cos(2^s\mu) - \frac{\eps}{\mu} \sin(2^s\mu)
\end{array}\right)
\]
with $\mu=\sqrt{1-2\eps^2}$ and we analyze the error growth due to the squaring process in Fig.~\ref{fig.longtime}.
We observe that neither Pad\'e nor Taylor methods are sensitive w.r.t. the small parameter, whereas the splitting method drastically improves when decreasing $\varepsilon$.
The splitting method is only of second order and thus used with a second order Pad\'e method $r_2$ (using the fourth order method $r_4$ leaves error plot unchanged).
Notice that for the small perturbation $\varepsilon=10^{-3}$, the splitting with $r_2(\eps B)$ is more accurate than the fourth-order Pad\'e $r_4(A)$ which comes at nearly twice the computational cost (1 inversion vs. 1 inversion and 1 dense product). According to Theorem~\ref{theorem1}, the error of $S_{(2,2)}^{[2,a]}$ is the sum of a local error proportional to $h^{3}\varepsilon$ and a global error proportional to $n h^{3} \varepsilon^2$, with $n=2^s$. Fig.~\ref{fig.longtime} shows the results obtained for different values of $\varepsilon^2$ and $s$ which clearly show both error sources.
\begin{figure}[!ht]
\centering
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=2,
minor y tick num=2,
minor z tick num = 2,
scale only axis,
font=\footnotesize,
}
}
\setlength\figurewidth{.4\textwidth
\setlength\figureheight{.2\textwidth}
\tikzsetnextfilename{figs/longtime_final}
\includegraphics{longtime_final}
\caption{\label{fig.longtime} Error in the approximation to $\e^{2^sA}$ with $A$ given by \eqref{eq.example} for different values of $\varepsilon$ and number of squaring, $s$, in double-logarithmic axes. The bottom figures show that the error of the splitting methods is proportional to $\varepsilon$ for small $s$ (local error) and proportional to $\varepsilon^2$ for large values of $s$ (global error)}
\end{figure}
\section{Splitting methods for scaling and squaring}\label{sec:splitting}
Taking into account the numerical effort established in the introduction, we derive methods which are optimal for the problem at hand.
The optimization principle becomes clear at the example of the two versions of Strang's second-order splitting method
\def\e{e}
\begin{align}\label{eq:StrangABA}
& &S_2^{[2,a]}&=
\e^{\frac{h}2 D} \ \e^{h \varepsilon B} \ \e^{\frac{h}2 D}
=\mathcal{D}_{h/2}\mathcal{B}_{h}\mathcal{D}_{h/2},\\
\label{eq:StrangBAB}
\text{and} & &
S_2^{[2,b]}&=
\e^{\frac{h}2 \varepsilon B} \ \e^{h D} \ \e^{\frac{h}2 \varepsilon B}
=\mathcal{B}_{h/2}\mathcal{D}_{h}\mathcal{B}_{h/2},
\end{align}
which differ in computational cost:
Using the notation $\mathcal{D}_h=e^{hD}$, $\mathcal{B}_h=e^{hB}$, and keeping in mind that $\mathcal{D}_{h}$ is a sparse matrix while $\mathcal{B}_{h}$ is dense, the dominant numerical cost amounts to a single exponential with $\cost(S_2^{[2,a]})=\cost(\mathcal{B}_{h})$ for the first version, whereas the latter requires an additional matrix product, $\cost(S_2^{[2,b]})=\cost(\mathcal{B}_{h/2})+\cost(\mathcal{B} \mathcal{B})$.
Furthermore, the large dominant part $D$ is multiplied by $1/2$ before exponentiation in the cheaper variant which is advantageous in the sense of the scaling process.
We follow a variety of strategies in order to develop new methods and group them according to the splitting terminology, keeping in mind that the costly parts are products and exponentials of the dense matrices $\mathcal{B}$ and $B$, respectively.
\subsection{Standard splittings}
As we have discussed for the Strang splitting $S_2^{[2,b]}$, despite the appearance of $B$ in two exponents, only one exponential actually has to be computed which is then stored and reused for the second identical exponent.
Generalizing this principle, we search for splitting methods $a_i,b_j$ where all $b_j=b$ are identical to reduce the computational effort which now comes solely from the dense-matrix multiplications.
A composition that is also symmetric in the coefficients $a_j$ will reduce a great number of error terms (since even powers in $h$ disappear) and additionally the amount of (cheap) exponentials $\cD$ to be computed.
Next, we derive a particular family of splittings which can be understood in analogy to squarings and allows to reduce the necessary products.
\subsubsection{Modified squarings}
We propose to replace a given number of squarings by a one-step splitting method which has the benefit of free parameters to minimize the error.
For illustration, let us compute a squaring step, $h=2^{-1}$, of the standard Strang method,
\begin{equation} \label{eq.2Strangs}
(e^{h/2A}e^{hB}e^{h/2A})^2 = e^{\frac14 A}e^{\frac12 B}e^{\frac12 A}e^{\frac12 B}e^{\frac14 A},
\end{equation}
which we then contrast with a general splitting method at the same cost (one exponential and one product) without squaring ($h=1$),
\begin{equation} \label{eq.2Steps}
e^{a_2 A}e^{\frac12 B}e^{a_1 A}e^{\frac12 B}e^{a_2 A}.
\end{equation}
It is evident that (\ref{eq.2Steps}) includes (\ref{eq.2Strangs}) as a special case (choosing $a_1=1/2, a_2=1/4$) and we use the example \eqref{eq.example} to illustrate the gains in accuracy. Fig.~\ref{fig.modified} shows that the performance is very sensitive to the choice of the free parameter and the method of effective order $(4,2)$ is very close to the optimal one.
\begin{figure}[!ht]
\begin{center}
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=2,
minor y tick num=2,
minor z tick num = 2,
scale only axis,
font=\footnotesize,
}
}
\setlength\figurewidth{.4\textwidth
\setlength\figureheight{.4\textwidth}
\tikzsetnextfilename{figs/squaring}
\includegraphics{squaring}
\end{center}
\caption{\label{fig.modified}Modified squarings. All methods apart from $r_2(A)$ (green solid) have approximately the same numerical cost since the split uses 2nd order pad\'e}
\end{figure}
A larger number of squarings $s$ can be replaced by a recursive procedure,
$$
X_0 = e^{h b \eps B}, \qquad
X_k = X_{k-1}e^{a_k h D}X_{k-1}, \qquad k=1,\ldots, s
$$
and $Y_{s} = e^{a_{s+1}h}X_{s}e^{a_{s+1}h}$ where $b=1/2^s$.
The costly multiplications occur in the consecutive steps, $X_k$, where we recycle already computed blocks while introducing free parameters $a_k$ at negligible extra effort. As a result, the cost of the algorithm is
\[
c(Y_{s}) = s + c(e^{h b \eps B})
\]
where it usually suffices to approximate $e^{h b \eps B}$ with a second or fourth-order Pad\'e method, so $c(e^{h b \eps B},r_2)=\frac43$ and $c(e^{h b \eps B},r_4)=1+\frac43$.
For consistency, the coefficients $a_k$ have to satisfy
$$
\left( 2^{s-1} a_1 + \cdots+ 2 a_{s-1} + a_{s} \right)+ 2a_{s+1} = \sum_{k=1}^{s} 2^{s-k} a_k + 2 a_{s+1}= 1.
$$
Notice that the choice $a_{s+1}=1/2^{s+1}$, $a_k=1/2^s$ for $k=1,\ldots, {s}$, corresponds to the standard scaling and squaring applied to the Strang method \eqref{eq:StrangABA}.
In the following, we have collected the most efficient splitting methods for an increasing numbers of products $s=0,1,2,3,4$.
We have observed in the numerical experiments that for $s>4$, the gain w.r.t. to standard scaling and squaring is marginal, and they are not considered in this work.
However, the parameter $h$ demonstrates how any such method can be combined with standard scaling and squaring.
This procedure is equivalent to consider the partition $s=s_1+s_2$ where the first $s_1$ squarings are carried out with the recursive algorithm with $b=1/2^{s_1}$ and we continue with the remaining standard $s_2$ squarings with $h=1/2^{s_2}$.
\paragraph{$s_1=0$}
Strang $S_2^{[2,a]}$ with
local order $\cO(\eps h^3)$.
\paragraph{$s_1=1$}
After imposing symmetry, one free parameter remains and is used to obtain (4,2) methods \cite{laskar01hos,mclachlan95cmi},
\begin{equation}\label{eq:42ABA}
Y_1 =
\mathcal{D}_{ha_2}\mathcal{B}_{h/2}\mathcal{D}_{ha_1} \mathcal{B}_{h/2}\mathcal{D}_{ha_2},
\end{equation}
where $a_2=(3-\sqrt{3})/6, \ a_1=1-2a_2$ and with
local order $\cO(\eps h^5+\eps^2 h^3)$.
\paragraph{$s_1=2$}
Allowing an additional product, at $b=1/4$, we have
\beq\label{eq.five}
Y_2 = \cD_{a_3h} (\cB_{h/4} \cD_{a_2h} \cB_{h/4}) \cD_{a_1h} (\cB_{h/4} \cD_{a_2h} \cB_{h/4}) \cD_{a_3h}.
\eeq
Optimizing the free parameters $a_3, a_2$, (where for consistency $a_1=1-2(a_3+a_2)$) we can construct fourth-order methods, although complex-valued, with $a_3 = \frac1{10}(1-i/3), a_2 = \frac{2}{15}(2+i)$ and their complex conjugates $a_i^*$ \cite{castella09smw}.
Alternatively, there are six real-valued (6,2) methods, the best of which is given in Table~\ref{tab.coefs}.
\paragraph{$s_1=3$}
The three parameters for $Y_3$ can be used to produce complex-valued methods of order (6,4) or real-valued methods of order (8,2), the ones with smallest error coefficients can be found in Table~\ref{tab.coefs}.
\paragraph{$s_1=4$}
The next iteration yields a 17-stage method $Y_4$.
Its four parameters can be used to cancel the error coefficients $e_{3,1}, e_{3,2}, e_{5,1}, e_{7,1}$
for 48 complex (8,4) methods, or a $(10,2)$ method with positive real coefficients, see Table~\ref{tab.coefs}.
\begin{table}[!ht]
\caption{\label{tab.coefs}Modified squarings with and without commutators. In the right column, the corresponding computational cost is given together with the number of omitted solutions of the order conditions.}
\noindent{\footnotesize
\begin{tabular}{ll}
\toprule
$Y_2$, order (6,2
& $\cost(\cB_{h/4}) + 2\cost(\cB \cB)$\\[1mm]
$a_1 = \sqrt{(5-\sqrt{5})/30}, \
a_2 = \sqrt{ (5-2\sqrt{5})/15} $
&
[7 solutions omitted]\\
\midrule
$Y_3$, order (8,2)\\[2mm]
$a_1= 0.153942020841153420134790213164$ & only positive solution\\
$a_2= 0.089999237645462605679630986655$ & [47 omitted]\\
$a_3= 0.102244554291437558627161030779$ \\
$a_4=\frac12- (4 a_1 + 2a_2 + a_3)/2$.\\[1mm]
$Y_3$, order (6,4) & $\cost(\cB_{h/8}) + 3\cost(\cB \cB)$\\[2mm]
$a_1 = 0.13534452760420860194 + 0.06201309787740406230i$ & [7 omitted]\\
$a_2 = 0.13027125534284511606 - 0.10310039626441585374i$\\
$a_3 = 0.099062332740825337251 - 0.015885424766237390724i$\\
$a_4=\frac12- (4 a_1 + 2a_2 + a_3)$\\[1mm]
\midrule
$Y_4$, order (10,2)& $\cost(\cB_{h/16}) + 4\cost(\cB \cB) $\\[2mm]
$a_1 = 0.077255933048297137202077893145$ & only positive solution\\
$a_2 = 0.0444926322393204245189059370354$& [383 omitted]\\
$a_3 = 0.051080773613693429438027986467$ \\
$a_5 = 0.0254553659841308990458390646508$\\
$a_4 = 1 - 8a_1 - 4a_2 - 2a_3 - 2a_5$\\[1mm]
$Y_4$, order (8,4)\\[1mm]
$a_1 = 0.06782965853562196485274129 + 0.03038453954138687801299186i$ & [47 omitted]\\% =a2
$a_2 = 0.06477414774829711915884478 - 0.05170904068177844632921239i$ & \\% =a3
$a_3 = 0.04963134399080347125041612 + 0.00584283681423207753349501i$ \\
$a_5 = 0.02474856149827627051056177 - 0.00610084851840072905292033i$ \\% =a1
$a_4 = 1 - 8a_1 - 4a_2 - 2a_3 - 2a_5$\\
\midrule
\midrule
$\tilde{Y}_2$, order (6,4), minimizing $\cO(\varepsilon^2 h^5)$\\[2mm]
$a_1=(1-a_2-2a_3)/2$\\
$a_2=0.47071989362081947165$\\
$a_3 = 0.04898669326146179875$\\
$\beta = -0.002320917859694561351$\\
$\gamma=0.0000329546718228203782$\\[1mm]
$\tilde{Y}_2$, order (8,4) &[47 omitted] \\[1mm]
$a_1=0.3602258146389491220734647$ \\
$a_2=1 - 2(a_3 + a_1)$\\
$a_3=0.0766102130069293861483005$ \\
$\beta = -0.00103637077918270398691258$\\
$\gamma = 0.000010240482532598594411391$\\
\bottomrule
\end{tabular}
}
\end{table}
\subsection{Modified splittings}
A drastic improvement on the previous methods can be made through the use of commutators.
The special structure of the matrix allows for the fast computation of certain commutators, namely the ones that contain the matrix $B$ only once.
The inclusion of these commutators in the scheme will not only allow to reduce the number of error terms but also to reach order 4 using only real coefficients.
Since we are interested in symmetric methods of up to order (6,4), the relevant terms are
\begin{align*}
[D, [D, B]] &= DDB - 2DBD + BDD, \\
[D^4, B]
&= DDDDB - 4DDDBD + 6DDBDD - 4 DBDDD + BDDDD,
\end{align*}
and neglecting the numerical cost of summation and multiplication by a sparse matrix $D$, it is clear that the exponential
\[
e^{\alpha h B + \beta h^3 [D,[D,B]] + \gamma h^5 [D, [D, [D, [D, B]]]]} = \tilde{\cB}_{\alpha,\beta,\gamma}
\]
can be evaluated at the same cost as $\cB_{\alpha h}$.
Along the lines of the modified squarings, we have derived the following compositions which require only one exponentials $\tilde{B}$ at a fixed number of products.
The substitution $Y_s\to\tilde{Y}_s$ indicates the replacement of $B$ by $\tilde{B}$.
\paragraph{$s=0$} Strang's method can be made into a (6,2) scheme with
\beq\label{eq:strang62}
\tilde{Y}_0 = \cD_{h/2} \tilde{\cB}_{1, 1/24, 1/1920}\cD_{h/2}.
\eeq
We stress that, in principle, a method of order $(2n,2)$ can be constructed using only a single exponential, however, at the expense of increasingly complicated commutators, $[D,[D,[\ldots,[D,B]]\cdots]$ whose computational complexity cannot be neglected anymore.
\paragraph{$s=1$}
Replacing $\cB_{h/2}$ by $\tilde{\cB}$ in \eqref{eq:42ABA}, we obtain the (6,4) method
\beq
\tilde{Y}_1 = \cD_{ha_2}\tilde{\cB}\cD_{ha_1} \tilde{\cB}\cD_{ha_2},
\eeq
where $a_2=1/6, \ a_1=2/3$ and $\tilde{\cB}_{1/2, -1/144, 121/311040}$ with unchanged effort $\cost(\cB_{h/2})+\cost(\cB \cB)$.
\paragraph{$s=2$}
Using one additional multiplication, we reach $\tilde{Y}_2$,
which can be tuned to be of order (8,4) or (6,4) while minimizing the error at $\mathcal{O}(\varepsilon^2 h^5)$, see Table~\ref{tab.coefs}.
We have also analyzed other classes of splitting and composition methods. The methods obtained showed a worst performance on the numerical examples tested in this work. The schemes obtained are, however, collected in the appendix for completeness.
\section{Error analysis}
Our methods have proven successful for a low to medium accuracy since the high-order Pad\'e methods are hard to beat at round-off precision.
In a first step, we derive new scaling estimates for Pad\'e methods for lower precision requirements following \cite{higham09tsa}.
Let $\theta_m(u)$ be the largest value of $\|A\|$ s.t. the Pad\'e scheme $r_{2m}$ has precision at least $u$, i.e.,
\[
\forall A, \|A\|\leq \theta_m \ :\ r_{2m}(A)=e^{A+E}, \ \text{s.t.}\ \|E\|\leq u.
\]
The new $\theta_m$ are given in Table~\ref{tab.theta}.
\begin{table}\centering\footnotesize
\caption{\label{tab.theta}Theta values for diagonal Pad\'e of order $2m$ with minimum number of products.
The numbers highlighted in boldface correspond to the minimal cost $\pi_{2m}-\log_2(\theta_{2m})$ }
\newcolumntype{H}{@{}>{\lrbox0}l<{\endlrbox}}
\newcolumntype{D}{>{$}r<{$}}
\begin{tabular}{*{8}{D}*{5}{H}D}
\toprule
u\backslash m &1 & 2& 3 & 4& 5& 6& 7& 8& 9& 10& 11& 12& 13\\ \midrule
\leq2^{-53}&
3.65\en8&5.32\en4&1.50\en2&8.54\en2&2.54\en1&5.41\en1&9.50\en1&1.47&2.10&2.81&3.60&4.46&\bf 5.37\\
\leq1\en10 &
3.46\en5&1.64\en2&1.47\en1&4.73\en1&9.98\en1&1.69 &\bf 2.51&3.44&4.44&5.51&6.62&7.76&8.94\\
\leq1\en6&
3.46\en3&1.64\en1&6.80\en1&1.49 &\bf 2.48&3.58 &4.76 &5.98&7.24&8.52&9.81&1.11\ep1&1.24\ep1\\
\bottomrule
\end{tabular}
\end{table}
It is clear that the number of necessary scalings for a sought precision is $s=\lceil{\log_2(\|A\|/\theta_m)\rceil}\in\mathbb{N}_0$ and taking into account the number of multiplications $\pi_m$ needed with each method, a global minimum $s+\pi_m$ can be found at each precision.
We will focus our attention on the medium precision range $u\leq 10^{-6}$, where the 10th order method $r_{10}$ is optimal among the Pad\'e schemes.
In analogy to the error control for Pad\'e methods, we discuss the backward error of the previously obtained splitting methods.
The BCH formula, in the form \eqref{eq.backward}, already gives us a series expansion of the remainder $E$,
\beq\label{eq.error}
E = \sum_{i=p+1}\sum_{j=1} h^i f_{i,j} \mathbf{C}_{i,j}.
\eeq
However, the expansion is difficult to compute for $i>15$ with exponentially growing effort in the symbolic computation.
Further complications arise from the nature of the expansion: it involves commutators $\mathbf{C}_{i,j}$ in $D,B$ which we have to estimate.
For most cases, the roughest (although sharp) estimate
\beq\label{eq.rough}
\|[D,B]\|=\|DB-BD\|\leq 2\eps\|D\|^2, \qquad \eps = \|B\|/\|D\|,
\eeq
is way to loose to give accurate results.
Having in mind matrices with asymmetric spectra, i.e., small positive and large negative eigenvalues,
the following estimate is more useful \cite[Theorem 4]{kittaneh09nif},
\[
\|[D,B]\|\leq \|B\| (d^+-d^-),
\]
where the numerical range of $D$ (or easier: the eigenvalues) lies within $[d^-,d^+]$, which corresponds to a factor 2 gain in the estimate.
In any case, we can refine the estimate by recycling the calculations for the modified splittings, $[D,[D,B]]$, $[D,[D,[D,[D,B]]]]$ and intermediate steps, $[D,B]$, etc.
Then, we estimate the most relevant commutators, recalling the notation $[D^2,B]=[D,[D,B]]$,
\begin{align*}
\|[B,[D,B]]\|&\leq 2\|[D,B]\| \|B\|, \\
\|[B,[D,[D,[D,B]]]]\|&\leq
2 \| [D,B]\|\, \|[D,[D,B]]\| ,
\\
\|[D,[B,[D,[D,B]]]]\|&
\leq 2 \| [D,B]\|\, \|[D,[D,B]]\|
,\\
\|[B,[B,[D,[D,B]]]]\|&\leq 4 \| B\|^2 \ \|[D,[D,B]]\|,\\
\|[D,[D,[D,[D,[D,[D,B]]]]]]\|&\leq (d^+-d^-)^2\|[D,[D,[D,[D,B]]]]\|.
\end{align*}
The splitting methods studied in this work can be classified by their order and the leading error commutators are collected in Table~\ref{tab.split.commutators}.
In principle, one could use the error terms at the next larger power in $h$ to estimate the quality of this truncation, but for practical purposes and $h\ll 1$, numerical experiments show that the simpler bounds are sufficient to get a reasonable recommendation for the number of squarings.
For illustration, we print the expansion \eqref{eq.error} for the method
\eqref{eq:strang62}
\begin{align}\label{eq.err62}
E^{[6,2]}(h)\leq \tilde{E}^{[6,2]}=&\; 3.11\en{6} h^7 \|[D^6,B] \|
+ 8.33\en{2} h^3 \|[B,[D,B]]\| \\ \nonumber
&+ h^5 (1.39\en{3} \|[B,[D^3,B]]\| + 5.56\en{3} \|[[B,D],[D^2,B]]\|)\\ \nonumber
& + h^5 (5.56\en{3} \|[B^2,[D^2,B]]\| + 2.78\en{3} \|[[B,D],[B^2,D]]\|\\ \nonumber
&+\mathcal{O}\left(\eps h^9 + \eps^2 h^7 + \eps^3 h^7\right)
\intertext{and for method $\tilde{Y}_2$ of order (6,4) from Table~\ref{tab.coefs},}
\label{eq.err64}
E^{[6,4]}(h)\leq \tilde{E}^{[6,4]}=&\; 3.49\en{5} h^7 \|[D^6,B] \| \\ \nonumber
& + h^5 (1.70\en{3} \|[B,[D^3,B]]\| + 1.39\en{3} \|[[B,D],[D^2,B]]\|)\\ \nonumber
& + h^5 (1.39\en{3} \|[B^2,[D^2,B]]\| + 4.63\en{4} \|[[B,D],[B^2,D]]\|\\ \nonumber
&+\mathcal{O}\left(\eps h^9 + \eps^2 h^7 + \eps^3 h^7\right) .
\end{align}
Then, the following algorithm suggests itself:
Compute the commutators needed for the modified squarings, estimate their norms and finally evaluate the polynomials $\tilde{E}(h)$ to find an upper bound for $h$ such that the local error remains below given accuracy $u$. This $h$ translates directly to the number of external squarings $s_2=\lceil\log_2(h)\rceil$ and now, it only remains to sum the computational cost originating from the number of dense products and exponentials to find the overall most efficient method for a particular set of matrices $D,B$.
In contrast to the static Pad\'e case, where there is a single best method by just fixing the precision, this procedure is more flexible and chooses - at virtually no extra cost - the best method for the given matrix algebra structure.
Furthermore, we can establish a threshold for the size of the small parameter $\varepsilon$ in order to decide when splittings should be preferred over Pad\'e methods.
For example, let $u=10^{-6} (10^{-4})$ be the desired precision, we then know that $r_{10}$ ($r_{10}$) is optimal and the largest value the norm $\theta =\|A\|$ can take is $\theta_5=2.48 (\theta_5=3.85)$.
Given that $r_{10}$ requires three multiplications, we use the splitting method $\tilde{Y}_0$ with three squarings to yield a method of the same computational cost. In \eqref{eq.err62}, this corresponds to taking $h=2^{-3}$.
Applying the roughest possible estimate \eqref{eq.rough} to $\tilde{E}^{[6,2]}(2^{-3})$, we obtain a polynomial in $\eps$ which takes values below $u$ for $\eps\leq 0.01 (0.05)$.
In practice, the norm estimates are sharper since we can use the commutators that have been computed in the algorithm and we expect an even larger threshold for $\eps$.
\begin{table}[!tbh]\caption{\label{tab.split.commutators}Leading error commutators at given order.
}
\begin{tabular}{l*{3}{>{$}l<{$}}}
\toprule
order & \eps^1 & \eps^2 & \eps^3\\ \midrule
$(2n,2)$ & [D^{2n},B] & [B,[D,B]] & [B,[B,[D,[D,B]]]]\\
$(2n,4)$ & [D^{2n},B] & [B,[D,[D,[D,B]]]] , [D,[B,[D,[D,B]]]] & [B,[B,[D,[D,B]]]]\\
\bottomrule
\end{tabular}
\end{table}
\section{Numerical results}
In a couple of test scenarios, we attempt to provide an idea about when our new methods are superior to standard Pad\'e methods.
In each setting, we define a different matrix $D$ which will be perturbed by a matrix $B$, s.t.
\[
B_{i,j}=k(i-j)/(i+j)
\]
and $k$ is chosen to satisfy $\varepsilon=\|B\|_1/\|D\|_1$ for the parameter set $\varepsilon=10^{-1}, 10^{-2}, 10^{-3}$.
We measure the relative error in the 1-norm, $\| S_p^{[m]} - e^A\|_1/\|e^A\|_1$ for all methods where the exact solution is computed by a high-order Pad\'e method and all splittings use the second-order scheme $r_2$ to approximate the exponential $\exp({2^{-s}B})$.
\subsection{Rotations}
Letting
\[
D=i \diag\{-25,-24.5,\ldots ,24.5,25\}
\]
with $i=\sqrt{-1}$, the performance of Pad\'e methods of order 10 and 26, together with the 16th-order Taylor method using 6 products is studied. Fig.~\ref{fig.case1} shows the relative error (in logarithmic scale) versus cost (number of matrix--matrix multiplications) for different choices of the scaling parameter, $s$. The horizontal line shows the tolerance desired for the numerical experiments.
It is evident that, as expected, the Pad\'e method $r_{10}$ is the most efficient among these standard schemes and will be used for reference in later experiments.
For illustration, Fig.~\ref{fig.case1} also includes two modified squaring methods without commutators ($Y_2$, order (6,2) and $Y_3$, order (6,4) from Table~\ref{tab.coefs}), both of which are more efficient than $r_{10}$ in the lower precision range. Notice that, since $A$ isa complex matrix, to use splitting methods with complex coefficients does not increase the cost of the algorithms in this case.
Furthermore, the standard methods are insensitive w.r.t. the small parameter $\varepsilon$, whereas the splitting methods improve as $\varepsilon$ decreases.
\begin{figure}[!ht]
\centering
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=1,
minor y tick num=4,
minor z tick num = 2,
scale only axis,
font=\footnotesize
}
}
\setlength\figurewidth{.24\textwidth
\setlength\figureheight{.4\textwidth}
\tikzsetnextfilename{figs/case1a}
\includegraphics{case1a}
\caption{\label{fig.case1} Relative error (in logarithmic scale) versus computational cost given by the number of dense matrix-matrix products for the standard Pad\'e and Taylor methods $r_{10}, r_{26}, T_{16}$, and the splitting methods $Y_2$ and $Y_3$ of order (6,2) and (6,4), respectively, without commutators from Table~\ref{tab.coefs}. }
\end{figure}
In a second experiment in Fig.~\ref{fig.case1b}, we use the same matrices as before but choose the most efficient splitting methods with commutators, $\tilde{Y}_0$ and $\tilde{Y}_1$.
Using the local error estimates in \eqref{eq.err62} and \eqref{eq.err64}, we indicate the point which corresponds to the optimal number of squarings for the splitting methods and compare it with the recommended squaring parameter for Pad\'e $r_{10}$.
For a relatively large parameter $\varepsilon$ in the left panel of Fig.~\ref{fig.case1b}, the method $r_{10}$ is still superior but is already equaled in terms of computational cost for a smaller perturbation in the center plot, but at higher accuracy. As $\varepsilon$ becomes smaller in the right panel, we achieve higher accuracy at lower computational cost, saving one product for $\tilde{Y}_1$ and two products for $\tilde{Y}_2$, respectively.
\begin{figure}[!ht]
\centering
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=1,
minor y tick num=4,
minor z tick num = 2,
scale only axis,
font=\footnotesize
}
}
\setlength\figurewidth{.24\textwidth
\setlength\figureheight{.4\textwidth}
\tikzsetnextfilename{figs/case1b}
\includegraphics{case1b}
\caption{\label{fig.case1b} The solid lines show the relative global error $e^{A}$ after squaring versus the overall computational cost and the dashed curves depict the relative local error in $e^{2^{-s}A}$ (before squaring) which is used for the error estimate, both for Pad\'e and the splittings.
The filled markers indicate the position of the recommended (automatic) algorithm.
}
\end{figure}
In the next plot, Fig.~\ref{fig.case1c}, we increase the norm of the matrix and set $D_2=100 D$, and $B$ is scaled accordingly to maintain the quotient $\|B\|_1/\|D_2\|_1=\varepsilon$.
The implications are a substantial increase in the number of necessary squarings with prior scaling and corresponds to a long-time integration in which we observe the favorable behavior expected from Fig.~\ref{fig.longtime}.
The gain with respect to Pad\'e's method is striking as $\varepsilon$ decreases.
\begin{figure}[!ht]
\centering
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=1,
minor y tick num=4,
minor z tick num = 2,
scale only axis,
font=\footnotesize
}
}
\setlength\figurewidth{.24\textwidth
\setlength\figureheight{.4\textwidth}
\tikzsetnextfilename{figs/case1c}
\includegraphics{case1c}
\caption{\label{fig.case1c} Same as Fig.~\ref{fig.case1b} for an exponential of a large norm matrix, with diagonal part $D_2=100D$.
}
\end{figure}
\subsection{Dissipation}
A less favorable problem for our algorithm is given using a stiff matrix with large positive and negative eigenvalues,
\[
D=\diag\{15,14.5,\ldots, -14.5,-15\}.
\]
The perturbation $B$ is scaled as before to $\|B\|/\|D\|=\varepsilon$. Fig.~\ref{fig.case1} shos the results obtained.
Again, our methods perform well for low accuracies for not too large perturbations and improve as $\varepsilon$ becomes smaller.
\begin{figure}[!ht]
\centering
\pgfplotsset{every axis plot/.append style={line width=1.2pt, mark size=2pt},
tick label style={font=\footnotesize},
every axis/.append style=
minor x tick num=1,
minor y tick num=4,
minor z tick num = 2,
scale only axis,
font=\footnotesize
}
}
\setlength\figurewidth{.24\textwidth
\setlength\figureheight{.4\textwidth}
\tikzsetnextfilename{figs/case2}
\includegraphics{case2}
\caption{\label{fig.case2} Same as Fig.~\ref{fig.case1b} but for the stiff matrix case $D=\diag\{15,14.5,\ldots, -14.5,-15\}$. }
\end{figure}
\section{Conclusions}
We have proposed a new recursive algorithm based on splitting methods for the computation of the exponential of perturbed matrices which can be written as the sum $A=D+\varepsilon B$ of a sparse and efficiently exponentiable matrix $D$ with sparse exponential $e^D$ and a dense matrix $\varepsilon B$ which is of small norm in comparison with $D$.
We have considered the scaling and squaring technique but replacing the Pad\'e or Taylor methods to compute the exponential of the scaled matrix by an appropriate splitting methods tailored for this class of matrices. We have proposed a recursive algorithm which allows to save computational cost and still leaves some free parameters for optimization. An important feature of splitting methods for perturbed problems is that the error is a sum of a local error of order ${\cal O}(\varepsilon)$ plus a global error of order ${\cal O}(\varepsilon^2)$ and this allows to build new methods with high performance when low to medium accuracy is desired.
The new schemes are built taking into account that the dominant computational cost arises from the computation of dense matrix products and we present a modified squaring which takes advantage of the smallness of the perturbed matrix $B$ in order to reduce the number of squarings necessary.
The recursive character of the modified squarings implies only light memory requirements.
Theoretical results on local error and error propagation for splitting methods are complemented with numerical experiments and show a clear improvement over existing and highly optimized Pad\'e methods when low to medium precision is sought.
\Appendix
\section{Further approaches}
In this subsection, we collect results on approaches that are successful in the context of splittings for ordinary differential equations, however, have been found less efficient on the numerical experiments than the methods presented before.
\subsection{On processing}\label{sec.processing}
A basic property of the adjoint action,
$$
e^{P}Ye^{-P} = e^{\mathrm{ad}_P}Y = Y + [P, Y] + \frac12 [P, [P, Y]] + \cdots
$$
together with the cheap computability of the commutator $[D,B]=DB-BD$ motivates the use of \defn{processing techniques}, well-known for the numerical integration of differential equations, to eliminate error terms.
The idea is now based on the observation that $(XYX^{-1})^N = XY^NX^{-1}$ and essentially corresponds to a change of basis in which the error propagation (recall that large $s$ can be regarded as a (long-) time integration using a small time-step $h=1/2^s$) is expected to be less severe.
The modified Strang algorithm \eqref{eq:strang62} has leading error proportional to
\[
[B,[D,B]], \quad [B,[D,[D,[D,B]]]], \quad [D,[D,[B,[D,B]]]].
\]
The first two of which can be eliminated using a processor with $P=\alpha [D,B] + \beta [D,[D,[D,B]]]$, thus motivating the ansatz
\[
e^{\alpha \eps h^2 [D,B] + \beta \eps h^4 [D,[D,[D,B]]]} \tilde{Y}_{s} e^{-\alpha \eps h^2 [D,B] - \beta \eps h^4 [D,[D,[D,B]]]}.
\]
The norm of the outer exponents is small and a low order Pad\'e approximation, say $r_2(P)$, usually provides sufficient accuracy.
Therefore, at the expense of one exponential, one multiplication and one inversion (which is performed together with the multiplication, as for the Pad\'e methods, $(\cB\cD)\tilde{\cB}^{-1}$), we get two free parameters, $\alpha,\beta$.
Using the kernel $\tilde{Y}_0$, we reach order (6,4), whereas $\tilde{Y}_1$ is sufficient for order (10,4) and (6,6,4), see Table~\ref{tab.coefs2}.
\subsection{More exponentials}
For problems where complex coefficients $a_j$ lead to a substantial increase in computational complexity (e.g., when $A,B\in\R^{n\times n}$) or matrix commutators are not desirable, it could be advantageous to allow negative values for some $b_j$.
A first example is the four-stage method
\beq\label{eq.neg4}
S^{[4]}_4 = \cD_{ha_1} \cB_{hb_1} \cD_{ha_2} \cB_{hb_2} \cD_{ha_2} \cB_{hb_1} \cD_{ha_1}.
\eeq
This scheme requires two exponentials, two products and has two free parameters which can produce a fourth-order method with real coefficients $a_j,b_j$, known as triple jump
\cite{creutz89hhm,suzuki90fdo,yoshida90coh},
see Table~\ref{tab.coefs2}.
Another product is necessary to compute the six-stage composition
\
S^{[6]}_{(6,4)}= \cD_{ha_1} (\cB_{hb_1} \cD_{ha_2} \cB_{hb_1})\cD_{ha_3}\cB_{hb_2} \cD_{ha_3} (\cB_{hb_1} \cD_{ha_2}\cB_{hb_1})\cD_{ha_1}.
\
Three free parameters are sufficient to construct (6,4) methods, however, with complex time-steps.
The real-valued fourth-order method minimizing the error at $\mathcal{O}(\varepsilon h^5)$ can be found in Table~\ref{tab.coefs2}.
An additional stage with a grouping similar to the modified splittings,
\
S^{[7]}_{(6,4)}= \cD_{ha_1} (\cB_{hb_1} \cD_{ha_2} \cB_{hb_2}\cD_{ha_2}\cB_{hb_1}) \cD_{ha_3}
(\cB_{hb_1} \cD_{ha_2} \cB_{hb_2}\cD_{ha_2}\cB_{hb_1})
\cD_{ha_1},
\]
requires the same number of products but has real solutions of order (6,4). Among the four real-valued solutions, the one minimizing the error at $\mathcal{O}(\varepsilon h^7)$ is printed in Table~\ref{tab.coefs2}.
We have found that supposedly clever re-utilization of exponentials by setting $b_j$ to be a rational multiple of an already computed exponent $b_k$ are not competitive since - at its very best - one can save the computation of an exponential at the cost of an inversion ($b_j=-b_k$) or a matrix product ($b_j=2b_k$), however, the direct use of the sufficiently accurate $r_2$ Pad\'e method needs only one inversion.
\subsection{Splitting for low-order Pad\'e}
Technically, the stated splitting orders assume the exact computation of all exponentials, but in practice, the cheap underlying Pad\'e scheme $r_2$ has accuracy limit $\cO(\eps^3h^3)$.
Since we assumed $\eps$ to be a small parameter, comparable to $h^2$, it could be regarded as $\cO(\eps h^7)$.
Instead of switching to the more precise $r_4$ method ($\cO(\eps^5 h^5)$) for the exponential $\cB$,
(using $r_2$ for the processor has error $\cO(h^6 \eps^3)$ and is therefore sufficient),
we attempt to use a free parameter to decrease the $r_2$-related error in $\cB$ to $h^5\varepsilon^5$.
The procedure is based on the observation that the approximant $r_2(h\eps B)$ can be expressed as a single exponential
$$
r_2(h\eps B) = e^{h\eps B + h^3 \eps^3 C + \cO(h^5 \eps^5)}
$$
for some matrix $C$.
Notice that the exponent can be expanded in odd powers of $h$ since $r_2$ is symmetric.
Now, we simply add the (unknown) matrix $C$ to the algebra and in addition to the previous order conditions, we have to solve $\sum_{i=1}^m b_i^3 = 0$.
It is clear that condition $b_i=1/m$ has to be dropped and at least three exponentials $\cB_{b_jh}$ are necessary.
We embark by modifying \eqref{eq.neg4} to
\beq\label{eq.neg4mod}
\Psi^{[4,mod]} = \cD_{ha_1} \tilde{\cB}_1 \cD_{ha_2} \tilde{\cB}_2 \cD_{ha_2} \tilde{\cB}_1 \cD_{ha_1}.
\eeq
Using two exponentials (inversions) and two multiplications, we have six free parameters and only one additional equation.
The freedom in the parameters allows to construct real-coefficient methods of order (10,4)
and alternatively, at order (8,4), a method minimizing the squared error polynomials $e_{5,2}$ at $\eps^2 h^5$, see Table~\ref{tab.coefs2}.
\begin{table}[!ht]
\caption{\label{tab.coefs2}Further splitting methods, including several exponentials and processing techniques.}
\noindent{\footnotesize
\begin{tabular}{ll}
\toprule
$S^{[4]}$, 4 stages, order 4 & 2 exp, 2 prod\\[1mm
$a_1= \frac16 (2 + 1/2^{1/3} + 2^{1/3})$, $b_1= \frac13 (2 + 1/2^{1/3} + 2^{1/3})$ & 2 complex sol. omitted
\\[1mm]
$S^{[6]}$, 6 stages, order 4 & 2 exp, 3 prod\\[1mm]
$a_1=0.19731107566242791631$,& [minimizes $\mathcal{O}(\varepsilon h^5)$]\\
$a_2=0.38252646594731312955$, \\
$a_3=(1-2a_1-2a_2)=-0.079837541609741045862$,\\
$b_1=0.42519341909910345071$,\\ $b_2=1-4b_1=-0.70077367639641380284$.
\\[1mm]
$S^{[7]}$, 7 stages, order (6,4) & 2 exp, 3 prod\\[1mm]
$a_1=0.35937529621978708941$, &{ [minimizes $\mathcal{O}(\varepsilon h^7)$]}\\
$a_2=-0.098379231055234835826$, \\
$a_3=(1-2a_1-4a_2)=0.67476633178136516448$,\\
$b_1=0.67702963544760500586$,\\ $b_2=1/2-2b_1=-0.85405927089521001173$.\\
\midrule\midrule
Processed $e^{xh^2[D,B]+yh^4[D,[D,[D,B]]]}\tilde{Y}_0e^{-xh^2[D,B]-yh^[D,[D,[D,B]]]}$ & 2 exp, 1 prod, 1 inv\\[1mm]
Order (6,4)\\[1mm]
$a_1=1/2$,
$\beta=-1/24$, $\gamma=31/5760$
$x=-1/12$, $y=1/120$.\\
\midrule
Processed $e^{xh^2[D,B]+yh^4[D,[D,[D,B]]]}\tilde{Y}_1e^{-xh^2[D,B]-yh^4[D,[D,[D,B]]]}$ & 2 exp, 1 prod, 1 inv\\[1mm]
Order (6,6,4)\\[1mm]
$a_2=0.2587977340833403434530275$, \\
$\beta=-0.005227683364583625421653925$, \\
$\gamma=0.0000329546718228203782$,\\
$x=-0.02303276685416841919659022$,\\
$y=0.0007499977372301362425777840$.\\[1mm]
Order (10,4)\\[1mm]
$a_2=0.250225501288894385213924$,\\
$\beta=-0.0052083460460411565905784$,\\
$\gamma=0.0000329546718228203782$,\\
$x=-0.0208897086555569296368143$, \\
$y=0.0000573371861339342917744$\\
\midrule\midrule
$\cD_{ha_1} \tilde{\cB}_1 \cD_{ha_2} \tilde{\cB}_2 \cD_{ha_2} \tilde{\cB}_1 \cD_{ha_1}$ & 2 exp, 2 prod\\[1mm]
real (10,4) based on $r_2$, & \\
$d_2 = -0.0017987433839305087766, c_2 = -0.14389703981903926044$,\\
$d_1 = 0.000039345117326816272608, c_1 = -0.0079989398412468330564$,\\
$b_2 = -0.58268652153120735848, a_2 = 0.50468619989723192191$\\[1mm]
(8,4) minimizing $e_{5,2}$,\\
$d_2 = 0.009460956758445480826, c_2 = -0.03780196888453765108$, \\
$d_1 = 0.0011653151315644152329,c_1 = -0.061046475308497637733$,\\
$b_2 = -0.58268652153120735848, a_2 = 0.50468619989723192191$\\
\bottomrule
\end{tabular}
}
\end{table}
|
\section{Introduction}
The question of why some materials form a disordered glass rather than a crystalline solid when they are cooled or compressed is one of the most pressing questions in both physics and materials science. A glass, like a crystalline solid, has a yield stress, i.e., it responds like an elastic solid when subjected to stress below the yield stress. However, on examining the microscopic structure of a glass (quantified via a suitable two-point correlation function or structure function, such as the static structure factor $S(k)$, that can be measured in a scattering experiment \cite{HM}), one finds no real difference between the structure of the glass and the same material at a slightly higher temperature, when it is a liquid. In order to discern the difference between a glass and a liquid from examining the microscopic structure, one approach is to determine the {\em dynamic} structure function. In a liquid, the particles are able to rearrange themselves over time, so that their subsequent positions become decorrelated from their earlier locations. On the other hand, in a glass, the particle positions remain strongly correlated to their locations at an earlier time. The standard picture of this phenomenon is that the particles become trapped within a `cage' of neighboring particles so that in the glass the probability of a particle escaping is negligibly small \cite{HM}. Thus, in a glass the particles can be thought of as frozen in a disordered arrangement, instead of forming a periodic or crystalline lattice.
Much insight into the formation and the statistical and thermodynamic properties of glasses has been gained in recent years from the study of colloidal suspensions, because of our ability to observe and track individual colloids in suspension with a confocal microscope \cite{WeeksetalScience2000}. In this paper we investigate the structure and phase behavior of a simple two-dimensional (2D) model colloidal fluid composed of ultra-soft particles that are able to interpenetrate. We first study the solidification of the one-component system, which generally forms a regular crystalline solid. We then investigate binary mixtures which form disordered solids much more readily, and relate the disorder we find to the solidification process when the system is quenched from the liquid state. In particular, we examine how solidification fronts propagate into the unstable liquid, and how this dynamical process can lead to disorder in the model \cite{ARTK12}. Our study of this system is based on density functional theory (DFT) \cite{Evans79, Evans92, HM, lutsko10} and dynamical density functional theory (DDFT) \cite{MaTa99, MaTa00, ArEv04, ArRa04}.
DFT is an obvious theoretical tool for studying the microscopic structure and phase behavior of confined fluids, because it provides a method for calculating the one-body (number) density $\rho(\rr)$ of a system confined in an external potential $\Phi(\rr)$. The density profile $\rho(\rr)$ gives the probability of finding a particle at position $\rr$ in the system and is obtained by minimizing the grand potential functional $\Omega[\rho]$ with respect to variations in $\rho(\rr)$ \cite{Evans92,HM}. Typically, this is done numerically, and one must discretise the density distribution $\rho(\rr) \to \rho_p$, recording it on a set of grid points (the index $p$ enumerates the grid points). One then numerically solves the discretised equations for the set $\{\rho_p\}$. An alternative approach is to assume the density profile $\rho(\rr)$ takes a specific functional form, parametrized by a set of parameters $\{\alpha_p\}$, and then seek the values of the parameters $\{\alpha_p\}$ minimizing the grand potential functional. This alternative technique is often used in studies of crystallization, where the density profile is (for example) assumed to be a set of Gaussian functions, centered on a set of lattice sites \cite{HM}.
Over the years, DFT has been used by several groups studying the properties of glassy systems. Wolynes and coworkers \cite{StWo84, SSW85, XiWo01} developed a successful model of hard-sphere glass formation based on the idea that the glass may be viewed as a system that is `frozen' onto a (random close-packed) nonperiodic lattice. This approach is based on the DFT theory for crystallization \cite{HM} and was followed up by a number of other investigations \cite{BaCo86, Lowen90, KaDa01, KaDa02, KiMu03, YYO07, CKDKS05, YOY08} extending and applying the method. All of these studies show that the free energy landscape may exhibit minima corresponding to the particles becoming localized (trapped) on a nonperiodic lattice. One limitation of these approaches is that the system is constrained by the choice of the nonperiodic lattice (or, in the case of the approach in Ref.\ \cite{YOY08}, by the fixed boundary particles). However, in the present work, rather than imposing a particular (nonperiodic) lattice structure on the system, we use DDFT to describe the solidification after the uniform liquid is deeply quenched to obtain the structure of the crystal or disordered solid that is formed as an {\em output}.
Here, we consider the case when the uniform fluid is quenched to a state point where the crystal is the equilibrium phase and we examine how the solid phase advances into the liquid phase, with dynamics described by DDFT. Our work here builds on earlier studies \cite{TBVL09, GaEl11, ARTK12} employing the phase field crystal (PFC) model \cite{ELWRGTG12} to explore a similar situation. The PFC free energy functional consists of a local gradient expansion approximation \cite{ARTK12, ELWRGTG12} and is arguably the simplest DFT that is able to describe both the liquid and crystal phases and the interface between them. In Refs.\ \cite{GaEl11, ARTK12} it was shown that the solidification front speed can be calculated by performing a marginal stability analysis, based on a dispersion relation obtained by linearising the DDFT (see Sec.\ \ref{sec:VA} below). The most striking result of the work in \cite{GaEl11, ARTK12} is the observation that the wavelength of the density modulations created behind such an advancing solidification front is not, in general, the same as that of the equilibrium crystal. Thus, for the system to reach the equilibrium crystal structure after such a solidification front passes through the system, significant rearrangements must occur and defects and disorder often remain. This conclusion, based on a marginal stability calculation in one dimension (1D), was confirmed in 2D PFC numerical simulations \cite{ARTK12}. In the present work we consider the same type of situation using a more sophisticated nonlocal DFT for fluids of soft penetrable particles. For this model fluid, we find that when the fluid is deeply quenched, the marginal stability analysis correctly predicts the solidification front speed, giving the same front speed as we obtain from direct numerical simulations. However, for shallow quenches we find that the front propagates via a nonlinear mechanism rather than the linear mechanism that underpins the marginal stability analysis, and that the 1D marginal stability analysis fails to predict the correct front speed. The overall picture that we observe is similar to that predicted for 2D systems on the basis of amplitude equations by Hari and Nepomnyashchy \cite{HN00}, as discussed further in the Appendix.
We also present results for a binary mixture of soft particles that exhibits several different competing crystal structures, including several hexagonal phases and a square phase. We find that when a solidification front advances through such a mixture a highly disordered state results, consisting of a patchwork of differently ordered regions, some that are square and others that are hexagonally coordinated. Thus, the solidification process generates disordered structures in a completely natural way.
This paper is structured as follows. In Sec.\ \ref{sec:model} we describe the model soft core fluids considered in this paper and briefly describe the Helmholtz free energy functional that we use as the basis of our DFT and DDFT calculations for the density profile(s) of the liquid and solid phases. In Sec.\ \ref{sec:fluid_structure} we examine the structure of the uniform fluid. We obtain and compare results for the radial distribution function $g(r)$, comparing results from a simple DFT that generates the Random Phase Approximation (RPA) closure to the Ornstein-Zernike equation with results from the Hyper-Netted Chain (HNC) closure approximation which is very accurate for soft systems, and find very good agreement between the two, thus validating the simple DFT that we use. In Sec.\ \ref{sec:phase_behavior} we present results for the equilibrium phase behavior of the one-component fluid, calculating the phase diagram. Then in Sec.\ \ref{sec:dynamics} we briefly describe the DDFT for the non-equilibrium fluid and calculate the dispersion relation for fluid mixtures. In Sec.\ \ref{sec:1comp_fronts} we briefly discuss the marginal stability analysis for determining front speeds and compare the results with those from 2D DDFT computations, and show that the solidifications fronts do not generate density modulations with the same wavelength as the equilibrium crystal. This leads to disorder, and we present results showing how the one-component system is able to rearrange over time to produce a well-ordered crystal, with only a few defects. In Sec.\ \ref{sec:bin_system} we present our results for a binary mixture of soft particles in which a solidification front can generate a solid with persistent disorder. Section \ref{sec:conc} contains concluding remarks. The Appendix describes an amplitude equation approach that helps explain the relation between the linear and nonlinear solidification fronts that we observe.
\section{Model fluid}
\label{sec:model}
In this paper we study 2D soft penetrable particles and their mixtures. We model the particles as interacting via the pair potential
\begin{equation}
v_{ij}(r)=\epsilon_{ij} e^{-(r/R_{ij})^n},
\label{eq:pair_pot}
\end{equation}
where the index $i,j=1,2$ labels particles of the two different species. The parameter $\epsilon_{ij}$ defines the energy for complete overlap of a pair of particles of species $i$ and $j$ and $R_{ij}$ defines the range of the interaction. We also consider a one-component fluid, and in this case omit the indices -- i.e., we write the interaction between the particles as $v(r)=\epsilon e^{-(r/R)^n}$. The case $n=2$ corresponds to the Gaussian core model (GCM) \cite{Stillinger76, BLHM01, LLWL00, Likos01} and larger values of $n$ define the so-called generalised exponential model of index $n$ (GEM-$n$). In this paper we focus on the cases $n=4$ and $n=8$. Penetrable spheres correspond to the limit $n\to\infty$. Such soft potentials provide a simple model for the effective interactions between polymers, star-polymers, dendrimers and other such soft macromolecules in solution \cite{Likos01, DaHa94, likos:prl:98, LBHM00, BLHM01, JDLFL01, Dzubiella_2001, LBFKMH02, likos:harreis:02, GHL04, MFKN05, Likos06, LBLM12}. For such particles one may approximate the intrinsic Helmholtz free energy of the system as \cite{Likos01}
\begin{eqnarray}\notag
\mathcal{F}[\{\rho_i(\rr)\}]=k_BT \sum_{i=1}^2\int \dr \rho_i(\rr)\left(\log[\rho_i(\rr)\Lambda_i^2]-1\right)\\
+ \frac{1}{2}\sum_{i,j}\int\dr\int\dr'\rho_i(\rr)v_{ij}(|\rr-\rr'|)\rho_j(\rr'),
\label{eq:DFT}
\end{eqnarray}
where $T$ is temperature, $k_B$ is Boltzmann's constant and $\Lambda_i$ is the (irrelevant) thermal de Broglie wavelength for species $i$. Henceforth we set $\Lambda_i=R_{11}=1$. The free energy is a functional of the one-body density profiles $\rho_1(\rr)$ and $\rho_2(\rr)$, where $\rr=(x,y)$. The first term in Eq.~\eqref{eq:DFT}, $\mathcal{F}_{id}$, is the ideal gas (entropic) contribution to the free energy while the second term $\mathcal{F}_{ex}$ is the contribution from the interactions between particles. The equilibrium density distribution is that which minimizes the grand potential functional
\begin{equation}
\Omega[\{\rho_i(\rr)\}]=\mathcal{F}[\{\rho_i(\rr)\}]+\sum_{i=1}^2\int\dr\rho_i(\rr)(\Phi_i(\rr)-\mu_i),
\label{eq:GPF}
\end{equation}
where $\mu_i$ are the chemical potentials and $\Phi_i(\rr)$ is the external potential experienced by particles of species $i$. When evaluated using the equilibrium density profiles, the grand potential functional gives the thermodynamic grand potential of the system. For a system in the bulk fluid state (i.e., where $\Phi_i(\rr)=0$), the minimizing densities are independent of position, $\rho_i(\rr)=\rho_i^b$. However, at other state points, for example, when the system freezes to form a solid, $\Omega$ is minimised by nonuniform density distributions, exhibiting sharp peaks.
The free energy functional in Eq.~\eqref{eq:DFT} generates the random phase approximation (RPA) for the pair direct correlation functions,
\begin{equation}
c^{(2)}_{ij}(\rr,\rr')\equiv-\beta\frac{\delta^2 F_{ex}}{\delta \rho_i(\rr) \delta \rho_j(\rr')}=-\beta v_{ij}(|\rr-\rr'|),
\label{eq:RPA}
\end{equation}
where $\beta\equiv 1/k_BT$. For three-dimensional (3D) systems of soft-particles such as those considered here, the simple approximation for the free energy in \eqref{eq:DFT} is known to provide a good approximation for the fluid structure and thermodynamics, as long as $\beta \epsilon$ is not too large and the density is sufficiently high, i.e., when the average density in the system $\rho R^2>1$ and the particles experience multiple overlaps with their neighbors -- the classic mean-field scenario \cite{Likos01}. Below, we confirm that this approximation is also good in 2D, by comparing results for the fluid structure with results from the more accurate HNC approximation. This simple DFT has been used extensively with great success to study the phase behavior and structure of soft particles and their mixtures \cite{ArEv01, ArEv02, AER02, ALE02, ALE04, GAL06, MGKNL06, MGKNL07, MoLi07, LMGK07, MCLFK08, LMMGK08, OvLi09b, OvLi09, TMAL09, CML10, MaLi11, NKL12, CPPR12, Pini14}. However, the DFT in \eqref{eq:DFT} is unable to describe the solid phases of the GCM, i.e., GEM-2 -- in order to calculate the free energy and structure of the solid phases of the GCM, one must introduce additional correlation contributions to the free energy \cite{Archer05c}. In contrast, when $n>2$, the approximation in Eq.~\eqref{eq:DFT} is able to provide a good account of the free energy and structure of the solid phase in 2D whenever $\beta \epsilon\sim O(1)$ or smaller. Away from this regime, other approaches are needed \cite{MGKNL06, MCLFK08, ZCM10, ZC12, WS13, WS14}.
\section{Structure of the fluid}
\label{sec:fluid_structure}
The pair correlations in a fluid may be characterised by the radial distribution functions $g_{ij}(r)=1+h_{ij}(r)$, where $h_{ij}(r)$ are the fluid total correlation functions \cite{HM}. These are related to the fluid direct pair correlation functions $c_{ij}^{(2)}(r)$ via the Ornstein-Zernike equation, which for a binary fluid is
\begin{equation}
h_{ij}(r)=c_{ij}^{(2)}(r)+\sum^2_{k=1} \rho_{k} \int d \rr' c^{(2)}_{ik}(|\rr-\rr'|)h_{jk}(\rr').
\label{eq:OZ}
\end{equation}
This equation, together with the exact closure relation
\begin{equation}
c_{ij}^{(2)}(r)=-\beta v_{ij}(r)+h_{ij}(r)-\ln(h_{ij}(r)+1)+b_{ij}(r),
\label{eq:CR}
\end{equation}
may be solved for $h_{ij}(r)$ and hence $g_{ij}(r)$. However, the bridge functions $b_{ij}(r)$ in Eq.\ \eqref{eq:CR} are not known exactly and so approximations are required. For different interactions between particles, various approximations for $b_{ij}(r)$ have been developed \cite{HM}. For fluids of soft particles, the HNC approximation, which consists of setting $b_{ij}(r)=0$, has been shown to be very accurate \cite{Likos01}. Below, we compare the results for $g(r)$ for the one-component fluid, obtained from the HNC closure with those obtained from the simple approximate DFT in Eq.\ \eqref{eq:DFT}. These are obtained via the so-called ``test particle method'' which consists of fixing one of the particles in the fluid and then calculating the density profiles $\rho_{i}(r)$ in the presence of this fixed particle. One then uses the Percus result $g_{ij}(r)=\rho_i(r)/\rho^b$, where the fixed particle is of species $j$. The equilibrium fluid density profiles are those which minimise the grand free energy, i.e., they satisfy the Euler-Lagrange equations
\begin{equation}
\frac{\delta \Omega}{\delta \rho_i(\rr)}=0.
\label{eq:EL}
\end{equation}
From Eqs.\ \eqref{eq:DFT} and \eqref{eq:GPF} we obtain
\begin{equation}
k_BT\ln\rho_i+\sum_j \int d \rr' \rho_j(\rr')v_{ij}(|\rr-\rr'|)+\Phi_{i}(\rr)-\mu_i=0.
\label{eq:EL2}
\end{equation}
In the test particle situation, we set the external potentials equal to those corresponding to fixing one of the particles, i.e., $\Phi_{i}(\rr)=v_{ik}(r)$, for a fixed particle of species $k$. Using the conditions that as $r\rightarrow \infty$, $\Phi_{i}(\rr)\rightarrow 0$ and $\rho_i(\rr)\rightarrow \rho_i^b$, we can eliminate the chemical potentials $\mu_i$ from Eq.~\eqref{eq:EL2} and obtain
\begin{eqnarray}
k_BT\ln \left(\frac{\rho_i(r)}{\rho_i^b}\right)&+&\sum_j \int d \rr' (\rho_j(\rr')-\rho_j^b)v_{ij}(|\rr-\rr'|)\nonumber\\
&+&v_{ik}(r)=0.
\label{eq:NoPot}
\end{eqnarray}
We solve these equations using standard Picard iteration to obtain the density profiles $\rho_{ik}(r)$, where the index $k$ denotes the species held fixed. It is worth noting that if we replace the density profiles in Eq.~\eqref{eq:NoPot} by the total correlation functions, i.e., using $\rho_{ik}(r)=\rho_i^bg_{ik}(r)$, where $g_{ik}(r)=1+h_{ik}(r)$, we can rewrite Eq.~\eqref{eq:NoPot} in the form
\begin{equation}
h_{ik}(r)=c_{ik,HNC}^{(2)}(r)+\sum_j \rho_j^b \int d \rr' h_{ij}(\rr')c_{ij,RPA}^{(2)}(\rr-\rr')
\label{eq:TP}
\end{equation}
(cf.~Eq.~\eqref{eq:OZ}), where $c_{ij,HNC}^{(2)}(r)$ denotes the HNC closure approximation for the pair direct correlation function (i.e., setting $b_{ij}(r)\equiv 0$ in Eq.~\eqref{eq:CR}) and $c_{ij,RPA}^{(2)}(r)=-\beta v_{ij}(r)$ denotes the RPA approximation. In Fig.~\ref{RPAvHNC} we compare results from the HNC closure of the OZ equation and the RPA test particle results for a one-component fluid with chemical potential $\mu=0$ and various values of $\beta \epsilon$. We see that the agreement between the two is very good, even at low temperatures such as $\beta\epsilon=10$, where one might expect the RPA to fail.
\begin{figure}[h!]
\includegraphics[width=1.\columnwidth]{HNCvRPA_EPS1_5_10.pdf}
\caption{(Color online) The radial distribution function $g(r)$ for a GEM-4 fluid with bulk chemical potential $\mu=0$ obtained from the HNC closure to the OZ equation (dashed lines) and from the RPA DFT via the test particle method (solid lines), for several values of $\beta\epsilon$. For clarity, the results for $\beta\epsilon=1$ and 5 have been shifted vertically. The results correspond to the state points $(\beta\epsilon,\rho^bR^2)=(1,0.36)$, $(5,0.14)$ and $(10,0.088)$. As $\beta \epsilon$ increases, the RPA approximation becomes increasingly poor; nevertheless, even for (fairly low density) state points such as $\beta\epsilon=10$ the agreement is surprisingly good -- recall that the RPA approximation improves as the density is increased.}
\label{RPAvHNC}
\end{figure}
\section{Equilibrium fluid phase behavior}
\label{sec:phase_behavior}
\begin{figure}[t]
\includegraphics[width=1.\columnwidth]{phase_diag.pdf}
\caption{(Color online) Phase diagrams of the one-component 2D GEM-4 and GEM-8 model fluids. The solid lines are the binodals, i.e., loci of coexisting liquid and solid phases. The dashed lines are the spinodal-like instability lines along which the metastable liquid phase becomes linearly unstable.}
\label{fig:phase_diag}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.9\columnwidth]{free_interface_N512_a.png}
\caption{(Color online) Equilibrium density profile at the free interface between coexisting liquid and solid phases in the GEM-4 model when $\beta \epsilon=1$ and $\beta\mu=17.0$.}
\label{fig:free_interface}
\end{figure}
Having established that the simple RPA approximation for the free energy \eqref{eq:DFT} gives a good description of the structure of the bulk fluid, we now apply this to determine the phase diagram of the one-component GEM-4 and GEM-8 models -- in particular, to determine where the fluid freezes to form a crystal. We calculate the density profile of the uniform solid by solving the Euler-Lagrange equation \eqref{eq:EL} using a simple iterative algorithm on a 2D discretised grid with periodic boundary conditions. The uniform density system is linearly unstable at higher densities (this notion is discussed further below) and so for these state points it is easy to calculate the density of the crystal phase. An initial condition consisting of a line along which the density is higher than elsewhere, plus an additional small random number to break the symmetry of the profile, is sufficient. The density profile of the crystal obtained at higher densities is then continued down to lower densities where the liquid and crystal phases coexist.
Two phases coexist when the temperature, pressure and chemical potential of the two phases are equal. The densities of the coexisting liquid and crystal states in the 2D GEM-4 and GEM-8 models are displayed as a function of temperature in Fig.\ \ref{fig:phase_diag}. Qualitatively, the phase diagram is very similar to that found previously for the system in three dimensions (3D) \cite{MGKNL06, MGKNL07, LMGK07}. However, in the 2D case there is only one solid phase, unlike in 3D, where the system can form both fcc and bcc crystals, depending on the state point. The GEM-4 particles freeze at a higher density than the GEM-8 particles, because the GEM-4 potential is softer.
In Fig.\ \ref{fig:free_interface} we display a plot of the equilibrium density profile for the interface between the [1,1] crystal surface and the liquid. This density profile is for the GEM-4 model at temperature $\beta \epsilon=1$. At this temperature the chemical potential at coexistence is $\beta\mu=17$ and the densities of the coexisting liquid and solid phases are $\rho_{\rm l}R^2=5.48$ and $\rho_{\rm s}R^2=5.73$, respectively. Similar stationary fronts at present even when the two competing phases are not exactly at thermodynamic coexistence -- a consequence of pinning of the front to the hexagonal structure to its left.
\section{Theory for the non-equilibrium system}
\label{sec:dynamics}
To extend the theory to non-equilibrium conditions, we assume the particles obey Brownian dynamics, modelled via overdamped stochastic equations of motion:
\begin{equation}
\dot{\rr}_l= -\Gamma_l\nabla_l U(\{\rr_l\},t) + \Gamma_l{\bf X}_l(t).
\label{eq:EOM}
\end{equation}
Here the index $l=1,..,N$ labels the particles, with $N\equiv N_1+N_2$ the total number of particles in the system and $N_i$ the number of particles of species $i$. The potential energy of the system is denoted by $U(\{\rr_l\},t)$, $\nabla_l\equiv\partial/\partial\rr_l$, ${\bf X}_l(t)$ is a white noise term and the friction constant $\Gamma_l^{-1}$ takes one of two values, $\Gamma_1^{-1}$ or $\Gamma_2^{-1}$, depending on the particle species. The quantities $\Gamma_i^{-1}$ characterise the drag of the solvent on particles of species $i$. The dynamics of a fluid of Brownian particles can be investigated using DDFT \cite{MaTa99,MaTa00,ArEv04,ArRa04}, which builds upon equilibrium DFT and takes as input the equilibrium fluid free energy functional. The two-component generalization of DDFT takes the form \cite{Archer05, Archer05b}
\begin{equation}
\frac{\partial\rho_i(\rr,t)}{\partial t} =
\Gamma_i \nabla\cdot\left[\rho_i(\rr,t)\nabla\frac{\delta\Omega[\{\rho_i(\rr,t)\}]}{\delta\rho_i(\rr,t)}\right],
\label{eq:DDFT}
\end{equation}
where $\rho_i(\rr,t)$ are now the time-dependent non-equilibrium fluid one-body density profiles. To derive the DDFT we use the approximation that the non-equilibrium fluid two-body correlations are the same as those in the equilibrium fluid with the same one-body density distributions \cite{MaTa99, MaTa00, ArEv04, ArRa04}.
\subsection{Fluid structure and linear stability}
\label{sec:VA}
We first consider the stability properties of a uniform fluid with densities $\rho_1^b$ and $\rho_2^b$, following the presentation in Refs.~\cite{ArEv04,ARTK12} (see also \cite{Evans79,Evans:TDGammaMolecP1979}). We set the external potentials $\Phi_i(\rr,t)=0$ and consider small density fluctuations $\tilde{\rho}_i(\rr,t)=\rho_i(\rr,t)-\rho_i^b$ about the bulk values. From Eq.~\eqref{eq:DDFT} we obtain
\begin{eqnarray}\notag
\frac{\beta}{\Gamma_i} \frac{\partial \tilde{\rho}_i(\rr,t)}{\partial t}
\, =\, \nabla^2 \tilde{\rho}_i(\rr,t)
\,-\, \rho_i^b \nabla^2 c_i^{(1)}(\rr,t) \\
-\, \nabla . [ \, \tilde{\rho}_i(\rr,t) \nabla c_i^{(1)}(\rr,t) \, ],
\label{eq:DDFT_2}
\end{eqnarray}
where
\begin{equation}
c_i^{(1)}(\rr)\equiv-\beta \frac{\delta (\mathcal{F}-\mathcal{F}_{id})}{\delta \rho_i(\rr)}
\end{equation}
are the one-body direct correlation functions \cite{Evans79,Evans92}. Taylor-expanding $c_i^{(1)}$ about the bulk values gives
\begin{equation}
c_i^{(1)}(\rr)=c_i^{(1)}(\infty)
+\sum_{j=1}^2\int \dr' \frac{\delta c_i^{(1)}(\rr)}{\delta
\rho_j(\rr')} \Bigg \vert_{\{\rho_i^b\}} \tilde{\rho}_j(\rr',t)+O (\tilde{\rho}^2),
\label{eq:c_1_expansion}
\end{equation}
where $c_i^{(1)}(\infty) \equiv c_i^{(1)}[\{\rho_i^b\}]=-\beta \mu_{i,ex}$ and $\mu_{i,ex}$ is the bulk excess chemical potential of species $i$. Since $\frac{\delta c_i^{(1)}(\rr)}{\delta \rho_j(\rr')}= c_{ij}^{(2)}(\rr,\rr')$, Eq.\ \eqref{eq:DDFT_2} yields, to linear order in $\tilde{\rho}_i$,
\begin{eqnarray}\notag
\frac{\beta}{\Gamma_i} \frac{\partial \tilde{\rho}_i(\rr,t)}{\partial t}
\, =\, \nabla^2 \tilde{\rho}_i(\rr,t)\hspace{2.5cm} \\
-\sum_j \rho_i^b \nabla^2 [ \, \int \, \dr' c_{ij}^{(2)}(|\rr-\rr'|
\tilde{\rho}_j(\rr',t) \, ].
\label{eq:DDFT_linear}
\end{eqnarray}
A spatial Fourier transform of this equation yields an equation for the time evolution of the Fourier transform $\hat{\rho}_j(\kk,t)=\int \dr \exp({\rm i} \kk.\rr) \tilde{\rho}_j(\rr,t)$, where ${\rm i}=\sqrt{-1}$. We obtain
\begin{equation}
\frac{\beta}{\Gamma_i} \frac{\partial \hat{\rho}_i(k,t)}{\partial t}
=-k^2 \hat{\rho}_i(k,t)
+\rho_i^b\sum_j k^2 \, \hat{c}_{ij}(k) \hat{\rho}_j(k,t),
\label{eq:DDFT_linear_FT}
\end{equation}
where $\hat{c}_{ij}(k)\equiv\int \dr \exp({\rm i} \kk.\rr) c^{(2)}_{ij}(r)$ is the Fourier transform of the pair direct correlation function. If we assume that the time dependence of the Fourier modes follows $\hat{\rho}_i(k,t)\propto\exp[\omega(k)t]$ we obtain \cite{PBMT05, RAT11,RATK12,LK13}
\begin{equation}
\mathbf{1}\omega(k) \hat{\rho} = \mathbf{M} \cdot \mathbf{E} \hat{\rho} ,
\label{eqMatrixForm}
\end{equation}
where $\hat{\rho}\equiv (\hat{\rho}_1,\hat{\rho_2})$ and the matrices $\mathbf{M}$ and $\mathbf{E}$ are given by
\begin{eqnarray}
\mathbf{M} &=& \matTT{-k_BT\Gamma_1\rho_1^b k^2}{\hspace{5mm}0}{0}
{-k_BT\Gamma_2\rho_2^b k^2}, \\
\mathbf{E} &=& \matTT{\,\left[\frac{1}{\rho_1^b}-\hat{c}_{11}(k)\right]}{-\hat{c}_{12}(k)}
{-\hat{c}_{21}(k)}{\left[\frac{1}{\rho_2^b}-\hat{c}_{22}(k)\right]\,}.
\end{eqnarray}
It follows that
\begin{equation}
\omega(k) = \frac{1}{2}\mbox{Tr}(\mathbf{M \cdot E}) \pm
\sqrt{\frac{1}{4}\mbox{Tr}(\mathbf{M \cdot E})^2 - |\mathbf{M \cdot E}|}.
\label{eqBetaTwoComp}
\end{equation}
where $|\mathbf{M \cdot E}|$ denotes the determinant of the matrix $\mathbf{M \cdot E}$.
When $\omega(k)<0$ for all wave numbers $k$, the system is linearly stable. If, however, $\omega(k)>0$ for any wave number $k$ then the uniform liquid is linearly unstable. Since $\mathbf{M}$ is a (negative definite) diagonal matrix its inverse $\mathbf{M^{-1}}$ exists for all nonzero densities and temperatures, enabling us to write Eq.\ \eqref{eqMatrixForm} as a generalised eigenvalue problem:
\begin{equation}
(\mathbf{E} - \mathbf{M^{-1}} \omega) \hat{\rho} = 0.
\end{equation}
As $\mathbf{E}$ is a symmetric matrix, all eigenvalues are real as one would expect for a relaxational system. It follows that the threshold for linear instability is determined by $|\mathbf{E}| = 0$, i.e., by the condition
\begin{equation}
D(k)\equiv [1-\rho_1^b\hat{c}_{11}(k)][1-\rho_2^b\hat{c}_{22}(k)]-\rho_1^b\rho_2^b\hat{c}_{12}^2(k)=0.
\label{eq:lin_instab_crit}
\end{equation}
The partial structure factors $S_{ij}(k)$ for an equilibrium fluid mixture are given by \cite{HM, ArEv01, ALE02, ALE04, BhTh70, Salmon06}
\begin{eqnarray}
\nonumber
S_{11}(k) & = & 1 + \rho_1^b{\hat h}_{11}(k),\\
\nonumber
S_{22}(k) & = & 1 + \rho_2^b{\hat h}_{22}(k),\\
S_{12}(k) & = & \sqrt{\rho_1^b\rho_2^b}{\hat h}_{12}(k),
\end{eqnarray}
where ${\hat h}_{ij}(k)$ are the Fourier transforms of $h_{ij}(r)$, i.e., of the fluid pair correlation functions. These are related to the pair direct correlation functions $c_{ij}^{(2)}(r)$ through the Ornstein-Zernike (OZ) equations \cite{HM,ArEv01}. In Fourier space the OZ equations are
\begin{equation}
{\hat{h}}_{ij} (k) \, =\, \frac{N_{ij}(k)}{D(k)},
\end{equation}
with the three numerators given by
\begin{eqnarray}
N_{11} (k) &=& {\hat{c}}_{11}(k)+ \rho_2^b[{\hat{c}}_{12}^2(k)-
{\hat{c}}_{11}(k){\hat{c}}_{22}(k)]\notag \\
N_{22} (k) &=& {\hat{c}}_{22}(k)+ \rho_1^b[{\hat{c}}_{12}^2(k)-
{\hat{c}}_{11}(k){\hat{c}}_{22}(k)]\notag \\
N_{12} (k) &=& {\hat{c}}_{12}(k).
\end{eqnarray}
Since for an equilibrium fluid $S_{11}(k)>0$, $S_{22}(k)>0$ and $S_{11}S_{22}-S_{12}^2>0$ for all values of $k$, it follows that $D(k)>0$ and hence that $\omega(k)<0$ for all wave numbers $k$. Thus all Fourier modes decay over time. Within the present RPA theory for GEM-$n$ particles $\hat{c}_{ij}(k)=-\beta \hat{v}_{ij}(k)$, where $\hat{v}_{ij}(k)$ are the Fourier transforms of the pair potentials in Eq.~\eqref{eq:pair_pot}, and for sufficiently high densities $D(k)$ dips below zero. Thus $\omega(k)>0$ for a band of wave numbers around $k\approx k_c$, indicating that the fluid has become linearly unstable.
For a one-component fluid, i.e., in the limit of $\rho_2^b\to 0$, we find that the fluid is stable when $[1-\rho^b\hat{c}(k)]>0$ but becomes linearly unstable when $[1-\rho^b\hat{c}(k)]<0$ \cite{ArEv04, ARTK12}. The loci $D(k=k_c)=0$ for both the GEM-4 and GEM-8 models are displayed as dashed lines in Fig.~\ref{fig:phase_diag}. In both cases the line along which the liquid phase becomes linearly unstable is located well inside the region where the crystal is the equilibrium phase.
\section{Solidification fronts in the one-component GEM-4 model}
\label{sec:1comp_fronts}
\begin{figure}[t]
\includegraphics[width=1.\columnwidth]{front_speed.pdf}
\includegraphics[width=1.\columnwidth]{front_speed_mu.pdf}
\caption{\label{fig:speed} (Color online) The front speed (a) as a function of the density of the metastable liquid into which the front propagates and (b) as a function of the chemical potential, for a GEM-4 fluid with temperature $k_BT/\epsilon=1$. The red solid line is the result from the marginal stability analysis and the black dashed line is the result from numerical computations from profiles such as that displayed in Fig.\ \ref{fig:front_profile}. The black circles denote (a) the densities $\rho_{\rm l}$, $\rho_{\rm s}$ at liquid-solid coexistence, and (b) the coexistence value $\beta \mu\approx17.0$.}
\end{figure}
\begin{figure*}
\includegraphics[width=0.99\textwidth]{1.png}
\includegraphics[width=0.9\textwidth]{front.pdf}
\includegraphics[width=0.9\textwidth]{front_log.pdf}
\caption{\label{fig:front_profile} (Color online) Density profile across a solidification front advancing from left to right into an unstable GEM-4 liquid with bulk density $\rho R^2=8$ and temperature $k_BT/\epsilon=1$, calculated from DDFT. The top panel shows the full 2D density profile $\rho(x,y)$ while the panel below shows the 1D density profile $\rho(x)$ obtained by averaging over the $y$-direction, perpendicular to the front. The bottom panel shows $\ln(|\rho(x)-\rho^b|R^2)$ in order to reveal the small amplitude oscillations at the leading edge of the advancing front.}
\end{figure*}
When the system is linearly unstable, any localised density modulation will grow and advance into the unstable uniform liquid phase. In Refs.\ \cite{GaEl11, ARTK12}, a marginal stability analysis was used to calculate the speed of such a front for the PFC model. Such a calculation allows one to obtain the speed of a front that has advanced sufficiently far for all initial transients to have decayed, so that the front attains a stationary front velocity. In 1D the speed $c$ with which the front advances into the unstable liquid may be obtained by solving the following set of equations \cite{DL83, BBDKL85, GaEl11, ARTK12}:
\begin{eqnarray}\label{eq:speed1}
{\rm i}c+\frac{d \omega(k)}{dk}=0\\
{\rm Re}[{\rm i}ck+\omega(k)]=0,
\label{eq:speed2}
\end{eqnarray}
corresponding to a front solution moving with speed $c$ that is marginally stable to infinitesimal perturbations in its frame of reference. In such a front the density profiles are coupled (via the solution of the linear problem (\ref{eqMatrixForm})) and both take the form $\tilde{\rho}(\rr,t) = \rho_{\rm front}(x-ct)$, where $\rho_{\rm front}(x-ct) \sim \exp(-k_{\rm im}x)\sin(k_{\rm r} (x-ct)+{\rm Im}[\omega(k)]t)$. Here $k_{\rm r}$ and $k_{\rm im}$ are the real and imaginary parts of the complex wave number $k\equiv k_{\rm r}+{\rm i}k_{\rm im}$. The speed calculated from this approach for the one-component GEM-4 model is displayed as the solid red line in Fig.\ \ref{fig:speed}(a) as a function of the density of the unstable liquid and in (b) as a function of the chemical potential $\mu$, both for $\beta \epsilon=1$. We also display the front speed calculated numerically using DDFT in 2D. Figure \ref{fig:front_profile} shows typical 2D and 1D density profiles used for determining the front speed $c$. The figure shows that the invasion of the metastable liquid state in fact occurs via a {\it pair} of fronts, the first of which describes the invasion of the liquid state by an unstable pattern of stripes, while the second describes the invasion of the unstable stripe pattern by a stable hexagonal state. By ``stripes'' we mean a density profile with oscillations perpendicular to the front, but no density modulations parallel to the front. This double front structure complicates considerably the description of the invasion process in 2D (see Appendix). Figure \ref{fig:speed} shows measurements of the speed of propagation of the hexagons-to-stripes front, obtained by comparing profiles like that in Fig.\ \ref{fig:front_profile}(a) at two successive times and determining the speed of advance of the hexagonal state when it first emerges from the unstable stripe state. The speed of the stripe pattern is harder to measure since the pattern is itself unstable and so never reaches a substantial amplitude. For this reason we measure the speed of the stripe-to-liquid front from plots of the logarithm of the density fluctuations (Fig.~\ref{fig:front_profile}(c)) which emphasizes the spatial growth of the smallest fluctuations at the leading edge of the front.
For $\beta\epsilon=1$ the uniform liquid is linearly stable for $\beta\mu\lesssim 19.6$ and unstable for $\beta\mu\gtrsim19.6$. The marginal stability prediction, obtained by solving Eqs.~(\ref{eq:speed1}) and (\ref{eq:speed2}), predicts that the 1D speed increases with $\beta\mu$ (or with increasing density $\rho$) in a square-root manner, as indicated by the solid red line in Fig.~\ref{fig:speed}. Since the theory is 1D this prediction applies to the invasion of the liquid state by the stripe pattern. Despite this we find that the prediction correctly describes the speed of the hexagons-to-stripes front for $\beta\mu \gtrsim21.5$ (i.e.\ for $\rho R^2 \gtrsim7$), as measured in numerical simulations of the DDFT for the GEM-4 fluid, suggesting that the two fronts are locked together and that the front speed is selected by linear processes at the stripe-to-liquid transition, i.e., the resulting double front is a {\it pulled} front \cite{saarloos}. For smaller values of $\beta\mu$ the speed of the hexagonal state departs substantially from the marginal stability prediction and the stripe section is swallowed by the faster moving hexagons-to-liquid front. Indeed, for $\beta\mu\lesssim 19.6$ (i.e.\ for $\rho R^2\lesssim 6.38$) the stripe state is absent altogether, as can be verified by performing a parallel study in one spatial dimension. The bifurcation to stripes is therefore supercritical. The hexagons-to-liquid front present in the metastable regime below the onset of linear instability of the liquid state is stationary at the Maxwell point at $\beta\mu\approx 17.0$, corresponding to the location of thermodynamic coexistence between the liquid and hexagonal states. For $\beta\mu > 17.0$ the hexagonal state advances into the liquid phase (the opposite occurs for $\beta\mu < 17.0$) and the hexagons-to-liquid front is {\it pushed} \cite{saarloos}: in this regime the front propagates via a nonlinear process since the liquid phase is linearly stable. The situation is more subtle when plotted as a function of the liquid density $\rho R^2$: when the liquid density takes a value in the interval $5.48\lesssim \rho R^2\lesssim 5.73$, i.e., between the densities of the liquid and crystalline states at coexistence, one cannot define a unique front speed. In this regime any front between these two states will slow down and, in any finite domain, eventually come to a halt. This occurs because the density $\rho_0$ of the liquid state into which the front moves is less than the density $\rho_\mathrm{s}$ of the crystal at coexistence but larger than the density $\rho_\mathrm{l}$ of the liquid at coexistence. In this situation, the moving `front' has a substructure consisting of two transitions: one from $\rho_\mathrm{s}$ to a depletion zone of a density close to $\rho_\mathrm{l}$ and another one from the depletion zone to the initial $\rho_0$. As the depletion zone widens in time and limits the diffusion from the region of density $\rho_0$ to the crystalline zone of density $\rho_\mathrm{s}$ the front slows down. In a finite system, the depletion zone moves and extends until it reaches the boundary and the system equilibrates in a state partitioned between a liquid with density $\rho_\mathrm{l}$ and crystal with density $\rho_\mathrm{s}$ with a stationary front between them. For a PFC model the role of the depletion zone in crystal growth is discussed in Ref.~\cite{TGTD2011sm}.
The speed of the hexagons-to-liquid front in the regime $17.0\lesssim \beta\mu\lesssim 19.6$ is determined uniquely (see Appendix). Refs.~\cite{HN00} and \cite{DSSS} predict that this is no longer the case for $\beta \mu \gtrsim 19.6$, but in practice we find that the front has a well-defined speed, possibly as a result of pinning of the stripes-to-liquid front to the stripes behind it, and of the hexagons-to-stripes front to the heterogeneity on either side. Both effects are absent from the amplitude equation formulation employed in Refs.~\cite{HN00} and \cite{DSSS}, that we analyse in the Appendix. Moreover, when the hexagon speed reaches the speed predicted by the marginal stability theory for the stripe state, the two fronts appear to lock and thereafter move together. In the theory based on amplitude equations summarized in the Appendix, the interval of stripes between the two fronts appears to have a unique width, depending on $\beta\mu$, a prediction that is consistent with our DDFT results. We have not observed the ``unlocking'' of the hexagons-to-stripes front from the stripes-to-liquid front noted in Ref.\ \cite{HN00} at yet larger values of $\beta\mu$. Possible reasons for this are discussed in the Appendix.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{k_star.pdf}
\caption{(Color online) The wave number $k^*$ of the stripe state produced behind the front as a function of density for the GEM-4 fluid with $\beta\epsilon=1$, obtained from Eq.~(\ref{eq:kbehind}) together with the wave numbers $k_{\rm r}$ of the 1D oscillations at the leading edge of the front and $k_{\rm eq}\equiv 2 \pi/\lambda$, where $\lambda$ is the distance between lattice planes in the {\em equilibrium} hexagonal state. This wavelength is very different from the wavelength of the oscillations produced by the advancing front, $2\pi/k^*$.
}
\label{fig:k_star}
\end{figure}
It is clear, therefore, that the 1D analysis based on Eqs.~(\ref{eq:speed1}) and (\ref{eq:speed2}) allows us to calculate the front speed when the unstable liquid is quenched deeply enough so that fronts propagate via linear processes. In addition to the front speed $c$ this analysis gives $k_{\rm r}$, the wave number of the growing perturbation at the leading edge of the front and $k_{\rm im}$, which defines the spatial decay length of the density oscillations in the forward direction. Within the 1D description the pattern left behind by the front is a large amplitude periodic state with wave number $k^*$, say. When no phase slips take place, this wave number is given by the expression \cite{BBDKL85,ARTK12}
\begin{equation}
k^*=k_{\rm r}+\frac{1}{c}{\rm Im}[\omega(k)].
\label{eq:kbehind}
\end{equation}
The wave number $k^*$ differs in general from $k_{\rm r}$. Moreover, as demonstrated in Ref.\ \cite{ARTK12} and confirmed in Fig.\ \ref{fig:k_star} for a GEM-4 crystal with temperature $\beta \epsilon=1$, the wavelength $2\pi/k^*$ of the density modulation that is created by the passage of the front can be very different from the scale $2\pi/k_{\rm eq}$ of the minimum free energy crystal structure which corresponds here to hexagonal coordination. The propagation of the solidification front therefore produces a frustrated structure that leads to the formation of defects and disorder in the crystal. Thus, we identify two sources of frustration: the wave number mismatch and the competition between the stripe state deposited by the advancing front and its subsequent transformation into a 2D hexagonal structure with a different equilibrium wavelength. Both effects generate disorder behind the advancing front and significant rearrangements in the structure of the modulation pattern occur as the system attempts to lower its free energy via a succession of local changes in the wavelength of the density modulation \cite{ARTK12}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{6_sidebar.png}
\includegraphics[width=0.9\columnwidth]{8_sidebar.png}
\caption{(Color online) Density profiles obtained from DDFT for an unstable GEM-4 fluid with bulk density $\rho_0R^2=8$. To facilitate clear portrayal of the front structure we plot the quantity $\ln(R^2|\rho(\rr)-\rho_0|)$. Solidification is initiated along the vertical line $x=0$ at time $t^*=0$. This produces two solidification fronts, one moving to the left, the other to the right, moving away from the line $x=0$. The upper profile is for the time $t^*=1$ and the lower for $t^*=1.4$. We see significant disorder as the front creates density modulations that are not commensurate with the equilibrium crystal structure.}
\label{fig:front_profiles}
\end{figure}
\begin{figure}
\includegraphics[width=0.9\columnwidth]{rho8_Angle.pdf}
\includegraphics[width=0.99\columnwidth]{rho8_triangulation12.pdf}
\includegraphics[width=0.99\columnwidth]{rho8_triangulation23.pdf}
\caption{(Color online) Top panel: the angle distribution $p(\theta)$ at times $t^*=2.2$, 3.2 and 4.4 after the initiation of a solidification front for a GEM-4 fluid with bulk density $\rho_0R^2=8$ (cf.\ Fig.\ \ref{fig:front_profiles}) computed from the triangles of a Delauney triangulation on the density peaks of the profile from DDFT (middle panel: $t^*=2.2$, bottom panel: $t^*=4.4$).}
\label{fig:Delauney}
\end{figure}
This ageing process can be rather slow \cite{ARTK12}. We illustrate its properties in Figs.~\ref{fig:front_profiles} and \ref{fig:Delauney}. Figure \ref{fig:front_profiles} displays the density profile in a part of the domain as computed from DDFT, and confirms the presence of substantial disorder in the crystalline structure close behind the advancing solidification front. There are actually two fronts in the profiles displayed in Fig.\ \ref{fig:front_profiles}, moving to the left and to the right away from the vertical line $x=0$, where the fronts are initiated at time $t=0$. Although there is substantial disorder close behind the front, further back the crystal has had time to rearrange itself into its equilibrium structure, thereby reducing the free energy. Overall, the process is similar to that observed in the PFC model \cite{ARTK12}. We quantify the rearrangement process using Delauney triangulation \cite{delauney}, as shown in Fig.~\ref{fig:Delauney}. Figure \ref{fig:Delauney}(a) displays the bond angle distribution $p(\theta)$ obtained from Delauney triangulation on the peaks of the density profile at various times after the solidification front was initiated. The distribution $p(\theta)$ has a single peak centered near $60^\circ$, which is not surprising since the triangulation on a hexagonal crystal structure yields equilateral triangles. The initial structure has a significant number of (penta-hepta) defects. Over time, the number of these defects gradually decreases, as shown by the fact that the width of the peak in $p(\theta)$ decreases over time, but the defects never completely disappear. These results show that the one-component GEM-4 system is able to rearrange itself after solidification to form a reasonably well-ordered polycrystalline structure, albeit with defects, but with the equilibrium scale $2\pi/k_{\rm eq}$ present throughout the domain.
\section{Results for a binary system}
\label{sec:bin_system}
\begin{figure}
\includegraphics[width=1.\columnwidth]{phase_diag2.pdf}
\caption{The linear stability limit for a binary mixture of GEM-8 particles with $\beta \epsilon=1$ and $R_{22}/R_{11}=1.5$ and $R_{12}/R_{11}=1$, plotted in the total density $\rho\equiv\rho_1+\rho_2$ vs concentration $\phi\equiv\rho_1/\rho$ plane. The black circles denote the state points corresponding to the density profiles displayed in Fig.~\ref{fig:binary_mixture_profiles}.}
\label{fig:lin_stab_line}
\end{figure}
\begin{figure*}
\includegraphics[width=0.65\columnwidth]{dens_rho4_x0_0_small_dx0_055.png}
\includegraphics[width=0.65\columnwidth]{dens_rho4_x0_10_dx0_055_a.png}
\includegraphics[width=0.65\columnwidth]{dens_rho4_x0_25_small_dx0_055.png}
\includegraphics[width=0.65\columnwidth]{dens_rho4_x0_5_ordered.png}
\includegraphics[width=0.65\columnwidth]{dens_rho4_x0_90_small_dx0_055.png}
\caption{Equilibrium crystal structures for a GEM-8 binary mixture with $\beta \epsilon_{ij}=\beta \epsilon=1$ for all $i,j=1,2$, $R_{22}/R_{11}=1.5$, $R_{12}/R_{11}=1$, with total average density $\rho R_{11}^2=4$ and concentrations (a) to (e) $\phi=0$, 0.1, 0.25, 0.5 and 0.9. The structures are shown in terms of the quantity $[\rho_1(\rr)-\rho_2(\rr)]R_{11}^2$, with regions where $\rho_1(\rr)>\rho_2(\rr)$ colored black. All profiles correspond to local minima of the free energy, but we have not checked whether they correspond to global minima at the given state points. We observe a binary square lattice structure in (c), a binary hexagonal lattice structure in (b) and (d), and a simple hexagonal lattice in (a) and (e), where the minority species particles occupy the same lattice sites as the majority species particles, in contrast to the lattice structures in (b)--(d). The density profiles of species 1 or 2 in case (e) are very similar to the profile shown, the only difference being the height of the density peaks.}
\label{fig:binary_mixture_profiles}
\end{figure*}
\begin{figure}
\includegraphics[width=0.9\columnwidth]{31_sidebar.png}
\includegraphics[width=0.9\columnwidth]{151_sidebar.png}
\caption{Snapshots of a solidification front in a GEM-8 mixture with $\beta \epsilon_{ij}=\beta \epsilon=1$ for all $i,j=1,2$, $R_{22}/R_{11}=1.5$ and $R_{12}/R_{11}=1$, advancing from left to right into an unstable fluid with $\rho R_{11}^2=8$ and $\phi=0.5$, in terms of the quantity $[\rho_1(\rr)-\rho_2(\rr)]R_{11}^2$. Density peaks of species 1 are colored black while the peaks of species 2 are white. The front was initiated at time $t=0$ along the line $x=0$. The top profile corresponds to time $t^*=0.6$ while the lower profile corresponds to $t^*=3$.}
\label{fig:front_bin}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.66\columnwidth]{10Voronoi.pdf}
\includegraphics[width=0.66\columnwidth]{10Triangles.pdf}
\includegraphics[width=0.66\columnwidth]{50_50_t2.pdf}
\includegraphics[width=0.66\columnwidth]{1999Voronoi.pdf}
\includegraphics[width=0.66\columnwidth]{1999Triangles.pdf}
\includegraphics[width=0.66\columnwidth]{50_50_t400.pdf}
\caption{(Color online) Analysis of the density peaks in the density profile in a GEM-8 mixture with $\beta \epsilon_{ij}=\beta \epsilon=1$ for all $i,j=1,2$, $R_{22}/R_{11}=1.5$ and $R_{12}/R_{11}=1$ and average total density $\rho R_{11}^2=8$ and concentration $\phi=0.5$, formed by a solidification front initiated along the line $x=25$ at time $t=0$. The diagrams along the top row correspond to time $t^*=2$, shortly after the solidification front has exited the domain and before the structure has had time to relax, while the diagrams along the bottom row correspond to time $t^*=400$, when the profiles no longer change in time -- the system has reached a minimum of the free energy. Left: Voronoi diagrams -- the construction reveals the disorder created by the front. The hexagons and squares correspond to the two competing crystal structures. Middle: Delauney triangulation -- domains of the hexagonal phase (equilateral triangles) are highlighted in red, while the remainder, including the right-angled triangles of the square phase, are shown in black. Right: the density maxima are color-coded according to the triangle type they belong to: right-angled triangles are black, equilateral are grey (red online) and scalene are open circles. Comparing the upper to the lower diagrams, we see that over time there is an increase in the size of the domains of the two different crystal structures.}
\label{fig:Morgan_analysis}
\end{figure*}
\begin{figure}[htbp]
\centering
\includegraphics[width=1.\columnwidth]{rho8_Angle_2Comp.pdf}
\caption{(Color online) Time evolution of the bond angle distribution function from Delauney triangulation, corresponding to the results in Fig.~\ref{fig:Morgan_analysis}.}
\label{fig:bond_angle_Morgan_analysis}
\end{figure}
Our results from the previous section and also those in \cite{ARTK12} indicate that solidification fronts for systems that have been deeply quenched in general do not produce density modulations with the wavelength of the equilibrium crystal structure. In the quenched one-component fluid discussed in the previous section, the system is subsequently able to rearrange to form the crystal, with only a few defects remaining. However, this begs the interesting question whether in some systems the density peaks are not able to rearrange so that the disorder generated by the solidification front remains. What is well known from the glass transition literature is that quenched binary mixtures are far more likely than one-component systems to form a glass instead of an ordered crystal -- see, for example, Ref.\ \cite{BHHP87}. In order to pursue this idea, we have performed similar computations for a binary mixture of GEM-8 particles with $\beta \epsilon_{ij}=\beta \epsilon=1$ for all $i,j=1,2$ and $R_{22}/R_{11}=1.5$ and $R_{12}/R_{11}=1$. In Fig.\ \ref{fig:lin_stab_line} we display the linear instability threshold for different values of the concentration $\phi\equiv\rho_1/\rho$, where $\rho\equiv\rho_1+\rho_2$ is the total density and $\rho_1$, $\rho_2$ are the densities of the two components of the mixture. For state points above the linear instability threshold line in Fig.\ \ref{fig:lin_stab_line} the uniform fluid is unstable and the system freezes to form a periodic solid. This line is obtained by tracing the locus defined by $D(k_c)=0$, where $D(k)$ is given by Eq.\ \eqref{eq:lin_instab_crit} and $k_c\neq0$ is the wave number at the minimum of $D(k)$ (i.e.\ $\frac{d}{dk}D(k=k_c)=0$). The cusp in the linear instability threshold in Fig.~\ref{fig:lin_stab_line} is a consequence of a crossover from linear instability at one length scale to linear instability at a different lengthscale. At the cusp point, which is at $\rho R_{11}^2=3.77$ and $\phi=0.708$, the system is marginally unstable at two length scales \cite{ARK13}.
This binary mixture exhibits at least four different crystalline phases; examples of these are displayed in Fig.\ \ref{fig:binary_mixture_profiles}. Owing to the fact that the number of potential crystal structures for binary systems of soft-core particles is rather large, we have not attempted to calculate the full phase diagram for this system or the location of the phase transitions between the different structures observed. For clarity the figure shows the quantity $[\rho_1(\rr)-\rho_2(\rr)]R_{11}^2$ with regions where the density of species 1 is higher than that of species 2 indicated in black. For large values of the concentration $\phi$, the system forms a simple hexagonal crystal that is essentially the same as that formed by the pure species 1 system. The minority species 2 particles simply join in low concentration the density peaks formed by the majority species 1 particles -- see Fig.\ \ref{fig:binary_mixture_profiles}(e). Similarly, for very low concentrations $\phi$, the system forms a simple hexagonal crystal, essentially that formed by the pure species 2 system -- see Fig.\ \ref{fig:binary_mixture_profiles}(a). However, for intermediate densities the system forms a binary hexagonal crystal structure, where the two different particle species sit on different lattice sites. Examples of this crystal structure are displayed in Figs.\ \ref{fig:binary_mixture_profiles}(b) and (d). We also observe a square crystal structure -- see Fig.\ \ref{fig:binary_mixture_profiles}(c) -- in which the two different species also reside on different lattice sites.
When the system contains roughly the same number of each species of particles, i.e.\ $\phi\approx0.5$, we find that either the square or the binary hexagonal crystal structures can be formed, depending on initial conditions, indicating that there is close competition between these two different crystal structures. This can also be seen in Fig.\ \ref{fig:front_bin}, where we display profiles calculated from DDFT after the uniform fluid is quenched to this state point and a solidification front is initiated along the line $x=0$. These profiles reveal that the front generates regions of both square and hexagonal crystalline structures. Furthermore, the system is highly disordered, as one might expect based on the demonstration in Sec.\ \ref{sec:1comp_fronts} that the density modulations created behind a solidification front in a deeply quenched system do not have the same wavelength as the equilibrium crystal. Thus, significant rearrangements are needed to get to the equilibrium structure. In the present case, there are two competing structures (squares and hexagons) and the resulting profile contains a mixture of the two. However, because the system is a binary mixture, it is unable to rearrange over time and so significant disorder remains indefinitely. In Figs.\ \ref{fig:Morgan_analysis} and \ref{fig:bond_angle_Morgan_analysis} we display a more detailed analysis of the structure created by the solidification front, and how this structure evolves over time. This analysis is based on performing a Delauney triangulation on the structures that are formed and determining its dual, the Voronoi diagram \cite{delauney}. To do this we first calculate the locations of all the peaks in the total density profile $\rho(\rr)\equiv\rho_1(\rr)+\rho_2(\rr)$. We include all maxima where the density at the maximum point is $>50R_{11}^{-2}$, and construct the Delauney triangulation and the Voronoi diagram on this set of points. The Voronoi diagrams are displayed on the left in Fig.~\ref{fig:Morgan_analysis} while the center panels display the Delauney triangulation. The upper diagrams correspond to a short time $t^*=2$ after the front was initiated along a line down the centre of the system while the lower profiles correspond to a much later time, $t^*=400$, which is roughly when the structure ceases to evolve in time. In the Voronoi diagram we observe regions of both squares and hexagons and in between these different regions we see various different polyhedra corresponding to the defects along the (grain) boundaries between the regions of different crystal structure and/or orientation. These different crystal regions can also be observed in the Delauney triangulation as regions made up of equilateral triangles (coloured red online), corresponding to the hexagonal structure, and regions of right-angled triangles, corresponding to the square crystal structure. The boundaries between these regions contain scalene triangles. In the right hand panels in Fig.\ \ref{fig:Morgan_analysis} we display the density maxima in $\rho(\rr)$. These are color-coded according to the nature of the local crystal structure around that point. The square crystal regions are displayed as black circles, the hexagonal regions as grey circles (red online) and the density peaks with neither square nor hexagonal local coordination are plotted as open circles. The criteria for deciding to which subset a given density peak belongs is based on the Delauney triangulation: any given triangle with corner angles $\theta_1$, $\theta_2$ and $\theta_3$ is defined as equilateral if $|\theta_i-\theta_j|<5^{\circ}$ for all pairs $i,j=1,2,3$. The vertices of these triangles are colored black. Similarly, triangles are defined as right-angled if for the largest angle $\theta_1$ we have $|\theta_1-90^\circ |<5^{\circ}$ AND for the other two angles $|\theta_2-\theta_3|<5^{\circ}$. The vertices of these triangles are colored grey (red online). The remaining vertices which fall into neither of these categories are displayed as open circles. We see that there are roughly equal-sized regions of both square and hexagonal ordering. The typical size of these different regions increases with the elapsed time after the solidification front has passed through the system. Likewise, the number of maxima that do not belong to either crystal structure (open circles) decreases with elapsed time, as the system seeks to minimize its free energy.
In Fig.\ \ref{fig:bond_angle_Morgan_analysis} we plot the distribution function $p(\theta)$ for the different bond angles obtained from Delauney triangulation, for three different times after the initiation of the solidification front. It has three maxima: one near 45$^\circ$, another at 60$^\circ$ and the other near 90$^\circ$. The peak at 60$^\circ$ is the contribution from the regions of hexagonal ordering (equilateral triangles) and the two peaks at 45$^\circ$ and 90$^\circ$ come from the regions of square ordering (right-angled triangles in the Delauney triangulation). The peak at 45$^\circ$ is, of course, twice as high as the peak at 90$^\circ$. We also observe that the peaks are much broader at short times, $t^*=1$, $2$, after the solidification front was initiated, than in the final structure from time $t^*=400$. These results provide an indication of the degree of disorder and number of defects in the system; the fact that the peaks become sharper over time is a consequence of the fact that the amount of disorder in the system decreases over time. Nonetheless, the peaks in $p(\theta)$ are still rather broad in the final state, indicating that significant strain and disorder remain in the structure.
\section{Concluding remarks}
\label{sec:conc}
In this paper we have seen that a deep quench generates a solidification front whose speed is correctly predicted from the dispersion relation using the marginal stability Ansatz. The front leaves behind a nonequilibrium crystalline state with many defects and a characteristic scale that differs substantially from the wavelength of the crystal in thermodynamic equilibrium. Subsequent aging generates domains with different orientations but in one-component systems the number of defects continues to decrease over time. In two-component systems different crystalline phases may compete, providing an additional source of disorder in the system, and the minority species may block rearrangement of the particles, thereby freezing the disorder in place, and leaving an amorphous solid with glass-like structure.
When the quench is shallow, the speed of the solidification front is slow and the amount of disorder generated by its passage is reduced. However, in this regime the front speed in a 2D system is no longer correctly predicted by the 1D marginal stability condition because the front becomes a pushed front, i.e., its speed is determined by nonlinear processes. As a result the speed becomes an eigenvalue of a nonlinear eigenvalue problem as summarized in the Appendix. The solution of this problem reproduces the qualitative features of Fig.~\ref{fig:speed} computed from numerical simulations of the DDFT for a one-component GEM-4 system (see Appendix), thereby providing support for this interpretation of Fig.~\ref{fig:speed}.
In particular, in the region of the phase diagram where the liquid is linearly stable and solidification fronts propagate via nonlinear processes, solidification must be nucleated -- a process that requires the system to surmount a free energy barrier. Once initiated, the resulting solidification front generates disorder in the system by the processes discussed above. However, in addition to these the nucleation process itself may play an important role as discussed in Refs.~\cite{TPTTG11, BaKl13, TXX14, GrTo14}. These studies show that the critical nucleus is likewise a structure that may be incommensurate with the equilibrium crystal structure so that the nucleation process itself can generate disorder in the system. This is especially so as one approaches the linear stability threshold, where the critical nucleus is predicted to have an `onion'--like structure \cite{BaKl13}. The second shell of the `onion' is incompatible with the equilibrium crystal structure, potentially leading to the growth of an amorphous phase, a suggestion supported by recent experimental results \cite{TXX14, GrTo14}. While one-component systems may subsequently be able to rearrange to form a well-ordered crystal, binary systems appear unable to escape the resulting disordered structure.
In the present work, we have studied solidification using DDFT with solidification initiated along a straight line (cf.\ Figs.~\ref{fig:front_profile} and \ref{fig:front_profiles}). The resulting fronts are straight, enabling us to study the front speed and wavenumber selection. For example, the fronts in the linearly unstable liquid in Fig.\ \ref{fig:front_profiles} are initiated by adding a small zero-mean random perturbation along the line $x=0$ to the initially uniform density profile. In reality, however, solidification fronts are initiated throughout the system at random locations, determined by the fluctuations in the system. This is equivalent to initiating fronts simultaneously at many points in the system. These fronts then propagate through the system, colliding and interacting, leading to the formation of the solid phase. To model this process, we add a small zero-mean random perturbation to the initial density profile at all points in the system. The final $t\to\infty$ density profiles produced in this way (not displayed) are very similar to those produced by initiating the solidification front along a single line. If instead of DDFT we employed kinetic Monte Carlo, or Brownian dynamics or even molecular dynamics computer simulations to study solidifation in systems of particles interacting via the potentials in Eq.\ \eqref{eq:pair_pot}, we would first equilibrate the system in the liquid phase at a higher temperature and then quench to a temperature where a solid forms. The dynamics following such a quench is very similar to that predicted by DDFT from an initial density profile with random noise at all points in the system, as is the case for the related soft-core fluid model discussed in Ref.\ \cite{ARK13}. We are thus confident that DDFT gives an excellent description of the system.
We mention, finally, that the behavior of the 2D PFC model studied in \cite{ARTK12} is qualitatively different from the 2D DDFT model studied here. In the PFC model there is a temperature-like parameter $r<0$, such that $(r+1)/2$ is the coefficient of the $\phi^2$ term in the PFC free energy. For the larger values of $|r|$ considered in \cite{ARTK12}, the linear instability threshold lies within the thermodynamic coexistence region between the liquid phase and the hexagonal crystalline phase. Thus, for these values of $|r|$, the hexagonal phase advances into the liquid at a well-defined speed determined by a linear mechanism as described by the marginal stability analysis. This is in contrast to the present DDFT model where the linear instability boundary lies outside the coexistence region (Fig.~\ref{fig:phase_diag}) and fronts between the hexagonal and liquid phases can propagate with speed determined either by a linear or a nonlinear mechanism, depending on parameters. However, for smaller values of $|r|$ the linear instability line in the PFC model does lie outside the coexistence region \cite{WK07, TP14} and in this case the behavior of the PFC system should be similar to that observed in the present study.
{\bf Acknowledgements}: The work of EK supported in part by the National Science Foundation under grant No. DMS-1211953 and a Chaire d'Excellence Pierre de Fermat of the R\'egion Midi-Pyr\'en\'ees, France. AJA and UT thank the Center of Nonlinear Science (CeNoS) of the University of M\"unster for recent support of their collaboration. We thank the anonymous referees for comments that helped shape the discussion in Sec.\ \ref{sec:conc}.
\section{Appendix: 2D front propagation into an unstable state}
\label{Appendix}
Figure \ref{fig:speed}(b) shows the front velocity $c$ as function of the chemical potential $\mu$ as computed from direct numerical simulations of a GEM-4 fluid with temperature $k_BT/\epsilon=1$ and compares the result with the prediction of the marginal stability calculation reported above (red solid line). The latter agrees well with the measured speed for larger values of $\mu$ but there is a substantial disagreement near threshold.
\begin{figure}[t]
\includegraphics[width=0.9\hsize]{frontSpeed_over_gamma_for_lambda1-2_version2.pdf}
\includegraphics[width=0.9\hsize]{front_equal_c.pdf}
\caption{(Color online) (a) The speeds $c_{\rm s}$, $c_{\rm hs}$ and $c_{\rm h}$ defined in the text as a function of $\gamma$ computed from the model system (\ref{A})--(\ref{B}) for $\lambda=1$ and $\lambda=2$. The results for $\lambda=1$ agree with those in Ref.~\cite{HN00}. The full range of values of $\gamma$ is shown including the Maxwell points $\gamma_{\rm M}$, where $c=0$, and the location of the critical values $\gamma_1$ and $\gamma_2$, where $c_{\rm h}=c_{\rm s}$ and $c_{\rm hs}=c_{\rm s}$, respectively. (b) The location of the Maxwell point and the critical values $\gamma_1$ and $\gamma_2$ as a function of the nonlinear coupling coefficient $\lambda$. The dotted line shows $-2.5\gamma_1$ and indicates that, in the range considered, the ratio $\gamma_M/\gamma_1$ is nearly constant.
}
\label{fig:speedHN}
\end{figure}
The reason for this discrepancy was elucidated by Hari and Nepomnyashchy \cite{HN00}, following earlier work by Csah\'ok and Misbah \cite{CM99}. The results of \cite{HN00} were largely confirmed in subsequent work by Doelman et al \cite{DSSS}. The work of Hari and Nepomnyashchy is based on a detailed study of a set of model equations describing the spatial modulation of a pattern of (small amplitude) hexagons:
\begin{eqnarray}\notag
\frac{\partial A_k}{\partial t}=&\gamma A_k+\frac{\partial^2 A_k}{\partial x_k^2}+A^*_{[k-1]}A^*_{[k+1]}\\
&-(|A_k|^2+\lambda|A_{[k-1]}|^2+\lambda|A_{[k+1]}|^2)A_k,
\end{eqnarray}
for $k=0,1,2$, where the $A_k$ are the complex amplitudes of the three wavevectors ${\bf n}_0\equiv (1,0)k_c$, ${\bf n}_1\equiv (-1,\sqrt{3})k_c/2$, ${\bf n}_2\equiv (-1,-\sqrt{3})k_c/2$ \cite{GSK}, and $x_k\equiv{\bf x}\cdot{\bf n}_k$. Here $k_c$ is the critical wave number at onset of the hexagon-forming instability ($\gamma=0$), and $[k\pm1]\equiv(k\pm1) ({\rm mod} 3)$. These equations constitute a gradient flow with free energy
\begin{eqnarray}
{\cal F}\equiv\int_{-\infty}^{\infty}L(x,t)\,dx,
\end{eqnarray}
where
\begin{eqnarray}
L=\sum_{k=0}^2\frac{1}{2}|\frac{\partial A_k}{\partial x_k}|^2-V
\end{eqnarray}
and
\begin{eqnarray}\notag
V\equiv&&\sum_{k=0}^2\biggl(\frac{1}{2}\gamma|A_k|^2-\frac{1}{4}|A_k|^4\biggr)+A_0^*A_1^*A_2^*\\ \notag
&&-\frac{\lambda}{2}\biggl(|A_0|^2|A_1|^2+|A_1|^2|A_2|^2+|A_2|^2|A_0|^2\biggr).
\end{eqnarray}
We focus on planar fronts perpendicular to ${\bf n}_0\equiv (1,0)k_c$ and thus suppose that the dynamics is independent of the variable $y$ along the front. Symmetry with respect to $y\rightarrow -y$ implies that $A_1=A_2\equiv B$, say. Absorbing the wave number $k_c$ in the variable $x$, and writing $A_0\equiv A$ we obtain the equations
\begin{equation}
\frac{\partial A}{\partial t}=\frac{\partial^2 A}{\partial x^2}+\gamma A+B^2-A^3-2\lambda AB^2 \label{e1}
\end{equation}
\begin{equation}
\frac{\partial B}{\partial t}=\frac{1}{4}\frac{\partial^2 B}{\partial x^2}+\gamma B+AB-(1+\lambda)B^3-\lambda A^2B.\label{e2}
\end{equation}
In writing these equations we have assumed that $A$ and $B$ are real to focus on the behavior of the amplitudes, thereby setting the phase $\Phi\equiv{\rm arg}(A)+2{\rm arg}(B)$ that distinguishes so-called up-hexagons from down-hexagons to zero \cite{GSK}.
These equations have solutions in the form of regular hexagons $(A,B)=(A_{\rm h}^{\pm},A_{\rm h}^{\pm})$, stripes $(A,B)=(A_{\rm s},0)$ and the homogeneous liquid state $(A,B)=(0,0)$, where
\begin{eqnarray}
A_h^{\pm}=\frac{1\pm\sqrt{1+4\gamma(1+2\lambda)}}{2(1+2\lambda)},\qquad A_s=\sqrt{\gamma},
\end{eqnarray}
corresponding to the critical points of the potential $V(A,B)=\frac{1}{2}\gamma(A^2+2B^2)+AB^2-[\frac{1}{4}A^4+\lambda A^2B^2+\frac{1}{2}(1+\lambda)B^4]$. The bifurcation to hexagons at $\gamma=0$ is transcritical and for $\gamma<0$ there are two hexagon branches: an unstable branch of small amplitude hexagons $A_{\rm h}^-$ and a stable branch of large amplitude hexagons $A_{\rm h}^+$. These annihilate at a saddle-node bifurcation at $\gamma=\gamma_{\rm sn}\equiv -\frac{1}{4(1+2\lambda)}$. Note that without loss of generality we have taken $A_{\rm h}^{\pm}$ and $A_{\rm s}$ to be positive since negative values can be compensated for by choosing $\Phi=\pi$, i.e., by an appropriate spatial translation.
The large amplitude hexagons $A_{\rm h}^+$ and the homogeneous state coexist stably in the subcritical regime, $-\frac{1}{4(1+2\lambda)}<\gamma<0$; the liquid state becomes unstable when $\gamma>0$. A front traveling with speed $c$ to the right, connecting $A_{\rm h}^+$ on the left with the liquid state $A=0$ to the right, takes the form
\begin{eqnarray}
A(x,t)={\tilde A}(\xi),\quad B(x,t)={\tilde B}(\xi),\quad \xi\equiv x-ct,
\end{eqnarray}
where
\begin{eqnarray}
\frac{\partial^2{\tilde A}}{\partial \xi^2}+c\frac{\partial{\tilde A}}{\partial \xi}+\gamma {\tilde A}+{\tilde B}^2-{\tilde A}^3-2\lambda {\tilde A}{\tilde B}^2=0,\label{A}\\
\frac{1}{4}\frac{\partial^2 {\tilde B}}{\partial \xi^2}+c\frac{\partial {\tilde B}}{\partial \xi}+\gamma {\tilde B}+{\tilde A}{\tilde B}-(1+\lambda){\tilde B}^3-\lambda {\tilde A}^2{\tilde B}=0 \label{B}
\end{eqnarray}
with the boundary conditions
\begin{eqnarray}\notag
{\tilde A}={\tilde B}=A_h^+ \quad{\rm as}\quad \xi\rightarrow-\infty,\\
{\tilde A}={\tilde B}=0 \quad{\rm as}\quad \xi\rightarrow\infty.
\end{eqnarray}
The speed $c$ vanishes in the subcritical regime when $\gamma=\gamma_M<0$ defined by the requirement $V(A_{\rm h},A_{\rm h})=V(0,0)=0$ and is positive for $\gamma>\gamma_{\rm M}$ ($V(A_{\rm h},A_{\rm h})<0$) and negative for $\gamma<\gamma_{\rm M}$ ($V(A_{\rm h},A_{\rm h})>0$). An elementary calculation gives $\gamma_{\rm M}=-\frac{2}{9(1+2\lambda)}$; $\gamma_{\rm M}$ thus corresponds to the Maxwell point between the trivial state $(0,0)$ and the hexagonal state $(A_{\rm h}^+,A_{\rm h}^+)$. Note that $\gamma_{\rm M}/\gamma_{\rm sn}=8/9$, independently of the value of $\lambda$. This prediction of the amplitude equations compares well with our numerical results for a GEM-4 mixture for which the chemical potential $\beta\mu_{\rm sn}\approx16.5$ and $\beta\mu_{\rm M}\approx16.8$ while the linear instability threshold corresponds to $\beta\mu_{\rm lin}\approx19.6$. Thus $(\mu_{\rm M}-\mu_{\rm lin})/(\mu_{\rm sn}-\mu_{\rm lin})\approx0.90$, very close to the predicted value $8/9$.
The situation is more complicated in the supercritical regime where $\gamma>0$ because this regime contains supercritical (but unstable!) stripes oriented parallel to the front. As a result one now finds fronts that connect the hexagonal structure to the stripe pattern and the stripe pattern to the liquid state, in addition to the front connecting the hexagonal structure and the (now unstable) liquid state. The marginal stability condition implies that stripes invade the homogeneous state with speed $c_{\rm s}=2\sqrt{\gamma}$, while an analogous calculation shows that the hexagons invade the unstable stripes with speed $c_{\rm hs}=[\sqrt{\gamma}-(\lambda-1)\gamma]^{1/2}$. This speed exceeds $c_{\rm s}$ in the interval $0<\gamma<\gamma_2\equiv(\lambda+3)^{-2}$, i.e., at $\gamma_2$ one has $c_{\rm hs}=c_{\rm s}$. The dependence of the speeds $c_{\rm hs}$ and $c_{\rm s}$ on $\gamma$ is shown in Fig.~\ref{fig:speedHN}(a) for $\lambda=1$ and $\lambda=2$.
It is evident that the speed $c_{\rm s}$ cannot be selected when $\gamma$ is too close to threshold $\gamma=0$ since $c$ must be positive for all $\gamma>\gamma_{\rm M}$. In the spatial dynamics picture of the front one seeks a heteroclinic connection between $(\tilde{A},\tilde{A})=(A_{\rm h},A_{\rm h})$ and $(0,0)$. Near $(0,0)$ we have the asymptotic behavior
\begin{eqnarray}
\tilde{A}\sim e^{\kappa_A\xi}\qquad \tilde{B}\sim e^{\kappa_B\xi},\qquad {\rm as}\quad \xi\rightarrow\infty,
\end{eqnarray}
where
\begin{eqnarray}
\kappa_A^{\pm}=-\frac{c}{2}\pm\frac{1}{2}\sqrt{c^2-4\gamma},\quad \kappa_B^{\pm}=-2c\pm 2\sqrt{c^2-\gamma}.
\end{eqnarray}
Evidently, for $\gamma<0$ the stable manifold of $(0,0)$ is two-dimensional, and since one expects the heteroclinic to connect to $(0,0)$ along the slow direction one anticipates that the solution will approach $(0,0)$ in the ``$A$" direction, with $\tilde{A}\sim e^{\kappa_A^-\xi}$ as $\xi\rightarrow\infty$. However, as soon as $\gamma>0$ the stable manifold of $(0,0)$ becomes four-dimensional, and the slowest direction is suddenly $\tilde{A}\sim e^{\kappa_B^+\xi}$. Hari and Nepomnyashchy \cite{HN00} solve the problem (\ref{e1})--(\ref{e2}) numerically and find that for $c<2\sqrt{\gamma_1}$ the front speed departs from the prediction $c=c_{\rm s}$ and instead follows a speed $c=c_{\rm h}$ for which the asymptotic behavior of the front continues to be $\tilde{A}\sim e^{\kappa_A^-\xi}$ as $\xi\rightarrow\infty$, thereby providing a smooth connection to the speed computed for $\gamma<0$. We refer to the value of $\gamma$ at which $c_{\rm h}=c_{\rm s}$ as $\gamma=\gamma_{1}$.
Hari and Nepomnyashchy \cite{HN00} also show that in the region $\gamma_{1}<\gamma<\gamma_2$ both the front connecting the hexagonal state to the stripes and the front connecting the stripes to the liquid state travel with the same speed $c_\mathrm{s}$. As a result the width of the stripe region between the hexagons and the liquid state remains constant; in numerical simulations this width was observed to be independent of the initial conditions adopted, despite the nonuniqueness of the overall front solution, and to increase with $\gamma$. Finally, for $\gamma>\gamma_2$ the front speed $c_{\rm s}>c_\mathrm{hs}$ and the front connecting the stripes to the liquid state outruns the hexagons invading the stripes and the width of the stripe interval in front of the hexagons grows without bound. In our models this behavior was not observed.
Figure \ref{fig:speedHN}(a) shows the computed fronts speeds as a function of the bifurcation parameter $\gamma$ for two values of the single nonlinear coupling coefficient $\lambda$ which is unknown for our GEM-4 model. In both cases the results behave qualitatively like those obtained from DDFT of this model system. In particular, we see that the speed $c_{\rm h}$ of the (pushed) hexagons increases monotonically from zero at the Maxwell point $\gamma_{\rm M}<0$ and terminates on the 1D stripe speed $c_{\rm s}$ obtained from the marginal stability at $\gamma=\gamma_1>0$; both $\gamma_{\rm M}$ and $\gamma_1$ decrease in magnitude as $\lambda$ increases (Fig.~\ref{fig:speedHN}(b)) and this is so for the point $\gamma=\gamma_2$ corresponding to the condition $c_{\rm s}=c_{\rm hs}$ as well. We mention that behavior similar to Fig.~\ref{fig:speedHN}(a) occurs even in 1D, provided only that the stripe state bifurcates subcritically before turning around towards larger values of the forcing parameter \cite{CC97}.
However, despite its qualitative success the model system (\ref{A})--(\ref{B}) fails in one key respect: it is not possible to match quantitatively the DDFT results for a shallow quench (Fig.~\ref{fig:speed}(b)) with the predictions of the model (Fig.~\ref{fig:speedHN}(b)). Specifically, the model predicts that $|\gamma_M|/\gamma_1\approx 2.5$ over the entire range of nonlinear coefficients $\lambda$ in Fig.~\ref{fig:speedHN}(b) while Fig.~\ref{fig:speed}(b) indicates that $|\gamma_M|/\gamma_1\approx 1.4$. For smaller $\lambda$ the ratio becomes yet larger. There are several issues that might contribute to this quantitative mismatch. First, the amplitude equations omit the phenomena of locking of the stripes-to-liquid front to the stripes behind the front, and of locking of the hexagons-to-stripes front to the heterogeneity ahead and behind the front. This is a consequence of modeling periodic structures using constant amplitude states, i.e., by spatially homogeneous states, resulting in the absence of the so-called nonadiabatic effects. Second, the amplitude equations are derived for nonconserved systems, while the DDFT system exhibits conserved dynamics. In the latter case we expect the equations for the amplitudes $A$ and $B$ to be coupled to a large scale mode, much as discussed in the work of Refs.~\cite{maco00,TARGK}. These aspects of the problem will be discussed in a future publication.
|
\section{Introduction}
Recently, the phenomenon of exciton-polariton condensation in quantum wells formed by semiconductor
microcavities bordered by Bragg mirrors has been found \cite{Kasprzak, Balili}. It spurred considerable
fundamental and applied interest as a new experimental realization of
Bose-Einstein condensate and a perspective base for optical switches, miniature terahertz lasers and ultra-sensitive gyroscopes \cite{Keeling2011,Carusotto2013}.
Condensation centers (CCs) may be localized by artificially
created spatial structures or disorder \cite{Balili, Lai}. Every
CC interacts with incoherent environment, getting excited by an
incoherent pumping source and decaying due to light radiation,
also interacting with their neighbors.
Till now, most results demonstrating rich dynamics of such systems
referred to the spatially homogeneous case. Only recently Aleiner,
Altshuler, and Rubo studied a pair of nonidential CCs, interacting
through mixed Josephson and radiative coupling \cite{Aleiner}. In
particular, they determined conditions of transition to
synchronous regime that corresponds to laser generation, found
bistability regimes and the region of chaotic behavior. At the
same time, the systems containing a greater number of CC
were left beyond the focus of the study. Moreover, the implemented
approach made use of the formal analogy to equations describing a
specific spin system and cannot not be directly applied to the
systems with greater number of CC \cite{Aleiner}. Noteworthy, the studies addressing very similar equations for coupled Van-der-Pol oscillators with generic nonlinearity and coupling terms also demonstrated a rich dynamics of synchronization, multistability, and chaotic regimes \cite{Ivanchenko,Kuznetsov,Astakhov,Emelianova}. The number of oscillators, though, was still confined to a few.
In this paper we extend and study the model of interacting underdamped (weakly lasing) exciton-polariton CCs to Ginzburg-Landau-type oscillatory arrays with the spatial inhomogeneity, known as diatomic in lattice dynamics. (Linear stability analysis of plane waves in spatially homogenous
DGLE with insights into soliton and complex pattern formation has been presented in previous studies \cite{Aranson, Ravoux, Kofane2006, Kofane2008}.) Developing mode formalism for active lattices, we derive and analyze the mode excitation and instability conditions in dependence on the parameters of the system. We study transition to complex multi-mode dynamics to find the two regimes of chaos: local and global in mode space. The crossover between them appears to be related to progressing dimensionality of modulational instability of the seed mode. Transition to global chaos is accompanied by the loss of synchronization.
We generalize the model of two interacting underdamped CCs \cite{Aleiner} to a discrete Ginzburg-Landau type equation array:
\begin{equation}\label{2}
\begin{split}
\dot
z_n=&\left(-g+i\Delta(-1)^n\right)\frac{z_n}{2}-(\beta+i\alpha)|z_n|^2\frac{z_n}{2}-\\&-(\gamma-iJ)\frac{z_{n-1}+z_{n+1}}{2},
\end{split}
\end{equation}
where $z_n=\sqrt{m}e^{i\varphi}$, $m$ and $\varphi$ are occupation
and phase of the $n$-th CC respectively, $n=1,\ldots,N$, periodic
boundary conditions are assumed $z_{N+1} \equiv z_1$.
Parameter $g$ is a net local linear dissipation rate (the
difference between the escape and incoming rates of the bosons for
a CC). Throughout the paper we assume that dissipation in an
individual CC prevails, that is $g>0$, and lasing becomes possible
only due to interaction between CCs. Parameters $\alpha$ and
$\beta$ describe nonlinear frequency shift and dissipation,
$\gamma$ and $J$ are ``radiative'' \cite{Aleiner} and Josephson
coupling between adjacent CCs, respectively, all assumed to take
non-negative values. $\Delta$ is frequency mismatch between
neighbor oscillators. Without the loss
of generality we set $\beta=J=1$.
The paper is organized as follows. In section \ref{sec2} we
introduce the mode formalism and employ it to study excitation and modulational
instability of plane waves in spatially homogeneous (subsection
\ref{sec2a}) and diatomic (subsection \ref{sec2b}) arrays. In
section \ref{sec3} we study the transition from local to multi-mode chaos and demonstrate its connection to
growing number of unstable directions for the most unstable waves.
\section{Mode formalism and modulational instability}\label{sec2}
\subsection{Spatially homogeneous case}\label{sec2a}
Let us first conduct a mode analysis for the homogeneous case
$\Delta=0$. The mode variables $a_k$ are introduced by
\begin{equation} \label{3}
z_n=\frac{1}{\sqrt{N}}\sum_k a_k e^{i k n},
\end{equation}
where $k=\frac{2\pi q}{N}$ are the mode wavenumbers with integer
$q=-\frac{N-1}{2},\ldots,\frac{N-1}{2}$ for $N$ odd, or
$q=-\frac{N}{2}+1,\ldots,\frac{N}{2}$ for $N$ even. Defined this
way, the wavenumber $k$ runs a set of $N$ values in the first
Brillouin zone $k\in (-\pi,\pi]$. This transform turns (\ref{2})
into \begin{multline}\label{4} \dot a_k = -
\left[\frac{g}{2} + (\gamma-i)\cos k \right] a_k -\\- \frac{1+ i
\alpha}{2N} \sum_{k_1 k_2 k_3} a_{k_1} a_{k_2} a_{k_3}^{\ast}
\delta_{2\pi} (k_1+k_2-k_3-k).
\end{multline}
Here the selective interaction is defined by $\delta_{2\pi}(k)=1$
if $k$ is equal to zero or a multiple of $2\pi$, and
$\delta_{2\pi}(k)=0$ otherwise. Note, that our special
statements related to the mode $k=\pi$ refer only to the case of
even $N$, when this mode exists.
These dynamical equations readily yield the excitation condition
of modes
\begin{equation}\label{5}
\frac{g}{2}+\gamma\cos k<0,
\end{equation}
which is wavenumber dependent, and is achieved, in general, with decrease of losses $g$ (or increase in pumping) or increase in radiative coupling $\gamma$.
It is easy to see that $\pi$-mode is the first one to get excited if $g<2\gamma$ and has the largest increment.
Conversely, if $g>2\gamma$ oscillations damp out and the zero equilibrium state is stable.
Noteworthy, Eq. (\ref{4}) has $N$ single-mode manifolds, namely,
solutions $a_{k_0}\neq 0,\quad a_{k\neq k_0}=0$, where $k_0$
is any permitted wavenumber. The dynamics on each manifold is described by
equation
\begin{equation}\label{6}
\dot a_{k_0} = - \left[\frac{g}{2} + (\gamma-i)\cos k_0 \right]
a_{k_0} - \frac{1+ i \alpha}{2N} |a_{k_0}|^2 a_{k_0}.
\end{equation}
If the excitation condition (\ref{5}) for that mode is
fulfilled, the dynamics on the manifold converges to a limit cycle of Eq. (\ref{6})
\begin{equation}\label{11}
a_{k_0}=A\sqrt{N}e^{i\Omega t},\quad A\in R,
\end{equation}
where
\begin{equation}\label{7}
A^2=-g-2\gamma\cos k_0,
\end{equation}
which is stable in the
corresponding manifold. In direct space (\ref{2}) this periodic solution
corresponds to a plane wave
\begin{equation}\label{8}
z_n(t)=Ae^{i\Omega t+ik_0n}.
\end{equation}
To investigate linear stability of these plane waves it
suffices to consider single or two side modes
perturbations, as determined by selective interaction in
(\ref{4}).
Elementary analysis of a single mode perturbations $a_k$
to the original mode $a_{k_0}$ yields a necessary stability criterion
\begin{equation}\label{9}
\frac{g}{2}+\gamma+2\gamma\cos k_0 < 0,
\end{equation}
determined again by the balance of net losses $g$, radiative
coupling $\gamma$, and is seed wave number $k_0$
specific.
Comparing (\ref{9}) to the excitation condition (\ref{5}), one
immediately recognizes that all modes except the one with the largest increment $k_0=\pi$ are
unstable on their excitation. Going away from bifurcation with
decreasing losses $g$ (equivalently, increasing pumping) or
increasing radiative coupling $\gamma$, this instability vanishes.
The exact stability criterion follows
from additionally introducing perturbations as pairs of modes,
with wave numbers symmetric relative to the wave number of the
seed wave. Linearizing Eq. (\ref{4}) in the neighborhood of a
single-mode periodic solution (\ref{11}) one obtains the increment
of a pair of modes $k=k_0+c$ and $k^{'}=k_0-c$ (see Appendix for
details):
\begin{equation}\label{10}
\begin{array}{l}
p=\frac{1}{2}Re\left[-2A^2+\lambda_k+\lambda_{k^{'}}^*+[(2A^2+\right.\\\left.-\lambda_k-\lambda_{k^{'}}^*)^2-4(i\Omega-\lambda_k+\right.\\\left.+dA^2)(-i\Omega-\lambda_{k^{'}}^*+d^*A^2)+d
d^*A^4]^{1/2}\right],
\end{array}
\end{equation}
where $\lambda_k=-\left[\frac{g}{2} +
(\gamma-i)\cos k\right]$, $d=1+i\alpha$, $\Omega=\cos k_0
-\alpha A^2/2$.
If $p<0$ for all side mode pairs, then the plane wave is linearly stable. Conversely, if $p>0$ at least for one pair of side modes the solution is unstable.
Eqs. (\ref{5}) and (\ref{10}) impose simultaneous conditions on
existence and stability of single-mode solutions, and it
is instructive for understanding to study some special cases.
First, one can explicitly resolve it for the mode $k_0=\pi$ to
derive the following condition on the strength of radiative
coupling:
\begin{equation}\label{10a}
\gamma>\alpha, \ \gamma>g/2.
\end{equation}
For $\gamma<\alpha$ and $\gamma>g/2$ the $\pi$-mode solution
exists, but is unstable. Second, assuming the balance of linear
loss and pumping $g=0$, when all modes within
$k\in[\pi/2,\pi]$ exist, according to (\ref{5}), one can determine
the boundary $k^*$ above which the modes are stable:
\begin{equation}\label{10b}
\sin^2 k^*\approx\frac{\gamma-\alpha}{\gamma-\alpha+\gamma(1+\alpha^2)},
\end{equation}
the approximation here implying $c\rightarrow0$ for the side modes. Note also an agreement with the stability condition for the $\pi$-mode (\ref{10a}), the last stable one as nonlinearity $\alpha$ increases (\ref{10b}).
\begin{figure}[ht!!!]
\begin{center}
(a)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg}
(b)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg2}
(c)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kalpha}
\caption{Stability of single-mode solutions with wave numbers $k$ as the net damping rate $g$ or conservative nonlinearity $\alpha$ vary: linearly stable (I), linearly unstable (II) and absent (III). In region (II) color codes the number of unstable directions for a specific seed wave number, as given by Eq. (\ref{10}), normalized by the system size. At relatively weak nonlinearity (a) $\alpha=0.5$, $\gamma=2$ all three regions are present, while for strong nonlinearity (b) $\alpha=20$, $\gamma=2$ the region of stable single-mode solutions may not appear, in accordance with (\ref{10b}).
In general, increasing nonlinearity destabilizes single-mode solutions (c) $\gamma=1$, $g=0.05$.}\label{fig1}
\end{center}
\end{figure}
To get deeper insight into the mode dynamics, we plot diagrams in the wave number -- parameter space, distinguishing three regions, where modes with respective wave number are linearly stable (I), linearly unstable (II), and do not exist (III). The borders between these regions are obtained according to mode excitation (\ref{5}) and stability (\ref{10}) conditions. In region (II) the color codes the relative number of unstable directions for a specific seed wave number, as given by Eq. (\ref{10}).
Studying these diagrams for the net losses $g$, the simplest to control in experiment, we recover two generic cases, in accordance to Eq.(\ref{10a}), see Fig.\ref{fig1}. For relatively weak conservative nonlineary with respect to radiative coupling ($\gamma>\alpha$) all the three regions (I), (II) and (III) are present (Fig.\ref{fig1}(a)). On decreasing net losses $g$ (increasing pumping), one observes gradual birth of single-mode solutions (transition from (III) to (II)) and stabilization of some about wave number $k=\pi$ (transition from (II) to (I)). In (I) one gets high multistability of single-mode solutions, in addition to possible multi-mode attractors. Mode with $k=\pi$ is the only one to pass from (III) to (I) directly.
Conversely, relatively strong conservative nonlinearity
$\gamma<\alpha$ destabilizes single-mode solutions and only
regions (II) and (III) are observed (Fig.\ref{fig1}(b)).
Note that in the region (II) multi-mode solutions may develop,
but we leave them beyond the scope of the present study.
In general, when single-mode solutions are possible
($\gamma>g/2$), weak nonlinearity allows for a family of stable
plane waves about $k=\pi$, region (I), and multi-stability, while
increasing nonlinearity destabilizes them and evokes region (II),
see Fig.\ref{fig1}(c).
\subsection{Diatomic array}\label{sec2b}
Let us analyze existence and stability of plane wave solutions in
diatomic arrays ($\Delta\neq0$). It is convenient to introduce a
binary elementary cell: $x_n=z_{2n-1}$, $y_n=z_{2n}$ and variables
$a_k^+$,$a_k^-$
\begin{equation}\label{12}
\left\{
\begin{aligned}
x_n&=\frac{1}{\sqrt{N}}\sum\limits_{k}\left(\xi_k^+ a_k^+ +\xi_k^-
a_k^-\right)e^{ikn}\\
y_n&=\frac{1}{\sqrt{N}}\sum\limits_{k}\left(\eta_k^+ a_k^+
+\eta_k^- a_k^-\right)e^{ikn},
\end{aligned}
\right.
\end{equation}
where $k$ is the mode wave number, twice the respective wave number in the original lattice (\ref{2}):
\begin{equation}
\label{12aaa}
k_{\Delta\neq0}=2k_{\Delta=0}.
\end{equation}
\begin{figure*}[ht!!!]
\begin{center}
(a)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg4}
(b)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg5}
(c)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg7}
(d)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kg8}
(e)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kalpha2}
(f)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{n_vs_kalpha10}
\caption{Effects of the frequency mismatch $\Delta$ on stability and existence of single-mode solutions. Notations and color coding inherit Fig.\ref{fig1}. Keeping in mind the relation between the mode wave numbers for homogeneous and diatomic arrays (\ref{12aaa}), compare (a,b,f) to Fig.\ref{fig1}(a,b,c), respectively, and observe qualitatively new features in (c,d). Here
(a)~$\alpha=0.5$, $\Delta=3.5$, $\gamma=2$, (b)~$\alpha=20$,
$\Delta=1$, $\gamma=2$, (c)~$\alpha=1.55$, $\Delta=2$, $\gamma=2$,
(d)~$\alpha=3.1$, $\Delta=2$, $\gamma=2$, (e)~$\gamma=\Delta=1$, $g=0.5$, (f)~$\gamma=\Delta=1$, $g=0.05$.}\label{fig3}
\end{center}
\end{figure*}
Coefficients $\xi_k^{\pm}$, $\eta_k^{\pm}$ are calculated to ensure that transformation
(\ref{12}) is canonical in conservative limit ($g=\gamma=\beta=0$)
and diagonalizes the linear part of the dynamical
equations (\ref{2}):
$$\xi_k^{\pm}=\frac{\xi_k^{\pm
0}}{|\xi_k^{\pm 0}|^2+|\eta_k^{\pm 0}|^2},\quad
\eta_k^{\pm}=\frac{\eta_k^{\pm 0}}{|\xi_k^{\pm 0}|^2+|\eta_k^{\pm
0}|^2},$$
$$\xi_k^{\pm
0}=\frac{(\gamma-i)(1+e^{-ik})}{i\Delta\mp
2\sqrt{(\gamma-i)^2\cos^2\frac{k}{2}-\frac{\Delta^2}{4}}},\quad\eta_k^{\pm
0}=1,\quad k\neq\pi,$$
$$\xi_{\pi}^{+0}=1,\eta_{\pi}^{+0}=0,\xi_{\pi}^{-0}=0,\eta_{\pi}^{-0}=1.$$
Dynamics of modes follow then
\begin{equation}\label{13}
\begin{split}
&\dot a_k^s=-i\omega_k^s
a_k^s-\frac{1+i\alpha}{2N}\times\\\times &\sum\limits_{k_1 k_2
k_3}\sum\limits_{s_1 s_2 s_3}[D_{k_1 k_2 k_3 k}^{s_1 s_2 s_3
s}a_{k_1}^{s_1}a_{k_2}^{s_2}a_{k_3}^{s_3*}\delta(k_1+k_2-k_3-k)].
\end{split}
\end{equation}
Here $s_i$ denotes the signature ``$+$'' or ``$-$'', and
coefficients $D_{k_1 k_2 k_3 k}^{s_1 s_2 s_3 s}$ determine
selective interaction between the modes:
$$D_{k_1 k_2 k_3 k}^{s_1 s_2 s_3
+}=\frac{1}{\xi_k^-\eta_k^+
-\eta_k^-\xi_k^+}(\xi_k^-\eta_{k_1}^{s_1}\eta_{k_2}^{s_2}\eta_{k_3}^{s_3*}-\eta_k^-\xi_{k_1}^{s_1}\xi_{k_2}^{s_2}\xi_{k_3}^{s_3*}),$$
$$D_{k_1 k_2 k_3 k}^{s_1 s_2 s_3 -}=\frac{1}{\xi_k^-\eta_k^+
-\eta_k^-\xi_k^+}(\eta_k^+\xi_{k_1}^{s_1}\xi_{k_2}^{s_2}\xi_{k_3}^{s_3*}-\xi_k^+\eta_{k_1}^{s_1}\eta_{k_2}^{s_2}\eta_{k_3}^{s_3*}),$$
Frequencies $\omega_k^s$ follow dispersion equation
\begin{equation}\label{12a}
\left(-i\omega_k^{\pm}+\frac{g}{2}\right)^2=(\gamma-i)^2\cos^2\frac{k}{2}-\frac{\Delta^2}{4}.
\end{equation}
Eq. (\ref{13}) gives excitation condition for modes $a_k^s$, simultaneously met for both signatures $s$:
\begin{equation}\label{14}
\mbox{Im}(\omega_k^s)>0
\end{equation}
It demonstrates that diatomic system possesses $N$ two-mode manifolds, i.e.
solutions of the form $a_{k_0}^{\pm}\neq 0$, $a_{k\neq
k_0}^{\pm}=0$, rather than single-mode ones as in the homogeneous case. The only exceptions
correspond to $a_{\pi}^+$ and $a_{\pi}^-$ modes.
Excitation threshold can be found letting $\mbox{Im}(\omega_k^s)=0$, which after some algebra gives
\begin{equation}\label{14a}
(g/2)^4+\left[\frac{\Delta^2}{4}+(1-\gamma^2)\cos^2\frac{k}{2}\right](g/2)^2-\gamma^2 \cos^4\frac{k}{2}=0,
\end{equation}
and several corollaries follow. First, for all $\gamma\neq0,
k\neq\pi$ there exists a unique threshold $g=g^*(k)$,
below which the zero state becomes unstable with respect to
$a_{k_0}^{\pm}$. Moreover, one finds that this threshold is
maximized by $k=0$, in agreement to the homogeneous limit
$\Delta=0$, where the $\pi$-mode is the first to get excited
(recall that we currently work with the binary elementary cell
and wave numbers double (\ref{12aaa})). Addressing the
effect of the frequency mismatch, one readily recognizes that the
excitation threshold $g=g^*(k,\Delta)$ is a decreasing function of
$\Delta$, as given by (\ref{14a}). It means that non-zero spatial
inhomogeneity requires to reduce net losses in order to obtain
excitation of modes. Notably, in absence of dissipative coupling
$\gamma=0$ the zero solution will always be stable.
Once excitation condition (\ref{14}) for a particular wave number $k_0$ is attained, there emerges a periodic trajectory in the corresponding two-mode manifold:
\begin{equation}\label{15}
a_{k_0}^{\pm}=\sqrt{N}A_{1,2}e^{-i\Omega t},
\end{equation}
whose amplitudes $A_{1,2}$ get saturated by nonlinear dissipation
in (\ref{13}) and can be numerically calculated (see Appendix). In
the original equations (\ref{2}) one recovers a plane wave
solution
\begin{equation}\label{16} \left\{
\begin{aligned}
x_n=B_1 e^{ik_0 n-i\Omega t}\\
y_n=B_2 e^{ik_0 n-i\Omega t}
\end{aligned}
\right.
\end{equation}
where $B_1=\xi_k^+A_1+\xi_k^-A_2$, $B_2=\eta_k^+A_1+\eta_k^-A_2$.
As before, we study modulational instability of these waves against arbitrary small perturbations in the side modes $a_k^{\pm}$, $k\neq k_0$, and $k_0$ referring to the seed mode. Evolution of perturbations follows from
linearization of Eq. (\ref{13}) in neighborhood of a periodic solution (\ref{15}). It is represented by mutually independent sets of dynamical equations for quadruplets of side modes $a_k^{\pm}$, $a_{k^{'}}^{\pm}$, where
$k=k_0+c$, $k^{'}=k_0-c$ (see Appendix for details).
The seed mode solution of (\ref{13}) is linearly stable if all increments of the side modes are negative and, additionally, stability within the two-mode manifold (\ref{15}) is provided. If an increment in at least one side modes quadruplet is positive or instability within the manifold develops, the seed solution becomes unstable.
Since analytical studies in the general case are extremely complicated, we further resort to a computationally exact numerical analysis (numerical calculation of the amplitudes of pure mode solutions and diagonalization of stability matrices about them).
The resulting typical stability diagrams demonstrate important qualitative distinctions from the spatially homogeneous case, which are illustrated in
Fig.~\ref{fig3}.
First, spatial inhomogeneity decreases the regions, where mode solutions exist, stable (I) or unstable (II), in agreement to (\ref{14a}), compare the thresholds in $g$ for each $k$ in Fig.\ref{fig1}(a) and Fig.\ref{fig3}(a). At the same time, inhomogeneity may improve stability of these solutions at relatively high nonlinearities, leading to emergence of the region (I) within (II), compare Fig.\ref{fig1}(b) and Fig.\ref{fig3}(b). Two principally new features may appear in $(k,g)$ diagrams. For $\gamma>\alpha$, when all the three regions (I)-(III) are present for $\Delta=0$, inhomogeneity can decrease the stability region (I), making it completely detached from $g=0$ axis (Fig.\ref{fig3}(c)). For $\gamma<\alpha$ the stability region (I) may emerge in two disconnected areas within (II), as demonstrated in Fig.\ref{fig3}(d). Turning to $(k,\alpha)$ diagrams, one observes pictures either qualitatively similar to $\Delta=0$ (compare Fig.\ref{fig1}(c) and Fig.\ref{fig3}(e)), or a stabilizing effect of inhomogeneity, leading to emergence of the second stability island (I) within (II) (Fig.\ref{fig3}(e)).
\section{Instability and mode dynamics: from local to global chaos}\label{sec3}
Having obtained stability conditions for plane wave
solutions, we turn to the dynamics, when instability occurs. While
there appears to be a variety of complex dynamical regimes, we
focus on transition from a few to multi-mode oscillations related
to the development of dynamical chaos. This crossover presents a
special interest due to the physically relevant interpretation of
emerging broad spectrum lasing of exciton-polariton condensates.
Evolution from arbitrary initial conditions depends on attractors
of the system and their basin of attraction. If plane
waves about $k=0$ in the diatomic array are linearly stable
(region (I) in Fig.\ref{fig3}), then initial conditions slightly
perturbed from such a mode will converge to it. In this region,
therefore, multistability takes place. As the parameters vary, the
stability interval in wave numbers may decrease until the most
stable mode $k=0$ undergoes modulational instability.
\begin{figure*}[th!!!]
\begin{center}
(a)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{sp1_1}
(b)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{2x3}
(c)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{2x7}
(d)\includegraphics[width=0.9\columnwidth,keepaspectratio,clip]{PLam}
\caption{From stable plane waves to
multi-mode chaos: (a) color coded distribution of mode space
excitations $\ln|a_k^+|^2$ ($|a_k^+|$ time-averaged after a
transient time) as nonlinearity $\alpha$ increases and
modulational instability develops, solid lines mark modulational
instability domain of the mode $k_0=0$; (b) and (c) color coded
spatio-temporal dynamics of $\mbox{Re}(z_n)$ for $\alpha=1$
(somewhat above instability threshold) and $\alpha=2.667$ (deep
into the instability region), respectively; (d) time-averaged mode
participation number $P$, the number of unstable directions $M$ of the mode $k_0=0$, the number
of positive Lyapunov exponents $L$, and the standard deviation of average oscillator frequencies versus nonlinearity $\alpha$ (note its upscaling). Here
$\gamma=0.5$, $g=0.0667$, $\Delta=0.667$, the
system size equals $N=400$. }\label{fig4}
\end{center}
\end{figure*}
To study this transition one needs to characterize the distribution of energy density in mode space. For that it is convenient to calculate participation number, the quantity defined as
\begin{equation}
P=\frac{\left(\sum_k (|a_k^+|^2+|a_k^-|^2)\right)^2}{\sum_k
(|a_k^+|^4+|a_k^-|^4)}
\end{equation}
and estimating the number of effectively excited modes.
Fig.\ref{fig4} demonstrates a characteristic evolution of the mode participation number on
an attractor as it evolves from the stable $k=0$ mode into the instability region, while the on-site nonlinearity $\alpha$ increases.
There we observe that the excitation in the mode space closely follows modulational instability boundary for the side modes (Fig.\ref{fig4}(a)).
In direct space collective oscillations demonstrate the crossover from distorted plain waves (somewhat above instability threshold, Fig.\ref{fig4}(b)) to strongly chaotic patterns (deep in the instability region, Fig.\ref{fig4}(c)). Remarkably, the transition does not appear entirely smooth: in the interval $\alpha\approx 0.6\ldots 1.1$ the excitation is concentrated about the seed mode $k_0=0$, while at $\alpha\approx1.1$ quite a sharp break up towards a broad excitation across the whole instability region is detected.
To quantify the relation between instability and multi-mode dynamics we plot corresponding mode participation number $P$, the number of side modes $M$ with positive increment, the number $L$ of positive Lyapunov exponents $\lambda_i>0$, and standard deviation of average oscillator frequencies $\sigma_\omega$ against nonlinearity strength $\alpha$ in Fig.\ref{fig4}(d). The results suggest that modulational instability of $k_0=0$ mode leads first to quasiperiodic oscillations on a torus of a low effective dimension, evidenced by small participation number in the mode space $P\approx1$ and absence of reliably positive Lyapunov exponents in $0.6<\alpha<0.87$. Then one observes transition to dynamical chaos that may be characterized as ``local'' in the mode space in $0.97<\alpha<1.04$, since $P\approx1$ remains. In direct space this is seen as chaotic modulations of a plain wave solution (Fig.\ref{fig4}(b)). Finally, the transition to ``global'' chaos occurs at $\alpha\approx1.04$, manifested by an abrupt increase of participation number to $P\approx10$. Simultaneously, synchronization of oscillation frequencies in the direct space is broken up. Further, dimensionality of chaos and the number of effectively excited modes grow in parallel (Fig.\ref{fig4}(d)), oscillations in the direct space become irregular (Fig.\ref{fig4}(c)).
\section{Conclusion}
We have studied collective dynamics in oscillatory arrays with
diatomic spatial inhomogeneity, radiative and Josephson coupling,
conservative and dissipative nonlinearities. We developed the mode
analysis for such systems to derive mode excitation and
modulation instability conditions. It was shown that the most
persistent and stable mode has the wave number $k_0=0$ in the
binary node basis (corresponding to the $k_0=\pi$ mode in the
limit of a homogeneous array in a singe-node basis). Generically,
the number of stable modes in vicinity of $k_0=0$ increases as the
net local damping (balance of pumping and dissipation) decreases,
giving rise to high-order multistability. Increasing conservative
nonlinearity supports the development of modulation instability of
the modes. The frequency mismatch between neighbor sites imposes
more stringent conditions on mode existence, as compared to the
homogeneous array. At the same time, inhomogeneity can
stabilize the modes at high nonlinearity,
which in the homogeneous case would be unstable. We also
identified the cases when the region of stable mode solutions
consists of two disjoint islands.
We explored the dynamics beyond the mode stability region and demonstrated several crossovers as the nonlinearity and degree of seed mode instability grow. First, the dynamics remains regular and quasiperiodic, localized in the mode space. Second, the transition to low-dimensional chaos, also localized in the mode space, is observed. Above the last threshold the mode participation number increases abruptly, oscillations span over most of the modulation instability wave interval, producing multi-dimensional chaos and the break up of synchronization.
Beside their fundamental interest for nonlinear dynamics, these
results are applicable to understanding and designing
collective behavior of active structured physical media with
spatial inhomogeneity, such as exciton-polaritons condensation
arrays, our primary system of interest, or actively coupled
waveguides.
A.T., O.K., and M.I. acknowledge support of Russian Foundation for Basic Research grant No. 13-02-97028. M.I. also acknowledges support of Russian Ministry of Science and Education (research of Section III supported by Research Assignment No. 1.115.2014/K). We thank Yu. Rubo, S. Flach, and N. Berloff for insightful discussions.
|
\section{Introduction}\label{intro}
\subsection{Confined atoms in $d=3$ dimensions}
\noindent In order to fix ideas we considered first a simple model \cite{hall1981} for a soft confined atom obeying
a Schr\"odinger equation in $d=3$ dimensions of the form $(-\Delta +V)\psi = E\psi,$
where $V(r)$ is an attractive central potential with Coulomb and confining terms. If we assume a wave function of the form
$\psi(r) = Y_{\ell}^{m}(\theta,\phi)r^{\ell}\exp(-g(r)),$ then we find the radial eigenequation implies
\begin{equation}
(E-V(r))r = rg''(r)+2(l+1)g'(r) -r(g'(r))^2.
\end{equation}
If we now choose $g(r) = \frac{1}{2}(vr+\omega r^2)$, $v>0$, $\omega >0,$ we obtain the
following family of exact solutions
\begin{equation}\label{modelsol}
V(r)=v\left(-\frac{\ell+1}{r} + \omega r\right) + \omega^2 r^2,\quad
E = (3+2\ell)\omega - v^2/4,\quad l = 0,1,2,\dots
\end{equation}
The lowest radial excitations of the familiar Coulomb and oscillator problems are
recovered from the special cases $\omega =0$ or $v = 0.$ Such specific exact solutions
allow for analytical reasoning and explorations. In addition, the explicit results provide relevant test
problems for the complementary approaches that must be used to complete the solution space.
The question of the existence of exact solutions and the methods for finding them are
therefore an important part of the overall task. As in the choice of the function $g(r)$ in
the simple illustration above, it is
often the case in this context that exactness has something to do with polynomials. Thus part of the paper involves the issue of
when an ordinary differential equation admits polynomial solutions. As we shall see, the `asymptotic iteration method' \cite{aim} plays a role both in the construction of exact solutions and also in finding approximations for arbitrary values of the problem parameters.
\subsection{Formulation of the problem in $d$ dimensions}
\noindent The Schr\"odinger equation in $d>1$ dimensions,
in atomic units $\hbar=2\mu=1$, with a spherically symmetric
potential $V(r)$ can be written as \cite{dong2011}
\begin{equation}\label{Sch_eq}
H\psi\equiv \left[-\Delta_d +V(r)\right]\psi({\bf r})=E\psi(\bf{r}),
\end{equation}
where $\Delta_d$ is the $d$-dimensional Laplacian operator and ${\bf r}=(x_1,x_2,\dots,x_d)$, $r^2=\|{\bf r}\|^2=\sum_{i=1}^d x_i^2$. The quantum wave function $\psi$ is an element of the Hilbert space $L^2(\Re^d).$ The principal class of spherically symmetric confining potentials we
shall consider has the form
\begin{equation}\label{gn_pot}
V(r)=\dfrac{a}{r+\beta}+c\,r+b^2\,r^2, \quad \beta> 0, ~b>0.
\end{equation}
Thus $V(r)$ is continuous and $V(r)\rightarrow \infty$. Consequently, by Theorem XIII.67 of Reed-Simon-IV \cite{simon},
we know that $H$ has
purely discrete eigenvalues and a complete set of eigenfunctions. Meanwhile, for $d \ge 3$, by Theorem XIII.69 of the same reference, we have a similar conclusion if we admit the Coulomb singularity in $V(r)$ by allowing $\beta = 0.$ For the case of hard confinement, $r\le R < \infty$
with Dirichlet boundary conditions at $r = R$,
and $ \beta >0,$ so that $V(r)$ is continuous, we know from Theorem 23.56 of Ref.\cite{gs} that again $H$ has a purely discrete spectrum.
These general results cover the cases we consider in this paper.
A comparable class of potentials has been carefully analysed in Refs.\cite{bulla,alb}. In order to express (\ref{Sch_eq}) in terms of
$d$-dimensional spherical coordinates $(r, \theta_1, \theta_2,
\dots, \theta_{d-1})$, we separate variables using
\begin{equation}\label{gs_Sch_eq}
\psi({\bf r})=r^{-(d-1)/2}\,u(r)\, Y_{\ell_1,\dots,\ell_{d-1}}(\theta_1\dots\theta_{d-1}),
\end{equation}
where $Y_{\ell_1,\dots,\ell_{d-1}}(\theta_1\dots\theta_{d-1})$ is a
normalized spherical harmonic \cite{atkin} with characteristic value $\ell(\ell+d-2),$ and $\ell=\ell_1=0, 1, 2, \dots$ (the angular quantum numbers). One obtains the
radial Schr\"odinger equation as
\begin{equation}\label{gs_Sch_eq1}
\left[-{d^2\over dr^2}+{(k-1)(k-3)\over
4r^2}+V(r)-E\right] u_{n\ell}^{(d)}(r)=0, \quad\quad
\int_0^\infty \left\{u_{n\ell}^{(d)}(r)\right\}^2dr=1, ~~~
u_{n\ell}^{(d)}(0)=0,
\end{equation}
where $k=d+2\ell$. Since the potential $V(r)$ is less singular than the centrifugal term,
$$u(r)\sim A\,r^{(k-1)/2},\qquad r\rightarrow 0,\qquad \mbox{where $A$ is a constant}.
$$
We note that the Hamiltonian and boundary
conditions of \eqref{gs_Sch_eq1} are invariant under the transformation
$$(d, \ell)\rightarrow (d\mp2, \ell\pm 1), $$
thus, given any solution for fixed $d$ and $\ell$, we can immediately
generate others for different values of $d$ and $\ell$. Further, the
energy is unchanged if $k=d+2\ell$ and the number of nodes $n$ is
constant. Repeated application of this transformation produces a
large collection of states; this has been discussed, for example, in Ref. \cite{doren1986}.
\vskip0.1true in
\noindent In the present work we study the exact and approximate solutions of the Schr\"odinger eigenproblem generated by a confined soft-core Coulomb potential in $d$-dimensions, where $d>1.$ As we have discussed above, for the cases we consider, the spectrum of this problem is discrete, all
eigenvalues are real and simple, and they can be arranged in an
increasing sequence $\lambda_0<\lambda_1<\dots\rightarrow \infty$. The paper is organized as follows. In section \ref{sec2}, we set up the Schr\"odinger equation for the potential \eqref{gn_pot} and discuss the correspondence second-order differential equation. In section \ref{method}, we present our method of solution that relies on the analysis of polynomial solutions of the differential equation
\begin{equation}\label{DE_general_sitting}
\left(\sum_{i=0}^{k}a_{k,i}\, r^{k-i} \right)y''+\left(\sum_{i=0}^{k-1}a_{k-1,i}\, r^{k-1-i} \right)y'-\left(\sum_{i=0}^{k-2}\tau_{k-2,i}\, r^{k-2-i} \right)y=0,\quad k\geq 2
\end{equation}
and different variants of this general differential-equation class. We discuss in particular necessary and sufficient conditions on the equation parameters for it to have polynomial solutions. A brief review of the asymptotic iteration method (AIM) is presented in section \ref{AIM}. In section \ref{app1}, the exact and approximate solutions for the problem are discussed, based on the results of section \ref{sec2}; and approximate solutions are found for arbitrary potential parameters $a,b,c$ and $\beta$ by an application of AIM . An analysis of the corresponding exact and approximate solutions for the pure confined Coulomb case $\beta = 0$ is presented in section \ref{app2}. The `hard confinement' case, that is to say when the same system confined to the interior of an impenetrable spherical box of radius $R,$ is discussed in section \ref{spec}. In each of these sections, the results obtained are of two types: exact analytic results that are valid when certain
parametric constraints are satisfied, and accurate numerical values for arbitrary sets of potential parameters.
\vskip0.1true in
\section{Setting up the differential equation}\label{sec2}
\noindent In this section, we consider the $d$-dimensional radial Schr\"odinger equation for $d>1$ :
\begin{equation}\label{Sch}
\left[-{d^2\over dr^2}+{(k-1)(k-3)\over 4r^2}+{a\over r+\beta}+c\,r+b^2r^2\right]u_{nl}^d(r)=E_{nl}^d u_{nl}^d(r),\quad \beta>0,\quad a\in(-\infty,\infty),\quad 0<r<\infty, \quad u(0) = 0.
\end{equation}
We note first that the differential equation (\ref{Sch}) has one regular singular
point at $r = 0$ with exponents given by the roots of the indicial equation
\begin{equation}\label{indicial}
s(s-1)-{1\over 4}(k-1)(k-3)=0,
\end{equation}
and an irregular singular point at $r = \infty$. For large $r$, the differential equation \eqref{Sch} assumes the asymptotic form
\begin{equation}\label{Sch_asy}
\left[-{d^2\over dr^2}+c\,r+b^2r^2\right]u_{nl}^d(r)\approx 0
\end{equation}
with an asymptotic solution
\begin{equation}\label{asy_sol}
u_{nl}^d(r)\approx \exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right).
\end{equation}
The roots $s$ of Eq.(\ref{indicial}), namely,
\begin{align*}
s_1={1\over 2}(3-k),\quad s_2={1\over 2}(k-1).
\end{align*}
determine the behaviour of $u_{nl}^d(r)$ as $r$ approaches $0$, only $s\geq 1/2$ is acceptable, since only in this case is the mean value of the kinetic energy finite \cite{landau}.
Thus, the exact solution of (\ref{Sch}) will assume the form
\begin{equation}\label{gen_sol}
u_{nl}^d(r)=r^{(k-1)/2}\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)~f_n(r),\quad k=d+3l,
\end{equation}
where we note that $u_{nl}^d(r)\sim r^{(k-1)/2}$ as $r\rightarrow 0$. On insertion of this ansatz wave function into (\ref{Sch}), we obtain the differential equation for $f_n(r)$ as
\begin{align}\label{secondorderde}
-4b^2\,r&\,(r+\beta)\,f_n''(r)+\left(8b^3\,r^3+4b\,(c+2b^2\,\beta)\,r^2+4b\,(b+c\beta-b\,k)\,r+4b^2\beta(1-k)\right)f_n'(r)
\notag\\
&+\left(\left[4b^2(b\,k-E)-c^2\right]\,r^2+(4 a b^2 - \beta c^2 - 4 b^2 \beta\,E + 2 b c (k-1) + 4 b^3 \beta k)\,r+2b\beta\,c(k-1)
\right)f_n(r)=0.
\end{align}
In the next section, we study the polynomial solutions of this differential equation which itself lies within a larger class of differential equations given by
\begin{align}\label{gen_DE}
(a_{4,2}r^2&+a_{4,3}r)y^{\prime \prime}+(a_{3,0}r^3+a_{3,1}r^2+a_{3,2}r+a_{3,3})y'-(\tau_{2,0}r^2+\tau_{2,1} r+\tau_{2,2})y=0,
\end{align}
where $\tau_{2,0},\tau_{2,1},\tau_{2,2}$ and $a_{i,j}$ are real constants for $i=3,4$ and $j=0,1,2,3$.
\vskip0.1true in
\section{The method of solution}\label{method}
\noindent The necessary condition (\cite{h2010}, Theorem 6) for polynomial solutions $y(r)=\sum_{k=0}^n c_k r^k$ of the second-order linear differential equation \eqref{gen_DE} is
\begin{equation}\label{poly_cond}
\tau_{2,0}=n\,a_{3,0},\quad n=0,1,2,\dots,
\end{equation}
provided $a_{3,0}^2+\tau_{2,0}^2\neq 0$. The polynomial coefficients $c_n$ then satisfy the four-term recurrence relations
\begin{align}\label{rec_rel}
((n-2)\, a_{3,0}-\tau_{2,0})\,c_{n-2}&+((n-1)\,a_{3,1}-\tau_{2,1})\,c_{n-1}+(n(n-1)\,a_{4,2}+ n\,a_{3,2}-\tau_{2,2})\,c_n\notag\\
&+(n(n+1)\,a_{4,3}+(n+1)\, a_{3,3} )\,c_{n+1}=0,\qquad c_{-2}=c_{-1}=0.
\end{align}
The proof of \eqref{rec_rel} follows from an application of the Frobenius method.
We note that the recurrence relations \eqref{rec_rel} can be written as a system of linear equations in the unknown coefficients $c_i$, $i=0,\dots, n$ given by
\begin{align}\label{lin_sys}
\begin{bmatrix}
\gamma_0&\delta_0&0&0&0&0 &\cdots&0&0 \\
\beta_1&\gamma_1&\delta_1&0&0&0&\cdots&0&0 \\
\alpha_2&\beta_2&\gamma_2&\delta_2&0&0&\cdots &0&0 \\
0&\alpha_3&\beta_3&\gamma_3&\delta_3&0&\cdots&0 &0 \\
\vdots & \vdots & \ddots & \ddots&\ddots&\ddots&\cdots&\vdots&\vdots\\
0&0&0&0&0&0 &\cdots&\beta_n&\gamma_n \\
\end{bmatrix}\begin{bmatrix}
c_0\\
c_1\\
c_2 \\
c_3 \\
\vdots\\
c_n \\
\end{bmatrix}=\begin{bmatrix}
0\\
0 \\
0 \\0 \\
\vdots\\
0 \\
\end{bmatrix}
\end{align}
where
\begin{align}\label{entries}
\gamma_n&=n(n-1)a_{4,2}+ na_{3,2}-\tau_{2,2},~\delta_n=n(n+1)a_{4,3}+(n+1) a_{3,3}, ~\beta_n= (n-1)a_{3,1}-\tau_{2,1},~
\alpha_n=(n-2) a_{3,0}-\tau_{2,0}.
\end{align}
Thus, for zero-degree polynomials $c_0\neq 0$ and $c_n=0,~n\geq 1$, we must have $\gamma_0=\beta_1=\alpha_2=0$, thus, in addition to the necessary condition $\tau_{2,0}=0$, the following two conditions become sufficient
\begin{equation}\label{poly0}
\tau_{2,2}=0,~\tau_{2,1}=0.
\end{equation}
For the first-degree polynomial solution, $ c_1\neq 0,$ and $c_n=0,~n\geq 2$, we must have
$\gamma_0\,c_0+\delta_0\,c_1=0,\,\beta_1\,c_0+\gamma_1\,c_1=0,\,\alpha_2\,c_0+\beta_2\,c_2=0$ and $\alpha_3\,c_1=0$, thus, in addition to the necessary condition $\alpha_3=0$ or
\begin{equation}\label{poly1_cond}
\tau_{2,0}=a_{3,0},
\end{equation}
it is also required that the following two $2\times 2$-determinants simultaneously vanish
\begin{equation}\label{cond_poly1}
\left|\begin{array}{ccc}
-\tau_{2,2}&a_{3,3}\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}
\end{array}\right|=0,\quad
\mbox{and}\quad
\left|\begin{array}{ccc}
-\tau_{2,2}&a_{3,3}\\
-a_{3,0}& a_{3,1} -\tau_{2,1}
\end{array}\right|=0.
\end{equation}
For the second-degree polynomial solution, $c_2\neq0$ and $c_n=0$ for $n\geq 3$, it is necessary that $\gamma_0\,c_0+\delta_0\,c_1+\mu_0\,c_2=0,~\beta_1\,c_0+\gamma_1\,c_1+\delta_1\,c_2=0,
\alpha_2\,c_0+\beta_2\,c_1+\gamma_2\,c_2=0,\, \alpha_3\,c_1+\beta_3\,c_2=0,$ and $\alpha_4\,c_3=0$, from which we have the necessary condition
\begin{equation}\label{poly2_cond}
\tau_{2,0}=2\,a_{3,0}
\end{equation}
along with the vanishing of the two $3\times 3$-determinants
\begin{equation}\label{cond_poly2}
\left|\begin{array}{ccc}
-\tau_{2,2}&a_{3,3}&0\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}\\
-2a_{3,0}&a_{3,1} -\tau_{2,1}&2a_{4,2}+2a_{3,2}-\tau_{2,2}
\end{array}\right|=0\quad\mbox{and}\quad
\left|\begin{array}{ccc}
-\tau_{2,2}&a_{3,3}&2a_{4,4}\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}\\
0&-a_{3,0}&2a_{3,1}-\tau_{2,1}
\end{array}\right|=0,
\end{equation}
For the third-degree polynomial solution, $c_3\neq0$ and $c_n=0$ for $n\geq 4$, we then have the necessary condition
\begin{equation}\label{cond_poly3}
\tau_{2,0}=3\,a_{3,0}
\end{equation}
along with the vanishing of the two $4\times 4$-determinants,
\begin{equation}\label{cond_poly3_1}
\left|\begin{array}{cccc}
-\tau_{2,2}&a_{3,3}&0&0\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}&0\\
-3a_{3,0}&a_{3,1} -\tau_{2,1}&2a_{4,2}+2a_{3,2}-\tau_{2,2}&3a_{3,3}+6a_{4,3}\\
0&-2a_{3,0} &2 a_{3,1} - \tau_{2,1}&3 a_{3,2} + 6 a_{4,2} - \tau_{2,2}
\end{array}\right|=0
\end{equation}
and
\begin{equation}\label{cond_poly3_2}
\left|\begin{array}{cccc}
-\tau_{2,2}&a_{3,3}&0&0\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}&0\\
-3a_{3,0}&a_{3,1} -\tau_{2,1}&2a_{4,2}+2a_{3,2}-\tau_{2,2}&3a_{3,3}+6a_{4,3}\\
0&0&-a_{3,0} &3a_{3,1} - \tau_{2,1}
\end{array}\right|=0
\end{equation}
For the fourth-degree polynomial solution ($n=4$), $c_4\neq 0$ and $c_n=0$ for $n\geq 5$, we then have the necessary condition
\begin{equation}\label{cond_poly3}
\tau_{2,0}=4\,a_{3,0}
\end{equation}
along with the vanishing of the two $5\times 5$-determinants,
\begin{equation}\label{cond_poly3_1}
\left|\begin{array}{ccccc}
-\tau_{2,2}&a_{3,3}&0&0&0\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}&0&0\\
-4a_{3,0}&a_{3,1} -\tau_{2,1}&2a_{4,2}+2a_{3,2}-\tau_{2,2}&3a_{3,3}+6a_{4,3}&0\\
0&-3a_{3,0} &2 a_{3,1} - \tau_{2,1}&3 a_{3,2} + 6 a_{4,2} - \tau_{2,2}&4a_{3,3}+12a_{4,3}\\
0&0&-2 a_{3,0}&3 a_{3,1}- \tau_{2,1}& 4 a_{3,2} + 12 a_{4,2} - \tau_{2,2}
\end{array}\right|=0
\end{equation}
and
\begin{equation}\label{cond_poly3_2}
\left|\begin{array}{ccccc}
-\tau_{2,2}&a_{3,3}&2a_{4,4}&0&0\\
-\tau_{2,1}& a_{3,2} -\tau_{2,2}&2a_{4,3}+ 2a_{3,3}&0&0\\
-4a_{3,0}&a_{3,1} -\tau_{2,1}&2a_{4,2}+2a_{3,2}-\tau_{2,2}&3a_{3,3}+6a_{4,3}&0\\
0&-3a_{3,0} &2 a_{3,1} - \tau_{2,1}&3 a_{3,2} + 6 a_{4,2} - \tau_{2,2}&4a_{3,3}+12a_{4,3}\\
0&0&0&- a_{3,0} &4 a_{3,1} - \tau_{2,1}
\end{array}\right|=0
\end{equation}
Similar expressions for higher-order polynomial solutions can be easily obtained. The vanishing of these determinants can be regarded as the sufficient conditions under which the coefficients $\tau_{2,1}$ and $\tau_{2,2}$ of Eq. \eqref{secondorderde} can be expressed in terms of the other parameters.
\vskip0.1true in
\section{The asymptotic iteration method and some related results}\label{AIM}
\noindent The asymptotic iteration method (AIM) is an iterative algorithm originally
introduced \cite{aim} to investigate the analytic and approximate solutions of the differential
equation
\begin{equation}\label{AIM_Eq}
y''=\lambda_0(r) y'+s_0(r) y,\quad\quad ({}^\prime={d\over dr})
\end{equation}
where $\lambda_0(r)$ and $s_0(r)$ are $C^{\infty}-$differentiable
functions. A key feature of this method is to note the invariant
structure of the right-hand side of (\ref{AIM_Eq}) under further
differentiation. Indeed, if we differentiate \eqref{AIM_Eq} with
respect to $r$, we obtain
\begin{equation}\label{first_der}
y^{\prime\prime\prime}=\lambda_1(r)\, y^\prime+s_1(r)\, y
\end{equation}
where $\lambda_1= \lambda_0^\prime+s_0+\lambda_0^2$ and
$s_1=s_0^\prime+s_0\lambda_0.$ Further differentiation of
equation \eqref{first_der}, we obtain
\begin{equation}\label{second_der}
y^{(4)}=\lambda_2(r)\, y^\prime+s_2(r)\, y
\end{equation}
where $\lambda_2= \lambda_1^\prime+s_1+\lambda_0\lambda_1$ and
$s_2=s_1^\prime+s_0\lambda_1.$ Thus, for $(n+1)^{th}$ and
$(n+2)^{th}$ derivative of (\ref{AIM_Eq}), $n=1,2,\dots$, we have
\begin{equation}\label{the_nth_der}
y^{(n+1)}=\lambda_{n-1}(r)\,y^\prime+s_{n-1}(r)\, y
\end{equation}
and
\begin{equation}\label{the_np1th_der}
y^{(n+2)}=\lambda_{n}(r)\,y^\prime+s_{n}(r)\,y
\end{equation}
respectively, where
\begin{equation}\label{AIM_seq}
\lambda_{n}=
\lambda_{n-1}^\prime+s_{n-1}+\lambda_0\lambda_{n-1}\hbox{ ~~and~~
} s_{n}=s_{n-1}^\prime+s_0\lambda_{n-1}.
\end{equation}
From \eqref{the_nth_der} and \eqref{the_np1th_der}, we have
\begin{equation}\label{poly_Co}
\lambda_n y^{(n+1)}- \lambda_{n-1}y^{(n+2)} = \delta_ny {\rm
~~~where~~~}\delta_n=\lambda_n s_{n-1}-\lambda_{n-1}s_n.
\end{equation}
Clearly, from \eqref{poly_Co} if $y$, the solution of
(\ref{AIM_Eq}), is a polynomial of degree $n$, then $\delta_n\equiv
0$. Further, if $\delta_n=0$, then $\delta_{n'}=0$ for all $n'\geq
n$. In an earlier paper \cite{aim}, we proved the principal theorem
of AIM, namely
\vskip0.1in
\begin{theorem} Given $\lambda_0$ and $s_0$ in
$C^{\infty}(a,b),$ the differential equation (\ref{AIM_Eq}) has the
general solution
\begin{equation}\label{AIM_solution}
y(r)= \exp\left(-\int\limits^{r}{s_{n-1}(t)\over \lambda_{n-1}(t)} dt\right) \left[C_2
+C_1\int\limits^{r}\exp\left(\int\limits^{t}(\lambda_0(\tau) +
2{s_{n-1}\over \lambda_{n-1}}(\tau)) d\tau \right)dt\right]
\end{equation}
if for some $n>0$
\begin{equation}\label{tm_cond}
\delta_n=\lambda_n s_{n-1}-\lambda_{n-1}s_n=0.
\end{equation}
where $\lambda_n$ and $s_n$ are given by \eqref{AIM_seq}.
\end{theorem}
\vskip0.1true in
\noindent Recently, it has been shown \cite{aim1} that the termination condition \eqref{tm_cond} is necessary and sufficient for the differential
equation \eqref{AIM_Eq} to have polynomial-type solutions of degree at most $n$, as we may conclude from Eq.\eqref{poly_Co}. The application of AIM to a number of problems has been outlined in many publications. The applicability of the method is not restricted to a particular class of differentiable functions (e.g. polynomials or rational functions), rather, it can accommodate any type of differentiable function. The fast convergence of the iterative scheme depend on a suitable choice for the starting values of $r=r_0$ and the correct asymptotic solutions near the boundaries \cite{saad2008}.
\section{Exact and approximate solutions for the soft-confined softcore Coulomb potential}\label{app1}
\noindent Comparing equation \eqref{secondorderde} with \eqref{gen_DE} and using parameters given by
\begin{align}\label{parameters}
a_{4,2}&=-4b^2,\qquad a_{4,3}=-4b^2\beta, \notag\\
a_{3,0}&=8b^3,\qquad a_{3,1}=4b\,(c+2b^2\,\beta),\qquad a_{3,2}=4b\,(b+c\beta-b\,k),\qquad a_{3,3}=4b^2\beta(1-k),\notag\\
\tau_{2,0}&=c^2+4b^2(E_{nl}^d-b\,k),\qquad \tau_{2,1}=-4 a b^2 + \beta c^2 + 4 b^2 \beta\,E_{nl}^d - 2 b c (k-1) - 4 b^3 \beta k,\qquad \tau_{2,2}=-2b\beta\,c(k-1),
\end{align}
the exact solution of \eqref{Sch} assumes the following form
\begin{equation}\label{gen_sol_exp}
u_{n\ell}^d(r)=r^{(k-1)/2}\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)~\sum_{i=0}^{n'} {C}_i\,r^i,\quad k=d+3l,
\end{equation}
where $f_{n'}(r)=\sum_{i=0}^{n'} {C}_i\,r^i$ and $n$ counts the number of zeros of $f_{n'}(r)=0$, hence the number of nodes in the wave function solution. The coefficients $ {C}_i$ can be easily evaluated using the four-term recurrence relations \eqref{rec_rel},
\begin{align}\label{rec_relation}
4 b^2 (i-2 - n') {C}_{i-2} &+ (2 a b + c ( k -3+ 2 i) + 4 b^2 \beta ( i-1 - n')) {C}_{i-1} +
(\beta c (k-1 + 2 i)-2 b i (k-2 + i)) {C}_i \notag\\
&-
2 b \beta (i+1) (k-1 + i) {C}_{i+1}=0,\quad {C}_{-1}=0,~~ {C}_0=1,~~ {C}_1=c/(2b),~~i\geq 2,
\end{align}
using the necessary condition
\begin{equation}\label{gen_poly_cond}
E_{n\ell}^d=b\,(2n'+k)-\dfrac{c^2}{4b^2},\quad n'=0,1,2,\dots.
\end{equation}
The potential parameters $a,b,c$ and $\beta$ satisfy
sufficient conditions according to the following scenarios: for a zero-degree polynomial solution, $n'=0$, if $f_0(r)=1$, the ground-state solution of equation \eqref{Sch} is given by
\begin{equation}\label{zero_poly_sol}
u_{0\ell}^d(r)=r^{(k-1)/2}\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right),
\end{equation}
with ground-state eigenenergy
\begin{equation}\label{ground_state_energy}
E_{0\ell}^d=b\,k-\dfrac{c^2}{4b^2}\quad\mbox{subject to the parameter conditions}\quad c=\dfrac{2ab}{1-k}\quad\mbox{and}\quad \beta=0.
\end{equation}
In the next section, we shall focus on the case of $\beta=0$ which corresponds to the Coulomb potential perturbed by an added polynomial in $r$. In the rest of this section, we shall assume $\beta>0.$ For a first-degree polynomial solution, $n'=1$,
\begin{equation}\label{first_poly_sol}
f_0(r)=1+\dfrac{c}{2b}\,r,
\end{equation}
and the exact solution wave function of equation \eqref{Sch} reads
\begin{equation}\label{wf1_gen_sol}
u_{0\ell}^d(r)=r^{(k-1)/2}\left(1+\dfrac{c}{2b}\,r\right)\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right),\qquad E_{0\ell}^d=b\,(k+2)-\dfrac{c^2}{4b^2},
\end{equation}
subject to the following two conditions related the potential parameters
\begin{equation}\label{cond1and2}
4a\,b^2+\beta\,(c^2(k+1)-8b^3)=0\quad\mbox{and}\quad 2abc+c^2(k+1)-8b^3=0.\
\end{equation}
Since, by assumption $b>0$ and $\beta>0$, $c=2b/\beta >0$ and the polynomial solution $f_i(r)=f_0(r)$ has no roots, in which case $E_{0\ell}^d$ represent a ground-state solution of the Schr\"odinger equation \eqref{Sch} subject to the parameters $a$, $b$ and $c$ satisfying the conditions given by \eqref{cond1and2}. In summary, the exact solutions of Schr\"odinger's equation
\begin{equation}\label{Sch_Exact_1}
\left[-{d^2\over dr^2}+\dfrac{(k-1)(k-3)}{4r^2}+{2b\beta^2-(k+1)\over \beta(r+\beta)}+\dfrac{2b}{\beta}\,r+b^2r^2\right]u_{0\ell}^d(r)=E_{0\ell}^d u_{0\ell}^d(r),\quad \beta>0,\quad 0<r<\infty.
\end{equation}
is explicitly given by
\begin{equation}\label{case1_gen_sol}
u_{0\ell}^d(r)=r^{(k-1)/2}\left(1+\dfrac{r}{\beta}\right)\exp\left(-\dfrac{r}{\beta}-\dfrac{b}{2}\,r^2\right),\qquad E_{0l}^d=b\,(k+2)-\dfrac{1}{\beta^2},
\end{equation}
In Figure \ref{Fig1}, we display the un-normalized
ground-state solution using \eqref{case1_gen_sol} for $b=\beta=1$ and different values of $k$.
\begin{figure}[!h]
\centering
\includegraphics[scale=1.2]{wf1.eps}
\caption{Un-normalized ground state wave functions as given by \eqref{wf1_gen_sol} for specific values of $b=\beta=1$ and different values of $k$.}\label{Fig1}
\end{figure}
For the rest of the spectrum we use the asymptotic iteration method as described in section \ref{AIM}, starting with
\begin{equation}\label{aim_gen_sol}
u_{nl}^d(r)=r^{(k-1)/2}\left(1+\dfrac{c}{2b}\,r\right)\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)\,g(r)
\end{equation}
where $g(r)=1$ corresponds to the exact solution \eqref{wf1_gen_sol}, equation \eqref{Sch_Exact_1} yields the second-order differential for $g(r)$ as
\begin{equation}\label{dE_g}
g''(r)=\left(\dfrac{2}{\beta} + \dfrac{1 - k}{r} + 2 b r -\dfrac{2}{\beta + r}\right)g'(r)+\left(2b-E+b\,k-\dfrac{1}{\beta^2}\right)g(r).
\end{equation}
Hence, we may initiate AIM with
\begin{equation}\label{aim_start}
\lambda_0(r)=\dfrac{2}{\beta} + \dfrac{1 - k}{r} + 2 b r -\dfrac{2}{\beta + r}\qquad\mbox{and}\qquad
s_0(r)= 2b-E+b\,k-\dfrac{1}{\beta^2}
\end{equation}
The question is then to find the initial value $r_0$ that stabilizes the computation of the termination-condition roots \eqref{tm_cond}. To this end, we take the highest of the absolute values among all the roots of
$$V(r)-E_{0\ell}^d= \dfrac{(k-1)(k-3)}{4r^2}+{2b\beta^2-(k+1)\over \beta(r+\beta)}+\dfrac{2b}{\beta}\,r+b^2r^2 - \left(b\,(k+2)-\dfrac{1}{\beta^2}\right)=0$$
which yields $r_0\sim 3$, henceforth we shall fix $r_0$ at $r_0 = 3$ for all of our numerical computations. In Table \ref{table:tab1}, we report our results from AIM for first 12 decimal places. The eigenvalue reported in Table \ref{table:tab1} were computed using Maple version 16 running on an IBM architecture personal computer and we have chosen a high-precision environment. In order to accelerate our computation we have written our own code for a root-finding algorithm instead of using the default procedure {\tt Solve} of \emph{Maple 16}. The results of AIM may be obtained to any degree of precision, although we have reported our results to only the first twelve decimal places,
\begin{table}[!h] \caption{The first few Eigenenergies $E_{n0}^{d=4}$ of Schr\"odinger equation \eqref{Sch_Exact_1}. The initial value utilize AIM is $r_0=3$. The subscript $N$ refers to the number of iteration used by AIM.\\ }
\centering
\begin{tabular}{||c |p{2.2in}||c| p{2.2in}|}
\hline
$n$&$E_{n0}^{d=4}$&$n$&$E_{n0}^{d=5}$\\ \hline
$0$&$~5.000~000~000~000_{N=3,Exact}$&$0$&$~6.000~000~000~000_{N=3,Exact}$\\ \hline
$1$&$10.223~655~148~231_{N=90}$&$1$&$11.139~009~555~512_{N=81}$\\ \hline
$2$&$15.140~755~138~866_{N=91}$&$2$&$16.025~939~658~710_{N=83}$\\ \hline
$3$&$19.899~975~543~589_{N=92}$&$3$&$20.771~696~017~356_{N=85}$\\ \hline
$4$&$24.559~997~330~221_{N=92}$&$4$&$25.425~173~414~020_{N=110}$\\ \hline
$5$&$29.150~691~578~737_{N=118}$&$4$&$30.012~690~909~013_{N=109}$\\
\hline
\end{tabular}
\label{table:tab1}
\end{table}
\vskip0.1true in
\noindent For a second-degree polynomial solution, $n'=2$, of \eqref{secondorderde}, we have
\begin{equation}\label{second_poly_sol}
f_i(r)=1+\dfrac{c}{2b}\,r+\dfrac{4ab^2+\beta((k+1)c^2-16b^3)}{8\,b^2\beta\, k}\, r^2,
\end{equation}
and the exact solution of equation \eqref{Sch} reads
\begin{equation}\label{wf2_gen_sol}
u_{i\ell}^d(r)=r^{(k-1)/2}\left(1+\dfrac{c}{2b}\,r+\dfrac{4ab^2+\beta((k+1)c^2-16b^3)}{8\,b^2\beta\, k}\, r^2\right)\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right),\qquad E_{i\ell}^d=b\,(k+4)-\dfrac{c^2}{4b^2},
\end{equation}
where $i$ counts the number of roots of the polynomial solution \eqref{second_poly_sol} subject to the simultaneous conditions relating the parameters $a,~b,~c$ and $\beta$,
\begin{align}\label{second_poly_cond}
4 a b^2 (-4 b k + 3 \beta c (1 + k)) + \beta^2 c (c^2 (1 + k) (3 + k) - 16 b^3 (3 + 2 k))&=0,\notag\\
8 a^2 b^3 + 2 a b (-16 b^3 \beta + \beta c^2 (1 + k) + 2 b c (3 + k)) +
\beta c (c^2 (1 + k) (3 + k) - 16 b^3 (3 + 2 k))&=0.
\end{align}
In particular, for
$$a=4 b \beta-\frac{\beta c^2}{4 b^2}+\left(\frac{c }{b}-\frac{2 }{\beta}-\frac{\beta c^2}{4 b^2}\right)\,k$$
The exact solution of the Schr\"odinger equation for $0<r<\infty$
\begin{equation}\label{Sch_Exact_2}
\left[-{d^2\over dr^2}+\dfrac{(k-1)(k-3)}{4r^2}+\dfrac{4 b \beta-\frac{\beta c^2}{4 b^2}+\left(\frac{c }{b}-\frac{2 }{\beta}-\frac{\beta c^2}{4 b^2}\right)\,k}{r+\beta}+c\,r+b^2r^2\right]u_{il}^d(r)=\left(b\,(k+4)-\dfrac{c^2}{4b^2}\right)u_{il}^d(r)
\end{equation}
is
\begin{equation}\label{wf2_gen_sol_2}
u_{il}^d(r)=r^{(k-1)/2}\left(1+\dfrac{c}{2b}\,r+\frac{\beta\,c-2 b}{2 b \beta^2}\, r^2\right)\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)
\end{equation}
subject to the relation among the parameters $b$, $c$ and $\beta$ given by
\begin{equation}\label{root_eq}
-16 b^3 k+4 b^2 \beta c (3+5 k)+b\,\beta^2 \left(32 b^3-8 c^2 (1+k)\right)+c\,\beta^3 \left(c^2 (1+k)-8 b^3\right)=0
\end{equation}
As an example, for $b=c=1$ and $k$, the roots of equation \eqref{root_eq} for $\beta>0$ are
$\beta_1=0.760~237~519~523~249~5$ and $\beta_2=3.854~071~917~077~363~6$. We display in Figure \ref{wfcase2} the exact solutions as given by \eqref{wf2_gen_sol_2}.
\begin{figure}[ht]
\centering
\includegraphics[scale=1.2]{wfcase2.eps}
\caption{Un-normalized ground-state and first-excited wave functions as given by \eqref{wf2_gen_sol_2} for $k=4, b=c=1$ and specific values of $\beta$.}\label{wfcase2}
\end{figure}
\noindent For a third-degree polynomial solution, $n'=3$, of equation \eqref{secondorderde},
\begin{equation}\label{third_poly_sol}
f_i(r)=1+\dfrac{c}{2b}\,r+\dfrac{4ab^2+\beta((k+1)c^2-24b^3)}{8\,b^2\beta\, k}\, r^2+\dfrac{4 a b^2 (3 \beta c (1+k)-4 b k)+\beta^2 c \left(c^2 (1+k) (3+k)-8 b^3 (9+7 k)\right)}{48 b^3\beta^2 k (1+k)}\,r^3,
\end{equation}
and the exact solution of equation \eqref{Sch} reads
\begin{align}\label{wf2_gen_sol}
u_{i\ell}^d(r)&=r^{(k-1)/2}\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)\notag\\
&\left(1+\dfrac{c}{2b}\,r+\dfrac{4ab^2+\beta((k+1)c^2-24b^3)}{8\,b^2\beta\, k}\, r^2+\dfrac{4 a b^2 (3 \beta c (1+k)-4 b k)+\beta^2 c \left(c^2 (1+k) (3+k)-8 b^3 (9+7 k)\right)}{48 b^3\beta^2 k (1+k)}\,r^3\right),
\end{align}
where
\begin{equation}\label{third_poly_En}
E_{il}^d=b\,(k+6)-\dfrac{c^2}{4b^2},
\end{equation}
and the parameters $a,b,c$ and $\beta$ satisfy, by means of \eqref{cond_poly3_1} and \eqref{cond_poly3_2}, the conditions
\begin{align}\label{third_poly_cond}
48\, a^2\, b^4\, \beta\, (1 + k) &-
8\, a\, b^2\, (48\, b^3\, \beta^2\, (1 + k) - 12\, b^2\, k\, (1 + k) +
8\, b\, \beta\, c\, k\, (2 + k) - 3\, \beta^2\, c^2\, (1 + k)\, (3 + k)) \notag\\
&+
\beta^3\, (576\, b^6\, (1 + k) + c^4\, (1 + k)\, (3 + k)\, (5 + k) -
16\, b^3\, c^2\, (24 + 5\, k\, (5 + k)))=0\notag\\ \notag\\
8\, a^2\, b^3\, (-4\, b\, k &+ 3\, \beta\, c\, (1 + k)) +
2 \,a\, b\, (\left[\beta^2\, c^3\, (3 + k)-48\, b^4 \beta\right]\, (1 + k) -
8\, b^2\, c\, k\, (5 + k) + 6\, b\, \beta\, c^2\, (1 + k)\, (5 + k) \notag\\
&-
8\, b^3\, \beta^2\, c\, (9 + 7 k)) +
\beta^2\, (\left[576\, b^6 + c^4\, (3 + k)\, (5 + k)\right](1+k) -
16\, b^3\, c^2\, (24 + 5\, k\, (5 + k)))=0.
\end{align}
\vskip0.1true in
\noindent For arbitrary values of the potential parameters $a$, $b$, $c$ and $\beta$ that do not necessarily obey the above conditions, we may use AIM directly to compute the eigenvalues \emph{accurately}, as the zeros of the termination condition (\ref{tm_cond}). The method above can also be used to verify the exact solutions we have obtained earlier. For arbitrary parameters, we employ AIM with
\begin{equation}\label{aim_start}
\lambda_0(r)=\dfrac{1-k}{r}+2\,b\,r+\dfrac{c}{b}\qquad\mbox{and}\qquad
s_0(r)= \dfrac{c(k-1)}{2\,b\,r}+\dfrac{a}{r+\beta}+\dfrac{4b^3\,k-c^2-4b^2E_{nl}^d}{4b^2}.
\end{equation}
and compute the AIM sequences $\lambda_n$ and $s_n$ as given by Eq.(\ref{AIM_seq}). We note that for given values of the potential parameters $a$, $b,$ and of $k=d+2\ell$, the termination condition $\delta_n=\lambda_n s_{n-1}-\lambda_{n-1}s_n=0
$ yields again an expression that depends on both $r$ and $E$. Thus, in order to use AIM as an approximation technique for computing the eigenvalues $E$ we need to feed AIM with a suitable initial value of $r=r_0$ that could stabilize AIM (that is, to avoid oscillations). Again, for our calculations in Table \ref{table:tab2}, we have used $r_0=3$.
\begin{table}[!h] \caption{Eigenvalues $E_{0\ell}^{d=4,5}$ for $V(r)=1/(r+1)+r+r^2$ and different values of the angular momentum $\ell$. The initial value used by AIM is $r_0=3$. The subscript $N$ refers to the number of iteration used by AIM.\\ }
\centering
\begin{tabular}{||c |p{3.6in}||c| p{3.6in}|}
\hline
$\ell$&$E_{0\ell}^{d=4}$&$l$&$E_{0\ell}^{d=5}$\\ \hline
$0$&$~5.743~064~598~822_{N=75}=E_{01}^{2}$&$0$&$~6.881~699~763~857_{N=69}=E_{01}^{3}$\\ \hline
$1$&$~8.010~441~473~733_{N=63}=E_{02}^{2}=E_{00}^{6}$&$1$&$~9.131~165~616~720_{N=56}=E_{02}^{3}=E_{00}^{7}$\\ \hline
$2$&$10.245~221~261~283_{N=52}=E_{03}^{2}=E_{01}^{6}=E_{00}^{8}$&$2$&$11.353~616~525~901_{N=48}=E_{03}^{3}=E_{01}^{7}=E_{00}^{9}$\\ \hline
$3$&$12.457~128~050~974_{N=44}=E_{04}^{2}=E_{02}^{6}=E_{01}^{8}=E_{00}^{10}$&$3$&$13.556~369~149~873_{N=41}=E_{04}^{3}=E_{02}^{7}=E_{01}^{9}=E_{00}^{11}$\\ \hline
$4$&$14.651~834~191~239_{N=39}=E_{05}^{2}=E_{03}^{6}=E_{02}^{8}=E_{01}^{10}=E_{00}^{12}$&$4$&$15.743~928~601~250_{N=38}=E_{05}^{3}=E_{03}^{7}=E_{02}^{9}=E_{01}^{11}=E_{00}^{13}$\\ \hline
$5$&$16.832~989~791~994_{N=35}=E_{06}^{2}=E_{04}^{6}=E_{03}^{8}=E_{02}^{10}=E_{01}^{12}=E_{00}^{14}$&$5$&$17.919~302~156~447_{N=32}=E_{06}^{3}=E_{04}^{7}=E_{03}^{9}=E_{02}^{11}=E_{01}^{13}=E_{00}^{15}$\\ \hline
\hline
\end{tabular}
\label{table:tab2}
\end{table}
\section{Exact and approximate solutions for the pure Coulomb
potential plus linear and oscillator radial terms}\label{app2}
\noindent In this section, we focus our attention on the case of $\beta=0$, specifically we study the exact and approximate eigenenergies of a hydrogenic atom with a Coulomb potential \cite{castro} in the presence of an external linear term and an harmonic oscillator. We have
\begin{align}\label{Pert_coul}
V(r)=\dfrac{a}{r}+c\,r+b^2\,r^2,\qquad a\neq 0,\quad b>0.
\end{align}
This soft-confined potential has been the subject of intensive study over the past few decades in a wide range of contexts \cite{roy1988,bessis1987,roy1990,hall2004,hall1996}. In the light of solutions to the equation \eqref{gen_DE}, we discuss the quasi-exact solutions of Schr\"odinger equation for the potential \eqref{Pert_coul} and their connection with the solution of the biconfluent heun equation \cite{ron1995,dutra,leaute1986, leaute1990,arriola1991,ovsiyuk, caruso} where we extend the some of the known results to arbitrary dimensions and provide a compact analytic solutions that we use to verify our approximation method using AIM. For this purpose, we set $\beta=0$ in the differential equation \eqref{secondorderde} to obtain
\begin{align}\label{Pert_coul_1}
r&\,f_n''(r)+\left(-2b\,r^2-\dfrac{c}{b}\,r+k-1\right)f_n'(r)
+\left(\left(E-b\,k+\dfrac{c^2}{4b^2}\right)\, r-a+\dfrac{(1-k)\,c}{2b}
\right)f_n(r)=0.
\end{align}
This equation can easily be compared with \eqref{gen_DE} for $a_{4,3}=a_{3,3}=\tau_{3,3}=0$, in which case equation \eqref{gen_DE} reduces to
\begin{align}\label{gen_DE_sp}
a_{4,2}r& y^{\prime \prime}+(a_{3,0}r^2+a_{3,1}r+a_{3,2})y'+(-\tau_{2,0}r-\tau_{2,1})y=0,
\end{align}
with polynomial solutions $y_n=\sum_{j=0}^n {C}_j\, r^j$ only if $\tau_{2,0}=n\,a_{3,0},$ and polynomial coefficients $ {C}_j$ that satisfy the three-term recurrence relations, derived by use of the Frobenius method,
given by
\begin{align}\label{rec_rel_sp}
((n-1) a_{3,0}-\tau_{2,0}) {C}_{n-1}+(na_{3,1}-\tau_{2,1}) {C}_{n}&+(n+1)(na_{4,2}+ a_{3,2}) {C}_{n+1}=0,\qquad {C}_{-1}=0,\quad {C}_0=1.
\end{align}
In this case, the first few polynomials are
\begin{align*}
f_0(r)&=1\qquad{providing}\quad \tau_{2,1}=0,\\
f_1(r)&=
1+\frac{\tau_{2,1}}{a_{3,2}}\, r\qquad{providing}\quad
\left|\begin{array}{ccc}
-\tau_{2,1}&a_{3,2}\\
-a_{3,0}& a_{3,1} -\tau_{2,1}
\end{array}\right|=0,\\
f_2(r)&=1+\dfrac{\tau_{2,1}}{a_{3,2}}\, r+\dfrac{
2a_{3,0}a_{3,2}-a_{3,1}\tau_{2,1}+\tau_{2,1}^2}{2a_{3,2}(a_{3,2}+a_{4,2})}\,r^2 \qquad{providing}\quad\left|\begin{array}{cccc}
-\tau_{2,1} &a_{3,2} &0\\
-2 a_{3,0}& a_{3,1} - \tau_{2,1}& 2 (a_{3,2} + a_{4,2})\\
0&-a_{3,0}&2 a_{3,1} - \tau_{2,1}
\end{array}\right|=0,\\
f_3(r)&=1+\dfrac{\tau_{2,1}}{a_{3,2}}\, r+\dfrac{
3a_{3,0}a_{3,2}-a_{3,1}\tau_{2,1}+\tau_{2,1}^2}{2a_{3,2}(a_{3,2}+a_{4,2})}\,r^2 +\dfrac{a_{3,0} (-6 a_{3,1} a_{3,2} + 7 a_{3,2} \tau_{2,1} + 4 a_{4,2} \tau_{2,1}) +
\tau_{2,1} (2 a_{3,1}^2 - 3 a_{3,1} \tau_{2,1} + \tau_{2,1}^2)}{6 a_{3,2} (a_{3,2} + a_{4,2}) (a_{3,2} + 2 a_{4,2})}\,r^3,\notag\\
&\qquad{providing}\quad\left|\begin{array}{ccccc}
-\tau_{2,1} &a_{3,2} &0&0\\
-3 a_{3,0} & a_{3,1} - \tau_{2,1} & 2 a_{3,2} + 2 a_{4,2}&0\\
0&-2 a_{3,0}& 2 a_{3,1} - \tau_{2,1}&3 a_{3,2} + 6 a_{4,2}\\
0&0&-a_{3,0}& 3 a_{3,1} - \tau_{2,1}\\
\end{array}\right|=0.
\end{align*}
On other hand, as noted earlier, \eqref{Pert_coul_1} is a special case of the biconfluent Heun differential equation \cite{ron1995,rovder,hautot1969}
\begin{equation}\label{bheun}
zf^{\prime\prime}(z) +(1+\alpha-\beta z-2z^2)f^\prime(z)+\left[(\gamma-\alpha-2)\,z-\dfrac12(\delta+(1+\alpha)\beta)\right]f(z)=0.
\end{equation}
Indeed, by a simple comparison, with $z=\sqrt{b}\,r$, between \eqref{bheun} and \eqref{Pert_coul}, we find by using
\begin{equation}\label{eq50}
\alpha=k-2,\quad \beta=\dfrac{c}{b^{3/2}},\quad \gamma=\dfrac{E}{b}+\dfrac{c}{4b^3},\qquad \delta = \dfrac{2a}{\sqrt{b}}.
\end{equation}
that we can express the analytic solutions of \eqref{gen_DE_sp} in terms of the Bi-confluent Heun functions \cite{ron1995,rovder,hautot1969} as
\begin{align}\label{bheun_func}
f(r)=H_B\left(k-2,\dfrac{c}{b^{3/2}},\dfrac{E}{b}+\dfrac{c}{4b^3}, \dfrac{2a}{\sqrt{b}},\sqrt{b}\, r\right)
\end{align}
with polynomial solutions providing $E_{nl}^d=b\,(2n'+k)-{c^2}/{4b^2},n'=0,1,2,\dots$. To this end, the polynomial solutions of the differential equation
\begin{align}\label{Pert_coul_2}
r&\,f_{n'}''(r)+\left(-2b\,r^2-\dfrac{c}{b}\,r+k-1\right)f_{n'}'(r)
+\left(2\,b\,n'\, r-a+\dfrac{(1-k)\,c}{2b}
\right)f_{n'}(r)=0,\qquad n'=0,1,2,\dots
\end{align}
are
\begin{align}\label{bheun_func}
f_{n'}(r)=H_B\left(k-2,{c}\,{b^{-3/2}},2\,n'+k, 2\,a\,b^{-1/2},\sqrt{b}\, r\right)=\sum_{j=0}^{n'} {C}_j\, r^j
,\qquad n'=0,1,2,\dots.
\end{align}
where the coefficients $ {C}_j$ are easily computed by means of the three-term recurrence relations, using \eqref{rec_rel_sp},
\begin{align}\label{Re-Rel}
(j + 1)(j + k - 1)\, {C}_{j+1} + \left(\dfrac{(1 - k)c}{2b} - a -\dfrac{c}{b}j\right) {C}_j +
2b(n' - j + 1)\, {C}_{j-1}= 0,\quad {C}_{-1} = 0,\quad {C}_0 = 1,
\end{align}
subject to the termination condition $ {C}_{j+1}=0.$
Thus, the first few polynomial solutions are given explicitly as
\begin{align}
&f_0(r)=1,\qquad{providing }\qquad 2ab+(k-1)\,c=0,\label{case1}\\ \notag\\
&f_1(r)=1+\dfrac{2ab+(k-1)\,c}{2b(k-1)}\,r,\qquad{providing}\qquad 8b^3(1-k)+4b^2a^2+4kcba+c^2(k^2-1)=0,\label{case2}\\ \notag\\
&f_2(r)=1+\dfrac{2ab+(k-1)\,c}{2b(k-1)}\,r+\dfrac{16b^3(1-k)+4b^2a^2+4kcba+c^2(k^2-1)}{8b^2\,k\,(k-1)}\, r^2,\qquad{providing}\notag\\
&32a(1-2k)b^4+8 \left(a^3 - 2 c (-1 + k) (3 + 2 k)\right)\,b^3+12 a^2 c (1 + k)\,b^2-2 a c^2 (1 - 3 k (2 + k))\,b+c^3(k^2-1)(k+3)=0. \label{case3}
\end{align}
For arbitrary values of the potential parameters, we may initiate the asymptotic iteration method to solve the eigenvalue problem independently of the above mentioned constraints. Although, AIM was applied previously to study this potential \cite{barakat,amore}, we claim that we obtain here more accurate and consistent numerical results. Using AIM with
\begin{equation}\label{aim_start}
\lambda_0(r)=\dfrac{1-k}{r}+2\,b\,r+\dfrac{c}{b} \qquad\mbox{and}\qquad s_0(r)= \dfrac{c(k-1)+2ab}{2\,b\,r}+\dfrac{4b^3\,k-c^2-4b^2E_{nl}^d}{4b^2},
\end{equation}
and computing the AIM sequences $\lambda_n$ and $s_n$ using \eqref{AIM_seq}, we evaluate, recursively, the roots of the termination condition \eqref{tm_cond}, starting with the initial value $r_0=3$, similar to the technique used to report Table \ref{table:tab2}. In Table \ref{table:difdn3}, we use AIM to verify the `exact' ground state energy \eqref{case1} for $a=b=1$, then apply AIM to the higher excited states. In Table \ref{table:difdn4}, it is clear that we have greatly improved on the earlier AIM results of Barakat \cite{barakat}. These results also highlight the conclusion obtained by Amore et. al. \cite{amore} on the fast convergence of AIM for this particular problem. In Table \ref{table:difdn5} using the Riccati-Pad\'e method (RPM), we report a simple comparison comparing our results with those obtained earlier by Amore et. al. \cite{amore}. An immediate reason for the improvement noted in the results of Tables \ref{table:difdn4} and \ref{table:difdn5} a consequence of the appropriate structures of the asymptotic solutions near zero and infinity \eqref{gen_sol_exp}. This illustrates the importance of using a more adequate asymptotic solution \cite{saad2008} that usually yields better stability, convergence, and accuracy of AIM.
\begin{table}[h] \caption{Eigenvalues $E_{n0}^{d=3,4,5,6}$ for $V(r)=1/r+cr+r^2$ where $c$ is determined from \eqref{case1}. The initial value used by AIM is $r_0=3$. The subscript $N$ refers to the number of iteration used by AIM.\\ }
\centering
\begin{tabular}{|c|c |p{2.5in}||c|c| p{2.5in}|}
\hline
$c$&$n$&$E_{n0}^{d=3}$&$c$&$n$&$E_{n0}^{d=4}$\\ \hline
$-1$&0&$~2.750~000~000~000~000~000_{N=3,exact}$&$-2/3$&0&$~3.888~888~888~888~888~889_{N=3~Exact}$\\
~&1&$~6.105~909~691~182~920~708_{N=64}$&~&1&$~7.485~841~099~550~171~275_{N=59}$\\
~&2&$~9.615~295~284~487~204~826_{N=62}$&~&2&$11.169~992~576~098~137~834_{N=58}$\\
~&3&$13.210~469~278~706~047~371_{N=61}$&~&3& $14.905~199~749~925~709~834_{N=57}$ \\
~&4&$16.860~555~849~138~091~010_{N=59}$&~&4&$18.674~207~558~484~831~292_{N=57}$\\
~&5&$20.549~102~541~238~464~811_{N=57}$&~&5&$22.467~438~445~946~572~167_{N=55}$\\
~&6&$24.266~299~867~653~311~177_{N=59}$&~&6&$26.279~004~288~368~339~228_{N=54}$\\ \hline
\hline
$c$&$n$&$E_{n0}^{d=5}$&$c$&$n$&$E_{n0}^{d=6}$\\ \hline
$-1/2$&0&$~4.937~500~000~000~000~000_{N=3,exact}$&$-2/5$&0&$~5.960~000~000~000~000~000_{N=3~Exact}$\\
~&1&$~8.655~823~170~124~162~086_{N=56}$&~&1&$~9.749~149~491~375~024~656_{N=50}$\\
~&2&$12.428~555~074~489~786~355_{N=54}$&~&2&$13.574~797~401~850~504~632_{N=50}$\\
~&3&$16.234~977~694~977~922~106_{N=52}$&~&3& $17.424~191~007~631~759~307_{N=49}$ \\
~&4&$20.064~504~343~130~534~075_{N=52}$&~&4&$21.290~390~825~325~282~344_{N=48}$\\
~&5&$23.910~981~048~253~203~499_{N=53}$&~&5&$25.169~188~410~054~967~854_{N=48}$\\
~&6&$27.770~499~635~352~076~648_{N=50}$&~&6&$29.057~829~460~632~247~615_{N=48}$\\ \hline
\hline
\end{tabular}
\label{table:difdn3}
\end{table}
\begin{table}[!h] \caption{A comparison between selected eigenenergies calculated by Barakat \cite{barakat} and in the present work.\\ }
\centering
\begin{tabular}{|c|c|c|c|c|p{2.0in}| p{0.5in}|}
\hline
$n$&$\ell$&$a$&$c$&$b$&$E_{nl}^{3}$&$\epsilon_{Barakat}$\\ \hline
$0$&$0$&$-2$&$0.894~42$&$\sqrt{0.2}$&$0.341~633~800~749~479~644_{N=45}$& $0.341~64$\\ \hline
$0$&$1$&$-2$&$0.447~22$&$\sqrt{0.2}$&$1.986~079~419~684~181~694_{N=33}$&$1.986~06 $\\ \hline
$0$&$2$&$-2$&$0.298~14$&$\sqrt{0.2}$&$3.019~378~385~388~245~576_{N=30}$& $3.019~38 $\\ \hline
$0$&$3$&$-2$&$0.223~60$&$\sqrt{0.2}$&$3.962~403~145~424~275~957_{N=37}$& $3.962~42 $\\ \hline
\hline
$0$&$0$&$-2$&$8.944~28$&$2\sqrt{5}$&$12.416~411~447~380~603~566_{N=83}$& $~6.208~20$\\ \hline
$0$&$1$&$-2$&$4.472~14$&$2\sqrt{5}$&$22.110~682~447~511~408~897_{N=72}$&$22.110~64 $\\ \hline
$0$&$2$&$-2$&$2.981~42$&$2\sqrt{5}$&$31.193~837~323~750~444~679_{N=64}$& $31.193~86 $\\ \hline
$0$&$3$&$-2$&$2.236~06$&$2\sqrt{5}$&$40.186~716~024~771~681~149_{N=60}$& $20.093~36 $\\ \hline
\hline
\end{tabular}
\label{table:difdn4}
\end{table}
\begin{table}[!h] \caption{A comparison between eigenenergies as obtained by Amore et. al. \cite{amore} using Riccati–Pad\'e method (RPD) and those of the present work $E_{AIM}$, for particular values of the parameter $c$ in the potential $V(r)=-2/r+c\,r+\sqrt{2}\, r^2$.\\ }
\centering
\begin{tabular}{|c|p{2.0in}|p{2.4in}|}
\hline
$c$&$E_{RPM}$&$E_{AIM}$\\ \hline
$-4$&$-2.343~347~169~439~4$& $-2.343~347~169~439~302~087~596~937_{N=93}$\\ \hline
$-2$&$-0.452~373750~381~743~8$& $-0.452~373~750~381~743~858~907~206_{N=89}$\\ \hline
$~~2$&$~~2.665~690~984~529~681~669~8$& $~~2.665~690~984~529~681~669~856~944_{N=74}$\\ \hline
$~~4$&$~~4.029~812~452~923~474~112~0$& $~~4.029~812~452~923~474~111~929~868_{N=66}$\\ \hline
\hline
\end{tabular}
\label{table:difdn5}
\end{table}
\section{Exact and approximate solutions with hard confinement $r \le R$.}\label{spec}
\noindent In this section, we turn our attention to study the $d$-dimensional radial Schr\"odinger equation
\begin{equation}\label{Sch_conf}
\left[-{d^2\over dr^2}+{(k-1)(k-3)\over 4r^2}+V(r)\right]u_{nl}^d(r)=E_{nl}^d\, u_{nl}^d(r),\qquad \int_0^R |u_{nl}^d(r)|^2dr<\infty, ~~~
u_{n\ell}^{(d)}(0)=u_{n\ell}^{(d)}(R)=0,
\end{equation}
with the potential
\begin{equation}\label{pot_conf}
V(r)=\left\{ \begin{array}{ll}
\dfrac{a}{r+\beta}+c\,r+b^2r^2, &\mbox{ if\quad $0<r< R$}, \\ \\
\infty, &\mbox{ if\quad $r\geq R$,}
\end{array} \right.
\end{equation}
where $u_{nl}^d(0)=u_{nl}^d(R)=0$. We employ the following ansatz for the wave function
\begin{equation}\label{wf_conf}
u_{nl}^d(r)=r^{(k-1)/2}\,(R-r)\,\exp\left(-\dfrac{c}{2b}\,r-\dfrac{b}{2}\,r^2\right)f_n(r),\quad k=d+2l,
\end{equation}
where the $(R-r)$ factor is inserted to ensure the vanishing of the radial wave function $u_{nl}^d(r)$ at the boundary $r=R$. On substituting \eqref{wf_conf} into \eqref{Sch_conf}, we obtain the following second-order differential equation for the functions $f_n(r)$,
\begin{align}\label{DE_conf}
\bigg(&-4 b^2 r^3 + 4 b^2 (R-\beta)\, r^2 + 4 b^2 \beta R\, r\bigg)f_n''(r)+\bigg(8 b^3 r^4 +4 b (c + 2 b^2 (\beta - R)) \,r^3-4 b (b - \beta c + b k + 2 b^2 \beta R + c R)\,r^2\notag\\
&-4 b (\beta c R + b (\beta + \beta k + R - k R))\, r + 4\, b^2 \beta\, (k-1)\, R\bigg)f_n'(r)\notag\\
&+\bigg[\left( 4 b^3 (2 + k)-c^2 - 4 b^2 E \right)\,r^3+(4 a b^2 + 2 b c (1 + k) + (c^2 + 4 b^2 E) (R-\beta) +
4 b^3 (\beta (k+2) - k R))\,r^2\notag\\
&+(2 b\, (\beta\, c\, (1 + k)-2 b\, (k-1) )+(\beta\, (c^2 + 4 b^2 E)-4 a b^2 - 2 b c (k-1) - 4 b^3 \beta k)R)\,r-2 b \beta (k-1) (2 b + c R)\bigg]f_n(r)=0.
\end{align}
This differential equation goes beyond the equation discussed in section \ref{method}, so we introduce another more general class of differential equation that that allows us to analyze the polynomial solutions of \eqref{DE_conf}.
\begin{theorem}\label{thmVII.1}
The second-order linear differential equation
\begin{align}\label{DE543}
(a_{5,0}\, r^5+ a_{5,1}\, r^4 + a_{5,2}\, r^3 + a_{5,3}\, r^2 + a_{5,4}\, r + a_{5,5} )
f''(r)&+(a_{4,0} r^4+ + a_{4,1} r^3 + a_{4,2} r^2 + a_{4,3} r +a_{4,4})f'(r)\notag\\
&-(\tau_{3,0} r^3 +\tau_{3,1}\,r^2 +\tau_{3,2}\,r+\tau_{3,3})f(r)=0
\end{align}
has a polynomial solution
$y(r)=\sum_{j=0}^nc_j r^j$,
if
\begin{equation}\label{DEE543_Nec_cond}
\tau_{3,0}=n\,(n-1)\,a_{5,0}+n\,a_{4,0},\qquad n=0,1,2,\dots,
\end{equation}
provided $a_{5,0}^2+a_{4,0}^2+\tau_{3,0}^2\neq 0$. The polynomial coefficients $c_n$ then satisfy the following six-term recurrence relation
\begin{align}\label{Rec_conf}
((j-3)&(j-4)a_{5,0}+(j-3) a_{4,0}-\tau_{3,0})\,c_{j-3}+((j-2)(j-3)a_{5,1}+ (j-2)a_{4,1}-\tau_{3,1})\,c_{j-2}\notag\\
&+((j-1)(j-2)a_{5,2}+ (j-1)a_{4,2}-\tau_{3,2})\,c_{j-1}+(j(j-1)a_{5,3}+j a_{4,3}-\tau_{3,3} )\,c_{j}\notag\\
&+(j(j+1)a_{5,4}+(j+1)a_{4,4})\,c_{j+1}+(j+1)(j+2)\,a_{5,5}\,c_{j+2}=0
\end{align}
with
$c_{-3}=c_{-2}=c_{-1}=0$. In particular, for the zero-degree polynomials $f_0(r)=1$ where $c_0=1$ and $c_n=0,~n\geq 1$, we must have $\tau_{3,0}=0$ along with
\begin{equation}\label{Case1_conf}
\tau_{3,1}=0,\quad\tau_{3,2}=0,\quad\tau_{3,3}=0.
\end{equation}
For the first-degree polynomial solution $$ f_1(r)=1+\dfrac{\tau_{3,3}}{a_{4,4}}\, r,$$ where $c_0=1,~ c_1=\tau_{3,3}/a_{4,4},$ and $c_n=0,n\geq 2$, we must have $
\tau_{3,0}=a_{4,0}
$
along with the vanishing of the three $2\times 2$-determinants, simultaneously,
\begin{equation}\label{Case2_conf}
\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}\\
-\tau_{3,2}& a_{4,3} -\tau_{3,3}
\end{array}\right|=0,\qquad
\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}\\
-\tau_{3,1}& a_{4,2} -\tau_{3,2}
\end{array}\right|=0,\qquad
\mbox{and}\qquad
\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}\\
-a_{4,0}& a_{4,1} -\tau_{3,1}
\end{array}\right|=0.
\end{equation}
For the second-degree polynomial solution, $$f_2(r)=1+\frac{(a_{4,4}+a_{5,4})\tau_{3,3}-a_{5,5} \tau_{3,2}}{a_{4,4} (a_{4,4}+a_{5,4})+a_{5,5} (\tau_{3,3}-a_{4,3})}\,r+\frac{a_{4,4}\tau_{3,2}+\tau_{3,3} (\tau_{3,3}-a_{4,3})}{2 (a_{4,4} (a_{4,4}+a_{5,4})+a_{5,5} (\tau_{3,3}-a_{4,3}))}\,r^2$$ where $c_n=0$ for $n\geq 3$, we must have
$
\tau_{3,0}=2a_{5,0}+2a_{4,0}$
along with the vanishing of the three $3\times 3$-determinants, simultaneously,
\begin{align}\label{Case3_conf}
&\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}\\
-\tau_{3,2}& a_{4,3} -\tau_{3,3}&2a_{5,4}+ 2a_{4,4}\\
-\tau_{3,1}&a_{4,2} -\tau_{3,2}&2a_{5,3}+2a_{4,3}-\tau_{3,3}
\end{array}\right|=0,\quad
\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}\\
-\tau_{3,2}& a_{4,3} -\tau_{3,3}&2a_{5,4}+ 2a_{4,4}\\
-2a_{5,0}-2a_{4,0}&a_{4,1} -\tau_{3,1}&2a_{5,2}+2a_{4,2}-\tau_{3,2}
\end{array}\right|=0,
\end{align}
and
\begin{align}
\left|\begin{array}{ccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}\\
-\tau_{3,2}& a_{4,3} -\tau_{3,3}&2a_{5,4}+ 2a_{4,4}\\
0&-2a_{5,0}-a_{4,0}&2a_{5,1} +2a_{4,1}-\tau_{3,1}
\end{array}\right|=0,
\end{align}
For third-degree polynomial solution,
\begin{align*}
&f_3(r)=1+\dfrac{2a_{5,5}^2 \tau_{3,1}+(a_{4,4}+a_{5,4}) (a_{4,4}+2a_{5,4})\tau_{3,3}+a_{5,5} \left(\tau_{3,3}^2-2 (a_{4,3}+a_{5,3}) \tau_{3,3}-(a_{4,4}+2a_{5,4}) \tau_{3,2}\right)}{a_{4,4}^3+3a_{4,4}^2a_{5,4}+2a_{5,5} (-a_{4,3}a_{5,4}+a_{4,2} a_{5,5}-a_{5,5}\tau_{3,2}+a_{5,4}\tau_{3,3})+a_{4,4} \left(2a_{5,4}^2+a_{5,5} (2 \tau_{3,3}-3a_{4,3}-2a_{5,3})\right)}\,r\notag\\
&+\dfrac{a_{4,4}^2\tau_{3,2}+2\tau_{3,3} (-a_{4,3}a_{5,4}+a_{4,2}a_{5,5}-a_{5,5}\tau_{3,2}+a_{5,4}\tau_{3,3})+a_{4,4} \left(-2 a_{5,5}\tau_{3,1}+2a_{5,4}\tau_{3,2}-a_{4,3}\tau_{3,3}+\tau_{3,3}^2\right)}{2 \left(a_{4,4}^3+3a_{4,4}^2 a_{5,4}+2 a_{5,5} (-a_{4,3}a_{5,4}+a_{4,2} a_{5,5}-a_{5,5}\tau_{3,2}+a_{5,4}\tau_{3,3})+a_{4,4} \left(2 a_{5,4}^2+a_{5,5} (2\tau_{3,3}-3 a_{4,3}-2 a_{5,3})\right)\right)}\,r^2\notag\\
&+\dfrac{\mu}{6 \left(a_{4,4}^3+3 a_{4,4}^2 a_{5,4}+2 a_{5,5} (-a_{4,3}a_{5,4}+a_{4,2}a_{5,5}-a_{5,5}\tau_{3,2}+a_{5,4} \tau_{3,3})+a_{4,4} \left(2a_{5,4}^2+a_{5,5} (-3a_{4,3}-2a_{5,3}+2\tau_{3,3})\right)\right)}\,r^3,
\end{align*}
where
\begin{align}
\mu&=2a_{4,4}^2\tau_{3,1}+2a_{4,4}a_{5,4}\tau_{3,1}-2 a_{4,4} (a_{4,3}+a_{5,3})\tau_{3,2}
-2 a_{5,5} (a_{4,3} \tau_{3,1}+\tau_{3,2} (-a_{4,2}+\tau_{3,2}))\notag\\
&+(2a_{4,3} (a_{4,3}+a_{5,3})-2a_{4,2} (a_{4,4}+a_{5,4})+2a_{5,5}\tau_{3,1}+3 a_{4,4} \tau_{3,2}+2 a_{5,4} \tau_{3,2}) \tau_{3,3}-(3a_{4,3}+2 a_{5,3}) \tau_{3,3}^2+\tau_{3,3}^3.\end{align}
where $c_n=0$ for $n\geq 4$,
we must have
$
\tau_{3,0}=6a_{5,0}+3a_{4,0}$
along with the vanishing of the three $4\times 4$-determinants, simultaneously,
\begin{align}\label{Case3_conf}
&\left|\begin{array}{cccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}&0\\
-\tau_{3,2}&a_{4,3}-\tau_{3,3}&2a_{5,4}+2a_{4,4}&6a_{5,5}\\
-\tau_{3,1}&a_{4,2} - \tau_{3,2}& 2a_{4,3} + 2a_{5,3} - \tau_{3,3}&3a_{4,4} + 6a_{5,4}\\
-6a_{5,0} - 3a_{4,0}& a_{4,1} - \tau_{3,1}& 2a_{4,2} + 2a_{5,2} - \tau_{3,2}&
3a_{4,3} + 6 a_{5,3} - \tau_{3,3}\\
\end{array}\right|=0,\notag\\ \notag\\
&\left|\begin{array}{cccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}&0\\
-\tau_{3,2}&a_{4,3}-\tau_{3,3}&2a_{5,4}+2a_{4,4}&6a_{5,5}\\
-\tau_{3,1}&a_{4,2} - \tau_{3,2}& 2a_{4,3} + 2a_{5,3} - \tau_{3,3}&3a_{4,4} + 6a_{5,4}\\
0&- 6a_{5,0} -2 a_{4,0}& 2a_{4,1} + 2a_{5,1} - \tau_{3,1} & 3a_{4,2} + 6a_{5,2} - \tau_{3,2}\\
\end{array}\right|=0,\notag\\
\end{align}
and
\begin{align}
\left|\begin{array}{cccc}
-\tau_{3,3}&a_{4,4}&2a_{5,5}&0\\
-\tau_{3,2}&a_{4,3}-\tau_{3,3}&2a_{5,4}+2a_{4,4}&6a_{5,5}\\
-\tau_{3,1}&a_{4,2} - \tau_{3,2}& 2a_{4,3} + 2a_{5,3} - \tau_{3,3}&3a_{4,4} + 6a_{5,4}\\
0&0&- a_{4,0}-4a_{5,0}& 3a_{4,1} + 6a_{5,1} - \tau_{3,1}\\
\end{array}\right|=0,\notag\\
\end{align}
and so on, for higher-order polynomial solutions. The vanishing of these determinants can be regarded as the conditions under which the coefficients $\tau_{3,1}$, $\tau_{3,2}$ and $\tau_{3,3}$ of Eq.(\ref{DE543}) are determined in terms of the other coefficients.
\end{theorem}
\vskip0.1true in
\begin{proof} The proof of this theorem is rather lengthy: it employs the asymptotic iteration method in a similar way to the approach used by Saad {\it et al} in (2014) (\cite{saad2013}, Appendix A).
\end{proof}
\noindent We shall first verify the conclusions of this theorem regarding equation \eqref{DE_conf} by using the asymptotic iteration method followed by an analysis of the solutions for arbitrary parameters. To this end, we employ AIM for \eqref{DE_conf} using
\begin{equation}\label{aim_conf}
\lambda_0=\frac{c}{b}+\frac{1-k}{r}+2 b r-\frac{2}{r-R},\quad s_0=b (2+k)-\frac{c^2}{4 b^2}-E+\dfrac{a}{\beta+r}+\frac{(k-1) (2 b+c\,R)}{2 bR\, r}+\frac{b(1- k)+R(c +2 b^2 R)}{b \, R\, (r-R)},
\end{equation}
and by means of
\begin{align}\label{coeff}
a_{5,0}&=a_{5,1}=a_{5,5}=0,\quad a_{5,2}=-4b^2,\quad a_{5,3}=4b^2(R-\beta),\quad a_{5,2}=4b^2\beta\,R\notag\\
a_{4,0}&=8b^3,\quad a_{4,1}=4 b (c + 2 b^2 (\beta - R)),\quad a_{4,2}=-4 b (b - \beta c + b k + 2 b^2 \beta R + c R)\notag\\
a_{4,3}&=-4 b (\beta c R + b (\beta + \beta k + R - k R)),\quad a_{4,4}= 4\, b^2 \beta\, (k-1)\, R\notag\\
\tau_{3,0}&= -(4 b^3 (2 + k)-c^2 - 4 b^2 E),\quad \tau_{3,1}=-(4 a b^2 + 2 b c (1 + k) + (c^2 + 4 b^2 E) (R-\beta) +
4 b^3 (\beta (k+2) - k R)),\notag\\
\tau_{3,2}&=-(2 b (\beta\, c\, (1 + k)-2 b\, (k-1) )+(\beta\, (c^2 + 4 b^2 E)-4 a b^2 - 2 b c (k-1) - 4 b^3 \beta k)R),\quad \tau_{3,3}=2 b \beta (k-1) (2 b + c R),
\end{align}
the necessary condition for the existence of polynomial solutions $f_n(r)=\sum_{k=0}^n c_k r^k$ of Eq. \eqref{DE_conf} becomes
\begin{equation}\label{En_aim_conf}
E_{n\ell}^d=b\, (2n'+k+2)-\dfrac{c^2}{4b^2},\qquad k=d+2l,
\end{equation}
where $n'$ refers to the degree of the polynomial solution of equation \eqref{DE_conf} and is not necessarily equal to the number of nodes $n$ of the wave function. It is clear from \eqref{coeff}, there is no zero-degree polynomial solution available. For the first-degree polynomial solution, we have
\begin{equation}\label{En_aim_conf_case1}
E_{n\ell}^d=b\, (4+k)-\dfrac{c^2}{4b^2},\quad f_1(r)=1+\left(\frac{c}{2 b}+\frac{1}{R}\right)\,r
\end{equation}
providing
\begin{align}\label{exact_conf_cond}
4 a b^2+2 b c (3+k)+\left(2 a b c+c^2 (3+k)\right) R+4 b^2\, c\, R^2&=0\notag\\
2 b (\beta c (3+k)-4 b k)+c (\beta c (3+k)-4 b k)R+\left(16 b^3-2 a b c+4 b^2\beta c-c^2 (1+k)\right)R^2&=0\notag\\
8 b^2\, \beta\, k+4 b\, \beta\, c\, k\, R+\left(4 a\, b^2+\beta \left(c^2 (1+k)-16\, b^3\right)\right)R^2&=0.
\end{align}
In Table \ref{table:difdn6}, we report the exact eigenvalues $E_{00}^{3}=7b-c^2/(4b^2)$ using the roots of the equations given by \eqref{exact_conf_cond} and the results obtained by AIM initiated with $r_0=R/2$ for different values of $R$ and $\beta$, where we have fixed $k=3$. For arbitrary values of the potential parameters, we can employ AIM initiated with \eqref{aim_conf} to obtain accurate eigenvalues as the roots of the termination condition \ref{tm_cond}. Some of the these results are reported in Table \ref{table:difdn6}. There is an interesting additional application of AIM for these confining potentials: it is possible to use the termination condition to find the proper radius of confinement $R$ for a particular energy; in other words, we may regard the termination condition as function of $(r,R)$ given a particular energy $E$. Consider for example $\beta=a=b=-c=1, k=3$, and $E=9$, what is the radius of confinement for this particular case? The direct application of AIM implies that $R=1.074~414~209~270~221~205$ while for $E=10$, the proper radius of confinement $R=1.016~954~256~339~063~400.$ The method can be easily generalized for arbitrary values of the parameters.
\begin{table}[!h] \caption{A comparison between selected eigenenergies calculated using AIM with the exact values obtained as the roots of equation \eqref{exact_conf_cond}.\\ }
\centering
\begin{tabular}{|c|c|c|c|c|p{1.7in}|c|| }
\hline
$a$&$b$&$c$&$\beta$&$R$&$E_{AIM}$& $E_{exact}$\\ \hline
${20}/{3}$&${11}/{6}$&$-{11}/{6}$&2&1&$12.583~333~333~333~333~333_{N=3}$&$151/12$\\ \hline
${28}/{15}$&${13}/{30}$&$-{13}/{90}$&3&2&$~3.005~555~555~555~555~556_{N=3}$&$541/180$\\ \hline
${55}/{63}$&${47}/{252}$&$-{47}/{1512}$&4&3&$~1.298~611~111~111~111~111_{N=3}$&$187/144$\\ \hline
${143}/{90}$&${71}/{360}$&$-{71}/{1350}$&5&3&$~1.362~777~777~777~777~778_{N=3}$&$2453/1800$\\ \hline
${91}/{180}$&${37}/{360}$&$-{37}/{3600}$&5&4&$~0.716~944~444~444~444~444_{N=3}$&$2581/3600$\\ \hline
${14}/{15}$&${13}/{120}$&$-{13}/{720}$&6&4&$~0.751~388~888~888~888~889_{N=3}$&$541/720$\\ \hline
\hline
\end{tabular}
\label{table:difdn6}
\end{table}
\begin{table}[h] \caption{Eigenvalues $E_{n\ell}^{d=3}$ for $V(r)=1/(r+1)-r+r^2$ for different radius of confinement $R$ . The initial value employed by AIM is $r_0=R/2$. The subscript $N$ refers to the number of iterations used by AIM.\\ }
\centering
\begin{tabular}{|c|c |p{2.5in}||c|c| p{2.5in}|}
\hline
$R$&$n$&$E_{n0}^{d=3}$&$R$&$n$&$E_{n0}^{d=3}$\\ \hline
$1$&0&$~10.328~716~871~106~505~751_{N=36}$&$2$&0&$~3.105~413~452~488~593~322_{N=56}$\\
~&1&$~39.987~716~212~123~541~087_{N=37}$&~&1&$10.692~851~715~920~035~023_{N=55}$\\
~&2&$~89.345~269~629~504~444~833_{N=36}$&~&2&$23.063~954~484~826~017~705_{N=57}$\\
~&3&$158.435~845~294~778~224~568_{N=41}$&~&3& $40.347~069~688~147~624~366_{N=55}$ \\
\hline
\hline
$R$&$\ell$&$E_{0\ell}^{d=3}$&$R$&$\ell$&$E_{0\ell}^{d=3}$\\ \hline
$1$&0&$10.328~716~871~106~505~751_{N=36}$&$2$&0&$~3.105~413~452~488~593~322_{N=56}$\\
~&1&$20.608~236~713~301~997~322_{N=36}$&~&1&$~5.819~309~536~633~945~722_{N=55}$\\
~&2&$33.620~107~194~959~911~851_{N=36}$&~&2&$~9.196~161~676~541~214~605_{N=56}$\\
~&3&$49.228~314~838~693~690~037_{N=36}$&~&3& $21.521~551~806~858~223~355_{N=59}$ \\
\hline
\hline
\end{tabular}
\label{table:difdn6}
\end{table}
\section{Conclusion}\label{conc}
\noindent In this work exact and approximate solutions of Schr\"odinger's equation with softcore Coulomb potentials under hard and soft confinement were found. These problems generate an interesting class of differential equation that goes beyond the classical problems which have solutions of hypergeometric type. In this paper the problems were analyzed as special cases of a very general scheme for the study of linear second-order differential equations with polynomial coefficients that admit polynomial solutions. Necessary and sufficient conditions are derived for the existence of such solutions. The methods presented in this work allow us to obtain compact algebraic expressions for the exact analytical solutions. These are then verified by the asymptotic iteration method. In cases where the parametric conditions for exact polynomial solutions are not met, the asymptotic iteration method is employed directly to find highly accurate numerical solutions. In this work, the asymptotic iteration method served two purposes. The first was to confirm the validity of the sufficient conditions obtained analytically. The second is to provide approximate solutions to the eigenvalue problems, whether potential parameters are specially restricted or freely chosen. For both purposes, the method proves to be extremely effective and provides very accurate results. It is also clear from the present work that the method and the analytic expressions obtained for the different classes of the differential equations can be easily adapted to study other eigenproblems appearing in theoretical physics.
\section{Acknowledgments}
\medskip
\noindent Partial financial support of this work under Grant Nos. GP3438 and GP249507 from the
Natural Sciences and Engineering Research Council of Canada
is gratefully acknowledged by us (respectively RLH and NS).
|
\section{Introduction}
Variable selection with regularized regression has been one of the hot topics in machine learning and statistics. Regularized regressions identify outcome associated features and estimate nonzero parameters simultaneously, and are particularly useful for high-dimensional BIGDATA with small sample sizes. In many real applications, such as bioinformatics, image and signaling processing, and engineering, a large number of features are measured, but only a small number of features are associated with the dependent variables. Including irrelevant variables in the model will lead to overfitting and deteriorate the prediction performance. Therefore, different regularized regression methods have been proposed for variable selection and model construction. $L_0$ regularized regressions, which directly penalize the number of non-zero parameters, are the most essential sparsity measure. Several popular information criteria, including Akaike information criterion (AIC) (Akaike 1974), Bayesian information criterion (BIC) (Schwarz 1978), and risk inflation criteria (RIC) (Foster and George 1994), are based on $L_0$ penalty and have been used extensively for variable selections. However, solving a general $L_0$ regularized optimization is NP hard and computational challenging. Exhaustive search with AIC or BIC over all possible combinations of features is computationally infeasible with high-dimensional BIGDATA.
Different alternatives have been proposed for the regularized regression problem. One common approach is to replace $L_0$ by $L_1$. $L_1$ is known as the best convex relaxation of $L_0$. $L_1$ regularized regression (Tibshirani 1996) is convex and can be solved by an efficient gradient decent algorithm. Minimizing $L_1$ is equivalent to minimizing $L_0$ under certain conditions. However, the estimates of $L_1$ regularized regression are asymptotically biased, and LASSO may not always choose the true model consistently (Zou 2006). Experimental results by Mancera and Portilla (2006) also posed additional doubt about the equivalence of minimizing $L_1$ and $L_0$. Moreover, there were theoretical results (Lin {et al.}, 2010) showing that while $L_1$ regularized regression never outperforms $L_0$ by a constant, in some cases $L_1$ regularized regression performs infinitely worse than $L_0$. Lin {et al.} (2010) also showed that the optimal $L_1$ solutions are often inferior to $L_0$ solutions found using greedy classic stepwise regression, although solutions with $L_1$ penalty can be found effectively.
More recent approaches aimed to reduce bias and overcome discontinuity include SCAD (Fan and Li, 2001), $L_p$ $p \in (0,1]$ regularization (Liu {et al.}, 2010; Mazumder {et al.}, 2011), and MC+ (Zhang, 2010). Even though there are some effects for solving the $L_0$ regularized optimization problems (Dicker {et al.}, 2012; Lu $\&$ Zhang, 2013), $L_0$ was either approximated by a continuous smooth function, or transformed into a much larger ranking optimization problem. To the best of our knowledge, there is no method that optimizes $L_0$ directly.
In this paper, we propose an efficient EM algorithm ($L_0$EM) that directly solves the $L_0$ regularized regression problem. $L_0$EM effectively deals with $L_0$ optimization by solving a sequence of convex $L_2$ optimizations and is efficient for high dimensional data. It also provides a natural solution to all $L_p$ $p\in [0,2]$ problems, including LASSO with $p=1$, elastic net with $p \in [1, 2]$ (Zou $\&$ Zhang 2009), and the combination of $L_1$ and $L_0$ with $p \in (0, 1]$ (Liu $\&$ Wu, 2007). While the regularized parameter $\lambda$ for LASSO must be tuned through cross-validation, which is time-consuming, the optimal $ {\lambda} $ with $L_0$ regularized regression can be pre-determined with different model selection criteria such as AIC, BIC and RIC. We demonstrate our methods through simulation and high-dimensional genomic data. The proposed methods identify the non-zero variables with less-bias and outperform the LASSO method by a large margin. They can also choose the biologically important genes and pathways effectively.
\section{Methods}
Given a $n\times 1$ dependent variable $\mathbf{y}$, and an $n \times m$ feature matrix $X$, a linear model is defined as
\[ \mathbf{y} = X{\theta} + \varepsilon, \]
where n is the number of samples and m is the number of variables and $n \ll m$, ${\theta} = [{\theta}_1, \ldots, {\theta}_m]^t$ are the m parameters to be estimated, and $\varepsilon \sim N (0, {\sigma}^2I_n)$ are the random errors with mean $0$ and variance ${\sigma}^2$. Assume only a small subset of $\{\mathbf{x}_j\}_{j=1}^m$ has nonzero ${\theta}_j$s. Let $R \subseteq \{1, \ldots, m\}$ be the subset index of relevant variables with ${\theta}_j \neq 0$, and $O \subseteq \{1, \ldots, m\}$ be the index of irrelevant features with $0$ coefficients, we have $ R \cup O = \{1, 2, \ldots, m\}$, $X_R \cup X_O = X$, and ${\theta}_R \cup {\theta}_O = {\theta}$, where ${\theta}_O =0$. The error function for $L_1$ regularized regression is
\begin{equation} E = \frac{1}{2}||\mathbf{y} - X{\theta}||^2 +\frac{ {\lambda} }{2}||{\theta}||_0 =\frac{1}{2}\sum_{i=1}^n (y_i -\sum_{j=1}^m {\theta}_j x_{ij})^2 + \frac{ {\lambda} }{2} \sum_{j =1}^m I({\theta}_j \neq 0), \label{eq1} \end{equation}
where $||{\theta}|| _0 =\sum_{j =1}^m I({\theta}_j \neq 0) = |R|$ counts the number of nonzero parameters.
One observation is that equation (\ref{eq1}) is equivalent to the following equation (\ref{eq2}), when reaching the optimal solution.
\begin{equation} E = \frac{1}{2}||\mathbf{y} - X{\theta}||^2 + \frac{ {\lambda} }{2}||{\theta}||_0 = \frac{1}{2}||\mathbf{y} -X{\theta}||^2 + \frac{ {\lambda} }{2} \sum_{j \in R} 1 = \frac{1}{2}||\mathbf{y} -X{\theta}||^2 + \frac{ {\lambda} }{2} |R|, \label{eq2} \end{equation}
because ${\theta}_O$ is a zero vector. Our $L_0$EM methods will be derived from equation (\ref{eq2}).
We can rewrite equation (\ref{eq2}) as the following two equations:
\begin{align} E &= \frac{1}{2}||\mathbf{y} -X{\theta}||^2 + \frac{ {\lambda} }{2} \sum_{j \in R} \frac{{\theta}_j^2}{\eta_j^2} \label{eq3} \\
\eta & ={\theta}. \label{eq4}
\end{align}
Given $\eta_j$, equation (\ref{eq3}) is a convex quadratic function and can be optimized by taking the first order derivative:
\begin{align} \frac{\partial E}{\partial {\theta}_R} &= {\lambda} {\theta}_R\oslash \eta_R^2 - X_R^t(\mathbf{y} -X{\theta}) = 0, \label{eq5} \\
\textrm{ and } \;\; \frac{\partial E}{\partial {\theta}_O} & = \mathbf{0}, \;\; \textrm{ as } \;\; {\theta}_O =\eta_O= \mathbf{0}, \label{eq6} \end{align}
where $\oslash$ indicates element-wise division. Rewriting (\ref{eq5}) and (\ref{eq6}), we have
\begin{align}
{\lambda} {\theta}_R -\eta_R^2\odot X_R^t (\mathbf{y} - X{\theta}) = 0, \label{eq7} \\
{\lambda} {\theta}_O - \eta_O^2\odot X_O^t(\mathbf{y} - X {\theta}) = 0, \;\; \forall \;\; {\lambda} > 0 \label{eq8}
\end{align}
where $\odot$ is element-wise multiplication, $\eta_R^2 \odot X_R^t = [\eta^2_R \odot \mathbf{x}_{1R}^t, \ldots, \eta^2_R\odot \mathbf{x}_{nR}^t ]$, and $\eta_O^2 \odot X_O^t = [\eta^2_O \odot \mathbf{x}_{1O}^t, \ldots, \eta^2_O\odot \mathbf{x}_{nO}^t ] = \mathbf{0}$.
Let $ X_{\eta}^t = \eta^2\odot X^t$ and combining equations (\ref{eq7}) and (\ref{eq8}) together, we have
\begin{equation} \eta^2\odot\frac{\partial E}{\partial {\theta}}= {\lambda} {\theta} -\eta^2\odot X^t(\mathbf{y} - X{\theta}) = {\lambda} {\theta} - X_{\eta}^t(\mathbf{y} - X{\theta}) = 0. \label{eq9} \end{equation}
Solving Equation (\ref{eq9}), we have the following explicit solution.
\begin{align}
{\theta} &= (X_{\eta}^tX + {\lambda} I) ^{-1}X_{\eta}^t\mathbf{y} \label{eq10} \\
\eta &= {\theta}, \label{eq11}
\end{align}
where equation (\ref{eq10}) can be considered as the M-step of the EM algorithm maximizing $-E$, and equation (\ref{eq11}) can be regarded as the E-step with $E(\eta)={\theta}$. Equations (\ref{eq10}) and (\ref{eq11}) together can also be treated as a fixed point iteration method in nonlinear optimization.
\begin{thm}
\label{theorem1}
Given an input matrix $X$, output matrix $\mathbf{y}$, and initialized solution ${\theta}^0$, the nonlinear system determined by equations (\ref{eq10}) and (\ref{eq11}) will converge to a unique solution, as long as the regularized parameter $0 < {\lambda} < ||X_{\theta}^tX||$, and the estimated solution is closer to the true solution after each iterative EM step.
\end{thm}
\noindent\textbf{Proof:} Equations (\ref{eq10}) and (\ref{eq11}) are the same as:
\begin{equation} {\theta} = (X_{{\theta}}^tX + {\lambda} I) ^{-1}X_{{\theta}}^t\mathbf{y} =({\theta}^2\odot X^tX + {\lambda} I)^{-1}({\theta}^2\odot X^t)\mathbf{y} \label{eq12} \end{equation} First, $ G({\theta}) = ({\theta}^2\odot X^tX + {\lambda} I)^{-1}({\theta}^2\odot X^t)\mathbf{y}$ is Lipschitz continuous for ${\theta} \in R^m$, and
\begin{align} \frac{\partial G({\theta})} { \partial {\theta}} & =({\theta}^2\odot X^tX + {\lambda} I)^{-2}[({\theta}^2\odot X^tX + {\lambda} I)(2{\theta}\odot X^t)\mathbf{y} - 2{\theta}\odot X^tX ({\theta}^2\odot X^t)\mathbf{y}] \nonumber \\
& = ({\theta}^2\odot X^tX + {\lambda} I)^{-2}[2 {\lambda} {\theta}\odot X^t \mathbf{y}] = 2 {\lambda} ({\theta}^2\odot X^tX + {\lambda} I) ^{-1}\mathbf{1}_m \nonumber \\
& = 2 {\lambda} (X_{{\theta}}^tX + {\lambda} I)^{-1}\mathbf{1}_m, \label{eq13}
\end{align}
where $I$ is the identity matrix and $\mathbf{1}_m=[1, \ldots, 1]^t$ is a $m$-dimensional vector of $1$s, and we substitute equation (\ref{eq12}) into equation (\ref{eq13}) to get the result.
Because $ {\lambda} < ||X_{{\theta}}^tX||$, it is clear from equation (\ref{eq13}) that
\[ \left|\frac{\partial G({\theta})}{\partial {\theta}_j}\right| = \frac{2 {\lambda} }{||X_{{\theta}}^tX|| + {\lambda} } < \frac{2 {\lambda} }{ {\lambda} + {\lambda} } = 1,\]
$\forall$ $j =1,\ldots, m$. Therefore, there is a Lipschitz constant
\[ {\gamma} = \left|\left|\frac{\partial G({\theta})}{\partial {\theta}}\right |\right|_{\infty} =\max_j \{\left|\frac{\partial G({\theta})}{\partial {\theta}_j}\right|\} < 1. \]
Now given the initial value for equations (\ref{eq10}) and (\ref{eq11}) $\eta = {\theta}^0 \in R^m$, the sequence $\{{\theta}^{r}\}$ remains bounded because $\forall$ $i =1,\ldots, r$,
\begin{align*} ||{\theta}^{i+1} -{\theta}^i||_{\infty}&=||G({\theta}^i) - G({\theta}^{i-1})||_{\infty}\simeq ||\frac{\partial G({\theta})}{\partial {\theta}}({\theta}^{i} -{\theta}^{i-1})||_{\infty} \le \left|\left|\frac{\partial G({\theta})}{\partial {\theta}}\right|\right|_{\infty}||{\theta}^{i} -{\theta}^{i-1}||_{\infty}\nonumber \\
& = {\gamma} ||{\theta}^{i} - {\theta}^{i-1}||_{\infty}\le \ldots \le {\gamma}^i||{\theta}^1 -{\theta}^0||_{\infty}.
\end{align*}
and therefore
\begin{align*}
||{\theta}^r - {\theta}^0||_{\infty} &= ||\sum_{i=0}^{r-1}({\theta}^{i+1}-{\theta}^i)||_{\infty} \le ||{\theta}^{1} -{\theta}^0||_{\infty}\sum_{i=0}^{r-1} {\gamma}^i \\
& \le \frac{||{\theta}^1-{\theta}^0||_{\infty}}{(1-{\gamma})}.
\end{align*}
Now $\forall$ $r, k \ge 0$,
\begin{align*}
||{\theta}^{r+k} - {\theta}^r||_{\infty} &=||G({\theta}^{r+k-1}) - G({\theta}^{r-1}||_{\infty} \le {\gamma} ||{\theta}^{r+k-1}- {\theta}^{r-1}||_{\infty} \\
& \le {\gamma} ||G({\theta}^{r+k-2}) - G({\theta}^{r-2})||_{\infty} \le {\gamma}^2||{\theta}^{r+k-2}- {\theta}^{r-2}||_{\infty} \\
& \le\ldots \le {\gamma}^r||{\theta}^k -{\theta}^0||_{\infty} \le \frac{{\gamma}^r||{\theta}^1-{\theta}^0||_{\infty}}{1-{\gamma}}.
\end{align*}
Hence, \[ \lim_{r, k \rightarrow \infty}||{\theta}^{r+k} - {\theta}^r||_{\infty} = 0, \]
and therefore $\{{\theta}^r\}$ is a Cauchy sequence that has a limit solution ${\theta}^*$.
Next the uniqueness of the solution is easy to show. Assuming there were two solutions ${\theta}^*$ and ${\theta}^{\diamond}$, then
\begin{equation} ||{\theta}^{*} - {\theta}^{\diamond}||_{\infty} = ||G({\theta}^{*}) - G({\theta}^{\diamond})||_{\infty} \le {\gamma} |{\theta}^{*} -{\theta}^{\diamond}||_{\infty}. \label{eq14}\end{equation}
Since ${\gamma} < 1$, equation (\ref{eq14}) can only hold, if $||{\theta}^{*} -{\theta}^{\diamond}||_{\infty} = 0$. i.e. ${\theta}^{*} = {\theta}^{\diamond}$, so the solution of the EM algorithm is unique.
Finally, the EM algorithm will be closer to the true solution at each step, because
\[ ||{\theta}^{r+1} - {\theta}^{*}||_{\infty} =||G({\theta}^{r}) - G({\theta}^{*}||_{\infty} \le {\gamma}||{\theta}^{r} -{\theta}^*||_{\infty}. \]
\begin{lemma}
\label{lemma1}
Assuming that relevant features are independent, i. e. $\mathbf{x}_i^t\mathbf{x}_j = 0$, $\forall$ $i\neq j$ $\&$ $ i, j \in \{1, \ldots, m\}$, then the maximal regularized parameter $ {\lambda} $ can be determined by \[ {\lambda} _{\max} = \max\left\{\frac{(\mathbf{x}_j^t\mathbf{y})^2}{4\mathbf{x}_j^t\mathbf{x}_j}\right\}_{j=1}^m\]
\end{lemma}
\noindent\textbf{Proof:} For each feature $\mathbf{x}_j$ and corresponding coefficient ${\theta}_j$, equations (\ref{eq9}) and (\ref{eq11}) can be rewritten as
\begin{align*}
\eta_j^2\frac{\partial E({\theta})}{\partial {\theta}_j} &= {\lambda} {\theta}_j - \eta_j^2\mathbf{x}_j^t (\mathbf{y} - X{\theta}) = 0 \\
\eta_j &= {\theta}_j, \;\;\; \forall \;\; j \in \{1, \ldots, m\}. \end{align*}
The above two equations are the same as:
\begin{equation} {\lambda} {\theta}_j - {\theta}_j^2\mathbf{x}_j^t (\mathbf{y} - X{\theta}) = 0. \label{eq15}\end{equation}
If ${\theta}_j =0$, then any $ {\lambda} >0$ will satisfy equation (\ref{eq15}). On the other hand, if
${\theta}_j \neq 0$, because $\mathbf{x}_i^t\mathbf{x}_j = 0$, equation (\ref{eq15}) becomes the following quadratic
equation:
\begin{equation} (\mathbf{x}_j^t\mathbf{x}_j) {\theta}_j^2 - (\mathbf{x}_j^t\mathbf{y}){\theta}_j + {\lambda} = 0. \label{eq16} \end{equation}
One necessary condition for equation (\ref{eq16}) to have a solution is:
\[ (\mathbf{x}_j^t\mathbf{y})^2 - 4(\mathbf{x}_j^t\mathbf{x}) {\lambda} \ge 0, \;\;\; \Rightarrow \;\; {\lambda} \le \frac{(\mathbf{x}_j^t\mathbf{y})^2}{4(\mathbf{x}_j^t\mathbf{x}_j)} \]
Therefore the maximal $ {\lambda} $ is \[ {\lambda} _{\max} = \max\left\{\frac{(\mathbf{x}_j^t\mathbf{y})^2}{4\mathbf{x}_j^t\mathbf{x}_j}\right\}_{j=1}^m.\]
If $ {\lambda} > {\lambda} _{\max}$, equation (\ref{eq15}) holds only if all ${\theta}_j = 0$, $\forall$ $j =1, \ldots, m$.
Both Theorem \ref{theorem1} and Lemma 1 \ref{lemma1} provide some useful guidance for implementing the method and choosing the regularized parameter $ {\lambda} $. Theorem \ref{theorem1} shows that the EM algorithm always converges to a unique solution, given a certain $ {\lambda} $ and initial solution ${\theta}^0$, and the estimated value is closer to the true solution after each EM iteration. Note that different initial values may still reach different solution, because of the non-convex $L_0$ penalty. Therefore, it is critical to choose a good initial value. Our experiences with the method indicate that initializing with the estimates from $L_2$ based ridge regression will usually lead to quick converge and super performance. The EM algorithm is as follows.
\begin{tabular}{l}
\textbf{The $L_0$EM Algorithm:}\\\hline
Given a $0< {\lambda} \le {\lambda} _{\max}$, small numbers $\epsilon$ and $\varepsilon$,\\
and training data $\{X, \mathbf{y}\}$,\\
Initializing ${\theta} =(X^tX + {\lambda} I)^{-1}X^t\mathbf{y}$, \\
While 1,\\
\;\;\;\;\; E-step: $ \mathbf{\eta} = {\theta} $ \\
\;\;\;\;\; M-step: $X_{\eta}^t = \eta^2\odot X^t = [\eta^2\odot\mathbf{x}_1 ,\ldots, \eta^2\odot \mathbf{x}^t_n]$\\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\text{ }\text{ } \text{ }$ {\theta} = (X_{\eta}^tX + {\lambda} I) ^{-1}X_{\eta}^t\mathbf{y} $\\
\;\;\;\;\;\;if $||{\theta} -\eta|| < \varepsilon$, Break; End\\
End \\
${\theta}(|{\theta}| <\epsilon) = 0$. \\ \hline\\
\end{tabular}
\noindent Similar procedures can be extended to general $L_p$, $p \in [0, 2]$ without much difficulty. $L_p$ based EM algorithm $L_p$EM is reported in Appendix.
\noindent\textbf{Consistency and Oracle Property:}
Let $\theta_0$ be the true parameter value. The following conditions will be used later for theoretical properties of the $L_0$-regularized estimator of $\theta_0$.
\noindent {\it CONDITIONS}
\begin{itemize}
\item[(C1)] $ln(m)=o(n)$ as $n \to \infty$.
\item[(C2)] There exists a constant $K>0$ such that
$\lambda_{max}(\frac{ X^tX}{n})\le K < \infty$ for large $n$, where for any matrix $B$, $\lambda_{max}(B)$ denotes the largest eigenvalue of $B$.
\item[(C3)] $\frac{\max_j || \mathbf{x}_j||}{\sqrt{n}} =O(\sqrt{ln(mn)})$ or $O(1)$ as $n,m\to \infty$.
\item[(C4)] There exists a constant $c>0$ such that $\frac{\min_j ||\mathbf{x}_j||^2}{n} \ge c>0 $ for large $n,m$.
\item[(C5)] $\mu(X) \equiv \max_{1\le i<j\le m} \frac{|\mathbf{x}_i^t \mathbf{x}_j | }{||\mathbf{x}_i||\cdot ||\mathbf{x}_j||}=O(\sqrt{\frac{ln(m)}{n}})$.
\item[(C6)] $||\theta_0||_0=O(1)$.
\end{itemize}
The above conditions are very mild. Condition (C1) trivially holds for $m\le n$ and for $m>n$. In particular, (C1) is satisfied even for ultra-high dimensional case such as $m=exp(n^{\alpha})$ for $0<\alpha<1$. (C2) is a standard condition for linear regression. Chi (2013, Section 3.2) gives examples satisfying(C3)-(C4). For example, (C3) and (C4) trivially hold if $||\mathbf{x}_i = \sqrt{n}$ for all $j=1,\ldots,m$. (C5) is referred to as the coherence condition under which the covariates are not highly colinear; see Bunea et al. (2007), Candes and Plan (2009), and Chi (2013). (C6) implies that the model is sparse.
The following theorem is a direct consequence of Chi (2013).
\begin{thm}[{\em Consistency}]
\label{theorem2}
Assume that conditions (C1)-(C6) hold.
Let $n(\nu)=(1-\nu)[1 + 1/\mu(X)]$ for some $0<\nu<1$.
For any $0<q<\frac{1}{2}$, let
$\lambda =\frac{3ln(m/q)}{\nu [1+\mu(X) ]} \frac{\max_j || \mathbf{x}_j||^2}{\min_j ||\mathbf{x}_j||^2}$
and
\begin{equation}
\nonumber
\hat\theta = arg \min_{||\theta||_0 \le n(\nu)} E_n(\theta),
\end{equation}
Then, with probability tending to 1,
\begin{align}
\label{equation17}
|| \hat{\theta} - \theta_0|| = O_p(\sqrt{\frac{\ln(nm)}{n}})
\end{align}
\end{thm}
\begin{prf} Note that the normal linear model in this paper is a special case of
the exponential model of Chi (2009): $p_t(y) =\exp(ty -\Lambda(t))$
with $t=\frac{\mathbf{x}_i^t \theta}{\sigma^2}$ and $\Lambda(t) = \frac{\sigma^2 t^2}{2}$.
Then, (\ref{equation17}) follows immediately from Theorem 3.1 of Chi (2009).
\end{prf}
\noindent\textbf{Model Recovery:} Next we show that $L_0$-regularized regression recovers the true model under mild conditions.
\begin{thm}[{\em Oracle Property}]
\label{theorem3}
Assume that conditions (C1)-(C6) hold.
Let
$
A = \{1 \leq j \leq m: \theta_{0j} \neq 0\},
$ and
$
A^c = \{1, 2, \ldots, m\} \backslash A.
$
Then, the minimizer $\hat{\theta}$ in Theorem \ref{theorem2} must satisfy $\hat{\theta}_j = 0$ for $j \in A^c$.
\end{thm}
\begin{prf} Let $\alpha_n=\sqrt{\frac{\ln(nm)}{n}}$.
For any $\theta$ such that $||\theta - \theta_0|| < C \alpha_n$ for some constant $C>0$ and $\displaystyle \sum_{j \in A^c} I(\theta_j \neq 0) \ge 1$, let
\begin{align*}
\tilde{\theta}_j = \left\{ \begin{array}{rcl}
\theta_j & \mbox{if} & j \in A \\
0 & \mbox{if} & j \in A^c
\end{array}\right.
\end{align*}
Then,
\begin{align*}
& E_n(\theta) - E_n(\tilde{\theta}) \\
& = \frac{1}{2n} (\theta - \tilde{\theta})^TX^TX(\theta-\tilde{\theta}) - \frac{1}{n}(\theta-\tilde{\theta})^TX^T(y-X\tilde{\theta}) + \frac{\lambda}{2} (||\theta||_0-||\tilde{\theta}||_0) \\
& = \frac{1}{2n}(\theta-\tilde{\theta})^TX^TX(\theta-\tilde{\theta}) - \frac{1}{n}(\theta-\tilde{\theta})^TX^T(X\theta_0+\epsilon - X \tilde{\theta}) + \frac{\lambda}{2}(||\theta||_0 - ||\tilde{\theta}||_0) \\
& = \frac{1}{2} (\theta-\tilde{\theta})^T \left(\frac{X^TX}{n}\right)(\theta-\tilde{\theta}) - (\theta-\tilde{\theta})^T \left(\frac{X^TX}{n}\right)(\theta_0-\tilde{\theta}) + \\
& +\frac{1}{\sqrt{n}} (\theta-\tilde{\theta})^T \cdot \frac{1}{\sqrt{n}} X^T\epsilon + \frac{\lambda}{2} (||\theta||_0-||\tilde{\theta}||_0) \\
& = I_1 + I_2 + I_3 + I_4
\end{align*}
Because $||\tilde{\theta}-\theta_0|| \leq ||\theta-\theta_0||$, we have $\theta-\tilde{\theta} = O(\alpha_n)$.
Thus,
$I_1 = O(\alpha_n^2)$ and
$I_2 = O(\alpha_n^2)$.
Moreover,
\begin{align*}
\bigg| \bigg| \frac{1}{\sqrt{n}} \epsilon^tX \bigg| \bigg|
= O_p(\sqrt{k\sigma^2}), \hspace{.2in}\mbox{ as } n \to \infty
\end{align*}
where $k=rank(X)\le n$.
Hence,
\begin{align*}
|I_3| \leq \frac{1}{\sqrt{n}} ||\theta-\tilde{\theta}|| \cdot \bigg| \bigg| \frac{1}{\sqrt{n}} X^T\epsilon \bigg| \bigg| = O(\alpha_n)\cdot O_p(\sqrt{k/n}) = O_p(\alpha_n).
\end{align*}
Furthermore,
\begin{align*}
I_4 & = \frac{\lambda}{2} ( ||\theta||_0 - ||\tilde{\theta}||_0) \\
& = \frac{\lambda}{2} \sum_{j=1}^m [I(\theta_j \neq 0) - I(\tilde{\theta}_j \neq 0)] \\
& = \frac{\lambda}{2} \left[\sum_{j \in A} 0 \right] + \frac{\lambda}{2} \sum_{j \in A^c} [I(\theta_j \neq 0) - 0] \\
& = \frac{\lambda}{2} \sum_{j \in A^c} I(\theta_j \neq 0) \geq \frac{\lambda}{2} \cdot 1 > 0.
\end{align*}
By conditions (C3)-C(5), $\lambda = O(ln(m) \cdot ln(nm))$. Therefore, the first three terms $I_1$, $I_2$ and $I_3$ are dominated by $\lambda$ in probability as $n\to\infty$.
Therefore, with probability tending to 1,
\begin{align}
E_n(\theta) - E_n(\tilde{\theta}) > 0.
\end{align}
This completes the proof of Theorem \ref{theorem3}.
\end{prf}
\noindent\textbf{Determination of $ {\lambda} $:} The regularized $ {\lambda} $ determines the sparsity of the model. The standard approach for choosing $ {\lambda} $ is cross-validation and the optimal $ {\lambda} $ is determined by the minimal mean squared error (MSE) of the test data ($ MSE = \sum(y_i -\hat{y_i})^2/n$). One could also adapt the stability selection (SS) approach for $ {\lambda} $ determination (Liu {et al.} 2010; Meinshausen, 2010). It chooses the smallest $ {\lambda} $ that minimizes the inconsistences in number of nonzero parameters with cross-validation. We first calculate the mean and standard deviation (SD) of the number of nonzero parameters for each $ {\lambda} $, and then find the smallest $ {\lambda} $ with 0 SD, where 0 SD indicates that all models in k-fold cross validation has the same number of nonzero estimates. Our experiences indicate that the larger $ {\lambda} $ chosen from both minimal MSE and stability selection ($ {\lambda} = \max \{ {\lambda} _{mse}, {\lambda} _{ss}\}$) has the best performance. Choosing optimal $ {\lambda} $ from cross-validation is computationally intensive and time consuming. Fortunately, unlike LASSO, identifying the optimal $ {\lambda} $ for $L_0$ does not require to use cross validation. The optimal $ {\lambda} _{opt}$ can be determined by variable selection criteria. The optimal $ {\lambda} _{opt}$ can be directly picked using AIC, BIC, or RIC criteria with $ {\lambda} _{opt} = 2$, $\log n$, or $2\log m$, respectively. Each of these criteria is known to be optimal under certain conditions. This is a huge advantage of $L_0$, especially for BIGDATA problems.
\section{Simulations}
To evaluate the performance of $L_0$ and $L_1$ regulation, we assume a linear model $\mathbf{y} = X{\theta} +\varepsilon$, where the input matrix $X$ is from Gaussian distribution with mean $\mu = \mathbf{0}$ and different covariance structures $\Sigma$, where $\Sigma(i, j) = r^{|i-j|}$ with $r = 0, 0.3, 0.6, 0.8$ respectively. The true model is $\mathbf{y} = 2\mathbf{x}_1 - 3\mathbf{x}_2 + 4\mathbf{x}_ 5 + \varepsilon$ with $\varepsilon \sim N(0, 1)$. Therefore, only three features are associated with output $\mathbf{y}$, and the rest of the ${\theta}_i$s are zero. In our first simulation, we first compare $L_0$ and $L_1$ regularized regression with a relative small number of features $m = 50$ and a sample size of n =100. Five-fold cross validation is used to determine the optimal $ {\lambda} $ and compare the model performance. We seek to fit the regularized regression models over a range of regularization parameters $ {\lambda} $. Each $ {\lambda} $ is chosen from $ {\lambda} _{\min} = 1e-4$, to $ {\lambda} _{\max}$ with 100 equally log-spaced intervals, where $ {\lambda} _{\max} = \max\{X^t\mathbf{y}\}$ for $L_1$ and $\max\left\{\frac{(\mathbf{x}_j^t\mathbf{y})^2}{4\mathbf{x}_j^t\mathbf{x}_j}\right\}$ for $L_0$. Lasso function in the statistics toolbox of MATLAB (www.mathworks.com) is used for comparison. Cross-validation with MSE is implemented nicely in the toolbox. The computational results are reported in Table 1.
\begin{table}[t!]
\begin{center}
\begin{tabular}{c|ccc|ccc} \hline
\multirow{2}{*}{r} & \multicolumn{3}{c|}{$L_0$} & \multicolumn{3}{c}{$L_1$}\\\cmidrule{2-7}
&$\#$ SF & MSE & $||\hat{{\theta}} -{\theta}||$ & $\#$ S.F. & MSE & $||\hat{{\theta}}-{\theta}||$ \\ \hline
0& $3.39(\pm 1.1)$ & $1.01 (\pm 0.14)$ & $0.206(\pm 0.12)$ & $14.5(\pm 3.45)$ & $1.19(\pm 0.19)$ & $0.38(\pm 0.1)$\\
0.3 &$3.37(\pm 0.9)$ & $1.02 (\pm 0.16)$ &$0.23(\pm 0.12)$ & $14.5(\pm 2.91)$ & $1.21(\pm 0.19)$ & $0.41 (\pm 0.19)$ \\
0.6& $3.49(\pm 1.7)$ & $1.02 (\pm 0.23)$ & $0.23(\pm 0.16)$ & $13.5(\pm 3.0)$ & $1.26(\pm 0.2)$ &$0.54 (\pm 0.15)$ \\
0.8 &$3.32(\pm 0.9)$ & $1.06 (\pm 0.15)$ & $0.28(\pm 0.21)$ &$11.7(\pm 2.69)$ & $1.3(\pm 0.21)$ & $0.89(\pm 0.25)$ \\\hline
\end{tabular}
\caption{\footnotesize Performance measures for $L_0$ and $L_1$ regularized regression over 100 simulations, where values in the parenthesis are the standard deviations, and $\#$ SF: number of average selected features; MSE: Average mean squared error; $||\hat{{\theta}} -{\theta}||$: average absolute bias when comparing true and estimated parameters.}
\end{center}
\end{table}
Table 1 shows that $L_0$ outperforms LASSO in all categories by a substantial margin, when using the popular test MSE measure for model selection. In particular, the number of variables selected by $L_0$ are very close to the true number of variables (3), while LASSO selected more than 11 features on average with different correlation structures (r = 0, 0.3, 0.6, 0.8). The test MSEs and bias both increase with the growth of correlation among features for both $L_0$ and LASSO, but the test MSE and bias of $L_0$ are substantially lower than these of LASSO. The maximal MSE of $L_0$ is 1.06, while the smallest MSE of $L_1$ is 1.19, and the largest bias of $L_0$ is 0.28, while the smallest bias of LASSO is 0.38. In addition (results are not shown in Table 1), $L_0$ correctly identifies the true model 81, 74, 81, and 82 times for r = 0, 0.3, 0.6, and 0.8 respectively over 100 simulations, while LASSO never chooses the correct model. Therefore, compared to $L_0$ regularized regression, LASSO selects more features than necessary and has larger bias in parameter estimation. Even though it is possible to get a correct model with LASSO using a larger $ {\lambda} $, the estimated parameters will have a bigger bias and worse predicted MSE.
The same parameter setting is used for our second simulation, but the regularized parameter $ {\lambda} $ is determined by the larger $ {\lambda} $ from both minimal MSE and stability selection ($ {\lambda} = \max\{ {\lambda} _{MSE}, {\lambda} _{SS}\}$). The
computational results are reported in Table 2.
\begin{table}[thb!]
\begin{center}
\begin{tabular}{c|ccc|ccc} \hline
\multirow{2}{*}{r} & \multicolumn{3}{c|}{$L_0$} & \multicolumn{3}{c}{$L_1$}\\\cmidrule{2-7}
&$\#$ SF & MSE & $||\hat{{\theta}} -{\theta}||$ & $\#$ S.F. & MSE & $||\hat{{\theta}}-{\theta}||$ \\ \hline
0& $3.09(\pm 0.53)$ & $1.04 (\pm 0.15)$ & $0.18(\pm 0.11)$ & $13.3(\pm 4.56)$ & $1.21(\pm 0.17)$ & $0.39(\pm 0.1)$\\
0.3 &$3.08(\pm 0.54)$ & $1.04 (\pm 0.15)$ &$0.17(\pm 0.07)$ & $14.5(\pm 4.20)$ & $1.22\pm 0.17)$ & $0.42 (\pm 0.19)$ \\
0.6& $3.10 (\pm 0.46)$ & $1.07 (\pm 0.17)$ & $0.21(\pm 0.10)$ & $13.8(\pm 5.4)$ & $1.27(\pm 0.47)$ &$0.57 (\pm 0.25)$ \\
0.8 &$3.02(\pm 0.14)$ & $1.04 (\pm 0.14)$ & $0.26(\pm 0.13)$ &$13.4(\pm 4.91)$ & $1.25(\pm 0.21)$ & $0.74(\pm 0.25)$ \\\hline
\end{tabular}
\caption{\footnotesize Performance measures for $L_0$ and $L_1$ regularized regression with $ {\lambda} = \max\{ {\lambda} _{mse}, {\lambda} _{ss}\}$ over 100 simulations, where values in the parenthesis are the standard deviations, and $\#$ SF: number of average selected features; MSE: Average mean squared error; $||\hat{{\theta}} -{\theta}||$: average absolute bias when comparing true and estimated parameters.}
\end{center}
\end{table}
Table 2 shows that the average number of associated features is much closer to 3 with sightly larger test MSEs. The maximal average number of features is 3.1 with $r = 0.6$, reduced from 3.49 with the test MSE only. In fact, with this combined model selection criteria and 100 simulations, $L_0$EM identified the true model with three nonzero parameters 95, 95, 95, and 97 times respectively (not shown in the table), while LASSO did not choose any correct models. The average bias of the estimates with $L_0$EM is also reduced. These indicate that the combination of test MSE and stability selection in cross-validation leads to better model selection results than MSE alone with $L_0$EM. However, the computational results did not improve much with LASSO. Over 13 features on average were selected under different correlation structures, suggesting that LASSO inclines to select more spurious features than necessary. A much more conservative criteria with larger $ {\lambda} $ is required to select the right number of features, which will induce larger MSE and bias, and deteriorate the prediction performance.
\subsection*{Simulation with high- dimensional data}
Our third simulation deals with high-dimensional data with the number of samples $n =100$, and the number of features $m = 1000$. The correlation structure is set to $ r =0, 0.3, 0.6$, and the same model $\mathbf{y} = 2 \mathbf{x}_1 - 3\mathbf{x}_2 + 4\mathbf{x}_5 + \varepsilon $ was used for evaluating the performance of $L_0$ and $L_1$. The simulation was repeated 20 times. The computational results are reported in Table 3.
\begin{table}[thb!]
\begin{center}
\begin{tabular}{c|c|ccc} \hline
& Measures & r = 0 & r = 0.3 & r = 0.6 \\ \hline
\multirow{4}{*}{$L_0$} & $\#$ SF & $3(\pm 0)$ & $2.9(\pm 0.47)$ & $2 (\pm 0.73)$ \\
& $||\hat{{\theta}} -{\theta}||$ & $0.14 (\pm 0.09)$ & $0.39(\pm 0.63)$ & $1.69(\pm 1.25)$ \\
& Test MSE & $1.14(\pm 0.34) $ & $1.59 (\pm 1.3)$ & $ 2.8 (\pm 1.72)$ \\
& $\#$ True Model & 20/20 & 15/20 & 5/20 \\ \hline
\multirow{4}{*}{$L_1$} & $\#$ SF & $24(\pm 18.4)$ & $31.3(\pm 20.7)$ & $36.7 (\pm 16.5)$ \\
& $||\hat{{\theta}} -{\theta}||$ & $0.57 (\pm 0.11)$ & $0.73(\pm 0.13)$ & $1.14(\pm 0.25)$ \\
& Test MSE & $1.50(\pm 0.25) $ & $1.63 (\pm 0.29)$ & $ 1.92(\pm 0.41)$ \\
& $\#$ True Model & 0/20 & 0/20 & 0/20 \\ \hline
\end{tabular}
\caption{\footnotesize Performance measures for $L_0$ and $L_1$ regularized regression with cross validation and $ {\lambda} = \max\{ {\lambda} _{mse}, {\lambda} _{ss}\}$ over 20 simulations and the sample size of $n=100$, and $m =1000$, where values in the parenthesis are the standard deviations, and $\#$ SF: number of average selected features; MSE: Average mean squared error; $||\hat{{\theta}} -{\theta}||$: average absolute bias when comparing true and estimated parameters.}
\end{center}
\end{table}
Table 3 shows that $L_0$ outperforms LASSO by a large margin when correlations among features are low. When there is no correlation among features, 20 out of 20 simulations identify the true model with $L_0$, and 15 out of 20 simulations choose the correct model when $r = 0.3$, while LASSO again chooses more features than necessary and no true model was found under any correlation setting. However, when correlations among features are large with $r =0.6$, the results are mixed. $L_0$ can still identify 5 out of 20 correct models, but the test MSE and bias of the parameter estimate of $L_0$ are slightly large than those of LASSO. In addition, we notice that $L_0$ is a more sparse model when correlation increases, indicating that $L_0$ tends to choose independent features. The regularization path of $L_0$ regression is shown in Figure 1.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.55]{regpath3}
\caption{\footnotesize Regularized path for $L_0$ penalized regression with n=100, m =1000, and r = 0.3}
\end{figure}
As shown in the top panel of Figure 1, the three associated features first increase their values when $ {\lambda} $ goes larger, and then go to zero when $ {\lambda} $ becomes extremely big, while the rest of the irrelevant features all go to zero when $ {\lambda} $ increases. Unlike LASSO, which shrinks all parameters uniformly, $L_0$ will only forces the estimates of irrelevant features go to zero, while keep the estimates of relevant features to their true value. This is the well-known Oracle property of $L_0$. For this specific simulation, the three parameters $[\hat{{\theta}_1}, \hat{{\theta} _2}, \hat{{\theta}_5}] = [ 1.85, -2.94, 4.0]$, very close to their true values $[2, -3, 4]$. The middle and bottom panels are the test MSE and the standard deviation of the number of nonzero variables. The optimal $ {\lambda} $ is chosen from the the larger $ {\lambda} $ with minimal test MSE and stability selection as shown in the vertical lines of Figure 1.
\subsubsection*{$L_0$ regularized regression without cross validation} Choosing the optimal parameter $ {\lambda} _{opt}$ with cross-validation is time consuming, especially with BIGDATA. As we mentioned previously, the optimal $ {\lambda} $ can be picked from theory instead of cross validation. Since we are dealing with the $n \ll m$ BIGDATA problem, RIC with $ {\lambda} _{opt} = 2 \log m$ is penalized too much for such problem. So computational results with AIC and BIC without cross validation are reported in Table 4.
\begin{table}[t!]
\begin{center}
\begin{tabular}{c|c|ccc} \hline
& Measures & r = 0 & r = 0.3 & r = 0.6 \\ \hline
\multirow{4}{*}{AIC} & $\#$ SF & $3.26(\pm 0.54)$ & $3.72(\pm 1.94)$ & $4.8 (\pm 2.77)$ \\
& $||\hat{{\theta}} -{\theta}||$ & $0.19 (\pm 0.09)$ & $0.36(\pm 0.58)$ & $1.02(\pm 1.2)$ \\
& MSE$^*$ & $0.96(\pm 0.14) $ & $1.02 (\pm 0.31)$ & $ 1.27 (\pm 0.51)$ \\
& $\#$ True Model & 78/100 & 73/100 & 59/100 \\ \hline
\multirow{4}{*}{BIC} & $\#$ SF & $3.0(\pm 0.0)$ & $3.0(\pm 0.38)$ & $2.89 (\pm 0.80)$ \\
& $||\hat{{\theta}} -{\theta}||$ & $0.16 (\pm 0.08)$ & $0.45(\pm 0.69)$ & $1.80(\pm 1.20)$ \\
& MSE$^*$ & $0.97(\pm 0.15) $ & $1.29 (\pm 0.81)$ & $ 2.48(\pm 1.17)$ \\
& $\#$ True Model & 100/100 & 94/100 & 53/100 \\ \hline
\end{tabular}
\caption{\footnotesize Performance measures for $L_0$ regularized regression with AIC and BIC over 100 simulations with $n=100$, and $m =1000$, where values in the parenthesis are the standard deviations, and $\#$ SF: number of average selected features; MSE$^*$: In-sample average mean squared error; $||\hat{{\theta}} -{\theta}||$: average absolute bias when comparing true and estimated parameters.}
\end{center}
\end{table}
Table 4 shows that $L_0$ regularized regression with AIC and BIC performs very well, when compared with the results from computationally intensive cross-validation in Table 3. Without correlation, BIC identifies the true model ($100\%$), which is the same as cross-validation in Table 3, and better than AIC's $78\%$. The bias of BIC (0.16) is only slightly higher than that of cross-validation (0.14), but lower than that of AIC (0.19). Even though MSE$^*$s with AIC and BIC are in-sample mean squared errors, which are not comparable to the test MSE with cross validation, larger MSE$^*$ with BIC indicates that BIC is an more stringent criteria than AIC and selects less variables. With mild correlation ( r = 0.3) and some sacrifices in bias and MSE$^*$ , BIC seems to perform the best in variable selection, since the average number of features selected is exactly 3 and $94\%$ of the simulations recognize the true model, while AIC chooses more features (3.72) than necessary and only $73\%$ of the simulations are right on targets. Cross validation is the most tight measure with 2.9 features on average and $75\%$ of the simulations finding the correct model. When the correlations among the variables are high (r= 0.6), the results are mixed. Both BIC and AIC correctly identify more than half of the true models, while cross validation only recognizes $25\%$ (5/20) of the model correctly. Therefore, comparing with the computationally intensive cross validation, both BIC and AIC perform reasonable well. The computational results of BIC is comparable to the results of cross validation, while the computational time is only $1/500$ of the time for cross validation, if the free-parameter $ {\lambda} _{opt}$ is chosen from 100 candidate $ {\lambda} $s with 5-fold cross validation.
\subsection*{Simulations for graphical models}
One important application of $L_0$ regularized regression is to detect high-order correlation structures, which has numerous real-world applications including gene network analysis. Given a matrix $X$, letting $\mathbf{x}_j$ be the $j$th variable, and $X_{-j}$ be the remaining variables, we have $P(\mathbf{x}_j|X_{-j}) \sim N(X_{-j}{\theta}, {\sigma}^2),$ where the coefficients ${\theta}$ measures the partial correlations between $\mathbf{x}_j$ and the rest variables. Therefore, the high-order structure of X has been determined via a series of $L_1$ regularized regression for each $\mathbf{x}_j$ with the remaining variables $X_{-j}$ (Peng {et al.} 2009; Liu $\&$ Ihler, 2011). The collected regression nonzero coefficients are the edges on the graph. The drawback of such approach is computationally intensive, because the regularized parameter $ {\lambda} $ for $L_1$ have to be determined through cross validation. For instance, given a matrix $X$ with 100 variables, to find the optimal $ {\lambda} _{opt}$ from 100 candidate $ {\lambda} $s with 5-fold cross validation, 500 models need to be evaluated for each variable $\mathbf{x}_j$. Therefore a total of $500\times 100 = 50000$ models have to be estimated to detect the dependencies among $X$ with LASSO. It usually takes hours to solve this problem. However, only 100 models are required to identify the same correlation structure with $L_0$ regularized regression and AIC or BIC. Solving such a problem with $L_0$ without cross-validation only takes less than one minute. Finally, negative correlations between genes are difficult to confirm and seemingly less ‘biologically
relevant’ (Lee {et al.}, 2004). Most national databases are constructed with similarity (dependency) measures. it is straight forward to study only the positive dependency by simply setting ${\theta}({\theta} <0) = 0$ in the EM algorithm.
We simulate two network structures similar to those in Zhang $\&$ Mallick (2013) (i) Band 1 network, where $\Sigma$ is a covariance matrix with ${\sigma}_{ij} = 0.6^{|i-j|}$, so $A =\Sigma^{-1}$ has a band 1 network structure, and (ii) A more difficult problem for a Band 2 network with weaker correlations, where $A = -\Sigma^{-1}$ with $a_{ij} =\left\{\begin{array}{rl}
0.25 & \textrm{ if } |i-j| = 1,\\
0.4 & \textrm{ if } |i-j| =2, \\
0 & \textrm{ Otherwise. }
\end{array} \right. $
The sample sizes are n = 50, 100, and 200, respectively and the number of variables is $m = 100$. $L_0$ regularized regression with AIC and BIC is used to detect the network (correlation) structure. The consistence between the true and predicted structures is measured by the area under the ROC curve (AUC), false discovery (positive) rate (FDR/FPR), and false negative rate (FNR) of edges.
The computational results are shown in Table 5.
\begin{table}[t!]
\begin{center}
\begin{tabular}{c|ccc|ccc} \hline
& \multicolumn{3}{c|}{AIC} & \multicolumn{3}{c}{BIC}\\ \hline
Band 1 & AUC & FDR($\%$) & FNR ($\%$) & AUC & FDR ($\%$) & FNR ($\%$) \\ \hline
$n=50$& $.95(\pm .01)$ & $.29 (\pm .08)$ & $9.4(\pm 2.6)$ & $.90(\pm .02)$ & $.10(\pm .05)$ & $20(\pm 3.6)$\\
100 &$.99(\pm .005)$ & $.20 (\pm .06)$ &$1.2(\pm 1.1)$ & $.991(\pm .007)$ & $.03(\pm .03)$ & $1.8 (\pm 1.3)$ \\
200& $.999(\pm .0003)$ & $.20 (\pm .05)$ & $0(\pm 0)$ & $.9999(\pm .0005)$ & $.01(\pm .01)$ &$.01(\pm .10)$ \\ \hline
Band 2 & AUC & FPR($\%$) & FNR ($\%$) & AUC & FPR ($\%$) & FNR ($\%$) \\ \hline
$n =50$ & $.82(\pm .01)$ & $.10 (\pm .05)$ & $36.7(\pm 1.5)$ & $.803(\pm .008)$ & $.02(\pm .02)$ & $ 39.3(\pm1.5)$ \\
100 & $.84(\pm.01)$ & $.11(\pm .04)$ & $32.7(\pm 1.9)$ &$.83(\pm .01)$ & $.03(\pm .02)$ & $ 34.9(\pm 1.6)$ \\
200 & $.93(\pm .01)$ & $.11(\pm .04)$ & $14.2 (\pm 2.4)$ & $.82(\pm .01)$ & $.03 (\pm .02)$ & $ 36.7 (\pm 1.8)$ \\ \hline
\end{tabular}
\caption{\footnotesize Performance measures for $L_0$ regularized regression for graphical structure detection over 100 simulations, where values in the parenthesis are the standard deviations.}
\end{center}
\end{table}
Table 5 shows that both AIC and BIC performed well. Both achieved at least 0.90 AUC for Band 1 network and 0.8 AUC for Band 2 network with different sample sizes. AIC performed slightly better than BIC, especially for Band 2 network with weak correlations and small sample sizes. This is reasonable because BIC is a heavier penalty and forces most of the weaker correlations with $a_{ij} = 0.25$ to 0. In addition, BIC has slightly larger AUCs for Band 1 network with strong correlation $r =0.6$ and larger sample size (n=100, 200). One interesting observation is that the FDRs of both AIC and BIC are well controlled. The maximal FDRs of AIC for the Band 1 and 2 networks are $0.29\%$ and $0.2\%$, while the maximal FDRs of BIC are only $0.1\%$, and $0.03\%$ respectively. Controlling false discovery rates is crucial for identifying true associations with high-dimensional data in bioinformatics. In general, AUC increases and both FDR and FNR decrease, as the sample sizes become larger, except for Band 2 network with BIC. The performance of BIC is not necessary better with large sample size, since the penalty $ {\lambda} $ increases with the sample size.
\section{Real Application}
The purpose of this application is to identify subnetworks and study the biological mechanisms of potential prognostic biomarkers for ovarian cancer with multi-source gene expression data. The ovarian cancer data was downloaded from the KMplot website(www.kmplot.com/ ovar) (Gyorffy {et al.} 2012). They originally got the data from searching Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/geo/) and The Cancer Genome Atlas (TCGA; http:// cancergenome.nih.gov) with multiple platforms. All collected datasets have raw gene expression data, survival information, and at least 20 patients available. They merged the datasets across different platforms carefully. The final data has 1287 patients samples, and 22277 probe sets representing 13435 common genes. We identified 112 top genes that are associated with patient survival times using univariate COX Regression. We constructed a co-expression network from the 112 genes with $L_0$ regularized regression and identified biologically meaningful subnetworks (modules) associated with patient survival. Network is constructed with positive correlation only and BIC. The computational time for constructing such network is less than 2 seconds. One survival associated subnetwork we identified is given in Figure 2.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.5]{OvSubnet4}
\caption{\footnotesize Subnetwork constructed with $L_0$ penalized regression, multi-source gene expression profiling, and BIC}
\end{figure}
The 22 genes on the subnetwork were then uploaded onto STRING (http://string-db.org/). STRING is an online database for exploring known and predicted protein-protein interactions (PPI). The interactions include direct (physical) and indirect (functional) associations. The predicted methods for PPI implemented in STRING include text mining, national databases, experiments, co-expression, co-occurrence, gene fusion, and neighborhood on the chromosome. The PPI network for the 22 genes are presented in Figure 3.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.6]{ppnet2}
\caption{\footnotesize Known and predicted protein protein interactions with the 22 genes on the subnetwork of Figure 2, where nodes represent proteins (genes) and edges indicate the direct (physical) and indirect (functional) associations. Stronger associations are represented by thicker lines.}
\end{figure}
Comparing Figure 3 and Figure 2, We conclude that the 22 identified genes on the subnetwork of Figure 2 are functioning together and have enriched important biological interactions and associations. Ninteen out of 22 genes on the survival associated subnetwork also have interactions on the known and predicted PPI network, except for genes LRRC15, ADAM12, and NKX3-2. Even though they are not completely identical, many interactions on our subnetwork can also be verified on the PPI interaction network of Figure 3. For instance, collagen COL5A2 is the most important genes with the largest number of degrees (7) on our subnetwork. Six out of 7 genes that link to COL5A2 also have direct edges on the PPI network. Those direct connected genes (proteins) include FAP, CTSK, VCAN, COL1A1, COL5A1, and COL11A1. The remaining gene SNAI2 was indirectly linked to COL5A2 through FBN1 on the PPI network. In addition, one of the other important genes with the degree of the node (6) is Decorin (DCN). 4 out of 6 genes directly connected to DCN on our subnetwork were confirmed on the PPI network, including FBN1, CTSK, LUM, and THBS2. The remain two genes (SNAI2, and COLEC11) are indirectly connected to DCN on the PPI network. As indicated on Figure 2, the remaining 5 important genes with degree of node 4 are POSTN, CTSK, COL1A1, COL5A1, and COL10A1, and 8 genes with degree of node 3 are FBN1, LUM, LRRC15, COL11A1, THBS2, SPARC, COL1A2, and FAP, respectively. Furthermore, those 22 genes are involved in the biological process of GO terms, including extracellular matrix organization and disassembly and collagen catabolic, fibril, and metabolic processes. They are also involved in several important KEGG pathways including ECM-receptor interaction, Protein Digestion and Absorption, Amoebiasis, Focal Adhesion, and TGF-beta Signaling pathways. Finally, a large proportion of the 22 genes are known to be associated with poor overall survival (OS) in ovarian cancer. For instance, VCAN and POSTN were demonstrated in \textit{vitro} to be involved in ovarian cancer invasion induced by TGF-$ {\beta} $ signaling (Yeung {et al.}, 2013), and COL11A1 was shown to increase continuously during ovarian cancer progression and to be highly over-expressed in recurrent metastases. Knockdown of COL11A1 reduces migration, invasion, and tumor progression in mice (Cheon {et al.} 2014). Other genes such as FAP, CTSK, FBN1, THBS2, SPARC, and COL1A1 are also known to be ovarian cancer associated (Riester {et al.}, 2014; Zhao {et al.}, 2011; Zhang {et al.}, 2013; Gardi {et al.}, 2014; Tang $\&$ Feng 2014; Yu {et al.}, 2014). Those genes contribute to cell migration and the progression of tumors and may be potential therapeutic targets for ovarian cancer. Further studies with the rest of the genes on the subnetwork are required to explore their biological mechanisms and potential clinical applications.
\section{Conclusions}
We proposed an efficient EM algorithm for variable selection with $L_0$ regularized regression. The proposed algorithm finds the optimal solutions of $L_0$, through solving a sequence of $L_2$ based ridge regressions. Given an initial solution, the algorithm will be guaranteed to converge to a unique solution under mild conditions, and the EM algorithm will be closer to the optimal solution after each iteration. Asymptotic properties, namely consistency and oracle properties are established under mild conditions. Our method apply to fixed, diverging, and ultra-high dimensional problems. We compare the performance of $L_0$ regularized regression and LASSO with simulated low and high dimensional data. $L_0$ regularized regression outperforms LASSO by a substantial margin under different correlation structures. Unlike LASSO, which selects more features than necessary, $L_0$ regularized regression chooses the true model with high accuracy, less bias, and smaller test MSE, especially when the correlation is weak. Cross-validation with the computation of the entire regularization path is computationally intensive and time consuming. Fortunately $L_0$ regularized regression does not require it. The optimal $ {\lambda} _{opt}$ can be directly determined from AIC, BIC, and RIC. Those criteria are optimal under appropriate conditions. We demonstrate that both AIC and BIC performed well when compared to cross-validation. Therefore, there is a big computational advantage of $L_0$, especially with BIGDATA. In addition, We demonstrate that $L_0$ regularized regression controls the false discovery (positive) rate (FDR) well with both AIC and BIC with the simulation of graphical models. The FDR is very low under different sample sizes with both AIC and BIC. Controlling FDR is crucial for biomarker discovery and computational biology, because further verifying the candidate biomarkers is time-consuming and costly. We applied our proposed method to construct a network for ovarian cancer from multi-source gene-expression data, and identified a subnetwork that is important both biologically and clinically. We demonstrated that we can identify biologically important genes and pathways efficiently. Even though we demonstrated our method with gene expression data, the proposed method can be used for RNA-seq, and metagenomic data, given that the data are appropriately normalized.
\section*{Appendix}
The proposed approach for $L_0$ regularized regression method can be extended to a general $L_p$ $p \in [0, 2]$ naturally. Mathematically, the general $L_p$ problem can be defined as:
\[ E = \frac{1}{2}||\mathbf{y} - X{\theta}||^2 + \frac{ {\lambda} }{2}\sum_{j=1}^m|{\theta}|^p, \]
which is equivalent to
\begin{align*} E &= \frac{1}{2}||\mathbf{y} -X{\theta}||^2 + \frac{ {\lambda} }{2} \sum_{j \in m} \frac{{\theta}_j^2}{\eta_j^{2-p}} \\
\eta & ={\theta}.
\end{align*}
Similar ideas in the manuscript can be used to get the the following equation for the general $L_p$EM method:
\[ \eta^{2-p}\odot\frac{\partial E}{\partial {\theta}}= {\lambda} {\theta} -\eta^{2-p}\odot X^t(\mathbf{y} - X{\theta}) = {\lambda} {\theta} - X_{\eta}^t(\mathbf{y} - X{\theta}) = 0, \]
where $X_{\eta}^t = [\eta^{2-p}\odot\mathbf{x}_1^t, \ldots, \eta^{2-p}\odot\mathbf{x}_n^t]$.
Solving Equation (\ref{eq9}), we have the following explicit solution.
\begin{align*}
{\theta} &= (X_{\eta}^tX + {\lambda} I) ^{-1}X_{\eta}^t\mathbf{y} \\
\eta &= {\theta},
\end{align*}
The general $L_p$EM algorithm is as follows:
\begin{tabular}{l}
$L_p$\textbf{EM Algorithm:}\\\hline
Given a $0< {\lambda} \le {\lambda} _{\max}$,and $p \in [0, 2]$, small numbers $\epsilon$ and $\varepsilon$,\\
and training data $\{X, \mathbf{y}\}$,\\
Initializing ${\theta} =(X^tX + {\lambda} I)^{-1}X^t\mathbf{y}$, \\
While 1,\\
\;\;\;\;\; E-step: $ \mathbf{\eta} = {\theta} $ \\
\;\;\;\;\; M-step: $X_{\eta}^t = \eta^{2-p}\odot X^t = [\eta^{2-p}\odot\mathbf{x}_1 ,\ldots, \eta^{2-p}\odot \mathbf{x}^t_n]$\\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\text{ }\text{ } \text{ }$ {\theta} = (X_{\eta}^tX + {\lambda} I) ^{-1}X_{\eta}^t\mathbf{y} $\\
\;\;\;\;\;\;if $||{\theta} -\eta|| < \varepsilon$, Break; End\\
End \\
${\theta}(|{\theta}| <\epsilon) = 0$. \\ \hline\\
\end{tabular}
\subsection*{References}
\begin{description}
\item[ ]Akaike, H. (1974). A new look at the statistical model identification. IEEE T. Automat. Contr. 19, 716–723.
\item[ ]Bunea, F., Tsybakov, A, and Wegkamp, M. (2007). Sparsity oracle inequalities for the Lasso.
{\sl Electron. J. Stat.}, {\bf 1}, 169-194.
\item[ ]Candes, E.J. and Plan, Y. (2009). Near-ideal model selection by $l_1$ minimization. {\sl Ann. Statist.}, {\bf 37}(5A), 2145-2177.
\item[ ]Cheon DJ, Tong Y, Sim MS, Dering J, Berel D, Cui X, Lester J, Beach JA, Tighiouart M, Walts AE, Karlan BY, Orsulic S. (2014), A collagen-remodeling gene signature regulated by TGF-β signaling is associated with metastasis and poor survival in serous ovarian cancer. Clin Cancer Res. 2014 Feb 1;20(3):711-23. doi: 10.1158/1078-0432.
\item[ ]Chi Z. (2009). $L_0$ regularized estimation for nonlinear models that have sparse underlying linear structures. arXiv:0910.2517v1 [math.ST] 14 Oct 2009.
\item[ ]Dicker, L., Huang, B. and Lin, X. (2012). Variable selection and estimation with the seamless-$L_0$
penalty. Statistica Sinica. In press. doi: 10.5705/ss.2011.074.
\item[ ]Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. J.
Am. Stat. Assoc. 96, 1348–1361.
\item[ ]Foster, D. and George, E. (1994). The risk inflation criterion for multiple regression. Ann. Statist. 22, 1947–1975.
\item[ ]Gardi NL, Deshpande TU, Kamble SC, Budhe SR, Bapat SA. (2014), Discrete molecular classes of ovarian cancer suggestive of unique mechanisms of transformation and metastases. Clin Cancer Res. 2014 Jan 1;20(1):87-99.
\item[ ]Gyorffy B, L$\acute{a}$nczky A, Sz$\acute{a}$ll$\acute{a}$si Z. (2012), Implementing an online tool for genome-wide validation of survival-associated biomarkers in ovarian-cancer using microarray data from 1287 patients. Endocr Relat Cancer. 2012 Apr 10;19(2):197-208. doi: 10.1530/ERC-11-0329.
\item[ ]Lee HK, Hsu AK, Sajdak J, Qin J, Pavlidis P (2004), Coexpression analysis of human genes across many microarray data sets. Genome Res. 2004 Jun;14(6):1085-94.
\item[ ] Lin, D., Foster, D. P., $\&$ Ungar, L. H. (2010). A risk ratio comparison of l0 and l1 penalized regressions. University of Pennsylvania, Tech. Rep.
\item[ ]Liu H, Roeder K, and Wasserman L. (2010) Stability approach to regularization selection for high dimensional graphical models. Advances in Neural Information Processing Systems, 2010.
\item[ ]Liu Q, Ihler A (2011), Learning scale free networks by reweighted $L_1$ regularization. AISTATS (2011).
\item[ ] Liu Y, Wu Y (2007), Variable Selection via A Combination of the $L_0$ and $L_1$ Penalties, Journal of Computational and Graphical Statistics, 16 (4): 782–798.
\item[ ]Liu Z, Lin S, Tan M. (2010) Sparse support vector machines with Lp penalty for biomarker identification. IEEE/ACM Trans Comput Biol Bioinform. 7(1): 100-7.
\item[ ]Lu H and Zhang Y (2013), Sparse Approximation via Penalty Decomposition Methods, SIAM Journal on Optimization, 23(4):2448-2478.
\item[ ]Mancera L and Portilla J. $L_0$ norm based Sparse Representation through Alternative Projections. In Proc. ICIP, 2006
\item[ ]Mazumder R, Friedman, JH, Hastie T (2011), SparseNet : Coordinate Descent with Non-Convex Penalties, JASA, 2011.
\item[ ]Meinshausen, N. \& Bühlmann, P. Stability selection. J. R. Stat. Soc. Series B Stat. Methodol. 72, 417–473 (2010).
\item[ ]Peng,J, Wang P, Zhou N, and Zhu J. (2009), Partial correlation estimation by joint sparse regression models. JASA, 104 (486):735-746, 2009.
\item[ ]Riester M, Wei W, Waldron L, Culhane AC, Trippa L, Oliva E, Kim SH, Michor F, Huttenhower C, Parmigiani G, Birrer MJ. (2014), Risk prediction for late-stage ovarian cancer by meta-analysis of 1525 patient samples.
J Natl Cancer Inst. 2014 Apr 3;106(5). pii: dju048. doi: 10.1093/jnci/dju048.
\item[ ]Schwarz, G. (1978). Estimating the dimension of a model. Ann. Statist. 6, 461–464.
\item[ ]Tang L, Feng J. (2014), SPARC in Tumor Pathophysiology and as a Potential Therapeutic Target. Curr Pharm Des. 2014 Jun 19. [Epub ahead of print].
\item[ ]Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. J. Roy. Stat. Soc. B. 58, 267–288.
\item[ ]Yeung TL, Leung CS, Wong KK, Samimi G, Thompson MS, Liu J, et al. TGF-beta modulates ovarian cancer invasion by upregulating CAF-derived versican in the tumor microenvironment. Cancer Res 2013;73:5016–28.
\item[ ]Yu PN, Yan MD, Lai HC, Huang RL, Chou YC, Lin WC, Yeh LT, Lin YW. (2014), Downregulation of miR-29 contributes to cisplatin resistance of ovarian cancer cells. Int J Cancer. 2014 Feb 1;134(3):542-51.
\item[ ]Zhang W, Ota T, Shridhar V, Chien J, Wu B, Kuang R. (2013), Network-based survival analysis reveals subnetwork signatures for predicting outcomes of ovarian cancer treatment. PLoS Comput Biol. 2013;9(3):e1002975. doi: 10.1371/journal.pcbi.1002975.
\item[ ]Zhao G, Chen J, Deng Y, Gao F, Zhu J, Feng Z, Lv X, Zhao Z (2011), Identification of NDRG1-regulated genes associated with invasive potential in cervical and ovarian cancer cells. Biochem Biophys Res Commun. 2011 Apr 29;408(1):154-9.
\item[ ]Zou, H. (2006). The adaptive lasso and its oracle properties. J. Am. Stat. Assoc. 101, 1418–1429.
\item[ ]Zou, H. and Zhang, H. (2009). On the adaptive elastic-net with a diverging number of parameters. Ann.
Statist. 37, 1733–1751.
\item[ ]Zhang, C. (2010). Nearly unbiased variable selection under minimax concave penalty. Ann. Statist. 38,
894–942.
\end{description}
\end{document}
|
\section{Introduction}
Applications of cohomology to deformations of $C^*$-algebras and von Neumann algebras have been studied
for decades, and yet they remain an active area of research in this field. Amongst more recent contributions,
we would like to mention the works of Buss and Exel on inverse semigroups, \cite{BE}, and of Kumjian, Pask
and Sims on higher-rank graphs, \cite{KPS}. Often deformation of the $C^*$-algebra is related to a
cohomological perturbation of another underlying object. A typical example of such process comes from
a {\em twisted} (semi)group action leading to the {\em twisted} crossed product.
In the present paper, we introduce a cohomology theory for product systems of Hilbert bimodules over
discrete semigroups, as defined by Fowler in \cite{F}. Interestingly, better understanding of twisting of
semigroup actions was one of the motivations behind the very introduction of such product systems, \cite{FR}.
In Section 3, we take the classical point of view, \cite{B,H}, and define cohomology groups of a product
system $X$ via the $\ext$ functor applied to a suitable module $\Mf$ of a ring $\Rf$ naturally associated
with $X$. First examples include cohomologies of groups, graphs, and certain product systems
arising from semigroup actions on abelian groups.
In Section 4, we restrict attention to a certain class of product systems arising from irreversible algebraic dynamics,
corresponding to actions of discrete semigroups $P$ on compact groups.
For such product systems, we construct explicitly a free resolution of module $\Mf$ and thus obtain
working formulae for cocycles and coboundaries. The construction of the
resolution takes advantage of the fact that all fibers $X_p$
of these systems $X= \bigsqcup_{p\in P}X_{p}$
are free modules over the coefficient $C^*$-algebra $A$. To each $2$-cocyle $\xi$
we associate a twisted product system $X^\xi$. The twisting is obtained by perturbing multiplication between
the fibers. Then each twisted product system $X^\xi$ gives rise to several $C^*$-algebras, including the
Toeplitz algebra $\T(X^\xi)$ and the Cuntz-Pimsner algebra $\OO(X^\xi)$. These algebras may be considered
twisted versions of the Toeplitz algebra $\T(X)$ and the Cuntz-Pimsner algebra $\OO(X)$, respectively,
associated with the original product system $X$.
In Section 5, we test this deformation procedure on the product system $X$ whose Cuntz-Pimsner algebra
$\OO(X)$ coincides with Cunt's algebra $\Q_\Nb$ associated to the $ax+b$-semigroup over $\Nb$, see
\cite{C,HLS1}. We look at certain numerical $2$-cocycles $\xi$ and show that the corresponding twisted
$C^*$-algebras $\OO(X^\xi)$ are purely infinite and simple.
\medskip\noindent
{\bf Acknowledgements.}
The first named author would like to thank faculty and staff of the Department of Mathematics and
Computer Science at the University of Southern Denmark in Odense for their warm hospitality during her
sabbatical stay there in 2013--2014. The third named author is grateful to Alex Kumjian for useful
discussions of graph cohomology, and to Nicolai Stammeier for valuable discussions and information on
irreversible algebraic dynamics and their product systems.
\section{Preliminaries on product systems}\label{Prelim-1}
Let $A$ be a $C^*$-algebra and $X$ be a complex vector space with a right action of $A$.
Suppose there is an $A$-valued inner product $\langle \cdot , \cdot \rangle _A$ on $X$ which is
conjugate linear in the first variable and satisfies
\begin{description}
\item{(i)} $\langle\xi, \eta\rangle _A =\langle\eta , \xi\rangle_A^*$,
\item{(ii)} $\langle\xi, \eta\cdot a\rangle_A=\langle\xi, \eta\rangle_A\, a$,
\item{(iii)} $\langle\xi, \xi\rangle_A \geq 0 $, and $\langle\xi, \xi\rangle_A=0 \; \Longleftrightarrow\; \xi=0$,
\end{description}
for $\xi,\eta\in X$ and $a\in A$. Then $X$ becomes a right Hilbert $A$-module
when it is complete with respect to the norm given by
$\|\xi\|:=\|\langle\xi,\xi\rangle _A\|^{\frac{1}{2}}$ for $\xi\in X$.
A module map $T:X\rightarrow X $ is said to be adjointable if there is a map $T^*:X\rightarrow X$
such that
$$
\langle T\xi,\zeta\rangle_A=\langle\xi,T^*\zeta\rangle_A
$$
for all $\xi,\eta\in X$. An
adjointable map is automatically norm-bounded, and the set $\LL(X)$ of all adjointable
operators on $X$ endowed with the operator norm
is a $C^*$-algebra. The rank-one operator $\theta_{\xi, \eta}$ defined on $X$ as
$$
\theta_{\xi,\eta}(\zeta)=\xi\langle \eta,\zeta\rangle_A \; \; \text{for} \; \xi, \eta, \zeta\in X,
$$
is adjointable and we have $\theta_{\xi,\eta}^*=\theta_{\eta,\xi}$. Then
$\K(X)=\clsp\{\theta_{\xi, \eta}\mid \xi, \eta\in X\}$ is the ideal of compact operators in $\LL(X)$.
Suppose $X$ is a right Hilbert $A$-module. A $*$-homomorphism
$\varphi:A\rightarrow\LL(X)$ induces a left action of $A$
on a $X$ by $a\xi:=\varphi(a)\xi$, for $a\in A$ and $\xi\in X$.
Then $X$ becomes a right Hilbert $A$--$A$-bimodule (or
$C^*$-correspondence over $A$). The standard bimodule $_AA_A$ is
equipped with $\langle a, b\rangle _A=a^*b$, and the right and left actions are simply given
by right and left multiplication in $A$, respectively.
For right Hilbert $A$--$A$-bimodules $X$ and $Y$, the balanced tensor product $X\otimes_AY$ becomes
a right Hilbert $A$--$A$-bimodule with the right action from $Y$, the left action implemented by the
homomorphism $A\ni a \mapsto\varphi(a)\otimes_A \id_Y\in\LL(X\otimes_A Y)$, and the $A$-valued
inner product given by
$$
\langle\xi_1\otimes_A\eta_1 , \xi_2\otimes_A\eta_2\rangle_A=
\langle \eta_1, \langle\xi_1,\xi_2\rangle_A\cdot\eta_2\rangle_A,
$$
for $\xi_i\in X$ and $\eta_i\in Y$, $i=1,2$.
Let $P$ be a multiplicative semigroup with identity $e$, and let $A$ be a $C^*$-algebra.
For each $p\in P$ let $X_p$ be a complex vector space.
Then the disjoint union $X := \bigsqcup_{p\in P}X_{p}$ is a {\em product system} over $P$
if the following conditions hold:
\begin{description}
\item{(PS1)} For each $p\in P\setminus\{e\}$, $X_p$ is a right Hilbert $A$--$A$-bimodule.
\item{(PS2)} $X_e$ is the standard bimodule $_AA_A$.
\item{(PS3)} $X$ is a semigroup such that $\xi\eta\in X_{pq}$ for $\xi\in X_p$ and $\eta\in X_q$,
and for $p, q\in P\setminus\{e\}$, this product extends to an isomorphism
$F^{p, q} : X_p\otimes_A X_q\rightarrow X_{pq}$ of right Hilbert $A$--$A$-bimodules.
If $p$ or $q$ equals $e$ then the corresponding product in $X$ is induced by the left
or the right action of $A$, respectively.
\end{description}
\begin{remark}\rm
For $p\in P$, there are maps
$F^{p, e}:X_p\otimes_A X_e\rightarrow X_p$ and $F^{e, p}:X_e\otimes_A X_p\rightarrow X_p$
by multiplication, ie $F^{p, e}(\xi\otimes a)=\xi\, a$ and $F^{e, p}(a\otimes\xi)=a\, \xi$
for $a\in A$ and $\xi\in X_p$. Note that $F^{p,e}$ is always isomorphism. However, $F^{e,p}$ is
isomorphism only if $\overline{\varphi(A)X_p}=X_p$ or, in the terminology from \cite{F}, if $X_p$ is
essential.
\end{remark}
For each $p\in P$, we denote by $\langle\cdot,\cdot\rangle_p$ the $A$-valued
inner product on $X_p$ and by $\varphi_p$ the $*$-homomorphism from $A$ into $\LL(X_p)$.
Due to associativity of the multiplication on $X$, we have
$\varphi_{pq}(a)(\xi\eta)=(\varphi_p(a)\xi)\eta$ for all $\xi\in X_p$, $\eta\in X_q$, and $a\in A$.
For each pair $p,q\in P\setminus\{e\}$,
the isomorphism $F^{p,q} : X_p\otimes_A X_q\rightarrow X_{pq}$
allows us to define a $*$-homomorphism $i_p^{pq}:\LL(X_{p})\to\LL(X_{pq})$ by
$i_p^{pq}(S)=F^{p,q}(S\otimes_A I_q)(F^{p,q})^*$ for $S\in\LL(X_p)$.
In the case $r\neq pq$ we define $i_p^r:\LL(X_p)\to\LL(X_r)$ to
be the zero map $i_p^r(S)=0$ for all $S\in\LL(X_p)$. Further, we let $i_e^q=\varphi_q$.
\smallskip
Let $X=\sqcup_{p\in P}X_p$ be a product system over $P$ of right Hilbert $A$--$A$-bimodules.
A map $\psi$ from $X$ to a $C^*$-algebra $C$ is a {\em Toeplitz
representation} of $X$ if the following conditions hold:
\begin{description}
\item{(T1)} for each $p\in P\setminus\{e\}$, $\psi_p:=\psi\vert_{X_p}$ is linear,
\item{(T2)} $\psi_e:A\longrightarrow C$ is a $*$-homomorphism,
\item{(T3)} $\psi_p(\xi)\psi_q(\eta)=\psi_{pq}(\xi\eta)\; \; $ for $ \; \xi\in X_p$, $\eta\in X_q$, $p,q\in P$,
\item{(T4)} $\psi_p(\xi)^*\psi_p(\eta)=\psi_e(\langle\xi, \eta\rangle_p)$ for $\xi, \eta\in X_p$.
\end{description}
As shown in \cite{P}, for each $p\in P$ there exists a $*$-homomorphism
$\psi^{(p)} : \K(X_p) \longrightarrow C$ such that $\psi^{(p)}(\theta_{\xi,\eta})=
\psi_p(\xi)\psi_p(\eta)^*\, , \; \text{for}\; \xi,\eta\in X_p$. A Toeplitz representation $\psi$ is
{\it Cuntz-Pimsner covariant}, \cite{F}, if
\begin{description}
\item{(CP)} $\psi^{(p)}(\varphi_p(a))=\psi_e(a)$ for $a\in\varphi_p^{-1}(\K(X_p))$ and all $p\in P$.
\end{description}
The Toeplitz algebra $\T(X)$ associated to the product system $X$ was defined by Fowler as the universal
$C^*$-algebra for Toeplitz representations, \cite{F}. The Cuntz-Pimsner algebra $\OO(X)$ is
universal for the Cuntz-Pimsner covariant Toeplitz representations. A number of other related constructions
exist in the literature, we do not discuss in here. However, we would like to mention co-universal
algebras studied by Carlsen, Larsen, Sims and Vittadello in \cite{CLSV}, and reduced Cuntz-Pimsner
algebras investigated by Kwa\'{s}niewski and Szyma\'{n}ski in \cite{KS}.
\section{A cohomology for product systems}\label{cohodefi}
Let $X$ be a product system of Hilbert bimodules over a semigroup $P$ and with the coefficient
(unital) $C^*$-algebra $A$. Then the direct sum of $A-A$-bimodules
\begin{equation}\label{r}
\Rf := \bigoplus_{p\in P} X_p
\end{equation}
becomes a ring graded over $P$ with the multiplication borrowed from $X$. We assume that there exists
a unital $A$-bimodule map $\Psi:\Rf\to A$ such that
\begin{equation}\label{psiidentity}
\Psi(xy) = \Psi(x\Psi(y)),
\end{equation}
for all $x,y\in\Rf$. Then $A=X_e$ becomes a left $\Rf$-module, with the $\Rf$-action $\rh$ given
by the composition of the multiplication in $\Rf$ with $\Psi$, i.e.
\begin{equation}\label{action}
x\rh a := \Psi(xa),
\end{equation}
for $x\in\Rf$, $a\in A$. We denote this module $\Mf$ and define the $n^{\rm th}$-cohomology group
of the product system $X$ relative to $\Psi$ as
\begin{equation}\label{hnext}
H_\Psi^n(X) := \ext_\Rf^n(\Mf,\Mf),
\end{equation}
cf. \cite{H}, \cite{B}.
\begin{example}\label{cohgraphs}\rm
Let $E$ be a {\em finite} directed graph, with vertices $E^0$, edges $E^1$, and range and source
mappings $r:E^1\to E^0$ and $s:E^1\to E^0$, respectively. Let $X_1$ be the standard Hilbert bimodule
associated with $E$, \cite{K}, with the finite-dimensional coefficient algebra $A$ generated by vertex projections.
Let $X$ be the product system over the additive semigroup $\Nb$ generated by $X_1$.
For each $n\in\Nb$, $X_n$ is the $\Cb$-span of paths of length $n$ (paths of length zero being vertices). Multiplication in ring $\Rf$ is simply given by concatenation of directed paths. For a path $\mu$ we set
$\Psi(\mu):=s(\mu)$. Then for a path $\mu$ and a vertex $v$ we have $\mu \rightharpoonup v = s(\mu)$ if
$v=r(\mu)$, and $0$ otherwise.
\end{example}
\begin{example}\label{groupcoho}\rm
Let $G$ be a countable group. We set $P=G$ and $X_g=\Cb g$ for all $g\in G$.
Then $X$ is a product system with the usual group algebra multiplication and the inner products
$\la zg, wg \ra_g = \overline{z}w1$, for $g\in G$ and $z,w\in\Cb$.
We have $\Rf=\Cb G$, the usual complex group algebra. In this case, $\Psi$ is the
trivial representation of $\Cb G$ and $\Mf$ is the trivial module.
\end{example}
\begin{example}\label{cohcuntz}\rm
Here we consider a product system studied in \cite{L} in connection with Exel's approach to semigroup
crossed products via transfer operator, \cite{E}, and in \cite{Y} and \cite{HLS1} in connection with Cuntz's
algebra $\Q_\Nb$, \cite{C}. The product system $X$ is over the multiplicative semigroup $\Nb^\times$.
The coefficient algebra $A$ is $C(\Tb)$,
and each fiber $X_p$ is a free left $A$-module of rank one with a basis vector $\mathbbm{1}_p$. The right
action of $A$ is determined by $\mathbbm{1}_p a=\alpha_p(a)\mathbbm{1}_p$, where $\alpha_p:A\to A$
is an endomorphism such that $\alpha_p(a)(z)=a(z^p)$ for $a\in A$ and $z\in\Tb$. The inner product in
fiber $X_p$ is given by $\la a\mathbbm{1}_p, b\mathbbm{1}_p \ra_p = L_p(a^*b)$, where $L_p:A\to A$ is
a transfer operator for $\alpha_p$ such that $L_p(a)(z)=\frac{1}{p}\sum_{w^p=z}a(w)$. Fibers are multiplied
according to the rule $(a\mathbbm{1}_p)(b\mathbbm{1}_q)=a\alpha_p(b)\mathbbm{1}_{pq}$.
It is shown in
\cite[Lemma 3.1]{HLS1} that the left action of $A$ on each fiber is by compact operators. In fact,
this product system belongs to the class of singly generated product systems of finite type, as introduced
in \cite[Definition 3.5]{HLS2}.
We set $\Psi(a\mathbbm{1}_p):=a$, for $p\in\Nb^\times$ and $a\in A$. Then the action of $\Rf$ on $\Mf$ is determined by $\mathbbm{1}_p\rightharpoonup a=\alpha_p(a)$, for $p\in\Nb^\times$ and $a\in\Mf$.
\end{example}
\section{Irreversible algebraic dynamics}
In this section, we consider irreversible dynamical systems corresponding to injective homomorphisms
of abelian groups. We follow the approach of Stammeier, \cite{St} (see also \cite{BLS}), building on the
works of Exel and Vershik, \cite{EV}, Cuntz and Vershik, \cite{CV}, and Carlsen and Silvestrov, \cite{CaSil}.
In particular, we use the description of the product systems naturally arising from such dynamics,
due to Stammeier, \cite{Storal}.
Let $G$ be a countable abelian group, and let $P$ be a semigroup with identity $e$.
Let $\theta$ be an action of $P$ on $G$ by injective group homomorphisms.
We denote by $A:=C^*(G)$ the group $C^*$-algebra of $G$. For each $p\in P$ let
$X_p$ be a free left $A$-module of rank one with a basis element $\mathbbm{1}_p$. The right
action of $A$ on $X_p$ is determined by $\mathbbm{1}_p a=\theta_p(a)\mathbbm{1}_p$, $a\in A$.
The inner product in $X_p$ is defined as
\begin{equation}\label{condexpsubgroup}
\la a\mathbbm{1}_p, b\mathbbm{1}_p\ra_p := \theta_p^{-1}E_p(a^*b),
\end{equation}
for $p\in\Nb^\times$ and $a,b\in A$. Here $E_p:C^*(G)\to C^*(\theta_p(G))$ is the conditional
expectation given by restriction. For $a=g$ and $b=h$ with $g,h\in G$, this yields
\begin{equation}\label{Stinnerpr}
\la g\mathbbm{1}_p, h\mathbbm{1}_p\ra_p = \left\{ \hspace{-2mm}\begin{array}{ll}
{\theta_p^{-1}(g^{-1}h)} & \text{if } g^{-1}h\in\theta_p(G), \\ 0 &
\text{otherwise}. \end{array} \right.
\end{equation}
If index $[G:\theta_p(G)]$ is finite, then in the dual picture, with $\hat{\theta}_p$
acting on $C(\widehat{G})$, this inner product corresponds to the transfer operator given by averaging
over the finitely many inverse image points, \cite{Storal}. Finally, fibers are multiplied according to the rule
\begin{equation}\label{fibermulti}
(a\mathbbm{1}_p)(b\mathbbm{1}_q)=a\theta_p(b)\mathbbm{1}_{pq}.
\end{equation}
In this case, ring $\Rf$ is the skew product $\Zb G\rtimes_\theta P$, with multiplication
$$ (gp)(hq) = (g\theta_p(h))(pq), $$
$g,h\in G$, $p,q\in P$. We take $\Psi(g{\mathbbm 1}_p):=g$,
$g\in G$, $p\in P$. Then the action of $\Rf$ on $\Mf$ is given by
$$ (g{\mathbbm 1}_p) \rh h = g\theta_p(h), $$
$g,h\in G$, $p\in P$. Example \ref{cohcuntz} from Section 3 arises as a special case of this construction.
\smallskip
Now, we describe an acyclic, free resolution of the $\Rf$-module $\Mf$.
To this end, we define a complex of $\Rf$-modules and maps
\begin{equation}\label{resolution}
\ldots \overset{\partial_2}{\longrightarrow} \Ff_2
\overset{\partial_1}{\longrightarrow} \Ff_1
\overset{\partial_0}{\longrightarrow} \Ff_0
\overset{\partial_{-1}}{\longrightarrow} \Mf
\longrightarrow 0,
\end{equation}
as follows. We let $\Ff_0$ be a free left $\Rf$-module of rank $1$ with a basis element $[\;\;]$.
For $n\geq 1$, we let $\Ff_n$ be a free left $\Rf$-module with a basis
\begin{equation}\label{basis}
\{ [p_1,\ldots,p_n] \mid p_k\in P, k=1,\ldots,n\}.
\end{equation}
The maps $\partial_*$ are defined as $\Rf$-module homomorphisms such that
$$ \begin{aligned}
\partial_{-1}([\;\;]) & = 1, \\
\partial_0([p]) & = (\mathbbm{1}_p - 1)[\;\;],
\end{aligned} $$
and for $n\geq 2$ we set
$$ \begin{aligned}
\partial_{n-1}( [p_1,\ldots,p_n]) & = \mathbbm{1}_{p_1}[p_2,\ldots,p_n] \\
& + \sum_{i=1}^{n-1}(-1)^i [p_1,\ldots,p_{i-1}, p_i p_{i+1},p_{i+2},\ldots,p_n] \\
& + (-1)^n[p_1,\ldots,p_{n-1}].
\end{aligned} $$
A routine calculation shows that
$$ \partial_n\partial_{n+1}=0 $$
for all $n\geq -1$.
To show that complex (\ref{resolution}) is acyclic, we construct splitting homotopies.
That is, we define abelian group homomorphisms $h_{-1}:\Mf\to\Ff_0$ and $h_n:\Ff_n\to\Ff_{n+1}$,
$n\geq 0$, such that
$$ \begin{aligned}
\partial_{-1}h_{-1} & = \id_{\Mf}, \\
\partial_n h_n + h_{n-1}\partial_{n-1} & = \id_{\Ff_n}, \;\;\; \text{for } n\geq 0.
\end{aligned} $$
For example, we may take
$$ \begin{aligned}
h_{-1}(a) & = a[\;\;], \\
h_0(a\mathbbm{1}_p[\;\;]) & = a[p], \\
h_n(a\mathbbm{1}_{p_0}[p_1,\ldots,p_n]) & = a[p_0,p_1,\ldots,p_n], \;\;\; n\geq 1,
\end{aligned} $$
for $a\in C^*(G)$.
\smallskip
Now, applying the $\hom_{\Rf}(*,\Mf)$ functor to chain complex (\ref{resolution}), with $\Mf$ deleted,
we obtain the following complex of homogeneous cochains:
\begin{equation}\label{cochains}
0 \longrightarrow \hom_{\Rf}(\Ff_0,\Mf)
\overset{\partial_0^*}{\longrightarrow} \hom_{\Rf}(\Ff_1,\Mf)
\overset{\partial_1^*}{\longrightarrow} \ldots
\end{equation}
By definition, we have
\begin{equation}\label{homogeneouscocycles}
H_\Psi^n(X) = \frac{\ker(\partial_n^*)}{\im(\partial_{n-1}^*)}.
\end{equation}
Restricting in (\ref{cochains}) elements of $\hom_{\Rf}(\Ff_n,\Mf)$ to the basis (\ref{basis})
of the free $\Rf$-module $\Ff_n$, we obtain the following complex of inhomogeneous cochains:
\begin{equation}\label{inhomogeneous}
0 \longrightarrow C^0(P,\Mf)
\overset{\partial^0}{\longrightarrow} C^1(P,\Mf)
\overset{\partial^1}{\longrightarrow} C^2(P,\Mf)
\overset{\partial^2}{\longrightarrow} \ldots
\end{equation}
Here we denote:
$$ \begin{aligned}
C^0(P,\Mf) & = \Mf, \\
C^n(P,\Mf) & = \{ \xi: \bigtimes^n P \to \Mf \}, \;\;\; n\geq 1.
\end{aligned} $$
The cochain maps are:
$$ \begin{aligned}
\partial^0(a)(p) & = \theta_p(a) - a, \\
\partial^n(\xi)(p_1,\ldots,p_{n+1}) & = \theta_{p_1}(\xi(p_2,\ldots,p_{n+1})) \\
& + \sum_{i=1}^n (-1)^i \xi(p_1,\ldots,p_{i-1}, p_i p_{i+1}, p_{i+2},\ldots,p_{n+1}) \\
& + (-1)^{n+1} \xi(p_1,\ldots,p_n),
\end{aligned} $$
for $n\geq1$, $a\in \Mf$, $\xi\in C^n(P,\Mf)$, $p$ and $p_1,\ldots,p_{n+1}\in P$. We have
\begin{equation}\label{inhomogeneouscocycles}
H_\Psi^n(X) \cong \frac{\ker(\partial^n)}{\im(\partial^{n-1})}.
\end{equation}
Now, let $\xi:P\times P\to A_{\sa}$ be a normalized (i.e. $\xi(p,q)=0$ if $p=1$ or $q=1$) $2$-cocycle with
self-adjoint values. We define a new product system $X^\xi$ over $P$ and with coefficients in $A$, as follows.
For each $p\in P$, fiber $X^\xi_p$ coincides with $X_p$ (but we denote the generator by
${\mathbbm 1}_p^\xi$ to avoid
confusion). However, the multiplication between fibers is twisted by $\xi$ according to the rule
\begin{equation}\label{twistedmulti}
(a{\mathbbm 1}_p^\xi)(b{\mathbbm 1}_q^\xi) := \exp(i\xi(p,q))a\theta_p(b){\mathbbm 1}_{pq}^\xi.
\end{equation}
It is not difficult to verify that $X^\xi$ satisfies axioms (PS1)--(PS3) of a product system, given in Section 2 above.
Consequently, the corresponding Toeplitz and Cuntz-Pimsner algebras $\T(X^\xi)$ and $\OO(X^\xi)$,
respectively, may be considered as $\xi$-twisted versions of $\T(X)$ and $\OO(X)$, respectively.
\begin{proposition}\label{trivialcocycles}
Let $\xi,\eta$ be normalized, self-adjoint $2$-cocycles such that $[\xi]=[\eta]$ in $H^2_\Psi(X)$.
Then the corresponding twisted product systems $X^\xi$ and $X^\eta$ are isomorphic.
\end{proposition}
\begin{proof}
By hypothesis, there is a $\psi:P\to \Mf$ such that $\xi-\eta=\partial^1(\psi)$. Replacing $\psi$ with
$1/2(\psi+\psi^*)$ if necessary, we may assume that $\psi(p)$ is self-adjoint for all $p\in P$. Define a map
$X^\xi\to X^\eta$ so that $a{\mathbbm 1}^\xi_p \mapsto \exp(i\psi(p))a{\mathbbm 1}^\eta_p$ for all
$p\in P$, $a\in A$. One easily verifies that this map yields the required isomorphism between $X^\xi$ and $X^\eta$.
\end{proof}
\section{Twisted ${\mathcal Q}_{\Nb}$}
In this section, we apply the twisting procedure described in Section 4 to the product system $X$ discussed
in Example \ref{cohcuntz} from Section 3. We begin by having a quick look at the $0-$,
$1-$ and $2-$cohomology groups. The $0$-cohomology is clear. Indeed, it follows from
(\ref{inhomogeneouscocycles}) that we simply have
$$ \begin{aligned}
H^0_\Psi(X) & = \{ a\in A \mid \alpha_p(a)=a, \forall p\in\Nb^\times \} \\
& = \Cb 1.
\end{aligned} $$
Now, let $\xi(1)=0$ and for each {\em prime} $p\in\Nb^\times$ let $\xi(p)\in A$ be arbitrary.
Let $1\neq q\in\Nb^\times$ have prime factorization $q=p_1\cdots p_m$, with
$p_1\leq p_2\leq\ldots\leq p_m$. Proceeding by induction on $m$, define
$\xi(q):=\alpha_{q/p_m}(\xi(p_m)) + \xi(q/p_m)$. Then $\xi:q\mapsto \xi(q)$, $q\in\Nb^\times$, is a $1$-cocycle.
For $\xi$ to be a $1$-coboundary, there must exist a function $\psi\in C(\Tb)$ such that for
all prime $p\in\Nb^\times$ and all $z\in\Tb$ we have
$$ \psi(z) = \psi(z^p) - \xi(p)(z). $$
To construct such a $\psi$, fix a prime $p$ for a moment and define $\psi(z)$ for $z\in \Tb$ such that
$z^{p^k}=1$, by induction on $k$, as follows.
$$ \begin{aligned}
\psi(1) &: = 0, \\
\psi(z) & := \psi(z^p) - \xi(p)(z).
\end{aligned} $$
In this way, $\psi$ is densely defined on $\Tb$ at all roots of unity. It follows that $\xi$ is a $1$-cocycle if and
only if $\psi$ can be extended to a continuous function on the entire cirle $\Tb$.
\smallskip
For a $2$-cocycle $\xi:\Nb^\times \times \Nb^\times \to A$,
suppose $\psi:\Nb^\times \to A$ is such that $\xi=\partial^1(\psi)$. Then for any two primes
$p,q$ we must have
$$ \begin{aligned}
\psi(pq) & = \alpha_p(\psi(q)) + \psi(p) - \xi(p,q) \\
& = \alpha_q(\psi(p)) + \psi(q) - \xi(q,p),
\end{aligned} $$
and hence
\begin{equation}\label{pq2coboundary}
(\psi(q)(z^p) - \psi(q)(z)) - (\psi(p)(z^q) - \psi(p)(z)) = \xi(p,q) - \xi(q,p)
\end{equation}
for all $z\in\Tb$. Thus, for $\xi$ to give a non-zero element in $H^2_\Psi(X)$, it suffices to have
$\xi(p,q)(1) \neq \xi(q,p)(1)$ for some primes $p$ and $q$. For a more specific example,
let $\xi:\Nb^\times \times \Nb^\times \to \Cb 1$ be a map such that
\begin{equation}\label{bicharacter}
\xi(mn,k)=\xi(m,k)+\xi(n,k) \;\;\; \text{and} \;\;\; \xi(m,nk)=\xi(m,n)+\xi(m,k).
\end{equation}
Then $\xi$ is a $2$-cocycle. For example, given two distinct primes $p$ and $q$ and complex numbers
$a,b,c,d$, we can set
\begin{equation}\label{pqbicharacter}
\xi(mp^kq^l,np^rq^j) := (ak+bl)(cr+dj),
\end{equation}
with $m,n$ relatively prime with both $p$ and $q$. By the above, if $ad\neq bc$ then $\xi$ is not
a coboundary.
\smallskip
Let $\xi:\Nb^\times \times \Nb^\times \to \Rb1$ be a $2$-cocycle defined in (\ref{pqbicharacter}), with
$a,b,c,d$ real numbers. We denote by $u$ the standard unitary generator of $A=C(\Tb)$ and for $m\in\Nb^\times$ we denote by $s_m$ the canonical image of ${\mathbbm 1}_m^\xi$ in
${\mathcal Q}_\Nb^\xi:=\OO(X^\xi)$. (Of course, $s_m$ depends also on $\xi$. We do not indicate this
explicitly to lighten the notation.)
Similarly to \cite{C} and \cite{HLS1}, each $s_m$ is an isometry
and the following relations hold:
\begin{description}
\item{(QX1)} $s_m s_n= e^{i(ak+bl)(cr+dj)}s_{mn}$,
\item{(QX2)} $s_mu^l=u^{ml}s_m$, for all $l\in\Zb$,
\item{(QX3)} $\displaystyle{\sum_{k=0}^{m-1}u^ks_ms_m^*u^{-k}=1,}$
\end{description}
where $k,r$ are the numbers of $p$-factors of $m$ and $n$, respectively, and $l,j$ are the
numbers of $q$-factors of $m$ and $n$, respectively.
\begin{proposition}\label{simplicity}
$C^*$-algebra ${\mathcal Q}_\Nb^\xi$ is simple.
\end{proposition}
A proof of simplicity of ${\mathcal Q}_\Nb^\xi$, claimed in Proposition \ref{simplicity} above, may be given
as an application of \cite[Theorem 5.10]{KS}. This requires
showing minimality and topological aperiodicity (in the sense of Definition 5.7 and Definition 5.3 of \cite{KS},
respectively) of the underlying product system $X^\xi$. Since both proofs are essentially the same as
those from \cite[Section 6.5]{KS} (treating the case of untwisted ${\mathcal Q}_\Nb$), we omit the details.
\smallskip
We want to investigate the structure of $C^*$-algebra $\Q_\Nb^\xi$ a little bit further. To this end, we note that
$X^\xi$ is a {\em regular} product system (i.e. the left action $\varphi_m$ on each fiber $X_m^\xi$ is injective
and by compacts, see \cite[Definition 3.1]{KS}) over an Ore semigroup $\Nb^\times$. Thus, it follows
from a very general argument, \cite[Lemma 3.7]{KS}, that
$$ \Q_\Nb^\xi=\clsp\{as_ms_n^*b \mid m,n\in\Nb^\times, a,b\in A\}. $$
Furthermore,
$$ \F_\Nb^\xi := \clsp\{as_ms_m^*b \mid m\in\Nb^\times, a,b\in A\} $$
is a unital $*$-subalgebra of $\Q_\Nb^\xi$. Since the $\xi$-twist does not affect $\F_\Nb^\xi$, this algebra
is unchanged by introduction of the cocycle. In fact, as shown by Cuntz in \cite[Section 3]{C}, it is a simple
Bunce-Deddens algebra with a unique trace, \cite{BD,D}.
In the present situation, since the enveloping group $\Qb_+^\times$ of $\Nb^\times$ is amenable,
$\Q_\Nb^\xi=\OO(X^\xi)$ coincides with the reduced algebra $\OO((X^\xi)^r)$, \cite{Exame} and
\cite{KS}, and with the co-universal algebra $\N\OO_{X^\xi}^r$, \cite{CLSV}. Thus, there exists
a faithful conditional expectation $E:\Q_\Nb^\xi\to\F_\Nb^\xi$ onto $\F_\Nb^\xi$ such that for all
$m,n\in\Nb^\times$, $a,b\in A$ we have
$$ E(as_ms_n^*b)=0 \text{ if } m\neq n. $$
Let $\D_\Nb^\xi$ be the $C^*$-subalgebra of $\F_\Nb^\xi$ generted by all projections $u^ks_ms_m^*u^{-k}$,
that is
$$ \D_\Nb^\xi := \clsp\{u^ks_ms_m^*u^{-k} \mid m\in\Nb^\times, k\in\Zb\}. $$
Then, as in \cite[Section 3]{C}, $\D_\Nb^\xi$ is commutative and there exists a faithful conditional
expectation $F:\F_\Nb^\xi\to\D_\Nb^\xi$ onto $\D_\Nb^\xi$ such that for all $m\in\Nb^\times$, $k,l\in\Zb$
we have
$$ F(u^ks_ms_m^*u^{-l})=0 \text{ if } k\neq l. $$
The composition $G:=F\circ E$ yields a
faithful conditional expectation from $\Q_\Nb^\xi$ onto $\D_\Nb^\xi$. We also recall from
\cite[Lemma 3.2(a)]{C}, that for all $k\in\Zb$ and $m,n\in\Nb^\times$, we have
\begin{equation}\label{cuntzequation}
u^ks_ms_m^*u^{-k} = \sum_{j=0}^{n-1}u^{k+jm}s_{mn}s_{mn}^*u^{-k-jm}.
\end{equation}
One immediate consequence of this identity is that
\begin{equation}\label{zeroequ}
s_r^*u^ts_r=0\; \text{ unless $t$ is divisible by $r$}.
\end{equation}
Another one is the identity
\begin{equation}\label{lcm}
s_ms_m^*s_ns_n^* = s_{m\vee n}s_{m \vee n}^*,
\end{equation}
where symbol $\vee$ denotes the least common multiple of two positive integers.
\begin{lemma}\label{projectioninequality}
Let $k,l\in\Zb$ and $m,n\in\Nb^\times$. Then
$$ u^ks_ms_m^*u^{-k} \leq u^ls_ns_n^*u^{-l} $$
if and only if both $m$ and $k-l$ are divisible by $n$.
\end{lemma}
\begin{proof}
By (\ref{cuntzequation}), we have
$$ \begin{aligned}
u^ks_ms_m^*u^{-k} & = \sum_{j=0}^{n-1}u^{k+jm}s_{mn}s_{mn}^*u^{-k-jm}, \\
u^ls_ns_n^*u^{-l} & = \sum_{j=0}^{m-1}u^{l+jn}s_{mn}s_{mn}^*u^{-l-jn}.
\end{aligned} $$
Thus, $ u^ks_ms_m^*u^{-k} \leq u^ls_ns_n^*u^{-l} $ if and only if for each $j\in\{0,\ldots,n-1\}$
there is a $j'\in\{0,\ldots,m-1\}$ such that $k+jm=l+j'n$ in $\Zb_{mn}$. This clearly implies the claim.
\end{proof}
Now, we will show that $C^*$-algebra $\Q_\Nb^\xi$ is purely infinite,
as in the untwisted case, \cite[Theorem 3.4]{C}.
Our proof immitates the classical argument of Cuntz, \cite{Co2}, employed also in \cite[Theorem 2.6]{CV},
and relies on the following technical lemma.
\begin{lemma}\label{cuntzlemma}
Let $Q$ be a non-zero projection in $\D_\Nb^\xi$, and let $k_0,l_0\in\Zb$, $m_0,n_0\in\Nb^\times$ be
such that either $k_0\neq l_0$ or $m_0\neq n_0$. Then there exist $k\in\Zb$ and $m\in\Nb^\times$
such that
\begin{description}
\item{(i)} $u^ks_ms_m^*u^{-k} \leq Q$, and
\item{(ii)} $(u^ks_ms_m^*u^{-k})(u^{k_0}s_{m_0}s_{n_0}^*u^{-l_0})(u^ks_ms_m^*u^{-k}) =0$.
\end{description}
\end{lemma}
\begin{proof}
By the definition of $\D_\Nb^\xi$, there exist $k'\in\Zb$ and $m'\in\Nb^\times$ such that
$u^{k'}s_{m'}s_{m'}^*u^{-k'} \leq Q$. Thus, it suffices to work with $u^{k'}s_{m'}s_{m'}^*u^{-k'}$
instead of $Q$.
If $u^{k'}s_{m'}s_{m'}^*u^{-k'}$ is not a subprojection
of either $u^{k_0}s_{m_0}s_{m_0}^*u^{-k_0}$ or $u^{l_0}s_{n_0}s_{n_0}^*u^{-l_0}$, then
to have (i) and (ii) satisfied it suffices
to take $k\in\Zb$ and $m\in\Nb^\times$ such that either $u^ks_ms_m^*u^{-k} \leq
u^{k'}s_{m'}s_{m'}^*u^{-k'}(1 - u^{k_0}s_{m_0}s_{m_0}^*u^{-k_0})$ or $u^ks_ms_m^*u^{-k}
\leq u^{k'}s_{m'}s_{m'}^*u^{-k'}(1 - u^{l_0}s_{n_0}s_{n_0}^*u^{-l_0})$, respectively.
Now, we may assume that $u^{k'}s_{m'}s_{m'}^*u^{-k'}$ is a subprojection
of both $u^{k_0}s_{m_0}s_{m_0}^*u^{-k_0}$ and $u^{l_0}s_{n_0}s_{n_0}^*u^{-l_0}$.
Thus, by virtue of Lemma \ref{projectioninequality}, both $m'$ and $k'-k_0$ are divisible by $m_0$,
while both $m'$ and $k'-l_0$ are divisible by $n_0$. Hence
$$ \begin{aligned}
(u^{k'}s_{m'}s_{m'}^*u^{-k'}) & (u^{k_0}s_{m_0}s_{n_0}^*u^{-l_0})(u^{k'}s_{m'}s_{m'}^*u^{-k'}) \\
& = u^{k'}s_{m'}s_{m'}^*s_{m_0}u^{(k_0-k')/m_0 - (l_0-k')/n_0}s_{n_0}^*s_{m'}s_{m'}^*u^{-k'}
\end{aligned} $$
is a partial isometry with the domain projection
$$ g=(u^{k'}s_{m'}s_{m'}^*u^{-k'}) u^{l_0-n_0(k_0-k')/m_0}s_{n_0(m'/m_0 \vee m'/n_0)}
s_{n_0(m'/m_0 \vee m'/n_0)}^*u^{-(l_0-n_0(k_0-k')/m_0)} $$
and the range projection
$$ f=(u^{k'}s_{m'}s_{m'}^*u^{-k'})u^{k_0-m_0(l_0-k')/n_0}s_{m_0(m'/m_0 \vee m'/n_0)}
s_{m_0(m'/m_0 \vee m'/n_0)}^*u^{-(k_0-m_0(l_0-k')/n_0)}. $$
Clearly, both $g$ and $f$ are subprojections of $u^{k'}s_{m'}s_{m'}^*u^{-k'}$. If
either $g\neq u^{k'}s_{m'}s_{m'}^*u^{-k'}$ or $f\neq u^{k'}s_{m'}s_{m'}^*u^{-k'}$ then we can
argue as above. So suppose that both $g=u^{k'}s_{m'}s_{m'}^*u^{-k'}$ and $f=
u^{k'}s_{m'}s_{m'}^*u^{-k'}$. Then by Lemma \ref{projectioninequality}, $m'$ is divisible by both
$n_0(m'/m_0 \vee m'/n_0)$ and $m_0(m'/m_0 \vee m'/n_0)$. This can only happen if $m_0=n_0$.
Now, since $m_0=n_0$, $0\neq k_0-l_0$ is divisible by $m_0$.
If we take $r\in\Zb$ relatively prime with $k_0-l_0$, then
$$ (u^{k'}s_rs_r^*u^{-k'}) (u^{k_0}s_{m_0}s_{m_0}^*u^{-l_0})(u^{k'}s_rs_r^*u^{-k'}) =
u^{k'}s_rs_r^*u^{k_0-l_0}s_rs_r^*s_{m_0}s_{m_0}^*u^{-k'}=0 $$
by (\ref{zeroequ}). Thus, in this case, it suffices to put $k=k'$ and $m=r \vee m'$.
\end{proof}
\begin{theorem}\label{purelyinfinite}
$C^*$-algebra ${\mathcal Q}_\Nb^\xi$ is purely infinite.
\end{theorem}
\begin{proof}
Let $0\neq x\in\Q_\Nb^\xi$. Since $\Q_\Nb^\xi$ is simple, to show it is purely infinite as well we must find
elements $T,R$ such that $TxR$ is invertible, \cite{Cpi}.
We have $0\neq G(xx^*)\geq 0$. Thus there exists a projection $Q\in\D_\Nb^\xi$
such that $G(xx^*)$ is invertible in $Q\D_\Nb^\xi$. So let $d$ be a positive element of $\D_\Nb^\xi$ such that
$G(dxx^*d)=d^2G(xx^*)=Q$.
Now, take a small $\epsilon>0$. There exists a finite collection $m_j,n_j\in\Nb^\times$, $k_j,l_j\in\Zb$,
$\lambda_j\in\Cb$ such that
$$ ||dxx^*d - \sum_j \lambda_j u^{k_j}s_{m_j}s_{n_j}^* u^{-l_j}|| < \epsilon. $$
Applying conditional expectation $G$ we get
$$ ||Q - \sum_{j:\; m_j=n_j, k_j=l_j} \lambda_j u^{k_j}s_{m_j}s_{m_j}^* u^{-k_j}|| < \epsilon. $$
Combining the two preceding inequalities, we see that
\begin{equation}\label{inequality}
||dxx^*d - Q - \sum_{j:\;m_j\neq n_j\; {\text or }\;k_j\neq l_j}\lambda_j u^{k_j}s_{m_j}s_{n_j}^* u^{-l_j}||
< 2\epsilon.
\end{equation}
Now, applying repeatedly Lemma \ref{cuntzlemma}, we find a $k\in\Zb$ and an $m\in\Nb^\times$ such that
$u^ks_ms_m^*u^{-k} \leq Q$ and $(u^ks_ms_m^*u^{-k})(u^{k_j}s_{m_j}s_{n_j}^*u^{-l_j})
(u^ks_ms_m^*u^{-k})=0$ for all $j$ with $m_j\neq n_j$ or $k_j\neq l_j$. Thus inequality
(\ref{inequality}) yields
$$ ||(u^ks_ms_m^*u^{-k})dxx^*d(u^ks_ms_m^*u^{-k}) - u^ks_ms_m^*u^{-k} || <2\epsilon. $$
Setting $T:=s_m^*u^{-k}d$ and $R:=x^*du^ks_m$ we have
$$ || TxR - 1 || <2\epsilon, $$
and $TxR$ is invertible if $\epsilon \leq 1/2$. This proves that $\Q_\Nb^\xi$ is purely infinite.
\end{proof}
|
\section{Introduction} \label{intro}
This paper is concerned with boundary-value problems
for partial differential equations of elliptic type coefficients of
which contain an indeterminacy. Such a situation is quite
typical in real-life problems where parameters of mathematical
model cannot be determined exactly and instead one knows only that the
coefficients belong to a certain set of ``admissible'' data $\Lambda$.
In view of this fact,
instead of a single exact solution ``$u$'', we have to consider a
``set of solutions'' (we denote it by $\indset( \Lambda )$). As a
result, the error control problem comes in a more complicated
form in which approximation errors must be evaluated together with
errors arose due to indeterminant data (various approaches that
can be used for such an analysis are exposed in, e.g.,
\cite{HlChBa,MaRe1,MaRe2,RE3}).
In this paper, we
establish explicit relations between the sets $\Lambda$
and $\indset(\Lambda)$ for the reaction-diffusion problem
with mixed Dirichl\'et--Robin boundary conditions
(we call it $\mathcal{P}$) defined by the system
\ben
\label{basic1}
-\dvg( A \nabla u ) + \rho u & = & f \quad \textrm{ in } \Omega \\
\label{basic2}
u & = & 0\quad \textrm{ on } \Gamma_1 \\
\label{basic3}
n \cdot A \nabla u & = & F \quad \textrm{ on } \Gamma_2 \\
\label{basic4}
\alpha u + n \cdot A \nabla u & = & G \quad \textrm{ on } \Gamma_3 .
\een
Here, $\Omega\in\Rd$ is a bounded and connected domain with Lipschitz
continuous boundary
$\Gamma_1 \cup \Gamma_2 \cup \Gamma_3$ and $f \neq 0$.
We assume that exact $A$, $\rho$, and $\alpha$ are unknown. Instead,
we know that $A\in \Lambda_A$, $\rho\in\Lambda_\rho$, and $\alpha\in\Lambda_\alpha$,
where
\be
\Lambda_A & := & \{ A \in L_2(\Omega,\Md) \; | \; A = A_0 + \delta_1 \Psi, \; ||\Psi||_{L_\infty(\Omega,\Md)} \leq 1 \} \\
\Lambda_\rho & := & \{ \rho \in L_2(\Omega) \; | \; \rho = \rho_0 + \delta_2 \psi_\rho , \; ||\psi_\rho||_{L_\infty(\Omega)} \leq 1 \} \\
\Lambda_\alpha & := & \{ \alpha \in L_2(\Gamma_3) \; | \; \alpha = \alpha_0 + \delta_3 \psi_\alpha , \; ||\psi_\alpha||_{L_\infty(\Gamma_3)} \leq 1 \} .
\ee
In other words, we assume that the sets of admissible data
are formed by some (limited) variations of some
known ``mean'' data (which are denoted by subindex 0). The parameters
$\delta_i$, $i=1,2,3$
represent the magnitude of these variations.
Thus, in the case considered,
\be
\Lambda:=\Lambda_A\times \Lambda_\rho\times \Lambda_\alpha.
\ee
We note that such a presentation of the data arises in
many engineering problems
where data are given in a form \verb+mean+$\pm$\verb+error+. The
solution associated to non-perturbed data $A_0$, $\rho_0$, and $\alpha_0$
is denoted by $u_0$.
Our goal is to give computable estimates of the radius of $\indset(\Lambda)$,
(we denote this quantity by $r_{\indset}$). The value of $r_{\indset}$ has
a large significance for practical applications because
it shows an accuracy limit defined by the problem statement.
Attempts to find approximate solutions having approximation
errors lesser then $r_{\indset}$ have no practical sense.
The generalized statement of Problem ($\mathcal P$)
is as follows:
Find $u \in V_0$ such that
\begin{equation} \label{varform}
a(u,w) = l(w) \quad \forall w \in V_0 ,
\end{equation}
where space $V_0$ and functionals $a: V_0 \times V_0 \rightarrow R$
and $l: V_0 \rightarrow R$ are defined by the relations
\be
V_0 & := & \{ w \in H^1(\Omega) \; | \; w_{|_{\Gamma_1}} = 0 \} , \\
a(u,w) & := & \IntO A \nabla u \cdot \nabla w \; dx + \IntO \rho uw \; dx + \IntG{3} \alpha uw \; ds, \\
l(w) & := & \IntO fw \; dx + \IntG{2} Fw \; ds + \IntG{3} Gw \; ds .
\ee
We assume that
\be
\blow{1} |\xi|^2 \leq & A_0 \xi \cdot \xi & \leq \bup{1} |\xi|^2
\quad \forall \xi \in \Rd \quad \textrm{ on } \Omega ,\\
\blow{2} \leq & \rho_0 & \leq \bup{2} \quad \textrm{ on } \Omega, \\
\blow{3} \leq & \alpha_0 & \leq \bup{3} \quad \textrm{ on } \Gamma_3 ,
\ee
where $\blow{i}>0$.
In view of the above-stated conditions, the ``mean'' problem is evidently
elliptic and has a unique solution $u_0$. The condition
\be
0 \leq \delta_i < \blow{i},\qquad i=1,2,3
\ee
guarantees that the perturbed problem remains elliptic and possesses
a unique solution $u$.
For each $A,\rho,\alpha\in \Lambda$, the corresponding
problem of $\mathcal{P}(\Lambda)$ is natural to analyze using the (energy) norm
\begin{equation}
\label{1.2}
\NT v \NT_{A,\rho,\alpha}^2 :=
a(v,v)= \IntO A \nabla v \cdot \nabla v \; dx + \IntO \rho v^2 \; dx + \IntG{3} \alpha v^2 \; ds .
\end{equation}
For the sake of simplicity we will also use an abridged notation
$\NT v \NT$ for the norm $\NT v \NT_{A,\rho,\alpha}$.
For the ``mean'' problem,
we use the norm \mbox{$\NT v \NT_{A_0,\rho_0,\alpha_0} $}, which is
also denoted by $\NT v \NT_{0}$.
It is easy to see that the norms
$\NT v \NT_{0}$ and $\NT v \NT$
are
equivalent and satisfy the relation
\begin{equation}
\label{norm_equi}
\underline{{\rm C}} \NT v \NT^2\;
\leq\; \NT v \NT_{0}^2 \;\leq \;\overline{{\rm C}} \NT v \NT^2 ,
\end{equation}
where
\begin{equation}
\overline{{\rm C}} := \max\limits_{ i \in \{ 1,2,3 \} } \frac{\bup{i}}{\blow{i}-\delta}
\quad \textrm{and} \quad
\underline{{\rm C}} := \min\limits_{ i \in \{ 1,2,3 \} } \frac{\blow{i}}{\bup{i}+\delta_i}
\end{equation}
These constants $\overline{{\rm C}}$ and $\underline{{\rm C}}$
depend only on the problem data and indeterminacy range. They play
an important role in our analysis.
Now, we can define the quantity we are interested in:
\begin{equation}
r_{\indset} :=
\sup\limits_{\tilde u \in \indset} \NT u_0 - \tilde u \NT_{0}.
\end{equation}
A normalized counterpart of $r_\indset$
is defined by the relation
\be
\hat r_{\indset} :=
\sup\limits_{\tilde u \in \indset} \frac{ \NT u_0 - \tilde u \NT_0 }
{ \NT u_0 \NT_0 }.
\ee
\section{Lower bound of $r_{\indset}$}
Problem ${\mathcal P}$ has a variational statement and the solution $u$
can be considered as a minimizer of the functional
\be
J(v):=\frac12 a(v,v)-l(v)
\ee
on the set $V_0$. Using this statement, we can easily derive
computable lower bounds of the difference between
$u$ and an arbitrary function $v \in V_0$ in terms of the energy norm
(see \cite{RE} where such bounds have been derived for a wide
class of variational problems).
First, we use the identity
\begin{equation}
\label{2.1}
\NT u - v \NT^2 = a(u-v,u-v) = 2 ( J(v) - J(u) ),
\end{equation}
which for quadratic functionals was established in \cite{Mi1}.
Let $w$ be an arbitrary function in $V_0$. Then,
\be
J(v)-J(u) \geq J(v) - J(v+w)
\ee
and by (\ref{2.1}) we conclude that
\begin{equation}
\label{2.2}
\NT u - v \NT^2 \geq -a(w+2v,w) + 2 l(w) \quad \forall w \in V_0 .
\end{equation}
We note that for $w = u-v$ the estimate (\ref{2.2})
holds as equality, so that there is no ``gap'' between the left- and
right-hand sides of (\ref{2.2}). The right hand side of (\ref{2.2})
is explicitly computable. It provides the so--called
functional error minorant, which we denote by
$\mathcal{M}_\ominus^{A,\rho,\alpha}(v,w)$
(if no confusion may arise, we also
use a simplified notation $\mathcal{M}_\ominus(v,w)$).
This functional serves as the main tool when deriving the lower bound
of $r_{\indset}$.
\begin{Theorem}
\label{Th2.1}
Assume that all the assumptions of Section \ref{intro} hold. Then
\begin{equation}
\label{2.3}
r_\indset^2 \geq \underline{{\rm C}}
\sup_{w \in V_0} {\mathrm M}^{r_\indset}_\ominus (u_0,w) ,
\end{equation}
where $w$ is an arbitrary function in $V_0$ and
\begin{multline}
\label{2.4}
{\mathrm M}^{r_\indset}_\ominus (u_0,w) := -\NT w \NT_{0}^2
+ \delta_1 \IntO | \nabla w + 2 \nabla u_0 | \; | \nabla w | \; dx + \\
+ \delta_2 \IntO | ( w + 2 u_0 ) w | \; dx
+ \delta_2 \IntG{3} | ( w + 2 u_0 ) w | \; ds .
\end{multline}
\end{Theorem}
\proofbegin
We have
\begin{equation}
\label{2.5}
r_{\indset}=
\sup\limits_{\tilde u \in \indset} \NT u_0 - \tilde u \NT_{0}\geq
\underline {C} \sup\limits_{\tilde u \in \indset} \NT u_0 - \tilde u \NT.
\end{equation}
On the other hand
\be
\sup\limits_{\tilde u \in \indset} \NT u_0 - \tilde u \NT^2 & = &
\sup\limits_{\tilde u \in \indset} \left\{ \sup\limits_{w \in V_0}
\mathcal{M}_\ominus(u_0,w) \right\} \\
& = &
\sup\limits_{w \in V_0} \left\{ \sup\limits_{A \in \Lambda_A, \rho \in \Lambda_\rho, \alpha \in \Lambda_\alpha} \mathcal{M}_\ominus^{A,\rho,\alpha}(u_0,w) \right\} .
\ee
and we conclude that
\begin{equation}
\label{2.6}
r_{\indset}^2 \geq \underline{{\rm C}}
\sup\limits_{w \in V_0} \left\{
\sup\limits_{
A \in \Lambda_A, \rho \in \Lambda_\rho, \alpha \in \Lambda_\alpha
}
\mathcal{M}_\ominus^{A,\rho,\alpha}(u_0,w)
\right\} .
\end{equation}
Now our goal is to estimate the right-hand side of (\ref{2.6})
from below. For this purpose, we exploit the structure of the minorant,
which allows to explicitly
evaluate effects caused by indeterminacy of the coefficients.
It is easy to see that the minorant can be represented
as follows:
\begin{multline}
\label{2.7}
{\mathcal M}_\ominus^{A,\rho,\alpha}(u_0,w) = - \IntO (A_0+\delta_1 \Psi) (\nabla w + 2 \nabla u_0 ) \cdot \nabla w \, dx \\
- \IntO (\rho_0+\delta_2 \psi_\rho) (w + 2 u_0 ) w \, dx \\
- \IntG{3} (\alpha_0+\delta_3 \psi_\alpha) (w + 2 u_0 ) w \, ds + 2 l(w)
\end{multline}
Note that
\be
\IntO ( A_0 \nabla u_0 \cdot \nabla w dx + \rho_0 u_0w ) dx +
\IntG{3} \alpha_0 u_0w ds=l(w).
\ee
Hence,
\begin{multline}
\label{2.8}
{\mathcal M}_\ominus^{A,\rho,\alpha}(u_0,w) = - \IntO A_0 \nabla w \cdot \nabla w \, dx - \IntO \rho_0 w^2 \, dx - \IntG{3} \alpha_0 w^2 \, ds \\
- \delta_1 \IntO \Psi ( \nabla w + 2 \nabla u_0 ) \cdot \nabla w \; dx
- \delta_2 \IntO \psi_\rho ( w + 2 u_0 ) w \; dx\\
- \delta_3 \IntG{3} \psi_\alpha ( w + 2 u_0 ) w \; ds
\end{multline}
and we obtain
\begin{multline}
\label{2.9}
\sup\limits_{
A \in \Lambda_A, \rho \in \Lambda_\rho, \alpha \in \Lambda_\alpha
}
\mathcal{M}_\ominus^{A,\rho,\alpha}(u_0,w)=-\NT w\NT^2_0+\\
+\delta_1
\sup\limits_{
|\Psi|\leq 1}
\IntO \Psi ( \nabla w + 2 \nabla u_0 ) \cdot \nabla w \; dx
+ \delta_2
\sup\limits_{
|\psi_\rho|\leq 1
}
\IntO \psi_\rho ( w + 2 u_0 ) w \; dx\\
+ \delta_3
\sup\limits_{
|\psi_\alpha|\leq 1
}
\IntG{3} \psi_\alpha ( w + 2 u_0 ) w \; ds.
\end{multline}
The integrand
of the first integral in the right-hand side of (\ref{2.9}) can be
presented as $\Psi:\tau$,
where
$$
\tau = \nabla w \otimes (\nabla w +
2 \nabla u_0)
$$
and $\otimes$ stands for the diad product.
For the first supremum we have
\begin{equation}
\label{2.10}
\sup_{|\Psi| \leq 1} \left\{
\IntO \Psi : \tau \; dx \right\} =
\IntO |\tau|
\; dx \; .
\end{equation}
Analogously, we find that
\ben
\label{2.11}
\sup\limits_{
|\psi_\rho|\leq 1
}
\IntO \psi_\rho ( w + 2 u_0 ) w \; dx\leq
\IntO | (w + 2 u_0 ) w | dx,
\\
\label{2.12}
\sup\limits_{
|\psi_\alpha|\leq 1
}\IntO \psi_\rho ( w + 2 u_0 ) w \; dx\leq
\IntG{3} |( w + 2 u_0 ) w| ds.
\een
By (\ref{2.9})--(\ref{2.12}), we arrive at the relation
\begin{multline}
\label{2.13}
\sup\limits_{A,\rho,\alpha}{\mathcal M}_\ominus^{A,\rho,\alpha}(u_0,w)
= - \NT w\NT^2_0+ \delta_1 \IntO |( \nabla w + 2 \nabla u_0 )
\otimes \nabla w| \; dx\\
+ \delta_2 \IntO | ( w + 2 u_0 ) w | \; dx
+ \delta_3 \IntG{3} |( w + 2 u_0 ) w | \; ds .
\end{multline}
which together with (\ref{2.6}) leads to \eqref{2.3}.
\proofend
\vspace{10pt}
Theorem \ref{Th2.1} gives a general form of the lower bound
of $r_\indset$. Also, it creates a basis for practical computation
of this quantity. Indeed, let $V_{0h}\subset V_0$ be a finite dimensional
space. Then
\begin{equation}
\label{2.14}
r_\indset^2 \geq \underline{{\rm C}}
\sup_{w \in V_{0h}} {\mathrm M}^{r_\indset}_\ominus (u_0,w).
\end{equation}
It is worth noting that the wider set $V_{0h}$ we take the better
lower bound of the radius we compute. However, as it is shown below,
a meaningful lower bound can be deduced even analytically.
\begin{Corollary}
\label{lb_rough}
Under assumptions of Theorem \ref{Th2.1},
\begin{equation}
\label{2.15}
r_\indset^2
\geq
\underline{{\rm C}} \; r_{\indset \ominus}^2
\quad \textrm{ and } \quad
\hat r_{\indset}^2
\geq
\underline{{\rm C}} \; \hat r_{\indset \ominus}^2 ,
\end{equation}
where
\begin{equation}
\label{2.16}
r_{\indset \ominus}^2 = \frac{\NT u_0 \NT^4_\delta }
{ \NT u_0\NT_0 ^2 - \NT u_0\NT_\delta ^2 }\geq \frac{\Theta^2}{1-\Theta} \NT u_0 \NT_0 ^2 ,
\end{equation}
where
\be
\NT u_0 \NT^2_\delta:=
\delta_1 ||\nabla u_0||_\Omega^2 + \delta_2 ||u_0||_\Omega^2 + \delta_3 ||u_0||_{\Gamma_3}^2
\ee
and
\be
\Theta:=\min\limits_{ i \in \{ 1,2,3 \} } \frac{\delta_i}{\bup{i}}.
\ee
For the normalized radius, we have
\begin{equation} \label{2.17}
\hat r_{\indset \ominus}^2 =\frac{\Theta^2}{1-\Theta}.
\end{equation}
\end{Corollary}
\proofbegin Use \eqref{2.3} and set
\begin{equation}
\label{2.18}
w = \lambda u_0 ,
\end{equation}
where $\lambda \in {\mathbb R}$. Then we observe that,
\ben
\label{2.19}
&& r_{\indset}^2 \geq \underline{{\rm C}} \Bigg( - \lambda^2 \NT u_0 \NT_0 ^2
+ \lambda (\lambda + 2) \NT u_0 \NT^2_\delta \Bigg) \; .
\een
The right hand side of (\ref{2.19}) is a quadratic function with respect to $\lambda$.
It attains its maximal value if
\be
\lambda \NT u_0 \NT^2_0=(\lambda+1)\NT u_0 \NT^2_\delta,
\ee
i.e., if for
\be
\lambda=\frac{\NT u_0 \NT^2_\delta}{\NT u_0 \NT^2_0-\NT u_0 \NT^2_\delta} .
\ee
Substituting this $\lambda$, we arrive at \eqref{2.16}. Note that
\begin{multline}
\NT u_0 \NT_0 ^2 = \IntO ( A_0 \nabla u_0 \cdot \nabla u_0 + \rho_0 u_0^2\; dx + \IntG{3} \alpha_0 u_0 \; ds
\geq \\
\geq \IntO ( \blow{1} \nabla u_0 \cdot \nabla u_0 + \blow{2} u_0^2 ) \,dx +
\IntG{3} \blow{3} u_0 \,ds
> \\
>\delta_1 ||\nabla u_0||_\Omega^2 + \delta_2 ||u_0||_\Omega^2 + \delta_3 ||u_0||_{\Gamma_3}^2=
\NT u_0 \NT^2_\delta,
\end{multline}
so that $\lambda$ (and the respective lower bound) is positive.
Moreover,
\begin{multline}
\label{2.21}
\NT u_0 \NT^2_\delta
\geq
\frac{\delta_1}{\bup{1}} \IntO A_0 \nabla u_0 \cdot \nabla u_0 \; dx
+ \frac{\delta_2}{\bup{2}} \IntO \rho_0 u_0^2 \; dx
+ \frac{\delta_3}{\bup{3}} \IntG{3} \alpha_0 u_0^2 \; ds \\
\geq
\Theta \NT u_0 \NT_0 ^2.
\end{multline}
Also,
\begin{multline}
\label{2.22}
\NT u_0\NT_0 ^2 - \NT u_0\NT_\delta ^2\\=
\IntO (A_0-\delta_1I) \nabla v \cdot \nabla v \; dx + \IntO (\rho_0-\delta_2) v^2 dx +
\IntG{3} (\alpha_0-\delta_3) v^2 \, ds\\
\geq
\left(1-\frac{\delta_1}{\bup{1}}\right)
\IntO A_0\nabla v \cdot \nabla v \, dx + \left(1-\frac{\delta_2}{\bup{2}}\right)
\IntO \rho_0 v^2 dx +
\left(1-\frac{\delta_3}{\bup{3}}\right)\IntG{3} \alpha_0 v^2 \, ds\\
\geq \max\limits_{i=1,2,3}
\left(1-\frac{\delta_i}{\bup{i}}\right)\NT u_0\NT^2_0=(1-\Theta)\NT u_0\NT^2_0.
\end{multline}
By (\ref{2.21}) and (\ref{2.22}), we arrive at relation
\begin{equation}
\label{2.23}
r_{\indset \ominus}^2 \geq \frac{\Theta^2}{1-\Theta} \NT u_0 \NT_0 ^2,
\end{equation}
which implies (\ref{2.16}) and (\ref{2.17}).
\section{Upper bound of $r_{\indset}$}
A computable upper bound
of $r_{\indset}$ can be derived with the help of a posterioiri error majorant
of the functional type, which are derived by purely functional methods without
attracting any information on the mesh and method used. For a wide class of problems they were
derived in \cite{RE,RE1,RE2} by variational techniques and in \cite{RE2,RS,RST}
by transformations of integral identities (see also the papers cited therein).
Below
we derive functional error majorant for our class of problems using the latter method
based on transformations of the respective integral identity.
After that, we use it's properties to derive the desired upper bound in \ref{der_ub}.
\subsection{Error majorant}
\label{der_maj}
Let $v\in V_0$ be an admissible approximation of the exact solution $u$
(generated by $A$, $\varrho$, and $\alpha$).
From \eqref{varform} it follows that for any $w\in V_0$
\begin{multline}
a(u-v,w)= \IntO fw \; dx + \IntG{2} Fw \; ds + \IntG{3} Gw \; ds - \\
-\IntO A \nabla v \cdot \nabla w \; dx - \IntO \rho v w \; dx - \IntG{3} \alpha vw \; ds + \\
+ \IntO \left( \dvg(y) w + y \cdot \nabla w \right) \; dx - \int\limits_{\Gamma_2\cup\Gamma_3}
( y\cdot\nu)\, w \; ds\,,
\end{multline}
where $\nu$ denotes unit outward normal to $\Gamma$
and
\be
y\in H^+(\dvg,\Omega):=
\{y\in H(\dvg,\Omega)\,\mid\,y\cdot\nu\in L^2(\Gamma_2\cup\Gamma_3)\}.
\ee
We note that the last line is zero for all $y \in H(\Omega,\dvg)$
(in view of the integration-by-parts formula). We regroup the terms and rewrite the relation
as follows:
\begin{equation}
a(u-v,w)= I_1 + I_2 + I_3 + I_4 ,
\end{equation}
where
\be
I_1 & := & \IntO r_1(v,y)w \; dx := \IntO (f - \rho v + \dvg(y) ) w \; dx ,\\
I_2 & := & \IntO r_2(v,y)w \; dx := \IntG{3} (G-\alpha v-y\cdot\nu\,) w \; ds ,\\
I_3 & := & \IntG{2} (F-y\cdot\nu\,) w \; ds ,\\
I_4 & := & \IntO (y - A \nabla v) \cdot \nabla w \; dx .
\ee
Now we can estimate each term separately by the Friedrichs and trace inequalities
(which holds due to our assumptions concerning $\Omega$). We have
\be
|| w ||_\Omega^2 \leq C_1(\Omega) || \nabla w ||_\Omega^2 \quad \forall w \in V_0, \\
|| w ||_{2,\Gamma_2}^2 \leq C_2(\Omega,\Gamma_2) || \nabla w ||_\Omega^2 \quad \forall w \in V_0, \\
|| w ||_{2,\Gamma_3}^2 \leq C_3(\Omega,\Gamma_3) || \nabla w
||_\Omega^2 \quad \forall w \in V_0. \ee When estimating the
integrands of $I_1$ and $I_2$, we introduce additional functions
$\mu_1$ and $\mu_2$, which have values in $[0, 1]$.
\begin{multline*}
I_1 = \IntO \frac{\mu_1}{\sqrt{\rho}} r_1(v,y) \sqrt{\rho} w \; dx + \IntO (1-\mu_1) r_1(v,y) w \; dx \\
\leq || \frac{\mu_1}{\sqrt{\rho}} r_1(v,y) ||_\Omega \; || \sqrt{\rho} w ||_\Omega +
\sigma_1|| (1-\mu_1) r_1(v,y) ||_\Omega \; \left(\IntO A\nabla w\cdot\nabla w \,dx\right)^{1/2}
\end{multline*}
and
\be
I_2 & = & \IntG{3} \frac{\mu_2}{\sqrt{\alpha}} r_2(v,y) \sqrt{\alpha} w \; dx + \IntG{3} (1-\mu_2) r_2(v,y) w \; dx \\
& \leq & || \frac{\mu_2}{\sqrt{\alpha}} r_2(v,y) ||_{\Gamma_3} \;
|| \sqrt{\alpha} w ||_{\Gamma_3} +\\
&&+ || (1-\mu_2) r_2(v,y) ||_{\Gamma_3} \; \sigma_3
\left(\IntO A\nabla w\cdot\nabla w \,dx\right)^{1/2},
\ee
and
\begin{equation}
I_3 \leq || F-y\cdot\nu\, ||_{\Gamma_2}
\sigma_2
\left(\IntO A\nabla w\cdot\nabla w \,dx\right)^{1/2},
\end{equation}
where
\be
\sigma_1=\sqrt{ \frac{C_1(\Omega)}{\blow{1}} },\quad
\sigma_2=\sqrt{ \frac{C_1(\Omega)C_2(\Omega,\Gamma_2)}{\blow{1}} },
{\rm and}\quad
\sigma_3=\sqrt{ \frac{C_1(\Omega)C_3(\Omega,\Gamma_3)}{\blow{1}} }.
\ee
We note that a similar approach
was used in \cite{RS} for the reaction-diffusion problem and in \cite{RST}
for the generalized Stokes problem.
In these publications it was shown that splitting of the
residual term (performed with the help of a single function $\mu$) allows to obtain error majorants
that are insensitive with respect to small values of the lower term coefficient and at the same
time sharp (i.e., have no irremovable gap between the left- and right-hand sides). In our case,
we have two lower terms, so that we need two functions $\mu_1$ and $\mu_2$ to split the respective
residual terms.
The term $I_4$ is estimated as follows:
\begin{equation}
I_4 \leq D(\nabla v,y)^{\frac{1}{2}} \left(\IntO A\nabla w\cdot\nabla w \,dx\right)^{1/2},
\end{equation}
where
\begin{equation}
D(\nabla v, y) = \IntO ( y-A\nabla v ) \cdot ( \nabla v - A^{-1} y ) \; dx .
\end{equation}
We collect all the terms and obtain
\begin{multline}
a(u-v,w) \leq \Big( D(\nabla v,y)^{1/2} +
\sigma_1 || (1-\mu_1) r_1(v,y) ||_\Omega + \\
+ \sigma_3 || (1-\mu_2) r_2(v,y) ||_{\Gamma_3} + \sigma_2 || F-y\cdot\nu\, ||_{\Gamma_2} \Big) \left(\IntO A\nabla w\cdot\nabla w \,dx\right)^{1/2} + \\
+ || \frac{\mu_1}{\sqrt{\rho}} r_1(v,y) ||_\Omega \; || \sqrt{\rho} w ||_\Omega +
|| \frac{\mu_2}{\sqrt{\alpha}} r_2(v,y) ||_{\Gamma_3} \; || \sqrt{\alpha} w ||_{\Gamma_3} .
\end{multline}
Set $w=u-v$ and
use the Cauchy-Schwartz inequality
\begin{equation}
\label{CS}
\sum\limits_{i=1}^d x_i y_i \leq \sqrt{ \sum\limits_{i=1}^d x_i^2}\;
\sqrt{ \sum\limits_{i=1}^d {y_i^2}}.
\end{equation}
Then, we arrive at the estimate
\begin{multline}
\label{3.8}
\NT u-v\NT ^2 \leq \Big( D(\nabla v,y)^{1/2} + \sigma_1 || (1-\mu_1) r_1(v,y) ||_\Omega + \\
+ \sigma_3 || (1-\mu_2) r_2(v,y) ||_{\Gamma_3} + \sigma_2 || F-y\cdot\nu\, ||_{\Gamma_2} \Big)^2 \\
+ || \frac{\mu_1}{\sqrt{\rho}} r_1(v,y) ||_\Omega^2 + || \frac{\mu_2}{\sqrt{\alpha}} r_2(v,y) ||_{\Gamma_3}^2 .
\end{multline}
It is worth remarking, that the estimate (\ref{3.8}) provides a guaranteed upper bound
of the error for {\em any} conforming approximation of the problem (\ref{basic1}--\ref{basic4}). The
estimate has a form typical for all functional a posteriori estimates: it is presented by the sum
of residuals of the basic relations with multipliers that depend on the constants in the respective
functional (embedding) inequalities for the domain and boundary parts.
However, for our subsequent goals, it is desirable to have the majorant in a form that
involve only quadratic terms. Such a form can be easily derived from (\ref{3.8})
if we square both parts and apply the algebraic inequality
\eqref{CS} to the first term (with multipliers $\gamma_i>0$, $i=1,2,3,4$). Then, we obtain
\begin{multline}
\label{3.9}
\NT u-v\NT ^2 \leq \kappa
\Bigl( \gamma_1 D(\nabla v,y) + \gamma_2 ||(1-\mu_1) r_1(v,y) ||_\Omega^2\\
+ \gamma_3 || (1-\mu_2) r_2(v,y) ||_{\Gamma_3}^2 + \gamma_4 || F-y\cdot\nu\, ||_{\Gamma_2}^2 \Bigr) + \\
+ || \frac{\mu_1}{\sqrt{\rho}} r_1(v,y) ||_\Omega^2 + || \frac{\mu_2}{\sqrt{\alpha}} r_2(v,y) ||_{\Gamma_3}^2 .
\end{multline}
where
\be
\kappa :=\frac{1}{\gamma_1} + \frac{\sigma^2_1}{ \gamma_2} +
\frac{\sigma^2_2}{ \gamma_3} + \frac{\sigma^2_3}{\gamma_4}.
\ee
We note that (\ref{3.9}) coincides with (\ref{3.8}) if
\ben
\gamma_1=\bar\gamma_1 & := & D(\nabla v,y)^{-1/2} , \\
\gamma_2=\bar\gamma_2 & := & \frac{ \sigma_1 }{ ||(1-\mu_1) r_1(v,y) ||_\Omega } , \\
\gamma_3=\bar\gamma_3 & := & \frac{ \sigma_2 }{ ||(1-\mu_2) r_2(v,y) ||_{\Gamma_3} } , \\
\gamma_4=\bar\gamma_4 &: = & \frac{ \sigma_3 }{ || F-y\cdot\nu\, ||_{\Gamma_2} } .
\een
Certainly, the estimate (\ref{3.9}) looks more complicated with respect
to (\ref{3.8}). However, it has an important advantage: the
weight functions $\mu_1$ and $\mu_2$ enter it as quadratic integrands, so that we can
easily find their optimal form adapted to a particular $v$ and the respective
error distribution.
In the simplest case, we take $\mu_1=\mu_2=0$, which yields the estimate
\begin{multline} \label{3.14}
\NT u-v\NT ^2 \leq \kappa
\times \Big( \gamma_1 D(\nabla v,y) + \gamma_2 || r_1(v,y) ||_\Omega^2 +\\
+\gamma_3 || r_2(v,y) ||_{\Gamma_3}^2 + \gamma_4 || F-y\cdot\nu\, ||_{\Gamma_2}^2 \Big).
\end{multline}
Another estimate arises if we set $\mu_1=\mu_2=1$.
In this case, the terms with factors $\sigma_1$ and $\sigma_3$ in (\ref{3.8})
are equal to zero, so that subsequent relations do not contain the terms with
multipliers formed by $\gamma_2$ and $\gamma_3$. Hence, we arrive at the estimate
\begin{multline}
\label{3.15}
\NT u-v\NT ^2 \leq \Big( \frac{1}{\gamma_1} + \frac{\sigma^2_3}{\gamma_4} \Big) \times \Big( \gamma_1 D(\nabla v,y) + \gamma_4 || F-y\cdot\nu\, ||_{\Gamma_2}^2 \Big) + \\
+ || \frac{1}{\sqrt{\rho}} r_1(v,y) ||_\Omega^2 + || \frac{1}{\sqrt{\alpha}} r_2(v,y) ||_{\Gamma_3}^2 .
\end{multline}
The estimate \eqref{3.15} involves ``free'' parameters $\gamma_1$ and $\gamma_4$
and a ´´free'' vector--valued function $y$ (which can be thought of as an image of the true flux).
There exist a combination of these free parameters that makes the left-hand side
of the estimate equal to the right-hand one. Indeed, set $y=A \nabla u$. Then
\be
r_1(v,y) & = & \rho (u-v) \quad{\rm in}\;\Omega , \\
r_2(v,y) & = & \alpha (u-v) \quad{\rm on}\;\Gamma_3 , \\
F - y\cdot\nu & = & 0\quad{\rm on}\;\Gamma_2
\ee
and we find that for $\gamma_4$ tending to infinity and for any $\gamma_1>0$ the right-hand side of (\ref{3.15})
coincides with the energy norm of the error.
However, the estimate (\ref{3.15}) has a drawback: it is sensitive
with respect to $\rho$ and $\alpha$
and may essentially overestimate the error if $\rho$ or $\alpha$ are small.
On the other hand, the right-hand side of \eqref{3.14} is stable with respect to small values of
$\rho$ and $\alpha$. Regrettably, it does not possess the ``exactness'' (in the above-discussed sense)
because it may have a ``gap'' between the left- and right-hand sides for any $y$.
An upper bound of the error that combines positive features of (\ref{3.14}) and (\ref{3.15})
can be derived (as in \cite{RS,RST}) if a certain optimization procedure is used in order to
select optimal functions $\mu_1$ and $\mu_2$.
If $y$ is given, then optimal $\mu_1$ and $\mu_2$ can be found analytically. It is not difficult
to see that $\mu_1$ must minimize the quantity
\begin{equation}
\IntO\left(
\kappa \gamma_2 (1-\mu_1)^2 + \frac{\mu_1^2}{\rho} \right) r_1(v,y)^2 dx.
\end{equation}
The quantity attains its minimum with
\begin{equation}
\mu_1(x)=\mu_1^{opt}(x): = \frac{ \kappa \gamma_2 }{ \kappa \gamma_2 + \rho(x)^{-1}} \quad \textrm{ in } \Omega .
\end{equation}
Similarly, we find that the integrals associated with $\Gamma_3$ attain minimum if
\begin{equation}
\mu_2(x)=\mu_2^{opt}(x) := \frac{ \kappa \gamma_3}{ \kappa \gamma_3+ \alpha(x)^{-1} } \quad \textrm{ on } \Gamma_3 .
\end{equation}
Substituting these values to \eqref{3.9} results in the estimate
\begin{multline}
\label{3.19}
\NT u-v\NT ^2 \leq \kappa \Big( \gamma_1 D(\nabla v , y)
+ \gamma_2 || \frac{ \sqrt{ \kappa^2 \gamma_2^2 \rho + 1 } }{ \kappa \gamma_2 \rho + 1 } \; r_1(v,y) ||_\Omega^2 + \\
+ \gamma_3 || \frac{ \sqrt{ \kappa^2 \gamma_3^2 \alpha + 1 } }{ \kappa \gamma_3 \alpha + 1 } \; r_2(v,y) ||_{\Gamma_3}^2
+ \gamma_4 || F-y\cdot\nu\, ||_{\Gamma_2}^2 \Big) .
\end{multline}
\begin{Remark}
For practical computations, it may be easier to use \eqref{3.9} and directly
minimize its right-hand side with respect to $\mu_i$, $\gamma_i$, and $y$ using the following
iteration procedure:
\begin{enumerate}
\item Keep $\gamma_i$ and $\mu_j$ fixed in \eqref{3.9} and minimize resulting quadratic functional of $y$ in suitable finite subspace. This task can be reduced to solving a system of linear equations.
\item Compute $\gamma_i^{opt}$.
\item Compute $\mu_j^{opt}$ and repeat from step 1.
\end{enumerate}
\end{Remark}
We denote the right hand side of \eqref{3.9} by ${\mathcal{M}_\oplus(v,y,\gamma,\mu_1,\mu_2)}$.
This error majorant provides a guaranteed upper bound of the error,
i.e.,
\begin{equation}
\label{3.20}
\NT u-v\NT ^2 \leq {\mathcal{M}_\oplus(v,y,\gamma,\mu_1,\mu_2)}.
\end{equation}
It is exact (in the above-discussed sense). Indeed, for
$y=A\nabla u$ and $\mu_1=\mu_2=1$ we obtain
\begin{equation}
\label{3.21}
\inf\limits_{\gamma_i>0}{\mathcal{M}_\oplus(v,A \nabla u,\gamma,1,1)}= \NT u-v\NT ^2 .
\end{equation}
Also, it directly follows from the structure of (\ref{3.19}) that the right-hand side
is insensitive to small values of $\rho$ and $\alpha$.
\subsection{Upper bound}
\label{der_ub}
In this section, we
derive an upper bound of $r_{\indset}$. For this purpose, we use the majorant
${\mathcal{M}_\oplus(v,y,\gamma,\mu_1,\mu_2)}$. Since the majorant
nonlinearly depends on $A$, $\rho$, and $\alpha$, taking supremum over the respective
indeterminacy sets imposes a more complicated task than that for the minorant.
For a class of diffusion problems this task was solved in \cite{RE,RE3}. Below, we
deduce a simpler estimate, which can be easily exploited in practical computations
and serves as a natural counterpart for the lower bound derived in Corollary 2.1.
\begin{Proposition} \label{comp}
Assume that all the assumptions of Section \ref{intro} hold. Then
\begin{equation} \label{ub00}
r_{\indset}^2 \leq \overline{{\rm C}} \; r_{\indset \oplus}^2
\quad \textrm{and} \quad
\hat r_\indset^2 \leq \overline{{\rm C}} \; \hat r_{\indset \oplus}^2 ,
\end{equation}
where
\begin{equation}
\label{3.23}
r_{\indset \oplus}^2 = \frac{\delta_1^2}{\blow{1}-\delta_1} ||\nabla u_0 ||^2_\Omega + \frac{\delta_2^2}{\blow{2}-\delta_2} || u_0 ||_\Omega^2 + \frac{\delta_3^2}{\blow{3}-\delta_3} || u_0 ||_{\Gamma_3}^2
\end{equation}
and
\begin{equation}
\label{3.24}
\hat r_{\indset \oplus}^2 =
\max\limits_{i \in \{ 1,2,3\}} \frac{\delta_i^2}{\blow{i} (\blow{i}-\delta_i)}.
\end{equation}
\end{Proposition}
\proofbegin
By properties of the majorant, we have
\be
\sup\limits_{\tilde u \in \indset} \NT u_0-\tilde u\NT^2
& = &
\sup\limits_{\tilde u \in \indset} \left\{ \inf\limits_{y,\mu_i,\gamma_j} {\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,y,\gamma,\mu_1,\mu_2)} \right\} \\
& \leq &
\inf\limits_{y,\mu_i,\gamma_j} \left\{ \sup\limits_{A,\rho,\alpha} {\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,y,\gamma,\mu_1,\mu_2)} \right\} .
\ee
Applying \eqref{norm_equi}, we obtain
\begin{equation}
\label{3.25}
r_\indset^2 \leq \overline{{\rm C}} \inf\limits_{y,\mu_i,\gamma_j} \left\{ \sup\limits_{A,\rho,\alpha} {\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,y,\gamma,\mu_1,\mu_2)} \right\} .
\end{equation}
Our task is to explicitly estimate the term in brackets.
For this purpose, we estimate from above the last two terms of the majorant and represent
it in the form
\begin{multline} \label{3.26}
{\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,y,\gamma,\mu_1,\mu_2)} \leq\\ \kappa
\Bigg( \gamma_1 D(\nabla v , y)
+ \left\| \sqrt{ \gamma_2 \kappa (1-\mu_1)^2 +
\frac{\mu_1^2}{\kappa(\blow{2} -\delta_2)}} \; r_1(v,y) \right\|_\Omega^2 + \\
+ \left\|
\sqrt{ \gamma_3 \kappa (1-\mu_2) +
\frac{\mu_2}{\kappa(\blow{3} -\delta_3)}} \; r_2(v,y) \right\|_{\Gamma_3}^2
+ \gamma_4 || F-y\cdot\nu\, ||_{\Gamma_2}^2 \Bigg) .
\end{multline}
Now, we find upper bounds with respect to
$A \in \Lambda_A$, $\rho \in \Lambda_\rho$, and $\alpha \in \Lambda_\alpha$ separately.
First, we consider the term $D$ generated by $A$ and $A^{-1}$:
\begin{multline}
\label{3.27}
\sup\limits_{A \in \Lambda_A} D(\nabla u_0, y) =
\sup\limits_{|\Psi|<1} \IntO (A_0+\delta_1 \Psi)^{-1} |(A_0+\delta \Psi)\nabla u_0 -y|^2 \; dx \\
\leq
\frac{1}{\blow{1}-\delta_1} \sup\limits_{|\Psi|<1} \left\{ ||A_0 \nabla u_0 -y ||^2 +
2 \delta_1 \IntO \Psi \nabla u_0 \cdot (A_0 \nabla u_0 - y) \; dx +
\delta_1^2 || \Psi \nabla u_0 ||^2 \right\} \\
\leq
\frac{1}{\blow{1}-\delta_1} \left( ||A_0 \nabla u_0 -y ||^2_\Omega +
2 \delta_1 \IntO |\nabla u_0 | \; | A_0 \nabla u_0 - y | \; dx +
\delta_1^2 || \nabla u_0 ||^2_\Omega \right).
\end{multline}
For the term related to the error in equilibrium equation, we have
\begin{multline}
\label{3.28}
\sup\limits_{\rho \in \Lambda_\rho} || r_1^{\rho}(u_0,y) ||_\Omega^2 =
\sup\limits_{|\psi_2|<1} \IntO (f- (\rho_0+\delta_2 \psi_2) u_0 + \dvg y )^2 \; dx \\
=
\sup\limits_{|\psi_2|<1} \IntO \left( \dvg y - \dvg( A_0 \nabla u_0) -\delta_2 \psi_2 u_0 \right)^2 \; dx
\\
\leq
|| \dvg(y-A_0 \nabla u_0) ||_\Omega^2 + 2 \delta_2 \IntO | \dvg(y-A_0 \nabla u_0) | \; |u_0| \; dx
+ \delta_2 || u_0 ||^2.
\end{multline}
Similarly, for the term related to the error in Robin boundary condition we have
\begin{multline}
\label{3.29}
\sup\limits_{\alpha \in \Lambda_\alpha} || r_2^{\alpha}(u_0,y) ||_{\Gamma_3}^2 \leq
\left\| \frac{\partial (y-A_0 \nabla u_0)}{\partial \nu} \right\|_{\Gamma_3}^2 \\
+ 2 \delta_3 \IntG{3}\left| \frac{\partial (y-A_0 \nabla u_0)}{\partial \nu} \right| \; | u_0 | \; ds + \delta_3^2 ||u_0||_{\Gamma_3}^2.
\end{multline}
It is clear, that for $y=y_0:=A_0 \nabla u_0$, the estimates (\ref{3.27})--(\ref{3.29})
attain minimal values. In addition, we
set in \eqref{3.26} $\mu_1=\mu_2 =1$ and find that
\begin{multline} \label{3.30}
{\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,A_0\nabla u_0,\gamma,1,1)}\leq\\
\leq
\kappa\left(
\frac{\delta_1^2 \gamma_1}{\blow{1}-\delta_1} || \nabla u_0 ||_\Omega^2
+ \frac{\delta_2^2}{\blow{2} -\delta_2} || u_0 ||_\Omega^2 +
\frac{\delta_3^2}{\blow{3} -\delta_3} || u_0 ||_{\Gamma_3}^2\right) .
\end{multline}
Now we tend $\gamma_2$,$\gamma_3$ and $\gamma_4$ (which are contained in $\kappa$) to infinity.
Then, \eqref{3.30} and \eqref{3.25} imply \eqref{3.23}.
An upper bound for the normalized radius follows from the relation
\begin{multline*}
{\mathcal{M}_\oplus^{A,\rho,\alpha}(u_0,A_0 \nabla u_0,\gamma,1,1)} \leq \\
\leq \frac{\delta_1^2}{\blow{1} (\blow{1}-\delta_1)}
\IntO A \nabla u_0 \cdot \nabla u_0 \; dx +
\frac{\delta_2^2}{\blow{2} (\blow{2}-\delta_2)} || \sqrt{\rho_0} u_0 ||_\Omega^2 +
\frac{\delta_3^2}{\blow{3} (\blow{3}-\delta_3)} || \sqrt{\alpha_0} u_0 ||_{\Gamma_3}^2 \leq \\
\leq \max\limits_{i \in \{ 1,2,3\}} \frac{\delta_i^2}{\blow{i} (\blow{i}-\delta_i)} \NT u_0 \NT ^2,
\end{multline*}
which leads to \eqref{3.24}. \proofend
\begin{Remark}
\label{b_ok}
The normalized lower bound in Corollary 2.1 is
less than the upper bound established in Proposition 3.1. Indeed, using
\begin{equation*}
\left(
1 - \min\limits_i \frac{\delta_i}{\bup{i}}
\right)^{-1}
=
\left(
\max\limits_i \left( 1 - \frac{\delta_i}{\bup{i}} \right)
\right)^{-1}
=
\left(
\min\limits_i \frac{\bup{i}}{\bup{i} - \delta_i}
\right)
\end{equation*}
to $\hat r_{\indset \ominus}^2$, we arrive at the following relation between bounds:
\begin{multline*}
\frac{ \overline{{\rm C}} \; \hat r_{\indset \oplus}^2 }{ \underline{{\rm C}} \; \hat r_{\indset \ominus}^2 } =
\frac{ \left( \max\limits_{i} \frac{\bup{i}}{\blow{i} - \delta_i} \right) \left( \max\limits_{i} \frac{\delta_i^2}{\blow{i} (\blow{i}-\delta_i)} \right) }{ \left( \min\limits_{i} \frac{\blow{i}}{\bup{i} + \delta_i} \right) \left( \min\limits_{i} \frac{\delta_i^2}{\bup{i}^2} \right) \left( \min\limits_{i} \frac{\bup{i}}{{\bup{i}-\delta_i}} \right) } = \\
= \left( \max\limits_{i } \frac{\bup{i}}{\blow{i} - \delta_i} \right) \left( \max\limits_{i} \frac{\delta_i^2}{\blow{i} (\blow{i}-\delta_i)} \right) \left( \max\limits_{i } \frac{\bup{i} + \delta_i}{\bup{i}} \right) \left( \max\limits_{i} \frac{\bup{i}^2}{\delta_i^2} \right) \left( \max\limits_{i} \frac{{\bup{i}-\delta_i}}{\bup{i}} \right) .
\end{multline*}
Maximums can be estimated from below by setting $i=1$ everywhere. The expression simplifies to
\begin{equation}
\frac{ \overline{{\rm C}} \; \hat r_{\indset \oplus}^2 }{ \underline{{\rm C}} \; \hat r_{\indset \ominus}^2 }
\geq
\frac{\bup{1}}{\blow{1}}
\left(
\frac{\bup{1}+\delta_1}{\blow{1}-\delta_1}
\right) \geq 1 .
\end{equation}
\end{Remark}
|
\section{Introduction}
In the past decade, there have been many revelations of the importance of non-coding RNAs to cellular regulatory functions and thus a growing interest in computational prediction of RNA tertiary structure \cite{LaingAndSchlick2010}, \cite{LeontisAndWesthof2012}. Nevertheless, RNA tertiary structure prediction from a single RNA sequence is a significant challenge. One major unresolved issue is in the immense space of tertiary conformations even for a short RNA sequence. Existing methods usually employ random sampling algorithms for computation feasibility, which assemble sampled tertiary motifs into native-like structures \cite{DasAndBaker2007}, \cite{DingEtAl2008}, \cite{JonikasEtAl2009}, \cite{ParisienAndMajor2008}, \cite{PopendalEtAl2012}, \cite{SharmaEtAl2008}. To reduce the chance to miss native structures, the assembly algorithms have mostly been guided with constraining structural models. For example, MC-Fold/MC-Sym \cite{ParisienAndMajor2008} assumes the tertiary structure consists of 4-nt cyclic tertiary motifs constructible from the predicted secondary structure. Rosetta \cite{DasAndBaker2007,DasEtAl2010} {\it de novo} assembles tertiary structure from a database of 3-nt tertiary fragments. Other methods follow samplings that preserve the secondary structure \cite{BidaAndMaher2012}, \cite{PopendalEtAl2012,ReinharzEtAl2013} or intervention from human experts \cite{JossinetEtAl2010}, \cite{MartinezEtAl2008}. However, these constraining models do not necessarily ensure that native conformations are
examined. In particular, the state-of-the-art methods have yet to deliver the desired prediction accuracy for RNA sequences of lengths beyond 50 \cite{LaingAndSchlick2010}.
In this work, we introduce a novel method to predict nucleotide interactions from sequences as a key step toward accurate {\it ab initio} prediction of tertiary structure.
Accurate knowledge of the nucleotide interactions is crucial to predicting the tertiary structure of an RNA and subsequently predicting its functional roles. To predict nucleotide interactions, our method is guided by a novel graph model called a {\it backbone $k$-tree}, for small integer $k$, to globally constrain the nucleotide interaction relationships (NIRs) that constitute the tertiary structure. In such a $k$-tree graph, nucleotides are organized into groups of size $k+1$, such that NIRs are permitted only for nucleotides belonging to the same group and groups are connected to each other with a tree topology (see section 2). This model was inspired by our recent discovery of the small treewidth of the NIR graphs for more than 3,500 RNA chains extracted from 1,984 resolved RNAs (Figure 1). We have been able to develop dynamic programming algorithms with $O(n^{k+1})$ time and space complexities, efficient for small $k$, to compute the optimal backbone $k$-tree spanning over the nucleotides on the query sequence, given a scoring function \cite{DingEtAl2014,DingEtAl2014a}
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.40]{tw_dis.png}
\caption{Treewidth distribution of NIR graphs of more than 3,500 chains deriving from 1,984 resolved RNA tertiary structures in the RNA Structure Atlas \cite{SarverEtAl2008}. The RNAs with treewidth larger than 18 are omitted due to their very small number. These treewidths are actually upper bounds computed by a program \cite{BoalaenderAndKoster2010}; it is likely that the exact treewidths of the NIR graphs may actually be smaller.}
\label{tw}
\end{figure}
To ensure that the computed optimal $k$-tree can actually yield the set of nucleotide interactions that constitutes the native tertiary structure, our method defines the scoring function over detailed patterns of nucleotide interactions within every group of $k+1$ nucleotides. We consider nucleotide interactions from the established geometric nomenclatures \cite{LeontisAndWesthof2001} and nucleotide interaction families \cite{LeontisEtAl2002}, \cite{StombaughEtAl2009}, \cite{ZirbelEtAl2009}, including base-base, base-phosphate, and base-ribose as well as base-stacking interactions. To test our method, we adopted an improved 3-tree model and pre-computed candidates of interaction patterns for every group of 4 nucleotides, by searching through RNA Structure Atlas \cite{SarverEtAl2008}; this contains annotated atom-level nucleotide interactions for nearly 3,000 resolved tertiary structures.
We trained artificial neural networks (ANNs) to compute the confidence of every given nucleotide interaction and the confidence of every admissible nucleotide interaction pattern for every group of 4 given nucleotides. We filtered out unlikely interaction patterns and kept only those with high confidences. With this 3-tree model, our algorithm efficiently predicts an optimal set of nucleotide interactions from the query sequence within computational time $O(c^5Mn^3)$, where $M$ is a constant and $c\leq 20$ is the maximum number of candidate interaction patterns for one group of 4 nucleotides. We have implemented the algorithm into a program called BkTree, which may use known or predicted canonical (i.e., {\it cis} Watson-Crick) base pairs on the query sequence.
To evaluate our method for nucleotide interaction prediction, we tested BkTree on a benchmark set of 43 high resolution RNAs, which had been used to survey a number of state-of-the-art tertiary structure prediction methods \cite{LaingAndSchlick2010}. The resolved, atom-level interactions were extracted with FR3D \cite{SarverEtAl2008}. BkTree performed impressively well across the set of tested RNAs (Table~\ref{withBbsTable}), achieving the averaged sensitivity, PPV, and MCC values of 0.86, 0.78, and 0.82, respectively (discounting the input canonical base pairs). In comparison with previous programs MC \cite{ParisienAndMajor2008}, Rosetta \cite{DasAndBaker2007}, and NAST \cite{JonikasEtAl2009} that all assumed the secondary structure as a part of the input \cite{LaingAndSchlick2010}, it is clear that BkTree outperformed the other three programs in the MCC measure on this set of benchhmark RNAs (Table~\ref{avg4methods}, Figure~\ref{laingMCCCompare}). In particular, on the four representative RNAs that contain typical helices and junctions \cite{LaingAndSchlick2010}, BkTree gave the best performance on all but one RNA, for which BkTree acquired a higher sensitivity value but lower PPV than the MC program, resulting in a slightly lower MCC value (Table~\ref{compare_Laing}).
To evaluate the significance of our method to 3D conformation prediction, we used the program MC-Sym to model 3D conformations from the interactions predicted by BkTree and calculated RMSDs against the resolved structures. Since MC-Sym requires secondary structure for 3D conformation modeling, we identified 30 RNAs from the benchmark set for which their secondary structures are covered by the BkTree-predicted nucleotide interactions together with its input canonical base pairs.
For the 4 representative RNAs listed in Table~\ref{compare_Laing}, BkTree outperforms MC and Rosetta on 3 of them.
\section{Model and Methods}
In this work, we consider all known types of
nucleotide interactions of atomic-resolution \cite{LeontisAndWesthof2001}, \cite{LeontisEtAl2002}, \cite{ZirbelEtAl2009}.
In particular, with the base triangle model consisting of Watson-Crick (W), Hoogsteen (H), and sugar (S) edges, base-base interactions has been fully characterized into rich 12 geometric types and 18 interaction families \cite{LeontisAndWesthof2001}, \cite{LeontisEtAl2002}, according to involved edges, {\it cis} or {\it trans}, and parallel or anti-parallel, observed in crystal structures. For example the cWW family contains, in addition to the canonical (i.e., {\it cis} Watson-Crick) base pairs, many non-canonical base-base interactions through W edges. More recently, classifications of nucleotide interactions have been extended to base-backbone interactions. There are 10 families identified for base-phosphate interactions based on the position of the interacting hydrogen atom in the base \cite{ZirbelEtAl2009}. Similarly, 9 additional families have been identified for base-ribose interactions \cite{ZirbelEtAl2011}. A few base stacking interactions have also been classified.
Table 1 summarizes these classes of nucleotide interactions, which also includes the backbone interaction between two neighboring nucleotides.
\begin{table}[!ht]
\centering
\caption{Categories, types and families of RNA nucleotide interactions, mostly summarized from works \cite{LeontisAndWesthof2001}, \cite{LeontisEtAl2002}, \cite{ZirbelEtAl2009,ZirbelEtAl2011}. It also includes the phosphodiester interaction between two neighboring nucleotides.}
\begin{tabular}{|cl|c|}
\hline
\multicolumn{1}{|c|}{Categories} & \multicolumn{1}{c|}{Types (Interaction Families)} & Number \\ \hline\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{Base pairs}} & cWW, tWW, cWH, tWH, cHW, tHW, cWS, tWS, cSW, & \multirow{2}{*}{18} \\
\multicolumn{1}{|c|}{} & tSW, cHH, tHH, cHS, tHS, cSH, tSH, cSS, tSS & \\ \hline
\multicolumn{1}{|c|}{Base-phosphates} & 0BPh, 1BPh, 2BPh, 3BPh, 4BPh, 5BPh, 6BPh, 7BPh, 8BPh, 9BPh & 10 \\ \hline
\multicolumn{1}{|c|}{Base-riboses} & 0BR, 1BR, 2BR, 3BR, 4BR, 5BR, 6BR, 7BR, 9BR & 9 \\ \hline
\multicolumn{1}{|c|}{Bases stackings} & s35, s53, s33, s55 & 4 \\ \hline
\multicolumn{1}{|c|}{Backbone-backbone} & phosphodiester & 1 \\ \hline
\end{tabular}
\label{all_inters}
\end{table}
\subsection{Backbone $k$-Tree Model}
Let the query RNA sequence be $S=S_1 S_2, \dots S_n$, where $S_i \in \{{\tt A, C, G, U}\}$, for $1\leq i \leq n$. We denote an interaction between
the $i$th and $j$th nucleotides, where $i<j$, with triple $\langle{S_i}^{(i)}, {S_j}^{(j)}, t\rangle$, for some interaction type $t$ shown in Table~\ref{all_inters}. Note that there are possibly two or more simultaneous interactions between the two nucleotides.
Given the native tertiary structure of the sequence $S$, we
model the {\it nucleotide interaction relationship}s (NIRs) within the tertiary structure with a graph $G=(V, E)$, where $V=\{ {S_i}^{(i)}: 1\leq i \leq n\}$, such that $({S_i}^{(i)}, {S_j}^{(j)})$ is an edge in $E$ if and only if $i\not= j$ and $\langle{S_i}^{(i)}, {S_j}^{(j)}, t\rangle$ is an interaction for some $t$.
We call $G$ {\it the NIR graph} of the sequence with the given structure. Because every two consecutive nucleotides are connected with the phosphodiester bond, every NIR graph of $n$ vertices contains all edges $({S_i}^{(i)}, S_{i+1}^{(i+1)})$, for $1\leq i \leq n-1$. These edges are called {\it backbone edges}.
In our recent investigation \cite{DingEtAl2014}, we constructed NIR graphs for all RNAs whose tertiary structures were known from RNA Structure Atlas \cite{ReinharzEtAl2013}. We discovered that an overwhelming majority of these RNAs are of small treewidths (Figure~\ref{tw}). Treewidth is a graph metric, which intuitively indicates how much a graph is tree-like. If a graph has treewidth bounded by $k$, any clique obtained by deleting vertices and edges and contracting edges of the graph can contain at most $k+1$ vertices \cite{ArnborgEtAl1990}. Thus the distribution of treewidths suggest that NIRs in the RNA tertiary structures are in general not arbitrarily complex.
The concept of treewidth originated from the algorithmic graph theory. It is closely related to, and may be better explained with the notion of {\it $k$-tree}, which is central to this work.
\begin{definition} \rm \cite{Patil1986}
\label{k-tree}
Let integer $k\geq 1$. The class of {\it $k$-trees} are graphs defined by the following inductive steps:
\begin{enumerate}
\item A $k$-tree of $k+1$ vertices is a clique of $k+1$ vertices;
\item A $k$-tree of $n$ vertices, for $n>k+1$, is a graph consisting of a $k$-tree $G$ of $n-1$ vertices and a vertex $v$, which does not occur in $G$, such that $v$ forms a $(k+1)$-clique with some $k$-clique already in $G$.
\end{enumerate}
\end{definition}
Figure \ref{fig_k_tree} shows a 3-tree with seven vertices in (a) and illustrates it in (b) with a tree-topology that connects the four 4-cliques in the graph.
\begin{figure}[!ht]
\centering
\includegraphics[width=\textwidth]{3treebb3tree.png}
\caption{(a) 3-tree of 7 vertices by Definition~\ref{k-tree}, with the order of forming the four 4-cliques: with initial clique $\{1, 2, 3, 6\}$ (black edges), vertex $5$ and blue edges added, then vertex $7$ and red edges added, and finally vertex 4 and green edges added. (b) Illustration of the graph of (a) with a tree-topology connecting the four 4-cliques. (c) A backbone 3-tree for sequence {\tt AUUGGCA}, of the same topology as shown in (a); backbone edges are in bold.}
\label{fig_k_tree}
\end{figure}
By \cite{vanLeeuwen1990}, for any $k \geq 1$, a graph is of treewidth $\leq k$ if and only if it is a subgraph of a $k$-tree. Therefore, NIR graphs for an overwhelming majority of known RNA tertiary structures are constrained in topology by $k$-trees, for small values of $k$. Because technically, every graph of treewidth bounded by $k$ can be augmented with additional edges into a $k$-tree, we adopt such $k$-trees as the model for NIRs of the RNA tertiary structure.
\begin{definition} \rm Let $k \geq 1$ be an integer. The {\it backbone $k$-tree} for an RNA sequence is an augmented NIR graph of the sequence, which is a $k$-tree.
\end{definition}
Figure 1(c) shows a backbone $3$-tree for sequence {\tt AUUGGCA}. Note that backbone $k$-trees differ from general $k$-trees in that a backbone $k$-tree has to the designated Hamiltonian path (consisting of all the backbone edges).
With the backbone $k$-tree model, in order to predict the set $I$ of nucleotide interactions from the query sequence, we propose to identify a backbone $k$-tree $G=(V, E)$ such that
\[ (S_i^{(i)}, S_j^{(j)}) \in E \mbox{ if and only if } \exists \, t\, \mbox{ } \langle S_i^{(i)}, S_j^{(j)}, t\rangle \in I\]
To ensure the identified $G$ actually corresponds to the set of interactions that constitute the native structure of the query sequence, we need to quantify nucleotide interactions for combinatorial optimization of such a backbone $k$-tree $G$, as explained in the subsequent sections.
\subsection{Quantification of Nucleotide Interactions}
\begin{definition} {\rm Let $q$ be a $(k+1)$-clique in a backbone $k$-tree of query sequence $S$.
An {\it interaction pattern} (ip) for clique $q$ is a set $P_q$ of interactions for the nucleotides in $q$ such that for every interaction $\langle {S_i}^{(i)}, {S_j}^{(j)}, t\rangle$ in $P_q$, both nucleotides ${S_i^{(i)}}$ and ${S_j}^{(j)}$ are in clique $q$.
}
\end{definition}
Given an ip $P_q$ for clique $q$, we define the {\it induced subgraph by $P_p$}, denoted with $B_{P_q} = (q, E_{B_{P_q}})$ to be a subgraph of $q$ such that edge $(S_i^{(i)}, S_j^{(j)}) \in E_{B_{P_q}}$ only if interaction $\langle S_i^{(i)}, S_j^{(j)}, t\rangle \in P_q$ for some $t$.
\begin{definition} {\rm Let $q$ be a $(k+1)$-clique in the in a backbone $k$-tree of query sequence $S$. The {\it confidence} of a given ip $P_q$ for clique $q$ is defined as
\begin{eqnarray}
\label{confidence}
f(q, P_q, S) = \sum_{\langle S_i^{(i)},S_j^{(j)}, t \rangle \in P_q} c_{q, B_{P_q}, t}^{(i,j)}
\end{eqnarray}
where $c_{q, B_{P_q}, t}^{(i,j)}$ is the {\it confidence} of interaction
$\langle S_i^{(i)}, S_j^{(j)}, t\rangle$ given $q$ and subgraph $B_{P_q}$ induced by $P_q$.}
\end{definition}
In the Section~\ref{section-algo}, we will introduce artificial neural networks (ANNs) to compute confidence $c_{q, B_{P_q}, t}^{(i,j)}$.
For every clique $q$, with ${\cal Q}(q)$, we denote the finite set of all ips for $q$. In the practical application, we may only include those ips in ${\cal Q}(q)$ which have ``high'' confidences (e.g., above certain threshold). Let $I$ be a set of interactions. By notation $I|_q$, we mean the maximal size subset of $I$ that is an ip for $q$.
\begin{definition} \rm Let $k$ be any fixed integer $\geq 2$. The {\it nucleotide interaction prediction} problem NIP$(k)$ is, given an input query sequence $S$, to identify a backbone $k$-tree $G^*=(V, E^*)$ as well as a set $I^*$ of nucleotide interactions that constitutes the tertiary structure of $S$, such that
every interaction $\langle S_i^{(i)},S_j^{(j)}, t \rangle \in I^*$ implies edge $(S_i^{(i)},S_j^{(j)}) \in E^*$ and
\begin{equation}
\label{optimization}
(I^*, G^*) = \arg \, \max\limits_{(I, G)} \,\{ \sum\limits_{q \mbox{ in } G, \, I|_q \in {\cal Q}(q)} f(q, I|_q, S)\}
\end{equation}
\end{definition}
\subsection{Overview of the Method}\label{section-overview}
Our method consists of three major components to solve the NIP$(k)$ problem, for any fixed $k\geq 2$. The first component is data repositories including NIPDB and NIPCCTable. NIPDB is a database of all possible interaction patterns (ips) for every $(k+1)$-clique, which was established by searching through the RNA Structure Atlas \cite{SarverEtAl2008}. For every such clique, its ips in NIPDB are extracted and ranked when the query sequence is preprocessed. NIPCCTable is a matrix for compatibility between every pair of ips for two cliques that share all but one nucleotide. The compatibility is checked by the dynamic programming algorithm computing the NIP$(k)$ problem.
The second component is a set of artificial neural networks (ANNs) to compute confidence for any given interaction type $t$ between any two given nucleotides $S_i^{(i)}$ and $S_j^{(j)}$ on the query sequence. The computed confidences for interactions are then used to compute confidence of an ip for every $(k+1)$-clique, as formulated in equation~(\ref{confidence}). For every such clique $q$, all ips of $q$ obtained from database NIPDB are ranked according to their confidence values. Often the number of ips with significant confidence values is small, e.g., $\leq 20$; ips of significant scores are included as ip candidates into the set ${\cal Q}(q)$ for $q$. The detailed construction of the ANNs will be described in the next section.
The third component is a dynamic programming algorithm solving the NIP$(k)$ problem, using the prepared data and preprocessing results from the first two components. From the input query sequence, the algorithm produces a backbone $k$-tree $G^*$ as well as a set $I^*$ of nucleotide interactions, maximizing the aggregate confidence value across all $(k+1)$-cliques in $G^*$ (see equations (\ref{optimization}) and (\ref{confidence})). The relationship between $G^*$ and $I^*$ is that, for every $(k+1)$-clique $q$ in the $k$-tree $G^*$, there is a maximal subset of nucleotide interactions $I|_q \subseteq I^*$ being an ip for $q$, such that $I^* = \bigcup_{q \mbox{ in } G^*} I^*|_q$. The next section describes the details of the dynamic programming algorithm.
\section{Algorithms}
\label{section-algo}
\subsection{ANNs for Computing Interaction Confidence}
Let the query sequence $S=S_1S_2\dots S_n$ of $n$ nucleotides, where $S_i \in \{{\tt A, C, G, U}\}$, for $1\leq i \leq n$. Technically we considered all $(k+1)$-cliques formed by $k+1$ vertices $\{S_{h_0}^{(h_0)}, S_{h_1}^{(h_1)}, \dots, S_{h_k}^{(h_k)}\}$, where $1\leq h_0 < h_1 < \dots < h_k \leq n$. Let $q=(V, E)$ be such a clique and $B_q = (V, E_{B_q})$, where $E_{B_q} \subseteq E$, be any subgraph of $q$. For every edge $(S_i^{(i)},S_j^{(j)})\in E_{B_q}$ and every possible interaction $\langle S_i^{(i)},S_j^{(j)}, t\rangle$ of type $t$, we constructed an ANN ${\cal N}_{q, B_{q},t}^{(i,j)}$ to calculate {\it confidence} $c_{q, B_{q},t}^{(i,j)}$ that interaction
$\langle {S_i}^{(i)}, {S_j}^{(j)}, t\rangle$ occurs in the subgraph $B_q$ of clique $q$.
Each ANN ${\cal N}_{q, B_{q},t}^{(i,j)}$ consists of an input layer, a hidden layer, and an output layer. The output layer is a single unit depicting a confidence value for interaction $\langle {S_i}^{(i)}, {S_j}^{(j)}, t\rangle$. The input layer consists of input units representing the selected global and local features shown in Table \ref{featureTable}. The features included the sequence length and the distance between the involved nucleotides as well as neighboring nucleotide types. In addition, we included the information of {\it assumed} canonical base pairs\footnote{These are known or predicted Watson-Crick and wobble base pairs. Note that they do not necessarily constitute all information about the secondary structure.} within the query sequence. The complete list of features selected for the trainings are given in the Table \ref{featureTable}.
\begin{table}[!ht]
\caption{Features selected from a given $(k+1)$-clique $q$ and given subgraph $B_q$ of $q$ for training ANN ${\cal N}_{q, B_{P_q},t}^{(i,j)}$. CBP is an abbreviation for canonical base pair. A {\it Component} (Cp) is defined as the maximal subsequence consisting of two or more nucleotides each involved in a CBP.}
\centering
\begin{tabular}{|c|c|p{10cm}|}
\hline
Feature & Value & \multicolumn{1}{c|}{Comments} \\ \hline\hline
Seq. length & An integer & Length of a training sequence containing $q$. \\ \hline
Distances & $k$ integers & Distances between every two nucleotides in the sequential order in $q$. \\ \hline
\multirow{3}{*}{Number of Cps} & \multirow{3}{*}{$k$ integers} & Number (one of $\{0, 1, 2, 3, -1\}$) of Cps on the subsequence between every two nucleotides in the sequential order. 3 means there are at least 3 Cps; $-1$ means the two nucleotides are neighboring nucleotides on the sequence. \\ \hline
\multirow{4}{*}{Neighbor nts.} & \multirow{4}{*}{$k+1$ 4-mers} & One $4$-mer (of letter {\tt A, C, G, U}) for every nucleotide in $q$, where the first two letters and the last two letter of the 4-mer indicate the two nts to the left and to the right of the nucleotide, respectively, and letter {\tt N} is used when there is no neighbor. \\ \hline
\multirow{4}{*}{Neighbor CBPs} & \multirow{4}{*}{$k+1$ 4-mers} & One 4-mer (of binary bits) for every nucleotide in $q$, where the first two bits and the last two bits of the 4-mer indicate the two nts to the left and to the right of the nucleotide are involved in CBPs, respectively, and letter {\tt N} is used when there is no neighbor. \\ \hline
\multirow{4}{*}{Edge properties} & \multirow{4}{*}{up to $\frac{k(k+1)}{2}$ integers} & For every edge in the subgraph $B_q$ of $q$, value 0 indicates both nts are involved in a CBP; -1 (resp. +1) indicates exclusively left (resp. right) nt is involved in a CBP; 2 indicates either is near a CBP; and -2 indicates both are far away (distant beyond 3 nts) from a CBP. \\ \hline
\end{tabular}
\label{featureTable}
\end{table}
We adopted conventional methods to construct and train the ANNs \cite{Mitchell1997}, typically the technique of back-propagation with gradient descent, using a fixed-size network. This is based on the calculation of the error by taking the first derivatives of half the Euclidean distance between the output and target and back-propagating it towards the input layer, over the whole training set. Each weight is then updated according to the error contribution of each unit, the error of each output unit and a learning rate. The logistic sigmoid was used as the activation functions for each unit. The updating is repeated until the training error converges to a minimum or the cross-validation error starts to rise, due to over-fitting. The learning rate $0.03$ was the value that yielded the best results for a subset of $895$ RNAs from RNA Structure Atlas.
The trained ANNs can be applied to compute confidence for interaction patterns. In particular, given a $(k+1)$-clique $q =\{S_{h_1}^{(h_1)}, S_{h_2}^{(h_2)}, \dots, S_{h_{k+1}}^{(h_{k+1})}\}$, $1\leq h_1 < \dots < h_{k+1} \leq n$, let $P_q$ be an ip for $q$ and let $B_{P_q}$ be the underlying graph for $P_q$, which is a subgraph of clique $q$. Then the trained ANN ${\cal N}_{q, B_{P_q},t}^{(i,j)}$ can be applied on each edge $(S_i^{(i)},S_j^{(j)}) \in E_{B_{P_q}}$ and each type $t$ to compute the confidence score $c_{q,B_{P_q}, t}^{(i,j)}$ for interaction $\langle {S_i}^{(i)}, {S_j}^{(j)}, t\rangle$. The confidence $f(q, P_q, S)$ of $P_q$ for $q$ is computed with the equation~(\ref{confidence}).
Then for $q$, all the ips $P_q$'s are ranked according to their confidences $f(q, P_q, S)$, and only significant top $m$ ips are included in the candidate set ${\cal Q}(q)$. We have chosen $m \leq 20$ in the performance evaluations as our experiments results had showed that a larger $m$ could not help to improve the results.
\subsection{Algorithm for NIP$(k)$ problem}
Roughly speaking, the algorithm for NIP$(k)$ problem considers every $(k+1)$-clique, from which recursive creations of more cliques are all examined. For every newly created clique $q$, all ips from ${\cal Q}(q)$
are considered but eventually exactly one of them is chosen for $q$. The algorithm follows the basic process of creating $k$-tree given in Definition 1. However, because the identified $k$-tree is a backbone $k$-tree that contains all backbone edges, the process is not straightforward. We need the following notations for an introduction to the algorithmic idea.
By {\it interval} $[i..j]$, for $i\leq j$, we mean the set of consecutive integers between $i$ and $j$, inclusive. Two intervals $[i..j]$ and $[h..l]$ are {\it non-overlapping} if either $j\leq h$ or $l\leq i$.
Formally, let the query sequence be $S=S_1S_2\dots S_n$ and $q$ be a clique formed by $k+1$ vertices $\{S_{h_1}^{(h_1)}, S_{h_2}^{(h_2)}, \dots, S_{h_{k+1}}^{(h_{k+1})}\}$, where $1=h_0 \leq h_1 < h_2 < \dots < h_{k+1} \leq n=h_{k+2}$. Let $A$ be a set of non-overlapping intervals and $P_q \in {\cal Q}(q)$ be an ip for clique $q$.
We define function $M(q, A, P_q, S)$
to be the maximum confidence of a $k$-tree constructed beginning from clique $q$, which includes all backbone edge $(S_i^{(i)}, S_{i+1}^{(i+1)})$ for integers $i$ and $i+1$ both contained in the same interval in $A$. Then we obtain the following recurrence:
\begin{eqnarray}
M(q, A, P_q, S) &=& \max\limits_{S_x^{(x)} \in q, \, S_y^{(y)} \not \in q, \, y \in [i..j] \in A, \, p=q|^x_y \,} \nonumber \\
&& \{ \max_{P_p \in {\cal Q}(p), \, {\cal R}(B, C), {\cal P}(P_q, P_p) } \{ M(p, B, P_p, S) + M(q, C, P_q, S) + f(q, P_q, S) \} \, \}\label{dp}
\end{eqnarray}
where abbreviations $q|^x_y = q \cup \{S_y^{(y)} \}\setminus \{S_x^{(x)}\}$, ${\cal P}(P_p, P_q)$ asserts that the chosen ip $P_p$ be compatible with $P_q$, and ${\cal R}(B, C)$ represents the choices of two sets of intervals, $B$ and $C$, which satisfy constraints
\begin{enumerate}[(a)]
\item $\{[i..y], [y..j]\} \subseteq B$, $\{[w..x], [x..z]\} \subseteq C$, for applicable $w$ and $z$; and
\item $B \cup C = A \cup \{[i..y], [y..j]\} \setminus \{[i..j]\}$, and $B \cap C = \emptyset$.
\end{enumerate}
Recurrence (\ref{dp}) gives an iterative process to produce a backbone $k$-tree. The intuitive idea is to create a new clique $p$ from $q$ by introducing a new nucleotide vertex $S_y^{(y)}$ to the partially constructed $k$-tree. This results in possibly two or more sub-$k$-trees, one starting from $p$ and the others from $q$ (but not including $S_y^{(y)}$). Since the two or more sub-$k$-trees will never join together again, interval sets are used to ensure backbone edges will be properly created.
Essentially, the constructed $k$-tree corresponding to the value of function $M(q,A,P_q,S)$ contains only those backbone edges that connect the nucleotides of indexes specified in the intervals in $A$. In particular, starting from clique $q$ of $k+1$ vertices $\{S_{h_1}^{(h_1)}, S_{h_2}^{(h_2)}, \dots, S_{h_{k+1}}^{(h_{k+1})}\}$, to compute an backbone $k$-tree that contains all the backbone edges, we need to set $A = \{ [h_i..h_{i+1}]: 0\leq i \leq k+1\}$, where $h_0 =1$ and $h_{k+2} = n$.
The confidence score of the produced $k$-tree is computed as the sum of confidence scores of ips chosen for all involved $(k+1)$-cliques. The chosen ips need to be compatible across the cliques when they share nucleotide interactions or even just nucleotides. This is ensured by the assertion ${\cal P}(P_q, P_p)$, which checks (1) $P_q$ and $P_p$ have the same set of interactions on the edges shared by cliques $q$ and $p$ by looking up table NIPCCTable; and (2) any pattern of interactions between a single nucleotide and multiple others has to exist in the structure database.
To complete the recurrence, we need the following base case:
\[M(q, A, P_q, S) = 0 \, \, \,\mbox{ if }A = \emptyset\]
To identify the desired backbone $k$-tree $G^*$, we maximize $M(q, A, P_q, S)$ over all starting clique $q$ and all ip $P_q \in {\cal Q}(q)$. The associated set $I^*$ of nucleotides is just the union of the chosen ips for all $(k+1)$-cliques in $G^*$.
Recurrence (\ref{dp}) naturally offers a dynamic programming solution. Function $M(q, P_q, A, S)$ can be computed by establishing a table with dimensions for $q, P_q$, and $A$. With the base cases, the table is computed bottom-up, from $A=\emptyset$, using the recurrence~(\ref{dp}).
\subsection{Improved Algorithms}
Simply implementing the above outlined algorithm would require $O(n^{k+1})$ memory space and $O(n^{k+2})$ computation time for every fixed value of $k$. Following the same idea but creating $(k+1)$-cliques from $k$-cliques instead leads
to an improved dynamic programming algorithm to solve the NIP$(k)$ problem, with a little more sophisticated steps to navigate through $k$-cliques. The improved algorithm uses $O(n^{k})$ amount of memory space and $O(n^{k+1})$ amount of time for every fixed value of $k$ \cite{DingEtAl2014,DingEtAl2014a}.
The efficiency can be further improved by demanding that every $(k+1)$-clique in backbone $k$-trees contains two consecutive nucleotides $S_i^{(i)}$ and $S_{i+1}^{(i+1)}$ for some $i$. That is, every interaction pattern for a $(k+1)$-clique always contains at least one backbone edge. This allows a further reduction of computation time to $O(n^k)$. Testing on the case $k=3$ has shown that the constrained backbone $3$-tree model maintains the similar capability to account for sophisticated nucleotide interactions as the ``standard'' backbone $3$-tree model. In addition the constraint may enforce the construction of the $3$-tree to follow backbone edges, providing more controls on the $3$-tree construction. Finally, the constraint also significantly reduced the number of cases that the ANNs need to consider in their construction.
\subsection{Implementation}
The NIPDB database construction was implemented by Python, where Prody package \cite{BakanEtAl2011} was adopted to search RNA Structure Atlas. Afterward, NIPCCTable, the matrix for ip consistence and compatibility was developed using Python. Training and building of ANNs were realized with WEKA package \cite{MarkHallEtAl2008}. Finally, confidences of ips admissible for every clique $(k+1)$-clique in the query sequence was computed by programs in Python.
We implemented in C++ the dynamic programming algorithm into a program called BkTree. We ran the evaluation tests on a Red Hat 4.8.2-7 server with 4 Intel Quad core X5550 Xeon Processors, 2.66GHz 8M Cache and 70GB Memory.
\section{Performance Evaluation}
\subsection{Test Data}\label{sectionTestData}
We implemented our method in the program BkTree. We evaluated our method through testing BkTree on a list of 43 RNAs
of high resolution structure data, which had been used as a benchmark set to evaluate a number of state-of-the-art tertiary structure prediction methods in the survey \cite{LaingAndSchlick2010}. 18 of the RNA sequences are of length $\geq 50$. In developing the ANNs for computing interaction confidences, 7 of these RNAs were not included in the training data.
\begin{table}[!ht]
\centering
\caption{Nucleotide interaction prediction results by BkTree on the benchmark set used in the survey \cite{LaingAndSchlick2010}. The number of canonical base pairs (CPBs) and number of non-canonical interactions (NCIs) are listed. The sensitivity (STY), PPV and MCC were calculated, excluding the canonical bases pairs used as a part of the input. The data of the 7 RNAs not used for training ANNs are displayed with the bold font.}
\label{withBbsTable}
\begin{tabular}{|c|c|c|c|c|c|c|l|}
\hline
PDB ID& Length & \# CBPs & \# NCIs & STY & PPV & MCC & \multicolumn{1}{c|}{Structure complexity} \\ \hline \hline
\textbf{2F8K} & \textbf{16} & \textbf{6} & \textbf{14} & \textbf{85} & \textbf{85} & \textbf{0.8571} & \textbf{Hairpin} \\ \hline
\textbf{2AB4} & \textbf{20} & \textbf{6} & \textbf{20} & \textbf{100} & \textbf{90} & \textbf{0.9534} & \textbf{Hairpin} \\ \hline
\textbf{361D} & \textbf{20} & \textbf{5} & \textbf{17} & \textbf{70} & \textbf{57} & \textbf{0.6351} & \textbf{Hairpin} \\ \hline
2ANN & 23 & 3 & 24 & 75 & 66 & 0.7071 & Hairpin \\ \hline
1RLG & 25 & 5 & 22 & 95 & 63 & 0.7793 & Hairpin, internal loop \\ \hline
2QUX & 25 & 9 & 22 & 90 & 71 & 0.8058 & Hairpin \\ \hline
387D & 26 & 4 & 23 & 86 & 68 & 0.7744 & Hairpin \\ \hline
1MSY & 27 & 6 & 39 & 97 & 92 & 0.9502 & Hairpin \\ \hline
1L2X & 28 & 8 & 34 & 88 & 88 & 0.8823 & Pseudoknot \\ \hline
2AP5 & 28 & 8 & 29 & 82 & 66 & 0.7427 & Pseudoknot \\ \hline
1JID & 29 & 8 & 31 & 93 & 72 & 0.8235 & Hairpin, internal loop \\ \hline
1OOA & 29 & 8 & 29 & 93 & 72 & 0.8242 & Hairpin, internal loop \\ \hline
430D & 29 & 6 & 37 & 94 & 77 & 0.8577 & Hairpin, internal loop \\ \hline
3SNP & 30 & 12 & 31 & 93 & 85 & 0.8932 & Hairpin, internal loop \\ \hline
2OZB & 33 & 10 & 33 & 93 & 79 & 0.8641 & Hairpin, internal loop \\ \hline
1MJI & 34 & 10 & 44 & 84 & 84 & 0.8409 & Hairpin, internal loop \\ \hline
1ET4 & 35 & 8 & 40 & 67 & 84 & 0.7546 & Pseudoknot \\ \hline
2HW8 & 36 & 12 & 44 & 93 & 80 & 0.8655 & Hairpin, internal loop \\ \hline
1I6U & 37 & 15 & 47 & 91 & 89 & 0.9053 & Hairpin, internal loop \\ \hline
1F1T & 38 & 10 & 38 & 81 & 63 & 0.7184 & Hairpin, internal loop \\ \hline
\textbf{1ZHO} & \textbf{38} & \textbf{13} & \textbf{46} & \textbf{95} & \textbf{83} & \textbf{0.8911} & \textbf{Hairpin, internal loop} \\ \hline
1S03 & 47 & 18 & 53 & 88 & 79 & 0.8404 & Hairpin, internal loop \\ \hline
1XJR & 47 & 15 & 55 & 83 & 80 & 0.8215 & Hairpin, internal loop \\ \hline
1U63 & 49 & 17 & 50 & 94 & 65 & 0.7833 & Hairpin, internal loop \\ \hline
2PXB & 49 & 16 & 66 & 98 & 94 & 0.9632 & Hairpin, internal loop \\ \hline
2FK6 & 53 & 20 & 58 & 77 & 70 & 0.7385 & Pseudoknot, 3-way junction \\ \hline
3E5C & 53 & 21 & 65 & 84 & 73 & 0.7877 & 3-way junction (riboswitch) \\ \hline
1MZP & 55 & 17 & 73 & 64 & 73 & 0.6876 & Hairpin internal \\ \hline
1DK1 & 57 & 24 & 65 & 100 & 89 & 0.9436 & 3-way junction \\ \hline
1MMS & 58 & 20 & 86 & 74 & 82 & 0.7814 & 3-way junction \\ \hline
3EGZ & 65 & 23 & 72 & 70 & 66 & 0.6849 & 3-way junction (riboswitch) \\ \hline
2QUS & 69 & 26 & 81 & 75 & 76 & 0.7577 & Pseudoknot, 3-way junction \\ \hline
1KXK & 70 & 28 & 87 & 96 & 92 & 0.9440 & Hairpin, internal loop \\ \hline
2DU3 & 71 & 27 & 75 & 78 & 70 & 0.7433 & 4-way junction (tRNA) \\ \hline
\textbf{2OIU} & \textbf{71} & \textbf{29} & \textbf{84} & \textbf{90} & \textbf{83} & \textbf{0.8692} & \textbf{3-way junction (riboswitch)} \\ \hline
1SJ4 & 73 & 19 & 83 & 78 & 81 & 0.7976 & Pseudoknot, 4-way junction \\ \hline
1P5O & 77 & 29 & 86 & 97 & 77 & 0.8716 & Hairpin, internal loop \\ \hline
3D2G & 77 & 28 & 103 & 80 & 88 & 0.8435 & 3-way junction (riboswitch) \\ \hline
2HOJ & 79 & 27 & 100 & 87 & 84 & 0.8572 & 3-way junction (riboswitch) \\ \hline
\textbf{2GDI} & \textbf{80} & \textbf{32} & \textbf{100} & \textbf{84} & \textbf{80} & \textbf{0.8197} & \textbf{3-way junction (riboswitch)} \\ \hline
2GIS & 94 & 36 & 125 & 87 & 82 & 0.8485 & Pseudoknot, 4-way junction (riboswitch) \\ \hline
1LNG & 97 & 38 & 124 & 85 & 79 & 0.8254 & 3-way junction (SRP) \\ \hline
\textbf{1MFQ} & \textbf{128} & \textbf{49} & \textbf{164} & \textbf{81} & \textbf{76} & \textbf{0.7895} & \textbf{3-way junction (SRP)} \\ \hline
\end{tabular}
\end{table}
Given the recent progress made in RNA secondary structure prediction \cite{LaingAndSchlick2010}, \cite{ReinharzEtAl2013}, we believe that canonical base pairs may be routinely predicted with a fair accuracy. Therefore, we have allowed the program BkTree to accept known or predicted canonical base pairs along with the query sequence as input. Note that the knowledge of canonical base pairs does not necessarily imply the whole secondary structure, which is often a part of input to most of the existing RNA 3D prediction methods. In our test, we extracted canonical base pairs of a RNA from FR3D analyzed interactions \cite{SarverEtAl2008}.
\subsection{Overall Performance}
We evaluated the quality of the predicted nucleotide interactions by the sensitivity (STY) and positive predictive value (PPV) against the FR3D-analyzed interactions \cite{SarverEtAl2008}. In order to take into account the effects of both true positive and false positive rates in one measure, the \textit{Matthews correlation coefficient} (MCC), defined in \cite{LaingAndSchlick2010} as MCC $:= \sqrt{\mathrm{PPV}\times \mathrm{STY}}$, was also calculated.
Table \ref{withBbsTable} summarizes the overall performance of BkTree on the benchmark set. On a large majority of RNAs, the sensitivity is decently high. Note that the STY and PPV calculations excluded the canonical base pairs.The sensitivity result indicates that our method has a high accuracy in identifying non-canonical interactions that may be crucial to tertiary structures. This is true even for those longer RNAs. We further note that for the 7 RNAs that were not included in the training data, BkTree also performed extremely well.
\subsection{Performance Comparison with Other Methods}
We compared our program BkTree with the programs MC, Rosetta, and NAST on the capability to predict nucleotide interactions. These other methods had been surveyed and evaluated in \cite{LaingAndSchlick2010} based on their ability to identify both base pairing and base stacking interactions. We removed base-phosphate and base-ribose interactions from our prediction results. We incorporated the canonical base pairs into our results because these other methods include all interactions from the input secondary structure.
Figure~\ref{laingMCCCompare} shows the MCC curves for MC, Rosetta, NAST, and BkTree on the benchmark set of RNAs. Data of RNAs failed by a program were not included in the calculation. We note that for every RNA, these other programs produced more than one conformation so the results were averaged for these comparisons. The figure demonstrates that BkTree overall outperformed the other three programs in predicting non-canonical base pairing and base stacking interactions.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.38]{mcc_4_methods.png}
\caption{Comparison of the MCC generated by MC, NAST, Rosetta and BkTree. The MCC of the 43 RNAs are calculated by including canonical base pairs in the results and sorted by their lengths. The plot was derived by merging the results obtained by BkTree and the data computed in the survey \cite{LaingPersonal2014,LaingAndSchlick2010}. In that survey, the tertiary structure predictions with the other 3 methods were based on resolved secondary structures and the secondary structures were included in the calculations. Therefore, the canonical base pairs were also been added to the perdiction results by BkTree. }
\label{laingMCCCompare}
\end{figure}
Table~\ref{avg4methods} gives comparisons on average performance between the four methods. In general, Bktree produced much better average results than Rosetta and NAST, and comparable average results with MC, for which BkTree shows better average STY value than MC, whereas MC gives better average PPV. On MCC values, BkTree had an edge over MC. On RNAs of length $\geq 50$, BkTree maintained almost the same average MCC as it did on the whole set.
\begin{table}[!ht]
\centering
\caption{Average performances of MC, Rosetta, NAST and BkTree, with results in two categories: average over all successfully resolved RNAs and average over all successfully resolved RNAs of length $>50$. The best performance data are displayed in bold. }
\begin{tabular}{c|c|c|c|c|c|c|c|c|}
\cline{2-9}
& \multicolumn{4}{c|}{All RNAs} & \multicolumn{4}{c|}{RNAs of length \textgreater50} \\ \cline{2-9}
& Success/Total & STY & PPV & MCC & Success/Total & STY & PPV & MCC \\ \hline \hline
\multicolumn{1}{|c|}{MC} & 21/43 & 80.7 & \textbf{86.2} & 0.8344 & 6/18 & 77.1 & \textbf{86.0} & 0.8145 \\ \hline
\multicolumn{1}{|c|}{Rosetta} & 43/43 & 62.8 & 80.3 & 0.7101 & 18/18 & 53.4 & 78.5 & 0.6474 \\ \hline
\multicolumn{1}{|c|}{NAST} & 30/43 & 44.5 & 68.2 & 0.5508 & 12/18 & 44.0 & 71.4 & 0.5604 \\ \hline
\multicolumn{1}{|c|}{BkTree} & 43/43 & \textbf{88.6} & 81.3 & \textbf{0.8482} & 18/18 & \textbf{86.0} & 82.7 & \textbf{0.8433} \\ \hline
\end{tabular}
\label{avg4methods}
\end{table}
\subsection{Significance to 3D conformation prediction}\label{3dConfSection}
To evaluate the significance of our method to 3D conformation prediction, we used MC-Sym \cite{ParisienAndMajor2008} to model 3D conformations from the interactions predicted by BkTree and calculated RMSDs against the resolved structures. We note that MC-Sym does not accept interactions of categories other than base-pair and base stacking; the correctly predicted base-phosphate and base-ribose interactions by our methods were discarded by MC-Sym to produce 3D folds. The \textit{deviation index} (DI) \cite{ParisienAndCruz2009}, a measure that accounts for both RMSD and MCC, defined as the quotient of them, was also calculated. Table \ref{compare_Laing} presents the performance values on the 4 representative RNAs chosen in \cite{LaingAndSchlick2010} which typically contain two hairpins and two junctions. Since both MC and Rosetta allow prediction of multiple optimal or suboptimal folds, we chose the averaged values of their solutions. We note that to model 3D conformations with MC using our predicted interaction data, we needed the secondary structure of the tested RNA to be covered by the input canonical base pairs together with the interactions predicted by BkTree. RNA 2QUS failed on this requirement. The averaged RMSDs achieved by BkTree for the rest 3 RNAs are significantly smaller than those achieved by MC and Rosetta.
\begin{table}[!ht]
\centering
\caption{List of performance values predicted using MC, Rosetta and BkTree on 4 representative RNAs chosen by \cite{LaingAndSchlick2010}. The results generated MC and Rosetta are obtained from the survey paper \cite{LaingPersonal2014,LaingAndSchlick2010}. For every RNA, the best results are displayed in bold.}
\begin{tabular}{cc|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\cline{3-17}
& & \multicolumn{5}{c|}{MC} & \multicolumn{5}{c|}{Rosetta} & \multicolumn{5}{c|}{BkTree} \\ \hline
\multicolumn{1}{|c|}{PDB} & Length & STY & PPV & MCC & RMSD & DI & STY & PPV & MCC & RMSD & DI & STY & PPV & MCC & RMSD & DI \\ \hline \hline
\multicolumn{1}{|c|}{1KXK} & 70 & 81 & 89 & 0.849 & 9.49 & 11.16 & 74 & 85 & 0.793 & 17.23 & 21.69 & 97 & 94 & \textbf{0.9589} & \textbf{8.33} & \textbf{8.68} \\ \hline
\multicolumn{1}{|c|}{1XJR} & 47 & 76 & 87 & 0.8131 & 8.74 & 10.74 & 71 & 83 & 0.7676 & 11.63 & 15.21 & 91 & 84 & \textbf{0.8782} & \textbf{6.00} & \textbf{6.83} \\ \hline
\multicolumn{1}{|c|}{2OIU} & 71 & 76 & 92 & 0.8361 & 16.85 & 20.14 & 63 & 87 & 0.7403 & 18.10 & 24.72 & 92 & 86 & \textbf{0.8925} & \textbf{13.21} & \textbf{14.8} \\ \hline
\multicolumn{1}{|c|}{2QUS} & 69 & 78 & 86 & \textbf{0.819} & 18.41 & 22.44 & 58 & 86 & 0.7062 & 15.73 & 22.80 & 80 & 80 & 0.8 & - & - \\ \hline
\end{tabular}
\label{compare_Laing}
\end{table}
\section{Discussion and Conclusion}\label{disSection}
Our method is the first to {\it ab initio} predict RNA non-canonical interactions of all types. Evaluation of the results have highlighted its potential as an important step toward accurate {\it ab initio} 3D structure prediction. We attribute the encouraging preliminary results to the recent growth of knowledge in high-resolution nucleotide interaction data as well as to the novel backbone $k$-tree modeling of nucleotide interaction relationships. The latter makes it possible to markedly reduce the space of solutions for the nucleotide interaction prediction problem to one that can be feasibly searched in polynomial time.
Our method differs from others also in its direct prediction of nucleotide interactions whereas the others mostly attempt 3D conformation construction before producing nucleotide interactions. The difference makes it difficult to compare their performances, especially when a 3D structure is not the direct output of a software, e.g., RNA-MoIP \cite{ReinharzEtAl2013}. Therefore, the MCC comparison with MC was probably more appropriate than the comparison with RNA-MoIP, since the results of MC were based on interactions from the RNA Structure Atlas and so did BkTree, while RNA-MoIP used Interaction Network Fidelity \cite{GendronEtAl2001} in calculating the MCC values. The contrast is more evident when using MC-Sym to model 3D conformations from interaction data predicted by BkTree. Even though the predicted base-phosphate and base-ribose interactions have to be discarded, the resulted RMSDs seem to correlate with the MCC values (Table~\ref{compare_Laing}).
The evaluation tests have also revealed some issues with BkTree. First, the complexity of structures has an impact on our prediction results. Typically, BkTree underperformed on some of the RNAs with pseudoknots or 4-way junctions. Table~\ref{compare_Laing} shows that BkTree loses to MC on MCC value for only one representative RNA 2QUS, which contains a pseudoknot. The underperformance is likely due to the 3-tree model that is a little too weak for complex structures. For example, the best 3-tree can include at most 83 interactions out of total 95 interactions of tRNA 2DU3, indicating a higher treewidth is needed for the NIR graph of this RNA. To improve prediction performance for such RNAs, an algorithm may need to be based on the backbone $4$-tree model. Our method is not ineffective for handling multi-way junctions or pseudoknots, e.g., RNA 2GIS in Table \ref{withBbsTable}. Fixing a specific $k$-tree model, it is the NIR graph treewidth of an RNA that determines the performance on the RNA.
Second, the NIR graph treewidth is also related to scalability of our method. The current algorithm for the nucleotide prediction problem has the complexity $O(n^3)$ for both time and memory requirements. With a large hidden constant in the polynomial, the implemented program BkTree typically runs in 2 to 3 hours on an RNA of length 100 and uses several Gigabytes of memory. This is because the current prototype has aimed at accuracy without optimization in computational efficiency. However, the problem (based on the $k$-tree model) has an inherent complexity of $O(n^k)$; our method is scalable to suit longer and more complex RNAs, e.g., which require the $4$-tree model.
Third, due to the lack of tools to model 3D conformations from nucleotide interactions of all types, it is an immediate future task of ours is to develop such a tool that can be pipelined with a program like BkTree for {\it ab initio} 3D structure prediction. We perceive such a task to be feasible. This is because the output of program BkTree contains not only the predicted nucleotide interactions but also a backbone $3$-tree that decomposes nucleotides according to their interconnectivity. The given 3-tree can be the basis for very efficient algorithms for computing a desirable optimization function on 3D conformations \cite{ArnborgAndProskurowski1989}.
\section*{Acknowledgments}
We thank Christian Laing for the provided raw data used in the survey paper \cite{LaingAndSchlick2010}. This work was supported in part by NSF IIS grant (award No: 0916250).
|
\section{Introduction}
One long-standing question in epidemiological research is how best to allocate limited amounts of vaccine or similar preventative measures in order to minimize the severity of an epidemic.
Discovering a way to reduce the severity of epidemics would benefit society by not only preventing loss of life, but also reducing the overall disease burden of those epidemics.
These burdens, or societal costs, can be seen in reduced economic productivity due to widespread sickness, or increased strain on health infrastructure.
Much of the literature on the problem of vaccine allocation has focused on influenza epidemics and has used mathematical models of epidemic spread to determine the effectiveness of proposed methods \cite{Matrajt2010,Medlock2009,Mylius2008,Tuite2010,Wallinga2010}.
Most investigations of vaccine allocation performed their analyses on geographically large scales and focused on determining which sub-populations should be prioritized for vaccination.
Since previous research focuses on large-scale, subpopulation-based models, our work explores alternative approaches to the question of vaccine allocation.
Much work has been devoted to the development of computational models for disease simulation.
Mikler \emph{et. al.} have developed multiple stochastic models of disease spread in a population, which use the standard SEIR(S) model \cite{Mikler2009, Reyes-Silveyra2011}.
Some of these models, such as their Global Stochastic Cellular Automata model \cite{Mikler2005}, can be adapted to the problem of vaccine allocation.
Simulation techniques relevant to our work have been developed by Indrakanti \cite{Indrakanti2012}, who implemented a SEIR-based framework for simulating epidemic spread within a county.
Our work applies computational models of epidemics to the problem of geographically allocating a limited number of vaccines within several Texas counties.
We developed a graph-based, stochastic model for epidemics that is based on the SEIR model and the work of Mikler \emph{et. al.} \cite{Mikler2005}.
Various centrality measures have been proposed over the years \cite{Nieminen1973,Bonacich1972,Bonacich2007,Freeman1977,Brandes2001}, mainly in the field of social network analysis, which provide means for determining which nodes in a graph are the most important.
Our model was then used to investigate various centrality-based vaccine allocation strategies, in which vaccines are allocated to various census blocks within the county in order of their centrality measure score.
This approach provides an alternative method for addressing the vaccine allocation problem, which can be combined with more conventional approaches to yield more effective epidemic suppression strategies.
In this paper, we present the results of several experiments with centrality-based vaccine allocation strategies.
These experiments were performed on graphs constructed from the census data of several Texas counties, and have yielded results that indicate directions for further study.
\section{Background}
\subsection{Vaccine Allocation}
Several authors have previously addressed the problem of vaccine allocation.
Medlock and Galvani have investigated the question of which groups should be prioritized for influenza vaccination throughout the United States in the event of a shortage \cite{Medlock2009}.
They developed a model, parameterized with real-world data, for analyzing various vaccination strategies under multiple effectiveness criteria.
Medlock and Galvani found that optimal epidemic control could be achieved by targeting adults between the ages of 30 and 39 and children, and that current CDC vaccine allocation recommendations were suboptimal under all considered measures.
Tuite \emph{et. al.} used similar methods to investigate vaccine allocation strategies in Canada during a H1N1 epidemic \cite{Tuite2010}.
They found that targeting high-risk groups, such as people with chronic medical conditions, and then targeting age groups, such as children, that are the most likely to develop complications after influenza infection will cause the greatest reduction in the number of deaths and serious illnesses caused by an influenza epidemic.
Similar work was performed by Matrajt and Longini \cite{Matrajt2010}.
In their work, Matrajt and Longini attempted to determine whether the optimal vaccine allocation strategy varies with the state of the epidemic.
They developed an SIR-based model to analyze the spread of influenza among and between several groups of people in idealized countries, and used criteria similar to those of Medlock and Galvani to ascertain the effectiveness of their methods.
The authors determined that the demographic structure of the population in an area significantly affects the optimal allocation strategy.
They also found that, at the beginning of an epidemic, those groups that are the most likely to transmit influenza should be targeted immediately, while the groups that are most vulnerable should be targeted after a point before, but near, the peak of the epidemic.
The work of Mylius \emph{et. al} modeled influenza epidemics and found that individuals that are most at risk for complications should be targeted if additional vaccines become available during an epidemic \cite{Mylius2008}.
Mylius \emph{et. al.} also found that schoolchildren should be prioritized for vaccination at the beginning of an epidemic, as they are heavily involved in disease transmission.
Their conclusions bear a close resemblance to those of Matrajt and Longini.
The problem of large-scale vaccine allocations were addressed by \cite{Matrajt2010,Medlock2009,Mylius2008,Tuite2010}.
They used different models to reach their conclusions, but used similar metrics, such as the reduction in the predicted number of deaths from a disease after the vaccination program, to assess the significance of their results.
None of these works addressed the question of vaccine allocation at the level of counties or other similar geographic divisions, nor did they address the vaccine allocation problem from a geographic perspective.
Another approach to the vaccine allocation problem, which we extend in our work, can be found in the work of Johnson \cite{Johnson2010}.
Johnson's approach, unlike those of many other researchers, focuses on vaccine allocation in the small scale.
Johnson generated synthetic social network graphs and used various measures of centrality to determine which individuals within those graphs should be vaccinated.
Her results showed that the optimal vaccination strategy depends on the structure of the social network it is applied to.
Our work adapts Johnson's methods to the problem of allocating vaccinations within a county.
\subsection{Centrality and its Applications}
\label{sec:centrality-app}
The notion of centrality has been used in various fields, most notably in the study of social networks.
Multiple centrality measures exist, all of which allow the vertices of a graph to be ranked in order of their importance.
Out-Degree and in-degree centralities were first defined by Nieminen \cite{Nieminen1973}.
In those measures, the centrality $c$ of a node $n$ in a graph $G = (V,E)$, where $V$ is the set of nodes in the graph, $u$ and $v$ are nodes in V, $E$ is the set of edges, and $|V|$ is the number of nodes is
\begin{equation}
c = \frac{d_n}{|V| - 1}
\end{equation}
where $d_n$ is the out-degree or in-degree of $n$.
Eigenvector centrality, first proposed by Bonacich \cite{Bonacich1972}, of a graph $G = (V,E)$ can be calculated by representing $G$ as an adjacency matrix $A$, where
\begin{equation}
A_{uv} =
\begin{cases}
1 & \text{if } (u,v) \in E \\
0 & \text{if } (u,v) \not\in E
\end{cases}
\end{equation}
Then, the eigenvector centrality $x$ is defined as
\begin{equation}
Ax = \lambda x
\end{equation}
where $\lambda$ is the largest eigenvalue of $A$.
Eigenvector centrality is useful because it gives a higher centrality score to a high-degree node that is connected to other high-degree nodes than to a high-degree node connected to low-degree nodes \cite{Bonacich2007}.
Another important measure of centrality is betweenness centrality.
Betweenness centrality, first described by Freeman \cite{Freeman1977}, is defined as follows, for a node $n$ in a graph $G = (V,E)$, where $d_{st}(n)$ is the number of shortest paths from $s \in V$ to $t \in V$ that include $n$, and $d_{st}$ is the number of shortest paths from $s$ to $t$ where the shortest path between two nodes is the path where the sum of the weights of its edges is minimized
\begin{equation}
\sum_{s \not = n \not = t \in V} \frac{d_{st}(n)}{d_{st}}
\end{equation}
Betweenness centrality gives higher centrality scores to nodes that are most likely to be involved in the transmission of information throughout a graph \cite{Brandes2001}.
Centrality has been applied to multiple epidemiological questions.
Rothenberg \emph{et. al.} applied various centrality measures to a social network graph generated from CDC data on the spread of HIV in Colorado Springs \cite{Rothenberg1995}.
Rothenberg analyzed the collected data under multiple centrality measures and found that all of the measures identified all but one of the HIV cases as non-central.
Their work also noted several differences in response patterns to the CDC questionnaire between people with high centrality and people with low centrality under all measures, which were closely correlated.
Similar work was also performed by Christley \emph{et. al.} \cite{Christley2005}.
Christley generated synthetic social networks and performed SIR-based simulations in order to study the applicability of centrality to the problem of identifying individuals at high risk of HIV infection.
They found that degree centrality performed no worse than other centrality measures, such as betweenness, but noted that the results might not be valid for larger datasets.
Centrality has also played a key role in Johnson's work, which was discussed previously \cite{Johnson2010}.
Johnson investigated how vaccinating those individuals identified as central in a social network would affect the spread of disease in that network.
Johnson found that vaccination strategies based on centrality measures were an effective means of mitigating outbreaks.
Related work was performed by Rustam, who applied centrality to the spread of viruses and worms in computer networks \cite{Rustam2006}.
He found that nodes with high centrality scores have a large amount of influence on the spread of a virus, especially when those nodes are rated central by multiple measures.
\subsection{Computational Simulation of Epidemics}
Several methods for simulating epidemics computationally are based on the SEIR model of epidemics \cite{Mikler2009,Reyes-Silveyra2011,Mikler2005,Indrakanti2012}.
In the SEIR model, the population is divided into four classes: susceptible individuals, who can become infected; exposed or latent individuals, who have been infected but are not capable of spreading the infection; infectious individuals, who can spread the disease to susceptible individuals; and recovered individuals, who can no longer be infected.
Mikler \emph{et. al.} proposed a SEIR-based model that uses a stochastic cellular automaton \cite{Mikler2005}.
The concepts used in their model are applicable to a wide range of epidemic simulations.
Agent-based and metapopulation models are the two dominant methods for computational epidemic simulation \cite{Ajelli2010}.
In agent-based models, each individual within the population is simulated.
Such models, which are often implemented using cellular automata \cite{Mikler2005,Mikler2009,Fu2003}, are useful because they provide insights into the progression of the disease.
However, they are impractical when large populations need to be simulated due to space required for information about each agent in the simulation.
Metapopulation models, on the other hand, break the population into subpopulations and then simulate the interactions between and within the subpopulations \cite{Indrakanti2012,Venkatachalam2006}.
Such models sacrifice some of the precision offered by agent-based modeling in exchange for the ability to simulate large populations.
Our work uses a metapopulation-based model, where each census block constitutes a subpopulation.
This helps to overcome the excessive computational resources that would be required to simulate hundreds of thousands of individuals.
A closely related model that we have adapted for our own work has been developed by Indrakanti \cite{Indrakanti2012}.
Indrakanti developed a county-level epidemic simulator, which uses census blocks as the base unit of simulation.
The model allows for contacts to be generated between any two census blocks, and uses an interaction coefficient to determine the likelihood of contact.
The model was used to conduct several experiments which found that epidemics with higher infectivity (likelihood of spread during a contact) reached their peak percentage of infected individuals earlier.
It also discovered that the distribution of population between census blocks had a significant effect on the spread of disease.
\section{Methods}
\subsection{Representing Counties}
The United States Census Bureau provides geographic data on US counties.
They subdivide counties into census blocks at the finest granularity, which are typically bounded by roads, streets, or water features.
Being that they may be varying sizes and shapes, rural census blocks are often significantly larger than urban ones.
Additionally, the population of a census block may vary greatly as there are many cencus-blocks that contain zero population, i.e. farmland, greenspace, or bodies of water, and others that contain high-desntiy residential structures, i.e. appartment buildings, that may have several hundred people \cite{cblockdef}.
This geographic data, along with the population of each census block, was obtained for multiple Texas counties, specifically Rockwall, Hays, and Denton County, and was stored in a PostGIS database.
The centroid of each census block, which is the average of the points defining the census block, was precomputed and stored within the database.
After the census block data was obtained, several methods for representing a county as a graph with census block nodes were investigated.
Rockwall County was used for these investigations due to its small size, allowing centrality measures to be computed quickly.
All of these representations were based on various methods of constructing a graph $G = (V,E)$ to represent a county, where $V$, the set of nodes, was the set of populated census blocks in the county.
Our first approach to representing counties as a graph, which was ineffective, was a representation where $E = \{(u,v) \mid u, v \in V, u \neq v, \textit{ST\_Touches}(u,v)\}$.
The \emph{ST\_Touches()} function, which determines whether two geographic entities are adjacent, was provided by PostGIS.
This representation used undirected edges.
Another of our early attempts was an approach where $E = \{(u,v) \mid u, v \in V, u \neq v, \textit{ST\_Distance}(\textit{Centroid}(u),v) < r\}$ for various constant values of $r$, where the \emph{ST\_Distance()} function was used to compute the shortest distance between the centroid of $u$ and the block $v$.
The \emph{Centroid()} function retrieves the centroid of a block from the database.
A variation of this approach where the $r$ value for each node $u \in V$ was lowered by multiplying it by
\begin{equation}
1 - \frac{\textit{Population}(u)}{1000}
\end{equation}
on the assumption that the population of highly-populated blocks was more likely to interact within the block than that of sparsely-populated blocks was also investigated.
All of these approaches were ineffective and gave unrealistic results, such as the identification of a large rural block with a population of 5 as the most central block within Rockwall County.
Therefore, they were rejected in favor of a representation based on weighted edges.
The results of our investigations led us to use the following method to represent US counties in our experiments.
In our final model, the county was represented as a directed graph $G = (V,E)$, where $V$ was the set of populated census blocks.
The set of edges $E$ was equal to $\displaystyle \{(u,v,W_{u \rightarrow v} \mid u \in V, v \in V, u \neq v\}$, where $\displaystyle W_{u \rightarrow v}$ was the weight of the edge from $u$ to $v$.
The weights were computed by the following formula, which was adapted from the work of Mikler \emph{et. al.}\cite{Mikler2005}: {\small
\begin{equation}
W_{u \rightarrow v} = \frac{\textit{Population}(u) \times \textit{Population}(v)}{10000 \times \textit{ST\_Distance}(\textit{Centroid}(u), v) + 0.00001}
\label{eqn:weights}
\end{equation}}
This definition assigns higher weights to blocks that are more likely to interact, a fact that is used in our simulator.
The multiplication by $1/10000$ was included to offset the behavior of \emph{ST\_Distance()}, which returns distance as real value less than 1.
If the distances were left unscaled, the division could have produced weights outside of the range that can be accurately represented by real values in a computer.
The addition of $0.00001$ was used to address cases where the centroid of block $u$ is located within block $v$, which ordinarily results in a distance of $0$, from producing an error.
Once these weights were computed, the graphs were represented using the NetworkX library \cite{Hagberg2008-networkx}.
\subsection{Centrality Measures}
We modified the centrality measures described in section \ref{sec:centrality-app} to work with our representation.
Code from \cite{Hagberg2008-networkx} was used as a basis for our work.
Out-degree and in-degree centralities were trivially adapted; the out-degree and in-degree of a node were redefined to be the sum of the weights of the out-edges or in-edges, respectively, instead of a count of those edges.
The implementation of eigenvector centrality in NetworkX already allowed for weighted edges, so it was used without modification.
Since that implementation of eigenvector centrality used only the out-edges for its computations, a variation of the measure, eigenvector-in centrality, was also tested where the in-edges were used instead.
Betweenness centrality was the most heavily modified measure.
Since betweenness centrality relies on shortest paths, the reciprocals of all edge weights were taken before centrality computation, so that the shortest path between nodes $s$ and $t$ is the one that the disease is most likely to spread across.
Afterwards, the betweenness centrality scores were computed by a variation of Brandes's algorithm \cite{Brandes2001}, where Dijkstra's algorithm was applied to each node in parallel as opposed to sequentially.
This modification was found to have no effect on the results of the centrality computation.
The centrality measure described above was reported as inverse betweenness centrality.
Finally, we defined random centrality, which is not a centrality measure, for use as a baseline.
In random centrality, each node is assigned a random real value in $[0,1)$, which is used as its centrality score.
The results of all centrality calculations except random centrality, which was re-computed before each set of simulations, were saved to avoid expensive recomputation when multiple experiments were performed on the same data.
\subsection{Graph-Based Stochastic Epidemic Simulation}
A stochastic epidemic simulator based on the SEIR model and the work of \cite{Indrakanti2012} was developed for use with our representation of counties.
The simulation began by constructing the graph $G = (V,E)$ as described above, and computing all centrality measures that were to be tested.
Each node $n \in V$ was labeled with the following attributes: $|N_n|$, the total population of the block, $|S_n| = |N_n|$, the number of susceptibles, and $|E_n|, |I_n|, |R_n| = 0$, where $|E_n|$ is the number of latent (exposed) individuals in the block, $|I_n|$ the number of infectious, and $|R_n|$ the number of recovered individuals.
Throughout the simulation, $|S_n + E_n + I_n + R_n| = |N_n|$.
In addition to these counts, each node was labeled with a mapping $\displaystyle M_{E,n} = \{ 0 \mapsto E_{0,n}, 1 \mapsto E_{1,n} \ldots T_{lp} \mapsto E_{T_{lp},n}\}$, where $T_{lp}$ is the latent period and $E_{m,n}$ is the number of people in block $n$ with $0 \leq m \leq T_{lp}$ days remaining until they transition to the next simulation state.
In this mapping, $\displaystyle \sum_{m=0}^{T_{lp}} E_{m,n} = |E_n|$.
A similar mapping was also produced for the infectious period, $\displaystyle M_{I,n} = \{ 0 \mapsto I_{0,n} \ldots T_{ip} \mapsto I_{T_{ip},n}$, where $T_{ip}$ is the infectious period.
For this mapping, $0 \leq m \leq T_{ip}$, similarly to $M_{E,n}$.
These attributes are illustrated in Figure \ref{fig:data-model}.
\begin{figure}
\centering
\includegraphics[scale=0.5]{data-model}
\caption{An illustration of the grouping of census blocks that have been split by the simulator.}
\label{fig:data-model}
\end{figure}
For the purposes of experimentation, 20 simulation sets were performed in parallel, with the random number generator re-seeded for each set.
Each set consisted of multiple simulations with varying values of available vaccine and centrality measures.
The simulations were performed on a server with four six-core Intel Xeon E7540 processors and 256 GB of RAM.
At the beginning of each set, a proportion $p$ of the census blocks were chosen randomly for initial infection, which remained constant throughout all simulations in the set.
The percentage of the population infected initially was varied between experiments.
Vaccines were distributed to the entire population of each chosen census until the total supply of vaccines was exhausted.
Census blocks were selected for vaccination in a decreasing order based on their centrality.
The centrality measure used to allocate vaccines was varied between experiments so that the measures could be compared for effectiveness.
The set of population (blocks) that was initially targeted for infection was held constant as the centrality measure and percent of population that was able to receive vaccination were varied.
This allowed us to more accurately determine which centrality measure or measures were most effective under varying conditions.
In experiments involving the prevention scenario, vaccines were distributed before the initial infection, while they were distributed six simulated days after initial infection in the intervention scenario.
The value of six days was chosen because at the beginning of day six the initial cases would have transitioned to the recovered state.
After the initial infections, which were recorded as day zero and placed those affected into $\displaystyle E_{T_{E},n}$, the simulation was executed until no latent or infectious individuals remained in the population.
During each day, after any scheduled vaccine distribution, each block's $M_E$ and $M_I$ is updated by setting
$M_E = \{ i \in M_E \mapsto M_E(i + 1), T_{lp} \mapsto 0 \}$, and similarly for $M_I$.
After these updates, the census blocks are iterated over.
For each block $n$ , $i$ contacts are simulated, where $i$ is $I_n$ multiplied by the contact rate, $20$ contacts/day, which was taken from \cite{Indrakanti2012}.
For each contact, a random real value in $[0,1)$ is generated.
If this number is greater than or equal to the mobility parameter, which was $0.99$, the contact takes place within $n$, otherwise, the contact takes place within a different census block.
The mobility parameter expresses the likelihood that an individual will contact someone who resides in a census block that the individual does not reside in.
In the case of non-local contacts, i.e. contacts made between agents in separate blocks, the external block is chosen by the following method.
First, the weights (given by Equation \ref{eqn:weights}) of the out-edges of the originating block are normalized by dividing by the out-degree of the block and then sorted in descending order.
These values were precomputed once as an optimization.
A random target value in $[0,1)$ is generated, and the list of weights is summed element by element until the sum is greater than or equal to the target value.
The node $t$ whose associated weight causes the summation to terminate is chosen as the block for contact.
This process is illustrated by Figure \ref{fig:infection-process}.
\begin{figure}
\centering
\includegraphics[scale=0.45]{infection-process}
\caption{An illustration of the infection process used by the simulator.}
\label{fig:infection-process}
\end{figure}
Once the block that will be contacted is chosen, a disease-spreading contact occurs if a random real number in $[0,1)$ is less than the transmissibility parameter, which was $0.05$.
For non-local contacts, this test is performed before block selection to speed up the simulation significantly.
When the contact occurs, one person is, if possible, removed from $S_t$, the susceptible population of the target block $t$, and added to $E_t$, the latent population.
The new infection is also added to $M_{Et}(T_{lp})$, which ensures that the newly-infected individual remains in the latent period for $T_{lp}$ days before becoming infectious.
After all the contacts are generated, the SEIR states are updated.
For each block $n \in V$, $M_{En}(0)$ people are removed from $E_n$, and added to $M_{In}(T_{ip})$ and $I_n$.
Similarly, $M_{In}(0)$ people are transferred from $I_n$ to $R_n$.
At the conclusion of the updates, the number of people throughout the county in each SEIR state and the percentage of the population in that state are reported.
In addition, the number of people who are infected, that is, either latent or infectious, is reported, along with the associated percentage.
\subsection{Experiments and Parameters}
In all of the experiments performed, the parameters listed in Table \ref{tab:params} were used.
These parameters were not chosen to reflect an extant disease, and were adapted from the work of Indrakanti \cite{Indrakanti2012}.
\begin{table}
\centering
\begin{tabular}{|l r|}
\hline
Contact rate & 20\\
Transmissibility & 0.05\\
Mobility & 0.99\\
Latent period ($T_{lp}$) & 2 days\\
Infectious period ($T_{ip}$) & 3 days\\
Percentage of blocks to infect ($p$) & 1\%\\ \hline
\end{tabular}
\caption{Parameters kept constant throughout all experiments. These parameters serve to characterize the disease being simulated.}
\label{tab:params}
\end{table}
Experiments were performed on Rockwall, Hays, and Denton counties using two epidemic scenarios for each county.
In the prevention scenario, vaccines were distributed before any infection took place, and 5 percent the population of the blocks that were selected for infection was initially infected.
In the intervention scenario 50 percent of the population of the infected blocks were infected initially, and at the beginning of the 6th day of the simulation vaccines were distributed.
These scenarios were intended to simulate a naturally-occurring epidemic of a disease such as influenza or a random, mass-exposure bio-terror attack using perhaps smallpox, respectively.
Each combination of county and scenario received its own run of 20 simulation sets.
Within each simulation set, the following experiments were performed.
The initially infected blocks were held constant for all of these experiments within a set.
The amount of vaccine available was varied between 30 percent, 50 percent, 75 percent, and 90 percent of the population of the county.
Out-degree, in-degree, eigenvector, eigenvector-in, inverse betweenness, and random centralities were tested.
\section{Results}
In Tables \ref{tab:peaksp} and \ref{tab:peaksi}, we present the average peak infected percentage from our experiments.
This data was obtained by first finding the maximum percent infected for each simulation.
Then, the maximums of each of the 20 simulations for each experiment were arithmetically averaged.
The maximum percentage infected at any given time indicates the severity of the epidemic and, consequently, the strain that epidemic places on health resources, such as hospitals.
These averages, along with their corresponding standard deviations, are reported in Table \ref{tab:peaksp} for the prevention scenario, and Table \ref{tab:peaksi} for the intervention scenario.
These results have been rounded to four decimal places to avoid the appearance of false precision.
We also obtained the total number of infected individuals for each simulation execution.
These totals allow us to determine the severity of infections that do not result in outbreaks.
These results were averaged together on a per-experiment basis and plotted on a log scale.
The plot for an intervention scenario in Denton County is presented in Figure \ref{fig:rptotals}.
We also plotted the average percentage of the population that was in each state during every day of the simulation for each experiment to better visualize the progress of the simulated epidemics.
Such a plot for a prevention scenario in Hays County at 50 percent vaccination is included as Figure \ref{fig:hinice}.
We used maps generated by QuantumGIS, which highlighted approximately the top 10 percent most central blocks according to multiple centrality measures, to ensure that our model was producing realistic results.
These maps were also used as an aid in the analysis of our results.
A map of Hays County, showing the 200 most central blocks according out-degree, in-degree, and inverse betweenness centralities, is included as Figure \ref{fig:haysmap}.
This map was created by obtaining the list of the 200 most central census blocks under each centrality measure and marking them with distinct colors.
Blocks that were central according to multiple centrality measures received their own colors.
\input{tables-2}
\begin{figure}
\centering
\includegraphics[scale=0.55]{denton-inter-counts-bars}
\caption{Average total number of infected individuals vs. number of vaccines for Denton County intervention scenario experiments, log scale.}
\label{fig:rptotals}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.4]{hays-prev-nicer}
\caption{The average percentage of the population that was latent, infectious, or infected (either latent or infectious) v. time for Hays County prevention scenario (50 percent vaccination). Points with a value of 0 were omitted from the graph.}
\label{fig:hinice}
\end{figure*}
\begin{figure}
\centering
\includegraphics[scale=0.3]{hays-map-legend}
\caption{200 most central blocks in Hays County according to multiple centrality measures.}
\label{fig:haysmap}
\end{figure}
\section{Discussion}
We found that in-degree and inverse betweenness centralities tended to be the most effective at containing epidemics.
At low vaccination levels, such as 30 percent and 50 percent, where almost all of the population becomes infected under all vaccination strategies, the peak infected percentage was used to compare the effectiveness of various strategies.
Lowering the number of infected people at one time reduces strain on public health infrastructure, allowing the cases that exist to be treated more effectively.
At higher vaccination levels, where no significant peak occurred, the state totals, along with the plots of the course of the epidemics, were used to determine effectiveness.
The peaks were not used to analyze vaccination effectiveness at the 75 percent and 90 percent intervention experiments, as the peaks were found to reflect the degree to which the disease had spread before the vaccines were distributed, and to have little relation to the effectiveness of allocation strategies.
These analyses allowed us to determine which of the tested vaccination strategies was most effective at containing the disease.
In all but one of our experiments, at least one of in-degree or inverse betweenness had lower peaks than the control, which was utilizing a random vaccination strategy.
Neither measure had a lower peak significantly more times than the other, which shows than inverse betweenness and in-degree centralities are both effective in various situations.
In-degree was significantly more effective in the intervention scenario, while inverse betweenness was more effective in the prevention scenario.
In-degree centrality allocated vaccines to those blocks that were the most vulnerable, that is, the most likely to be contacted.
Because it shielded vulnerable blocks from the disease, in-degree was more effective at containing epidemics that had already begun to spread throughout the county at the time of vaccination.
Inverse betweenness centrality, however, allocated vaccines to those blocks most likely to be involved in disease transmission.
This allocation strategy was more effective in the prevention scenario because it blocked off likely transmission paths, which would have allowed the epidemic to spread quickly.
Both of these targeting methods appear to be effective at reducing the severity of epidemics at the county level, though they are effective in different scenarios.
This conclusion agrees with that of Matrajt and Longini, who found that vaccinating the vulnerable and those likely to transmit the disease was an optimal strategy at the national level \cite{Matrajt2010}.
Out-degree centrality was consistently the least effective method of vaccination, with peaks higher than those of random vaccination in most of the 30 percent and 50 percent vaccination experiments.
This resulted from out-degree centrality's tendency to allocate vaccines to high-population areas that are likely to spread the disease once it reaches them, regardless of their probability of infection.
This result is supported by Figure \ref{fig:haysmap}, which shows out-degree targeting mainly small, urban blocks.
Figure \ref{fig:haysmap} also shows that out-degree centrality selected different blocks than the effective measures, which had many blocks in common.
At 75 percent vaccination, at least one of in-degree and inverse betweenness was more effective at halting disease spread than random vaccination in all experiments.
Similar to the 30 percent and 50 percent levels of vaccination, both measures were the most effective about half the time.
In-degree centrality tended to fall slightly faster than inverse betweenness at the beginning of the simulation, though inverse betweenness was more effective at stopping disease spread near the end of the simulation, a results that agrees with those of Matrajt and Longini \cite{Matrajt2010}.
Effectively no disease spread occurred under all measures, at 90 percent vaccination.
As at previous vaccination levels, in-degree and inverse betweenness were each most effective in half of the experiments.
The differences in effectiveness in the prevention and intervention scenarios were much less significant than at lower vaccination levels.
The lack of significant disease spread at 90 percent was a consequence of herd immunity, the primary factor influencing disease spread at high vaccination levels \cite{introepi}.
Due to the initial unvaccinated period, epidemics in the intervention scenario at 90 percent vaccination tended to last longer than those in the prevention scenario.
Eigenvector-based measures, especially eigenvector-in centrality, which targeted vulnerable blocks that were likely to spread the disease to other vulnerable blocks, showed an interesting trend: their effectiveness increased with county size and the prevalence of urban areas.
In fact, in the prevention scenario at 30 percent vaccination in Denton County, eigenvector-in centrality was the most effective, followed by inverse betweenness.
In Rockwall County, which is small and sparsely populated, these measures were the worst performers.
At 75 percent vaccination in Rockwall, eigenvector and eigenvector-in centralities produced epidemic curves similar to those observed at lower vaccination levels, which did not occur for the other measures.
In contrast, eigenvector-based measures, especially eigenvector-in centrality, were more effective in Hays and Denton counties, which are larger and more heavily urbanized.
In several experiments, eigenvector-in centrality outperformed random vaccination, though they were much less reliable than in-degree and inverse betweenness.
Eigenvector-in centrality's effectiveness increased in urbanized counties because eigenvector-based measures favor central blocks that are connected to other central blocks.
Clusters of vulnerable blocks, which eigenvector-in centrality selects, were prevalent in counties such as Hays and Denton than in Rockwall, leading to the increased effectiveness of that centrality measure.
However, as seen in Figure \ref{fig:rptotals}, eigenvector-based measures were still generally ineffective compared to other options.
In-degree centrality, at 30 percent and 50 percent vaccination, caused the longest delay in the peak of the epidemic.
This result arose because contacts within in-degree central blocks, which were more likely to occur than those within non-central ones, did not cause disease spread because those blocks were vaccinated.
This resulted in the disease taking longer to spread because it was forced into less probable paths.
In previous experiments with our simulator, we received unusual data, which we do not represent here.
Specifically, random vaccination was reported as the most effective method for both the intervention and prevention scenarios in Denton County.
These results were most likely a result of the stochastic nature of our simulator.
It is possible, however, that this result was caused by in-degree centrality's preference for the large number of highly-populated blocks in Denton County, which left large areas of the county unvaccinated.
Some of the limitations of our work are our assumptions that the population is homogenous within each block, that contact rates between census blocks are approximated by Equation \ref{eqn:weights}, and that the individual contact rate is constant.
Another limitation is that we are only testing one isolated county at a time.
These assumptions do not reflect reality, but are common approximations that have been used in computational models of epidemics such as the Global Stochastic Cellular Automata model \cite{Mikler2005}.
Future work will primarily focus on addressing the limitations described above.
Methods for simulating non-homoge-\ \ nous populations \cite{Reyes-Silveyra2011,Indrakanti2012} will be incorporated into our model.
Alternative centrality measures and more accurate weighting formulas will be investigated.
We will also attempt to determine what methods would produce results that surpass random vaccination in cases that are currently outliers.
\bibliographystyle{acm}
|
\section{Introduction}
In the \ProblemName{Channel Assignment} problem, we are given a symmetric weight function $w:V^2\rightarrow \mathbb{N}$ (we assume that $0\in\mathbb{N}$).
The elements of $V$ will be called vertices (as $w$ induces a graph on the vertex set $V$ with edges corresponding to positive values of $w$).
We say that $w$ is $\ell$-bounded when for every $x,y\in V$ we have $w(x,y)\le \ell$.
An assignment $c:V\rightarrow\{1,\ldots,s\}$ is called {\em proper} when for each pair of vertices $x,y$ we have $|c(x)-c(y)|\ge w(x,y)$.
The number $s$ is called the {\em span} of $c$. The goal is to find a proper assignment of minimum span.
Note that the special case when $w$ is $1$-bounded corresponds to the classical graph coloring problem.
It is therefore natural to associate the instance of the channel assignment problem with an edge-weighted graph $G=(V,E)$ where $E=\{uv\ :\ w(u,v)>0\}$ with edge weights $w_E:E\rightarrow \mathbb{N}$ such that $w_E(xy)=w(x,y)$ for every $xy\in E$ (in what follows we abuse the notation slightly and use the same letter $w$ for both the function defined on $V^2$ and $E$).
The minimum span is called also the span of $(G,w)$ and denoted by ${\rm span}(G,w)$.
It is interesting to realize the place \ProblemName{Channel Assignment} in a kind of hierarchy of constraint satisfaction problems. We have already seen that it is a generalization of the classical graph coloring.
It is also a special case of the constraint satisfaction problem (CSP).
In CSP, we are given a vertex set $V$, a constraint set $\mathcal{C}$ and a number of colors $d$.
Each constraint is a set of pairs of the form $(v, t)$ where $v\in V$ and $t\in \{1,\ldots,d\}$.
An assignment $c:V\rightarrow\{1,\ldots,d\}$ is {\em proper} if every constraint $A\in\mathcal{C}$ is satisfied, i.e.\ there exists $(v,t)\in A$ such that $c(v)\ne t$.
The goal is to determine whether there is a proper assignment.
Note that \ProblemName{Channel Assignment} corresponds to CSP where $d=s$ and every edge $uv$ of weight $w(uv)$ in the instance of \ProblemName{Channel Assignment} corresponds to the set of constraints of the form $\{(u,t_1),(v,t_2)\}$ where $|t_1-t_2|<w(uv)$.
Since graph coloring is solvable in time $O^*(2^n)$~\cite{bhk:coloring} it is natural to ask whether \ProblemName{Channel Assignment} is solvable in time $O^*(c^n)$, for some constant $c$.
Unfortunately, the answer is unknown at the moment and the best algorithm known so far runs in $O^*(n!)$ time (see McDiarmid~\cite{mcdiarmid}).
However, there has been some progress on the $\ell$-bounded variant.
McDiarmid~\cite{mcdiarmid} came up with an $O^*((2\ell+1)^n)$-time algorithm which has been next improved by Kral~\cite{kral} to $O^*((\ell+2)^n)$ and to $O^*((\ell+1)^n)$ by Cygan and Kowalik~\cite{cygan}. These are all dynamic programming (and hence exponential space) algorithms, and the last one applies the fast zeta transform to get a minor speed-up.
Interestingly, all these works show also algorithms which {\em count} all proper assignments of span at most $s$ within the same running time (up to polynomial factors) as the decision algorithm.
It is a major open problem (see~\cite{kral,cygan,dagstuhl}) to find such a $O(c^n)$-time algorithm for $c$ independent of $\ell$ or prove that it does not exist under a reasonable complexity assumption.
A complexity assumption commonly used in such cases is the Exponential Time Hypothesis (ETH), introduced by Impagliazzo and Paturi~\cite{IP01}.
It states that \ProblemName{3-CNF-SAT} cannot be computed in time $2^{o(n)}$, where $n$ is the number of variables in the input formula.
The open problem mentioned above becomes even more interesting when we realize that under ETH, CSP does not have a $O^*(c^n)$-time algorithm for a constant $c$ independent of $d$, as proved by Traxler~\cite{traxler}.
\heading{Our Results}
Our main result is a new $O^*((2\sqrt{\ell+1})^n)$-time algorithm for the $\ell$-bounded \ProblemName{Channel Assignment} problem. Note that this is the first algorithm which breaks the $(O(\ell))^n$ barrier.
Our algorithm follows the meet-in-the-middle approach (see e.g. Horowitz and Sahni~\cite{HS-JACM74}) and is surprisingly simple, so we hope it can become a yet another clean illustration of this beautiful technique. We show also its (more technical) counting version, which runs within the same time (up to a polynomial factor).
Although we were not able to show that the unrestricted \ProblemName{Channel Assignment} does not admit a $O(c^n)$-time for a constant $c$ under, say ETH, we were able to shed some more light at this issue.
Let us consider some more problems in the CSP hierarchy.
In the \ProblemName{$T$-Coloring}, introduced by Hale~\cite{Hale80}, we are given a graph $G=(V,E)$, a set $T\subseteq\mathbb{N}$, and a number $s\in\mathbb{N}$.
An assignment $c:V\rightarrow\{1,\ldots,s\}$ is proper when for every edge $uv\in E$ we have $|c(u)-c(v)|\not\in T$.
As usual, the goal is to determine whether there exists a proper assignment.
Like \ProblemName{Channel Assignment}, \ProblemName{$T$-Coloring} is a special case of CSP and generalizes graph coloring, but it is incomparable with \ProblemName{Channel Assignment}.
However, Fiala, Kr\'{a}l' and \v{S}krekovski introduced which is a common generalization of vertex list-coloring (a variant of the classical graph coloring where each vertex has a list, i.e., a set of allowed colors), \ProblemName{Channel Assignment} and \ProblemName{$T$-Coloring}.
The instance of the \ProblemName{Generalized List $T$-coloring} is a triple $(G,\Lambda,t,s)$ where $G=(V,E)$ is a graph, $\Lambda : V\rightarrow 2^{\mathbb{N}}$, $t : E \rightarrow 2^{\mathbb{N}}$ and $s\in\mathbb{N}$, where $\mathbb{N}$ denotes the set of all nonnegative integers.
An assignment $c:V\rightarrow\{1,\ldots,s\}$ is proper when for every $v\in V$ we have $c(v)\in\Lambda(v)$, and for every edge $uv\in E$ we have $|c(u)-c(v)|\not\in t(uv)$.
As usual, the goal is to determine whether there exists a proper assignment.
Similarly as in the case of \ProblemName{Channel Assignment}, we say that the instance of \ProblemName{Generalized List $T$-coloring} is $\ell$-bounded if
$\max \bigcup_{e\in E} t(e)\le \ell$.
Very recently, the \ProblemName{Generalized List $T$-coloring} was considered by Junosza-Szaniawski and Rzążewski~\cite{junosza-tcol}. They show \ProblemName{Generalized List $T$-coloring} can be solved in $O^*((\ell+2)^n)$ time, which matches the time complexity of the algorithm of Cygan and Kowalik~\cite{cygan} for \ProblemName{Channel Assignment} (note that an $\ell$-bounded instance of \ProblemName{Channel Assignment} can be seen as an $(\ell-1)$-bounded instance of \ProblemName{Generalized List $T$-coloring}).
In this work we show that most likely one cannot hope for am $O^*(c^n)$-time algorithm for \ProblemName{Generalized List $T$-coloring}.
We even consider a special case of \ProblemName{Generalized List $T$-coloring}, i.e.\ the non-list version where every vertex is allowed to have any color, so the instance is just a triple $(G,t,s)$. We call it \ProblemName{Generalized $T$-Coloring}. We show that, under ETH, \ProblemName{Generalized $T$-Coloring} does not admit a $2^{2^{o\left(\sqrt{n}\right)}} {\rm poly}(r)$-time algorithm, where $r$ is the size of the instance (including all the bits needed to represent the sets $t(e)$ for all $e\in E$). Note that this rules out an $O(n!)$ algorithm as well.
\heading{Organization of the paper}
In Section~\ref{sec:dp} we describe an $O^*((\ell+2)^n)$-time dynamic programming algorithm for $\ell$-bounded \ProblemName{Channel Assignment}. It is then used as a subroutine in the $O^*((2\sqrt{\ell+1})^n)$-time algorithm described in Section~\ref{sec:mim}. In Section~\ref{sec:count} we extend the algorithm from Section~\ref{sec:mim} to counting proper assignments of given span.
Finally, in Section~\ref{sec:hardness} we discuss hardness of \ProblemName{Generalized $T$-Coloring} under ETH.
\heading{Notation}
Throughout the paper $n$ denotes the number of the vertices of the graph under consideration.
For an integer $k$, by $[k]$ we denote the set $\{1, 2, \ldots, k\}$.
Finally, $\uplus$ is the disjoint sum of sets i.e. the standard sum of sets $\cup$
but with an additional assumption that the sets are disjoint.
\section{Yet another $O^*((\ell+2)^n)$-time dynamic programming}
\label{sec:dp}
In this section we provide a $O^*((\ell+2)^n)$-time dynamic programming algorithm for \ProblemName{Channel Assignment}.
It uses a different approach than e.g.\ the algorithm of Kral, and will be used as a subroutine in our faster algorithm.
For a subset $X\subseteq V$ and a function $f:X\rightarrow [\ell+1]$ let $\mathcal{A}_{X,f}$ be the set of all proper assignments $c:X\rightarrow\mathbb{N}$ of the graph $G[X]$ subject to the condition that for every $x\in X$ we have $c(x)\ge f(x)$.
For every subset $X\subseteq V$ and $f:X\rightarrow [\ell+1]$ we compute the value of $T[X,f]$ which is equal to the minimum span of an assignment from $\mathcal{A}_{X,f}$.
Clearly, the minimum span of $(G,w)$ equals to $T[V,f_1]$ where $f_1$ is the constant function which assigns $1$ to every vertex.
The values of $T[X,f]$ are computed by dynamic programming as follows.
First we initialize $T[\emptyset, e_{[\ell+1]}]=0$ (where $e_{[\ell+1]}$ is the only function $f:\emptyset\rightarrow [\ell+1]$).
Next, we iterate over all non-empty subsets of $V$ in the order of nondecreasing cardinality.
In order to determine the value of $T[X,f]$ we use the recurrence relation formulated in the following lemma.
Informally, it uses the observation that there is a minimum-span assignment $c$ such that the vertex $v\in X$ with minimum color $c(v)$ is {\em left-shifted}, i.e.\ $c(v)=f(v)$. Hence we can check all possibilities for $v$ and then the colors of all the other vertices from $X$ have lower bounds in range $\{f(v),\ldots,f(v)+\ell\}$, so we can translate the range back down to $\{1,\ldots,\ell+1\}$ and use the previously computed values of $T[X\setminus\{v\},\cdot]$.
\begin{lemma}
For a subset $X\subseteq V$, a function $f:X\rightarrow [\ell+1]$ and a vertex $v$ define the function $f_v:X\setminus\{v\}\rightarrow [\ell+1]$ given by the formula \[f_v(x)=1+\max\{w(v,x),f(x)-f(v)\} \text{\quad\quad for every $x\in X\setminus\{v\}$}.\]
Then,
\begin{equation}
\label{eq:dp}
T[X,f] = \min_{v\in X}(f(v) + T[X\setminus\{v\},f_v] - 1),
\end{equation}
\end{lemma}
\begin{proof}
Fix $v\in X$.
Denote ${\ensuremath{\mathcal{A}}}_{X,f,v} = \{c \in {\ensuremath{\mathcal{A}}}_{X,f}\ :\ c(v)=f(v)=\min_{x\in X} f(x)\}$.
Then, for every assignment $c\in{\ensuremath{\mathcal{A}}}_{X,f,v}$, for every $x\in X\setminus\{v\}$ we have $c(x)\ge f(v)+\max\{w(v,x),f(x)-f(v)\}$.
Hence, the minimum span of an assignment from ${\ensuremath{\mathcal{A}}}_{X,f,v}$ is equal to $f(v) + T[X\setminus\{v\},f_v]-1$.
It suffices to show that there is an assignment $c^*\in\mathcal{A}_{X,f}$ of minimum span such that $c^*(v)\in A_{X,f,v}$ for some $v\in X$.
Consider an arbitrary assignment $c^*\in\mathcal{A}_{X,f}$ of minimum span. Let $x\in X$ be the vertex of minimum color, i.e.\ $c^*(x)$ is minimum.
If $c^*(x)=f(x)$ we are done. Otherwise consider a new assignment $c^{**}$ which is the same as $c^*$ everywhere except for $x$ and $c^{**}(x)=f(x)$; then $c^{**}$ is proper since $c^*(x)$ is minimal and clearly $c^{**}\in\mathcal{A}_{X,f}$.
The span of $c^{**}$ is not greater than the span of $c^*$ (actually they are the same since $c^*$ has minimal span), so the claim follows.
\end{proof}
The size of the array $T$ is $\sum_{i=0}^n{n\choose i}(\ell+1)^i=(\ell+2)^n$. Computing a single value based on previously computed values for smaller sets takes $O(n^2)$ time, hence the total computation takes $O((\ell+2)^nn^2)$ time. As described, it gives the minimum span only, but we can retrieve the corresponding assignment within the same running time using standard techniques.
\section{The meet-in-the-middle speed-up}
\label{sec:mim}
In this section we present our main result, an algorithm for $\ell$-bounded \ProblemName{Channel Assignment} that applies the meet-in-the-middle technique. Roughly, the idea is to find partial solutions for all possible {\em halves} of the vertex set and then merge the partial solutions efficiently to solve the full instance.
For the clarity of the presentation we assume $n$ is even (otherwise we just add a dummy isolated vertex).
Before we describe the algorithm let us introduce some notation.
For a set $X\subseteq V$, by $\overline{X}$ we denote $V\setminus X$. Moreover, for a function $f:X\rightarrow[\ell+1]$ we define function $\overline{f}:\overline{X}\rightarrow[\ell+1]$ such that for every $v\in\overline{X}$,
\[\overline{f}(v) = 1 + \max(\{1 + w(uv)-f(u)\ :\ uv\in E,\ u\in X\}\cup\{0\}).\]
The values $T[X,f]$ are defined as in Section~\ref{sec:dp}. Our algorithm is based on the following observation.
\begin{lemma}
\label{lem:mim}
The span of $(G,w)$ is equal to
\[\min (T[X,f] + T[\overline{X},\overline{f}] - 1),\]
where the minimum is over all pairs $(X,f)$ where $X\in{V\choose n/2}$ and $f:X\rightarrow[\ell+1]$.
\end{lemma}
\begin{proof}
Let $c^*:V\rightarrow\mathbb{N}$ be a proper assignment of minimum span $s$.
Order the vertices of $V=\{v_1,\ldots,v_n\}$ so that for every $i=1,\ldots,n-1$ we have $c^*(v_i)\le c^*(v_{i+1})$.
Consider the subset $X=\{v_1,\ldots,v_{n/2}\}$.
Let $s_1 = c^*(v_{n/2})$.
Define $f:X\rightarrow[\ell+1]$ such that $f(x)= 1 + \min\{s_1-c^*(x),\ell\}$ for every $x\in X$.
From the definition of $T$ we have $T[X,f]\le s_1$ (because the assignment $x\mapsto 1+s_1-c^*(x)$ belongs to ${\ensuremath{\mathcal{A}}}_{X,f}$ and has span $s_1$).
Moreover, note that for every $v\in\overline{X}$ it holds that
\begin{equation*}
\begin{split}
c^*(v) & \ge \max(\{c^*(u)+w(uv)\ :\ uv\in E,\ u\in X\}\cup\{s_1\}) \\
& = \max(\{s_1 +w(uv) - f(u) + 1\ :\ uv\in E,\ u\in X\}\cup\{s_1\}) \\
& = s_1 - 1 + \overline{f}(v).
\end{split}
\end{equation*}
It follows that $s=\max_{v\in\overline{X}}c^*(v) \ge s_1 - 1 + T[\overline{X},\overline{f}] \ge T[X,f] + T[\overline{X},\overline{f}] - 1$.
Finally we show that $s > T[X,f] + T[\overline{X},\overline{f}] - 1$ contradicts the optimality of $c^*$.
Let $c_1\in{\ensuremath{\mathcal{A}}}_{X,f}$ be an assignment of span $T[X,f]$ and let $c_2\in{\ensuremath{\mathcal{A}}}_{\overline{X},\overline{f}}$ be an assignment of span $T[\overline{X},\overline{f}]$. Consider the following assignment $c:V\rightarrow\mathbb{N}$.
\[
c(x) = \begin{cases}
1 + T[X,f] - c_1(x) & \text{for $x \in X$} \\
T[X,f] + c_2(x) - 1 & \text{for $x \in \overline{X}$}
\end{cases}
\]
One can check that from the definition of $\overline{f}$ it follows that $c$ is a proper assignment.
Moreover, the span of $c$ is equal to $T[X,f] + T[\overline{X},\overline{f}] - 1$. Hence, if $s > T[X,f] + T[\overline{X},\overline{f}] - 1$ then $c^*$ is not optimal, a contradiction.
\end{proof}
From Lemma~\ref{lem:mim} we immediately obtain the following algorithm for computing the span of $(G,w)$:
\begin{enumerate}
\item Compute the values of $T[X,f]$ for all $X\in{V\choose \le n/2}$ and $f: X\rightarrow [\ell+1]$ using the algorithm from Section~\ref{sec:dp}.
\item Find the span of $(G,w)$ using the formula from Lemma~\ref{lem:mim}.
\end{enumerate}
Note that Step 1 takes time proportional to $\sum_{i=0}^{n/2}{n \choose i}(\ell+1)^in^2=O(2^n(\ell+1)^{n/2}n^2)$. The size of array $T$ is clearly
$O(2^n(\ell+1)^{n/2})$. In Step 2 we compute a minimum of ${n \choose n/2}(\ell+1)^{n/2}=O(2^n(\ell+1)^{n/2})$ values.
Hence the total time is $O(2^n(\ell+1)^{n/2}n^2)$.
As described, the above algorithm gives the minimum span only, but we can retrieve the corresponding assignment within the same running time using standard techniques.
We have just proved the following theorem.
\begin{theorem}
For every $\ell$-bounded weight function the channel assignment problem can be solved in $O(2^n(\ell+1)^{n/2}n^2)$ time.
\end{theorem}
\section{An Extension to Counting}
\label{sec:count}
In this section we present an extension of our meet-in-the-middle algorithm which finds the number of proper assignments of span $s$.
This is slightly more technical than the decision algorithm because we need to avoid counting the same assignment more than once. We assume here that $V=\{1,\ldots,n\}$ (we will use the fact that $V$ is linearly ordered).
For $X\in {V\choose n/2}$, function $f:X\rightarrow [\ell+1]$ and value $r=1,\ldots,s$ denote the set of all assignments from $\mathcal{A}_{X,f}$ with span $r$ by $\mathcal{A}_{X,f,r}$.
Let us denote $Q[X,f,r]=|\mathcal{A}_{X,f,r}|$. We will use the recurrence relation formulated in the following lemma.
\begin{lemma}
\label{lem:eq-dp2}
For a subset $X\in {V\choose \le n/2}$, a function $f:X\rightarrow [\ell+1]$ and a vertex $v$ define the function $f_v:X\setminus\{v\}\rightarrow [\ell+1]$ given by the formula
\[f_v(x)=\max\{f(x),1+w(vx),1+[x<v]\} \text{\quad\quad for every $x\in X\setminus\{v\}$}.\]
Also, for a function $f:X\rightarrow \{2,\ldots,\ell+1\}$ define the function $f_{\downarrow}:X\rightarrow [\ell+1]$ given by the formula
\[f_{\downarrow}(x)=\max\{f(x)-1,1\} \text{\quad\quad for every $x\in X$}.\]
Then, for every $X\in {V\choose n/2}$, $f:X\rightarrow [\ell+1]$ and $r=1,\ldots,s$
\begin{equation}
\label{eq:dp2}
Q[X,f,r] = \begin{cases}
\sum_{v\in f^{-1}(1)}Q[X\setminus\{v\},f_v,r] + [r>1]Q[X,f_{\downarrow},r-1] & \text{if $X\ne\emptyset$} \\
[r=1] & \text{otherwise}
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
The proof is by induction on $|X|+r$.
The formula~\eqref{eq:dp2} clearly holds when $X=\emptyset$, since there is exactly one assignment with empty domain, it is proper and its span is 1.
Assume $X\ne\emptyset$.
The set $\mathcal{A}_{X,f,r}$ partitions into two subsets $\mathcal{B}$ and $\mathcal{C}$, where $\mathcal{B}$ contains the assignments which assign color 1 to some vertex and $\mathcal{C}$ contains the remaining assignments.
We can further partition $\mathcal{B}=\bigcup_{v\in f^{-1}(1)}\mathcal{B}_v$, where
\[\mathcal{B}_v = \{c\in\mathcal{B}\ :\ \min c^{-1}(1)=v\}.\]
Define $\mathcal{B}'_v=\{c|_{X\setminus\{v\}}\ :\ c \in \mathcal{B}_v\}$.
Then $|\mathcal{B}'_v|=|\mathcal{B}_v|$.
Consider an arbitrary $c\in \mathcal{B}_v$.
Then for every $x\in X\setminus\{v\}$ we have $c(x)\ge f(x)$, $c(x)\ge f(v)+w(vx)=1+w(vx)$, and if $x<v$ then $c(x)\ge 2$.
In other words, for every $x\in X\setminus\{v\}$ we have $c(x)\ge f_v(x)$ and hence $c|_{X\setminus\{v\}} \in \mathcal{A}_{X\setminus\{v\},f_v,r}$.
It follows that $\mathcal{B}'_v \subseteq \mathcal{A}_{X\setminus\{v\},f_v,r}$.
It is also easy to verify that every assignment $c'\in\mathcal{A}_{X\setminus\{v\},f_v,r}$ can be extended to a proper assignment $c\in\mathcal{B}_v$ by putting $c(v)=1$ and $c|_{X\setminus\{v\}}=c'$.
Hence $\mathcal{A}_{X\setminus\{v\},f_v,r} \subseteq \mathcal{B}'_v$.
It follows that $\mathcal{B}'_v=\mathcal{A}_{X\setminus\{v\},f_v,r}$ and hence $|\mathcal{B}_v|=|\mathcal{A}_{X\setminus\{v\},f_v,r}|=Q[X\setminus\{v\},f_v,r]$, where the last equality follows from the induction hypothesis. We get $|\mathcal{B}|= \sum_{v\in f^{-1}(1)}Q[X\setminus\{v\},f_v,r]$.
If $r=1$ then ${\ensuremath{\mathcal{C}}}=\emptyset$. Assume $r>1$.
It is clear that the assignments in $\mathcal{C}$ are in 1-1 correspondence with the assignments in $\mathcal{C}'=\{c_{\downarrow}\ :\ c\in\mathcal{C}\}$ and the assignments in $\mathcal{C}'$ have span $r-1$. Hence $|\mathcal{C}|=|\mathcal{A}_{X,f_{\downarrow},r-1}|= Q[X,f_{\downarrow},r-1]$, where the last equality follows from the induction hypothesis.
To sum up, \[|\mathcal{A}_{X,f,r}|=|\mathcal{B}|+|\mathcal{C}|=\sum_{v\in f^{-1}(1)}Q[X\setminus\{v\},f_v,r] + [r>1]\cdot Q[X,f_{\downarrow},r-1],\] as required.
\end{proof}
With Lemma~\ref{lem:eq-dp2} it is easy to describe a dynamic programming algorithm which for every subset $X\in {V\choose n/2}$, function $f:X\rightarrow [\ell+1]$ and value $r=1,\ldots,s$ computes the value of $Q[X,f,r]$.
First we initialize $Q[\emptyset, \emptyset,r]=[r=1]$ for every $r=1,\ldots,s$ and next the values of $Q[X,f,r]$ are computed according to Formula~\eqref{eq:dp2}, using previously computed values of array $Q$; to this end we iterate over the triples $(X,f,r)$ in nondecreasing order of $|X|+r$.
The number of triples considered is $O(2^n(\ell+1)^{n/2} s)$ and processing each triple takes $O(n^2)$ time. We have just shown the following.
\begin{lemma}
\label{lem:compute-Q}
There is an $O(2^n(\ell+1)^{n/2} sn^2)$-time $O(2^n(\ell+1)^{n/2} s)$-space algorithm which finds the values of $Q[X,f,r]$ for all subsets $X\in {V\choose n/2}$, functions $f:X\rightarrow [\ell+1]$ and values $r=1,\ldots,s$.
\end{lemma}
If we use just the values of $Q[X,f,r]$ in the merge phase of the meet-in-the-middle approach, it is unclear how to avoid double-counting the same assignments.
To overcome this problem, for a subset $X\subseteq V$, a function $f:X\rightarrow [\ell+1]$ and a value $r=1,\ldots,s$ define $\mathcal{A^*}_{X,f,r}$ as the set of all proper assignments $c:X\rightarrow\mathbb{N}$ of the graph $G[X]$ such that $c$ has span $r$ and for every $x\in X$ , if $f(x)\le \ell$ then $c(x)=f(x)$ and otherwise $c(x)\ge f(x)$. Denote $Q^*[X,f,r]=\mathcal{A^*}_{X,f,r}$.
Observe the following.
\begin{observation}
\label{obs-Q*}
For a subset $X\subseteq V$ and a function $f:X\rightarrow [\ell+1]$ define the function $f_{\leftarrow\ell}:X\setminus f^{-1}([\ell])\rightarrow [\ell+1]$ given by the formula
\[f_{\leftarrow\ell}(x)=\max(\{f(y) + w(yx) - \ell\ :\ {y\in f^{-1}([\ell])}\}\cup\{1\}),\]
for every $x\in X\setminus f^{-1}([\ell])$. Then, for every $X\subseteq V$, $f:X\rightarrow [\ell+1]$ and $r=1,\ldots,s$
\begin{enumerate}[$(i)$]
\item if $f|_{f^{-1}([\ell])}$ is not a proper assignment then $Q^*[X,f,r]=0$;
\item if $f|_{f^{-1}([\ell])}$ is a proper assignment and $r\le \ell$ then
\[Q^*[X,f,r]=[f^{-1}(\{r\})\ne\emptyset\text{ and }f^{-1}(\{r+1,\ldots,\ell+1\})=\emptyset];\]
\item if $f|_{f^{-1}([\ell])}$ is a proper assignment and $r\ge \ell+1$ then
\begin{equation}
\label{eq:dp3}
Q^*[X,f,r] = Q[X\setminus f^{-1}([\ell]),f_{\leftarrow\ell},r-\ell].
\end{equation}
\end{enumerate}
\end{observation}
Now we proceed to the merge phase of our meet-in-the-middle algorithm.
For a function $f:X\rightarrow[\ell+1]$ we define function $\tilde{f}:\overline{X}\rightarrow[\ell+1]$ such that for every $v\in\overline{X}$,
\[\tilde{f}(v) = 1 + \max(\{1+w(uv)-f(u)\ :\ uv\in E,\ u\in X\}\cup\{[v<\max f^{-1}(1)]\}).\]
The role of the function $\tilde{f}$ is similar as $\overline{f}$ in determining the span using the meet-in-the-middle approach; the only difference is that if for some $x\in X$ we have $f(x)=1$ then for every $v\in\overline{X}$, if $v<x$ then $\tilde{f}(v)\ge 2$.
Informally, this helps us to avoid counting the same assignment once for every partition of the ``middle color'' into parts of relevant sizes.
Now we can formulate the counting counterpart of Lemma~\ref{lem:mim}.
\begin{lemma}
\label{lem:counting-mim}
For a given graph $G$, weight function $w$ and integer $s\in\mathbb{N}$ the number of
proper assignments of span $s$ is equal to
\[\sum_{s^*=1}^s\sum_{X\in {V\choose n/2}}\sum_{\substack{f:X\rightarrow[\ell+1]\\f^{-1}(1)\ne\emptyset}} Q^*[X,f,s^*]\cdot Q[\overline{X},\tilde{f},s-s^*+1].\]
\end{lemma}
\begin{proof}
Let $\mathcal{D}$ be the set of all proper assignments of span $s$.
For an assignment $c\in\mathcal{D}$ define a total order of $V$ as follows: for $i,j\in V$ we have $i\prec j$ iff $(c(i),i)\le_{{\rm lex}} (c(j),j)$, where $\le_{{\rm lex}}$ is the lexicographic order. Then $c$ defines a permutation of the vertices $v^c_1\prec v^c_2 \prec \cdots \prec v^c_n$. Then $\mathcal{D}=\biguplus_{s^*=1}^s\mathcal{D}_{s^*}$, where
\[\mathcal{D}_{s^*} = \{c\in\mathcal{D}\ :\ c(v^c_{n/2})=s^*\}\]
Moreover, ${\ensuremath{\mathcal{D}}}_{s^*}=\biguplus_{X\in {V\choose n/2}}{\ensuremath{\mathcal{D}}}_{s^*,X}$, where
\[{\ensuremath{\mathcal{D}}}_{s^*,X} = \{c\in{\ensuremath{\mathcal{D}}}_{s^*}\ :\ \{v^c_1,\ldots,v^c_{n/2}\}=X\}.\]
Finally, \[{\ensuremath{\mathcal{D}}}_{s^*,X}=\biguplus_{\substack{f:X\rightarrow[\ell+1]\\ f^{-1}(1)\ne\emptyset}}{\ensuremath{\mathcal{D}}}_{s^*,X,f},\]
where ${\ensuremath{\mathcal{D}}}_{s^*,X,f}$ is the set of assignments $c\in{\ensuremath{\mathcal{D}}}_{s^*,X}$ such that
for every $x\in X$, if $f(x)\le\ell$ then $c(x)=s^*-f(x)+1$ and if $f(x)=\ell+1$ then $c(x)\le s^*-f(x)+1$. Note that the condition $f^{-1}(1)\ne\emptyset$ is necessary to satisfy the defining condition of ${\ensuremath{\mathcal{D}}}_{s^*}$; in particular $v^c_{n/2}=\max f^{-1}(1)$.
Consider an arbitrary $c\in {\ensuremath{\mathcal{D}}}_{s^*,X,f}$.
Now observe that for every $v\in\overline{X}$ and $u\in X$ such that $uv\in E$, we have
$c(v)\ge \max\{c(u)+w(uv),s^*\}$. Moreover, if $v<v^c_{n/2}$, i.e.\ $v<\max f^{-1}(1)$, then $c(v)\ge s^*+1$. Hence,
\begin{equation*}
\begin{split}
c(v) & \ge \max(\{c^*(u)+w(uv)\ :\ uv\in E,\ u\in X\}\cup\{s^*+[v<\max f^{-1}(1)]\}) \\
& = \max(\{s^* +w(uv) - f(u) + 1\ :\ uv\in E,\ u\in X\}\cup\{s^*+[v<\max f^{-1}(1)]\}) \\
& = s^* - 1 + \tilde{f}(v).
\end{split}
\end{equation*}
It follows that $|{\ensuremath{\mathcal{D}}}_{s^*,X,f}|=Q^*[X,f,s^*]\cdot Q[\overline{X},\tilde{f},s-s^*+1]$, as required.
\end{proof}
From Lemma~\ref{lem:compute-Q}, Observation~\ref{obs-Q*} and Lemma~\ref{lem:counting-mim} we infer the following theorem.
\begin{theorem}
For every $\ell$-bounded weight function the number of all proper assignments of a given span can be computed in $O^*(2^n(\ell+1)^{n/2})$ time.
\end{theorem}
\section{Hardness of \ProblemName{Generalized $T$-Coloring}}
\label{sec:hardness}
In this section we give a lower bounds for the time complexity of \ProblemName{Generalized $T$-Coloring}, under ETH.
To this end we present a reduction from \ProblemName{SetCover}.
The instance of the decision version of \ProblemName{SetCover} consists of a family of sets $\mathcal{S}=\{S_1,\ldots,S_m\}$ and a number $k$.
The set $U=\bigcup\mathcal{S}$ is called \emph{the universe} and we denote $n=|U|$.
The goal is to decide whether there is a subfamily $\mathcal{C}\subseteq\mathcal{S}$ of size at most $k$ such that $\bigcup {\cal C} = U$ (then we say the instance is {\em positive}).
In the following lemma we reduce {\sc Set Cover} to the decision version of \ProblemName{Generalized $T$-Coloring}, where for a given instance $(G,w)$ and a number $s$ we ask whether there is a proper assignment of span at most $s$ (then we say the instance is {\em positive}). We say that an instance $(\mathcal{S},k)$ of \ProblemName{SetCover} is equivalent to an instance $(G,w,k)$ of \ProblemName{Generalized $T$-Coloring} when $(\mathcal{S},k)$ is positive iff $(G,w,k)$ is positive. For every edge $e$ of $G$, every pair $(e,d)$ for $d\in t(e)$ is called a {\em constraint}.
\begin{lemma}
\label{th:SC-to-TC}
Let $(\mathcal{S},k)$ be an instance of \ProblemName{SetCover} with $m$ sets
and universe of size $n$ and let $A\in[1,m]$ and $B\in[1,n]$ be two reals.
Then we can generate in polynomial time an equivalent instance of \ProblemName{Generalized $T$-Coloring} which has $O\left(\frac{n}{B} + \frac{m}{A} \cdot \max\{1, \log A\}\right)$ vertices, $O^*\left(2^A \cdot m^B\right)$ constraints and is $O\left( 2^A\cdot m^B\right)$-bounded.
\end{lemma}
\begin{proof}
For convenience we assume that $A$ and $B$ are natural numbers, since otherwise we round $A$ and $B$ down and the whole construction and its analysis is the same, up to some details.
In the proof we consider coloring of the vertices as placing the vertices on
a number line in such a way that every vertex is placed in the coordinate
equal to its color.
Let ${\ensuremath{\mathcal{S}}} = \{S_1,\ldots,S_m\}$.
We are going to construct a complex instance $(G=(V,E),t,s)$ of \ProblemName{Generalized $T$-Coloring}.
We describe it step-by-step and show some of its properties.
We begin by putting vertices $v_L$ and $v_R$ in $V$ and $t(v_Lv_R)=\{0,\ldots,s-2\}$, i.e.\ in every proper assignment $v_L$ has color $1$ and $v_R$ has color $s$, or the other way around; w.l.o.g.\ we assume the first possibility. We specify $s$ later.
In what follows, whenever we put a new vertex $v$ in $V$, we will specify the set $A(v)$ of its {\em allowed} colors.
Formally, this corresponds to putting $t(v_Lv)=\{d\in\{0,\ldots,s-1\}\ :\ d+1\not\in A(v)\}$.
Our instance will consist of three separate modules (the set choice module, the witness module and the parsimonious module).
By separate we mean they have disjoint sets of vertices $V_S$, $V_U$ and $V_P$ and moreover they have disjoint sets of allowed colors, i.e.\ for $i,j\in\{S,U,P\}$, when $x\in V_i$ and $y\in V_j$ for $i\ne j$ then $A(x)\cap A(y) = \emptyset$. However the modules will interfere with each other by forbidding some distances between pairs of vertices from two different modules.
\begin{figure}[t]
\begin{center}
\begin{minipage}{0.45\textwidth}
\centering
\def3in{2in}
\input{reduction-part-1.pdf_tex}
\caption{The set choice module.}
\end{minipage}
\vspace{0.1\textwidth}
\begin{minipage}{0.45\textwidth}
\centering
\def3in{2in}
\input{reduction-part-2.pdf_tex}
\caption{
The witness module. (The grey areas are the gaps between the
$m^B$ potentially allowed positions.)
}
\end{minipage}
\end{center}
\vspace{-2cm}
\end{figure}
\heading{The set choice module}
The first module represents the sets in ${\ensuremath{\mathcal{S}}}$.
For every $i=1,\ldots,\left\lceil \frac{m}{A} \right\rceil$ the set $V_S$ contains a vertex $s_i$.
Vertex $s_i$ represents the $A$ sets
\[{\ensuremath{\mathcal{S}}}_i=\{S_{(i - 1) \cdot A + 1}, S_{(i - 2) \cdot A + 2}, \ldots, S_{i \cdot A}\}\] (and the last vertex $s_{\left\lceil {m}/{A} \right\rceil}$ represents ${\ensuremath{\mathcal{S}}}_{\left\lceil {m}/{A} \right\rceil}=S_{(\left\lceil {m}/{A} \right\rceil-1)A+1}, \ldots, S_m$).
We also put $A(s_i)=\{1,\ldots,2^A\}$ for every $s_i\in V_S$.
The intuition is that the color $c\in[2^A]$ of a vertex $s_i$ corresponds to a subset ${\ensuremath{\mathcal{S}}}_i(c)\subseteq{\ensuremath{\mathcal{S}}}_i$, i.e.\ the choice of sets from ${\ensuremath{\mathcal{S}}}_i$ to the solution of \ProblemName{SetCover}.
\heading{The witness module}
Let denote the elements of the universe as $e_1, e_2, \ldots, e_n.$
For every $i=1,\ldots,\left\lceil \frac{n}{B} \right\rceil$ the set $V_U$ contains a vertex $u_i$.
Vertex $u_i$ represents the $B$ elements
\[U_i=\{e_{(i - 1) \cdot B + 1}, e_{(i - 2) \cdot B + 2}, \ldots, e_{i \cdot B}\}\] (and the last vertex $u_{\left\lceil {n}/{B} \right\rceil}$ represents $U_{\left\lceil {n}/{B} \right\rceil}=e_{(\left\lceil {n}/{B} \right\rceil-1)B+1}, \ldots, e_n$).
This time vertices $V_U$ do not need to have the same sets of allowed colors, but for every $u\in V_U$ we have $A(u) \subseteq \{1+i\cdot 2^A\ :\ i=1,\ldots,m^B\}$.
Note that every vertex has at most $m^B$ allowed colors and there are gaps of length $2^A-1$ where no vertex is going to be assigned.
We say that a sequence $(S_{w_1},\ldots,S_{w_B})\in{\ensuremath{\mathcal{S}}}^B$ is a {\em witness} for a vertex $u_i\in V_U$ when
\[U_i \subseteq \bigcup_{j=1}^B S_{w_j}.\]
For every $i=1,\ldots,m^B$ color $1+i\cdot 2^A$ corresponds to the $i$-th sequence in the set ${\ensuremath{\mathcal{S}}}^B$ (say, in the lexicographic order of indices); we denote this sequence by ${\ensuremath{\mathcal{W}}}_i$. Then, for every $u\in V_U$,
\[A(u) = \{1+i\cdot 2^A\ :\ \text{${\ensuremath{\mathcal{W}}}_i$ is a witness for u, }i=1,\ldots,m^B \}.\]
The intuition should be clear: color of a vertex $u_i\in V_U$ in a proper assignment represents the choice of at most $B$ sets in the solution of \ProblemName{SetCover} which cover $U_i$.
\heading{The interaction between the set choice module and the witness module}
As we have seen, every assignment $c$ of colors to the vertices determines a choice of a subfamily ${\ensuremath{\mathcal{S}}}(c)\subseteq{\ensuremath{\mathcal{S}}}$, where
${\ensuremath{\mathcal{S}}}(c)=\bigcup_{i=1}^{\left\lceil {m}/{A} \right\rceil}{\ensuremath{\mathcal{S}}}_i(c(i))$.
Similarly, $c$ determines a choice of a subfamily ${\ensuremath{\mathcal{S}}}'(c)\subseteq{\ensuremath{\mathcal{S}}}$, where
${\ensuremath{\mathcal{S}}}'(c)=\bigcup_{u\in V_U}{\ensuremath{\mathcal{W}}}_{c(u)}$.
It should be clear that we want to force that in every proper assignment ${\ensuremath{\mathcal{S}}}'(c)\subseteq {\ensuremath{\mathcal{S}}}(c)$. To this end we introduce edges between the two modules.
\begin{figure}[t]
\centering
\def3in{3in}
\input{reduction-part-1-and-2.pdf_tex}
\caption{
The interaction between a vertex $s_i$ in the set choice module and a vertex $u$ in the witness module.
All the drawn arcs are forbidden distances between $s_i$ and $u$.
Note that for every possible color $1+j\cdot 2^A$ of $u$ the subset of $[2^A]$ excluded
by the forbidden distances in $t(us_i)$ is exactly $F_{i,j}$.
}
\vspace{-0.5cm}
\end{figure}
For $i=1,\ldots,\left\lceil \frac{m}{A} \right\rceil$ and $j=1,\ldots,m^B$ define the following set of forbidden colors
\[F_{i,j} = \{c\in [2^A]\ :\ {\ensuremath{\mathcal{W}}}_j\cap{\ensuremath{\mathcal{S}}}_i \not\subseteq {\ensuremath{\mathcal{S}}}_i(c)\}.\]
The intuition is the following: If a proper assignment colors a vertex $u_i\in V_U$ with color $1+j\cdot2^A$ (i.e.\ it assigns the witness ${\ensuremath{\mathcal{W}}}_j$ to the set $U_i$) then it cannot color the vertex $s_i$ with colors from $F_{i,j}$ (i.e.\ choose this subsets of ${\ensuremath{\mathcal{S}}}_i$ corresponding to these colors), for otherwise ${\ensuremath{\mathcal{S}}}'(c)\not\subseteq {\ensuremath{\mathcal{S}}}(c)$.
{\bf Claim 1} Consider any proper assignment $c:V\rightarrow[s]$.
If for every $i=1,\ldots,\left\lceil \frac{m}{A} \right\rceil$ we have
$c(s_i)\not\in\bigcup_{u\in V_U} F_{i,c(u)}$, then ${\ensuremath{\mathcal{S}}}'(c)\subseteq {\ensuremath{\mathcal{S}}}(c)$.
{\em Proof of the claim:} Consider a set $S_t \in {\ensuremath{\mathcal{W}}}_{c(u)}$ for an arbitrary $u\in V_U$. Then $S_t\in {\ensuremath{\mathcal{S}}}_i$ for some $i$. From the assumption, $c(s_i)\not\in F_{i,c(u)}$, so ${\ensuremath{\mathcal{W}}}_{c(u)}\cap {\ensuremath{\mathcal{S}}}_i \subseteq {\ensuremath{\mathcal{S}}}_i(c)$. Hence, $S_t \in {\ensuremath{\mathcal{S}}}(c)$, as required.
Hence we would like to add some forbidden distances to our instance to make the assumption of Clam 1 hold. To this end, for every $u\in V_U$ and every $s_i\in V_S$ we put
\[t(us_i) = \bigcup_{j=1}^{m^B}\{1+j\cdot 2^A-f\ :\ f\in F_{i,j}\}.\]
In other words, for every possible color $1+j\cdot 2^A$ of $u$ we forbid all distances between $u$ and $s_i$ that would result in coloring $s_i$ with $F_{i,j}$. Then indeed the assumption from Claim 1 holds.
{\bf Claim 2} For any proper assignment $c:V\rightarrow[s]$ we have ${\ensuremath{\mathcal{S}}}'(c)\subseteq {\ensuremath{\mathcal{S}}}(c)$.
{\em Proof of the claim:} We need to verify the assumption in Claim 1. Assume for the contradiction that for some $i$ and some $u\in V_U$ we have $c(s_i)\in F_{i,c(u)}$.
Recall that in a proper assignment $c(u)=1+j\cdot 2^A$ for some $j=1,\ldots,m^B$.
Then $|c(u)-c(s_i)|=1+j\cdot 2^A-c(s_i)\in t(us_i)$, a contradiction.
{\bf Claim 3} For any proper assignment $c:V\rightarrow[s]$ we have ${\ensuremath{\mathcal{S}}}(c)$ covers the universe.
{\em Proof of the claim:} This is an immediate corollary from Claim 2 and the fact that every vertex $u\in V_U$ is colored with a color from $A(u)$.
{\bf Claim 4} For every cover ${\ensuremath{\mathcal{C}}}\subseteq{\ensuremath{\mathcal{S}}}$ of the universe, there is a proper assignment $c:V\rightarrow[s]$ such that ${\ensuremath{\mathcal{S}}}(c)={\ensuremath{\mathcal{C}}}$.
{\em Proof of the claim:} We color $v_L$ and $v_R$ with $1$ and $s$, and every vertex $s_i$ with the color from $[2^A]$ corresponding to the subset ${\ensuremath{\mathcal{S}}}_i\cap {\ensuremath{\mathcal{C}}}$ of ${\ensuremath{\mathcal{S}}}_i$.
For every set $U_i$ for every $e\in U_i$ we pick a set $S_e\in {\ensuremath{\mathcal{C}}}$ that contains $e$ and we build a witness ${\ensuremath{\mathcal{W}}}$ from the sets $S_e$. We color $u_i$ with the color $1+j\cdot 2^A$, where $j$ is the number of ${\ensuremath{\mathcal{W}}}$ in the lexicographic order of all witnesses. It remains to check that the resulting assignment $c$ is proper. The only nontrivial issue is whether for every $u\in V_U$ and $s_i\in V_S$ we have $|c(u)-c(s_i)|\not\in t(us_i)$. It is clear that $|c(u)-c(s_i)|\not\in\{c(u)-f\ :\ f\in F_{i,j}\}$, where $j$ is such that $c(u)=1+j\cdot 2^A$. However, for every $j'\ne j$ the set $\{1+j'\cdot 2^A-f\ :\ f\in F_{i,j'}\}$ is disjoint from $2^A$ (this is where we make use of the `gaps' of length $2^A-1$).
\heading{Bounding the number of sets chosen to the solution}
The last thing we need in a proper assignment $c$ is to keep the number of the sets in ${\ensuremath{\mathcal{S}}}(c)$ bounded by $k$. To this end we use the parsimonious module with the vertex set $V_P$.
The third parsimonious module consists of $\left\lceil \frac{m}{A} \right\rceil$
consecutive submodules and an additional free space of length $2k$ (meaning that for every $v\in V_P$ the set of allowed colors $A(v)$ contains this free space.
Between those submodules and the additional free space we put a gap of length
$2^A$, where no vertex can be assigned.
The intuition is that in a proper assignment $c$ the $i$-th submodule represents the number of sets from ${\ensuremath{\mathcal{S}}}_i$ chosen to the solution, i.e.\ $|{\ensuremath{\mathcal{S}}}_i(c_i)|$.
\begin{figure}[t]
\centering
\def3in{3in}
\input{reduction-part-3.pdf_tex}
\caption{The parsimonious module.\label{fig:disks}}
\vspace{-0.5cm}
\end{figure}
More precisely, $V_P=\biguplus_{i=1}^{\ceil{m/A}} V_i$, where $V_i$ is a set of $1 + \left\lfloor \log A \right\rfloor$ vertices
representing numbers $2^0, 2^1, \ldots, 2^{\left\lfloor \log A \right\rfloor}.$
Let For a vertex $x\in V_P$ let $r(x)$ denote the number represented by $x$.
For every two vertices $x,y\in V_P$ we define
\[t(xy)=\{0,\ldots,r(x)+r(y)-1\}.\]
It follows that we can interpret those vertices as disjoint disks with radii equal to
the represented numbers (see Fig.~\ref{fig:disks}).
Let $q=(1+m^B)2^A$, i.e.\ $q$ is the number of colors used by the first two modules.
For every $i$, we define $i$-th slot as the set of colors $\{q+1+(i-1)\cdot 4A,\ldots,q+i\cdot 4A\}$. Note that the length of each slot is $4A$. Define also the free space as $Q=\{q+\ceil{m/A}\cdot 4A+2^A+1,\ldots,q+\ceil{m/A}\cdot 4A+2^A+2k\}$.
Each vertex $x\in V_i$ is either in $i$-th slot or in the free space $Q$.
However, $x$ has exactly one allowed color in the $i$-th slot chosen so that we can put all the disks in the $i$-th slot and they will be disjoint.
Let $j$ be such that $r(x)=2^j$. Then we denote the allowed color by $a_x=q+(i-1)\cdot 4A+\sum_{r<j}2\cdot 2^r+2^j$.
In precise terms, $A(x)=\{a_x\}\cup Q$.
\begin{figure}[h!]
\centering
\def3in{3in}
\input{reduction-part-1-and-3.pdf_tex}
\caption{
The interaction between the set choice module and one of
the submodules of the parsimonious module.
Note that colors in $[2^A]$ are ordered according to the cardinality of the chosen
collection of sets ($0, 1, 1, 1, 2, 2, 2, 3$).
}
\vspace{-0.5cm}
\end{figure}
Vertices of the $i$-th submodule have some edges to the vertex $s_i$ of the
set choice module. As we mentioned, for a proper assignment $c$ the $i$-th submodule is going to be a counter representing the number of sets in ${\ensuremath{\mathcal{S}}}_i(c)$; in fact the vertex representing $2^j$ corresponds to the $j$-th bit of the counter.
So if $r(x)=2^j$ for $x\in V_i$, then $t(s_ix)$ contains all distances $d$ such that
$a_x-d$ is a color $b$ from $2^A$ such that the $j$-th bit of $|{\ensuremath{\mathcal{S}}}_i(b)|$ is 1.
Hence, in a proper assignment $c$, if the $j$-th bit of the number of sets in ${\ensuremath{\mathcal{S}}}_i(c)$ is 1 then $x$ is thrown away from the $i$-th slot and it is colored by a color from the free space $Q$.
However, the $|Q|=2k$ so the sum of the radii of the disks thrown out from its slots is at most $k.$ It follows that the total number of the chosen sets is also at most $k.$
Also, if there is a cover ${\ensuremath{\mathcal{C}}}\subseteq{\ensuremath{\mathcal{S}}}$ of the universe such that $|{\ensuremath{\mathcal{C}}}|\le k$, then for every $i$, if $|{\ensuremath{\mathcal{C}}}\cap{\ensuremath{\mathcal{S}}}_i|$ has 1 on the $j$-th bit we put the vertex of $V_i$ representing $2^j$ in $Q$. It is clear that since $|{\ensuremath{\mathcal{C}}}|\le k$ we have enough space for them in $Q$. Moreover, we do not violate any edge between these vertices and $V_S$ because of the gap $2^A$ inside the parsimonious module. Together with Claim 4 it implies that $({\ensuremath{\mathcal{S}}},k)$ is a YES-instance of \ProblemName{SetCover} iff $(G,t,s)$ is a YES-instance of \ProblemName{Generalized $T$-Coloring}, provided that $s$ is sufficiently large to provide disjoint intervals of colors for all the modules. From the construction we infer that it is sufficient to put $s = 2^A + 2^A \cdot m^B + 4A \cdot \left\lceil\frac{m}{A}\right\rceil + 2^A + 2k.$
\heading{Calculating the parameters}
Note that $s=O(2^Am^B)$ and in particular our instance is $O(2^Am^B)$-bounded.
Moreover, $|V|=
\left\lceil\frac{n}{B}\right\rceil +
\left\lceil\frac{m}{A}\right\rceil +
\left\lceil\frac{m}{A}\right\rceil \cdot
\left(1 + \left\lfloor\log A\right\rfloor \right)
+ 2
= O\left(\frac{n}{B} + \frac{m}{A} \cdot \left(1 + \log A\right) \right)
= O\left(\frac{n}{B} + \frac{m}{A} \cdot \max\left\{1, \log A\right\} \right)$.
Finally, the total number of constraints is bounded by
$
O^*(
(\frac{n}{B} + \frac{m}{A} \cdot \max\left\{1, \log A\right\} )^2
\cdot (2^A \cdot m^B)
)
= O^*\left(2^A \cdot m^B\right)$, i.e.,
the number of pairs of the vertices times the maximum forbidden distance $s-1$.
It ends the proof.
\ignore{
To denote every forbidden distance as well as the searched span in the query
it is sufficient to use $
\log\left(\left(( 2^A\cdot m^B\right)\cdot {\rm poly}(A, B, m,n) \right)
= A + B\log m + \log\left({\rm poly}(A, B, m, n)\right)
= O\left(A + B\log m + \log n\right) $
bits.
To denote the number of the vertices in the instance as well as to denote
the numbers of specific vertices it is sufficient to use
$O\left(\log(n+m)\right)$ bits.
So clearly to denote every number in our instance of \ProblemName{Generalized $T$-Coloring}
it suffices to use $O\left(A + B\log m + \log n\right)$ bits.
The last thing is to deal with the real values of $A$ and $B.$
We can perform all the above construction for integer numbers $\lfloor A\rfloor$
and $\lfloor B\rfloor$ instead of the real $A$ and $B.$
Then we obtain an instance with $
O\left(
\frac{n}{\lfloor B\rfloor} +
\frac{m}{\lfloor A\rfloor} \cdot \max\left\{1, \log \lfloor A\rfloor\right\}
\right)
= O\left(
\frac{n}{B} +
\frac{m}{A} \cdot \max\left\{1, \log A\right\}
\right)
$ vertices (because if $x \geq 1$ then
$\frac{1}{\lfloor x\rfloor} \leq \frac{2}{x}$),
with the overall number of the forbidden distances $
O^*\left(2^{\lfloor A\rfloor} \cdot m^{\lfloor B\rfloor}\right)
= O^*\left(2^A \cdot m^B\right)$
and with all the numbers in the instance with
$O\left(\lfloor A\rfloor + \lfloor B\rfloor\log m + \log n\right)
= O\left(A + B\log m + \log n\right)$
bits.
}
\end{proof}
\begin{corollary}
\label{cor:DS-to-TC}
Let $(G, k)$ be an instance of \ProblemName{Dominating Set} where $G$
is a graph on $n$ vertices and $k\in\mathbb{N}$.
Then, for any real number $A \in [1,n]$
we can generate in polynomial time an instance of \ProblemName{Generalized $T$-Coloring} with
$O\left(\frac{n}{A} \cdot \max\{1, \log A\}\right)$ vertices
and with $O^*\left((2n)^A\right)$ constraints
and such that all the numbers in the instance have
$O\left(A \cdot \max\left\{1, \log n\right\}\right)$ bits.
\end{corollary}
\begin{proof}
The instance of \textsc{Dominating Set} with $n$ vertices can be transformed to an equivalent instance of \textsc{Set Cover} with $n$ sets and also $n$ elements of
the universe in a standard way (the sets are exactly the neighborhoods of the vertices).
The number $k$ stays the same.
Therefore we can use the Lemma~\ref{th:SC-to-TC} with $A=B$ and $m=n.$
\end{proof}
\begin{theorem}
\label{th:DS-to-TC-complex}
If there exists an algorithm solving \ProblemName{Generalized $T$-Coloring}
in one of the following time complexities:
\begin{enumerate}[$(i)$]
\item $2^{2^{o\left(\sqrt{n}\right)}} {\rm poly}(r),$
\item $2^{n \cdot o\left({\log l}/{(\log\log l)^2}\right)} {\rm poly}(r),$
\end{enumerate}
where $n$ is the number of vertices in the input graph and $r$ is the bit size of the input, then there exists an algorithm solving \textsc{Dominating Set}
in time $2^{o(n)}.$
\end{theorem}
\begin{proof} We begin with proving $(i)$.
Let us assume that we have an algorithm solving \ProblemName{Generalized $T$-Coloring}
in time $2^{2^{f(n)}} {\rm poly}(r)$
where $f$ is some function such that $f(n) = o\left(\sqrt{n}\right).$
We can assume without loss of generality that $f$ is positive and
nondecreasing.
Let $C$ be a constant such that Corollary \ref{cor:DS-to-TC} will give us
always at most $C \cdot \frac{n}{A} \cdot \max\left\{1, \log A\right\}$
vertices.
Let $\alpha$ be a positive nondecreasing function such that
$\alpha(n) \leq \frac{\sqrt{n}}{f(Cn)}$ and $\alpha(n) = \omega(1).$
Such a function always exists because
$\frac{\sqrt{n}}{f(Cn)} = \frac{1}{\sqrt{C}} \cdot \frac{\sqrt{Cn}}{f(Cn)}
= \omega(1).$
For every instance of \ProblemName{Dominating Set} with $n$ vertices we can
take $A = \frac{n}{\alpha\left(\log^2 n\right) \log n}$ and use
Corollary \ref{cor:DS-to-TC} to obtain an instance of \ProblemName{Generalized $T$-Coloring}
with $O\left(\frac{n}{A} \log A\right)
= O\left(\alpha\left(\log^2 n\right) \log^2 n\right)$
vertices and
\[O^*\left((2n)^A\right) = O^*\left(2^{A + A\log n}\right)
= O^*\left(2^{O\left({n}/{\left(\alpha\log^2 n\right)}\right)}\right)
= 2^{o(n)}\]
constraints. Moreover the numbers in the instance have polynomial size,
so the size of the whole instance is $2^{o(n)}.$
Thus this instance can be built in
${\rm poly}\left(n, 2^{o(n)}\right) = 2^{o(n)}$ time.
Then we can solve this instance in
$2^{2^{f\left(C \cdot \frac{n}{A}\log A\right)}}
{\rm poly}\left(2^{o(n)}\right)$
time. But
$f\left(C \cdot \frac{n}{A} \log A\right)
\leq f\left(C \cdot \alpha\left(\log^2 n\right) \log^2 n\right)
\leq \frac{\sqrt{\alpha\left(\log^2 n\right) \log^2 n}}{
\alpha\left(\alpha\left(\log^2 n\right) \log^2 n\right)}
= $ \linebreak $\log n \cdot \frac{\sqrt{\alpha\left(\log^2 n\right)}}{
\alpha\left(\alpha\left(\log^2 n\right) \log^2 n\right)}
= \log n \cdot \frac{\sqrt{\alpha\left(\log^2 n\right)}}{
\alpha\left(\log^2 n\right)}
\cdot \frac{\alpha\left(\log^2 n\right)}{
\alpha\left(\alpha\left(\log^2 n\right) \log^2 n\right)}
\leq \frac{\log n}{\sqrt{\alpha\left(\log^2 n\right)}}
= o\left(\log n \right)$.
So the time of the whole procedure is
$2^{o(n)} + 2^{2^{o\left(\log n\right)}}{\rm poly}\left(2^{o(n)}\right)
= 2^{o(n)}.$
Now we focus on $(ii)$. Let us assume we have an algorithm solving \ProblemName{Generalized $T$-Coloring} in time
$2^{n \cdot f(l)} {\rm poly}(m)$ where $f$ is a positive function such that
$f(l) = o\left(\frac{\log l}{\log^2\log l}\right).$
Let $A = \frac{n}{\log^2 n}.$
For every instance of \ProblemName{Dominating Set} we can use
Corollary \ref{cor:DS-to-TC} to obtain an instance of
\ProblemName{Generalized $T$-Coloring} with
$O\left(\log^2 n \cdot \log \frac{n}{\log^2 n}\right)
= O\left(\log^3 n\right)$ vertices,
$O^*\left((2n)^\frac{n}{\log^2 n}\right)
= 2^{O\left(\frac{n}{\log n}\right)}
= 2^{o(n)}$ constraints and
every number with $O\left(\frac{n}{\log n}\right)$ bits.
We can obtain it in ${\rm poly}\left(n, 2^{o(n)}\right) = 2^{o(n)}$ time.
Note that then $\log l \leq C \frac{n}{\log n}$ for some constant $C.$
The function ${x}/{\log^2 x}$ is nondecreasing for big values of
$x$ so for big values of $n$ we have ${\log l}/{\log^2\log l}
\leq {C \frac{n}{\log n}}/{\log^2 \left(C \frac{n}{\log n}\right)}.$
So we can solve our instance of \ProblemName{Generalized $T$-Coloring} in time
\begin{equation*}
\begin{split}
& 2^{O\left(\log^3 n\right) \cdot
o\left({C \frac{n}{\log n}}/{\log^2\left(C\frac{n}{\log n}\right)}
\right)} \cdot {\rm poly}\left(2^{o(n)}\right)
= 2^{o\left({n\log^2 n}/{\log^2\left(C\frac{n}{\log n}\right)}\right)}
\cdot 2^{o(n)}
= \\
& 2^{o\left({n\log^2 n}/{\left(\log C + \log n - \log\log n\right)^2}
\right)} \cdot 2^{o(n)}
= 2^{o\left({n}/{\left(\frac{\log C}{\log n} + 1
- \frac{\log\log n}{\log n}\right)^2}\right)} \cdot 2^{o(n)} = \\
& 2^{o(n)} \cdot 2^{o(n)}
= 2^{o(n)}.
\end{split}
\end{equation*}
So we have solved the given instance of \textsc{Dominating Set} in time
$2^{o(n)} + 2^{o(n)} = 2^{o(n)}.$
\end{proof}
\begin{corollary}
\label{cor:TC-ETH}
There is no algorithm solving an $n$-vertex instance of \ProblemName{Generalized $T$-Coloring}
with bit size $r$ in any of the listed time complexities
\begin{itemize}
\item $2^{2^{o\left(\sqrt{n}\right)}} {\rm poly}(r),$
\item $2^{n \cdot o\left({\log l}/{\log^2\log l}\right)} {\rm poly}(r),$
\end{itemize}
unless the Exponential Time Hypothesis fails.
\end{corollary}
\begin{proof}
Under the ETH assumption there is no algorithm solving \textsc{Dominating Set}
in time $2^{o(n)}$ where $n$ is a number of the vertices (See~\cite{FominKW04}).
Therefore the claim follows immediately from Theorem \ref{th:DS-to-TC-complex}.
\end{proof}
Regarding the first claim the theorem above, we note that there is a $2^{O(n\log l)} {\rm poly}(r)$-time algorithm for \ProblemName{Generalized $T$-Coloring}, see~\cite{junosza-tcol}.
\bibliographystyle{abbrv}
|
\section{Introduction}
For any natural $d \ge k \ge 2$ we calculate the cohomology groups of the space
of homogeneous polynomials $\R^2 \to \R$ of degree $d$, which do not vanish
with multiplicity $\ge k$ on real lines. For $k=2$ this problem provides the
simplest example of the situation, when the ``finite-order'' invariants of
nonsingular objects are not a complete system of invariants.
The ``affine'' version of this problem (the calculation of the homology group
of the space of polynomials $\R^1 \to \R^1$ with leading term $x^d$ and without
$\ge k$-fold roots) was solved by V.~I.~Arnold in \cite{A89}, see also
\cite{V89}. As in these works, our present calculation is based on the study of
the discriminant set, i.e. of the set of polynomials with forbidden multiple
zeros.
The problem solved below turns out to be more complicated, because an essential
contribution to the homology group comes from the neighborhood of the
``infinitely degenerate'' polynomial equal identically to 0. By this reason,
the method of simplicial resolutions of the discriminant set, solving
immediately the ``affine'' problem, is replaced by its continuous analog:
conical resolution, used previously in \cite{V91}.
Here we have the simplest situation, when the invariants of ``finite order'' of
the space of nonsingular objects do not constitute a complete system of
invariants. Indeed, our spaces of nonsingular polynomials can be considered as
finite-dimensional approximations of the space $\F \sm \Sigma_k$ of smooth
functions $S^1 \to \R^1$ without $k$-fold zeros (if $d$ is odd, then with
values in a nontrivial line bundle); here $S^1$ is realized as a half of the
unit circle in $\R^2.$ By analogy with \cite{V90}, the cohomology classes of
``finite order'' of the space $\F \sm \Sigma_k$ are exactly those, which are
obtained by a natural stabilization of such cohomology groups for approximating
spaces. It turns out, that for $k=2$ and even $d$ all such 0-dimensional
cohomology classes (i.e. the invariants of degree $d$ polynomial functions $S^1
\to \R^1$ without multiple zeros) are polynomials of the number of zeros of
these functions, in particular they cannot separate everywhere positive
functions from the everywhere negative ones. For any finite even $d$ such
polynomial functions can be distinguished by a certain 0-dimensional cohomology
class, arising from the construction of the conical resolution, but this class
is not stable.
Note that a similar example, proving that the system of all finite-order
invariants of knots in $\R^3$ (see \cite{V90}) is not complete, is not
constructed yet.
\section{Main result}
{\sc Notation.} Denote by $HP_d$ the space of homogeneous polynomials $\R^2 \to
\R^1$ of degree $d$, and by $\Sigma_k \subset HP_d$ the set of polynomials
taking zero value with multiplicity $\ge k,$ $k \ge 2,$ on some line in $\R^2$.
For any topological space $X$, $B(X,j)$ denotes its $j$-th configuration space
(i.e. the space of $j$-point subsets in $X$, supplied with a natural topology).
\medskip
{\sc Main Theorem.} {\it 1. If $k$ is even, then the group $\tilde H^*(HP_d \sm
\Sigma_k)$ is free Abelian of rank $2[d/k]+1$, and its free generators have
dimensions $k-2, 2(k-2), \ldots, [d/k](k-2),$ $k-1, 2(k-2)+1, \ldots
,[d/k](k-2)+1$ and $d-2[d/k].$
2. If $k$ is odd and $d$ is not a multiple of $k$, then the group $\tilde
H^*(HP_d \sm \Sigma_k)$ is the direct product of following groups:
(a) for any $p= 1, 2, \ldots, [d/k]$ such that $d-p\cdot k$ is odd, $\Z$ in
dimension $p(k-2)$ and $\Z$ in dimension $p(k-2)+1$;
(b) for any $p= 1, 2, \ldots, [d/k]$ such that $d-p \cdot k$ is even, $\Z_2$ in
dimension $p(k-2)+1$;
(c) $\Z$ in dimension $d-2[d/k]$.
3. If $k$ is odd and $d$ is a multiple of $k$, the answer is almost the same as
in the case 2, only the summand $\Z_2$ in dimension $d-2(d/k)+1$ vanishes.}
\medskip
{\bf Main example.} Let $d$ be even and $k=2$. Then the space $HP_d \sm
\Sigma_k$ consists of $d/2+2$ connected components, two of which (corresponding
to everywhere positive and everywhere negative functions) are contractible, and
all the other are homotopy equivalent to a circle. This homotopy equivalence is
a composition of two: the first maps any polynomial to the collection of its
zero lines (i.e. to an element of the configuration space $B(\RP^1,2p)$ with
appropriate $p \in [1,d/2]$), and the second is the arrow in the following
well-known statement.
\medskip
{\sc Lemma 1.} {\it For any $j,$ there is a locally trivial fibre bundle
$B(S^1,j) \to S^1$, whose fiber is homeomorphic to an open $(j-1)$-dimensional
disc. This fibre bundle is trivial if $j$ is odd and is non-orientable if $j$
is even.} \quad $\Box$ \medskip
{\bf Caratheodory theorem.} In the proof of the main theorem we use the
following fact.
Suppose that a manifold (or finite CW-complex) $M$ is embedded generically in
the space $\R^N$ of a very large dimension, and denote by $M^{*r}$ the union of
all $(r-1)$-dimensional simplices in $\R^N$, whose vertices lie on this
embedded manifold (and the ``genericity'' of the embedding means that if two
such simplices intersect at a certain point, then their minimal faces,
containing this point, coincide). If $M$ is a semialgebraic variety, then by
the Tarski---Seidenberg lemma also $M^{*r}$ is, in particular it has a
triangulation.
\medskip
{\sc Proposition 1} (C.~Caratheodory, see also \cite{Vfil}). {\it The space
$(S^1)^{*r}$ is $PL$-homeomorphic to} $S^{2r-1}$. \quad $\Box$
\section{Proof of the main theorem}
Following \cite{A70}, we use the Alexander duality
\begin{equation}
\label{alex} H^l(HP_d \sm \Sigma_k) \simeq \bar H_{d-l}(\Sigma_k),
\end{equation}
where $\bar H_*$ is the notation for the Borel---Moore homology, i.e. the
homology of the one-point compactification modulo the added point.
To calculate the right---hand group in (\ref{alex}) we construct the {\it
conical resolution} of the space $\Sigma_k$. Let us embed the projective line
$\RP^1$ generically in the space $\R^N$ of a very large dimension, and for any
function $f \in \Sigma_k,$ not equal identically to zero, consider the simplex
$\Delta(f)$ in $\R^N,$ spanned by the images of all points $x_i \in \RP^1,$
corresponding to all possible lines, on which $f$ takes zero value with
multiplicity $\ge k.$ (The maximal possible number of such lines is obviously
equal to $[d/k].$) In the direct product $HP_d \times \R^N$ consider the union
of all simplices of the form $f \times \Delta(f),$ $f \in \Sigma_l \sm 0$. This
union is not closed: the set of its limit points, not belonging to it, is the
product of the point $0 \in HP_d$ and the union of all simplices in $\R^N,$
spanned by the images of no more than $[d/k]$ different points of the line
$\RP^1.$ By the Caratheodory's theorem, the latter union is homeomorphic to the
sphere $S^{2[d/k]-1}.$ We can assume that our embedding $\RP^1 \to \R^N$ is
algebraic, and hence this sphere is semialgebraic. Take a $2[d/k]$-dimensional
semialgebraic disc in $\R^N$ with boundary at this sphere (e.g., the union of
segments connecting the points of this sphere with a generic point in $\R^N$)
and add to the previous union of simplices in $HP_d \times \R^N$ the product of
the point $0 \in HP_d$ and this disc. The obtained set will be denoted by
$\sigma.$ \medskip
{\sc Lemma 2.} {\it The obvious projection $\sigma \to \Sigma_k$ is proper, and
the corresponding map of one-point compactifications of these spaces is a
homotopy equivalence.}
\medskip
This follows easy from the fact that this projection is a stratified map of
semialgebraic spaces, and the preimage of any point $\bar \Sigma_k$ is
contractible, cf. \cite{V}, \cite{Vfil}. \quad $\Box$ \medskip
The space $\sigma$ has a natural increasing filtration: its term $F_p,$ $p \le
[d/k],$ is the union of all $\le (p-1)$-dimensional faces of all simplices,
participating in our construction (or, which is the same, the closure of the
union of all simplices of the form $f \times \Delta(f)$ over all polynomials
$f$ having no more than $p$ forbidden multiple lines), and $F_{[d/k]+1} =
\sigma$.
\medskip
{\sc Lemma 3.} {\it For any $p =1, \ldots, [d/k],$ the term $F_p \sm F_{p-1}$
of our filtration is the space of a locally trivial fiber bundle over the
configuration space $B(\RP^1,p),$ whose fiber is the direct product of an
$(p-1)$-dimensional open simplex and an $(d+1-pk)$-dimensional real space. The
corresponding bundle of open simplices is orientable if and only if $p$ is odd
(i.e. exactly when the base configuration space is orientable), and the bundle
of $(d+1-pk)$-dimensional spaces is orientable if and only if the number
$k(d+1-pk)$ is even.
The last term $F_{[d/k]+1} \sm F_{[d/k]}$ of this filtration is homeomorphic to
the open $2[d/k]$-dimensional disc.}
\medskip
Indeed, to any configuration $(x_1, \ldots, x_p) \subset \RP^1$ there
corresponds the direct product of the interior part of the simplex in $\R^N,$
spanned by the images of points of this configuration under our embedding
$\RP^1 \to \R^N$, and the subspace in $HP_d,$ consisting of polynomials, having
$k$-fold zeros on corresponding $p$ lines in $\R^2.$ The assertion concerning
the orientations can be checked elementary. \quad $\Box$ \medskip
Consider the spectral sequence $E_{p,q}^r,$ calculating the group $\bar
H_*(\Sigma_k)$ and generated by this filtration. Its term $E_{p,q}^1$ is
canonically isomorphic to the group $\bar H_{p+q}(F_p \sm F_{p-1}).$ \medskip
\begin{figure}
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(100.00,150.00)
\thicklines \put(40.00,10.00){\vector(1,0){60.00}}
\put(40.00,10.00){\vector(0,1){145.00}} \thinlines
\put(40.00,56.00){\line(1,0){20.00}} \put(60.00,62.00){\line(-1,0){20.00}}
\put(40.00,62.00){\line(-1,0){32.00}} \put(8.00,68.00){\line(1,0){52.00}}
\put(43.00,6.00){\makebox(0,0)[cc]{\small $1$}}
\put(55.00,6.00){\makebox(0,0)[cc]{\small $3$}}
\put(49.00,6.00){\makebox(0,0)[cc]{\small $2$}}
\put(69.00,41.00){\line(0,1){30.00}} \put(75.00,71.00){\line(0,-1){30.00}}
\put(81.00,41.00){\line(0,1){30.00}} \put(68.00,68.00){\line(1,0){22.00}}
\put(90.00,62.00){\line(-1,0){22.00}} \put(68.00,56.00){\line(1,0){22.00}}
\put(100.00,6.00){\makebox(0,0)[cc]{$p$}} \put(46.00,111.00){\line(0,1){36.00}}
\put(52.00,147.00){\line(0,-1){36.00}} \put(58.00,111.00){\line(0,1){36.00}}
\put(60.00,143.00){\line(-1,0){45.00}} \put(15.00,137.00){\line(1,0){45.00}}
\put(60.00,131.00){\line(-1,0){45.00}} \put(15.00,125.00){\line(1,0){45.00}}
\put(60.00,119.00){\line(-1,0){45.00}} \put(15.00,113.00){\line(1,0){45.00}}
\put(49.00,122.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(43.00,140.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(36.00,151.00){\makebox(0,0)[cc]{$q$}}
\put(72.00,65.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(42.67,76.33){\circle*{0.94}} \put(49.00,76.33){\circle*{0.94}}
\put(55.00,76.33){\circle*{0.94}} \put(61.00,76.33){\circle*{0.94}}
\put(67.00,76.33){\circle*{0.94}} \put(73.00,76.33){\circle*{0.94}}
\put(79.00,76.33){\circle*{0.94}} \put(79.00,84.00){\circle*{0.94}}
\put(73.00,84.00){\circle*{0.94}} \put(67.00,84.00){\circle*{0.94}}
\put(61.00,84.00){\circle*{0.94}} \put(55.00,84.00){\circle*{0.94}}
\put(49.00,84.00){\circle*{0.94}} \put(42.67,84.00){\circle*{0.94}}
\put(43.00,134.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(46.00,111.00){\line(0,-1){18.00}} \put(52.00,93.00){\line(0,1){18.00}}
\put(58.00,111.00){\line(0,-1){18.00}} \put(60.00,107.00){\line(-1,0){45.00}}
\put(15.00,101.00){\line(1,0){45.00}}
\put(55.00,104.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(15.00,95.00){\line(1,0){45.00}}
\put(55.00,98.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(49.00,116.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(72.00,59.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(60.00,44.00){\line(-1,0){20.00}} \put(40.00,44.00){\line(-1,0){32.00}}
\put(8.00,50.00){\line(1,0){52.00}} \put(68.00,50.00){\line(1,0){22.00}}
\put(90.00,44.00){\line(-1,0){22.00}}
\put(78.00,47.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(40.00,38.00){\line(1,0){20.00}} \put(68.00,38.00){\line(1,0){22.00}}
\put(78.00,41.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(81.00,41.00){\line(0,-1){18.00}} \put(75.00,23.00){\line(0,1){18.00}}
\put(69.00,41.00){\line(0,-1){18.00}} \put(90.00,23.00){\line(-1,0){22.00}}
\put(60.00,23.00){\line(-1,0){45.00}} \put(15.00,38.00){\line(1,0){25.00}}
\put(84.00,26.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(28.00,26.00){\makebox(0,0)[cc]{\small $[d/k]-1$}}
\put(26.00,140.00){\makebox(0,0)[cc]{\small $d-(k-1)$}}
\put(26.00,122.00){\makebox(0,0)[cc]{\small $d-2(k-1)$}}
\put(26.00,104.00){\makebox(0,0)[cc]{\small $d-3(k-1)$}}
\put(25.00,47.00){\makebox(0,0)[cc]{\small $d-[d/k](k-1)$}}
\put(90.00,29.00){\line(-1,0){22.00}} \put(15.00,29.00){\line(1,0){45.00}}
\put(78.00,6.00){\makebox(0,0)[cc]{\small $[d/k]$}}
\put(8.00,38.00){\line(1,0){7.00}} \put(8.00,56.00){\line(1,0){32.00}}
\put(58.00,71.00){\line(0,-1){48.00}} \put(52.00,23.00){\line(0,1){48.00}}
\put(46.00,71.00){\line(0,-1){48.00}}
\end{picture}
\caption{First term of the spectral sequence for even $k$}
\end{figure} \begin{figure}
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(100.00,135.00)
\thicklines \put(40.00,10.00){\vector(1,0){60.00}}
\put(40.00,10.00){\vector(0,1){130.00}} \thinlines
\put(74.00,41.00){\line(0,1){30.00}} \put(80.00,71.00){\line(0,-1){30.00}}
\put(86.00,41.00){\line(0,1){30.00}} \put(100.00,6.00){\makebox(0,0)[cc]{$p$}}
\put(51.00,111.00){\line(0,1){17.00}} \put(57.00,128.00){\line(0,-1){17.00}}
\put(63.00,111.00){\line(0,1){17.00}}
\put(36.00,137.00){\makebox(0,0)[cc]{$q$}}
\put(77.00,65.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(48.00,76.33){\circle*{0.94}} \put(54.00,76.33){\circle*{0.94}}
\put(60.00,76.33){\circle*{0.94}} \put(66.00,76.33){\circle*{0.94}}
\put(72.00,76.33){\circle*{0.94}} \put(78.00,76.33){\circle*{0.94}}
\put(84.00,76.33){\circle*{0.94}} \put(84.00,84.00){\circle*{0.94}}
\put(78.00,84.00){\circle*{0.94}} \put(72.00,84.00){\circle*{0.94}}
\put(66.00,84.00){\circle*{0.94}} \put(60.00,84.00){\circle*{0.94}}
\put(54.00,84.00){\circle*{0.94}} \put(48.00,84.00){\circle*{0.94}}
\put(51.00,111.00){\line(0,-1){18.00}} \put(57.00,93.00){\line(0,1){18.00}}
\put(63.00,111.00){\line(0,-1){18.00}}
\put(60.00,104.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(60.00,98.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(77.00,59.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(40.00,44.00){\line(-1,0){32.00}} \put(73.00,50.00){\line(1,0){22.00}}
\put(95.00,44.00){\line(-1,0){22.00}} \put(73.00,38.00){\line(1,0){22.00}}
\put(86.00,41.00){\line(0,-1){18.00}} \put(80.00,23.00){\line(0,1){18.00}}
\put(74.00,41.00){\line(0,-1){18.00}} \put(95.00,23.00){\line(-1,0){22.00}}
\put(15.00,38.00){\line(1,0){25.00}}
\put(89.00,26.00){\makebox(0,0)[cc]{${\mathbb Z}$}}
\put(28.00,26.00){\makebox(0,0)[cc]{$[d/k]-1$}}
\put(25.00,47.00){\makebox(0,0)[cc]{\small $d-[d/k](k-1)$}}
\put(95.00,29.00){\line(-1,0){22.00}} \put(83.00,6.00){\makebox(0,0)[cc]{\small
$[d/k]$}} \put(83.00,41.00){\makebox(0,0)[cc]{${{\mathbb Z}_2}$}}
\put(54.00,116.00){\makebox(0,0)[cc]{${{\mathbb Z}_2}$}}
\put(22.00,122.00){\makebox(0,0)[cc]{\small $d-(d-2l)(k-1)$}}
\put(20.00,104.00){\makebox(0,0)[cc]{\small $d-(d-2l+1)(k-1)$}}
\put(8.00,38.00){\line(1,0){7.00}} \put(54.00,7.00){\makebox(0,0)[cc]{\small
$d-2l$}} \put(56.00,0.50){\framebox(17.00,4.50)[cc]{\small $d-2l+1$}}
\put(65.00,5.00){\line(-1,1){5.00}} \put(8.00,56.00){\line(1,0){32.00}}
\put(54.00,122.00){\makebox(0,0)[cc]{$0$}}
\put(83.00,47.00){\makebox(0,0)[cc]{$0$}} \put(14.00,23.00){\line(1,0){26.00}}
\put(50.00,23.00){\line(1,0){15.00}} \put(65.00,29.00){\line(-1,0){15.00}}
\put(40.00,29.00){\line(-1,0){26.00}} \put(50.00,38.00){\line(1,0){15.00}}
\put(65.00,44.00){\line(-1,0){15.00}} \put(50.00,50.00){\line(1,0){15.00}}
\put(65.00,56.00){\line(-1,0){15.00}} \put(73.00,56.00){\line(1,0){22.00}}
\put(95.00,62.00){\line(-1,0){22.00}} \put(63.00,71.00){\line(0,-1){48.00}}
\put(57.00,23.00){\line(0,1){48.00}} \put(51.00,71.00){\line(0,-1){48.00}}
\put(40.00,50.00){\line(-1,0){32.00}} \put(8.00,62.00){\line(1,0){32.00}}
\put(50.00,62.00){\line(1,0){15.00}} \put(73.00,68.00){\line(1,0){22.00}}
\put(65.00,68.00){\line(-1,0){15.00}} \put(40.00,68.00){\line(-1,0){32.00}}
\put(8.00,113.00){\line(1,0){32.00}} \put(50.00,113.00){\line(1,0){15.00}}
\put(65.00,119.00){\line(-1,0){15.00}} \put(40.00,119.00){\line(-1,0){32.00}}
\put(8.00,125.00){\line(1,0){32.00}} \put(50.00,125.00){\line(1,0){15.00}}
\put(65.00,95.00){\line(-1,0){15.00}} \put(40.00,95.00){\line(-1,0){32.00}}
\put(8.00,101.00){\line(1,0){32.00}} \put(50.00,101.00){\line(1,0){15.00}}
\put(65.00,107.00){\line(-1,0){15.00}} \put(40.00,107.00){\line(-1,0){32.00}}
\end{picture}
\caption{Term $E^1$ for odd $k$}
\end{figure}
{\sc Corollary.} {\it The term $E_{p,q}^1$ of our spectral sequence is as in
Fig. 1 for even $k$ and as in Fig. 2 is $k$ is odd, i.e.:
if $k$ is even or $d-p$ is odd, $1 \le p \le [d/k],$ then in the column
$E^1_{p,*}$ only the cells $E^1_{p,q}$ with $q=d-p(k-1)$ and $q=d-p(k-1)-1$ are
nontrivial; these two cells are isomorphic to $\Z$;
if the number $k(d-p-1)$ is odd, $1 \le p \le [d/k],$ then the column
$E^1_{p,*}$ contains unique nontrivial cell $E^1_{p,d-p(k-1)-1} \simeq \Z_2$;
for any $k$ and $d$ the unique nontrivial cell of the column $E^1_{[d/k]+1,*}$
is $E^1_{[d/k]+1,[d/k]-1} \simeq \Z;$
all other columns of the spectral sequence are trivial.} \quad $\Box$ \medskip
{\sc Remark.} The situation shown in columns $p=[d/k]$ and $p=[d/k]-1$ of Fig.
2 appears when $d-[d/k]$ is even. In fact the opposite also can happen, but in
the most interesting case, when $d$ is a multiple of $k,$ the situation is
exactly as in the picture. In this last case the unique nontrivial cell
$E^1_{[d/k]+1,[d/k]-1} \simeq \Z$ of the column $p=[d/k]+1$ lies in the same
horizontal row as the unique nontrivial cell $E^1_{[d/k],[d/k]-1} \simeq \Z_2$
of the column $p=[d/k].$
\medskip
{\sc Proposition 2.} {\it For any $d, k,$ except for the case when $k$ is odd
and $d$ is a multiple of $k$, our spectral sequence degenerates in the term
$E_1$, i.e. $E_1 = E_\infty$. In the exceptional case there is unique
nontrivial operator $d_1,$ acting from the cell $E^1_{d/k+1,d/k-1} \sim \Z$ to
the cell $E^1_{d/k,d/k-1} \sim \Z_2,$ killing the latter cell, after which the
spectral sequence also degenerates.} \medskip
Main theorem follows immediately from this proposition.
\medskip
{\it Proof of Proposition 2.} For almost all $d$ and $k$ the assertion follows
immediately from the explicit form of the term $E_1.$ The only two cases, when
nontrivial differentials could occur, are the case $k=2$ (a half of which is
already studied, see Main example in \S \ 2, and the remaining case of odd $d$
is even so easy) and the case when $d$ is a multiple of $k$. In the latter case
(for $k>2$) the unique nontrivial differential can act from the cell
$E^1_{d/k+1,d/k-1} \sim \Z$ to the cell $E^1_{d/k,d/k-1}.$ If $k$ is even, then
the latter cell is isomorphic to $\Z,$ and the corresponding term $F_{d/k} \sm
F_{d/k-1}$ of our filtration is the oriented bundle with fiber $\R^1,$ whose
base also is oriented and is, in its turn, the space of a fibre bundle with
base $B(S^1,d/k)$ and the fiber equal to an $(d/k-1)$-dimensional open simplex.
The boundary of the disc $F_{d/k+1} \sm F_{d/k}$ in this term of filtration
coincides with the zero section of the former (line) bundle, therefore it is
the boundary (modulo lower terms of the filtration) of the space of the bundle
with fiber $\R_+$.
In the case of odd $k$ similar line bundle is non-orientable, therefore when we
try to span this zero section by a similar chain over a maximal
simple-connected domain in the base, then we construct a homology between the
image of this section and a cycle generating the group $H_{2d/k-1}(F_{d/k} \sm
F_{d/k-1}) \sim \Z_2$. (An adequate picture here is the M\"obius band, whose
equator circle is the boundary of the disc $F_{d/k+1} \sm F_{d/k}$.) All the
further differentials are trivial by the dimensional reasons, and main theorem
is completely proved.
\medskip
{\sc Acknowledgment.} I thank A.~Durfee for a stimulating discussion and
G.~Kalai, who indicated me that Proposition 1 is due to Caratheodory.
|
\section{Introduction}
\unitlength 0.7mm \linethickness{0.6pt}
Sir Isaac Newton has proved (see \cite{Newton}, Lemma XXVIII of Book 1) that
the areas cut off by different lines from a convex bounded domain with
infinitely smooth boundary in ${\R^2}$ never define an algebraic function on
the space of lines\footnote{``There is no oval figure whose area, cut off by
right lines at pleasure, can be universally found by means of equations of any
number of finite terms and dimensions'', originally ``Nulla extat figura Ovalis
cujus area, rectis pro lubitu abscissam possit per aequationes numero
terminorum ac dimensionum finitas generaliter inveniri''}.
This fact contrasts to the Archimedes' theorem on sphere sections, implying
that the volumes cut off by planes from a ball in ${\R^3}$ depend algebraically
on these planes; it is easy to check that the last fact holds also for
arbitrary ellipsoids in odd-dimensional spaces.
In 1987, in connection with the 300th anniversary of the Newton's Book,
V.I.~Arnold has asked whether his result is true in the case of other
dimensions and general domains with smooth boundaries, see \cite{Arprob},
problems 1987-14, 1988-13 and 1990-27.
In 1988, solving this problem, I have extended the Newton's result to convex
domains in even-dimensional spaces and to arbitrary bounded domains with smooth
boundaries in ${\R}^2$, see \cite{Notices}, \cite{APLT}. In the present paper,
the same statement is proved for arbitrary bounded domains with smooth
boundaries in even-dimensional spaces. The proof is based on Picard-Lefschetz
theory and elementary facts on finite reflection groups.
\subsection{Definitions and main theorem}
Denote by $P_n$ the space of all affine hyperplanes in $\R^n$. It can be
considered as $\R P^n$ with one point removed, in particular as an algebraic
manifold. Given a compact domain $D \subset \R^n$, the corresponding two-valued
{\em volume function} $P_n \to \R$ associates with any hyperplane $L \in P_n$
the volumes of both parts cut off from $D$ by this hyperplane. The domain $D$
is called {\em algebraically integrable} if this function is algebraic, i.e.
there exists a non-trivial polynomial $\Phi$ in $n+2$ variables such that for
any real numbers $a_1, \dots, a_n, b$ and any value $V_i$, $i=1,2$, of the
volume function on the plane $L \in P_n$ defined by the equation $a_1 x_1 +
\dots + a_n x_n = b$, we have $\Phi(a_1, \dots , a_n, b, V_i)=0$. We are going
to prove the following fact.
\begin{theorem}
\label{mainth} If $n$ is even, then there is no algebraically integrable
bounded domain with $C^\infty$-smooth boundary in $\R^n$.
\end{theorem}
\begin{remark} \rm
1. For ovals (i.e. convex bounded domains) in $\R^2$ this theorem was proved
in \cite{Newton}, see also \cite{comment}, \cite{Notices}.
2. If $D$ is a bounded algebraically integrable domain with $C^\infty$-smooth
boundary in $\R^n$, then by projective duality and Tarski--Seidenberg lemma
this boundary $\partial D$ is semialgebraic, see \cite{Kvant}, \cite{Notices}.
Therefore it is enough to consider only the bodies with regular semialgebraic
boundaries in $\R^n$, i.e. to assume that $\partial D$ consists of several
smooth connected components of the zero locus of a real polynomial. Also, it is
enough to prove our theorem for any connected component of $D$ separately, so
we will assume that $D$ is connected. On the other hand, for any finite $m$
there exist algebraically integrable domains in $\R^{2k}$ with $C^m$-smooth
boundaries, see \cite{Appendix}. Therefore the condition of
$C^\infty$-smoothness in Theorem \ref{mainth} cannot be reduced.
3. By an Archimedes' theorem, the volume cut off by a plane from a ball in
$\R^3$ depends polynomially on the distance of the plane from the center, i.e.
on a two-valued algebraic function in $P_3$. We need also take into account all
planes not intersecting the ball, therefore we add two single-valued functions
$P_3 \to \R^1$ equal identically to 0 and to the volume of the ball, and obtain
a four-valued algebraic function proving the algebraic integrability of the
ball in $\R^3$. Moreover, the Archimedes' theorem can be easily extended to
balls and ellipsoids in any odd-dimensional spaces. It seems likely that there
are no other examples of irreducible integrable domains with smooth boundaries
in odd-dimensional spaces, but this conjecture is not yet proved.
\end{remark}
\section{Two main examples}
\subsection{Convex domains in $\R^{2k}$}
\label{convv} Let $n$ be an even number, and $D$ a convex domain in $\R^{n}$
bounded by a compact semialgebraic variety $\partial D$ without singular
points. Choose a linear function $l: \R^{n} \to \R^1,$ whose restriction to
$\partial D$ is a Morse function (such functions exist by the Sard's lemma
applied to the Gauss map $\partial D \to \R P^{n-1}$). Let $m<M$ be both
critical values of this restriction. Denote by $A$ the complexification of
$\partial D$ (i.e. the hypersurface in $\CC^n$ distinguished by the same
polynomial equation). Also we can and will consider $l$ as the restriction to
$\R^n$ of a linear function $(\CC^n,\R^n) \to (\CC^1,\R^1)$, which we'll denote
also by $l$.
For any $t \in (m,M)$ define $V(t)$ as the volume of the set $D \cap
l^{-1}((-\infty, t])$. If $D$ is algebraically integrable, then $V$ is an
algebraic function, in particular its analytic continuation to $\CC^1$ is
finite-valued.
Further, denote by $\CC P_n$ the space of all complex affine hyperplanes in
$\CC^n$. For any $X \in \CC P_n$, consider the group
\begin{equation}
\label{cont} H_n(\CC^n, A \cup X).
\end{equation}
\begin{lemma}
There is a well-defined linear function $H_n(\CC^n, A \cup X) \to \CC^1,$ whose
value on a relative homology class is equal to the integral of the volume form
\begin{equation}
\label{vol} dx_1 \wedge \dots \wedge dx_n
\end{equation}
along an arbitrary piecewise-smooth relative cycle representing this class.
\end{lemma}
\noindent {\it Proof.} This follows from the Stokes' theorem applied to the
holomorphic form (\ref{vol}), and from the fact that the integral of this form
along any singular $n$-chain in the $(n-1)$-dimensional complex variety $A \cup
X$ is equal to zero. \hfill $\Box$ \medskip
There is a Zariski open subset $\mbox{Reg} \subset \CC P_n$ such that all
planes $X \in \mbox{Reg}$ are transversal to (maybe singular) variety $A$:
indeed, for any stratum of any algebraic Whitney stratification of $A$ the set
of planes $X$ not transversal to this stratum is a semialgebraic subvariety of
positive codimension in $\CC P_n$. Then by Thom's isotopy lemma (see e.g.
\cite{GM}) there is a locally trivial fiber bundle over $\mbox{Reg}$, whose
fiber over the point $\{X\}$ is the pair $(\CC^n, A \cup X)$. Consider the
associated homological bundle over $\mbox{Reg}$, whose fiber over $\{X\}$ is
the group (\ref{cont}).
This fiber bundle is locally trivialized, i.e. it carries the flat {\em
Gauss-Manin connection} (see e.g. \cite{AGLV}) defined by continuous shifts of
cycles into the neighboring fibers in correspondence with any local
trivialization of the initial fiber bundle of pairs $(\CC^n, A \cup X)$. This
connection is well-defined because the homological classes of these moved
cycles do not depend on the exact choice of this local trivialization. In
particular, if we fix a point $\{X_0\} \in \mbox{Reg}$ and a class $\gamma \in
H_n(\CC^n, A \cup X_0)$, then a function $\beta_{\gamma}(X)$ arises in any
simply-connected neighborhood of $\{X_0\}$ in $\mbox{Reg}$: its value at the
point $X$ is equal to the integral of the form (\ref{vol}) along the cycle
obtained from $\gamma$ by the Gauss-Manin connection over any path connecting
$X_0$ and $X$ in our neighborhood. This function is analytic in this
neighborhood, and can be continued to an analytic function on entire manifold
$\mbox{Reg}$. The ramification of this function along paths in $\mbox{Reg}$
depends on the monodromy action of the group $\pi_1(\mbox{Reg}, \{X_0\})$ on
$H_n(\CC^n, A \cup X_0)$.
\begin{figure}
\begin{center}
\begin{picture}(110,105)
\bezier{70}(60,100)(30,100)(30,80) \bezier{70}(30,80)(30,63)(50,63)
\bezier{70}(50,63)(80,63)(80,84) \bezier{70}(80,84)(80,100)(60,100)
\put(20,40){\line(1,0){90}} \put(110,40){\line(-1,-2){20}}
\put(90,0){\line(-1,0){90}} \put(0,0){\line(1,2){20}}
\put(10,20){\line(1,0){90}} \put(30,20){\circle*{2}} \put(80,20){\circle*{2}}
\put(28,14){$m$} \put(78,13){$M$} \put(39,18.5){$\blacksquare$}
\put(39.5,21){\vector(-1,0){8}} \put(31.5,19){\vector(1,0){8}}
\put(30,21){\oval(3,3)[t]} \put(30,19){\oval(3,3)[b]} \put(39,15){$\tau $}
\put(28,25){$\alpha(m)$} \put(28.5,19){\line(0,1){2}}
\put(40,64){\line(0,1){32}} \put(40.5,19){\vector(1,0){38}}
\put(78.5,21){\vector(-1,0){36.7}} \put(80,21){\oval(3,3)[t]}
\put(80,19){\oval(3,3)[b]} \put(81.5,19){\line(0,1){2}}
\put(55,25){$\alpha(M)$} \put(40,61){\vector(0,-1){28}}
\put(102,17.5){${\mathbb R}^1$} \put(100,34){${\mathbb C}^1$}
\put(31,80){\line(1,1){8.5}} \put(31.5,84){\line(1,1){8}}
\put(32.2,88){\line(1,1){7.3}} \put(31.5,76){\line(1,1){8}}
\put(32,72){\line(1,1){7.5}} \put(34,68){\line(1,1){5.2}} \put(42,52){\Large
$l$}
\end{picture}
\caption{Picard-Lefschetz monodromy of integration cycles for a convex domain
in even-dimensional space} \label{onne}
\end{center}
\end{figure}
For instance, let $X_0 = X(\tau ) \in \mbox{Reg}$ be the hyperplane
$l^{-1}(\tau ) \subset \CC^n,$ $\tau \in (m,M) \subset \R^1$, and
$\gamma=\gamma(\tau ) \in H_n(\CC^n, A \cup X(\tau ))$ the class of the figure
$D \cap l^{-1}((-\infty,\tau ])$ oriented by the form (\ref{vol}). Denote by
${\mathcal Reg}$ the set of all $t\in \CC^1$ such that $X(t) \equiv l^{-1}(t)
\in \mbox{Reg}$. For all $t \in {\mathcal Reg}$ close to $\tau $, the function
$\beta_{\gamma}( X(t))$ coincides with the volume function $V(t)$, hence their
analytic continuations to entire ${\mathcal Reg}$ also coincide. So, in order
to investigate the ramification of the volume function, let us study the orbit
of this element $\gamma$ under the monodromy action of $\pi_1({\mathcal Reg})$.
It follows easily from the Picard-Lefschetz formula (see e.g. \cite{APLT}) that
the loop $\alpha(m) \in \pi_1({\mathcal Reg})$ (see Fig. \ref{onne}) moves the
element $\gamma$ to $-\gamma$, and the loop $\alpha(M)$ moves $\gamma$ to $2
[D] - \gamma$, where $[D]$ is the homology class of the entire domain $D$.
Hence the composition of these two operators moves $\gamma$ to $\gamma + 2[D]$.
The relative cycle $[D]$ does not depend on $t$ and is invariant under the
action of the group $\pi_1({\mathcal Reg})$, hence iterating the operator
$\alpha(M) \circ \alpha(m)$ we obtain consecutively $\gamma+ 4[D]$, $\gamma +
6[D]$, etc. But the volume of $D$ is positive, hence our analytic function
takes infinitely many different values at one and the same point $\tau $, and
cannot be algebraic.
\begin{remark} \rm
Picard-Lefschetz operators act differently in spaces $\R^n$ of different
parities, because the intersection form in the middle homology group of an
$(n-2)$-dimensional complex manifold is symmetric if $n$ is even and
antisymmetric if $n$ is odd. In particular, if $n$ is odd, then the loops
$\alpha(m)$ and $\alpha(M)$ act trivially on $\gamma(t)$, which makes the
Archimedes' example possible.
\end{remark}
\subsection{General compact domain with smooth algebraic boundary in $\R^2$}
Consider first the sample domain $D$ shown in Fig. \ref{two}. Let $l:\R^2 \to
\R^1$ be the projection downwards in the page, where the target line $\R^1$ is
oriented to the right. The restriction of this function to $\partial D$ is a
strictly Morse function. Let $\tau \in \R^1$ be a value greater than the
global minimum of $l$ on $\partial D$, but lower than all other its critical
values. For real $t \approx \tau $, let $ \gamma(t) \in H_2(\CC^2, A \cup
X(t))$ be the class of the positively oriented domain $D \cap l^{-1}((-\infty,
t])$, see Fig. \ref{two}a. Let us increase $t$ until the right-hand boundary
segment of this domain meets a critical point of $l|_{\partial D}$, shortly
before this meeting move $t$ into the complex domain in $\CC^1$ and go around
the corresponding critical value. The Gauss-Manin connection over entire this
path moves $\gamma(t)$ into the homology class in $H_2(\CC^2, A \cup X(t))$ of
the sum of two domains as shown in Fig. \ref{two}b; notice that these two
domains should be taken with opposite orientations, so that their common
boundary in $X(t)$ is a single oriented segment. Further, we decrease $t$ until
one of endpoints of this segment becomes again a critical point of
$l|_{\partial D}$, turn $t$ around the corresponding critical value (see Fig.
\ref{two}c), etc. It is important that when $t$ passes for the second time the
critical value at the local (but not global) maximum of $l$ on $\partial D$
(between Figs. \ref{two}c and \ref{two}d), the corresponding leaf of the
integral of $dx_1 \wedge dx_2$ along $\gamma(t)$ has a regular point: indeed,
the derivative of this integral over $t$ is equal to the length of the boundary
segment in $X(t)$. Therefore we do not need to make a whole circle around this
critical value: instead, we miss this critical value by a half-circle in either
half-plane of $\CC^1$ and continue to increase $t$ along the real line, see
Fig. \ref{two}d.
\unitlength 0.35mm
\begin{figure}
\begin{picture}(400,60)
\mbox{
\begin{picture}(80,60)
\bezier{70}(20,55)(-10,45)(20,35) \bezier{50}(20,35)(45,28)(25,23)
\bezier{50}(25,23)(10,17)(25,11) \bezier{50}(25,11)(35,7)(45,11)
\bezier{100}(45,11)(90,39)(50,55) \bezier{40}(50,55)(35,60)(20,55)
\put(14.5,37){\line(0,1){16}} \put(35,0){a} \put(7.5,42){\footnotesize $+$}
\end{picture}
} \mbox{
\begin{picture}(80,60)
\bezier{70}(20,55)(-10,45)(20,35) \bezier{50}(20,35)(45,28)(25,23)
\bezier{50}(25,23)(10,17)(25,11) \bezier{50}(25,11)(35,7)(45,11)
\bezier{100}(45,11)(90,39)(50,55) \bezier{40}(50,55)(35,60)(20,55)
\put(26,23){\line(0,1){33.3}} \put(35,0){b} \put(14,41){\footnotesize $+$}
\put(25.9,26){\scriptsize $-$}
\end{picture}
} \mbox{
\begin{picture}(80,60)
\bezier{70}(20,55)(-10,45)(20,35) \bezier{50}(20,35)(45,28)(25,23)
\bezier{50}(25,23)(10,17)(25,11) \bezier{50}(25,11)(35,7)(45,11)
\bezier{100}(45,11)(90,39)(50,55) \bezier{40}(50,55)(35,60)(20,55)
\put(24,12.5){\line(0,1){43}} \put(35,0){c} \put(14,41){\footnotesize $+$}
\put(24.8,26){\footnotesize $-$} \put(17.5,15){\scriptsize $+$}
\end{picture}
} \mbox{
\begin{picture}(80,60)
\bezier{70}(20,55)(-10,45)(20,35) \bezier{50}(20,35)(45,28)(25,23)
\bezier{50}(25,23)(10,17)(25,11) \bezier{50}(25,11)(35,7)(45,11)
\bezier{100}(45,11)(90,39)(50,55) \bezier{40}(50,55)(35,60)(20,55)
\put(50,15){\line(0,1){39}} \put(35,0){d} \put(32,37){\footnotesize $+$}
\end{picture}
}
\end{picture} \caption{Transformations of the integration cycle in $\R^2$}
\label{two}
\end{figure}
Finally our boundary segment in $X(t)$ shrinks to a point, namely to the global
maximum point of $l$ on $\partial D$. At the instant of this degeneration the
cycle obtained from $\gamma(t)$ by the Gauss-Manin connection over our path
would coincide with entire domain $D$. However we do not let $t$ to meet the
maximum value, but shortly before this rotate $t$ again in the complex domain,
and then go back along the same path. When $t$ returns to the initial position
$\tau $, the Gauss--Manin connection over the obtained loop in ${\mathcal Reg}$
moves the cycle $\gamma(\tau )$ to $2[D]-\gamma(\tau )$. The rest of the proof
repeats that in the previous example.
In the case of a general connected domain $D \subset \R^2$ with smooth
boundary, we proceed in the similar way. Namely, we take the exterior component
$\overline{\partial} D$ of $\partial D$ (which is homeomorphic to $S^1$) and a
linear function $l:(\CC^n,\R^n) \to (\CC, \R)$, whose restriction to
$\overline{\partial} D$ is Morse. Denote by $m<M$ the global extremal values of
$l$ on $\overline{\partial} D$. The corresponding global extremal points of $l$
break $\overline{\partial} D$ into two segments. Consider the set of pairs
$(a,b) \in \overline{\partial} D \times \overline{\partial} D$ such that $a$
belongs to one such part and $b$ to the other, and $l(a) = l(b)$. It is easy to
see that this set is a smooth one-dimensional submanifold with boundary, and
this boundary consists of two points $(a,a)$ where $a$ is either the global
minimum or the global maximum point of $l$ on $\partial D$. Although this
submanifold can be not connected, its boundary points should belong to one and
the same its connected component diffeomorphic to a segment. So, we start from
its point $(a,b)$ where $\{a,b\} = \partial D \cap l^{-1} (\tau)$, $\tau
\approx m$, which is close to one endpoint of this segment, and go along this
segment towards the other endpoint, watching the corresponding common values of
$l$ and making appropriate rotations in ${\mathcal Reg}$ each time when $a$ or
$b$ approaches a critical point of $l|_{\overline{\partial} D}$. This gives us
a path in ${\mathcal Reg}$. Gauss-Manin connection over this path moves the
cycle $\gamma(\tau )$ into the entire domain bounded by $\overline{\partial}
D$; in particular the value of the analytic continuation of the volume function
$V(t)$ along this path close to its endpoint is almost equal to the area of
this domain. Then we turn in $\CC^1$ around the maximal value $M$ of $l$ on
$\partial D$ and come back to $\tau $ along the same path. The monodromy along
this path in ${\mathcal Reg}$ moves the homology class $\gamma(\tau )$ into
$2[D] - \gamma(\tau )$. The rest of the construction is the same as above.
\begin{remark} \rm
To be rigorous, there can be other points of $\CC^1 \setminus {\mathcal Reg}$
on the segment $[m,M]$ apart from the critical values of the restriction of $l$
to $\overline{\partial} D$. Therefore, constructing our path in ${\mathcal
Reg}$, we need to avoid these points along small arcs in the complex domain.
This does not affect our consideration, because our integration cycle (and its
area function) behave regularly when $t$ passes these additional points.
\end{remark}
\section{Proof of the main theorem}
\subsection{On finite reflection groups}
We use the following well-known facts about finite reflection groups, see e.g.
\cite{bour}, \cite{AGLV}.
\begin{prop}
\label{refle} Let $\Z^N$ be an integer lattice with integer-valued symmetric
bilinear form $\langle \cdot , \cdot \rangle$; let $\{e_j\} \subset \Z^N$ be a
finite collection of elements of length $\sqrt{2}$ $($i.e. $\langle e_j, e_j
\rangle = 2$ for each $j)$ generating the entire $\Z^N$ as a $\Z$-module. Let
$G$ be the subgroup in $\mbox{SL}(N, \Z)$ generated by reflections
corresponding to these elements $e_j$, acting on $\Z^N$ by the formula
\begin{equation}
\label{reflect} R_j : a \mapsto a-\langle e_j, a \rangle e_j.
\end{equation}
Suppose that the orbits in $\Z^N$ of all generating elements $e_j$ under the
action of the group $G$ are finite. Then
1. The group $G$ is finite.
2. The form $\langle \cdot, \cdot \rangle$ is non-degenerate\footnote{and even
positive definite, which is less important for us now} : if $\langle e_j , a
\rangle = 0$ for any $j$, then $a=0$.
\end{prop}
\subsection{Reflection group related with a smooth semialgebraic domain in $\R^{2k}$}
Suppose that $n$ is even. Let $D$ be a bounded connected domain with
semialgebraic non-singular boundary in $\R^n$. This boundary can consist of
several connected components, but exactly one of them separates the entire $D$
and all other components of $\partial D$ from the infinity. Let us denote this
component by $\overline{\partial} D$. Let $A$ be the complexification of
$\overline{\partial} D$: this is a hypersurface in $\CC^n$, which can have
singularities apart from a neighborhood of $D$. Let us fix some complex
semialgebraic Whitney stratification of $A$ (see e.g. \cite{GM}), all whose
strata of dimension $<n-1$ do not meet $\partial D$.
\begin{definition} \rm
\label{gener} A linear function $l:(\CC^n,\R^n) \to (\CC^1,\R^1)$ with real
coefficients is {\em generic} with respect to the hypersurface $A$ if
1. Its restriction to $\partial D$ is a strictly Morse function;
2. For any critical value $t \in \R^1$ of this restriction, the corresponding
complex hyperplane $X(t) = l^{-1}(t) \subset \CC^n$ is transversal to all
strata of $A$ apart from the corresponding critical point of this restriction.
\end{definition}
Generic linear functions are dense in the space of all linear functions; let
$l$ be one of them. Recall the notation ${\mathcal Reg}$ for the set of all
values $t \in \R^1$ such that the plane $X(t)$ is transversal to $A$.
Let $a_j \in \overline{\partial} D$ be a critical point of
$l|_{\overline{\partial} D}$, and $t_j$ the corresponding critical value
$l(a_j)$. For any sufficiently small $\varepsilon>0$ and any $\nu>0$
sufficiently small with respect to $\varepsilon$, denote by $B_\varepsilon
(a_j)$ the ball of radius $\varepsilon$ in $\CC^n$ centered at $a_j$, and
suppose that $t \in \CC^1$ is an arbitrary point in the punctured
$\nu$-neighborhood of $t_j$. Then by Milnor's theorem (see e.g. \cite{AGLV})
the (reduced modulo a point) homology group $\tilde H_{n-2}(B_\varepsilon (a_j)
\cap A \cap X(t))$ is isomorphic to $\Z$. The chain of isomorphisms
\begin{equation}
\label{redu1}
\begin{array}{l} H_n(B_\varepsilon (a_j), B_\varepsilon(a_j) \cap (A
\cup X(t)))
\to H_{n-1}(B_\varepsilon (a_j) \cap (A \cup X(t))) \to \\
\to H_{n-1}(B_\varepsilon (a_j) \cap X(t), B_\varepsilon (a_j) \cap A \cap
X(t)) \to \tilde H_{n-2}(B_\varepsilon (a_j) \cap A \cap X(t))
\end{array}
\end{equation} (in which the first and the third arrows are boundary
homomorphisms, and the second one is a fragment of the exact sequence of the
pair $(B_\varepsilon(a_j) \cap (A \cup X(t)), B_\varepsilon(a_j) \cap A)$)
proves the same for its left-hand group
\begin{equation}
\label{local} H_n(B_\varepsilon (a_j), B_\varepsilon (a_j) \cap (A \cup X(t))).
\end{equation}
The monodromy action on these groups, defined by the rotation of $t$ around the
value $t_j$ in its $\nu$-neighborhood, commutes with all these isomorphisms and
takes one generator of this group (\ref{local}) into the other (i.e. it is
multiplication by $-1$).
Let us choose an arbitrary such generator $\delta_j$ for $t = t_j+ \nu/2$,
extend this choice by the Gauss-Manin connection to similar groups
(\ref{local}) for all $t \in (t_j, t_j+\nu)$, and define the function
$\varphi_j: (t_j, t_j+\nu) \to \CC^1$, whose value $\varphi_j(t)$ is equal to
the integral of the form (\ref{vol}) along this generator of (\ref{local}). It
is easy to calculate that this function is real or purely imaginary depending
on the parity of the Morse index of $l|_{\partial D}$ at $a_j$, and its
absolute value vanishes asymptotically as $(t-t_j)^{(n+1)/2}$ when $t$ tends to
$t_j$. Also, it is analytic and can be extended to a neighborhood of the point
$t_j + \nu/2$ in $\CC^1$. Let again $[m,M]$ be the segment of values
$l(\overline{\partial} D) \subset \R^1.$ Since the set $\CC^1 \setminus
{\mathcal Reg}$ is finite, there is a neighborhood $U([m,M])$ of this segment,
such that all points of this set in $U([m,M])$ are real; they include the set
$\Sigma$ of critical values of $l|_{\overline{\partial} D}$. We can and will
assume that $0 \in [m,M] \cap {\mathcal Reg}$. Let us connect the point $0$
with all points $t_j+ \nu/2$ by arbitrary paths in the domain $U([m,M]) \cap
{\mathcal Reg} \cap \{t: Im(t) > 0\}$: \unitlength 1.5mm \linethickness{0.6pt}
\begin{picture}(40,3)
\put(0,0){\line(1,0){40}} \put(19.3,-0.8){\Large $0$} \put(37,1){$\R$}
\put(2,0){\circle*{1.2}} \put(12,0){\circle*{1.2}} \put(33,0){\circle*{1.2}}
\put(11.75,0){\oval(16.5,5)[t]} \put(16.75,0){\oval(6.5,3)[t]}
\put(27.25,0){\oval(14.5,5)[t]} \put(2,0){\oval(3,3)[t]}
\put(12,0){\oval(3,3)[t]} \put(33,0){\oval(3,3)[t]}
\end{picture} \label{pic3} . Define the function germs
$\Phi_j: (\CC^1,0) \to \CC$ at the point $0$ as analytic continuations of
functions $\varphi_j$ along these paths. Their values at $0$ are equal to the
integrals of (\ref{vol}) along the relative cycles $\tilde \Delta_j$ in $\CC^n
\mbox{ mod } (A \cup X(0))$ obtained from the chosen generators of
(\ref{local}) by the Gauss-Manin connection over these paths. If we realise
this Gauss-Manin connection by the local trivialization of the fiber bundle of
pairs $(\CC^n, A \cup X(t))$ following from the Thom's isotopy lemma, then the
closures of these cycles $\tilde \Delta_j$ do not meet the singular locus of
$A$. Denote by ${\stackrel{\circ}{A}}$ the non-singular part of $A$, then the
cycles obtained in this way define certain elements $ \Delta_j \in H_n(\CC^n
\setminus \mbox{sing}(A), {\stackrel{\circ}{A}} \cup X(0))$.
Let $\mathfrak{F}$ be the subgroup of the space of all complex-valued function
germs at $0$, consisting of integer linear combinations of our function germs
$\Phi_j$. This subgroup is the image of a homomorphism of an integer lattice
(consisting of all integer linear combinations of symbols $\Phi_j$) into the
space of germs, hence also an integer lattice. For any $i, j$ define the scalar
product $\langle \Phi_i, \Phi_j \rangle$ as follows: we take two
$(n-2)$-dimensional cycles in ${\stackrel{\circ}{A}} \cap X(0)$ obtained from
$\Delta_i$ and $\Delta_j$ by the composite homomorphism
\begin{equation}
\label{reduc}
\begin{array}{r}
H_n(\CC^n \setminus \mbox{sing}(A), {\stackrel{\circ}{A}} \cup X(0)) \to
H_{n-1}({\stackrel{\circ}{A}} \cup X(0)) \to \\ \to H_{n-1}(X(0),
{\stackrel{\circ}{A}} \cap X(0)) \to H_{n-2}({\stackrel{\circ}{A}} \cap X(0))
\end{array}
\end{equation}
similar to (\ref{redu1}), calculate the intersection index of these
$(n-2)$-dimensional cycles in the complex variety ${\stackrel{\circ}{A}} \cap
X(0),$ and multiply this intersection index by $(-1)^{1+n/2}.$ Since $n$ is
even, $\langle \Phi_i, \Phi_j \rangle = \langle \Phi_j, \Phi_i \rangle$ for any
$i, j$. The scalar squares $\langle \Phi_j, \Phi_j \rangle$ of all generating
elements $\Phi_j$ are equal to $2$ (see e.g. \cite{AGLV}, \S II.1.3).
\begin{lemma}
\label{lem1} The scalar products $\langle \Phi_i, \Phi_j \rangle$ can be
extended by linearity $($in a unique way$)$ to a symmetric bilinear form on the
lattice $\mathfrak{F}$.
\end{lemma}
\noindent {\it Proof.} Suppose that the linear combination $c_1 \Phi_1 + \dots
+ c_r \Phi_r$ defines the identically zero function germ. Consider the loop in
$U([m,M]) \cap {\mathcal Reg}$, consisting of the distinguished path from $0$
to $t_j +\nu/2$, rotation around $t_j$ along the circle of radius $\nu/2$, and
return to $0$ by our distinguished path. By the Picard-Lefschetz formula, the
{\it analytic continuation of our zero germ} along this path turns it into the
germ $\Phi_j$ taken with the coefficient $\pm \langle (c_1 \Phi_1 + \dots + c_r
\Phi_r), \Phi_j \rangle.$ The germ $\Phi_j$ is not equal to identical zero,
hence this coefficient should vanish for any $j$. \hfill $\Box$ \medskip
Each basic element $\Phi_j$ defines a reflection in the space $\mathfrak{F}$
acting by
\begin{equation}
\label{PL2} \Phi \mapsto \Phi - \langle \Phi_j, \Phi \rangle \Phi_j;
\end{equation}
by Picard--Lefschetz formula it describes the analytic continuation of an
arbitrary germ $\Phi \in \mathfrak{F}$ along the loop constructed in the proof
of lemma \ref{lem1}. These reflections preserve the scalar product $\langle
\cdot , \cdot \rangle$ and thus generate a subgroup of the group of
automorphisms of $(\mathfrak{F}; \langle \cdot , \cdot \rangle$).
\begin{prop}
\label{pro3} If the domain $D$ is algebraically integrable, then the subgroup
generated by operators $($\ref{PL2}$)$ in the group of automorphisms of the
lattice $\mathfrak{F}$ is finite.
\end{prop}
\noindent {\it Proof}. The set of all complex hyperplanes $X(t) = l^{-1}(t),$
$t \in \CC^1$, forms a line in the space $\CC P_n$. We can assume that $\nu$ is
small enough, so that $(m,m+\nu) \subset {\mathcal Reg}$. Let $\tau =m+\nu/2$,
and denote by $\gamma(X(\tau ))$ the class in $H_n(\CC^n, A \cup X(\tau ))$ of
the positively oriented domain $D \cap l^{-1}((-\infty, \tau ])$. Consider also
the analytic function $\mbox{Vol}$ on $\mbox{Reg} \subset \CC P_n$ defined by
integrals of the form (\ref{vol}) along the similar cycles $\gamma(X) \in
H_n(\CC^n, A \cup X)$ obtained from $\gamma(X(\tau ))$ by Gauss-Manin
connection over different paths in $\mbox{Reg}$ connecting $X(\tau )$ and $X$.
\begin{lemma}
\label{lem2} The set of restrictions of different leaves of the function
$\mbox{\rm Vol}$ to neighborhoods of the point $0$ in the line $\CC^1 \subset
\CC P_n$ of planes $X(t)$ contains all function germs $\Phi_j: (\CC^1, 0) \to
\CC^1$ corresponding to all critical points of the restriction of $l$ to
$\overline{\partial} D$.
\end{lemma}
\noindent {\it Proof of lemma \ref{lem2}}. It is enough to prove that for any
$j$ the integral functions $\varphi_j:(t_j, t_j+\nu) \to \CC^1$ defined above
appears among restrictions of the function $\mbox{Vol}$ to the interval
$(t_j,t_j+\nu) \subset \CC^1$. Consider the set of all non-singular points of
the hypersurface $A$, at which the second fundamental form of this hypersurface
is non-degenerate. Its complement has complex codimension 1 in $A$, hence this
set is path-connected within any irreducible component of $A$. Since
$l|_{\overline{\partial} D}$ is a Morse function, this set contains all its
critical points (and all of them obviously belong to one and the same component
of $A$). Choose a smooth path in this set, connecting the critical points with
values $m$ and $t_j$. The tangent hyperplanes at these points define a path in
$\CC P_n$. Take a path in the set $\mbox{Reg} \subset \CC P_n$ escorting this
one in its thin neighborhood and connecting the planes $X(\tau )$ and
$X(t_j+\nu/2)$ in such a way that for any point $\{X\}$ of this path the
element of the group (\ref{cont}) obtained from $\gamma$ by the Gauss-Manin
connection over this path is realised by a cycle inside a small ball centered
at the neighboring tangency point of $A$ and a hyperplane parallel to $X$, in
particular this cycle for the endpoint of this path generates the group
(\ref{local}). This terminal cycle can coincide with the integration cycle
defining the function $\varphi_j$ or be opposite to it. In the first case
$\varphi_j$ is defined by the analytic continuation of the function
$\mbox{Vol}$ along this path, in the second one we need additionally go once
around the critical value $t_j$ in the set ${\mathcal Reg}$ of generic
hyperplanes $X(t)$. \hfill $\Box$
\begin{remark} \rm
In this lemma, we do not state that any germ $\Phi_j$ can be obtained from the
initial function at the point $\tau $ by the analytic continuation along a path
inside our line $\CC^1$. Moreover, the reflection group on $\mathfrak{F}$
generated by operators (\ref{PL2}) (defined by paths inside this domain only)
can be reducible, see subsection \ref{redred} below.
\end{remark}
\begin{remark} \rm
The path in $\CC P_n$ used in this proof can be chosen in an arbitrarily small
neighborhood of the set of $($complexifications of$)$ real hyperplanes. Indeed,
let us connect our critical points of $l|_{\overline{\partial} D}$ by a generic
path inside our component of $\overline{\partial} D$; then to avoid the set of
parabolic points of $A$ we can move this path slightly into the complex domain
in arbitrarily small neighbourhoods of points at which it crosses this set.
\end{remark}
Further, the set of analytic continuations of all function germs $\Phi_j$
contains all germs obtained from them by reflections (\ref{PL2}). Therefore if
the domain $D$ is algebraically integrable then the orbit of any element
$\Phi_j$ under the reflection group generated by operators (\ref{PL2}) is
finite. By proposition \ref{refle} the entire this reflection group is then
finite, in particular the symmetric form $\langle \cdot, \cdot \rangle$ on
$\mathfrak{F}$ is non-degenerate. Proposition \ref{pro3} is proved. \hfill
$\Box$ \medskip
\begin{prop}
\label{prp4} The class $[\overline{D}] \in H_n(\CC^n, A \cup X(0))$ of the
entire domain in $\R^n$ bounded by $\overline{\partial} D$ is equal to the sum
of classes $\Delta_j$ $($taken with appropriate signs$)$ over all critical
points of $l|_{\overline{\partial} D}$. In particular, the corresponding sum of
function germs $\pm \Phi_j$ is the constant function equal identically to the
volume bounded by $\overline{\partial} D$.
\end{prop}
This fact follows immediately from the following one (Lemma \ref{next} below).
For any critical value $t_j$ of $l|_{\overline{\partial} D}$ we have two
elements in the group $H_n(\CC^n, A \cup X(t_j+\nu/2))$: one (let us call it
$\nabla_+$) given by the positively oriented part of the half-space $\R^n\cap
l^{-1}((-\infty, t_j+ \nu/2])$ bounded by $\overline{\partial} D$, and the
other, $J(\nabla_-),$ obtained from the similar class $\nabla_- \in H_n(\CC^n,
A \cup X(t_j-\nu/2))$ by the Gauss--Manin connection over the arc of radius
$\nu/2$ in the upper half-plane of $\CC^1$.
Consider also absolute homology classes $\Pi_{\pm} \in H_{n-2}(A \cap X(t_j\pm
\nu/2))$ represented by naturally oriented real manifolds $\overline{\partial}
D \cap X(t_j\pm \nu/2)$ respectively. These classes are obtained from
$\nabla_\pm$ by maps similar to the composition (\ref{reduc}).
\begin{lemma}[see \cite{APLT}, Lemma 3.3 on page 121]
\label{next} 1. The difference of two relative homology classes $\nabla_+$ and
$J(\nabla_-)$ in $H_n(\CC^n, A \cup X(t_j+\nu/2))$ is equal to the image of the
vanishing cycle $\delta_j \in H_n(B_\varepsilon(a_j), A \cup X(t_j+\nu/2))$
under the identical embedding $B_\varepsilon(a_j) \hookrightarrow \CC^n$.
2. The difference in the group $H_{n-2}(A \cap X(t_j+\nu/2))$ of the class
$\Pi_+$ and the element obtained from $\Pi_-$ by the Gauss--Manin connection
over the arc of radius $\nu/2$ in the upper half-plane of $\CC^1$ can be
realized by a vanishing cycle generating the group $H_{n-2}(B_\varepsilon(a_j)
\cap A \cap X(t_j+ \nu/2)) \simeq \Z.$ \hfill $\Box$ \ $\Box$
\end{lemma}
\begin{remark} \rm
The earliest (very technical) proof of item 2 known to me is given in
\cite{Leray}; for another proof, deducing it directly from the Picard-Lefschetz
formula, see \cite{Petr} and \S V.3 of \cite{APLT}.
\end{remark}
Since the volume bounded by $\overline{\partial} D$ is positive, the sum
$\sum_j \pm \Phi_j$ mentioned in proposition \ref{prp4} is a non-zero element
of the lattice $\mathfrak{F}$. On the other hand, the image of the
corresponding cycle $[\overline{D}]$ under the map (\ref{reduc}) is equal to
zero, because the image of this cycle under the first arrow in (\ref{reduc}) is
a cycle in $A$, and the second arrow in (\ref{reduc}) is the reduction modulo
$A$. Therefore this non-zero element of $\mathfrak{F}$ belongs to the kernel of
the scalar product $\langle \cdot, \cdot \rangle$. Thus by proposition
\ref{refle} the group $G$ cannot be finite. Theorem \ref{mainth} is completely
proved. \hfill $\Box$
\subsection{Reducibility of the reflection group}
\label{redred}
The reflection group
\begin{equation}
\label{genref} (\mathfrak{F}; \langle \cdot , \cdot \rangle; \{R_j\})
\end{equation}
can be reducible, for instance this is the case if $D$ is a thin tubular
neighborhood of the standard circle embedded into $\R^4$. In this case we can
choose the linear function $l: \R^4 \to \R$ in such a way that its restriction
to $\partial D$ has four critical points $a_j$, $j = 1, 2, 3, 4$ with Morse
indices $0, 1, 2$ and $3$ respectively, and the critical values $t_2$ and $t_3$
at the points $a_2$ and $a_3$ almost coincide. Then we have the following
calculation.
\begin{prop}
For appropriate choice of orientations of vanishing cycles $\beta_i \in
H_2(B_\varepsilon (a_j) \cap A \cap X(t_j+\nu/2)) \simeq \Z,$ the intersection
matrix in $H_{2}(A \cap X(0))$ of cycles obtained from them by the Gauss-Manin
connection over any paths in the upper half of $U([m,M]) \cap {\mathcal Reg}$
$($see picture in page \pageref{pic3}$)$ is equal to
$$\begin{vmatrix}
\ 2 & -2 & \ 0 & \ 0 \\ -2 & \ 2 & \ 0 & \ 0 \\ \ 0 & \ 0 & \ 2 & -2 \\ \ 0 & \
0 & -2 & \ 2
\end{vmatrix}\ .
$$
In particular, in this case the entire action of the reflection group
$($\ref{genref}$)$ splits into the direct sum of two-dimensional ones, each of
which is isomorphic to one arising in the example of subsection \ref{convv}.
\end{prop}
\noindent {\it Sketch of the proof}. $\langle \beta_2, \beta_3 \rangle = 0$
because the corresponding critical points are distant, but their critical
values can be made arbitrarily close to one another by bending the function $l$
(which does not affect the intersection index). Further, we can assume that $0$
is slightly above the critical values $t_2$ and $t_3$. In this case the cycle
in the variety $A \cap X(0)$ obtained by the Gauss--Manin connection from
$\beta_4$ is presented by the set $\partial D \cap X(0)$ of all its real
points. Its intersection with the vanishing cycles $\beta_3$ and $\beta_2$ can
be calculated in local terms of Morse critical points $a_2$ and $a_3$ (see
\cite{GZ}, \cite{AC}, \cite{APLT}). Replacing $0$ by a value slightly below
$t_2$ and $t_3$ (which does not affect the intersection indices of cycles
obtained by Gauss--Manin connection in the upper half-plane) we obtain similar
indices for cycles obtained from $\beta_1, \beta_2$ and $\beta_3$.
Finally, we know from lemma \ref{next} that the cycle of real points in $A \cap
X(t_4 - \nu/2)$ can be realised as the sum of three cycles obtained by
Gauss-Manin connection in the upper half-plane from $\beta_1$, $\beta_2$ and
$\beta_3$. Hence, by the previous calculations, the cycle obtained from
$\beta_4$ has with this sum the same intersection index, as with $\beta_3$
only. This gives us zero for the remaining corner elements $(1,4)=(4,1)$ of the
intersection matrix. \hfill $\Box$ \medskip
In general, if the reflection group (\ref{genref}) is reducible, then the set
of all cycles $\Delta_j$ splits into collections of cycles generating germs
$\Phi_j$ that belong to different irreducible components. It is easy to see
that the sum of all function germs corresponding to all cycles $\Delta_j$ from
any such collection belongs to the kernel of the form $\langle \cdot, \cdot
\rangle$. Indeed, this sum for one such collection is obviously orthogonal to
all generators $\Phi_j$ from the other collections. On the other hand, this sum
is equal to the difference of the sum of all $\Phi_j$ over all collections and
the sum of such sums over all collections except for this one; both sums are
orthogonal to all $\Phi_j$ from our collection.
|
\section{Introduction}
Two famous theorems of Newton assert that
a) a homogeneous spherical layer in Euclidean space does not attract bodies
inside the sphere, and
b) exterior bodies are attracted by it to the center of the sphere as by the
point-wise particle whose mass is equal to the mass of the entire sphere.
Ivory [I] extended both these theorems to the attraction of ellipsoids, and
Arnold [A 82] extended the first of them to the attraction of arbitrary
hyperbolic hypersurfaces: such a surface does not attract the particles inside
the hyperbolicity domain; see also [G 84].
In any component of the complement of the attracting surface this attraction
force coincides with a real analytic vector-function; we investigate the
ramification of this function, in particular (following one another famous
theory of Newton, see [A 87], [AV]) the question if it is algebraic or not.
We describe the monodromy group responsible for the ramification and identify
it as a subgroup of the local monodromy group of a complex complete
intersection of codimension 2 in $\C^n$. Unlike the usual local monodromy
action, this monodromy representation is reducible: e.g. the
Newton--Ivory--Arnold theorem depends on the fact that the homology class of
the set of real points of a hyperbolic surface defines an invariant element of
this action (although this element is not equal to zero: indeed, otherwise even
the potential function of the force would be zero, and not only its gradient
field, which is wrong already in the Newton's case).
Although we consider mainly the orbit of a very special cycle, formed by all
real points of a hyperbolic polynomial, all our calculations can be applied to
more general situations, e.g. when the integration cycle is an arbitrary linear
combination of real components (maybe non-compact) of an algebraic hypersurface
in $\R^n$.
In the case of odd $n$, this group acts in a vanishing homology group with
twisted coefficients (so that the corresponding kernel form $r^{2-n}ds$ of the
potential function can be integrated correctly along its elements). In \S \ 2.3
we extend the standard facts concerning vanishing homology of complete
intersections to this group, cf. [Ph 65], [G 88].
There is a (non-formal) partition of all classes of isolated singularities of
complete intersections into series with varying dimension $n$ of the ambient
space $\C^n$ (but with the constant codimension $p$ of the complete
intersection), see [E], [AGLV]; e.g., all singularities given by $p$ generic
quadrics in the spaces $\C^n$ with different $n$ and fixed $p$ form such a
series. To any such series there corresponds a series of reflection groups,
also depending on the parameter $n$; for such $n$ that $n-p$ is even, these
groups coincide with the (standard) local monodromy groups of corresponding
singularities. The homology groups described in \S \ 2.3 fill in the gap: for
$n-p$ odd, the reflection group of the natural series coincides with the
monodromy action on such a twisted homology group of the corresponding
singularity. (In the marginal case $p=1$, all the reflection groups of the
series coincide, see [GZ], [G 88].)
This is a reason why the qualitative behavior of attraction forces in the
spaces of any dimension is essentially the same, unlike the usual situation
(see e.g. [P], [ABG], [A 87], [AV], [V 94]) when the functions given by
similar integral representations behave in very different way in the spaces of
dimensions of different parity.
For $n=2$ and arbitrary $d$, our monodromy group is finite, thus the analytic
continuation of the attraction force is finitely-valued, in particular (by the
Riemann's existence principle) algebraic, see \S \ 5.1 below. A realistic
estimate of the number of values of this continuation is given by Theorem 4. In
particular, we get a new series of examples when the attraction force coincides
with a {\em single-valued} (rational) vector-function outside the hyperbolicity
domain, see the Corollary to Theorem 4.
For $d=2$ and arbitrary $n>2,$ the monodromy group is infinite, and the orbit
of any integration cycle lies on an ellipsoidal cylinder in the vanishing
homology space. Fortunately, the integral of the attracting charge takes zero
value on the directing plane of this cylinder, thus the number of its values
along the elements of any orbit again is finite, see \S \ 5.2.
In all the other cases (when $d \ge 3$ and $n \ge 3$) it seems likely that the
monodromy group defined by the {\it generic} algebraic surface of degree $d$ in
$\R^n$ is large enough to ensure that the Newton's integral (and any other
non-zero linear form on the space of vanishing cycles) takes an infinite number
of values on the orbit of any non-invariant vector (and the unique invariant
vector is presented by the integration cycle corresponding to the hyperbolicity
domain of an hyperbolic charge). I can prove this conjecture only if the
additional restriction $d+n \ge 8$ is
satisfied\footnote{For a complete proof, removing this restriction, see the
Appendix to this article, written by W.~Ebeling}.
\medskip
Everywhere below all the homology groups $H_*(\cdot)$ are reduced modulo a
point.
\section{Vanishing homology and local monodromy of complete intersections}
Here we recall the basic facts about the local Picard--Lefschetz theory of
isolated singularities of complete intersections (see e.g. [H], [E], [AGLV])
and extend them to the case of twisted vanishing homology groups.
\subsection{Classical theory}
Let $f:(\C^n,0) \to (\C^p,0)$ be a holomorphic map, $f=(f_1, \ldots, f_p),$ and
suppose that the variety $f^{-1}(0)$ is an {\it isolated complete intersection
singularity} (ICIS) at $0$ (i.e. it is a smooth $(n-p)$-dimensional variety in
a punctured neighborhood of $0$). Suppose that the coordinates in $\C^p$ are
chosen generically, then the map $\tilde f \equiv (f_1, \ldots, f_{p-1}): \C^n
\to \C^{p-1}$ also defines an ICIS at $0$. Let $B$ be a sufficiently small
{\em closed} disc centered at the origin in $\C^n$, and $c=(c_1, \ldots, c_p)$
a generic point very close to the origin in $\C^p$. The corresponding
manifolds $X_f \equiv f^{-1}(c) \cap B$ and $\tilde X_{f} \equiv \tilde
f^{-1}(c_1, \ldots, c_{p-1}) \cap B$ are called the {\it Milnor fibres} of
$f$and $\tilde f$. Their homology groups are connected by the exact sequence
\begin{equation}
\cdots \to H_{n-p+1}(\tilde X_f) \to H_{n-p+1}(\tilde X_f,X_f)
\stackrel{\partial}{\longrightarrow} H_{n-p}(X_f) \to \cdots. \label{hamm}
\end{equation}
{\bf Proposition 1} (see [M], [H]). {\it The sequence (\ref{hamm}) is trivial
outside the fragment presented here. All groups in (\ref{hamm}) are free
Abelian. Moreover, the spaces $X_f$ and $\tilde X_f$ are homotopy equivalent to
the wedges of spheres of dimensions $n-p$ and $n-p+1$ respectively.}
\medskip
The rank of $H_{n-p}(X_f)$ is called the {\it Milnor number} of the complete
intersection $f$ and is denoted by $\mu(f)$. The Milnor numbers of all {\it
quasihomogeneous} complete intersections are calculated in [GH] (in [MO] for
$p=1$); we need the following special case of this calculation. \medskip
{\bf Proposition 2.} {\it 1. The Milnor number of a homogeneous function $f:
\C^n \to \C^1$ of degree $d$ with isolated singularity at 0 is equal to
$(d-1)^n$.
2. The Milnor number of a complete intersection $f = (f_1, f_2)$ with isolated
singularity at $0$, where the functions $f_1$ and $f_2$ are homogeneous of
degrees $a$ and $b$ respectively, is equal to $((a-1)^nb - (b-1)^na)/(a-b)$ if
$a \ne b$, and to $(a-1)^n(an-a+1)$ if $a = b$.} \medskip
The rank $\mu(f)+ \mu(\tilde f)$ of the middle group $H_{n-p+1}(\tilde
X_f,X_f)$ of (\ref{hamm}) is equal to the number of (Morse) critical points of
the restriction of $f_p$ on $\tilde X_f$. The generators of this group are
represented by the {\it Lefschetz thimbles} defined by the (non-intersecting)
paths in $\C^1$ connecting the non-critical value $c_p$ of this restriction
with all critical values, namely, any of these thimbles is an embedded disc
swept out by the one-parametric family of {\it vanishing spheres} lying in the
varieties $f^{-1}(c_1, \ldots, c_{p-1}, \tau),$ where $\tau$ runs over the
corresponding path in $\C^1:$ when $\tau$ tends to the endpoint (i.e. to a
critical value of this restriction) the cycles of this family contract to the
corresponding critical point. These vanishing spheres in the variety $X_f$
(which corresponds to the common starting point $c_p$ of these paths) generate
the group $H_{n-p}(X_f)$, while the elements of $H_{n-p+1}(\tilde X_f)$ define
relations among them.
\subsection{Picard--Lefschetz formula for standard homology}
Let $s \subset \C^1$ be the set of all these critical values, then the group
$\pi_1(\C^1 \setminus s)$ acts naturally on all groups of (\ref{hamm}). This
action commutes with all arrows in (\ref{hamm}) and is trivial on the left-hand
group $H_{n-p+1}(\tilde X_f)$. The action on the middle and right-hand groups
$H_{n-p+1}(\tilde X_f,X_f)$, $H_{n-p}(X_f)$ is determined by the
Picard--Lefschetz formula: a class $\delta \in H_{n-p+1}(\tilde X_f,X_f)$,
being transported along a {\em simple loop} (see [Ph 67], [V 94]) $\omega_i$,
corresponding to the path connecting $c_p$ with the $i$-th critical value,
becomes
$$\delta + (-1)^{(n-p+1)(n-p+2)/2}\langle
\partial \delta, \partial \delta_i \rangle \delta_i, $$
where $\delta_i$ is the class of the thimble defined by this path, $\partial$
is the boundary operator in (\ref{hamm}), and $\langle \cdot , \cdot \rangle$
is the intersection form in $H_{n-p}(X_f)$. In particular, a similar formula
describes the monodromy action of the same loop on $H_{n-p}(X_f)$: it sends an
element $\Delta$ of this group to
\begin{equation}
\Delta + (-1)^{(n-p+1)(n-p+2)/2}\langle \Delta, \Delta_i \rangle \Delta_i,
\label{plf1}
\end{equation}
where $\Delta_i \equiv \partial \delta_i$ is the sphere vanishing along this
path. \medskip
{\bf Proposition 3} (see e.g. [AGV]). {\it The intersection form $\langle \cdot
, \cdot \rangle$ is symmetric if $n-p$ is even and skew-symmetric if $n-p$ is
odd. The self-intersection index of any vanishing sphere is equal to $2$ if
$n-p \equiv 0 (mod \ 4)$ and to $-2$ if $n-p \equiv 2 (mod \ 4)$.} \medskip
In particular, if $n-p$ is even, then any transportation along a simple loop
$\omega_i$ acts on the group $H_{n-p}(X_f)$ (respectively, $H_{n-p+1}(\tilde
X_f,X_f)$) as the reflection in the hyperplane orthogonal to the vector
$\Delta_i$ (respectively, $\delta_i$) with respect to the intersection form in
the homology of $X_f$ (respectively, the form induced by the boundary operator
from this intersection form). The latter action is a central extension of the
former one.
\medskip
More generally, let $F$ be a $k$-parametric deformation of $f,$ i.e. a map
$F:\C^n \times \C^k \to \C^p$ such that $F(\cdot, 0) \equiv f$. For any
$\lambda \in \C^k$ lying in a sufficiently small neighborhood $D^k$ of the
origin, denote by $f_\lambda$ the map $F(\cdot, \lambda)$ and by $\tilde
f_\lambda$ the map $\C^n \to \C^{p-1}$ given by first $p-1$ coordinate
functions of $f_\lambda$. Set $X_{f,\lambda}= f^{-1}_{\lambda} \cap B$ and
$\tilde X_{f,\lambda}= \tilde f^{-1}_{\lambda} \cap B$. If $F$ is ``not very
degenerate'' then for almost all values of $\lambda$ these varieties are smooth
(with boundaries) and have the same topological type; e.g. the varieties $X_f,
\tilde X_f$ participating in (\ref{hamm}) appear in the $p$-parametric
deformation consisting of maps $f_\lambda \equiv (f_1-\lambda_1, \ldots,
f_p-\lambda_p)$ and correspond to the particular value $(\lambda_1, \ldots,
\lambda_p) = (c_1, \ldots, c_p).$ \medskip
{\bf Definition 1.} The {\it discriminant variety} $\Sigma(F)$ of $F$ is the
set of such $\lambda \in D^k$ that the topological type of the pair of
varieties $(\tilde X_{f,\lambda}, X_{f,\lambda})$ does not coincide with that
for all neighboring $\lambda$, i.e., either the origin in $\C^{p-1}$ is a
critical value of $\tilde f_\lambda$ or the origin in $\C^1$ is a critical
value of $f_p|_{\tilde X_{f,\lambda}}$. An exact sequence similar to
(\ref{hamm}) appears for any $\lambda \in D^k \setminus \Sigma(F)$, as well as
the monodromy action of the group $\pi_1(D^k \setminus \Sigma(F))$ on this
sequence.
Now suppose that the deformation $F$ keeps $\tilde f$ undeformed, i.e., $\tilde
f_\lambda \equiv \tilde f$ for any $\lambda$; in particular the action of this
group on the left-hand group in (\ref{hamm}) is trivial. A standard speculation
with the Zariski's theorem (see e.g. [AGV], [V]) allows us to reduce this
action to the above-considered action of the group $\pi_1(\C^1 \setminus s)$,
and thus to the Picard--Lefschetz operators.
\medskip
There is a natural map, {\it Leray tube operation}
\begin{equation}
t: H_{n-p}(X_f) \to H_{n-p+1}(\tilde X_f \setminus X_f) \label{leray}
\end{equation}
described e.g. in [Ph 67], [AGLV], [V 94]: for any cycle $\gamma$ in
$X_\lambda$ the cycle $t(\gamma)$ is swept out by the small circles in $\tilde
X_f \setminus X_f$ which are the boundaries of the fibres of the natural
fibration of the tubular neighborhood of $X_f$.
\subsection{Twisted vanishing homology of complete intersections}
Let $L_{-1}$ (respectively, $\pm \ZZ$) be the local system on $\tilde X_f
\setminus X_f$ with the fibre $\C^1$ (respectively, $ \ZZ ^1$) such that any
loop having an odd linking number with $X_f$ acts on this fibre as
multiplication by $-1$. In particular, $L_{-1} \equiv \pm \ZZ \otimes \C$.
\medskip
Consider the obvious homomorphism
\begin{equation}
j: H_{n-p+1}(\tilde X_f \setminus X_f, L_{-1}) \to H^{lf}_{n-p+1}(\tilde X_f
\setminus X_f, L_{-1}), \label{lf}
\end{equation}
where $H^{lf}_*(\cdot)$ denotes the homology of locally finite chains.
The Lefschetz thimbles define elements also in the right-hand group of
(\ref{lf}) and in the similar group $H^{lf}_{n-p+1}(\tilde X \setminus X_f, \pm
\ZZ ):$ indeed, they are embedded discs in $\tilde X_f\setminus X_f$ with
boundary in $X_f$, and thus their interior parts can be lifted to an arbitrary
leaf of the local system $L_{-1}$ or $\pm \ZZ $. For any such thimble $\delta_i
\in H^{lf}_{n-p+1}(\tilde X \setminus X_f, \pm \ZZ ) $ there is an element
$\kappa_i \in H_{n-p+1}(\tilde X \setminus X_f, \pm \ZZ ), $ the {\it vanishing
cycle} defined by the same path in $\C^1$, such that $j(\kappa_i) = 2
\delta_i$, see [Ph 65] and Fig. 1, where such a cycle in one-dimensional
$\tilde X$ is shown.
\bigskip
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(95.00,12.00)
\put(38.00,8.00){\vector(1,1){4.00}} \put(42.00,12.00){\vector(1,0){18.00}}
\put(60.00,8.00){\vector(-1,0){18.00}} \put(42.00,8.00){\vector(-1,1){4.00}}
\put(38.00,12.00){\vector(-1,0){18.00}} \put(20.00,8.00){\vector(1,0){18.00}}
\put(20.00,10.00){\oval(6.00,4.00)[l]} \put(60.00,10.00){\oval(6.00,4.00)[r]}
\put(20.00,5.00){\makebox(0,0)[cc]{$X_f$}}
\put(60.00,5.00){\makebox(0,0)[cc]{$X_f$}} \put(60.00,10.00){\circle*{1.50}}
\put(20.00,10.00){\circle*{1.50}} \put(95.00,10.00){\makebox(0,0)[cc]{Fig. 1}}
\end{picture}
{\bf Theorem 1.} {\it a) The homomorphism (\ref{lf}) is an isomorphism,
as well as the similar homomorphism of homology groups reduced mod $\partial
\tilde X_f,$
$$ H_{n-p+1}(\tilde X_f \setminus X_f,
\partial \tilde X_f \setminus \partial X_f; L_{-1})
\to H^{lf}_{n-p+1}(\tilde X_f \setminus X_f,
\partial \tilde X_f \setminus \partial X_f; L_{-1});$$
b) the dimensions of both groups (\ref{lf}) are equal to $\nu(f) \equiv \mu(f)+
\mu(\tilde f)$, and similar homology groups in all other dimensions are
trivial;
c) the right-hand group in (\ref{lf}) is freely generated by the Lefschetz
thimbles specified by an arbitrary {\em distinguished} (see e.g. [AGV], [AGLV])
system of paths connecting the noncritical value $c_p$ of $f_p|_{\tilde X}$
with all critical values.}
\medskip
{\bf Corollary.} {\it The left-hand group in (\ref{lf}) is generated by the
{\em vanishing cycles} defined by the same paths.} \medskip
{\it Proof of the theorem.} The fact that the map (\ref{lf}) (and also its
relative version) is isomorphic is a general algebraic fact, which is true for
all local systems $L_\alpha$ with monodromy indices $\alpha \ne 1$: this
follows from the comparison of the {\it Leray spectral sequences} (see e.g.
[GrH], \S \ III.5) calculating the indicated homology groups and applied to the
identical embedding $\tilde X_f \setminus X_f \to \tilde X_f$.
The assertion of statement b) concerning the right-hand group in (\ref{lf})
follows from the similar assertion concerning the non-twisted vanishing
homology group $H_{n-p+1}(\tilde X_f, X_f; \ZZ ) $ $\equiv
H^{lf}_{n-p+1}(\tilde X_f \setminus X_f, \ZZ)$ (see Proposition 1), the fact
that $\pm \ZZ \otimes \ZZ_2 = \ZZ \otimes \ZZ_2 = \ZZ_2$ (the constant local
system with fibre $\ZZ_2$) and from the formula of universal coefficients. The
same reasons prove that the $\ZZ_2$-torsion of the group $H^{lf}_{*}(\tilde X_f
\setminus X_f, \pm \ZZ)$ is trivial in all dimensions.
Statement c) follows now from the fact that the images of thimbles are linearly
independent already in the group $H^{lf}_{n-p+1}(\tilde X_f \setminus X_f, \pm
\ZZ) \otimes \ZZ_2$.
\medskip
Let $\Im$ be the subgroup in $H_{n-p+1}(\tilde X_f \setminus X_f, \pm \ZZ)$
generated by vanishing cycles $\kappa_i$ defined by all possible paths
(probably it coincides with entire $H_{n-p+1}(\tilde X_f \setminus X_f, \pm
\ZZ)$).
\medskip
{\bf Lemma 1.} {\it For any elements $\alpha, \beta \in \Im, $ their
intersection index is even.}
\medskip
Indeed, this index is equal to the (well-defined) intersection index of
$\alpha$ and $j(\beta),$ and $j(\beta) \in 2 H^{lf}_{n-p+1}(\tilde X_f
\setminus X_f, \pm \ZZ)$.
\medskip
Define the bilinear form $\langle \cdot , \cdot \rangle$ on $\Im$ equal to half
this intersection index.
\medskip
{\bf Proposition 4.} {\it The form $\langle \cdot , \cdot \rangle$ is symmetric
if $n-p$ is odd and is skew-symmetric if $n-p$ is even. For any basis vanishing
cycle $\kappa_i,$ $\langle \kappa_i, \kappa_i \rangle$ is equal to 2 if
$n-p\equiv 3(mod\ 4)$ and to $-2$ if $n-p\equiv 1(mod\ 4)$.}
\medskip
In the terms of this form, the monodromy action on the group $\Im$ is defined
by the same Picard--Lefschetz formula as before: the monodromy along the simple
loop $\omega_i$ takes a cycle $\kappa$ to
\begin{equation}
\kappa + (-1)^{(n-p+1)(n-p+2)/2}\langle \kappa, \kappa_i \rangle \kappa_i.
\label{pllf}
\end{equation}
\section{Surface potentials and Newton--Ivory--Arnold
theorem}
\subsection{Potential function of a surface}
Denote by $d V$ the volume differential form in ${\R}^n$, i.e. the form $dx_1
\wedge \ldots \wedge dx_n$ in the Euclidean positively oriented coordinates
$x_1, \ldots, x_n$.
Denote by $r$ the Euclidean norm in ${\R}^n$, \ $r = (x_1^2 + \cdots +
x_n^2)^{1/2}$, and by $C_n$ the area of the unit sphere in ${\R}^n$.
\medskip
{\bf Definition 2.} The {\it elementary Newton--Coulomb potential function},
or, which is the same, the {\it standard fundamental solution of the Laplace
operator} in ${\R}^n$, is the function equal to ${{1} \over {2\pi}}\hbox{ln} \,
r$ if $n=2$, and to $-r^{2-n}/((n-2)C_n)$ if $n \ge 3$. Denote this function by
$G$. \medskip
This function can be interpreted as the potential of the force of attraction by
a particle of unit mass placed at the origin, i.e., the attraction force of
this particle is equal to $\, -\hbox{grad} \ G$.
The attraction force with which a body $K$ with density distribution $P$
attracts a particle of unit mass placed at the point $x \in {\R}^n$ is equal to
minus the gradient of the corresponding {\it potential function}, whose value
at the point $x$ is equal to the integral over $K$ of the differential form
$G(x - z)P(z)d V(z)$ (if such an integral exists). \medskip
Let $F$ be a smooth function in Euclidean space ${\R}^n$, and $M_F$ the
hypersurface $\{F = 0\}$. Suppose that $\; \hbox{grad} \ F \ne 0$ at the points
of $M_F$, so that $M_F$ is smooth.
\medskip
{\bf Definition 3.} The {\it standard charge} $\omega_F$ on the surface $M_F$
is the differential form $d V/dF$, i.e. the $(n-1)$-form such that for any
tangent frame $(l_2, \ldots, l_n)$ of $M_F$ and a transversal vector $l_1$ the
product of the values $\omega_F(l_2, \ldots, l_n)$ and $(dF, l_1)$ is equal to
the value $dV(l_1, \ldots, l_n)$. The {\it natural orientation} of the surface
$M_F$ is the orientation defined by this differential form.
\medskip
In particular, the value at a point $x \not \in M_F$ of the limit of potential
functions of homogeneous (with density $1/\epsilon$) distributions of charges
between the surfaces $F=0$ and $F=\epsilon$ is equal to the integral of the
standard charge form
\begin{equation}
G(x - z)\omega_F(z) \label{surfpot}
\end{equation}
along the naturally oriented surface $M_F$.
In a similar way, any function $P$ on the surface $M_F$ defines the charge
$P\cdot \omega_F$, which is called the {\it standard charge with density} $P$;
the potential at the point $x$ of this charge is equal to the integral of the
form
\begin{equation}
G(x-z)P(z)\omega_F(z) \label{polpot}
\end{equation}
along the naturally oriented surface $M_F$. The attraction force of this charge
is equal to minus the gradient of this potential function.
In these terms, theorems of Newton and Ivory look as follows.
\medskip
{\bf Theorem.} {\it The potential of the standard charge of the sphere
(respectively, an ellipsoid) in ${\R}^n$ given by the canonical equation (i.e.
by a polynomial $F$ of degree 2) is equal to a constant inside the sphere (the
ellipsoid), while outside it coincides (up to multiplicative constant) with the
potential function defined by any smaller ellipsoid confocal to ours.} \medskip
Arnold extended the ``interior'' part of this theorem to all {\it hyperbolic}
layers. \medskip
{\bf Definition 4.} An algebraic hypersurface $M$ of degree $d$ in $\R P^n$ is
{\it strictly hyperbolic} with respect to a point $x \in \R P^n \setminus M$ if
any real line through $x$ intersects $M$ at exactly $d$ different real points.
A polynomial $F: \R^n \to \R$ is strictly hyperbolic with respect to the point
$x \in \R^n$ if the projective closure $\bar M_F$ of the corresponding surface
$M_F$ is.
\medskip
{\bf Proposition 5} (see e.g. [ABG]). {\it If a hypersurface $M \subset \R P^n$
is strictly hyperbolic with respect to a point $x$, then it is also strictly
hyperbolic with respect to any point in the same component of the complement of
$M$. Any strictly hyperbolic hypersurface is smooth.} \medskip
{\bf Definition 5.} The {\it hyperbolicity domain} of a surface $M$ is the
union of points $x$ such that $M$ is hyperbolic with respect to $x$. \medskip
{\bf Proposition 6} (see [N]). {\it The set of all hypersurfaces $M$ of given
degree $d$ in ${\R}P^n$, which are strictly hyperbolic with respect to a given
point x, is contractible (or, equivalently, the set of all polynomials of
degree $d$ defining them consists of two contractible components).} \medskip
In particular, all the strictly hyperbolic surfaces $M$ of a given degree $d$
in $\R P^n$ are situated topologically in the same way: if $d$ is even, then
$M$ is ambient (and even rigid) isotopic to the union of $[d/2]$ concentric
spheres lying in an affine chart in ${\R}P^n$; if $d$ is odd, then $M$ is
isotopic to the union of $[d/2]$ concentric spheres plus the improper
projective hyperplane. The hyperbolicity domain consists of the interior points
of the ``most interior'' spheroid. This spheroid is always convex in ${\R}P^n$,
in particular, the hyperbolicity domain in ${\R}^n$ may consist of at most two
connected components. \medskip
The hyperbolic surface $M_F$ separates the space $\R^n$ into {\it zones}: the
$k$-th zone consists of all points $x \in \R^n \setminus M_F$ such that the
minimal number of intersection points of $M_F$ with segments connecting $x$ and
points of the hyperbolicity domains is equal to $k$. In particular, the maximal
index $k$ of a zone is equal to $[d/2]+1$ if $d$ is odd and the hyperbolicity
domain in $\R^n$ consists of one component, and is equal to $[d/2]$ otherwise.
\medskip
Given a strictly hyperbolic polynomial $F$, let us fix some path-component of
its hyperbolicity domain in ${\R}^n$, and number the components of $M_F$
starting from the boundary of this component (which becomes number 1), its
neighboring component gets number 2, etc. \medskip
{\bf Definition 6.} The {\it Arnold cycle} of $F$ is the manifold $\bar M_F$,
oriented in such a way that in the restriction to its finite part $M_F$ all odd
components are taken with the natural orientation (see Definition 3), while all
even components are taken with the reversed orientations.
The {\it hyperbolic potential} (respectively, {\it hyperbolic potential with
density $P$}) of the surface $M_F$ at a point $x \in \R^n \setminus M_F$ is the
integral of the form (\ref{surfpot}) (respectively, (\ref{polpot})) along the
Arnold cycle. As usual, the {\it attraction forces} defined by these potentials
are equal to minus the gradients of the potential functions. \medskip
{\bf Lemma 2.} {\it This definition of the Arnold cycle is correct, i.e. the
orientations of different non-compact components of $M_F$ thus defined are the
restrictions of the same orientation of the corresponding components of $\bar
M_F$.} \medskip
The proof is immediate. \medskip
{\bf Theorem} (see [A 82]). {\it The hyperbolic potential of the surface $M_F$
(and moreover any hyperbolic potential with density $P$, where $P$ is a
polynomial of degree $\le d-2$) is constant inside the hyperbolicity domain.}
\medskip
(In other words, the points of the hyperbolicity domain are not attracted by
the standard charge on $M_F$ taken with sign 1 or $-1$ depending on the parity
of the number of the component on which this charge is distributed.) \medskip
The proof follows Newton's original proof: for any infinitesimally narrow cone
centred at the point $x$, whose direction is not asymptotic for the surface
$M_F$, the forces of attraction to the pieces of $M_F$ cut by the cone
annihilate one another. Indeed, let us restrict the polynomial $F$ to the line
$L$ in ${\R}^n$ through $x$ contained in this cone; then this attraction force
is equal to the solid angle of our cone multiplied by the sum of the numbers
$P(A_i)/ F'(A_i)$ over all zeros $A_i$ of the polynomial $F|_L$. The last sum
is zero because it is the sum of the residues of a rational function over all
its complex poles. \medskip
The restriction $\deg \ P \le d-2$ from the Arnold's theorem ensures that the
integration form (\ref{polpot}) is ``regular at infinity'', i.e. extends to a
holomorphic form on the projective hypersurface $\bar M_F$. Givental [G 84]
remarked that a similar statement is true for polynomial potentials of
arbitrary degree if the integration cycle $M_F$ is compact in $\R^n$: in this
case the potential function in the hyperbolicity domain coincides with a
polynomial of degree $\le \deg \ P - d + 2$.
\medskip
In other domains the potential also coincides with real analytic functions; in
the next sections we study the global behavior of these functions, in
particular their algebraicity. The ramification of these functions is defined
by the action of certain monodromy group on a certain homology group; in the
next \S \ 4 we define these objects, and in \S \ 5 we calculate this monodromy
group.
\section{Monodromy group responsible for the ramification of potentials}
\subsection{Homology groups}
For any point $x \in {\C}^n, \ x = (x_1, \ldots, x_n),$ denote by $S(x)$ the
cone in ${\C}^n$ given by the equation
\begin{equation}
(z_1 - x_1)^2 + \cdots + (z_n - x_n)^2 = 0. \label{cone}
\end{equation}
Denote by $@ \equiv @(x)$ a local system over ${\C}^n \setminus S(x)$ with
fibre ${\ZZ}$ such that the corresponding representation $\pi_1({\C}^n
\setminus S(x)) \to \ \hbox{\rm Aut}({\ZZ})$ maps the loops whose linking
numbers with $S(x)$ are odd to the multiplication by $-1$.
We specify this local system in such a way that integrals of the form $r(\cdot\
- x)dz_1 \wedge \ldots \wedge dz_n$ along the $(n-1)$-dimensional cycles with
coefficients in it are well defined. Namely, we consider the two-fold covering
over ${\C}^n \setminus S(x)$, on which this form is single-valued, and the
direct image in ${\C}^n \setminus S(x)$ of this bundle under the obvious
projection of this covering. The trivial ${\ZZ}$-bundle over ${\C}^n \setminus
S(x)$ is naturally included in this direct image as a subbundle; the desired
local system is the quotient bundle of these two local systems. Obviously,
integrals of the form $r(\cdot\ - x)dz_1 \wedge \ldots \wedge dz_n$ (and of its
products by all single-valued functions) along the piecewise smooth $n$-chains
with coefficients in this local system are well-defined, and if these chains
are cycles, these integrals depend only on their homology classes.
\medskip
Let $F:{\C}^n \to {\C}$ be a polynomial, $W_F \subset \C^n$ the set of its
zeros, and $\bar W_F$ the projective closure of $W_F$.
For any $x \in \C^n$ we denote by ${\mathcal H}(x)$ the group
\begin{equation}
H_{n-1}(W_F \setminus S(x), {\ZZ}) \label{abs}
\end{equation}
in the case of even $n$, and the group
\begin{equation}
H_{n-1}(W_F \setminus S(x), \ @(x)) \label{abstw}
\end{equation}
if $n$ is odd. \medskip
Similarly, denote by ${\mathcal PH}(x)$ the group
\begin{equation}
H_{n-1}(\bar W_F \setminus \bar S(x), {\ZZ}) \label{proj}
\end{equation}
in the case of even $n$, and the group
\begin{equation}
H_{n-1}(\bar W_F \setminus \bar S(x), \ @(x)) \label{projtw}
\end{equation}
in the case of odd $n$. \medskip
{\bf Definition 7.} If the polynomial $F$ is real (i.e., $F(\R^n) \subset \R$)
and strictly hyperbolic, then the Arnold cycle defines correctly an element of
the group ${\mathcal PH}(x)$ (and even of the group ${\mathcal H}(x)$ if $M_F$
is compact); these elements are called the {\it Arnold homology classes} and
are denoted by $PA(x)$ and $A(x)$ respectively.
\medskip
In the case of odd $n$, integrals of the form (\ref{polpot}) along
$(n-1)$-chains in $W_F \setminus S(x)$ with coefficients in $@(x)$ are well
defined, and the values of these integrals along the cycles depend only on
their homology classes in the group (\ref{abstw}). Moreover, if $\; \hbox{deg}
\, P \le d-2$, and hence the form (\ref{polpot}) is regular at infinity, then
it can be integrated along the chains in $\bar W_F \setminus \bar S(x)$, and
the integrals along the cycles depend only on their classes in the group
(\ref{projtw}).
In the case of even $n > 2$ the form (\ref{polpot}) is single-valued, and no
problems with the definition of similar integrals along the elements of the
group (\ref{abs}) (or even (\ref{proj}) if $\deg \, P \le d-2$) arise, and in
the exceptional case $n = 2$, when (\ref{polpot}) is logarithmic, we remember
that we are interested not in the potential, but in its first partial
derivatives with respect to the parameter $x$ (i.e. in the components of the
attraction force vector). Therefore we integrate not the form (\ref{polpot})
but its partial derivatives
$${x_i - z_i \over (x_1-z_1)^2+(x_2-z_2)^2} P(z) \omega_F, \ i
=1, 2\, ;$$ these forms are already single-valued and there is no problem in
integrating them along the elements of the group (\ref{abs}) (or (\ref{proj})
if $\ \deg \, P \le d-2$).
\subsection{Homological bundles}
For almost all $x \in {\C}^n$ the groups ${\mathcal H}(x)$ (respectively,
${\mathcal PH}(x)$) are naturally isomorphic to one another. The set of
exceptional $x$ (for which the pair $(\bar W_F \setminus \bar S(x), W_F
\setminus S(x))$ is not homeomorphic to these for all neighboring $x'$) belongs
to a proper algebraic subvariety in ${\C}^n$ consisting of three components:
a) $W_F$ itself,
b) the set of such $x$ that $S(x)$ and $W_F$ are tangent outside $x$ in
${\C}^n$, and
c) the set of such $x$ that the projective closure of $S(x)$ in ${\C}P^n$ is
``more nontransversal'' to the closure of $W_F$ at their infinitely distant
points.
For a generic $F$ the last component is empty, and the second is irreducible
provided additionally that $n \ge 3$. \smallskip
Denote this algebraic set of all exceptional $x \in {\C}^n$ by $\Sigma(F)$.
\medskip
Consider two fibre bundles over ${\C}^n \setminus \Sigma(F)$ whose fibres over
a point $x$ are the spaces $W_F \setminus S(x)$, $\bar W_F \setminus \bar
S(x)$, and associate with them the homological bundles whose fibres over the
same point are the groups ${\mathcal H}(x)$ and ${\mathcal PH}(x)$. As usual,
the Gauss--Manin connection in these bundles defines the monodromy
representations
\begin{equation}
\pi_1({\C}^n \setminus \Sigma(F)) \to \hbox{Aut} \ {\mathcal H}(x), \label{mon}
\end{equation}
\begin{equation}
\pi_1({\C}^n \setminus \Sigma(F)) \to \hbox{Aut} \ {\mathcal PH}(x).
\label{monproj}
\end{equation}
These representations obviously commute with the natural map ${\mathcal H}(x)
\to {\mathcal PH}(x).$
Let $u$ be the potential function of the polynomial charge $P\cdot \omega_F$,
i.e. the function defined for any $x$ by the integral of the form
(\ref{polpot}) along the Arnold cycle. The ramification of (the analytic
continuation of) the function $u$ depends on the monodromy action (\ref{mon})
(respectively, (\ref{monproj})) on the Arnold element in ${\mathcal H}(x)$
(respectively, in ${\mathcal PH}(x)$).
\medskip
Namely, for any multiindex $\nu \in {\ZZ}_+^n$ ($\nu \ne 0$ if $n = 2$)
consider the linear forms
\begin{equation}
N^{(\nu)} : {\mathcal H}(x) \to {\C}, \quad PN^{(\nu)} : {\mathcal PH}(x) \to
{\C}, \label{functs}
\end{equation}
whose values on the cycle $\gamma$ are equal to the integral along $\gamma$ of
the $\nu$-th partial derivative of the form (\ref{polpot}) with respect to the
parameter $x$. \medskip
{\bf Proposition 7.} {\it For any $\nu$ ($\ne 0$ if $n=2$) and $x \in {\R}^n
\setminus \Sigma(F)$, the $\nu$-th partial derivative of the potential function
of the standard charge of the compact hyperbolic surface $M_F$ with density $P$
is finite-valued at $x$ if and only if the linear form $N^{(\nu)}$ takes
finitely many values on the orbit of the cycle $A(x)$ under the action of the
monodromy group (\ref{mon}). If $P$ is a polynomial of degree $\le d-2-|\nu|$,
then the same is true for non-compact hyperbolic surfaces if we replace $A(x)$
by $PA(x)$, $N^{(\nu)}$ by $PN^{(\nu)}$, and the action (\ref{mon}) by
(\ref{monproj}).} \medskip
This is a tautology.
\subsection{The invariant cycle}
In this subsection we show that for any $F$ and $x \in \C^n \setminus
\Sigma(F)$ the representation (\ref{monproj}) has an invariant vector; if $F$
is a real hyperbolic polynomial and $x$ lies in its hyperbolicity domain, then
this cycle coincides with the Arnold homology class. \medskip
Denote by $PS \subset \C P^{n-1}$ the common ``infinite'' part of all cones
$\bar S(x) \subset \C P^n$ and by $@$ the local system over $\C P^{n-1}
\setminus PS$ such that any system $@(x)$ is induced from it by the obvious
projection with center $x$.
\medskip
{\bf Proposition 8.} {\it The groups
\begin{equation}
H_{n-1}({\C}P^{n-1} \setminus PS) \label{pr1}
\end{equation}
(if $n$ is even) and
\begin{equation}
H_{n-1}({\C}P^{n-1} \setminus PS, @) \label{pr2}
\end{equation}
(if $n$ is odd) are one-dimensional. The generators of all these groups are
presented by the class of the submanifold $\R P^{n-1} \subset \C P^{n-1}
\setminus PS.$}
\medskip
The proof is elementary.
\medskip
The obvious map $\Pi : \bar W_F \setminus \bar S(x) \to \C P^{n-1} \setminus
PS$ (projection from the center $x$) is a $d$-fold ramified covering of complex
(and thus oriented) manifolds. The variety $\Pi^{-1}(\R P^{n-1})$ admits thus
an orientation ($@(x)$-orientation if $n$ is odd) induced from the chosen
orientation of $\R P^{n-1}$; denote by $\Omega(x)$ the class of this variety in
the group ${\mathcal PH}(x)$.
\medskip
{\bf Proposition 9.} {\it 1. The classes $\Omega(x)$ for different $x$
constitute a section of the homology bundle over $\C^n \setminus \Sigma(F)$
with fibres ${\mathcal PH}(x)$, which is invariant under the Gauss--Manin
connection, in particular these classes are invariant under the representation
(\ref{monproj}).
2. If $F$ is a real hyperbolic polynomial and $x$ lies in its hyperbolicity
domain, then $\Omega(x)$ coincides with the Arnold homology class $PA(x)$.}
\medskip
This follows immediately from the construction.
\subsection{Reduced Arnold class}
For an arbitrary element $\gamma$ of the group ${\mathcal PH}(x)$, the
corresponding potential function $u_\gamma(x)$ can be defined as the integral
of the form (\ref{polpot}) along the cycle $\gamma$ (if this integral exists),
in particular the usual potential $u(x)$ coincides with $u_{PA(x)}(x)$.
In this subsection we for any point $x \in \R^n \setminus \Sigma(F)$ replace
the corresponding Arnold class $PA(x)$ by another class $P\tilde A(x),$ whose
potential function $u_{P\tilde A(x)}$ ramifies in exactly the same way, but
which is more convenient because (as we shall see later)
a) it is represented by a cycle lying in the ``finite'' part $W_F \setminus
S(x)$ of $\bar W_F \setminus \bar S(x)$ and thus defining an element $\tilde
A(x)$ of the (much better studied) group ${\mathcal H}(x)$, and
b) if $n$ is even, then this element $\tilde A(x)$ can be obtained by the
``Leray tube operation'' (\ref{leray}) from a certain homology class $\alpha(x)
\in H_{n-2}(W_F \cap S(x))$, so that the action (\ref{mon}) on it is reduced to
the similar action on this more standard group.
\medskip
Indeed, it follows from Proposition 9, that if the class $\gamma' \in {\mathcal
PH}(x)$ is obtained by the Gauss--Manin connection over some path in $\C^n
\setminus \Sigma(F)$ from the Arnold cycle $PA({\bf x})$, where ${\bf x}$ is a
point in the hyperbolicity domain of a compact hyperbolic surface, then the
potential function $u_{\gamma'}(x)$ is a single-valued holomorphic function in
$\C^n \setminus \Sigma(F)$. Therefore the ramification of our integrals defined
by the class $\gamma$ coincides with that defined by the class $\gamma -
\gamma'$ (if both integrals are well-defined).
For any point $x \in {\R}^n \setminus M_F$ we choose canonically some class
$\gamma'$ obtained in this way. Namely, we choose an arbitrary point ${\bf x}
\in \R^n$ in the hyperbolicity domain (if this domain has two components in
$\R^n$, then in the component closest to $x$, i.e. such that the segment
connecting $x$ and ${\bf x}$ has $\le [d/2]$ intersections with $M_f$ ). Then
connect $x$ with ${\bf x}$ by a complex line and take the path in this line
that goes from ${\bf x}$ to $x$ along the real segment and misses any point of
$W_F$ along a small arc in the {\it lower} complex half-line with respect to
this direction (i.e. the half-line into which the vector $i\cdot({\bf x}-x)$ is
directed). See Fig. 2.
\bigskip
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(92.00,13.00)
\put(10.00,10.00){\line(1,0){12.00}} \put(25.00,10.00){\oval(6.00,6.00)[b]}
\put(28.00,10.00){\line(1,0){14.00}} \put(45.00,10.00){\oval(6.00,6.00)[b]}
\put(48.00,10.00){\vector(1,0){13.00}} \put(61.00,7.00){\makebox(0,0)[cc]{$x$}}
\put(11.00,7.00){\makebox(0,0)[cc]{${\bf x}$}}
\put(92.00,8.00){\makebox(0,0)[cc]{Fig. 2}} \put(25.00,10.00){\circle*{1.33}}
\put(45.00,10.00){\circle*{1.33}}
\end{picture}
For any $x \in {\R}^n \setminus M_F$, denote by $PA_{\rm hyp}(x)$ the class in
${\mathcal PH}(x)$ obtained from $PA({\bf x})$ by the Gauss--Manin connection
over this path. We are interested in the monodromy of the class $PA(x) -
PA_{\rm hyp}(x)$, which will be called the {\it reduced Arnold class} and
denoted by $P\tilde A(x)$.
\subsection{Groups ${\mathcal H}(x)$ and the vanishing homology of complete
intersections.}
We shall consider especially carefully the case when the attracting surface
$W_F$ satisfies certain genericity conditions, namely, the following ones.
\medskip
We say that two holomorphic hypersurfaces in $\C^n$ are {\it simple tangent} at
their common point, if in some local holomorphic coordinates with origin at
this point one of them is given by the equality $z_n=0$, and the second by
$z_n= z_1^2+ \cdots + z_{n-1}^2.$
\medskip
{\bf Definition 8.} The polynomial $F$ (and the corresponding hypersurface
$W_F$) is $S$-{\it generic} if the projective closure $\bar W_F$ of $W_F$ is
smooth and transversal to the improper hyperplane ${\C}P^n \setminus {\C}^n$,
its ``infinite part'' $\bar W_F \setminus {\C}^n$ is transversal in the
improper hyperplane $\C P^{n-1}$ to the standard quadric $\{z_1^2 + \cdots +
z_n^2 = 0\}$, i.e. to the boundary of any cone $S(x)$, and additionally the set
of points at which $W_F$ is {\it simple} tangent to appropriate cones $S(x)$ is
dense in the set of all points of tangency of $W_F$ and these cones at their
nonsingular points.
\medskip
The transversality conditions from this definition can be reformulated as
follows: let $\bar F$ be the principal (of degree $d$) homogeneous part of
$F$, and $r^2 \equiv z_1^2 + \cdots + z_n^2$, then the function $\bar F$ has an
isolated singularity at $0$, and also the pair of functions $(\bar F, r^2)$
defines a (homogeneous) complete intersection with an isolated singularity at
$0$. \medskip
{\bf Theorem 2.} {\it Suppose that the algebraic surface $W_F = \{F=0\}$ in
${\C}^n$ is $S$-generic, $\hbox{\rm deg} \ F = d$. Then for a generic $x$ the
ranks of both groups (\ref{abs}), (\ref{abstw}) (in particular, of the group
${\mathcal H}(x)$) are equal to \ $(d-1)^n + (2(d-1)^n-d)/(d-2)$\ if $d
> 2$, and to $\; 2n\;$ if $d=2$.} \medskip
Indeed, the pair of functions $(F, r^2(\cdot \ -x))$ defining the manifolds
$W_F, S(x)$ is a perturbation of the complete intersection $(\bar F, r^2)$,
changing only terms of lower degree of these polynomials. Thus the pair $(W_F,
W_F \cap S(x))$ for smooth $W_F$ and nondiscriminant $x$ is homeomorphic to the
pair $(\tilde X_f, X_f)$ from (\ref{hamm}), and the local system $@(x)$ is
isomorphic to the system $\pm \ZZ$ on $\tilde X_f \setminus X_f,$ see \S \ 2.3.
For the group (\ref{abstw}) the assertion of the theorem follows now from
Theorem 1 and Propositions 1 and 2.
Denote by $\partial W_F$ the ``infinite part'' $\bar W_F \setminus {\C}^n$ of
$\bar W_F$. Then the group (\ref{abs}) is Poincar\'e--Lefschetz dual to the
group $H_{n-1}(\bar W_F, \partial W_F \cup (\bar W_F \cap \bar S(x)))$.
Consider the homological exact sequence of the triple $(\bar W_F, \partial W_F
\cup (\bar W_F \cap \bar S(x)), \partial W_F)$. By Proposition 1 and Poincar\'e
duality in the manifolds $W_F$, $W_F \cap S(x)$, the only nontrivial fragment
in this sequence is
\begin{eqnarray}
0 \to H_{n-1}(\bar W_F, \partial W_F) \to
H_{n-1}(\bar W_F, \partial W_F \cup (\bar W_F \cap \bar S(x))) \to \nonumber \\
\to H_{n-2}(\bar W_F \cap \bar S(x), \partial W_F \cap \bar S(x)) \to 0,
\label{exact}
\end{eqnarray}
and the assertion of our theorem about the group (\ref{abs}) follows from
Proposition 2. \medskip
{\bf Remark.} It is easy to see that the map
\begin{equation}
H_{n-2}(W_F \cap S(x)) \to H_{n-1}(W_F \setminus S(x)), \label{tube}
\end{equation}
conjugate with respect to Poincar\'e dualities to the third arrow in
(\ref{exact}), coincides with the Leray tube operation (\ref{leray}), in
particular in this case this operation is monomorphic.
\medskip
So we have identified the pair $(W_F, W_F \cap S(x))$ with a standard object of
the theory of singularities of complete intersections. The pair of functions
$(F, r^2(\cdot - x))$ defining this complete intersection participates in three
important families, which depend on $n, 1$ and $n+1$ parameters respectively.
Since all of them keep the first function $F$ unmoved, we describe only the
corresponding families of second components. The first family consists of all
functions $r^2(\cdot - \tilde x)$, $\tilde x \in \C^n;$ the second of all
functions $r^2(\cdot - x) - \tau,$ $\tau \in \C,$ and the third of all
functions
\begin{equation}
\label{ttt} \rho_\lambda \equiv z_1^2+ \cdots + z_n^2+ \lambda_1 z_1+ \cdots +
\lambda_n z_n + \lambda_0.
\end{equation}
Denote the parameter space of the third deformation by $T;$ the parameter
spaces $\C^n$ and $\C^1$ of the first and second families are obviously
included in it.
Define the set $\Sigma_T$ as the set of all such points $\lambda \in T$ that
the variety $W_F \cap \{\rho_\lambda=0\}$ is not smooth; the intersection of
$\Sigma_T$ with the parameter space of the first (respectively, the second)
subfamily coincides with $\Sigma(F)$ (respectively, the set $s$ of critical
values of the restriction of $r^2(\cdot - x)$ on $W_F$, see \S \ 2.2).
By the Zariski theorem, the obvious homomorphism $\pi_1(\C^1 \setminus s) \to
\pi_1(T \setminus \Sigma_T)$ is monomorphic, in particular the monodromy group
generated by the action of the latter group in ${\mathcal H}(x)$ coincides with
the standard monodromy group of the complete intersection $(\bar F, r^2)$
considered in \S \ 2.2, 2.3.
\medskip
{\bf Definition 9.} The monodromy group defined by the Gauss--Manin
representation $\pi_1(\C^1 \setminus s) \to Aut ({\mathcal H}(x))$ (or,
equivalently, $\pi_1(T \setminus \Sigma_T) \to Aut ({\mathcal H}(x))$) is
called the {\it big} monodromy group, while the similar monodromy group defined
by the natural action (\ref{mon}) is the {\em small} one.
\medskip
Below we shall see that the small monodromy group actually is a proper subgroup
of the big one. To describe it we need several more reductions and notions.
The subgroup ${\mathcal J}(x) \subset {\mathcal H}(x)$ for any $n$ is defined
as that generated by all vanishing cycles in $W_F \setminus S(x)$ defined by
all paths in $\C^1 \setminus s$ connecting $0$ with all the points of $s$, see
\S \ 2: for even $n$ it coincides with the image of the Leray tube map
(\ref{tube}), for odd $n$ it is just the group $\Im$ described in the end of \S
\ 2.3.
\medskip
On this subgroup there is a symmetric bilinear form $\langle \cdot , \cdot
\rangle$: in the case of odd $n$ it was defined before Proposition 4 (as half
the intersection index), and in the case of even $n$ it is induced by the tube
{\it mono}morphism (\ref{tube}) from the intersection index on the group
$H_{n-2}(W_F \cap S(x))$. By Propositions 3 and 4, for any vanishing cycle
$\alpha \in {\mathcal H}(x)$ $\langle \alpha, \alpha \rangle$ is equal to 2 if
$[{n+1 \over 2}]$ is odd and to $-2$ if $[{n+1 \over 2}]$ is even. \medskip
{\bf Lemma 3.} {\it For any $n$, the action of the big monodromy group on
${\mathcal H}(x)$ preserves the subgroup ${\mathcal J}(x)$ and the bilinear
form $\langle \cdot , \cdot \rangle$ on it.}
\medskip
This follows immediately from the Picard--Lefschetz formulae (\ref{plf1}),
(\ref{pllf}).
Now suppose that the polynomial $F$ is real and hyperbolic.
\medskip
{\bf Theorem 3.} {\it For any point $x$ from the $k$-th zone of $\ {\R}^n
\setminus \nobreak \Sigma(F)$, $k \le [d/2]$, the reduced Arnold class $P\tilde
A(x) = PA(x) - PA_{\rm hyp}(x)$ can be represented by a cycle with support in
$W_F \setminus S(x)$ which is homological in ${\mathcal H}(x)$ to the sum of
$k$ pairwise orthogonal vanishing cycles. In particular, its homology class
$\tilde A(x)$ belongs to the subgroup ${\mathcal J}(x)$, and its
self-intersection index $\langle \tilde A(x), \tilde A(x) \rangle$ is equal to
$2k$ if $[{n+1 \over 2}]$ is odd, and to $-2k$ if $[{n+1 \over 2}]$ is even.}
\medskip
Indeed, these vanishing cycles are constructed as follows. If the point $y \in
\R^n \setminus M_F$ is sufficiently close to a component of $M_F$, then in a
small disc $B \subset \C^n$ centered at $y$ the pair $(W_F, S(y))$ is
diffeomorphic to the pair consisting of the plane $\{x_1=1\}$ and the cone
$S(0);$ it is easy to see that both groups $H_{n-1}(B \cap W_F \setminus S(y))$
and $H_{n-1}(B \cap W_F \setminus S(y), @(y))$ are isomorphic to $\ZZ$ and
generated by vanishing cycles defined by the one-parametric family of maps $(F,
r^2(\cdot - y)-\tau),$ $\tau \in \C^n$ (in the first case this cycle is equal
to the tube around the vanishing cycle in $W_F \cap S(y)$).
\medskip
{\bf Lemma 4} (see [V 94], Lemma 2 in \S \ III.3.4). {\it If we go from the
hyperbolicity domain along a line in $\R^n$ and traverse a component of $M_F$,
then the Arnold class corresponding to the point after the traversing is equal
to the sum of this vanishing cycle and of the similar Arnold cycle for the
point before it transported by the Gauss--Manin connection over the arc of the
path from Fig. 2 connecting them.}
\medskip
In particular, the difference $PA(x)-PA_{hyp}(x)$ for $x$ from the $k$-th zone
is homologous to the sum of $k$ vanishing cycles; by construction all these
cycles lie in the finite domain $W_F \setminus S(x)$. The homology class of
this sum in ${\mathcal H}(x)$ is exactly the promised reduced Arnold class
$\tilde A(x)$, see \S \ 4.4. It remains only to prove that these cycles are
pairwise orthogonal. To do it, consider a model hyperbolic surface: the union
of $[d/2]$ concentric close spheres of radii $1, 1+\varepsilon, \ldots ,
1+([d/2]-1)\varepsilon$ (which do not intersect one another even in the complex
domain) and, if $d$ is odd, one plane distant from these spheres.
Although this surface is not $S$-generic, the above-described construction of
the cycle $\tilde A(x)$ can be accomplished for any point $x$ in the $k$-th
zone where $k \le [d/2]$ and, if $d$ is odd and $k=[d/2],$ then $x$ lies much
closer to the exterior ovaloid than to the additional plane. Then any of our
$k$ vanishing cycles lies on the complexification of its own sphere, in
particular they do not intersect one another, and our assertion is proved for
the (very degenerate) model hyperbolic surface. We can change this surface
arbitrarily weakly so that its closure $\bar W_F$ becomes $S$-generic and
transversal to $S(x)$, but the topological shape of the pair $(W_F, S(x))$ does
not change in a large ball in ${\C}^n$ containing all our $k$ vanishing cycles.
Therefore they have zero intersection indices also for a certain generic
hyperbolic polynomial. Finally, the set of nongeneric real hyperbolic
polynomials, all whose ``nongenericity'' lies in the complex domain, has
codimension at least 2 in the space of all strictly hyperbolic polynomials,
and, by Proposition 6, the space of pairs of the form $\{$a strictly hyperbolic
polynomial $F$ of degree $d$ in $\R^n$; a point $x$ of its $k$-th zone with $k
\le [d/2]\}$ is open and path-connected; this gives our assertion also for
arbitrary generic $F$.
\section{Description of the small monodromy group and finiteness
theorems in the cases $n=2$ and $d=2$}
\subsection{The two-dimensional case}
Let $n=2$. Denote by $\eta(F)$ the number of factors $x_1^2 + x_2^2$ in the
decomposition of the principal part $\bar F$ of the polynomial $F$ into the
simplest real factors. (Of course, if $\eta(F)>0$ then $F$ is not $S$-generic.)
\medskip
{\bf Theorem 4.} {\it The attraction force of the standard charge, distributed
on a hyperbolic curve $\{F=0\}$ of degree $d$ in $\R^2$ coincides in the $k$-th
zone with the sum of two algebraic vector-functions, any of which is $\le
(^{d-\eta(F)}_{\ \ k})$-valued. The same is true for the standard charge with
polynomial density $P$ of degree $\le d-2$.
If the hyperbolic curve $\{F=0\}$ is compact and the density function $P$ is
holomorphic, then the corresponding attraction force coincides in the $k$-th
zone with the sum of two analytic finite-valued (and even algebraic if $P$ is a
polynomial) vector-functions, any of which also is $\le (^{d-\eta(F)}_{\ \
k})$-valued.} \medskip
{\bf Corollary.} {\it If $d$ is even and $\bar F \equiv (x_1^2 + x_2^2)^{d/2}$,
then the attraction force coincides with a rational vector-function in the
``most nonhyperbolic'' $(d/2)$-th zone.} \medskip
{\bf Example.} If $d=2,$ then $\eta(F) \ne 0$ only in the Newtonian case (when
$M_F$ is a circle). In this case the attraction force is single-valued, in all
the other irreducible cases it is 4-valued in the 1-st zone. \medskip
Proof of Theorem 4. If $n=2$, then the surface $S(x)$ consists of two complex
lines through $x$, collinear to the lines $\{x_1 = \pm i\cdot x_2\}$. The
reduced Arnold class $\tilde A(x)$ corresponding to a point $x$ from the $k$-th
zone is represented by $2k$ small circles in $W_F \setminus S(x)$ around the
intersection points of these two lines with $W_F$: $k$ circles around the
points of any line. It follows from the construction of Arnold cycles that all
these circles close to one line are oriented in accordance with the complex
structure of the normal bundle of this line, while close to all points of the
other they are oriented clockwise. The total number of such intersection
points in the finite domain for any line is equal to $d-\eta(F)$. Moving the
point $x$ in ${\C}^2 \setminus \Sigma(F)$ we can only permute these $d-\eta(F)$
circles (and, if $W_F$ is smooth, all permutations can be realized). Therefore
the orbit of the monodromy group consists of $(^{d-\eta(F)}_{\ k})^2$ elements;
this implies Theorem 4. \medskip
{\bf Remark.} Already in this case we see that the small monodromy group
actually is smaller than the big one. Indeed, the standard (``big'') monodromy
group of the complete intersection $(\bar F, r^2)$ in $\R^2$ is just the
permutation group of all $2d$ points of the Milnor fibre. In particular, the
orbit of the reduced Arnold class from the $k$-th zone under this action
consists of $(^{\ \ 2d}_{k,k,2d-2k})$ points, which is much more than
$(^d_k)^2$ provided by Theorem 4 in the case $\eta(F)=0.$
\medskip
{\it Remark about Ivory's second theorem.} Given a hyperbolic surface, do there
exist other surfaces defining the same attraction force in some exterior zone?
If yes, these surfaces define the same ramification locus of the analytic
continuations of these forces. In the case of irreducible plane curves this
locus consists of $d(d-1)$ lines tangent to $W_F$ and parallel to the line $x_1
= i\cdot x_2$ plus $d(d-1)$ lines parallel to the line $x_1 = -i\cdot x_2$. If
$d=2$, the set of curves for which these ramification loci coincide consists of
all conics inscribed in a given rectangle whose sides are parallel to these two
directions. It is easy to see that this set is one-parametric and coincides
with the family of confocal conics. For larger $d$, such copotential families
do not exist or at least are exceptional, because the number $2d(d-1)$ of
conditions that the curves of such a family should satisfy becomes much
greater than the dimension of the space of curves.
\subsection{Reduction of the kernel of the form $\langle \cdot, \cdot
\rangle$ and the case of conical sections}
Denote by $\hbox{Ker} {\mathcal J}(x)$ the kernel of the bilinear form $\langle
\cdot,\cdot \rangle$ on the group ${\mathcal J}(x)$, i.e. the set of all
$\gamma \in {\mathcal J}(x)$ such that $\langle \gamma, \alpha \rangle =0$ for
any $\alpha$. By the Picard--Lefschetz formula, this subspace is invariant
under the monodromy action, and hence this action on the quotient lattice \quad
$\tilde {\mathcal J}(x) \equiv {\mathcal J}(x)/\hbox{Ker} {\mathcal J}(x)$
\quad is well defined. \medskip
{\bf Theorem 5.} {\it If $F$ is $S$-generic, $x \in {\C}^n \setminus
\Sigma(F)$, and $P$ a polynomial of degree $p$, then any form $N^{(\nu)}$ (see
(\ref{functs})) with $|\nu| \ge p+2-d$ takes zero value on $\ \hbox{\rm Ker}
{\mathcal J}(x)$.} \medskip
{\it Proof.} Let $n$ be even, so that ${\mathcal J}(x) = t(H_{n-2}(W_F \cap
S(x_0)))$, see (\ref{tube}). By Poincar\'e duality in $W_F \cap S(x_0)$, the
condition $\gamma \in \hbox{\rm Ker} \ {\mathcal J}(x_0)$ implies that the
cycle $t^{-1}(\gamma) \in H_{n-2}(W_F \cap S(x_0))$ is homologous in the
projective closure $\bar W_F \cap \bar S(x_0) \subset {\C}P^n$ of $W_F \cap
S(x_0)$ to a cycle which lies in the improper subspace $\bar W_F \cap \bar
S(x_0) \cap ({\C}P^n \setminus {\C}^n)$. The tube around this homology provides
the homology of $\gamma$ to some cycle belonging to $ \partial W_F \setminus
\bar S(x_0) \equiv (\bar W_F \setminus \bar S(x_0)) \cap ({\C}P^n \setminus
{\C}^n)$. The last space is an $(n-2)$-dimensional Stein manifold, thus
$\gamma$ is homologous to zero in $\bar W_F \setminus \bar S(x_0)$. On the
other hand, the forms $D^{(\nu)}_x|_{x=x_0}G(x-y)P(y)\omega_F(y)$ with $|\nu|
\ge 2+p-d$ can be extended to holomorphic forms on $\bar W_F \setminus \bar
S(x_0)$, thus their integrals along $\gamma$ are equal to zero.
In the case of odd $n$, the condition $\gamma \in Ker \ {\mathcal J}(x_0)$ also
implies that $\gamma$ is homologous in $\bar W_F \setminus \bar S(x_0)$ (as a
cycle with coefficients in $@(x_0) \otimes {\C}$) to a cycle in the improper
subspace: indeed, by Poincar\'e duality this condition implies that $\gamma$
defines a trivial element of the group $H_{n-1}^{lf}(\bar W_F \setminus S(x_0),
\partial W_F \setminus S(x_0); @(x_0))$, and hence, by the relative part of
Theorem 1a), also of the group $ H_{n-1}(W_F \setminus S(x_0),
\partial W_F \setminus S(x_0); @(x_0) \otimes {\C}))$. The rest
of the proof is the same as for even $n$. \medskip
{\bf Corollary.} {\it In the conditions of Theorem 5, the linear form
$N^{(\nu)}$ induces a form on the quotient lattice $\tilde {\mathcal J}(x)$,
and the number of different values of this form on any orbit of the monodromy
action on ${\mathcal J}(x)$ coincides with similar number for the induced form
and induced monodromy action on} $\tilde {\mathcal J}(x)$. \medskip
{\bf Theorem 6.} {\it For any $n \ge 3$ the potential of the standard charge
(\ref{surfpot}) distributed on a strictly hyperbolic surface $\{F=0\}$ of
degree $2$ in $\R^n$ coincides in the 1-st zone with an algebraic function.}
\medskip
{\bf Proposition 10.} {\it If $n$ is even, $n>2,$ and $F$ is a generic quadric
in ${\C}^n$, then the pair consisting of the corresponding lattice ${\mathcal
J}(x)$ and the bilinear form $\langle \cdot, \cdot \rangle $ on it coincides
with that defined by the extended root system $\tilde D_{n+1}$. For odd $n$
this pair is a direct sum of the lattice $\tilde D_{n+1}$ and the
$(n-1)$-dimensional lattice with zero form on it.} \medskip
This fact in the case of even $n$ and non-twisted homology is proved in [E],
and the calculation for odd $n$ is essentially the same.
\medskip
{\it Proof of Theorem 6.} If $F$ is a {\it generic} quadric, then by the
Proposition 10 the lowered form $\langle \cdot, \cdot \rangle $ on the quotient
lattice $\tilde {\mathcal J}(x)$ is isomorphic to the canonical form on the
lattice $D_{n+1}$, in particular is elliptic. Hence the orbit of any class in
this lattice (in particular of the coset of the reduced Arnold class) under the
reduced monodromy action is finite, and any linear form takes finitely many
values on it.
Finally, the {\it non-generic} quadric $F$ can be approximated by a
one-parameter family $F_\tau,$ $\tau \in (0, \epsilon]$, of generic quadrics.
The analytic continuation of the potential function $u = u(F)$ is equal to the
limit of similar continuations of potentials $u(F_\tau)$. Hence the number of
leaves of $u(F)$ is majorized by the (common) number of leaves of any of the
$u(F_\tau)$. \medskip
This proof estimates the number of leaves of potential functions of quadrics by
the numbers of elements of length $ \sqrt {-2}$ in the lattice $D_{n+1}$. As we
shall see in the next subsection, this majorization is not sharp: a more
precise upper bound is the number of integer points in the intersection of the
sphere of radius $\sqrt{-2}$ with a certain affine sublattice of corank $1$
that does not pass through the origin.
\subsection{Principal theorem on the small monodromy group}
The obvious map $\Pi : \bar W_F \setminus \bar S(x) \to \C P^{n-1} \setminus
PS$ (see \S \ 4.3) induces a homomorphism $\Pi_*$ of the group ${\mathcal
J}(x)$ to the group (\ref{pr1}) (if $n$ is even) or (\ref{pr2}) (if $n$ is
odd). Denote by ${\mathcal M}(x)$ the kernel of this homomorphism. \medskip
{\bf Theorem 7.} {\it Suppose that the polynomial $F:\C^n \to \C$ is
$S$-generic. Then for any $x \in \C^n \setminus \Sigma(F)$
a) the map $\Pi_*$ is epimorphic, in particular ${\mathcal M}(x)$ is a
sublattice of corank 1 in ${\mathcal J}(x)$;
b) ${\mathcal M}(x)$ is spanned by all vectors but one of some basis of
vanishing cycles in ${\mathcal J}(x)$;
c) the small monodromy group in ${\mathcal J}(x)$ is generated by reflections
(with respect to the form $\langle \cdot, \cdot \rangle$) in all the basis
vanishing cycles generating ${\mathcal M}(x)$;
d) the set of these basis vanishing cycles in ${\mathcal M}(x)$ is transitive
under the action of this small monodromy group;
e) the subgroup $\hbox{\rm Ker} \ {\mathcal J}(x) \subset {\mathcal J}(x)$
belongs to ${\mathcal M}(x)$.
If $F$ is a real hyperbolic polynomial of degree $d$, $x$ a point from the
$k$-th zone, $1 \le k \le [d/2],$ and $\tilde A(x)$ the corresponding reduced
Arnold cycle, then additionally
f) $\tilde A(x)$ belongs to $k$ times the generator of the quotient group
${\mathcal J}(x)/{\mathcal M}(x) \sim \ZZ $, in particular does not belong to
${\mathcal M}(x)$;
g) the linear form $\langle \tilde A(x), \cdot \rangle$ on ${\mathcal M}(x)$ is
not trivial.}
\medskip
For the proof of this theorem and next Theorem 8 see \S \ 6.
\medskip
{\bf Corollary.} {\it The orbit of any element of ${\mathcal J}(x)$ under the
small monodromy group lies in some affine hyperplane parallel to ${\mathcal
M}(x)$.}
\medskip
Indeed, this follows from Theorem 7c) and Picard--Lefschetz formula. \medskip
{\bf Definition 10.} A polynomial $P:\C^n \to \C$ is {\it very degenerate} with
respect to $W_F$ if it is equal to $0$ at all points $y \in W_F$ at which
appropriate surfaces of the form $S(x)$ are tangent to $W_F$ at their smooth
points. \medskip
{\bf Theorem 8.} {\it Suppose that $W_F$ is $S$-generic, and the polynomial $P$
is not very degenerate with respect to $W_F$. Then there exist multiindices
$\nu \in \ZZ_+^n$ with arbitrarily large $|\nu|$ such that for a generic $x \in
\C^n \setminus \Sigma(F)$, the restriction on ${\mathcal M}(x)$ of the linear
form $N^{(\nu)}$ (see (\ref{functs})) is not trivial.}
\subsection{Main conjectures}
{\bf Conjecture 1.} {\it If the hyperbolic polynomial $F$ of degree $d \ge 3$
in $\C^n,$ $n \ge 3,$ is $S$-generic, then the potential function of the
standard charge (\ref{polpot}) with not very degenerate $P$ does not coincide
with algebraic functions in the components of $\R^n \setminus \Sigma(F)$ other
than the hyperbolicity domain; moreover, the same is true for some arbitrarily
high partial derivatives of this potential function.} \medskip
Theorem 7 reduces this conjecture to the following Conjecture 2
(proved recently by W.~Ebeling, see the Appendix).
\medskip
{\bf Definition 11.} A triple $\ (A; \langle \cdot,\cdot \rangle ; g) \ $
consisting of an integer lattice $A$, an {\it even} integer-valued symmetric
bilinear form on it and a group $g \subset \hbox{Aut}(A)$ generated by the
reflections in hyperplanes orthogonal to several elements $a_i$ of length
$\sqrt{-2}$ in $A$, is called {\it completely infinite} if for any element $a
\in A$ such that not all numbers $\langle a, a_i \rangle$ are equal to $0$, any
nonzero linear form $A \otimes {\C} \to {\C}$ takes infinitely many values on
the orbit of $a$ under the action of the group $g$. \medskip
{\bf Conjecture 2.} {\it For any $S$-generic polynomial $F$ of degree $d \ge 3$
in ${\C}^n$, $n\ge 3$, the triple consisting of the group ${\mathcal M}(x)$,
the bilinear form equal (up to sign if $[{n+1 \over 2}]$ is odd) to the form
$\langle \cdot, \cdot \rangle$ defined before Lemma 3, and the ``small''
monodromy group on ${\mathcal M}(x)$, is completely infinite.}
\medskip
In [V 94] this conjecture was proved if additionally $n+d \ge 8$.
\medskip
{\bf Proposition 11.} {\it Conjecture 2 implies Conjecture 1.}
\medskip
{\it Proof.} Let $x$ be a nondiscriminant point in the $k$-th zone, $1 \le k
\le [d/2]$, for which the assertion of Theorem 8 with a certain $\nu$ is
satisfied. By Theorem 7b), g) there is a vanishing cycle $\Gamma \in {\mathcal
M}(x)$ such that $\langle \tilde A, \Gamma \rangle \ne 0$. By the
Picard--Lefschetz formula, the monodromy along the corresponding simple loop
takes $\tilde A$ to $\tilde A + \lambda \Gamma$, $\lambda \ne 0$. By Conjecture
2, for generic $x$ the form $N^{(\nu)}$ takes infinitely many values on the
orbit of the added term $\lambda \Gamma$ under the action of the small
monodromy group. On the other hand, this infinite number is estimated from
above by the number $q(q-1),$ where $q$ is the number of values of the form
$N^{\nu}$ on the orbit of $\tilde A(x)$, in particular this number $q$ is also
infinite.
Finally, for the points $x$ from the $([d/2]+1)$-th zone (if it exists) the
assertion of the Conjecture 1 follows from the fact that the potential function
defined by the charge (\ref{polpot}) obviously extends to an analytic function
on ${\R}P^n \setminus M_F$, hence its algebraicity in the $([d/2]+1)$-th zone
is equivalent to that in the zone separated from it by a piece of the improper
subspace in ${\R}P^n$; the number of the latter zone is surely less than
$[d/2]+1$.
\section{Proof of Theorems 7, 8}
All the main characters of statements a)--e) of Theorem 7 corresponding to all
$S$-generic $F$ of the same degree in $\C^n$ and all $x \not \in \Sigma(F)$ are
isomorphic to one another, therefore we can assume that $F$ is a real
hyperbolic polynomial and $x$ a real point.
The proof of statement e) follows immediately from that of Theorem 5.
Any induction step from the proof of Theorem 3 obviously increases the image of
$\tilde A(x)$ under the map $\Pi_*$ by a generator of the target homology
group; all such $k$ steps are locally topologically equivalent, and hence add a
fixed generator of this target group with the same sign. This proves statement
f) of Theorem 7, and statement a) is a direct corollary of it.
For any $k = 1, 2, \ldots, [d/2]$, and any point $x$ in the $k$-th zone,
consider the difference of the projective Arnold class $PA(x)$ and the element
in ${\mathcal PH}(x)$ obtained as in the definition of the reduced Arnold
cycles (i.e. by transportation along an arc in the lower complex half-line)
from a similar class $PA(x')$, $x'$ in the $(k-1)$-st zone. By Lemma 4, if $x$
and $x'$ are sufficiently close to one another and to the $k$-th component of
$M_F$ separating them, then this class can be realized by a cycle lying in a
small disc $B$ containing both these points $x, x'$. Denote by $a(x)$ the class
of this cycle in the group ${\mathcal H}(x)$; by continuity this class $a(x)$
is well defined also for arbitrary $x$ from the same zone (not necessarily
close to $M_F$). By Lemma 4, for all $x$ not in the hyperbolicity domain the
corresponding maps $\Pi_*$ send the elements $a(x)$ into the same element of
the group (\ref{pr1}) or (\ref{pr2}). \medskip
{\bf Theorem 9.} {\it If $n>2$, then
a) all classes $a(x)$, corresponding to all points $x \in {\R}^n \setminus
\Sigma(F)$, $x$ not in the hyperbolicity domain or in the $([d/2]+1)$-th zone,
can be obtained from one another by the Gauss--Manin connection in the homology
bundle $\{ {\mathcal H}(x) \to x \}$ over some path in ${\C}^n \setminus
\Sigma(F)$. These classes $a(x)$ do not belong to ${\mathcal M}(x)$, and any of
them, being added to the set of $\; \hbox{\rm dim} \, {\mathcal J}(x)-1$ basis
elements of ${\mathcal M}(x)$, mentioned in statements b), c) of Theorem 7,
completes this set to a basis in ${\mathcal J}(x)$;
\smallskip
b) for arbitrary $x$ in the $k$-th zone, $1 \le k \le [d/2]$, the linear form
$\langle a(x), \cdot \rangle$ on the group ${\mathcal M}(x)$, defined by our
bilinear form, is nontrivial.}
\subsection{Comparison of big and small monodromy groups}
Now we compare the fundamental groups of ${\C}^n \setminus \Sigma(F)$ and of
the complement of the discriminant variety $\Sigma_T$ of the deformation
(\ref{ttt}) of the complete intersection $(\bar F, r^2)$. Since $F$ is
$S$-generic, the set $\Sigma(F)$ consists of only two components, $W_F$ and the
set of $x \not \in W_F$ such that $S(x)$ is tangent to $W_F$; if $n>2$, then
the latter component is irreducible.
Let us choose the distinguished point ${\bf x}$ of the space $T \setminus
\Sigma_T$ in the hyperbolicity domain of the subspace ${\R}^n \setminus
\Sigma(F)$ . The group $\pi_1(T \setminus \Sigma_T)$ acts in the usual way on
the group ${\mathcal H}({\bf x})$ and generates the ``big'' monodromy group,
see \S \ 4.5.
Let $\Lambda$ be a generic 2-plane in $T$, and $L = \Lambda \cap {\C}^n$; $\bar
U$ a small neighbourhood of $L$ in the projective compactification of $T$, and
$U = \bar U \cap T$ the affine part of $\bar U$. Let $L'$ be a generic line in
$\Lambda$ through ${\bf x}$ sufficiently close to $L$, so that $L' \subset U$
and $L'$ intersects $\Sigma_T$ transversally. \medskip
{\bf Lemma 5.} {\it The obvious maps $\pi_1(L\setminus \Sigma(F)) \to
\pi_1({\C}^n \setminus \Sigma(F))$ and \\ $\pi_1(L' \setminus \Sigma_T) \to
\pi_1(U \setminus \Sigma_T) \to \pi_1(T \setminus \Sigma_T)$ are epimorphic.}
\medskip
The proof follows directly from the generalized Lefschetz theorem (see [GM]).
\medskip
Thus the small and big monodromy groups are generated by simple loops lying in
$L\setminus \Sigma_T$ and $L' \setminus \Sigma_T$, respectively. Let us compare
these collections of loops. \medskip
{\bf Lemma 6.} {\it The group ${\mathcal J}({\bf x})$ is generated by the
cycles vanishing along the paths of an arbitrary distinguished system in $L'$
connecting the distinguished point ${\bf x}$ with all points of $L' \cap
\Sigma_T$.} \medskip
Indeed, the group $\pi_1(L' \setminus \Sigma_T)$ acts on the group ${\mathcal
J}({\bf x})$; this monodromy action is described by the Picard--Lefschetz
formulae, see \S \ 2. Lemma 6 follows from these formulae, from Lemma 5, and
from the fact that the group ${\mathcal J}({\bf x})$ coincides with the linear
hull of the orbit of any vanishing cycle under the action of the big monodromy
group, see [Gab], [E].
\medskip
The set $L \cap \Sigma(F)$ consists of several points of two kinds: the points
of transversal intersection of $L$ and $W_F$ and points $x \not \in W_F$ such
that $S(x)$ is tangent to $W_F$. \medskip
{\bf Lemma 7.} {\it a) Close to a generic point $y$ of the submanifold $W_F
\subset {\C}^n \subset T$ (i.e. to a point at which the generating lines of the
cone $S(y)$ are transversal to $W_F$) the variety $\Sigma_T$ is smooth and has
simple tangency with ${\C}^n$ along $W_F$. In particular, the intersection of
$\Sigma_T$ with any 2-plane $\Lambda$ transversal to $W_F$ coincides close to
the points of $\Lambda \cap W_F$ with a smooth curve having simple tangency
with the line $\Lambda \cap {\C}^n \equiv L$; \smallskip
b) if $F$ is $S$-generic, then close to a generic point of the variety
$(\Sigma(F) \setminus W_F) \subset {\C}^n \subset T$ the variety $\Sigma_T$ is
smooth and intersects ${\C}^n$ transversally along $(\Sigma(F) \setminus
W_F)$.}
\medskip
The proof is immediate. \medskip
Thus the cardinality of $L' \cap \Sigma_T$ is equal to the cardinality of $L
\cap \Sigma(F)$ plus $\; \hbox{deg} \, F$: to any point of $L \cap (\Sigma(F)
\setminus W_F)$ there corresponds one close point of $L' \cap \Sigma_T$, while
to any point of $L \cap W_F$ there correspond two such points; see Fig. 3a.
\begin{figure}
\unitlength=0.80mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(147.00,75.00)
\put(10.00,42.00){\line(1,0){60.00}}
\put(33.00,41.00){\rule{2.00\unitlength}{2.00\unitlength}}
\put(10.00,46.00){\line(6,-1){60.00}} \put(18.00,51.00){\oval(10.00,18.00)[b]}
\put(18.00,42.00){\circle*{2.00}} \put(52.00,33.50){\oval(10.00,17.00)[t]}
\put(52.00,42.00){\circle*{2.00}} \put(58.00,55.00){\line(2,-5){11.67}}
\put(63.00,42.00){\circle{2.00}} \put(45.00,55.00){\line(-1,-6){4.67}}
\put(43.00,42.00){\circle{2.00}} \put(69.00,44.00){\makebox(0,0)[cc]{$L$}}
\put(69.00,39.00){\makebox(0,0)[cc]{$L'$}}
\put(34.00,39.00){\makebox(0,0)[cc]{${\bf x}$}}
\put(34.00,63.00){\makebox(0,0)[cc]{\large $\Lambda$}}
\put(32.00,4.00){\makebox(0,0)[cc]{\large $a$}}
\put(129.00,61.00){\rule{2.00\unitlength}{2.00\unitlength}}
\put(130.00,58.00){\makebox(0,0)[cc]{${\bf x}$}}
\put(140.00,62.00){\circle*{2.00}} \put(145.00,72.00){\circle{2.00}}
\put(145.00,52.00){\circle{2.00}} \put(125.00,50.00){\circle{2.00}}
\put(125.00,74.00){\circle{2.00}} \put(114.00,62.00){\circle*{2.00}}
\put(107.00,62.00){\circle*{2.00}}
\put(129.00,24.00){\rule{2.00\unitlength}{2.00\unitlength}}
\put(131.00,21.00){\makebox(0,0)[cc]{${\bf x}$}}
\put(146.00,35.00){\circle{2.00}} \put(144.00,14.00){\circle{2.00}}
\put(126.00,12.00){\circle{2.00}} \put(123.00,37.00){\circle{2.00}}
\put(130.00,25.00){\line(-1,-3){4.00}} \put(130.00,25.00){\line(-3,5){6.67}}
\put(131.00,26.00){\line(5,3){14.33}} \put(131.00,24.00){\line(4,-3){12.33}}
\put(140.67,27.00){\circle*{1.33}} \put(131.00,25.67){\line(1,0){6.00}}
\put(137.00,25.67){\line(3,1){3.67}} \put(131.00,24.67){\line(1,0){6.00}}
\put(137.00,24.67){\line(5,-3){4.67}} \put(141.67,22.00){\circle*{1.33}}
\put(129.00,25.67){\line(-1,0){10.00}} \put(119.00,25.67){\line(-3,2){4.00}}
\put(115.00,28.33){\circle*{1.33}} \put(129.00,24.67){\line(-1,0){9.67}}
\put(119.33,24.67){\line(-3,-1){5.00}} \put(114.00,22.67){\circle*{1.33}}
\put(129.00,24.67){\line(-3,-1){20.33}} \put(108.67,18.00){\line(0,1){10.33}}
\put(108.67,29.00){\circle*{1.33}} \put(130.00,24.00){\line(-3,-1){22.67}}
\put(107.33,16.33){\line(-1,6){1.33}} \put(106.33,24.00){\circle*{1.33}}
\put(95.00,24.00){\makebox(0,0)[cc]{$L'$}}
\put(95.00,62.00){\makebox(0,0)[cc]{$L$}}
\put(115.00,4.00){\makebox(0,0)[cc]{\large $b$}}
\end{picture}
\centerline{Fig. 3. Lines $L$ and $L'$ and discriminant points in them}
\end{figure}
Since the point ${\bf x}$ lies in the hyperbolicity domain, all points of $L
\cap W_F$ are real. For any such point $y$ belonging to the $k$-th component of
$M_F$, let $y_+ \in {\R}^n \setminus \Sigma(F)$ be a close point in the $k$-th
zone. For such a point $y_+$, the class $a(y_+)$ was defined before Theorem 9.
Let us agree to choose the distinguished system of paths in $L'$ in such a way
that the paths connecting ${\bf x}$ with any two points of $L' \cap \Sigma_T$
arising from the same point $y$ of $L \cap W_F$ go together up to a small
common neighborhood of these two points and are close to the real segment in
$L$ connecting ${\bf x}$ and $y$, while the paths in $L'$ connecting ${\bf x}$
with any other points of $L' \cap \Sigma_T$ do not touch this small
neighborhood; see Fig. 3b. \medskip
{\bf Definition 12.} A point of $L' \cap \Sigma_T$ is of the first kind
(respectively, of the second kind) if it arises from a close point of $W_F$
(respectively, of $\Sigma(F) \setminus W_F$) in $L$ after the move $L \to L'$.
A cycle in ${\mathcal J}({\bf x})$ vanishing over a path of our distinguished
system in $L'$ that connects ${\bf x}$ with a point $y \in \Sigma_T$ is called
a cycle of the first kind (respectively, of the second kind) if this point $y$
is of the first (respectively, the second) kind. \medskip
In Fig. 3b the points of $L \cap W_F$ and the points of the first kind in $L'$
are shown by small black circles, while the points of $L \cap (\Sigma(F)
\setminus W_F)$ and the points of the second kind in $L'$ are shown by white
circles.
\medskip
{\bf Lemma 8.} {\it a) Two cycles of the first kind in ${\mathcal H}({\bf x})$,
vanishing over two distinguished paths connecting ${\bf x}$ with two points of
$L' \cap \Sigma_T$ arising from the same close point $y$ of $L \cap W_F$,
coincide (maybe up to sign); \smallskip
b) this cycle coincides (maybe up to sign) with the cycle $a(y_+)$ transported
from the point $y_+$ to ${\bf x}$ along the path described in the definition of
the reduced Arnold class. In particular, the map $\Pi_*$ sends the homology
class of any such cycle into a generator of the corresponding group (\ref{pr1})
or (\ref{pr2}); \smallskip
c) the monodromy action in the group ${\mathcal H}({\bf x})$, defined by any
simple loop in $L \setminus \Sigma(F)$ going around some point of $L \cap W_F$,
is trivial; \smallskip
d) any cycle in ${\mathcal H}({\bf x})$ vanishing over a path in $L \setminus
\Sigma(F)$ connecting ${\bf x}$ with a point of $\Sigma(F) \setminus W_F$
belongs to the subspace ${\mathcal M}({\bf x})$. In particular, the same is
true for any cycle of the second kind defined by a path of our distinguished
system in $L' \setminus \Sigma_T$ connecting ${\bf x}$ with a point (of the
second kind) of $\Sigma_T$.} \medskip
{\it Proof.} Consider the space of complex lines through ${\bf x}$ transversal
to $\Sigma_T$ in the plane $\Lambda$. Obviously this space is a projective line
with several points removed, one of which is the point $\{L\}$. Consider a
small loop in this space, which starts and finishes at the point $\{L'\}$ and
goes once around the point $\{L\}$. This loop takes one of the two
distinguished paths from statement a) of the lemma into the other, thus this
statement follows.
Statement c) is a direct consequence of a). Indeed, the loop considered there
is homotopic in $\Lambda \setminus \Sigma_T$ to a loop in $L' \setminus
\Sigma_T$ which turns around two discriminant points defining the same
vanishing cycle, thus its monodromy action is equal to the square of the
reflection in the hyperplane orthogonal to this vanishing cycle.
Statement b) follows from Lemma 4 and the local shape of the pair $(W_F$,
${\mathcal S}(\lambda))$ where ${\mathcal S}(\lambda)$ is the variety of zeros
of the polynomial (\ref{ttt}) defined by the discriminant point $\lambda$ of
the first kind. The way in which the pairs of distinguished paths connecting
$x$ with different pairs of points of the first kind miss one another is not
important, because by the proof of Theorem 3 all the cycles of the first kind
that vanish over the paths going from ${\bf x}$ to the points arising from
different points of $L\cap W_F$ on the same side of ${\bf x}$ in $\; \hbox{Re}
\, L$ are pairwise orthogonal.
Statement d) of the lemma follows immediately from the constructions.
\medskip
Thus, the vanishing cycles of the first (respectively, second) kind are exactly
those that are sent by the map $\Pi_*$ into a generator of the group
(\ref{pr1}) or (\ref{pr2}) (respectively, into a zero class). \medskip
{\bf Lemma 9.} {\it Any vanishing cycle of the first kind in ${\mathcal J}({\bf
x})$ can be transformed into any other by a sequence of reflections in the
hyperplanes orthogonal to cycles of the second kind and to this cycle itself.}
\medskip
By the Picard--Lefschetz formula, this lemma follows from the next one.
\medskip
{\bf Lemma 9$'$.} {\it There exists a distinguished system of paths in $L'
\setminus \Sigma_T$ connecting ${\bf x}$ with all points of $L' \cap \Sigma_T$,
such that all vanishing cycles of the first kind defined by this system are
equal to each other.} \medskip
{\it Proof.} (This proof simulates that of the well-known fact that the
fundamental group of the complement of a smooth irreducible algebraic
hypersurface in ${\C}^n, \, n\ge 2$ is isomorphic to ${\ZZ}$.)
Let $y_1$ be any point of $L \cap W_F$. Let us fix an arbitrary path $\gamma_1$
in $L \setminus \Sigma_T$ connecting ${\bf x}$ with $y_1$. Denote by $A^n$ the
space of complex lines in ${\C}^n$, and by $\hbox{Reg}\, (\Sigma(F))$ the
subset of $A^n$ consisting of lines transversal to $\Sigma(F)$. Consider a path
$\chi_1:[0,1] \to A^n$ such that $\chi_1(0) = L$, $\chi_1([0,1)) \subset
\hbox{Reg}\,(\Sigma(F))$, the last point $\chi_1(1)$ is a line transversal to
$\Sigma(F)$ everywhere except for one point of simple tangency with $W_F$, and
one of the two points of $\chi_1(\tau) \cap W_F, \; \tau =1-\varepsilon$ that
coalesce at this tangency point is obtained from the point $y_1$ of the similar
set corresponding to the value $\tau = 0$ during the deformation of the set
$\chi_1(\tau) \cap W_F, \; \tau \in [0,1-\varepsilon]$.
Consider the continuous deformation $\gamma_1[\tau], \,\tau \in [0,1]$, of the
path $\gamma_1$ such that $\gamma_1[0] = \gamma_1$, $\gamma_1[\tau] \subset
\chi_1(\tau)$, and for any $\tau$ the path $\gamma_1[\tau]$ connects in
$\chi_1(\tau) \setminus \Sigma(F)$ a point of $\chi_1(\tau) \cap W_F$ with some
distinguished point ${\bf x}(\tau) \in \gamma_1[\tau] \setminus \Sigma(F)$,
${\bf x}(0)={\bf x}$. At almost the final instant $\tau=1- \varepsilon$, the
endpoint $\gamma_1[1-\varepsilon](1)$ of the path $\gamma_1[1-\varepsilon]$
lies very close to some other point of $\chi_1(1-\varepsilon)\cap W_F$ (with
which it coalesces at the instant $\tau = 1$). Connect this new point with
${\bf x}(1-\varepsilon)$ by a path $\gamma_2[1-\varepsilon]$ in
$\chi_1(1-\varepsilon)\setminus \Sigma(F)$ that goes very close to
$\gamma_1[1-\varepsilon]$ but does not intersect it except for the initial
point. Then construct a continuous family of paths $\gamma_2[\tau] \subset
\chi_1(\tau), \,\tau \in [0,1-\varepsilon]$, such that for any $\tau$ the
corresponding path $\gamma_2[\tau]$ connects a point of $\chi_1[\tau] \cap W_F$
with ${\bf x}(\tau)$ and does not intersect other points of $\chi_1(\tau) \cap
\Sigma(F)$ or of the path $\gamma_1[\tau]$. At the instant $\tau=0$ we get a
path $\gamma_2 \equiv \gamma_2[0] \subset L$ connecting ${\bf x}$ with some
point $y_2$ of $W_F$.
Then consider a new path $\chi_2:[0,1] \to A^n$, $\chi_2([0,1)) \subset
\hbox{Reg} \,(\Sigma(F))$, connecting $L$ with some new simple tangent line to
$W_F$ and having no extra nontransversalities with $\Sigma(F)$, in such a way
that at the last instant $\tau=1$ one of the two points of $\chi_2(\tau) \cap
W_F$ that coalesce at the tangency point is obtained by deformation along our
path $\chi_2$ from one of the points $y_1$ or $y_2$, and the other two points
of these two pairs do not coincide. Arguing as before, we construct a third
path in $L \setminus \Sigma(F)$, connecting ${\bf x}$ with some third point of
$L \cap W_F$, and so on.
After the $(d-1)$-th step we get a system of $d$ nonintersecting paths in $L
\setminus \Sigma(F)$, connecting ${\bf x}$ with all points of $L \cap W_F$.
Complete this family to any distinguished collection of paths connecting ${\bf
x}$ with all points of $L\cap \Sigma(F)$. For the close perturbation $L'
\subset T$ of $L$, take a close distinguished system of paths in $L'$,
connecting the point ${\bf x}$ with all points of $L' \cap \Sigma_T$ in such a
way that to any path in $L$ connecting ${\bf x}$ with $W_F$ there correspond
two paths connecting ${\bf x}$ with two close points of the first kind. This
system of paths is the desired one. For instance, the cycles vanishing along
the (perturbed) paths $\gamma_1$ and $\gamma_2$ define the same vanishing
homology class in ${\mathcal J}({\bf x})$: indeed, a similar assertion for the
cycles in the group ${\mathcal H}({\bf x}(1-\varepsilon)) \equiv H_{n-1}(W_F
\setminus S({\bf x}(1-\varepsilon)), {\ZZ})$ or $ H_{n-1}(W_F \setminus S({\bf
x}(1-\varepsilon)), @({\bf x}(1-\varepsilon)))$ is proved just as the statement
a) of Lemma 8, and for other values of $\tau \in [0,1-\varepsilon]$ it follows
by continuity. Lemmas $9'$ and 9 are thus proved.
\medskip
Now we are ready to prove statement b) of Theorem 7. Indeed, by Lemma 6 the
group ${\mathcal J}({\bf x})$ is generated by the vanishing cycles of the first
and second kind. By Lemma 9 and the Picard--Lefschetz formula, all vanishing
cycles of the first kind lie in the linear span of an arbitrary one of them
(for which we can take the class obtained by the Gauss--Manin connection from
$a(x)$, $x$ from the ${\bf k}$-th zone, $1 \le k \le [d/2]$, see statement b)
of Lemma 8) and the vanishing cycles of the second kind (which lie in
${\mathcal M}({\bf x})$, see statement d) of Lemma 8). \smallskip
Statement c) of Theorem 7 follows immediately from statement c) of Lemma 8, and
statement d) follows from the fact that the variety $\Sigma(F) \setminus W_F$
is irreducible. \medskip
{\it Proof of Theorem 9a).} We can assume that the points $y_1$ and $y_2$,
whose classes $a(y_1)$ and $a(y_2)$ we want to transfer to each other, lie very
close to the ``interior'' (i.e. closest to the hyperbolicity domain) components
of $M_F$ bounding corresponding zones. For such $y_i$ the class $a(y_i)$ is
realized by a cycle generating the group $ H_{n-1}(W_F \cap B \setminus
S(y_i))$ or $ H_{n-1}(W_F \cap B \setminus S(y_i), @(y_i))$, where $B$ is a
small neighbourhood of $y_i$; see Lemma 4. Thus, for the desired path
connecting $y_1$ and $y_2$ we can take the path that goes very close to the set
of generic points of $W_F$ (i.e. of such points $y$ close to which all the
generating lines of the cones $S(y)$ are transversal to $W_F$ and hence the
pairs $(W_F, S(y))$ have locally the same topological structure). \smallskip
Statement b) of Theorem 9 follows from Theorem 3 and the connectedness of
Dynkin diagrams of isolated singularities of complete intersections.
\medskip
{\it Proof of the statement g) of Theorem 7.}
First of all, this statement is true in the case when $M_F$ is an ellipsoid
with different eigenvalues. Indeed, by Theorem 3 in this case $\tilde A(x)$ is
a vanishing cycle, and the assertion follows from the connectedness of the
Dynkin diagram and the fact that the group ${\mathcal M}(x)$ is nontrivial for
such $F$, see e.g. [E].
For arbitrary $d$, consider the model (not $S$-generic) hyperbolic surface
$M_{F}$ consisting of $[d/2]$ ellipsoids $ \alpha_1 x_1^2 + \cdots + \alpha_n
x_n^2 = j, \quad j=1, 1+\varepsilon, \ldots, 1+([d/2]-1)\varepsilon, $ where
all $\alpha_i$ are positive and distinct, plus, if $d$ is odd, a distant
hyperplane. The class $\tilde A(x)$ for $x$ from the $k$-th zone, $1\le k \le
[d/2]$, is then equal to the sum of $k$ vanishing cycles, each of which lies in
the complexification of its own ellipsoid; see the proof of Theorem 3. By the
previous special case of a single ellipsoid, in each of these $k$ complexified
ellipsoids ${\mathcal E}_i$ there is a compact cycle $\Gamma$ defining an
element of the group $ H_{n-1}({\mathcal E}_i \setminus S(x))$ if $n$ is even,
or in $ H_{n-1}({\mathcal E}_i \setminus S(x), @(x))$ if $n$ is odd, such that
$\langle \tilde A(x), \Gamma \rangle \ne 0$ and the map $\Pi_*$ sends the
homology class of $\Gamma$ into the zero homology class.
Consider a perturbation of our model hyperbolic polynomial $F$ which replaces
it by a $S$-generic one and is so weak that it does not change the topology of
the variety $W_{F} \cup S(x)$ inside a sufficiently large disc, in which the
cycles $\Gamma$ and $\tilde A(x)$ lie. The cycle $\tilde \Gamma$ close to
$\Gamma$ in the moved manifold $W_F$ satisfies all the above conditions, and
statement g) of Theorem 7 is proved for {\em some} $S$-generic hyperbolic
polynomial. For an arbitrary such polynomial this statement follows from the
fact that all the generic surgeries separating different path-components of the
space of all strictly hyperbolic $S$-generic surfaces of given degree in $\R^n$
(these surgeries correspond to the smooth hyperbolic surfaces in $\R P^n$
simple tangent to the non-proper plane) preserve the homology classes $\tilde
A(x)$ (provided that the corresponding point $x$ and the distinguished point in
the hyperbolicity domain do not change in this surgery). (In formal terms, this
preservation means that these homology classes corresponding to the polynomials
before and after the surgery are transposed into one another by the natural
connection over any {\em short} connecting them path in the space of all {\em
complex} $S$-generic polynomials.)
\subsection{Proof of Theorem 8}
Let $c$ be a point of simple tangency of a cone $S(x_0)$ and $W_F$ such that
$P(c) \ne 0$. Let ${\Upsilon}$ be an affine complex line through $x_0$ in
$\C^n$, transversal to the common tangent hyperplane of $S(x_0)$ and $W_F$ at
$c$; let $\xi$ be an affine coordinate on it with the origin at $x_0$. Consider
the one-parametric family of surfaces $S(x(\xi)),$ $x(\xi) \in {\Upsilon}$. The
elements $S(x(\xi))$ of this family with $\xi$ from a small {\it punctured}
neighborhood of the origin are transversal to $W_F$ in a small disc $B$ centred
at $c$, and the vanishing element $\gamma(\xi)$ of the group $H_{n-1}(B \cap
W_F \setminus S(x(\xi)))$ (if $n$ is even) or $H_{n-1}(B \cap W_F \setminus
S(x(\xi)), @(x(\xi)))$ (if $n$ is odd) is well defined (up to sign) by this
family. By the Picard--Lefschetz formulae of \S \ 2, in both cases the
rotation of $\xi$ around $0$ sends $\gamma(\xi)$ to $-\gamma(\xi)$.
Define the function $\Xi(\xi), \, $ $\xi \in \C$, as the integral of the form
(\ref{polpot}) with $x=x(\xi) \in \Upsilon$ along the cycle $\gamma(\xi)$. It
is sufficient to prove that there are arbitrarily high derivatives of this
function not equal identically to 0. This follows from the next lemma.
\medskip
{\bf Lemma 11.} {\it The function $\Xi(\xi)$ is represented by a power series
of the variable $\sqrt{\xi}$, whose leading (of smallest degree) term with
non-zero coefficient has degree $1$.}
(Of course, all even powers of this series vanish.)
\medskip
{\it Proof.} Using the Leray residue theorem, we can replace the integral
(\ref{polpot}) along the cycle $\gamma(\xi)$ by the integral of the form
$G(x(\xi)-z)P(z)/F(z)dz_1 \wedge \ldots \wedge dz_n$ along the Leray tube $t
\gamma(\xi) \in H_n(B\setminus (W_F\cup S(x(\xi))))$ or $\in H_n(B\setminus
(W_F\cup S(x(\xi))), @(x(\xi)))$. Close to $c$ the holomorphic function $\C^n
\to \C$ is defined, which assigns to any point the coordinate $\xi$ of the
origin $x(\xi)$ of the cone $S(x(\xi))$ containing it. Choose this function for
the last local coordinate $w_n$ at $c$; by the Morse lemma we can choose the
remaining coordinates $w_1, \ldots, w_{n-1}$ in such a way that $W_F$ is
locally given by $w_n = w_1^2 + \cdots + w_{n-1}^2.$ In these coordinates our
differential form becomes
\begin{equation}
\label{integ} (w_n-\xi)^{-(n-2)/2}(w_n - w_1^2 - \cdots - w_{n-1}^2)^{-1}
I(w_1, \ldots, w_n)dw_1 \wedge \ldots \wedge dw_n,
\end{equation}
where the function $I$ does not vanish at $c$. Let $I= I_0 + I_1 + \cdots $ be
the expansion of $I$ into the sum of quasihomogeneous polynomials of degrees
$0, 1 , \ldots$ respectively with respect to the weights $\deg w_1 = \cdots =
\deg w_{n-1}=1, \deg w_n=2.$ Using the corresponding group of quasihomogeneous
dilations $(w_1, \ldots, w_{n-1}, w_n) \to (\tau w_1, \ldots, \tau w_{n-1},
\tau^2w_n)$ we see, that the integral along $t \gamma(\xi)$ of the form similar
to (\ref{integ}), in which $I_m$ is substituted instead of $I$, is a
homogeneous function in $\xi$ of degree $(m+1)/2$. It is easy to calculate that
this function corresponding to the constant polynomial $I_0 \ne 0$ is not the
identical zero function; this proves our Lemma.
|
\section{Introduction}
\noindent Throughout, all rings considered are commutative with unity and all modules are unital. The following diagram of ring homomorphisms
$$\xymatrix{ R \ar[d]^{\mu_2}\ar[r]^{\iota_2} & \ar[d]^{\mu_1}T\\A \ar[r]^{\iota_1}& B}$$
is called the pullback (or fiber product) of $\mu_1$ and $\iota_1$ if the homomorphism $\iota_2\times \mu_2: R\rightarrow T\times A$, $r\mapsto (\iota_2(r),\mu_2(r))$ induces an isomorphism of $R$ onto the subring of $T\times A$ given by $$\mu_1\times_{B}\iota_1:=\big\{(t,a)\mid \mu_1(t)=\iota_1(a)\big\}.$$
If $\mu_1$ is surjective and $\iota_1$ is injective, the above diagram is called a conductor square. In this setting, $\iota_2$ and $\mu_2$ are injective and surjective, respectively, and $\Ker(\mu_1)\cong\Ker(\mu_2)$. By abuse of notation, we view $R$ as a subring of $T$ making $\Ker(\mu_1)=\Ker(\mu_2)$ the largest common ideal of $R$ and $T$; it is called the conductor of $T$ into $R$.
\emph{Amalgamated algebras} are rings which arise as special pullbacks. Their introduction in 2007 by D'Anna and Fontana \cite{DF1,DF2} was motivated by a construction of D. D. Anderson \cite{An} related to a classical construction due to Dorroh \cite{Do} on endowing a ring (without unity) with a unity. The interest of these amalgamations resides, partly, in their ability to cover several basic constructions in commutative algebra, including classical pullbacks (e.g., $D+M$, $A+XB[X]$, $A+XB[[X]]$, etc.), Nagata's idealizations \cite{Hu,N} (also called trivial ring extensions which have been widely studied in the literature), and Boisen-Sheldon's CPI-extensions \cite{BSh2}. The following paragraphs collect background and main contributions on amalgamations.
Let $A$ be a ring, $I$ an ideal of $A$, and $\pi:A\rightarrow \frac{A}{I}$ the canonical surjection. The amalgamated duplication of $A$ along $I$, denoted by $A\bowtie I$, is the special pullback of $\pi$ and $\pi$; i.e., the subring of $A \times A$ given by $$A\bowtie I:=\pi\times_{\frac{A}{I}}\pi=\big\{(a,a+i)\mid a\in A, i\in I\big\}.$$
If $I^{2} = 0$, then $A\bowtie I$ coincides with Nagata's idealization $A\ltimes I$.
In 2007, the construction $A\bowtie I$ was introduced and its basic properties were studied by D'Anna and Fontana in \cite{DF1,DF2}. In the firs paper \cite{DF1}, they discussed the main properties of the amalgamated duplication in relation with pullback constructions and special attention was devoted to its ideal-theoretic properties as well as to the topological structure of its prime spectrum. In the second paper \cite{DF2}, they restricted their attention to the case where $I$ is a multiplicative canonical ideal of $A$, that is, $I$ is regular and every regular fractional ideal $J$ of $R$ is $I$-reflexive (i.e., $J=(I :(I :J))$). In particular, they examined contexts where every regular fractional ideal of $A\bowtie I$ is divisorial. Later in the same year, the amalgamated duplication was investigated by D'Anna in \cite{D} with the aim of applying it to curve singularities (over algebraic closed fields) where he proved that the amalgamated duplication of an algebroid curve along a regular canonical ideal yields a Gorenstein algebroid curve \cite[Theorem 14 and Corollary 17]{D}. In 2008, Maimani and Yassemi studied in \cite{MY} the diameter and girth of the zero-divisor graph of an amalgamated duplication. In 2010, Shapiro \cite{Sh} corrected Proposition 3 in \cite{D} and proved a pertinent result asserting that if $A$ is a one-dimensional reduced local Cohen-Macaulay ring and $A\bowtie I$ is Gorenstein, then $I$ must be regular. In 2012, in \cite{CJKM}, the authors established necessary and sufficient conditions for an amalgamated duplication of a ring along an ideal to inherit Pr\"ufer conditions (which extend the notion of Pr\"ufer domain to commutative rings with zero divisors). The new results yielded original and new families of examples issued from amalgamated duplications subject to various Pr\"ufer conditions.
In 2009 and 2010, D'Anna, Finocchiaro, and Fontana considered the more general context of amalgamated algebra $$A\bowtie^{f} J:=\big\{(a,f(a)+j)\mid a\in A, j\in J\big\}$$ for a given homomorphism of rings $f: A\rightarrow B$ and ideal $J$ of $B$. In particular, they have studied these amalgamations in the frame of pullbacks which allowed them to establish numerous (prime) ideal and ring-theoretic basic properties for this new construction. In \cite{DFF1}, they provided necessary and sufficient conditions for $A\bowtie^{f} J$ to inherit the notions of Noetherian ring, domain, and reduced ring and characterized pullbacks that can be expressed as amalgamations. In \cite{DFF2}, they provided a complete description of the prime spectrum of $A\bowtie^{f} J$ and gave bounds for its Krull dimension.
Let $\alpha: A\rightarrow C$, $\beta: B\rightarrow C$ and $f:A\rightarrow B$ be ring homomorphisms. In the aforementioned papers \cite{DFF1,DFF2}, the authors studied amalgamated algebras within the frame of pullbacks $\alpha\times \beta$ such that $\alpha=\beta\circ f$ \cite[Propositions 4.2 and 4.4]{DFF1}. In this work, we are interested in new constructions, called \emph{bi-amalgamated algebras} (or \emph{bi-amalgamations}), which arise as pullbacks $\alpha \times \beta$ such that the following diagram of ring homomorphisms
$$\xymatrix{ A \ar[d]^{g}\ar[r]^{f} & \ar[d]^{\alpha}B\\ C \ar[r]^{\beta}& D}$$
is commutative with $\alpha\circ \pi_B(\alpha\times \beta)=\alpha\circ f(A)$, where $\pi_B$ denotes the canonical projection of $B\times C$ over $B$. Namely, let $f: A\rightarrow B$ and $g: A\rightarrow C$ be two ring homomorphisms and let $J$ and $J'$ be
two ideals of $B$ and $C$, respectively, such that $f^{-1}(J)=g^{-1}(J')$. The \emph{bi-amalgamation of $A$ with $(B, C)$ along $(J, J')$ with respect to $(f,g)$} is the subring of $B\times C$ given by
$$A\bowtie^{f,g}(J,J'):=\big\{(f(a)+j,g(a)+j') \mid a\in A, (j,j')\in J\times J'\big\}.$$
This paper investigates ring-theoretic properties of \emph{bi-amalgamations} and capitalizes on previous works carried on various settings of pullbacks and amalgamations. In the second and third sections, we provide examples of bi-amalgamations and show how these constructions arise as pullbacks. The fourth section investigates the transfer of some basic ring theoretic properties to bi-amalgamations and the fifth section is devoted to the prime ideal structure of these constructions. All new results agree with recent studies in the literature on D'Anna-Finocchiaro-Fontana's amalgamations and duplications.
Throughout, for a ring $R$, $Q(R)$ will denote the total ring of quotients and $\Ze(R)$ and $\J(R)$ will denote, respectively, the set of zero divisors and Jacobson radical of $R$. Finally, $\Spec(R)$ shall denote the set of prime ideals of $R$.
\section{Examples of bi-amalgamations}
Notice, first, that every amalgamated duplication is an amalgamated algebra and every amalgamated algebra is a bi-amalgamated algebra, as seen below.
\begin{example}[The amalgamated algebra]\label{aa} Let $f:A\rightarrow B$ be a ring homomorphism and $J$ an ideal of $B$.
Set $I:=f^{-1}(J)$ and $\iota:=\id_A$. Thus,
\begin{eqnarray*}
A\bowtie^{\iota, f}(I,J) &=& \big\{(a+i,f(a)+j) \mid a\in A , (i,j)\in I\times J\big\} \\
&=&\big\{(a+i,f(a+i)+j-f(i))\mid a\in A , (i,j)\in I\times J\big\} \\
&=& \big\{(a,f(a)+j)\mid a\in A , j\in J\big\} \\
&=& A\bowtie^fJ.
\end{eqnarray*}
\end{example}
Further, the subring $f(A)+J$ of $B$ can be regarded as a bi-amalgamation; precisely:
\begin{remark}\label{f(A)+J}
Let $f:A\rightarrow B$ be a ring homomorphism and $J$ and ideal of $B$. Set $I:=f^{-1}(J)$ and consider the canonical projection $\pi: A\rightarrow A/I$. Then, one can easily check that
\begin{eqnarray*}
f(A)+J &\cong &\big\{(\bar{a},f(a)+j)\mid a\in A , j\in J\big\} \\
&= &A\bowtie^{\pi,f}(0,J).
\end{eqnarray*}
\end{remark}
In particular, Boisen-Sheldon's CPI-extensions \cite{BSh2} can also be viewed as bi-amalgamations.
\begin{example}[The CPI-extension]
Let $A$ be a ring and let $I$ be an ideal of $A$. Then $\overline{S}:=(A/I)\setminus\Ze(A/I)$ and $S:=\{s\in A \mid \bar{s}\in \overline{S}\}$ are multiplicatively closed subsets of $A/I$ and $A$, respectively. Let $\varphi: S^{-1}A\rightarrow Q(A/I)=(\overline{S})^{-1}(A/I)$ and $f: A\rightarrow S^{-1}A$ be the canonical ring homomorphisms. Then, the subring $$C(A,I):=\varphi^{-1}(A/I)=f(A) + S^{-1}I$$ of $S^{-1}A$ is called the CPI-extension of $A$ with respect to $I$ (in the sense of Boisen-Sheldon). Now, let $\pi: A\rightarrow A/I$ be the canonical projection. From Remark~\ref{f(A)+J}, we have
$$A\bowtie^{\pi,f}(0, S^{-1}I)\cong f(A) + S^{-1}I=C(A,I).$$
\end{example}
Other known families of rings stem from Remark~\ref{f(A)+J}; namely, those issued from extensions of rings $A\subset B$ (including classic pullbacks).
\begin{example}[The ring $A+J$]
Let $i:A\hookrightarrow B$ be an embedding of rings, $J$ and ideal of $B$, $I:=A\cap J$, and $\pi: A\rightarrow A/I$ the canonical projection. From Remark~\ref{f(A)+J}, the subring $A+J$ of $B$ can arise as a bi-amalgamation via $$A+J\cong A\bowtie^{\pi,i}(0,J)$$
and, consequently, so do most classic pullback constructions such as $A+XB[X]$ (via $A\subset B[X]$ and $XB[X]$), $A+XB[[X]]$ (via $A\subset B[[X]]$ and $XB[[X]]$), and $D+M$ (via $D\subset T$ and $M$ ideal of $T$ with $D\cap M=0$).
\end{example}
In the next section, as an application of Proposition~\ref{cara}, we will see that some
glueings of prime ideals \cite{Ped,Tam,Tra,Yan} can be viewed as bi-amalgamations. We
close this section with an explicit (non-classic pullback) example; namely, the ring
$R:=\Z[X]+(X^{2}+1)\Q[X]$ which lies between $\Z[X]$ and $\Q[X]$.
\begin{example}
Let $i:\Z[X]\hookrightarrow \Q[X]$ be the natural embedding and consider the ring homomorphism $\pi:\Z[X]\rightarrow \Z[i]$, $p(X)\mapsto p(i)$. Clearly, $(X^2+1)\Q[X]\cap \Z[X]=(X^2+1)$ and $\dfrac{\Z[X]}{(X^2+1)}\cong\Z[i]$ so that $$R:=\Z[X]+(X^{2}+1)\Q[X]\cong \Z[X]\bowtie^{\pi,i}\big(0,(X^2+1)\Q[X]\big).$$
\end{example}
\section{Pullbacks and bi-amalgamations}
\noindent Throughout, let $f: A\rightarrow B$ and $g: A\rightarrow C$ be two ring homomorphisms and $J,J'$ two ideals of $B$ and $C$, respectively, such that $I:=f^{-1}(J)=g^{-1}(J')$. Let $A\bowtie^{f,g}(J,J')$ denote the bi-amalgamation of $A$ with $(B, C)$ along $(J, J')$ with respect to $(f,g)$.
This section sheds light on the correlation between pullback constructions and bi-amalgamations. We first show how every bi-amalgamation can arise as a natural pullback.
\begin{proposition}\label{prop1}
Consider the ring homomorphisms $\alpha: f(A)+J\rightarrow A/I$, $f(a)+j\mapsto \bar{a}$ and $\beta: g(A)+J'\rightarrow A/I$, $g(a)+j'\mapsto \bar{a}$. Then, the bi-amalgamation is determined by the following pullback
$$\xymatrix{
A\bowtie^{f,g}(J,J') \ar@{>>}[d] \ar@{>>}[r] & \ar[d]^{\alpha}f(A)+J\\
g(A)+J' \ar[r]^{\beta} & A/I
}$$
that is
$$A\bowtie^{f,g}(J,J')=\alpha\times_{\frac{A}{I}}\beta.$$
\end{proposition}
\begin{proof}
Note that the mappings $\alpha$ and $\beta$ are well defined since $I:=f^{-1}(J)=g^{-1}(J')$ and are ring homomorphisms. Further, the inclusion $A\bowtie^{f,g}(J,J')\subseteq \alpha\times \beta $ is trivial. On the other hand,
$$\alpha\times_{\frac{A}{I}}\beta=\big\{(f(a)+j,g(b)+j')\mid a,b\in A,\ (j,j')\in J\times J',\ \alpha(a)=\beta(b)\big\}.$$
The condition $\alpha(a)=\beta(b)$ means that $f(b-a)\in J$ and $g(b-a)\in J'$. It follows that $g(b)+j'=g(a)+(j'+g(b-a))$ with $j'+g(b-a)\in J'$. Therefore, $\alpha\times \beta \subseteq A\bowtie^{f,g}(J,J')$.
\end{proof}
Next, we see how bi-amalgamations can be represented as conductor squares.
\begin{proposition}\label{conductor}
Consider the following ring homomorphisms
$$\begin{array}{cccc}
\iota_1: & \dfrac{A}{I} & \longrightarrow & \displaystyle\frac{f(A)+J}{J}\times \frac{g(A)+J'}{J'} \\
& \bar{a} & \longmapsto & \left(\overline{f(a)},\overline{g(a)}\right)\\
&&&\\
\mu_2: & A\bowtie^{f,g}(J,J') & \longrightarrow & \dfrac{A}{I} \\
& (f(a)+j,g(a)+j') & \longmapsto & \bar{a}
\end{array}$$
Then, the following diagram
$$\xymatrix@R=1.5cm @C=1.5cm{ A\bowtie^{f,g}(J,J') \ar@{>>}[d]^{\mu_2}\ar[r]^{\iota_2} & \ar@{>>}[d]^{\mu_1}(f(A)+J)\times (g(A)+J')\\
\displaystyle\frac{A}{I}\ar[r]^{\iota_1}& \displaystyle \frac{f(A)+J}{J}\times \frac{g(A)+J'}{J'}
}$$
is a conductor square with conductor $\Ker(\mu_1)=J\times J'$, where $\iota_2$ is the natural embedding and $\mu_1$ is the canonical surjection.
\end{proposition}
\begin{proof}
The mappings $\iota_1$ and $\mu_2$ are well defined since $I=f^{-1}(J)=g^{-1}(J')$ and are ring homomorphisms. Next,
set $R:=\mu_1\times \iota_1$ and let $a\in A$ and $(j,j')\in J\times J'$. Then
$$\iota_2\times \mu_2\big((f(a)+j,g(a)+j')\big)=\big((f(a)+j,g(a)+j'),\bar{a}\big)$$
with
$$\mu_1\big((f(a)+j,g(a)+j')\big)=\big(\overline{f(a)},\overline{g(a)}\big)=\iota_1(\bar{a}).$$
Thus, $\iota_2\times \mu_2\big(A\bowtie^{f,g}(J,J')\big)\subseteq R$. Now, let $\big((f(a)+j,g(a')+j'),\bar{b}\big) \in R$. Then
$$\big(\overline{f(a)},\overline{g(a')}\big)=\big(\overline{f(b)},\overline{g(b)}\big).$$
Hence, $f(a-b)\in J$ and $g(a'-b)\in J'$. Whence,
$$\iota_2\times \mu_2\big((f(b)+f(a-b)+j,g(b)+g(a'-b)+j')\big)=\big((f(a)+j,g(a')+j'),\bar{b}\big).$$
It follows that $\iota_2\times \mu_2$ induces an isomorphism of $A\bowtie^{f,g}(J,J')$ onto $R$ since $\iota_2\times \mu_2$ is injective. Consequently, the above diagram is a pullback. Moreover, it is clear that $\iota_1$ is injective and that $\Ker(\mu_1)=J\times J'=\Ker(\mu_2)$.
\end{proof}
The next result characterizes pullbacks that can arise as bi-amalgamations.
\begin{proposition}\label{cara}
Consider the following diagram
$$\xymatrix{ A \ar[d]^{g}\ar[r]^{f} & \ar[d]^{\alpha}B\\ C \ar[r]^{\beta}& D}$$
of ring homomorphisms and let $\pi:B\times C\rightarrow B$ be the canonical projection. Then, the following conditions are equivalent:
\begin{enumerate}
\item $\alpha\times_D \beta=A\bowtie^{f,g}(J,J')$, for some ideals $J$ of $B$ and $J'$ of $C$ with $f^{-1}(J)=g^{-1}(J')$;
\item The above diagram is commutative with $\alpha\circ \pi(\alpha\times_D \beta)=\alpha\circ f(A)$.
\end{enumerate}
\end{proposition}
\begin{proof}
$(1)\Rightarrow(2)$ Let $a\in A$. By hypothesis, $(f(a),g(a))\in \alpha\times_D \beta$ so that $\alpha\circ f(a)=\beta\circ g(a)$. Also, we have $\pi(\alpha\times_D \beta)=f(A)+J$. Further, for any $j\in J$, the fact $(j,0)\in A\bowtie^{f,g}(J,J')$ yields $\alpha(j)=\beta(0)=0$. Therefore, $\alpha\circ \pi(\alpha\times_D \beta)=\alpha\circ f(A)$, as desired.
$(2)\Rightarrow(1)$ Let $J:=\Ker(\alpha)$ and $J':=\Ker(\beta)$. By assumption, for each $x\in f^{-1}(J)$, $\beta\circ g(x)=\alpha\circ f(x)=0$. Then, $g(x)\in J'$ and hence $f^{-1}(J)\subseteq g^{-1}(J')$. Likewise for the reverse inclusion. Hence $f^{-1}(J)=g^{-1}(J')$. Next, let $(f(a)+j,g(a)+j')\in A\bowtie^{f,g}(J,J')$. We have
$$\alpha(f(a)+j)=\alpha\circ f(a)=\beta\circ g(a)=\beta(g(a)+j')$$
so that $A\bowtie^{f,g}(J,J')\subseteq\alpha\times_D \beta$. On the other hand, let $(b,c)\in \alpha\times_D \beta$. By assumption, there exists $a\in A$ such that
$$\alpha(b)=\alpha\circ \pi(b,c)=\alpha(f(a)).$$
Then, $b-f(a)\in J$. Moreover, we have
$$\beta(c)=\alpha(b)=\alpha(f(a))=\beta(g(a)).$$
Then, $c-g(a)\in J'$. It follows that
$$(b,c)=(f(a)+b-f(a),g(a)+c-g(a)\in A\bowtie^{f,g}(J,J').$$
Consequently, $\alpha\times_D \beta=A\bowtie^{f,g}(J,J')$, completing the proof of the proposition.
\end{proof}
In view of Example~\ref{aa}, Proposition~\ref{cara} recovers the special case of amalgamated algebras, as recorded in the next corollary.
\begin{corollary}[{\cite[Proposition 4.4]{DFF1}}]
Let $\alpha: A\rightarrow D$ and $\beta: B\rightarrow D$ be two ring homomorphisms.
Then, $\alpha\times_D \beta=A\bowtie^{f}J$, for some ideal $J$ of $B$ if and only if $\alpha=\beta\circ f$.
\qed\end{corollary}
We close this section with a brief discussion on Traverso's Glueings of prime
ideals \cite{Ped,Tam,Tra} which are special pullbacks \cite[Lemma 2]{Yan}. So, they can also be viewed as special bi-amalgamations if they satisfy Condition (2) of Proposition~\ref{cara}. Precisely, from \cite[Lemma 1]{Yan}, let $A$ be a Noetherian ring and $B$ an overring of $A$ such that $B$ is a finite $A$-module. Let $p\in\Spec(A)$ and let $p_{1}, ..., p_{n}$ be the prime ideals of $B$ lying over $p$. For each $i$, $\frac{A_{p}}{pA_{p}}$ is a subfield of $\frac{B_{p_{i}}}{p_{i}B_{p_{i}}}$, and let $\overline{\frac{b}{t}}^{i}$ denote the class of the element $\frac{b}{t}$ of $B_{p_{i}}$ modulo $p_{i}B_{p_{i}}$. The ring $A'$ obtained from $B$ by glueing over $p$
is the subring of $B$ (containing $A$) given by
$$A':=\left\{b\in B \mid \exists \frac{a_{o}}{s_{o}}\in A_{p}\ \text{with}\ \overline{\frac{b}{1}}^{i}=\overline{\frac{a_{o}}{s_{o}}}^{i}\ \forall i\ \text{and, for}\ \frac{a}{s}\in A_{p},\ \overline{\frac{b}{1}}^{i}=\overline{\frac{a}{s}}^{i}\Leftrightarrow\overline{\frac{b}{1}}^{j}=\overline{\frac{a}{s}}^{j}\ \forall i,j\right\}.$$
Now, consider the following diagram
$$\xymatrix@R=1.5cm @C=2.5cm{ A \ar[d]^{\mu}\ar[r]^{\iota} & \ar[d]^{\Phi}B\\ \frac{A_{p}}{pA_{p}} \ar[r]^{\Psi}& D:=\frac{B_{p_{1}}}{p_{1}B_{p_{1}}}\times \dots \times \frac{B_{p_{n}}}{p_{n}B_{p_{n}}}}$$
where $\iota$ is the natural embedding, $\mu(a)=\overline{\frac{a}{1}}\ \forall a\in A$, $\Phi(b)=(\overline{\frac{b}{1}}^{1}, ..., \overline{\frac{b}{1}}^{n})\ \forall b\in B$, and $\Psi(\overline{\frac{a}{s}})=(\overline{\frac{a}{s}}^{1}, ..., \overline{\frac{a}{s}}^{n})\ \forall \frac{a}{s}\in A_{p}$. Let $J:=\Ker(\Phi)$ and $J':=\Ker(\Psi)$ and note that $$p=\iota^{-1}(J)=\mu^{-1}(J').$$
\begin{corollary}\label{glu}
Under the above notation, the following assertions are equivalent:
\begin{enumerate}
\item $A'= A\bowtie^{\iota,\mu}(J,J')$;
\item For any $(\frac{a}{s},b)\in A_{p}\times B:\ a-sb\in\bigcap_{1\leq i\leq n}p_{i} \Rightarrow a-sa_{o}\in p$, for some $a_{o}\in A$.
\end{enumerate}
\end{corollary}
\begin{proof}
By \cite[Lemma 2]{Yan}, $A'$ can be identified with the pullback $\Phi\times_{D}\Psi$. Further, notice that $\Phi\circ\iota=\Psi\circ\mu$; i.e., the above diagram is commutative. Let $\pi: B\times \frac{A_{p}}{pA_{p}}\rightarrow B$ be the canonical projection and let $a\in A$. Then
$$\Psi\big(\overline{\frac{a}{1}}\big)=\big(\overline{\frac{a}{1}}^{1}, ..., \overline{\frac{a}{1}}^{n}\big)=\Phi(a)=\Phi\circ\pi\big(a,\overline{\frac{a}{1}}\big).$$
Hence $\Phi(A)\subseteq\Phi\circ\pi(\Phi\times_{D}\Psi)$. Therefore, by Proposition~\ref{cara}, (1) holds if and only if $\Phi\circ\pi(\Phi\times_{D}\Psi)\subseteq \Phi(A)$ if and only if for any $(\frac{a}{s},b)\in A_{p}\times B$, $\overline{\frac{a}{s}}^{i}=\overline{\frac{b}{1}}^{i}\ \forall i$ forces $\overline{\frac{a}{s}}^{i}=\overline{\frac{a_{o}}{1}}^{i}\ \forall i$, for some $a_{o}\in A$ if and only if (2) holds.
\end{proof}
For example, if $A := \Z$ and $p := 2\Z$, then for any finite $\Z$-module $B$ (e.g., $\Z[i]$)
Condition (2) of Corollary~\ref{glu} always holds since, for any $n\in \Z$ and $s\in\Z \setminus 2\Z$, $n-sn\in 2\Z$.
\section{Basic algebraic properties of bi-amalgamations}
\noindent Throughout, let $f: A\rightarrow B$ and $g: A\rightarrow C$ be two ring homomorphisms and $J,J'$ two ideals of $B$ and $C$, respectively, such that $I_{o}:=f^{-1}(J)=g^{-1}(J')$. Let
$$A\bowtie^{f,g}(J,J'):=\big\{(f(a)+j,g(a)+j') \mid a\in A, (j,j')\in J\times J'\big\}$$
be the bi-amalgamation of $A$ with $(B, C)$ along $(J, J')$ with respect to $(f,g)$.
This section studies basic algebraic properties of bi-amalgamations. Precisely, we investigate necessary and sufficient conditions for a bi-amalgamation to be a Noetherian ring, a domain, or a reduced ring. We will show that the transfer of these notions is made via the special rings $f(A)+J$ and $g(A)+J'$ (which correspond to $B$ and $C$, respectively, in the case when $f$ and $g$ are surjective).
We start with some basic ideal-theoretic properties of bi-amalgamations. For this purpose, notice first that $0\times J'$, $J\times 0$, and $J\times J'$ are particular ideals of $A\bowtie^{f,g}(J,J')$; and if $I$ is an ideal of $A$, then the set $$I\bowtie^{f,g}(J,J'):=\big\{(f(i)+j,g(i)+j')\mid i\in I, (j,j')\in J\times J'\big\}$$
is an ideal of $A\bowtie^{f,g}(J,J')$ containing $J\times J'$.
\begin{proposition}\label{prop cong}
Let $I$ be an ideal of $A$. We have the following canonical isomorphisms:
\begin{enumerate}
\item $\dfrac{A\bowtie^{f,g}(J,J')}{I\bowtie^{f,g}(J,J')}\cong\dfrac{A}{I+I_{o}}$.
\item $\displaystyle\frac{A\bowtie^{f,g}(J,J')}{0\times J'}\cong f(A)+J$ and $\displaystyle\frac{A\bowtie^{f,g}(J,J')}{J\times 0}\cong g(A)+J'$.
\item $\dfrac{A}{I_{o}}\cong\dfrac{A\bowtie^{f,g}(J,J')}{J\times J'}\cong \dfrac{f(A)+J}{J}\cong \dfrac{g(A)+J'}{J'}$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) Consider the mapping
$$\begin{array}{cccc}
\varphi: & A & \rightarrow & \displaystyle\frac{A\bowtie^{f,g}(J,J')}{I\bowtie^{f,g}(J,J')}\\
& a & \longmapsto & \overline{(f(a),g(a))}.
\end{array}$$
Clearly, $\varphi$ is a surjective ring homomorphism and one can check that $\Ker(\varphi)=I+I_{o}$.
(2) If $f(a)+j=0$ for some $a\in A$ and $j\in J$, then $g(a)+j'\in J'$ for any $j'\in J'$. So the kernel of the surjective canonical homomorphism $A\bowtie^{f,g}(J,J')\twoheadrightarrow f(A)+J$ coincides with $0\times J'$. Hence, the first isomorphism holds and the second one follows similarly.
(3) The first isomorphism is a particular case of (1) for $I=0$. Further, if $f(a)+j\in J$ for some $a\in A$ and $j\in J$, then $g(a)+j'\in J'$ for any $j'\in J'$. So the kernel of the canonical surjective homomorphism $$A\bowtie^{f,g}(J,J')\twoheadrightarrow\dfrac{f(A)+J}{J}$$ coincides with $J\times J'$.
\end{proof}
The fact that bi-amalgamations can be represented as pullbacks is an important tool that one can use to investigate the algebraic properties of these constructions. The following results give examples of this use.
\begin{proposition}\label{noeth}
Under the above notation, we have:
\begin{center} $A\bowtie^{f,g}(J,J')$ is Noetherian $\Leftrightarrow$ $f(A)+J$ and $g(A)+J'$ are Noetherian.\end{center}
\end{proposition}
\begin{proof}
In view of Proposition~\ref{prop cong}(2), we only need to prove the reverse implication. By Proposition~\ref{prop1},
$A\bowtie^{f,g}(J,J')=\alpha\times_{\frac{A}{I_{o}}}\beta$ determined by the ring homomorphisms $\alpha: f(A)+J\rightarrow A/I_{o}$, $f(a)+j\mapsto \bar{a}$ and $\beta: g(A)+J'\rightarrow A/I_{o}$, $g(a)+j'\mapsto \bar{a}$. Sine $f(A)+J$ is Noetherian, by \cite[Proposition 4.10]{DFF1}, it suffices to show that $\Ker(\beta)=J'$ is a Noetherian module over $A\bowtie^{f,g}(J,J')$ with the module structure induced by the surjective canonical homomorphism $A\bowtie^{f,g}(J,J')\twoheadrightarrow g(A)+J'$. But, under this structure, $A\bowtie^{f,g}(J,J')$-submodules of $J'$ correspond to subideals of $J'$ in the Noetherian ring $g(A) + J'$. This leads to the conclusion.
\end{proof}
In view of Example~\ref{aa}, Proposition~\ref{noeth} recovers the special case of amalgamated algebras, as recorded in the next corollary.
\begin{corollary}[{\cite[Proposition 5.6]{DFF1}}]
Under the above notation, we have:
\begin{center} $A\bowtie^{f}J$ is Noetherian $\Leftrightarrow$ $A$ and $f(A)+J$ are Noetherian.\end{center}
\qed\end{corollary}
As an illustrative example for Proposition~\ref{noeth} (of an original Noetherian ring which arises as a bi-amalgamation) is provided in Example~\ref{s4:exa1}.
Recall that the prime spectrum of a ring $R$ is said to be Noetherian if $R$ satisfies the ascending chain condition on radical ideals (or, equivalently, every prime ideal of $R$ is the radical of a finitely generated ideal) \cite{OP}. Let $\Spec(R)$ denote the prime spectrum of a ring $R$.
\begin{proposition}
Under the above notation, we have:
\begin{center} $\Spec\big(A\bowtie^{f,g}(J,J')\big)$ is Noetherian $\Leftrightarrow$ $\Spec\big(f(A)+J\big)$ and $\Spec\big(g(A)+J'\big)$ are Noetherian.\end{center}
\end{proposition}
\begin{proof}
$A\bowtie^{f,g}(J,J')=\alpha\times_{\frac{A}{I_{o}}}\beta$ via the homomorphisms $\alpha: f(A)+J\rightarrow A/I_{o}$, $f(a)+j\mapsto \bar{a}$ and $\beta: g(A)+J'\rightarrow A/I_{o}$, $g(a)+j'\mapsto \bar{a}$. So, by \cite[Corollary 1.6]{F}, the prime spectra of $A\bowtie^{f,g}(J,J')$ and $A/I_{o}$ are Noetherian if and only if so are the prime spectra of $f(A)+J$ and $g(A)+J'$. But, by Proposition~\ref{prop cong}(3) if the prime spectrum of $A\bowtie^{f,g}(J,J')$ is Noetherian, then so is the spectrum of $A/I_{o}$ since this notion is stable under homomorphic image. This leads to the conclusion.
\end{proof}
The next result characterizes bi-amalgamations without zero divisors.
\begin{proposition}\label{domain}
Under the above notation, the following assertions are equivalent:
\begin{enumerate}
\item $A\bowtie^{f,g}(J,J')$ is a domain;
\item ``$J=0$ and $g(A)+J'$ is a domain" or ``$J'=0$ and $f(A)+J$ is a domain."
\end{enumerate}
\end{proposition}
\begin{proof}
Assume that $A\bowtie^{f,g}(J,J')$ is a domain. If $J\neq 0$ and $J'\neq 0$, then for nonzero elements $j\in J$ and $j'\in J'$ we have $(0,j')(j,0)=(0,0)$. Therefore, one of $J$ and $J'$ must be null; in such case, $A\bowtie^{f,g}(J,J')$ collapse (up to an isomorphism) to $f(A)+J$ or $f(A)+J$ by Proposition~\ref{prop cong}(2). This leads to the conclusion.
\end{proof}
In view of Example~\ref{aa}, Proposition~\ref{domain} recovers the special case of amalgamated algebras, as recorded in the next corollary.
\begin{corollary}[{\cite[Proposition 5.2]{DFF1}}]
Under the above notation, assume $J\not=0$. Then:
\begin{center} $A\bowtie^{f}J$ is a domain $\Leftrightarrow$ $f^{-1}(J)=0$ and $f(A)+J$ is a domain.\end{center}
\qed\end{corollary}
The next result characterizes bi-amalgamations without nilpotent elements.
\begin{proposition}\label{reduced}
Under the above notation, consider the following conditions:
\begin{enumerate}
\item[(a)] $f(A)+J$ is reduced and $J'\cap \nil(C)=0$,
\item[(b)] $g(A)+J'$ is reduced and $J\cap \nil(B)=0$,
\item[(c)] $A\bowtie^{f,g}(J,J')$ is reduced,
\item[(d)] $J\cap \nil(B)=0$ and $J'\cap \nil(C)=0$.
\end{enumerate}
Then:
\begin{enumerate}
\item $(a)\ \text{or}\ (b) \Rightarrow (c) \Rightarrow (d)$.
\item If $I_{o}$ is radical, then the four conditions are equivalent.
\item If $f$ is surjective and $\Ker(f)\subseteq \Ker(g)$, then:
\begin{center} $A\bowtie^{f,g}(J,J')$ is reduced $\Leftrightarrow$ $B$ is reduced and $J'\cap \nil(C)=0$.\end{center}
\end{enumerate}
\end{proposition}
\begin{proof}
(1) Let $(f(a)+j,g(a)+j')\in \nil\big(A\bowtie^{f,g}(J,J')\big)$. Then $f(a)+j\in \nil(f(A)+J)=0$. Hence, $a\in I_{o}$. Thus, $g(a)+j'\in J'\cap \nil(C)=0$. Consequently, $\nil\big(A\bowtie^{f,g}(J,J')\big)=0$. This proves $(a)\Rightarrow (c)$. Likewise for $(b)\Rightarrow (c)$.
Let $j\in \nil(B)\cap J$. Therefore, there is a positive integer $n$ such that $0=(j^n,0)=(j,0)^n$ in $A\bowtie^{f,g}(J,J')$. It follows that $j=0$ and hence $\nil(B)\cap J=0$. Similarly, $\nil(C)\cap J'=0$. This proves $(c)\Rightarrow (d)$.
(2) Next, assume that $I_{o}$ is radical, $J\cap \nil(B)=0$, and $J'\cap \nil(C)=0$. Let $f(a)+j\in \nil(f(A)+J)$. Then, there is a positive integer $n$ such that $(f(a)+j)^n=0$. Hence, $f(a)^n\in J$ and thus $a^n\in I_{o}$; that is, $a\in I_{o}$. So, $f(a)+j\in J\cap \nil(B)=0$, as desired. This proves $(d)\Rightarrow (a)$. Likewise for $(d)\Rightarrow (b)$.
(3) In view of (1), it suffices to observe that $f(a^{n})=0$, for some positive integer, forces $(f(a),g(a))^{n}=0$, yielding $f(a)=0$.
\end{proof}
\begin{remark}\label{Bnil}
If $f(A)+J$ and $g(A)+J'$ are both reduced, then $A\bowtie^{f,g}(J,J')$ is reduced by Proposition~\ref{reduced}. The converse is not true in general. A counter-example (for the special case of amalgamated algebras) is given in \cite[Remark 5.5 (3)]{DFF1}.
\end{remark}
In view of Example~\ref{aa}, Proposition~\ref{reduced} recovers the special case of amalgamated algebras, as recorded in the next corollary.
\begin{corollary}[{\cite[Proposition 5.4]{DFF1}}]
Under the above notation, we have:
\begin{center} $A\bowtie^{f}J$ is reduced $\Leftrightarrow$ $A$ is reduced and $J\cap \nil(B)=0$.\end{center}
\qed\end{corollary}
As an illustrative example for Propositions \ref{noeth} \& \ref{domain} \& \ref{reduced}, we provide an original reduced Noetherian ring with zero divisors which arises as a bi-amalgamation.
\begin{example}\label{s4:exa1}
Consider the surjective ring homomorphism $f: \Z[X]\twoheadrightarrow \Z[\sqrt{2}]$, $p(X)\mapsto p(\sqrt{2})$ and the principal ideal $J:=(\sqrt{2})$ of $\Z[\sqrt{2}]$. Let $p\in \Z[X]$ and write it as $p=(X^2-2)q(X)+aX+b$ for some $a,b\in\Z$ and $q\in \Z[X]$. Then, one can verify that $p(\sqrt{2})\in J$ if and only if $b\in 2\Z$. That is,
$$I_{o}:=f^{-1}(J)=\big\{p\in \Z[X] \mid p(0)\in 2\Z\big\}.$$
Now, consider the ring homomorphism $\alpha: \Z[\sqrt{2}]\twoheadrightarrow \dfrac{\Z[X]}{I_{o}}$, $a+b\sqrt{2}\mapsto \bar{a}$. It follows, by Proposition~\ref{prop1} and Propositions \ref{noeth} \& \ref{domain} \& \ref{reduced}, that
$$\Z[X]\bowtie^{f,f}(J,J)=\alpha\times_{\frac{\Z[X]}{I_{o}}}\alpha=\big\{(a+b\sqrt{2}, c+d\sqrt{2})\mid a,b,c,d\in \Z,\ a-c\in 2\Z\big\}$$
is a reduced Noetherian ring that is not a domain (since $\Z[\sqrt{2}]$ is a Noetherian domain and $J\not=0$).
\end{example}
\section{The prime ideal structure of bi-amalgamations}
\noindent Throughout, let $f: A\rightarrow B$ and $g: A\rightarrow C$ be two ring homomorphisms and $J,J'$ two ideals of $B$ and $C$, respectively, such that $I_{o}:=f^{-1}(J)=g^{-1}(J')$. Let
$$A\bowtie^{f,g}(J,J'):=\big\{(f(a)+j,g(a)+j') \mid a\in A, (j,j')\in J\times J'\big\}$$
be the bi-amalgamation of $A$ with $(B, C)$ along $(J, J')$ with respect to $(f,g)$.
This section investigates the prime ideal structure of bi-amalgamations and their localizations at prime ideals. We also establish necessary and sufficient conditions for a bi-amalgamation to be local.
Next, we describe the prime (and maximal) ideals of bi-amalgamations. To this purpose, let's adopt the following notation:
$$\begin{array}{lcl}
Y &:= &\Spec(f(A)+J)\\
Y' &:= &\Spec(g(A)+J')
\end{array}$$
and, for $L\in Y$ and $L'\in Y'$, consider the prime ideals of $A\bowtie^{f,g}(J,J')$ given by:
$$\begin{array}{rcl}
\bar{L} &:= &\big(L\times (g(A)+J')\big)\cap\big(A\bowtie^{f,g}(J,J')\big)\\
&= &\big\{(f(a)+j,g(a)+j')\mid a\in A, (j,j')\in J\times J', f(a)+j\in L\big\},\\
\bar{L'} &:= &\big((f(A)+J)\times L'\big)\cap\big(A\bowtie^{f,g}(J,J')\big)\\
&= &\big\{(f(a)+j,g(a)+j')\mid a\in A, (j,j')\in J\times J', g(a)+j'\in L'\big\}.
\end{array}$$
The next two lemmas are needed for the proof of Proposition~\ref{spec}. Recall that if $I$ is an ideal of $A$, then $$I\bowtie^{f,g}(J,J'):=\big\{(f(i)+j,g(i)+j')\mid i\in I, (j,j')\in J\times J'\big\}$$
is an ideal of $A\bowtie^{f,g}(J,J')$. As an immediate consequence of Proposition~\ref{prop cong}(1), we have the following lemma.
\begin{lemma}\label{s5:lem1}
Let $I$ be an ideal of $A$. Then, $I\bowtie^{f,g}(J,J')$ is a prime (resp., maximal) ideal of $A\bowtie^{f,g}(J,J')$ if and only if $I+I_{o}$ is a prime (resp., maximal) ideal of $A$.
\qed\end{lemma}
An element of $Y$ (resp., $Y'$) containing $J$ (resp., $J'$) has a special form, as shown by the next lemma.
\begin{lemma}\label{s5:lem2}
Let $L\in Y$ (resp., $Y'$) containing $J$ (resp., $J'$). Then:
$$\bar{L}=f^{-1}(L)\bowtie^{f,g}(J,J')\ \big(\text{resp.,}\ =g^{-1}(L)\bowtie^{f,g}(J,J')\big).$$
\end{lemma}
\begin{proof}
Let $L\in Y$ containing $J$. Notice first that $f^{-1}(L)$ is a prime ideal of $A$ containing $I_{o}:=f^{-1}(J)$ so that $f^{-1}(L)\bowtie^{f,g}(J,J')$ is a prime ideal of $A\bowtie^{f,g}(J,J')$ by Lemma~\ref{s5:lem1}. Moreover, for any $a\in A$ and $j\in J$, one can easily see that $f(a)+j\in L$ if and only if $a\in f^{-1}(L)$. Thus, $\overline{L}=f^{-1}(L)\bowtie^{f,g}(J,J')$. Likewise for $L\in Y'$.
\end{proof}
\begin{proposition}\label{spec}
Under the above notation, let $P$ be a prime ideal of $A\bowtie^{f,g}(J,J')$. Then
\begin{enumerate}
\item $J\times J'\subseteq P$ $\Leftrightarrow$ $\exists!\ p\supseteq I_{o}$ in $\Spec(A)$ such that $P=p\bowtie^{f,g}(J,J').$\\
In this case, $\exists\ L\supseteq J$ in $Y$ and $\exists\ L'\supseteq J'$ in $Y'$ such that $P=\bar{L}=\bar{L'}.$
\vspace{.2cm}
\item $J\times J'\nsubseteq P$ $\Leftrightarrow$ $\exists!\ L\in Y$ (or $Y'$) such that $J\nsubseteq L$ (or $J'\nsubseteq L$) and $P=\bar{L}$.\\
In this case, $(A\bowtie^{f,g}(J,J'))_{P}\cong (f(A)+J)_L\ \big(\text{ or }\ (A\bowtie^{f,g}(J,J'))_{P}\cong (g(A)+J')_L\big).$
\vspace{.2cm}
\noindent Consequently, we have
$$\Spec\big(A\bowtie^{f,g}(J,J')\big)=\big\{\bar{L}\mid L\in \Spec\big(f(A)+J\big)\cup \Spec\big(g(A)+J'\big)\big\}.$$
\end{enumerate}
\end{proposition}
\begin{proof}
(1) We only need to prove ($\Rightarrow$). Assume $J\times J'\subseteq P$ and consider the ideal $p$ of $A$ given by
$$p:=\big\{a\in A\mid \exists\ (j,j')\in J\times J'\ \text{such that}\ (f(a)+j,g(a)+j')\in P\big\}.$$
Clearly, the fact $J\times J'\subseteq P$ forces $I_{o}\subseteq p$. Moreover, we have $P\subseteq p\bowtie^{f,g}(J,J')$. For the reverse inclusion, let $a\in p$. So there exists $(j_{1},j'_{1})\in J\times J'$ such that $(f(a)+j_{1},g(a)+j'_{1})\in P$. Hence, for every $(j,j')\in J\times J'$, we obtain $$(f(a)+j,g(a)+j')=(f(a)+j_{1},g(a)+j'_{1})+(j-j_{1},j'-j'_{1})\in P$$
since $J\times J'\subseteq P$. It follows that $$P= p\bowtie^{f,g}(J,J').$$
By Lemma~\ref{s5:lem1}, $p$ is a prime ideal of $A$. By Proposition~\ref{prop cong}(1), $p$ must be unique since it contains $I_{o}$.
Next, let $L:=f(p)+J$. One can verify that $L$ is a prime ideal of $f(A)+J$ with $p\subseteq f^{-1}(L)$. Now, let $a\in f^{-1}(L)$. Then $f(a)=f(x)+j$ for some $x\in p$ and $j\in J$. Hence $(a-x)\in I_{o}\subseteq p$, whence $a\in p$. So, $$f^{-1}(L)= p.$$ It follows, via Lemma~\ref{s5:lem2}, that $$\bar{L}=f^{-1}(L)\bowtie^{f,g}(J,J')=p\bowtie^{f,g}(J,J')=P.$$
Note that for $L':=g(p)+J'$, the same arguments lead to
$$P=\bar{L}=\bar{L'}.$$
(2) We only need to prove ($\Rightarrow$). Assume $J\times J'\nsubseteq P$. By Proposition \ref{conductor} and \cite[Lemma 1.1.4(3)]{FHP}, there is a unique prime $Q$ of $(f(A)+J)\times (g(A)+J')$ such that
$$P=Q\cap A\bowtie^{f,g}(J,J')\ \text{ with}\ \big((f(A)+J)\times (g(A)+J')\big)_{Q}=\big(A\bowtie^{f,g}(J,J')\big)_{P}.$$
Then either $Q=L\times (g(A)+J')$ for some prime ideal $L\in Y$ or $Q=(f(A)+J)\times L'$ for some prime ideal $L'\in Y'$. That is,
$$P=\bar{L}\ \text{ or }\ P=\bar{L'}.$$
Accordingly, we'll have
$$(A\bowtie^{f,g}(J,J'))_{P}\cong (f(A)+J)_L\ \text{ or }\ (A\bowtie^{f,g}(J,J'))_{P}\cong (g(A)+J')_{L'}$$
completing the proof of the proposition.
\end{proof}
Next, as an application of Proposition~\ref{spec}, we establish necessary and sufficient conditions for a bi-amalgamation to be local. Notice at this point that, in the presence of the equality $f^{-1}(J)=g^{-1}(J')$, $J\not=B$ if and only if $J'\not=C$.
\begin{proposition}\label{local}
Under the above notation, we have
\begin{enumerate}
\item $A\bowtie^{f,g}(J,J')$ is local $\Leftrightarrow$ $J\not= B$ and $f(A)+J$ \& $g(A)+J'$ are local.\\
Moreover, the maximal ideal of $A\bowtie^{f,g}(J,J')$ has the form $\m\bowtie^{f,g}(J,J')$, where $\m$ is the unique maximal ideal of $A$ containing $I_{o}$.
\bigskip
\item Suppose that $A$ is local. Then: $$A\bowtie^{f,g}(J,J')\ \text{is local}\ \Leftrightarrow J\times J'\subseteq Jac(B\times C).$$
\end{enumerate}
\end{proposition}
\begin{proof}
(1) Notice first that if $J=B$, (hence $J'=C$ and) then $A\bowtie^{f,g}(J,J')= B\times C$ which is never local. Assume that $A\bowtie^{f,g}(J,J')$ is local. Then $J\not= B$ and, by Proposition~\ref{prop cong}(2), both $f(A)+J$ and $g(A)+J'$ are local. Moreover, $I_{o}\not=A$. Therefore, there is $\m\supseteq I_{o}$ maximal in $A$. By Lemma~\ref{s5:lem1}, $\m\bowtie^{f,g}(J,J')$ is the maximal ideal of $A\bowtie^{f,g}(J,J')$. Then, the uniqueness of $\m$ is ensured by Proposition~\ref{prop cong}(1).
Next assume that $J\not= B$ and $f(A)+J$ \& $g(A)+J'$ are local. Let $M$ be a maximal ideal of $A\bowtie^{f,g}(J,J')$. We claim that $J\times J'\subseteq M$. Deny. Then, by Proposition~\ref{spec}(2), there is a unique prime $L$, say, of $f(A)+J$ such that $M=\bar{L}$ and $J\nsubseteq L$. Further, the uniqueness of $L$ and maximality of $M$ force $L$ to be a (in fact, the) maximal ideal of $f(A)+J$. It follows that $J\subseteq L$ (since $J\not= B$), the desired contradiction. Therefore, $$J\times J'\subseteq M.$$ So, by Proposition \ref{spec}(1), there is a (unique) prime ideal $\m$ of $A$ containing $I_{o}$ such that
$$M=\m\bowtie^{f,g}(J,J').$$
By Lemma~\ref{s5:lem1}, $\m$ is maximal in $A$. By Proposition~\ref{prop cong}(3), $\dfrac{A}{I_{o}}\cong\dfrac{f(A)+J}{J}$ is local with maximal ideal $\dfrac{\m}{I_{o}}$. This forces $M$ to be the unique maximal ideal of $A\bowtie^{f,g}(J,J')$.
(2) ($\Rightarrow$) In this direction we don't need the assumption ``$A$ is local." Assume that $A\bowtie^{f,g}(J,J')$ is local. By (1), necessarily, its maximal ideal contains $J\times J'$. Let $(j,j')\in J\times J'$ and $(b,c)\in B\times C$. Then, $(b,c)(j,j')\in J\times J'$. Thus, $(1,1)-(b,c)(j,j')$ is invertible in $A\bowtie^{f,g}(J,J')$ (and so in $B\times C$). Hence, $J\times J'\subseteq Jac(B\times C)$.
($\Leftarrow$) Assume that $A$ is local and $J\times J'\subseteq Jac(B\times C)$. Let $a$ be a unit of $A$. We claim that $(f(a)+j,g(a)+j')$ is a unit of $A\bowtie^{f,g}(J,J')$ for every $(j,j')\in J\times J'$. Indeed, $f(a)+j$ and $g(a)+j'$ are, respectively, units in $B$ and $C$ since $J\times J'\subseteq Jac(B\times C)$. Thus, there exist $u\in B$ and $v\in C$ such that $(f(a)+j)u=1$ and $(g(a)+j')v=1$. Hence,
$$(f(a)+j,g(a)+j')(f(a^{-1})-uf(a^{-1})j,g(a^{-1})-vg(a^{-1})j')=(1,1);$$
that is, $(f(a)+j,g(a)+j')$ is a unit of $A\bowtie^{f,g}(J,J')$. Next, let $(f(a)+j_1,g(a)+j'_1)$ be a nonunit element of $A\bowtie^{f,g}(J,J')$. So, $a$ is a nonunit of $A$. Moreover, for any $(f(b)+j_2,g(b)+j_2')\in A\bowtie^{f,g}(J,J')$, we have $$(1,1)-(f(b)+j_2,g(b)+j'_2)(f(a)+j_1,g(a)+j'_1)=(f(1-ba)+j_3,g(1-ba)+j'_3)$$
for some $j_{3}\in J$ and $j'_{3}\in J'$. Further, $1-ba$ is a unit of $A$ since $A$ is local. Hence, $(1,1)-(f(b)+j_2,g(b)+j'_2)(f(a)+j_1,g(a)+j'_1)$ is a unit of $A\bowtie^{f,g}(J,J')$. This proves that $A\bowtie^{f,g}(J,J')$ is local.
\end{proof}
In view of Example~\ref{aa}, Proposition~\ref{local} recovers the special case of amalgamated algebras and amalgamated duplications, as recorded in the next corollaries.
\begin{corollary}
Under the above notation, the following assertions are equivalent:
\begin{enumerate}
\item $A\bowtie^fJ$ is local;
\item $J\not=B$ and $A$ \& $f(A)+J$ are local;
\item $A$ is local and $J\subseteq Jac(B)$.
\end{enumerate}
\qed\end{corollary}
\begin{corollary}[{\cite[Corollary 6]{D} \& \cite[Theorem 3.5(1.e)]{DF1} \& \cite[Proposition 2.2]{DF2}}]
Let $A$ be a ring and $I$ a proper ideal of $A$. Then, $A\bowtie I$ is local if and only if $A$ is local.
\qed\end{corollary}
Next, we describe the localizations of $A\bowtie^{f,g}(J,J')$ at its prime ideals which contain $J\times J'$. Recall that, given a ring $R$, an ideal $I$ of $R$, and $S$ a multiplicatively closed subset of $R$ with $S\cap I =\emptyset$, then $S+I$ is a multiplicatively closed subset of $R$.
\begin{proposition}\label{localization}
Let $p$ be a prime ideal of $A$ containing $I_{o}$ and let $P:=p\bowtie^{f,g}(J,J')$. Consider the multiplicative subsets $S:=f(A-p)+J$ of $B$ and $S':=g(A-p)+J'$ of $C$. Let $f_p: A_p\rightarrow B_S$ and $g_p: A_p\rightarrow C_{S'}$ be the ring homomorphisms induced by $f$ and $g$. Then: $$f_p^{-1}(J_S)=g_p^{-1}(J'_{S'})=(I_{o})_p$$ and $$\big(A\bowtie^{f,g}(J,J')\big)_{P}\cong A_p\bowtie^{f_p,g_p}(J_S,J'_{S'}).$$
\end{proposition}
\begin{proof}
It is easy to show that $f_p^{-1}(J_S)=g_p^{-1}(J'_{S'})=(I_{o})_p$. Moreover, by Proposition \ref{prop1}, $A_p\bowtie^{f_p,g_p}(J_S,J'_{S'})$ is the fiber product of $\alpha: f_p(A_p)+J_S\rightarrow A_p/(I_{o})_p$ and $\beta: g_p(A_p)+J'_{S'}\rightarrow A_p/(I_{o})_p$. On the other hand, $\pi_B(A\bowtie^{f,g}(J,J')-P)=S$ and $\pi_C(A\bowtie^{f,g}(J,J')-P)=S'$. Then, the fact that $(A\bowtie^{f,g}(J,J'))_{P}$ is isomorphic to $A_p\bowtie^{f_p,g_p}(J_S,J'_{S'})$ follows from \cite[Proposition 1.9]{F}.
\end{proof}
\begin{remark}
If $P$ is a prime ideal of $A\bowtie^{f,g}(J,J')$ which contains $J\times J'$, then by Proposition \ref{spec},
there exists a (unique) prime ideal $p$ (which contains $I_{o}$) such that $P=p\bowtie^{f,g}(J,J')$. Thus, by Proposition \ref{conductor} and Proposition \ref{localization}, one can obtain
a conductor square of the form:
$$\xymatrix@R=2cm @C=2cm{ (A\bowtie^{f,g}(J,J'))_{P} \ar@{>>}[d]^{\mu_2}\ar[r]^{\iota_2} & \ar@{>>}[d]^{\mu_1}(f_p(A_p)+J_S)\times (g_p(A_p)+J'_{S'})\\
\displaystyle\frac{A_{p}}{I_{o}A_{p}}\quad \quad \ar[r]^{\iota_1}& \quad \quad \displaystyle \frac{A_{p}}{I_{o}A_{p}}\times \frac{A_{p}}{I_{o}A_{p}}}$$
\end{remark}
|
\section{Introduction}
Emerged in 2009 \cite{yuan2013real}, Real-Time Bidding (RTB) has become an important new paradigm in display advertising \cite{muthukrishnan2009ad,chakraborty2010selective,google2011arrival}.
For example, eMarketer estimates a 73\% spending growth on RTB in United States during 2013, which accounts for 19\% of the total spending in display advertising \cite{emarketer2013rtb}.
Different from the conventional negotiation or pre-setting a fixed bid for each campaign or keyword, RTB enables the advertisers to give a bid for every individual impression.
A concise interaction process between the main components of RTB ecosystem is shown in Figure~\ref{fig:interaction}.
Each ad placement will trigger an auction when the user visits an ad-supported site (e.g., web page, streaming videos and mobile apps). Bid requests will be sent via the ad exchange to the advertisers' buying systems, usually referred to as Demand-Side Platforms (DSPs).
Upon receiving a bid request, a DSP will calculate a bid as the response after holding an internal auction among all of its qualifying campaigns.
An auction will be held at each intermediary (ad networks, ad exchanges, etc.) and finally in the publishers' system.
Finally, the winner's ad will be shown to the visitor along with the regular content of the website.
It is commonly known that a long time page-loading would greatly reduce users' satisfactory \cite{muthukrishnan2009ad}, thus, DSPs are usually required to return a bid in a very short time frame (e.g. 100 ms). More detailed introductions to RTB could be found in \cite{Wang:RTB:Tutorial,yuan2013real}.
Algorithms employed by DSPs are expected to contribute a much higher return-on-investment (ROI) comparing with the traditional channels. It is crucial that such algorithms can quickly decide whether and how much to bid for a specific impression, given the contextual and behaviour data (usually referred to as user segments). This is apparently also an engineering challenge considering the billion-level bid requests that a DSP could normally see in a day.
Despite its popularity, the majority research activities on RTB have been limited in advertising technology companies \cite{muthukrishnan2009ad,lee2012estimating,perlich2012bid} so far. It is nearly impossible for researchers from academia to access the sensitive thus highly protected data.
Fortunately, a three-season global competition of RTB algorithms was held by iPinYou\footnote{\url{http://www.ipinyou.com.cn}} in 2013. As will be discussed in Section \ref{sec:task}, the competition task focuses on the bidding strategies from the DSP's perspective: it aims to maximise the campaign's Key-Performance-Indicator (KPI) with the budget and lifetime constraint by developing the bidding strategy. We refer such task as \emph{DSP Bid Optimisation} problem. In March 2014, the dataset used in the three seasons of the competition (about 35 GB) was released for the purpose of research.
To the best of our knowledge, this is the first large-scale real-world RTB dataset. We believe it will stimulate the interest of RTB research and development of DSP bidding algorithms in the whole data science research community, and further speed up the growth of RTB display advertising ecosystem. The dataset can be directly downloaded from our website of computational advertising research\footnote{\url{http://data.computational-advertising.org}}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/AdExchange-DSP}
\caption{A brief illustration of the interactions between user, ad exchange and DSP.}
\label{fig:interaction}
\end{figure}
In this paper, we first report a detailed statistical analysis of this dataset. Then, we formally present the research problem of DSP bid optimisation and its simple yet comprehensive evaluation protocol. Finally, we show the experimental results of some benchmark bidding strategies as well as the click-through rate (CTR) estimation models.
\section{The iPinYou RTB Dataset}\label{sec:dataset}
\subsection{iPinYou Demand-Side Platform}
iPinYou Information Technologies Co., Ltd (iPinYou) was founded in 2008 and is currently the largest DSP in China. iPinYou is headquartered in Beijing and has offices in Shanghai, Guangzhou and Silicon Valley. iPinYou has built world class RTB technology and algorithm, proprietary cloud computing platform and patented audience profiling technology. It has served over 1000 brands in IT, financial service, auto, consumer packaged goods, travel, electric commerce, gaming and more. It has also significantly improved the advertising effectiveness and fostered extensive partnerships with domestic mainstream media and private exchanges. It is established as a leading provider of audience based programmatic advertising technology.
\subsection{Data Format}
There are four different types of logs in the iPinYou dataset: bids, impressions, clicks, and conversions. The logs are organised on a row-per-record basis.
The feature description and example of each column of the ad log data are presented in Table~\ref{tab:data-format}. Generally, each record contains three kinds of information: (i) The auction and ad features (all columns except 3, 20 and 21). These features are sent to the bidding engine to make a bid response. (ii) The auction winning price (column 21), i.e. the highest bid from the competitors. If the bidding engine responses a bid higher than the auction winning price, the DSP will win this auction and get the ad impression. (iii) The user feedback (click and conversion) on the ad impression (column 3). If the DSP wins the auction, the user feedback on this ad impression can be checked to update the DSP performance.
Note that all numbers related to money (e.g., bid price, paying price and floor price) use the currency of RMB and the unit of Chinese fen $\times 1000$, corresponding to the commonly adopted cost-per-mille (CPM) pricing model. However, in our analysis the calculated numbers (e.g., cost, average CPM, and effective cost-per-click) are not multiplied by 1000.
\begin{table}[t]
\center
\caption{The log data format. Columns with $*$ are hashed or modified before releasing; columns with $\dag$ are only available in impression/click/conversion logs but not in bid logs. }
\label{tab:data-format}
\begin{tabular}{rll}
\small
Col \# & Description & Example \\ \hline
\\ [-2.0ex]
$^*$1 & Bid ID & 015300008...3f5a4f5121\\
2 & Timestamp & 20130218001203638\\
$^\dag$3 & Log type & 1\\
$^*$4 & iPinYou ID & 35605620124122340227135\\
5 & User-Agent & Mozilla/5.0 (compatible; \textbackslash \\
& & MSIE 9.0; Windows NT \textbackslash \\
& & 6.1; WOW64; Trident/5.0)\\
$^*$6 & IP & 118.81.189.*\\
7 & Region & 15\\
8 & City & 16\\
$^*$9 & Ad exchange & 2\\
$^*$10 & Domain & e80f4ec7...c01cd1a049\\
$^*$11 & URL & hz55b00000...3d6f275121\\
12 & Anonymous URL ID & Null\\
13 & Ad slot ID & 2147689\_8764813\\
14 & Ad slot width & 300\\
15 & Ad slot height & 250\\
16 & Ad slot visibility & SecondView\\
17 & Ad slot format & Fixed\\
$^*$18 & Ad slot floor price & 0 \\
19 & Creative ID & e39e178ffd...1ee56bcd\\
$^*$20 & Bidding price & 753\\
$^{*\dag}$21 & Paying price & 15\\
$^{*\dag}$22 & Key page URL & a8be178ffd...1ee56bcd\\
$^{*}$23 & Advertiser ID & 2345\\
$^{*}$24 & User Tags & 123,5678,3456\\
\end{tabular}
\end{table}
Along with Table~\ref{tab:data-format} we want to give more detailed description for some of the columns here: \\
\textsc{(c01)} The bid ID serves as the unique identifier of all event logs and could be used to join bids, impressions, clicks, and conversions together. \\
\textsc{(c02)} The column uses the format of yyyyMMddHHmmssSSS\footnote{\url{http://docs.oracle.com/javase/8/docs/api/java/time/format/DateTimeFormatter.html#patterns}}.\\
\textsc{(c03)} The possible values include: 1 (impression), 2 (click), and 3 (conversion). \\
\textsc{(c04)} The internal user ID set by iPinYou. \\
\textsc{(c05)} The column describes the device, operation system, and browser of the user. \\
\textsc{(c10)} The domain of the hosting webpage of the ad slot. The values were hashed. \\
\textsc{(c11)} The URL of the hosting webpage of the ad slot. The values were hashed. \\
\textsc{(c12)} When URL is not directly available to the DSP (e.g. masked by ad exchanges) this column will be used. The values are provided by ad exchanges. For one record, either URL or Anonymous URL ID is meaningful. \\
\textsc{(c16)} The column describes if the ad slot is above the fold (``FirstView'') or not (``SecondView'' to ``TenthView''), or unknown (``Na''). \\
\textsc{(c17)} Possible values include ``Fixed'' (fixed size and position), ``Pop'' (the pop-up window), ``Background", ``Float", and ``Na" which presents unknown cases. \\
\textsc{(c18)} Floor (or reserve) price of the ad slot. No bid lower than the floor price could win auctions. A linear scale normalisation was applied to this column. \\
\textsc{(c20)} The bid price from iPinYou for this bid request. \\
\textsc{(c21)} The paying price is the highest bid from competitors, also called market price and auction winning price. If this bid price is higher than the auction winning price, then this record will occur in impression log. \\
\textsc{(c24)} User tags (segments) in iPinYou's proprietary audience database. Only a part of the user tags are released in this dataset.
\subsection{Basic Statistics}\label{sec:basic-stats}
The advertisers\footnote{Every advertiser of this dataset has only one campaign. Thus, these two terms are equivalent in this scenario.} and their industrial categories are summarised in Table~\ref{tab:adv-fields}. Note that in the 1st season no advertiser ID was given.
The diversity of advertisers makes the dataset more interesting.
As we show later in the paper, ads from different fields have greatly different user response behaviour.
\begin{table}[t]
\center
\caption{Advertiser Fields}\label{tab:adv-fields}
\vspace{0px}
\begin{tabular}{c c l}
Advertiser ID & Season & Industrial Category\\ \hline
\\[-2.0ex]
1458 & 2 & Chinese vertical e-commerce\\
2259 & 3 & Milk powder\\
2261 & 3 & Telecom\\
2821 & 3 & Footwear\\
2997 & 3 & Mobile e-commerce app install\\
3358 & 2 & Software\\
3386 & 2 & International e-commerce\\
3427 & 2 & Oil\\
3476 & 2 & Tire
\end{tabular}
\end{table}
\begin{table*}[t]
\center
\caption{Dataset statistics}\label{tab:basic-stats}
\vspace{5px}
\small
\begin{tabular}{rrrrrrrrrrrr}
\multicolumn{12}{c}{\textbf{Training Data}} \vspace{5px} \\
Adv. & Period~ & Bids & Imps & Clicks & Convs & Cost & Win Ratio & CTR & CVR & CPM & eCPC\\ \hline
\\[-2.0ex]
1458 & 6-12 Jun. & 14,701,496 & 3,083,056 & 2,454 & 1 & 212,400 & 20.97\% & 0.080\% & 0.041\% & 68.89 & 86.55\\
2259 & 19-22 Oct. & 2,987,731 & 835,556 & 280 & 89 & 77,754 & 27.97\% & 0.034\% & 31.786\% & 93.06 & 277.70\\
2261 & 24-27 Oct. & 2,159,708 & 687,617 & 207 & 0 & 61,610 & 31.84\% & 0.030\% & 0.000\% & 89.60 & 297.64\\
2821 & 21-23 Oct. & 5,292,053 & 1,322,561 & 843 & 450 & 118,082 & 24.99\% & 0.064\% & 53.381\% & 89.28 & 140.07\\
2997 & 23-26 Oct. & 1,017,927 & 312,437 & 1,386 & 0 & 19,689 & 30.69\% & 0.444\% & 0.000\% & 63.02 & 14.21\\
3358 & 6-12 Jun. & 3,751,016 & 1,742,104 & 1,358 & 369 & 160,943 & 46.44\% & 0.078\% & 27.172\% & 92.38 & 118.51\\
3386 & 6-12 Jun. & 14,091,931 & 2,847,802 & 2,076 & 0 & 219,066 & 20.21\% & 0.073\% & 0.000\% & 76.92 & 105.52\\
3427 & 6-12 Jun. & 14,032,619 & 2,593,765 & 1,926 & 0 & 210,239 & 18.48\% & 0.074\% & 0.000\% & 81.06 & 109.16\\
3476 & 6-12 Jun. & 6,712,268 & 1,970,360 & 1,027 & 26 & 156,088 & 29.35\% & 0.052\% & 2.532\% & 79.22 & 151.98\\
Total & - & 64,746,749 & 15,395,258 & 11,557 & 935 & 1,235,875 & 23.78\% & 0.075\% & 8.090\% & 80.28 & 106.94\\
\end{tabular}
\begin{tabular}{rrrrrrrrrrr}
\multicolumn{11}{c}{}\\
\multicolumn{11}{c}{\textbf{Test Data}} \vspace{5px} \\
Adv. & Period~ & Imps & Clicks & Convs & Cost & CTR & CVR & CPM & eCPC & $N$ \\ \hline
\\[-2.0ex]
1458 & 13-15 Jun. & 614,638 & 543 & 0 & 45,216 & 0.088\% & 0.000\% & 73.57 & 83.27 & 0 \\
2259 & 22-25 Oct. & 417,197 & 131 & 32 & 43,497 & 0.031\% & 24.427\% & 104.26 & 332.04 & 1 \\
2261 & 27-28 Oct. & 343,862 & 97 & 0 & 28,795 & 0.028\% & 0.000\% & 83.74 & 296.87 & 0 \\
2821 & 23-26 Oct. & 661,964 & 394 & 217 & 68,257 & 0.060\% & 55.076\% & 103.11 & 173.24 & 1 \\
2997 & 26-27 Oct. & 156,063 & 533 & 0 & 8,617 & 0.342\% & 0.000\% & 55.22 & 16.17 & 0 \\
3358 & 13-15 Jun. & 300,928 & 339 & 58 & 34,159 & 0.113\% & 17.109\% & 113.51 & 100.77 & 2 \\
3386 & 13-15 Jun. & 545,421 & 496 & 0 & 45,715 & 0.091\% & 0.000\% & 83.82 & 92.17 & 0 \\
3427 & 13-15 Jun. & 536,795 & 395 & 0 & 46,356 & 0.074\% & 0.000\% & 86.36 & 117.36 & 0 \\
3476 & 13-15 Jun. & 523,848 & 302 & 11 & 43,627 & 0.058\% & 3.642\% & 83.28 & 144.46 & 10 \\
Total & - & 4,100,716 & 3,230 & 318 & 364,243 & 0.079\% & 9.845\% & 88.82 & 112.77 & - \\
\end{tabular}
\end{table*}
The basic statistical information is given in Table~\ref{tab:basic-stats}. Specifically, the ``Win Ratio" column is about the ad auction winning ratio with the default bidding strategy from iPinYou platform. Conversion rate (CVR) is with respect to the number of clicks (instead of impressions). Note that in the original record, there would be multiple clicks on the same impression. However, duplications are removed in our analysis to allow focus on the events themselves (whether users would click or convert, or not).
From Table~\ref{tab:basic-stats} we can see that (i) all the advertisers has CTR less than 0.1\% except for advertiser 2997 (0.444\%). Note that 0.1\% is usually around the average CTR for desktop display advertising in practice. The high CTR for advertiser 2997 confirms the difference of mobile environment where clicks are more easily generated possibly due to ``fat finger'' effect; (ii) Although the nine advertisers have similar CPM, their effective cost-per-click (eCPC), i.e. the expected cost for achieving one click, are fairly different. This could be caused by the target rule setting (i.e., the target user demographic information, location and time) and the market of each specific advertiser; (iii) Some advertisers do not record conversions. Even for the ones who did report conversions, their CVRs differ a lot, which could also be due to different market and conversion setting.
In the table of test data, there is a conversion weight factor for each advertiser, denoted by $N$. It shows the relative importance of a conversion against a click for each advertiser. For example, the weight factor for advertiser 3476 (tire, $N=10$) is much higher than that for advertiser 2259 (milk powder, $N=1$).
\subsection{User Feedback}\label{sec:ctr}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figs/icc-1458-3358-ctr-features-white}
\caption{CTR distribution against different features for advertiser 1458 and 3358. We choose only two advertisers here because of the page limit and presentation concern.}
\label{fig:ctr-1458-3358}
\end{figure}
Figure~\ref{fig:ctr-1458-3358} depicts some statistics of user feedback on advertiser 1458 and 3358. Specifically, the mean value with the standard error of CTR\footnote{We do not compare the CVR here because the conversion definitions across different advertisers are very different. For example, advertiser 2821 sells footwear and the conversion is defined as a purchase, while advertiser 3358 sells software and the conversion is a download.} against some features, such as the time, location, user-agent, publisher's ad slot size, ad exchanges, and user tags\footnote{Tags are re-indexed by the descending rank of their frequency.}.
We can see from Figure~\ref{fig:ctr-1458-3358} that for different advertisers, the same feature could have a different impact on the CTR: \\
\textsc{(i)} Advertiser 1458 has received the highest CTR on Monday and weekends while advertiser 3358 does on Tuesday and Wednesday. \\
\textsc{(ii)} The mobile users (on Andriod or iOS) are more likely to click the ads from Advertiser 1458 while PC users (on Mac and Windows) prefer the ads from Advertiser 3358. \\
\textsc{(iii)} The ad CTR from two advertisers are both volatile across different region locations, and the trend are different. \\
\textsc{(iv)} Ad slot size is correlated with the slot locations in the webpage and the design of creatives. We can see the banner ($1000\times 90$) and standard ($300\times 250$) slots generally have the highest CTR for both advertisers. \\
\textsc{(v)} Ad exchanges call for bids and host the auctions. Different publishers (or their supply-side platforms, i.e. SSPs) connect different exchanges, thus the CTR distribution on these exchanges are different. \\
\textsc{(vi)} The CTR against different user tags is depicted in a log scale because the difference is fairly large. For example, the vertical e-commerce advertiser 1458 could receive the CTR as high as 30\% on the users with the tag 38 (\texttt{In-market/clothing, shoes\&bags}) while only around 0.1\% CTR on the other users. The same volatility happens to advertiser 3358. It shows the importance of user segmentation for predicting their response for a specific ad campaign. Therefore, the advertisers can refine their targeting rules and bidding strategies based on the ad performance on different user segmentations. Such user segmentation data is often provided by a third-party data management platform (DMP) or DSPs themselves.
In sum, the above analysis suggests that the user response models need to be trained independently for each advertiser. It requires some non-trivial work \cite{ahmed2014scalable} to leverage data from the similar advertisers to improve the performance of prediction. In addition, advertisers may not allow the DSP to use their data to help other advertisers.
\subsection{Bidding Behaviour}\label{sec:market-price}
In the second price auctions, the second highest bids are defined as the \emph{market price} for the winner. If his/her bid is higher than the market price, the advertiser wins this auction and pays the market price. Market price is always modelled as a stochastic variable because it is almost impossible to analyse the strategy of each of the thousands of auction participators \cite{amin2012budget}. A higher market price reflects a more competitive the environment.
Here we have an investigation of the market price of advertiser 1458 and 3358 from the perspective of a DSP. The market price mean and standard error against different features are depicted in Figure~\ref{fig:price-1458-3358}, where we can see that just like the CTR plot, the market price has different trends against the same domain features on these two advertisers. For example, for advertiser 1458 the bid competitiveness in the morning is higher than that in the afternoon and evening, while it is inverse for advertiser 3358. In addition, for advertiser 1458 the competitiveness in ad exchange 1 is higher than that in ad exchange 2 and 3, while for advertiser 3358, the ad exchange 2 is the most competitive one.
Comparing the market price and CTR distribution on individual features, we find that the ratio of standard error to its mean of market prices is smaller than that of CTR. This is mainly because the click observations are binary value while market prices are integers.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figs/icc-1458-3358-price-features-white}
\caption{Market price distribution against different features for advertiser 1458 and 3358.}
\label{fig:price-1458-3358}
\end{figure}
\subsection{eCPC}\label{sec:ecpc}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figs/icc-3358-ecpc-features}
\caption{eCPC against different features for advertiser 3358. Zero-height bars mean there is no observed click for the specific feature.}
\label{fig:ecpc-3358}
\end{figure}
From the joint observation on Figure~\ref{fig:ctr-1458-3358} and \ref{fig:price-1458-3358}, if we regard user clicks as the return, we can find some imbalanced return-on-investment (ROI) across the cases with different features. For example, we consider the weekday features for advertiser 3358. It has the lowest CTR but the highest market price on Sunday. Compared with Sunday, it has a lower market price but an around three-time CTR on Thursday. Another example for advertiser 1458 is that it has a quite lower market price on ad exchange 3 than that on ad exchange 2, but a higher CTR there. Such ROI is effectively measured by eCPC, which shows the amount of money that needs to be spent to achieve one click. A lower eCPC suggests a more cost effective algorithm.
We depict the eCPC bars against different features for advertiser 3358 in Figure~\ref{fig:ecpc-3358}. Just as the first example above, advertiser 3358 suffers the highest eCPC on Sunday, while it is much cost effective on Monday and Wednesday. In fact, the eCPC varies greatly against almost every feature considered here. For example, for advertiser 3358, the advertising to iOS users is about 3 times cost effective than that to Windows users. Its two low ROI creatives have the size of $200\times200$ and $360\times300$. The auctions from ad exchange 3 are much more cost effective than those from ad exchange 2. The users with tag 22, 38, 44, 49, 51, and 65 show significantly higher interest on this advertiser's ads than other users.
Ideally, if certain kind of features brings a lower eCPC than average, the advertiser (or DSP) should allocate more budget (via bidding higher) in such auctions. The observation from Figure~\ref{fig:ecpc-3358} indicates that there is great optimisation potential. For example, if the advertiser reallocates part of the budget from Sunday to Monday, more clicks could be achieved with the same budget. But we cannot bid much higher on Monday, because the higher bid price results in higher cost, which always increases the eCPC. Thus, there is a trade-off between the achieved clicks and the eCPC. Due to the page limit, we do not show the eCPC performance for other advertisers. In fact, all advertisers have the inconsistent eCPC across different features, and the changing trends across the features are different, too.
From the above data analysis we can see the same features would have much different impact on the user feedback and market price of different advertisers, which results in different eCPC. Therefore, it is reasonable to independently build the user response models and bidding strategies for each advertiser.
\section{Task and Protocol}\label{sec:task}
\begin{figure*}
\vspace{-0pt}
\centering
\includegraphics[width=0.9\textwidth]{figs/DSP-train-test-flow-new}
\caption{The training framework and evaluation protocol.}
\label{fig:training-evaluation}
\end{figure*}
\subsection{Task Description}\label{sec:task}
The DSP bid optimisation task for a given advertiser refers to optimising a predefined KPI given a cost budget and the coming bid requests during the lifetime of the budget. For example, a straightforward KPI is the click (conversion) number and the task is to maximise the click (conversion) number with the given budget and coming bid requests. We write the general optimisation formula as:
\begin{align}
\max_{\text{bidding strategy}} &~~~ \text{KPI}\\
\text{subject to} &~~~ \text{cost} \leq \text{budget}.
\end{align}
In the iPinYou bidding algorithm competition, the KPI was a linear combination of the click number and conversion number:
\begin{align}
\max_{\text{bidding strategy}} &~~~ \text{\#click} + N \cdot \text{\#conversion} \label{eq:ipinyou-kpi} \\
\text{subject to} &~~~ \text{cost} \leq \text{budget},
\end{align}
where $N$ is the parameter showing the relative importance of conversions to clicks and it varies across different advertisers, as has been shown in Table~\ref{tab:basic-stats}. This KPI is practically meaningful since conversions are the final measure for advertising but they are usually sparse.
As shown in Figure~\ref{fig:interaction}, the input of the bidding strategy is a bid request (with the auction and ad information), the output is the bid response for this bid request. In addition, in the process of making a bid decision, the budget, current cost and achieved performance is accessible to the bidding engine.
Figure~\ref{fig:training-evaluation} presents a flow diagram about the bidding strategy training framework and the following test evaluation protocol. In the latter two subsections, we discuss these two parts respectively.
\subsection{Training Framework}\label{sec:training-framework}
As discussed in Section~\ref{sec:dataset}, given the training data, one can investigate the data to learn about the bid request feature distribution and derive its relationship with the user ad response (e.g., CTR) and the market price. As shown in the left part of Figure~\ref{fig:training-evaluation}, based on the components of the CTR estimator and the market price model, one can use the training data to perform the bidding function optimisation and finally obtain the bidding strategy.
With the feature engineering on the bid request data and the label extraction from user feedback data, one can train a CTR estimator to predict the probability of the user click on a given ad impression. Standard regression models such Logistic regression and tree models can be used here. We will later show the details in the experiment section.
Besides the CTR estimation, the market price distribution can also be estimated either by direct data statistics or a regression model, which is called bid landscape forecasting \cite{cui2011bid}. With the knowledge of the market price distribution, one can estimate the probability of winning a specific ad auction and the corresponding cost given a bid price.
A bidding function generally takes the predicted CTR (pCTR) and the market price distribution as input and outputs the bid price. As a framework description, here we do not specify the process of optimising the bidding function, which could be different for different models. In fact, the market price model has not been formally considered in previous bid optimisation work \cite{perlich2012bid,lee2012estimating}. In addition, other factors can also be added into this training framework.
\subsection{Evaluation Protocol}\label{sec:eval-protocol}
The evaluation protocol is illustrated as in the middle and right part of Figure~\ref{fig:training-evaluation}. Given the bidding strategy and a budget for the test period for a particular advertiser, we can perform a simulation via going through its bid logs and comparing the bidding results with impression, click, and conversion logs. A detailed description of each step is as follow: \\
\textsc{(0)} To initialise the bidding engine, such as setting a predefined budget, initialising the cost and performance (e.g., achieved click and conversion number) as zero.\\
\textsc{(1)} To pass the next bid request (in the ascending order of the timestamp) to the bidding engine. A bid request contains both contextual and behavioural data of the auction as well as the ad data as shown in Table~\ref{tab:data-format}. \\
\textsc{(2)} The bidding strategy computes a bid for this request (with the information of the budget, current cost and achieved performance). This step is highlighted in Figure~\ref{fig:training-evaluation} as it is what bid optimisation focuses on. Note that if the cost has been higher than the budget (i.e. the budget is run out), all the bid responses should be set to zero (i.e. to skip all the left bid requests).\\
\textsc{(3)} To simulate the auction by referencing the impression logs: if the bid price is higher than the logged auction winning price (i.e., paying price on column 21 in Table~\ref{tab:data-format}) and floor price (column 18 in Table~\ref{tab:data-format}), the bidding engine wins the auction and gets the ad impression. \\
\textsc{(4)} To match the click and conversion events in the logs for this impression if winning the auction. The performance of the bidding strategy is then updated and saved. The cost is also added by the paying price. \\
\textsc{(5)} To check whether there is any bid request left in test data to determine if the evaluation needs to terminate. Note that a tiny amount of over-spend from the last impression is possible but it is neglected in our evaluation.
Compared the procedures in Figure~\ref{fig:interaction} and \ref{fig:training-evaluation}, we can see step 1,2 and 3 are the same. The step 4 in evaluation flow actually merges the step 4 (win notice), 5 (ad delivery and tracking) and 6 (user feedback notice) in Figure~\ref{fig:interaction}, which is reasonable for offline evaluation since the log data has already collected the user feedback information.
It is worth noting that there are limitations of replaying logs for offline evaluation. Click and conversion events are only available for winning auctions (having impressions); there is no data for the lost auctions. Thus, there is no way to check if the performance could be improved if the algorithm bids higher to win those lost auctions.
However, this evaluation protocol do follow the convention of the offline evaluations from sponsored search \cite{zhang2012joint,graepel2010web}, recommender systems \cite{zhang2013optimizing} and Web search \cite{craswell2008experimental} where the objects (auctions) with unseen user feedback have to be ignored.
\section{Benchmark Experiments}
As discussed before, it is routine to perform the real-time bidding with a two stage process \cite{perlich2012bid,lee2012estimating}: (i) estimating the CTR/CVR of the ad impression being auctioned; (ii) making the bid decision based on the evaluation and other information. For iPinYou dataset, different advertisers have largely different setting on their conversions and more than half of them has no conversion record. Therefore, in our benchmark experiment, we will focus on the clicks. We first test the performance of two standard CTR estimators. Then we compared several bidding strategies based on the pCTR. Besides the total achieved clicks, the considered KPIs include conversions and the iPinYou KPI in Eq. (\ref{eq:ipinyou-kpi}).
\subsection{CTR Estimation}
\subsubsection{Compared Models and Feature Engineering}
In our benchmarking, we consider the following two CTR estimation models. Besides the model setting, the feature engineering details are also discussed respectively.
\noindent \textbf{Logistic Regression (\textsf{LR})} is a widely used linear model to estimate CTR in computational advertising \cite{Richardson2007a}. The loss is the cross entropy between the predicted click probability and the ground-truth result. In addition, L2 regularisation is used.
In our experiment, all the features for \textsf{LR} are binary. Specifically, we extract the weekday and hour feature from timestamps. User agent text is processed to extract the operation systems and browser brands as shown in Figure~\ref{fig:ctr-1458-3358}. The floor price is processed by buckets of 0, [1,10], [11,50], [51,100] and [101,+$\infty$). The tag list of each user is divided into binary features for each tag. We do not include the features of Bid ID, Log Type, iPinYou ID, URL, Anonymous URL ID, Bidding Price, Paying Price, Key Page URL because they are either almost unique for each case or meaningless to be added to \textsf{LR} training. Also we do not add combination features (e.g., weekday-region-tag) because there are many variants and tricks for adding high-order combination features, which is not recommended in benchmark experiment. In sum, we have 937,748 binary features for \textsf{LR} training and prediction.
\noindent \textbf{Gradient Boosting Regression Tree (\textsf{GBRT})} \cite{friedman2002stochastic} is a non-linear model widely used in regression and learning to rank applications \cite{burges2010ranknet}. Comparing to the linear model \textsf{LR}, \textsf{GBRT} has the advantage of learning the non-linear features, which can hardly be achieved by the feature engineering of \textsf{LR}. However, once finishing training, \textsf{GBRT} will get rid of most features and only keep a small part of features for prediction. In our experiment, we leverage the open-sourced \textsc{xgboost}\footnote{https://github.com/tqchen/xgboost} for implementation. Specifically, we set the max tree depth as 5 and train 50 trees with 0.05 learning rate.
Different with the binary features for \textsf{LR}, here every feature for \textsf{GBRT} is continuous. Specifically, for the indicator features (e.g., city=152) of a domain (e.g., city) we calculate the frequency and CTR based on a subset of the training file and let these two numbers be the values of two \textsf{GBRT} features (e.g., the frequency that city=152 is 348,432, the empirical CTR for the cases whose city=152 is 0.1\%). We do not make the frequency value for tag features. For the continuous feature (e.g., slot floor price), we directly use the specific value as the feature value (e.g., slot floor price=21). In sum, we have 78 features for \textsf{GBRT} training and prediction.
\subsubsection{Evaluation Measure}
The area under ROC curve (AUC) is a widely used measure for evaluating the ad CTR estimator \cite{graepel2010web,oentaryo2014predicting}. Besides AUC, the root mean square error (RMSE) is also chosen as the evaluation measure here as it is widely used in various regression tasks. Because of the huge imbalance of positive/negative cases in ad clicking, the empirically best regression model usually provides the pCTR very close to 0, which results in the RMSE having a quite small value and the improvement on RMSE is much slight, compared with AUC.
\begin{table}[t]
\center
\caption{CTR estimation performance.}
\label{tab:ctr-estimation}
\begin{tabular}{cc|cc|cc}
\small
& & \multicolumn{2}{c|}{AUC} & \multicolumn{2}{c}{RMSE}\\
Season & Adv. & \textsf{LR} & \textsf{GBRT} & \textsf{LR} & \textsf{GBRT}\\
\hline
2 & 1458 & 0.9881 & 0.9707 & 0.0191 & 0.0263\\
3 & 2259 & 0.6865 & 0.6791 & 0.0177 & 0.0176\\
3 & 2261 & 0.6238 & 0.5739 & 0.0168 & 0.0167\\
3 & 2821 & 0.6325 & 0.5820 & 0.0239 & 0.0238\\
3 & 2997 & 0.6039 & 0.5979 & 0.0582 & 0.0581\\
2 & 3358 & 0.9753 & 0.9722 & 0.0246 & 0.0279\\
2 & 3386 & 0.7908 & 0.7686 & 0.0284 & 0.0285\\
2 & 3427 & 0.9735 & 0.9342 & 0.0214 & 0.0245\\
2 & 3476 & 0.9625 & 0.9422 & 0.0230 & 0.0231\\
2 & Total & 0.9141 & 0.9200 & 0.0263 & 0.0260\\
3 & Total & 0.7615 & 0.7715 & 0.0268 & 0.0268\\
2,3 & Total & 0.8307 & 0.8518 & 0.0270 & 0.0263\\
\end{tabular}
\end{table}
\subsubsection{Results}
The experimental results for CTR estimation with \textsf{LR} and \textsf{GBRT} are shown in Table~\ref{tab:ctr-estimation}. From the results we can see different advertisers have quite large difference on the value of AUC and RMSE, due to the different user behaviour on their ads. For example, advertiser 2997 has the highest overall CTR (0.444\%) but the lowest observation number (see Table~\ref{tab:basic-stats}) which makes it more difficult to predict the CTR. In addition, both models achieve a much better performance on advertisers in season 2 than that in season 3. iPinYou technicians explained that this is due to the different user segmentation systems between season 2 and 3.
\subsection{DSP Bid Optimisation}
\subsubsection{Compared Bidding Strategies}
We compare the following bidding strategies in our benchmark experiment for DSP real-time bidding. The parameters of each bidding strategy are tuned using the training data. And the evaluation is performed on the test data.
\noindent \textbf{Constant bidding (\textsf{Const}).} Bid a constant value for all the bid requests. The parameter is the specific constant bid price.
\noindent \textbf{Random bidding (\textsf{Rand}).} Randomly choose a bid value in a given range. The parameter is the upper bound of the random bidding range.
\noindent \textbf{Bidding below max eCPC (\textsf{Mcpc}).} The goal of bid optimisation is to reduce the eCPC. In \cite{lee2012estimating}, given the advertiser's goal on max eCPC, which is the upper bound of expected cost per click, the bid price on an impression is obtained by multiplying the max eCPC and the pCTR. Here we calculate the max eCPC for each campaign by dividing its cost and achieved number of clicks in the training data. No parameter for this bidding strategy.
\noindent \textbf{Linear-form bidding of pCTR (\textsf{Lin}).} In the previous work \cite{perlich2012bid}, the bid value is linearly proportional to the pCTR under the target rules. The formula can be generally written as \texttt{bid=base\_bid$\times$pCTR/avgCTR}, where the tuning parameter \texttt{base\_bid} is the bid price for the average CTR cases.
Among the compared bidding strategies, only \textsf{Mcpc} is non-parametric and it does not consider the budget limits while the others do by tuning their parameters. In addition, \textsf{Mcpc} and \textsf{Lin} need to evaluate CTR for each impression. Thus we denote the two bidding strategies with \textsf{LR} CTR estimator as \textsf{Mcpc-L}, \textsf{Lin-L} and with \textsf{GBRT} CTR estimator as \textsf{Mcpc-G} and \textsf{Lin-G}.
\subsubsection{Experimental Setting}
The evaluation follows the protocol in Section~\ref{sec:eval-protocol}. The only issue discussed here is about the pre-set budget for each advertiser. In our experiment we set the budget for each advertiser as a proportion of the original total cost in the test log. Particularly, in order to check the bidding strategies' performance under different budget limits, we set the budget as 1/32, 1/8, and 1/2 of the original total cost in the test log. Note that we cannot set the budget higher than the original total cost because in such case, simply bidding as high as possible on each auction will make the DSP win all the auctions without running out of the budget.
\subsubsection{Results}
We list the achieved click performance for each algorithm under different budget limits in Table~\ref{tab:clk-perf}. We can see from Table~\ref{tab:clk-perf} that \textsf{Lin} and \textsf{Mcpc} generally work much better than \textsf{Const} and \textsf{Rand}, which verifies the importance of impression-level evaluation and real-time bidding. To some advertisers, such as 2259 and 2261 with 1/32 and 1/8 budget limits, \textsf{Mcpc} achieves fewer clicks than \textsf{Const} and \textsf{Rand}. This is because \textsf{Mcpc} is not adaptive with different budget limits, which results in running out of budget quite soon for the low budget settings. Moreover, \textsf{Lin} works much better than \textsf{Mcpc}. Compared with \textsf{Mcpc}, \textsf{Lin} has the ability to change the bidding scale by tuning its parameter \texttt{base\_bid}, which helps \textsf{Lin} adapt different budget limits against the coming ad auction volume. Such adaptivity is essential to DSP bidding strategies because the ad auction volume and market price could vary a lot as time goes by \cite{cui2011bid}. Generally, if the budget allocated per ad auction is high, then it is encouraged to bid high to achieve more user clicks and conversions, while if the budget allocated per ad auction is low, the bid price should be reduced to remain the high ROI.
\begin{table}[t]
\center
\vspace{-0pt}
\caption{Click numbers for each advertiser under different budget limits.}
\label{tab:clk-perf}
\small
\begin{tabular}{r|rrrrrr}
\multicolumn{7}{c}{\textbf{Budget (1/32)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 28 & 29 & 261 & 89 & 500 & 491\\
2259 & 12 & 11 & 7 & 6 & 16 & 15\\
2261 & 9 & 9 & 3 & 3 & 12 & 11\\
2821 & 37 & 44 & 16 & 16 & 57 & 42\\
2997 & 74 & 63 & 22 & 46 & 78 & 78\\
3358 & 11 & 13 & 85 & 83 & 278 & 260\\
3386 & 23 & 23 & 39 & 29 & 127 & 86\\
3427 & 21 & 21 & 67 & 43 & 321 & 294\\
3476 & 27 & 25 & 33 & 28 & 205 & 140\\
S2 & 110 & 111 & 485 & 272 & 1431 & 1271\\
S3 & 132 & 127 & 48 & 71 & 163 & 146\\
Total & 242 & 238 & 533 & 343 & 1594 & 1417\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/8)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 92 & 98 & 502 & 294 & 521 & 496\\
2259 & 28 & 27 & 25 & 27 & 37 & 32\\
2261 & 25 & 26 & 23 & 20 & 28 & 25\\
2821 & 88 & 88 & 61 & 84 & 109 & 91\\
2997 & 159 & 138 & 78 & 112 & 162 & 166\\
3358 & 50 & 59 & 310 & 272 & 314 & 284\\
3386 & 67 & 72 & 128 & 112 & 216 & 184\\
3427 & 56 & 59 & 361 & 205 & 358 & 321\\
3476 & 63 & 71 & 148 & 94 & 273 & 235\\
S2 & 328 & 359 & 1449 & 977 & 1682 & 1520\\
S3 & 300 & 279 & 187 & 243 & 336 & 314\\
Total & 628 & 638 & 1636 & 1220 & 2018 & 1834\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/2)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 322 & 315 & 502 & 503 & 540 & 524\\
2259 & 73 & 79 & 86 & 86 & 96 & 87\\
2261 & 66 & 67 & 70 & 66 & 72 & 67\\
2821 & 240 & 236 & 239 & 124 & 245 & 253\\
2997 & 366 & 355 & 329 & 112 & 358 & 359\\
3358 & 163 & 174 & 310 & 272 & 335 & 307\\
3386 & 276 & 277 & 336 & 280 & 401 & 369\\
3427 & 208 & 217 & 361 & 329 & 389 & 363\\
3476 & 195 & 186 & 290 & 274 & 301 & 279\\
S2 & 1164 & 1169 & 1799 & 1658 & 1966 & 1842\\
S3 & 745 & 737 & 724 & 388 & 771 & 766\\
Total & 1909 & 1906 & 2523 & 2046 & 2737 & 2608\\
\end{tabular}
\end{table}
Table~\ref{tab:cnv-perf} lists the conversion performance on four advertisers with conversion records. And Table~\ref{tab:score-perf} lists the corresponding iPinYou KPI score of $\text{\#click}+N\cdot\text{\#conversion}$, where $N$ for each advertiser has been shown in Table~\ref{tab:score-perf}. We can see the significant improvement of \textsf{Lin} against other bidding strategies. Another common important point from the results of these three KPIs is that in the lower budget setting, \textsf{Lin} achieves the higher improvement rate against other strategies. This is because when the budget is quite limited, it is more important to identify which cases are probably valuable and adaptively lower the overall bid price.
\begin{table}[t]
\center
\caption{Conversion numbers for each advertiser under different budget limits.}
\label{tab:cnv-perf}
\small
\begin{tabular}{r|rrrrrr}
\multicolumn{7}{c}{\textbf{Budget (1/32)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
2259 & 4 & 4 & 3 & 4 & 4 & 4\\
3476 & 1 & 2 & 1 & 1 & 3 & 3\\
2821 & 24 & 29 & 7 & 8 & 33 & 29\\
3358 & 4 & 4 & 14 & 14 & 53 & 51\\
S2 & 5 & 6 & 15 & 15 & 56 & 54\\
S3 & 28 & 33 & 10 & 12 & 37 & 33\\
Total & 33 & 39 & 25 & 27 & 93 & 87\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/8)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
2259 & 7 & 9 & 7 & 8 & 10 & 10\\
3476 & 3 & 3 & 4 & 3 & 8 & 5\\
2821 & 54 & 51 & 35 & 49 & 61 & 54\\
3358 & 10 & 11 & 56 & 52 & 57 & 53\\
S2 & 13 & 14 & 60 & 55 & 65 & 58\\
S3 & 61 & 60 & 42 & 57 & 71 & 64\\
Total & 74 & 74 & 102 & 112 & 136 & 122\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/2)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
2259 & 16 & 19 & 19 & 20 & 23 & 20\\
3476 & 6 & 8 & 9 & 7 & 11 & 7\\
2821 & 140 & 133 & 130 & 75 & 134 & 139\\
3358 & 28 & 31 & 56 & 52 & 58 & 57\\
S2 & 34 & 39 & 65 & 59 & 69 & 64\\
S3 & 156 & 152 & 149 & 95 & 157 & 159\\
Total & 190 & 191 & 214 & 154 & 226 & 223\\
\end{tabular}
\end{table}
\begin{table}[t]
\center
\caption{KPI score (\#clicks+$N\cdot$\#conversions) for each advertiser under different budget limits.}
\label{tab:score-perf}
\small
\begin{tabular}{r|rrrrrr}
\multicolumn{7}{c}{\textbf{Budget (1/32)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 28 & 29 & 261 & 89 & 500 & 491\\
2259 & 16 & 15 & 10 & 10 & 20 & 19\\
2261 & 9 & 9 & 3 & 3 & 12 & 11\\
2821 & 61 & 73 & 23 & 24 & 90 & 71\\
2997 & 74 & 63 & 22 & 46 & 78 & 78\\
3358 & 19 & 21 & 113 & 111 & 384 & 362\\
3386 & 23 & 23 & 39 & 29 & 127 & 86\\
3427 & 21 & 21 & 67 & 43 & 321 & 294\\
3476 & 37 & 45 & 43 & 38 & 235 & 170\\
S2 & 128 & 139 & 523 & 310 & 1567 & 1403\\
S3 & 160 & 160 & 58 & 83 & 200 & 179\\
Total & 288 & 299 & 581 & 393 & 1767 & 1582\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/8)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 92 & 98 & 502 & 294 & 521 & 496\\
2259 & 35 & 36 & 32 & 35 & 47 & 42\\
2261 & 25 & 26 & 23 & 20 & 28 & 25\\
2821 & 142 & 139 & 96 & 133 & 170 & 145\\
2997 & 159 & 138 & 78 & 112 & 162 & 166\\
3358 & 70 & 81 & 422 & 376 & 428 & 390\\
3386 & 67 & 72 & 128 & 112 & 216 & 184\\
3427 & 56 & 59 & 361 & 205 & 358 & 321\\
3476 & 93 & 101 & 188 & 124 & 353 & 285\\
S2 & 378 & 411 & 1601 & 1111 & 1876 & 1676\\
S3 & 361 & 339 & 229 & 300 & 407 & 378\\
Total & 739 & 750 & 1830 & 1411 & 2283 & 2054\\
\multicolumn{7}{c}{}\\
\multicolumn{7}{c}{\textbf{Budget (1/2)}}\\
Adv. & \textsf{Const} & \textsf{Rand} & \textsf{Mcpc-L} & \textsf{Mcpc-G} & \textsf{Lin-L} & \textsf{Lin-G}\\
\hline
\\[-2.0ex]
1458 & 322 & 315 & 502 & 503 & 540 & 524\\
2259 & 89 & 98 & 105 & 106 & 119 & 107\\
2261 & 66 & 67 & 70 & 66 & 72 & 67\\
2821 & 380 & 369 & 369 & 199 & 379 & 392\\
2997 & 366 & 355 & 329 & 112 & 358 & 359\\
3358 & 219 & 236 & 422 & 376 & 451 & 421\\
3386 & 276 & 277 & 336 & 280 & 401 & 369\\
3427 & 208 & 217 & 361 & 329 & 389 & 363\\
3476 & 255 & 266 & 380 & 344 & 411 & 349\\
S2 & 1280 & 1311 & 2001 & 1832 & 2192 & 2026\\
S3 & 901 & 889 & 873 & 483 & 928 & 925\\
Total & 2181 & 2200 & 2874 & 2315 & 3120 & 2951\\
\end{tabular}
\end{table}
\section{Other Research Topics}
Besides the CTR estimation and DSP bid optimisation problem, there are other potential research topics where this dataset can be used for experimentation.
\begin{itemize}
\item \textbf{Bid landscape modelling.} As previously mentioned, for RTB based display advertising, market price can varying quite a lot \cite{cui2011bid}. The current DSP work \cite{perlich2012bid,lee2012estimating} mainly focuses on the bidding strategy based on CTR/CVR estimation. However, another important factor for the bidding decision making is the market price. As a repeated auction game with the budget constraint, the optimal bidding strategy is not truth-telling while it does depend on the market price distribution. Thus the research problem of modelling the market price distribution is much meaningful for a better bidding strategy.
\item \textbf{Adaptive bid control.} Just like sponsored search, a practically significant problem is to adaptively control the scale of real-time bidding to exactly deliver the budget during the campaign's lifetime. For some advertisers setting a high bidding scale at the beginning, their budget could be run out just at the beginning of the campaign lifetime. The techniques of pacing \cite{lee2013real} could be used here. More generally, in \cite{amin2012budget}, the authors model the bid control problem in sponsored search via Markov decision processes (MDPs) where the left budget and ad auction volume are states while the bid prices are actions. Further research on adaptive bid control can be based on this dataset.
\end{itemize}
\section{Conclusions}
Due to the sensitivity, research on computational advertising, especially from academia, is seriously restricted by the data availability. With the publication of iPinYou dataset, we believe the research on RTB display advertising will be stimulated.
In this paper, we performed a detailed statistical analysis on iPinYou dataset; formally defined the bid optimisation problem; and presented a simple yet comprehensive offline evaluation protocol for bidding strategies. We conducted benchmark experiments on CTR estimation and DSP bid optimisation with several baseline algorithms. The dataset and report are hosted on our website for computational advertising research.
{
\bibliographystyle{abbrv}
|
\section{Introduction}
Broad absorption lines (BALs) in quasar spectra identify high velocity outflows that originate from the quasar accretion disk and could play a major role in feedback to galaxy evolution. In order to determine the viability of BAL outflows as a feedback mechanism, we need estimates of their mass outflow rates and kinetic energy yields. These quantities depend on the outflow speeds, which are easy to measure, plus the column densities and distances of the flows from the central SMBH, which are much more difficult to derive. Distances can be inferred in some cases from BAL variability \citep{Misawa07,Moe09,Capellupo11,Capellupo13,Hall11,RodriguezH11,RodriguezH13} or from excited state lines with photoionization modeling \citep{Moe09,Dunn10,Borguet13}. Column density estimates are hampered by saturation. In an unknown number of cases, the apparent optical depths of BAL troughs give just lower limits on the true optical depths and column densities because the absorbing gas only partially covers the background light source (e.g. \citealt{Hamann98,Arav99a,Gabel03}).
One way to overcome the problem of saturation in BALs is to search for absorption in low-abundance ions, such as \mbox{P\,{\sc v}}\ (P/C $\sim$ 0.001 in the Sun; \citealt{Asplund09}). \mbox{P\,{\sc v}}\ is a good choice because it has a resonance doublet at accessible wavelengths, 1118 and 1128 \AA, and its ionization is similar to much more abundant and commonly measured ions such as \mbox{C\,{\sc iv}}\ $\lambda$\lam1548,1551. \mbox{P\,{\sc v}}\ absorption should be present if the column densities in the outflows are large enough. Originally, detections of \mbox{P\,{\sc v}}\ absorption were interpreted as indications of a vast over-abundance of phosphorus (e.g. \citealt{Turnshek88}; \citealt{Junkkarinen97}; \citealt{Hamann98}). The main problem is a poor understanding of the true BAL optical depths due to partial covering of the background continuum source (\citealt{Hamann98,Arav99a,Hamann02,Gabel03}). However, if the relative abundances are roughly solar, the presence of \mbox{P\,{\sc v}}\ absorption implies that \mbox{C\,{\sc iv}}\ and other common BALs are extremely optically thick and the total outflow column densities are much larger than previously supposed. For example, assuming solar abundances and a standard ionizing spectrum in photoionization models, \citet{Hamann98} showed that the true \mbox{C\,{\sc iv}}\ optical depths in one BAL quasar are at least $\sim$800 times greater than \mbox{P\,{\sc v}}\ in idealized BAL clouds that are optically thin throughout the Lyman continuum. In other situations with total column densities up to $N_H \sim 4\times 10^{23}$ cm$^{-2}$, the ratio of \mbox{C\,{\sc iv}}\ to \mbox{P\,{\sc v}}\ optical depths might be as low as $\sim$100. \citet{Leighly09} and \citet{Leighly11} present similar results across a wide range of physical conditions. Using many more observational constraints on the outflow conditions in a particular quasar, \citet{Borguet12} estimated that the \mbox{C\,{\sc iv}} /\mbox{P\,{\sc v}}\ optical depth ratio should be $\sim$1200. Altogether these results show that even a weak \mbox{P\,{\sc v}}\ detection indicates a very saturated \mbox{C\,{\sc iv}}\ BAL and large total column densities.
As part of our survey to study BAL variability \citep{Capellupo11,Capellupo12,Capellupo13}, we discovered variability in \mbox{P\,{\sc v}}\ and other BALs in the quasar Q1413+1143. The \mbox{P\,{\sc v}}\ absorption was detected previously by \citet{Monier98}. However, this is the first report of \mbox{P\,{\sc v}}\ BAL {\it variability}, in this or any quasar, to our knowledge. It is significant because it confirms that the measured feature is indeed \mbox{P\,{\sc v}}\ in the quasar's BAL outflow rather than a coincidental broad blend of unrelated, intervening Ly$\alpha$ absorption lines in the Ly$\alpha$ forest. It is also significant because it favors models for the variability that involve clouds crossing the line of sight at high speeds \citep{Hamann08,Hall11,Capellupo12,Capellupo13,Wildy14}. We analyze new and existing spectra of Q1413+1143 to derive new constraints on the location and column densities of the BAL outflow.
Q1413+1143, also known as the Cloverleaf, is a well-studied lensed quasar at $z_e = 2.563$ \citep{Adelman08}, with four components of similar brightness. The faintest component is within 0.5 mag of the brightest component \citep{Turnshek97}. The components have a small separation; all are within 0''.6 of the image center \citep{Magain88}. The four components are also similar spectroscopically. Although the BEL equivalent widths are smaller, and there are some small differences in the \mbox{C\,{\sc iv}}\ absorption profile, in the faintest component \citep{Angonin90}, \citet{Hutsemekers10} attributes the spectral differences in this faint component to a long-term microlensing effect, which magnifies the continuum source in this component.
Section \ref{data} describes the data and properties of Q1413+1143. Section \ref{anal} discusses the variability in the \mbox{P\,{\sc v}}\ BAL and describes the constraints that the \mbox{P\,{\sc v}}\ detection places on the column density of the outflow in this quasar. Section \ref{discuss} discusses our estimates of the energetics of this flow and their implications.
\section{Data}
\label{data}
\begin{table}
\caption{Data Summary.}
\begin{tabular}{clccc}
\hline
Year & Telescope & Resolution & $\Delta\lambda$ & $\Delta$t \\
& & (km s$^{-1}$) & (\AA) & (yr) \\
\hline
1989.26 & Lick 3-m & 530 & 3793--6382 & ... \\
1994.98 & HST FOS & 250 & 3235--4781 & 1.6 yr \\
2000.29 & HST STIS & 560 & 2900--5709 & 3.1 yr \\
2006.31 & SDSS 2.5-m & 150 & 3803--9221 & 4.8 yr \\
2010.11 & KPNO 2.1-m & 280 & 3482--6232 & 5.9 yr \\
\hline
\end{tabular}
\label{tab}
\end{table}
During the BAL monitoring campaign described in \citet{Capellupo11,Capellupo12,Capellupo13}, we monitored 24 BAL quasars, including Q1413+1143, and we searched for variability in the \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ absorption lines. In Q1413+1143, we also detected a variable \mbox{P\,{\sc v}}\ absorption line. The monitoring campaign includes 7 epochs of ground-based data for this quasar, but only 3 have wavelength coverage that extends blue enough to include the \mbox{P\,{\sc v}}\ BAL. The first of these 3 spectra is originally from \citet{Barlow93}, observed in April 1989, and the second spectrum is from the Sloan Digital Sky Survey \citep{Adelman08} and was taken in April 2006. The most recent spectrum was taken in February 2011 at the KPNO 2.1m. As mentioned earlier, Q1413+1143 is lensed. However, given the small separation between the four components, these spectra include light from all four components.
We further supplement our dataset for Q1413+1143 with spectra from the HST archive. Q1413+1143 was observed by both FOS and STIS. The FOS data has individual observations of each of the four lens components. We use the sum of the four components to compare to the observations above that include integrated light from all the components. We note that the D component of the lens had a weaker Ly$\alpha$ \lam1216 emission line than the other components (see also, \citealt{Angonin90}), but the absorption in \mbox{P\,{\sc v}}\ is identical in all the components. For the STIS data, there are separate observations of just the A and B components of the lens. The two spectra are nearly identical, so we combined them to improve the signal-to-noise.
Table \ref{tab} summarizes all of the data covering \mbox{P\,{\sc v}}\ that we discuss in this paper. The first column gives the fractional years of each observation. The next three columns list the observatories, the spectral resolution of the data, and the wavelength coverage. The final column lists the rest-frame time-scale, $\Delta$$t$, in comparison to the first observation, from 1989.26. For more details on these data, see \citet{Capellupo11} for the Lick and SDSS spectra, \citet{Monier98} for the HST FOS spectra, \citet{Monier09} for the HST STIS spectra, and \citet{Capellupo13} for the KPNO spectrum.
For some of our discussion of the BAL outflow properties below, we estimate a bolometric luminosity based on $\lambda L_{\lambda}$(1500\AA), measured from the absolute flux-calibrated Lick spectrum, in a cosmology with $H_o = 71$ km s$^{-1}$\ Mpc, $\Omega_M = 0.3$, and $\Omega_{\Lambda}=0.7$. Using a standard bolometric correction factor, $L\approx 4.4\lambda L_{\lambda}$(1500\AA) \citep{Hamann11}, and dividing by a lens magnification factor of 11 \citep{Venturini03}, gives a bolometric luminosity of $L = 2.4 \times 10^{46}$ ergs s$^{-1}$ for Q1413+1143.
We approximate the black hole mass by assuming $L = 1/3L_{edd}$, which gives $M_{\mathrm{BH}} \approx 4.8 \times 10^{8}$ M$_{\sun}$.
\section{Analysis and Results}
\label{anal}
\subsection{Variability in a \mbox{P\,{\sc v}}\ absorption line}
\label{var}
\begin{figure*}
\centering
\includegraphics[width=135mm]{fig1_mod.eps}
\caption{Spectrum of Q1413+1143, showing all of the epochs outlined in Section 2 and Table
\ref{tab}. The middle shaded region marks the interval of variability in \mbox{Si\,{\sc iv}}\ in the blue side of the trough, and, in order of increasing wavelength, the other three shaded regions show the corresponding velocity intervals in \mbox{P\,{\sc v}}, \mbox{N\,{\sc v}}, and \mbox{C\,{\sc iv}}. The black dotted curve is the pseudo-continuum fit, and formal 1$\sigma$ errors are plotted across the bottom.}
\label{spectra}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=75mm]{fig2_mod.eps}
\caption{The BAL line profiles for, from top to bottom, \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}. The shaded regions here mark the same intervals as in Fig. \ref{spectra}.}
\label{stack}
\end{figure}
In the variability study described in \citet{Capellupo11,Capellupo12}, we find that Q1413+1143 varied in both \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}. We also detect the \mbox{P\,{\sc v}}\ absorption line with variability at velocities corresponding to the variability in both \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}, which we discuss here for the first time. Out of the 7 epochs of data we have for this object in our BAL monitoring programme, 3 of the spectra extend blue enough to show this \mbox{P\,{\sc v}}\ absorption (Table \ref{tab}).
We adopt the Lick spectrum here as the fiducial spectrum, for which we fit a pseudo-continuum in \citet{Capellupo11}. We use the same pseudo-continuum fit in this paper. This pseudo-continuum includes a power-law fit to the continuum in regions free of emission and absorption, and for Q1413+1143, we used the spectral regions at 1270$-$1290 and 1680$-$1700 \AA. We also fit multiple gaussians to the \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ broad-emission lines (BELs) to get a smooth fit to the profiles. The \mbox{Si\,{\sc iv}}\ and \mbox{C\,{\sc iv}}\ BELs were fit in \citet{Capellupo11} in order to properly measure the \mbox{Si\,{\sc iv}}\ and \mbox{C\,{\sc iv}}\ BALs. We do not fit the Ly$\alpha$, \mbox{N\,{\sc v}}, or \mbox{O\,{\sc vi}}\ BALs or adjacent emission lines in the Ly$\alpha$ forest because of the large uncertainties and, specifically, the severe blending of Ly$\alpha$ and \mbox{N\,{\sc v}}\ with each other and possibly with \mbox{Si\,{\sc iii}}\ $\lambda$ 1206. Also, the \mbox{O\,{\sc vi}}\ BAL is not covered in the Lick 1989.26 spectrum and, therefore, it is not useful for our variability analysis. While we do measure the \mbox{P\,{\sc v}}\ absorption, we do not fit \mbox{P\,{\sc v}}\ emission because it does not have a significant emission line. The black dotted curve in Fig \ref{spectra} shows the pseudo-continuum fit, combining the power-law continuum fit and the fits to the \mbox{Si\,{\sc iv}}\ and \mbox{C\,{\sc iv}}\ BELs.
The strongest BALs in our data are identified by horizontal solid lines in Fig. \ref{spectra}. Redward of the Ly$\alpha$ emission are the well-studied, aforementioned \mbox{Si\,{\sc iv}}\ and \mbox{C\,{\sc iv}}\ BALs. Blueward of Ly$\alpha$ are \mbox{O\,{\sc vi}}\ \lam1037, \mbox{S\,{\sc iv}}\ $\lambda$\lam1063,1073, \mbox{P\,{\sc v}}, and \mbox{N\,{\sc v}}\ $\lambda$\lam1239,1243 \citep{Leighly09,Baskin13}.
To compare the spectra from different epochs, we use a simple multiplicative shift, with no change to the slope, to match the spectra along regions free of emission and absorption. Fig. \ref{spectra} shows the Lick spectrum from 1989.26 (black curve), with the HST FOS (1994.98; purple curve), HST STIS (2000.29; green curve), SDSS (2006.31; red curve), and KPNO (2010.11; blue curve) spectra overplotted. The \mbox{C\,{\sc iv}}\ BAL has a longer wing on the blue side of the profile in the Lick spectrum, and the variability in the BAL extends to higher outflow velocities than in \mbox{Si\,{\sc iv}}. There also appears to be variability in \mbox{N\,{\sc v}}\ at $\sim$1140$-$1160\AA, corresponding to the variability in the wing of the \mbox{C\,{\sc iv}}\ line. We do not have coverage across \mbox{O\,{\sc vi}}\ in the Lick spectrum, so we cannot compare its variability properties to those of the other lines. We also note that there is variability between the Lick spectrum and the later spectra at $\sim$1580\AA, which may be caused by variability in \mbox{Fe\,{\sc ii}}\ emission lines and/or absorption in these features \citep{Vestergaard05}.
To compare the variability in the different lines, we mark the region of variability in the blue side of the \mbox{Si\,{\sc iv}}\ trough with a shaded bar. We then mark the corresponding velocities in the \mbox{P\,{\sc v}}, \mbox{N\,{\sc v}}, and \mbox{C\,{\sc iv}}\ BALs with shaded bars, with the widths of the bars adjusted for the different doublet separations in the lines. \mbox{C\,{\sc iv}}\ has the narrowest doublet separation at 498 km s$^{-1}$, and \mbox{Si\,{\sc iv}}\ and \mbox{P\,{\sc v}}\ have wider doublet separations of 1930 and 2670 km s$^{-1}$, respectively. Fig. \ref{stack} shows the \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}\ profiles for three of the epochs, namely, Lick 1989.26, SDSS 2006.31, and KPNO 2010.11, shown here again with black, red, and blue curves, respectively. We plot here the normalized spectra, after dividing by the pseudo-continuum fit shown in Fig. \ref{spectra}. The velocity scales in the three panels are defined based on the wavelength of the short-wavelength doublet member for each ion.
Fig. \ref{stack} shows clearly how the profiles of these three lines match in velocity space. The main difference, as mentioned above, is that \mbox{C\,{\sc iv}}\ has a more extended wing on the blue side of the profile and the variability extends to higher outflow velocities than in \mbox{Si\,{\sc iv}}\ and \mbox{P\,{\sc v}}. The velocities where variability occurred in the blue side of the \mbox{Si\,{\sc iv}}\ trough and in the \mbox{P\,{\sc v}}\ trough closely match. The \mbox{P\,{\sc v}}\ line also varies in the same sense as \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}; all of the lines get weaker in the blue side of the troughs between 1989.26 and the later epochs.
The changes in the \mbox{P\,{\sc v}}\ line profile occur specifically between the Lick and the FOS observations, which are separated by 1.6 yr in the rest-frame of the quasar. The \mbox{P\,{\sc v}}\ line did not vary between any of the later observations presented here. However, in \citet{Capellupo13}, we found variability on time-scales as short as 0.25 yr (91 days) in \mbox{C\,{\sc iv}}\ between 2009.22 and 2010.11.
\subsection{Line optical depths and total column density}
\label{tau}
Here we focus on the Lick 1989.26 observation to derive the column densities of the outflowing gas. We first derive the apparent optical depth versus velocity, $\tau_{a}(v)$, for the \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}\ BALs.
\begin{figure}
\centering
\includegraphics[width=65mm]{fig3_mod.eps}
\caption{The normalized BAL line profiles for, from top to bottom, \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}\ for the Lick 1989.26 observation (bold curves). The thin curves show the gaussian fits to the \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ profiles in the top two panels. In the bottom panel, the thin curve is the fit to \mbox{P\,{\sc v}}\ using the optical depth profile for \mbox{Si\,{\sc iv}}\ (Fig. \ref{tau_all}).}
\label{gauss}
\end{figure}
Using the normalized spectrum, we fit gaussians to the \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ BAL profiles. These gaussians do not represent any physical properties of the gas; they are simply used to define a smooth fit to the profiles (top two panels of Fig. \ref{gauss}). We then use the equation, $I = I_{o}e^{-\tau_{a}}$, to derive $\tau_{a}(v)$ for the \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ lines, where $I_{o} = 1$ for these normalized spectra. The lines we analyze here, \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and also \mbox{P\,{\sc v}}, are all doublets, so deriving $\tau_{a}(v)$ in this way includes contributions from both transitions at each velocity. We therefore use Equation (2) from \citet{Junkkarinen83} to remove the doublet structure. The solid and dotted curves in Fig. \ref{tau_all} show the run of apparent optical depths versus velocity, $\tau_{a}(v)$, for \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}, respectively, after removing the doublet structure.
\begin{figure}
\includegraphics[width=80mm]{fig4_mod.eps}
\caption{The apparent optical depth profiles for \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}, after removing the doublet structure in each line. The \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ profiles are derived from the gaussian fits to the residual intensities in the BALs as shown in Fig. \ref{gauss}, and the \mbox{P\,{\sc v}}\ profile is derived based on the \mbox{Si\,{\sc iv}}\ optical depth profile.}
\label{tau_all}
\end{figure}
To estimate $\tau_{a}(v)$ for \mbox{P\,{\sc v}}, we first take the optical depth profiles for \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ and broaden the profiles based on the increased doublet separation in \mbox{P\,{\sc v}}. We then calculate $I(v)$ from this adjusted $\tau_{a}(v)$, and fit the resulting \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}\ intensity profiles to the \mbox{P\,{\sc v}}\ BAL. The \mbox{Si\,{\sc iv}}\ $I(v)$ profile provided the best fit to the \mbox{P\,{\sc v}}\ line, with a small adjustment to the amplitude (bottom panel of Fig. \ref{gauss}; see also, \citealt{Junkkarinen97}). We then calculate the \mbox{P\,{\sc v}}\ optical depth profile using this adjusted \mbox{Si\,{\sc iv}}\ $I(v)$ profile and $I = I_{o}e^{-\tau_{a}}$, and the result is plotted in Fig. \ref{tau_all} as the dashed-dotted curve. We note that even though the adjusted \mbox{Si\,{\sc iv}}\ $I(v)$ profile provides the better fit to the \mbox{P\,{\sc v}}\ BAL, it does overestimate the depth of the absorption on the red side of the trough (Fig. \ref{gauss}). This will have a small effect on the column densities we derive, but it does not significantly affect our final results.
We derive apparent ionic column densities for \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}\ directly from the $\tau_{a}(v)$ profiles shown in Fig. \ref{tau_all}. These column densities, at least for \mbox{C\,{\sc iv}}\ and \mbox{Si\,{\sc iv}}, most likely underestimate the true column densities because the lines are probably very saturated (see Section 1 and the Discussion below). We calculate these values for reference using the following equation:
\begin{equation}
N = \frac{m_{e}c}{\pi e^{2}f\lambda_{0}} \int \tau(v)\ \mathrm{d}v ,
\label{eqn:N}
\end{equation}
where $f$ is the oscillator strength and $\lambda_{0}$ is the laboratory wavelength \citep{Savage91}. We obtain the following values for the ionic column densities: $N_{\mathrm{CIV}} \sim 1.3 \times 10^{16}$ cm$^{-2}$, $N_{\mathrm{SiIV}} \sim 3.6 \times 10^{15}$ cm$^{-2}$, and $N_{\mathrm{PV}} \sim 4.0 \times 10^{15}$ cm$^{-2}$.
Given the small number of lines we measure and the modest constraints they provide on the ionization and column densities, it is sufficient to rely on the published photoionization models by \citet{Hamann98}, \citet{Leighly09,Leighly11}, and \citet{Borguet12} to constrain the true optical depths in the \mbox{C\,{\sc iv}}\ BAL and the true total column density, $N_H$, in the Q1413+1143 outflow. The quasar studied by \citet{Borguet12} has much narrower absorption lines than Q1413+1143, particularly in \mbox{P\,{\sc v}}, and the ionic column densities for \mbox{C\,{\sc iv}}, \mbox{Si\,{\sc iv}}, and \mbox{P\,{\sc v}}\ are roughly an order of magnitude smaller in the quasar they studied. They also have many more lines detected that provide specific constraints on the outflow physical conditions. Based on those constraints, \citet{Borguet12} estimate that the true ratio of \mbox{C\,{\sc iv}}\ to \mbox{P\,{\sc v}}\ optical depths should be $\tau_{\mathrm{CIV}}$/$\tau_{\mathrm{PV}}$ $\sim$ 1200 for solar abundances.
\citet{Hamann98}, \citet{Leighly09}, and \citet{Leighly11} consider a wider range of physical conditions in BAL quasars that more closely resemble Q1413+1143, e.g., with broad \mbox{P\,{\sc v}}\ troughs, similar to \mbox{C\,{\sc iv}} , and limited ranges of only higher ions detected. In the BAL quasar studied in \citet{Hamann98}, $N_{\mathrm{CIV}}$ is roughly a factor of 2 smaller than in Q1413+1143, and $N_{\mathrm{SiIV}}$ and $N_{\mathrm{PV}}$ are roughly a factor of 4 smaller. \citet{Hamann98} showed that in the limit of low total column densities, where the gas is optically thin throughout the Lyman continuum, the optical depth ratio should be $\tau_{\mathrm{CIV}}$/$\tau_{\mathrm{PV}}$ $\ga$ 800 for solar abundances and any range of ionization conditions. However, those models can be unrealistic if the total columns are too low to yield a detectable \mbox{P\,{\sc v}}\ BAL. \citet{Hamann98} and \citet{Leighly09} estimate that the \mbox{P\,{\sc v}}\ detections in the quasars they studied require minimum total column densities of $N_H\ga 10^{22}$ and $\ga$$10^{22.2}$ cm$^{-2}$, respectively. In that situation, the optical depth ratio should be $\tau_{\mathrm{CIV}}$/$\tau_{\mathrm{PV}}$ $\sim$ 1000 near the column density lower limits and possibly as low as $\sim$100 in extreme circumstances with total columns up to $4\times 10^{23}$ cm$^{-2}$ (e.g., figures 6 in Hamann 1998 and 15 in Leighly et al. 2011). The optical depth ratio in Q1413+1143 should be somewhere in this range.
To estimate the total column density, we first note that the \mbox{P\,{\sc v}}\ BAL in Q1413+1143 is significantly deeper than the quasars measured by \citet{Leighly09} and especially \citet{Hamann98}, indicating roughly 2--4 times larger \mbox{P\,{\sc v}}\ optical depths and 2--4 times larger \mbox{P\,{\sc v}}\ column densities. If we scale up the column density constraints in \citet{Hamann98} and \citet{Leighly09} by a factor of $\sim$2 based on the stronger \mbox{P\,{\sc v}}\ BAL, we estimate that the total outflow column density in Q1413+1143 is conservatively $\log N_H > 22.3$ cm$^{-2}$.
\section{Discussion}
\label{discuss}
\subsection{Location of the outflow}
\label{loc}
In \citet{Capellupo12} and \citet{Capellupo13} we discussed the two most likely scenarios for explaining the variability in BALs (see also, \citealt{Lundgren07,Wildy14}). The first scenario is a change in the quasar's continuum flux causing global changes in the ionization of the outflowing gas. This could affect the line strengths (optical depths) by changing the column densities in particular ions. There is some evidence for this scenario in outflow lines that vary simultaneously across a wide range in velocities \cite[e.g.,][]{Hamann11,Capellupo12}. However, in Q1413+1143 the detection of a \mbox{P\,{\sc v}}\ BAL indicates that the \mbox{C\,{\sc iv}}\ BAL is very saturated. If the optical depth ratio is $\tau_{\mathrm{CIV}}$/$\tau_{\mathrm{PV}}$ $>$ 100 (Section 3.2 above), then the true optical depth across the variable portion of the \mbox{C\,{\sc iv}}\ trough is at least $\sim$70 (cf. Figs. 2 and 4). This is much too saturated to be susceptible to changes in the ionization. Similarly, \citet{Hamann08} and \citet{Capellupo13} identify other quasars where a high ratio of \mbox{Si\,{\sc iv}}\ to \mbox{C\,{\sc iv}}\ absorption strength indicates a saturated variable \mbox{C\,{\sc iv}}\ BAL, and they conclude that in these cases, the \mbox{C\,{\sc iv}}\ is likely too saturated to be affected by changes in the ionization of the gas.
Instead, the results presented here (and in \citealt{Hamann08} and \citealt{Capellupo13}) strongly support the scenario of clouds moving across our line-of-sight to the quasar continuum source as the cause of the observed variability (see also \citealt{Moe09}, \citealt{Leighly09}, \citealt{Hall11}, \citealt{Vivek12}, and refs. therein). A possible caveat to this interpretation might be that the absorbing region is inhomogeneous, such that it spans a wide range of column densities and line optical depths across our view of the continuum source. In that situation, changes in the continuum flux that change the absorbing region ionization could also change the projected area that has optical depth $\ga$1 in a given line (i.e., the observed covering fraction), without requiring tangential/crossing motions of the gas (see \citealt{Hamann11} and \citealt{Hamann12} for more discussion). However, that possibility seems to be contradicted here by our observations of 1) variability in only portions of the BAL troughs \citep[which is a common feature of BALs][]{Capellupo12,FilizAk13}, and 2) similar variabilities and observed covering fractions in all three lines of \mbox{P\,{\sc v}} , \mbox{Si\,{\sc iv}} , and \mbox{C\,{\sc iv}}\ that span a factor of $>$100 in optical depth.
We therefore adopt the moving cloud model and use the method described in \citet{Capellupo13} to estimate the location of the gas. This method adopts a simple scenario where a single outflow component of constant ionization and column density crosses our line of sight to the continuum source. The crossing speed we derive from the data depends on the geometry. \citet{Capellupo13} describes two models for the movement of outflow components across the continuum source (see their figure 14 for an illustration). First, a circular disc crosses a larger, circular continuum source along a path that continues through the centre of the background continuum source (the `crossing disc' model). In the second model, the absorber, which has a straight edge, moves across a square continuum source (the `knife edge' model). The `knife edge' model is less realistic, especially given the tendency for variability to occur in small portions of BAL troughs \citep{Gibson08,Capellupo11,FilizAk13}. In the `knife edge' model, the absorber continues to cover a greater percentage of the source over time, until it completely covers the source. This would eventually result in complete absorption at zero intensity. The `crossing disk' model is more realistic, especially in a scenario where separate outflow components, at different velocities, are crossing the background source, causing variability in different portions of the observed spectrum.
We therefore proceed with the `crossing disk' model to calculate the crossing speed of the outflow component. This speed is given by $D_{1500}\sqrt{\Delta A}$/$\Delta$$t$, where $D_{1500}$ is a characteristic diameter for the continuum source at 1500 \AA\ and $\Delta$$A$ is the change in the absorber line-of-sight covering fraction on a time-scale of $\Delta$$t$ \citep{Capellupo13}. For saturated lines like the \mbox{C\,{\sc iv}}\ BAL here, $\Delta A$ equals the change in absorption strength relative to a normalized continuum. We calculate a $D_{1500}$ of $\sim$0.004 pc using our estimated bolometric luminosity for Q1413+1143 of $L = 2.4 \times 10^{46}$ ergs s$^{-1}$ (\citealt{Peterson04,Bentz07,Gaskell08}; Hamann \& Simon, in preparation).
For $\Delta$$A$ and $\Delta$$t$, we use results from \citet{Capellupo13} and the current work.
We have more spectra that cover \mbox{C\,{\sc iv}}\ than \mbox{P\,{\sc v}}, and the shortest time-scale over which we detect \mbox{C\,{\sc iv}}\ variability is 0.25 yr. The portion of the \mbox{C\,{\sc iv}}\ BAL that varied had a change in absorption depth of $\Delta$$A$ $\sim$0.09 \citep{Capellupo13}. Since we only observed variability in \mbox{P\,{\sc v}}\ between the Lick 1989.26 and FOS 1994.98 spectra, the maximum variability time-scale for the \mbox{P\,{\sc v}}\ line during the period of our observations is 1.6 yr. The variability in \mbox{C\,{\sc iv}}\ on the shorter time-scale of 0.25 yr occurs over a much smaller portion of the BAL than on the longer time-scale of 1.6 yr. Furthermore, in general, variability over 1.6 yr is much more common than on time-scales of 0.25 yr \citep{Capellupo13}. Perhaps, the variability seen on the shorter time-scales is caused by smaller components in the flows at small distances, while the bulk of the flow, represented by the larger, more stable BAL feature, resides further out. We adopt the more conservative approach using 1.6 yr as the $\Delta$$t$ for determining the crossing speed of the outflow, and since the FOS data does not cover the \mbox{C\,{\sc iv}}\ absorption line, we adopt the value of $\Delta$$A$ $\sim$0.09. This gives a crossing speed of 750 km s$^{-1}$.
Continuing with the procedure of \citet{Capellupo13}, we assume the crossing speed is rougly equal to the Keplerian rotation speed around an SMBH, with a mass of $M_{BH} \sim 4.8\times10^{8}M_{\sun}$, to get the physical location of the gas. Using the crossing speed based on the `crossing discs' model and the time-scale of 1.6 yr, the distance is $\sim$3.5 pc from the central black hole. Since the variability may have occurred over a shorter time-scale, this distance is likely an upper limit.
\subsection{Energetics of the outflow}
\label{ener}
The energetics of quasar outflows, and thus their potential to affect the host galaxy and large-scale environments via feedback, is still poorly understood. To estimate the energetics, we first need to know the mass of the flow. If we estimate the flow geometry as part of a thin spherical shell, then the total mass is given by
\begin{equation}
M\ \approx\ 4 100 \, \left(\frac{Q}{15\%}\right) \left(\frac{N_{H}}{2 \times 10^{22}\,\mathrm{cm^{-2}}}
\right) \left(\frac{R}{3.5\,\mathrm{pc}}\right)^2 \ \ \mathrm{M_{\sun}},
\label{eqn:M}
\end{equation}
where $Q$ is the global covering fraction of the outflow, i.e. the fraction of 4$\pi$ steradians that the flow covers from the point-of-view of the central continuum source, $N_H$ is the total column density of the flow, and $R$ is the distance of the flow from the central SMBH \citep{Hamann00}. We adopt a value of $Q \sim$ 15\% based on the fraction of quasars that exhibit BALs in their spectra (\citealt{Hewett03}; \citealt{Reichard03b}; \citealt{Trump06}; \citealt{Knigge08}; \citealt{Gibson09}).
For most BAL quasars, it is difficult to constrain $N_H$ because few lines are measured and/or the broad profiles blend together and hide the doublet structure in well-studied BALs like \mbox{C\,{\sc iv}}. In most cases, the data yield only lower limits on the optical depths and the column densities of the outflowing gas. In this regard, the detection of a low-abundance line such as \mbox{P\,{\sc v}}\ is important because it allows us to estimate the true optical depth in the BALs and better constrain the total column density (see Section 1 and references therein). We adopt the value of $N_H \sim 2 \times 10^{22}$ cm$^{-2}$, based on the apparent optical depth we derive for \mbox{P\,{\sc v}}\ in Q1413+1143 and the photoionization models of \citet{Hamann98} and \citet{Leighly09} (Section 3.2).
The remaining variable is the radial distance, for which we adopt the value $R\sim 3.5$ pc derived above from the \mbox{C\,{\sc iv}}\ variations. Note that Eqn. \ref{eqn:M} provides only a crude approximation because the value of $N_H$ is a lower limit (Section 3.2) while the distance of 3.5 pc is likely an upper limit (Section 4.1). Nonetheless, we can see how changes to these parameters would alter the result. A firm upper limit on the total column in the \mbox{C\,{\sc iv}}\ BAL region is $N_H\la 10^{24}$ cm$^{-2}$ to avoid large electron scattering optical depths that would extinguish all of the near-UV flux. Outflows that are Compton thick are also much too mass loaded to be driven out by radiation pressure (see Hamann et al., in prep.). A lower limit on the radial distance is more difficult to establish, but at $R$ much smaller than a parsec, it becomes increasingly difficult to maintain the observed moderate degrees of ionization so close to the continuum source (\citealt{Hamann13}).
To estimate the mass-loss rates, $\dot{M}$, we divide the mass in Eqn. \ref{eqn:M} by a characteristic flow time, $t_{flow} \sim R/v$. We adopt a nominal value of 10 000 km s$^{-1}$\ for the velocity of the flow based on the velocity of the variable interval identified in Figs \ref{spectra} \& \ref{stack}. At a distance of 3.5 pc, $t_{flow}$ is $\sim$340 yr, and the mass-loss rate is therefore 12 M$_{\sun}$ yr$^{-1}$. For comparison, we estimate the black hole mass accretion rate with the equation $L = \eta\dot{M}_{\mathrm{acc}}c^2$, where $\eta$, the efficiency, is nominally 0.1 \citep{Peterson97}. For the luminosity, $L$, we use the bolometric luminosity of $\sim2.4 \times 10^{46}$ ergs s$^{-1}$ (Section \ref{data}), which gives a mass accretion rate, $\dot{M}_{\mathrm{acc}}$, of 4.2 M$_{\sun}$ yr$^{-1}$. Hence, the mass-loss rate through the outflow is roughly 3 times greater than the mass accretion rate onto the central black hole.
The kinetic energy of the outflow, as defined by $K = Mv^{2}/2$, is
\begin{equation}
K \approx\ 4 \times 10^{54} \, \left(\frac{M}{4100 \, \mathrm{M_{\sun}}}\right)
\left(\frac{v}{10\,000 \, \mathrm{km\,s^{-1}}}\right)^2 \, \mathrm{erg}.
\label{eqn:K}
\end{equation}
Dividing by the characteristic flow time, $t_{flow}$, gives a time-averaged kinetic luminosity of $\langle L_k\rangle \sim 4 \times 10^{44}$ ergs s$^{-1}$. The ratio of this kinetic energy luminosity to the quasar bolometric luminosity is therefore $\langle L_k\rangle/L \sim 0.02$. This is roughly a factor of 3 smaller than the ratio of $\langle L_k\rangle/L \sim 0.05$ that is typically cited as the ratio necessary for an outflow to be important for feedback (\citealt{Scannapieco04}; \citealt{DiMatteo05}; \citealt{Prochaska09}), but it exceeds the smaller ratio of $\langle L_k\rangle/L \sim$ 0.005 estimated theoretically by \citet{Hopkins10}. As mentioned above, our estimate of the column density is a lower limit, so if the column density is at least a factor of 2.5 greater, at the maximum distance of 3.5 pc, then the ratio of $\langle L_k\rangle/L$ would also exceed the larger threshold of 0.05. Therefore, the outflow might be important for feedback to the host galaxy.
Finally, we point out that the radial distances and kinetic energies derived here, and in other studies of outflow variability \citep{Misawa07,Capellupo11,Capellupo13,Hall11,Hamann11,Hamann13,RodriguezH11,RodriguezH13,Vivek12b}, are smaller than recent results that use excited-state lines to constrain the outflow densities and radial distances (via photoionization models, e.g., \citealt{Moe09,Dunn10, Borguet13, Arav13}). The excited-state absorber studies tend to find distances $\ga$1 kpc and kinetic energy luminosities that are important for feedback (sometimes by a large margin). It is not obvious how this apparent discrepancy between the distance results from variability and excited-state lines can be reconciled. It is possible that BAL outflows exist across a wide range of physical scales, perhaps manifesting themselves differently in different quasars. \citet{Moe09} argue that a variable component of one BAL outflow resides at distances $\la$0.1 pc while another component {\it in the same quasar} with excited-state lines is at $R\sim 3.3$ kpc. It will be an important test of all of these studies to look for variability in BALs known to have excited-state lines and, conversely, to search for excited-state lines to provide density constraints in BALs systems that vary.
\section{Conclusions}
Using the first reported detection of a variable \mbox{P\,{\sc v}}\ BAL, we constrain the mass and energetics of the outflow in Q1413+1143. The detection of \mbox{P\,{\sc v}}\ absorption, which indicates a very saturated \mbox{C\,{\sc iv}}\ BAL, and the similar variabilities in the \mbox{P\,{\sc v}}\ and \mbox{C\,{\sc iv}}\ BALs imply that the detected variability is caused by the outflow moving across our line-of-sight. This crossing clouds scenario provides a constraint on the distance of the outflow from the central SMBH, and the detection of the \mbox{P\,{\sc v}}\ BAL, along with photoionization models, constrains the true column density of the flow. Together, these quantities allow us to derive an estimate of the mass and energetics of the flow. Our upper limit for the distance of the flow and the lower limit on the column density gives $\langle L_k\rangle/L$ $\sim$ 0.02. This is between the ratio of 0.005 cited by \citet{Hopkins10} and 0.05 determined by earlier work \citep{Scannapieco04,DiMatteo05,Prochaska09} for an outflow to be sufficiently powerful to affect the host galaxy evolution. Therefore, the outflow in Q1413+1143 might be an important feedback mechanism to its host galaxy.
Using large samples of quasars from the ongoing SDSS-III BOSS survey, which includes many quasars at a large enough redshift to cover the \mbox{P\,{\sc v}}\ region, we can search for more detections of \mbox{P\,{\sc v}}\ absorption. We can also look for more cases of variable \mbox{P\,{\sc v}}\ absorption via repeat observations within SDSS-III and follow-up observations with other telescopes. Estimating column densities and distances for flows in more quasars will better able us to assess the typical strengths of these outflows and their viability as a feedback mechanism.
\section*{Acknowledgments}
We thank the referee for helpful comments on the manuscript.
FH acknowledges support from the USA National Science Foundation grant AST-1009628.
\bibliographystyle{mn2e}
|
\section{Introduction}
Multinucleon emission by electroweak probes is of much interest
nowadays~\cite{Gal11,For12,Mor12,Alv14}. Evidence of its presence in
the quasielastic (QE) peak region has been emphasized in the analysis
of recent neutrino and antineutrino scattering
experiments~\cite{Agu10,Agu13,Fio13,Abe13}. This has been confirmed
by theoretical calculations~\cite{Mar09,Nie11,Ama11,Lal12}, including
in the dynamics various nuclear effects such as meson-exchange
currents (MEC) with and without $\Delta$-isobar excitations,
final-state interactions (FSI), short-range correlations (SRC), the
random-phase approximation (RPA), effective interactions, {\it etc.}
These ingredients lead to discrepancies between the theoretical
predictions, and these need to be clarified in order to reduce the
systematic uncertainties in neutrino data
analyses~\cite{Gra13,Mar13b,Mar14,Ama12}.
The implementation of two-nucleon ejection in Monte Carlo (MC)
neutrino event generators requires an algorithm to generate events of
two-nucleon final states from given values of momentum and energy
transfer. The standard way to proceed, followed in
\cite{Sob12,GENIE,Katori13},
is to select two nucleons from the Fermi sea, invoke energy-momentum
conservation and compute the four-momentum of the final two-nucleon
state (selecting two nucleon momenta in the final state). In the CM
frame one assumes that
the two final nucleons move back-to-back with the same
given energy and opposite momentum. The emission angles are chosen
assuming an isotropic distribution in the CM. Once the final momenta
are given, a boost is performed to the Lab system to obtain the
momenta of the two ejected nucleons in this frame; these are then
further propagated in the MC cascade model.
We have recently studied the angular distribution in the Lab frame
corresponding to two-particle (2p) emission in the frozen nucleon
approximation~\cite{Rui14}, where the two nucleons are initially at
rest. This distribution appears in the phase-space integration of
the inclusive hadronic tensor in the 2p-2h channel. We found that
the angular distribution has singularities coming from the Jacobian
obtained by integration of the Dirac delta function of energy
conservation, where a denominator appears that can be zero for some
angles. This behaviour is due to the fact that for a fixed pair of
hole momenta ${\bf h}_1,~{\bf h}_2$, and for given momentum transfer, $q$,
and emission angle $\theta'_1$ of the first particle, there are two
solutions for the momentum of the ejected nucleon $p'_1$ that are
compatible with energy conservation. For a given value of the energy
transfer $\omega$, these two solutions collapse into only one for
the maximum allowed emission angle. For this angle there is a
minimum in the 2p-2h excitation energy, $E_{ex}$, as a function of
$p'_1$, and therefore the derivative that appears in the denominator
of the Jacobian is zero: $dE_{ex}/dp'_1=0$.
In \cite{Rui14} we showed that the divergence of the angular distribution
in the Lab system is of the type $\int_0^1 f(x) dx/\sqrt{x}$. Hence
it is integrable around zero, and we gave an analytic formula for
the integral around the divergence. The interest of the detailed study
of the angular integral was to reduce the CPU time in the calculation
of the hadronic tensor for inclusive neutrino scattering. Here a 7D
integral appears that has to be computed in a reasonable time in
order to use it to predict flux integrated neutrino cross sections,
where one additional integration is needed.
In this paper we show that the isotropic angular distribution in the
CM frame, as the one used in Monte Carlo generators~\cite{NuWro},
corresponds exactly to the angular distribution obtained by us in the
Lab system after integration of the Dirac delta function of
energy. Although this correspondence seems to be evident, in practice
it is not so obvious because in Monte Carlo generators no integration
of a delta function of energy is explicitly performed, or at least no
Jacobian is present in the algorithm to select the emission
angle~\cite{Sob12}. That means that the phase-space angular
distribution in the Monte Carlo codes is known except for a
normalization factor. Besides it was not evident earlier
why the divergence in
the angular distribution appears in the Lab system from a constant
distribution in the CM and how it can be handled by the Monte Carlo
procedure.
Furthermore, we also show that upon performing the phase-space integral
in the CM system one finds that the result is analytic if there is no Pauli
blocking, and we give a simple formula for it in the frozen nucleon
approximation. This integration method in the CM frame
provides an alternative way to compute the hadronic tensor
in neutrino and electron scattering.
The interest of the present study is directly linked to the
reliability of the frozen nucleon approximation to get sensible
results for intermediate to high momentum and energy transfers. This
was already applied to a preliminary evaluation of the hadronic tensor
in the case of the seagull current. Moreover the frozen nucleon
approximation is the leading term if the current is expanded in powers
of (h1,h2) around (0,0). An integral over the emission angle remains
to be performed. Under the assumption that the dependence of the
elementary hadronic tensor on the emission angle is soft, one could
factorize it out of the integral, evaluating it for some average
angle, say $(\theta_{Max}+\theta_{Min})/2$, times the phase-space
integral. In fact, the strong dependence of the electroweak matrix
elements comes from the $(q,\omega)$ dependence of the electroweak
form factor and not from the angular dependence for fixed
$(q,\omega)$. The validity of these assumptions will be verified in a
coming paper where the angular dependence of the elementary hadronic
tensor will be studied.
In Section II we present a detailed study of the general formalism
with explicit evaluation of the phase space and discussions on how to
perform explicitly the boost between the two reference frames, Lab and
CM. We introduce all of the variables required to analyze the 2p-2h
problem and make contact with the frozen nucleon approximation where
the calculations can be done in a straightforward way. Importantly, we
show that these ideas can be incorporated into fully relativistic
2p-2h analyses of neutrino reactions. In Section III we summarize our
basic findings and point out the main issues to be considered in
future work, {\it i.e.,} in any approach that attempts to take into
account two-nucleon ejection effects in lepton scattering reactions.
\section{Formalism}
\subsection{Lab frame}
The starting point is the 2p-2h hadronic tensor
for neutrino and electron scattering in the Lab system,
given in the Fermi gas by
\begin{eqnarray}
W^{\mu\nu}_{2p-2h}
&=&
\frac{V}{(2\pi)^9}\int
d^3p'_1
d^3h_1
d^3h_2
\frac{m_N^4}{E_1E_2E'_1E'_2}
\nonumber \\
&&
r^{\mu\nu}({\bf p}'_1,{\bf p}'_2,{\bf h}_1,{\bf h}_2)
\delta(E'_1+E'_2-E_1-E_2-\omega)
\nonumber\\
&&
\Theta(p'_1,p'_2,h_1,h_2) \,,
\label{hadronic2}
\end{eqnarray}
where $Q^\mu=(\omega,{\bf q})$ is the four momentum transfer,
$m_N$ is the nucleon mass, and $V$ is the volume of the system.
The four-momenta of the final particles and holes are
$P'_i=(E'_i,{\bf p}'_i)$, and $H_i=(E_i,{\bf h}_i)$, respectively.
Momentum conservation implies $\bf p'_2= h_1+h_2+q-p'_1$.
The initial Fermi gas ground state and Pauli blocking imply that
$h_i<k_F$, and $p'_i>k_F$. These conditions
are included in the $\Theta$ function,
defined as the product of step functions
\begin{eqnarray}
\Theta(p'_1,p'_2,h_1,h_2)
&=&
\theta(p'_2-k_F)
\theta(p'_1-k_F)
\nonumber\\
&\times&
\theta(k_F-h_1)
\theta(k_F-h_2) \,.
\end{eqnarray}
The function $r^{\mu\nu}({\bf p}'_1,{\bf p}'_2,{\bf h}_1,{\bf h}_2)$ is the
hadronic tensor for the elementary transition of a nucleon pair
with the given initial and final momenta,
summed over spin and isospin~\cite{Rui14}.
We choose the ${\bf q}$ direction to be along the $z$-axis. Then the above
integral is reduced to 7 dimensions. First there is a global rotational
symmetry over one of the azimuthal angles. We choose $\phi'_1=0$ and
multiply by a factor $2\pi$. Furthermore, the energy delta function
enables an analytic integration over $p'_1$.
This 7D integral has to be performed numerically~\cite{DePace03,Ama10a}.
Under some approximations~\cite{Donnelly:1978xa,Van80,Alb84,Gil97}
the number of dimensions can be further reduced,
but this cannot be done in the fully relativistic calculation.
In a previous paper~\cite{Rui14} we compared different methods to
evaluate the above integral numerically. In particular we studied the
special case of the phase-space function $F(q,\omega)$, obtained by
using a constant elementary tensor $r^{\mu\nu}=1$ (independent of the
kinematics), defined, except for a factor $V/(2\pi)^9$, as
\begin{eqnarray}
F(q,\omega)
&\equiv&
\int
d^3p'_1
d^3h_1
d^3h_2
\frac{m_N^4}{E_1E_2E'_1E'_2}
\nonumber \\
&&
\delta(E'_1+E'_2-E_1-E_2-\omega)
\Theta(p'_1,p'_2,h_1,h_2)
\nonumber\\
\label{phase}
\end{eqnarray}
with $\bf p'_2= h_1+h_2+q-p'_1$.
For fixed hole momenta, the energy of the two final particles is
\begin{equation}
E'=E'_1+E'_2=\sqrt{p'_1{}^2+m_N^2}+\sqrt{({\bf p}'-{\bf p}'_1)^2+m_N^2} \,,
\end{equation}
where
\begin{equation}
{\bf p}'={\bf h}_1+{\bf h}_2+{\bf q}
\end{equation}
is the final momentum of the pair.
For fixed emission angle $\theta'_1$, we integrate over $p'_1$
changing to the variable $E'$.
By differentiation we arrive at the following Jacobian
[note that the Jacobian of \cite{Lal12} agrees
with Eq. (\ref{jacobiano2})]
\begin{equation}
\left|\frac{dp'_1}{dE'}\right|
=\left| \frac{p'_1}{E'_1}-\frac{{\bf p}'_2\cdot\widehat{\bf p}'_1}{E'_2} \right|^{-1}
\label{jacobiano2}
\end{equation}
with $\widehat{\bf p}'_1\equiv {\bf p}'_1/p'_1$.
Now integration of the Dirac delta function of energy gives $E'=E_1+E_2+\omega$
and the phase-space function becomes
\begin{eqnarray}
F(q,\omega)
&=&
2\pi
\int
d^3h_1
d^3h_2
d\theta'_1\sin\theta'_1
\frac{m_N^4}{E_1E_2}
\label{integral7Drel}\\
&\times &
\sum_{\alpha=\pm}
\left.
\frac{p'_1{}^2}
{\left| \frac{p'_1}{E'_1}-\frac{{\bf p}'_2\cdot\widehat{\bf p}'_1}{E'_2} \right|}
\frac{ \Theta(p'_1,p'_2,h_1,h_2) }{ E'_1E'_2 }
\right|_{p'_1= p'_1{}^{(\alpha)}} \,,
\nonumber
\end{eqnarray}
where the sum inside the integral runs over the two solutions
$p'_1{}^{(\pm)}$ of the energy conservation equation which is
quadratic in $p'_1$. The explicit expressions of
the two solutions are given in~\cite{Rui14}.
In this paper we are interested in the angular dependence of the integrand.
We define the angular distribution function
for fixed values of $(q,\omega,{\bf h}_1, {\bf h}_2)$ as
\begin{eqnarray}
\Phi(\theta'_1)
&=&
\sin\theta'_1 \int p'_1{}^2 dp'_1 \delta(E_1+E_2+\omega-E'_1-E'_2)
\nonumber\\
&&\times
\Theta(p'_1,p'_2,h_1,h_2)
\frac{m_N^4}{E_1E_2E'_1E'_2}
\nonumber\\
&=&
\sum_{\alpha=\pm}
\left.
\frac{ m_N^4\sin\theta'_1 p'_1{}^2 \Theta(p'_1,p'_2,h_1,h_2) }
{ E_1E_2E'_1E'_2
\left| \frac{p'_1}{E'_1}-\frac{{\bf p}'_2\cdot\widehat{\bf p}'_1}{E'_2} \right|}
\right|_{p'_1= p'_1{}^{(\alpha)}}
\nonumber\\
&\equiv& \Phi_+(\theta'_1) + \Phi_-(\theta'_1) \, ,
\label{angular}
\end{eqnarray}
where $\Phi_{\pm}(\theta'_1)$ correspond to the two terms of the sum.
Once more ${\bf p}'_2={\bf h}_1+{\bf h}_2+{\bf q}-{\bf p}'_1$.
The function $\Phi(\theta'_1)$ thus measures the distribution of final
nucleons as a function of the angle $\theta'_1$. Note that this
function is computed analytically in the Lab system, given as a sum
over the two solutions of the energy conservation condition. Thus there are
really two distributions corresponding to the two possible energies of
final particles for a given emission angle. The angular distribution
is referred to the first particle. The second one is
determined by energy-momentum conservation.
In \cite{Rui14} it was shown that the angular distribution
in Eq.~(\ref{angular}) has divergences for some angles where the
denominator coming from the Jacobian is zero. Examples were given in
the frozen nucleon approximation. It was also shown that the
divergence is integrable, and an analytic formula was given for the
integral over $\theta'_1$ around the divergence. The integral in the
remaining intervals was performed numerically.
\subsection{Boost from the CM frame}
In Monte Carlo event generators the angular distribution is
obtained from an isotropic distribution in the CM frame, and then transformed
back to the Lab system. Here we show that our distribution is
recovered except for a normalization constant that we determine.
First we fix the kinematics of $(q,\omega,{\bf h}_1, {\bf h}_2)$. To simplify
our formalism, we consider the particular case of the frozen nucleon
approximation, {\it i.e.,} $h_1=h_2=0$. The general case can be done
similarly. The frozen nucleon approximation has the advantage that the total
final momentum is equal to ${\bf p}'= {\bf q}$ and hence the CM frame moves
in the z-direction (``upwards''). Therefore, the $x,y$ components are
invariant under the boost from the CM to the Lab frames. In
\cite{Rui14} it was shown that the frozen nucleon approximation gives an
accurate representation of the total phase-space function, so one
expects the angular distribution in the frozen nucleon approximation
to be representative of the general case.
Doubly-primed variables refer to the CM system.
The total final momentum is
\begin{equation}
{\bf p}''= {\bf p}_1''+{\bf p}_2''=0 \, ,
\end{equation}
and the total final
energy $E''$ is determined by invariance of the squared four momentum
\begin{equation}
E''= \sqrt{E'^2-{p'}^2} \, ,
\end{equation}
where $(E',{\bf p}')=(2m_N+\omega,{\bf q})$
are the final energy and momentum in the Lab frame.
In the CM frame the two final nucleons are assumed
to go back-to-back with the same momentum
and with the same energy
\begin{equation}
E''_1=E''_2= \frac{E''}{2} = \frac12 \sqrt{E'^2-{p'}^2} \, .
\end{equation}
The condition $E''_1>m_N$ restricts the allowed $(\omega,q)$ region
where the two-nucleon emission is possible.
Let $\theta''_1$ be the emission angle corresponding to the first
particle. To obtain the nucleon momentum in the Lab system we perform
a boost of the four vector $(P''_1)^\mu= (E''_1,{\bf p}''_1)$ back to the
Lab frame, that is moving downward the $z$-axis with dimensionless velocity $v$,
where this is the velocity of the CM system with respect to the Lab system,
given by
\begin{equation}
v= \frac{p'}{E'} \, .
\end{equation}
The boost transformation of the $(0,z)$ four-vector components
is given by a $2\times 2$ Lorentz matrix equation
\begin{equation}
\left(
\begin{array}{c} E'_1 \\ p'_{1z} \end{array}
\right)
=
\gamma \left(
\begin{array}{cc} 1 & v \\ v & 1 \end{array}
\right)
\left(
\begin{array}{c} E''_1 \\ p''_{1z} \end{array}
\right) \, ,
\end{equation}
where $\gamma\equiv 1/\sqrt{1-v^2}$. From here we get
\begin{eqnarray}
E'_1 &=& \gamma(E''_1+v p''_1\cos\theta''_1) \\
p'_{1}\cos\theta'_1 &=& \gamma(v E''_1+p''_1\cos\theta''_1) \, .
\end{eqnarray}
Therefore the momentum and angle in the Lab system are
\begin{eqnarray}
p'_1 &=& \sqrt{\gamma^2(E''_1+v p''_1\cos\theta''_1)^2-m_N^2}
\label{plab}\\
\cos\theta'_1 &=&
\frac{\gamma(v E''_1+p''_1\cos\theta''_1)}
{\sqrt{\gamma^2(E''_1+v p''_1\cos\theta''_1)^2-m_N^2}} \, .
\label{coslab}
\end{eqnarray}
\begin{figure}
\includegraphics[width=8cm, bb=130 240 430 780]{fig1.ps}\caption{
\label{fig1}
(Color online) Lab magnitudes as a function of CM magnitudes. The
momentum and energy transfer are $q=3$ GeV/c, and $\omega=2$ GeV. Top
panel: $\cos\theta'_1$ versus $\cos\theta''_1$. Middle panel:
$\theta'_1$ versus $\theta''_1$. Bottom panel: $p'_1$ versus
$\theta''_1$. }
\end{figure}
In Fig.~\ref{fig1} we show the Lab emission angle as a function of the
CM angle for momentum and energy transfers: $q=3$ GeV/c and $\omega=2$ GeV. We
choose in this case a high value of the momentum transfer to avoid
effects linked to Pauli blocking. The $\omega$ value is close to the QE
peak,
$\omega_{QE}=\sqrt{q^2+m_N^2}-m_N$,
and below it.
As the CM angle runs from 0 to 180 degrees, for
this kinematics the Lab angle starts growing, reaches a maximum and
then decreases. Therefore, for a given emission angle in the Lab system,
$\theta'_1$, there correspond two angles in the CM, that we denote
$(\theta''_1)^+$ and $(\theta''_1)^-$. They
differ in the value of the Lab momentum $p'_1$, that is plotted in the lower
panel of Fig.~1. Hence there are two different values of $p'_1$
for a given Lab angle. These two $p'_1$-values obviously
correspond to the two solutions, $(p'_1)^{\pm}$ of energy
conservation, appearing in the sum of the phase-space function in Eqs.~(\ref{integral7Drel},\ref{angular}). The momentum of the second
nucleon, $p'_2$, could be obtained by changing $\cos\theta''_1$ by
$(-\cos\theta''_1)$ in Eq.~(\ref{plab}). Therefore the range of
values it takes is the same as $p'_1$.
\subsection{Transformation of the angular distribution}
We assume that the angular distribution in the CM frame is independent of
the emission angle, except for Pauli blocking restictions,
\begin{equation}
n''(\theta''_1) = C \Theta(p'_1,p'_2,0,0) \, ,
\end{equation}
where $C$ is a constant that is determined below. The step function ensures
Pauli blocking. The angular
distribution in the Lab system, $n'(\theta'_1)$,
is obtained by imposing conservation of the
number of particles emitted within two corresponding solid angles $d\Omega'_1$
and $d\Omega''_1$, in the Lab and the CM systems
\begin{equation}
n'(\theta'_1)d\Omega'_1 = n''(\theta''_1)d\Omega''_1 \,.
\end{equation}
Since the boost conserves the azimuthal angle $d\phi''_1=d\phi'_1$, we
get the well-known transformation expression:
\begin{equation}
n'(\theta'_1)
=
\frac{C \Theta(p'_1,p'_2,0,0)}{\left|\frac{d\cos\theta'_1}{d\cos\theta''_1}\right|} \,.
\label{transformacion}
\end{equation}
The derivative in the Jacobian is computed by differentiation of
Eq.~(\ref{coslab}) with respect to $\cos\theta''_1$,
and can be written in the form
\begin{equation}
\frac{d\cos\theta'_1}{d\cos\theta''_1}
=
\gamma p''_1
\frac{p'_1-vE'_1\cos\theta'_1}{(p'_1)^2}
\end{equation}
Writing $\gamma$ in the form:
\begin{equation}
\gamma=\frac{E'}{\sqrt{E'{}^2-p'{}^2}} = \frac{E'}{2E''_1}
\end{equation}
we arrive at the following formula
for the angular distribution in the Lab frame
\begin{equation}
n'(\theta'_1)=
\frac{2E''_1}{E'p''_1}
\frac{(p'_1)^2}{|p'_1-vE'_1\cos\theta'_1|} C \Theta(p'_1,p'_2,0,0) \,.
\label{distribucion}
\end{equation}
Note that this distribution is not unique, because, as shown in
Fig.~1, there may be two different CM angles, and two different values
of $p'_1$ corresponding to the same Lab angle $\theta'_1$. Therefore
the are two possible angular distributions, and the total distribution
is given by the sum of the two,
\begin{equation}
n'(\theta'_1) =
n'_+(\theta'_1) +
n'_-(\theta'_1) \, ,
\end{equation}
where each partial distribution $n'_{\pm}(\theta'_1)$ corresponds to
Eq.~(\ref{distribucion}) using the $(p'_1)^{\pm}$ values, respectively.
\subsection{Equivalence of Lab distributions}
The next step is to compare the functions $n_{\pm}(\theta'_1)\sin\theta'_1$
with the angular distribution $\Phi_{\pm}(\theta'_1)$ computed for nucleons at
rest, $h_1=h_2=0$, given by Eq.~(\ref{angular})
\begin{equation}
\Phi_{\pm}(\theta'_1)
= \sin\theta'_1
\frac{m_N^2(p'_1)^2\Theta(p'_1,p'_2,0,0)}{|E'_2p'_1-E'_1{\bf p}'_2\cdot\widehat{\bf p}'_1|} \, ,
\label{phimasmenos}
\end{equation}
where $p'_1=(p'_1)^{\pm}$.
Using
\begin{equation}
{\bf p}'_2\cdot\widehat{\bf p}'_1 = q\cos\theta'_1-p'_1
\end{equation}
the denominator in Eq.~(\ref{phimasmenos}) can be written as
\begin{eqnarray}
E'_2p'_1-E'_1{\bf p}'_2\cdot\widehat{\bf p}'_1
&=&
E'p'_1-E'_1q\cos\theta'_1
\nonumber\\
&=&
E'(p'_1-E'_1 v\cos\theta'_1) \,.
\end{eqnarray}
Substituting in Eq.~(\ref{phimasmenos}) we obtain
\begin{equation}
\Phi_{\pm}(\theta'_1)
= \sin\theta'_1
\frac{m_N^2(p'_1)^2\Theta(p'_1,p'_2,0,0)
}{E'|p'_1-E'_1v\cos\theta'_1|} \,.
\end{equation}
Comparing with Eq.~(\ref{distribucion}), it follows that
\begin{equation}
n'_{\pm}(\theta'_1)\sin\theta'_1 = \Phi_{\pm}(\theta'_1)
\end{equation}
provided that
\begin{equation}
C = \frac{m_N^2}{2} \frac{p''_1}{E''_1} \,.
\end{equation}
\begin{figure}
\includegraphics[width=8cm, bb=130 240 430 780]{fig2.ps}\caption{
\label{fig2}
(Color online)
The two angular distributions $\Phi_{\pm}$ and the total,
in the Lab system, for two-nucleon emission in the frozen nucleon approximation.
The momentum transfer is
$q=3$ GeV/c and three values of $\omega=1800, 2000$
and 2200 GeV are considered.
}
\end{figure}
In Fig.~\ref{fig2} we show the two angular distributions
$\Phi_{\pm}(\theta'_1)$ for $q=3$ GeV/c and three values of
$\omega$. We can see that both distributions are zero above a maximum
allowed angle in the Lab system.
Both distributions
present a divergence (they are infinite) at that precise maximum
angle, because the derivative in the denominator of
Eq.~(\ref{transformacion}) is zero at that point. This is
in agreement with our previous work~\cite{Rui14} where we also
demonstrated that the divergence is integrable. The results of
Fig.~\ref{fig2} for the total distribution agree with the findings
of~\cite{Rui14}. In Fig.~\ref{fig2} we have not included Pauli
blocking in the plots of $\Phi_{\pm}$, but it is included in the total
distribution. We see that Pauli blocking only is effective in the last
case, $\omega=2200$ MeV, killing the divergence.
\subsection{Integration in the CM}
The method of the previous section can be reversed by making the
inverse boost from Lab to CM. This allows us to perform the integral over
$\theta'_1$ in Eq.~(\ref{integral7Drel}) using the CM emission angle,
by changing variables $\theta'_1\rightarrow\theta''_1$. Since this is
the inverse transformation applied in the previous sections, the
Jacobian cancels the denominator in Eq.~(\ref{integral7Drel}).
We start by fixing ${\bf h}_1$ and ${\bf h}_2$ and define
the phase-space integral over the final momenta
\begin{eqnarray}
G({\bf h}_1,{\bf h}_2,q,\omega)
&\equiv&
\int d^3p'_1d^3p'_2
\frac{m_N^2}{E'_1E'_2}
\Theta(p'_1,p'_2,h_1,h_2)
\nonumber\\
&&
\delta^4(H_1+H_2+Q-P'_1-P'_2) \,,
\end{eqnarray}
such that
\begin{equation}
F(q,\omega)=\int d^3h_1 d^3h_2 \frac{m_N^2}{E_1E_2}G({\bf h}_1,{\bf h}_2,q,\omega) \, .
\end{equation}
We recall from special relativity
that the integral measure
$\int d^3p/E$
is Lorentz invariant because of the result,
\begin{equation}
\int \frac{d^3p}{2E(p)}
=
\int d^4p \, \delta(p^\mu p_\mu -m_N^2)\theta(p^0) \,.
\end{equation}
Then we can write
\begin{eqnarray}
G({\bf h}_1,{\bf h}_2,q,\omega)
&=&
\int d^3p''_1d^3p''_2
\frac{m_N^2}{E''_1E''_2}
\Theta(p'_1,p'_2,h_1,h_2)
\nonumber\\
&&
\delta^4(H''_1+H''_2+Q''-P''_1-P''_2) \, ,
\end{eqnarray}
where the doubly-primed variables refer to the momenta in the CM frame. The CM is
defined by ${\bf p}''=({\bf h}_1+{\bf h}_2+{\bf q})''=0$. The step functions, which are not
invariant, must be computed in the Lab system, {\it i.e.,} the momenta
inside the integral have to be transformed back to the Lab system to
compute the argument of the step function. Integrating over ${\bf p}''_2$ we obtain
\begin{eqnarray}
G({\bf h}_1,{\bf h}_2,q,\omega)
&=&
\int d^3p''_1
\delta(E''-E''_1-E''_2)
\nonumber\\
&\times&
\frac{m_N^2}{E''_1E''_2}
\Theta(p'_1,p'_2,h_1,h_2)
\end{eqnarray}
with ${\bf p}''_2= -{\bf p}''_1$. Therefore, the CM energies satisfy the relationship
$E''_1=E''_2$, and we can write
\begin{eqnarray}
G({\bf h}_1,{\bf h}_2,q,\omega)
&=&
\int d^3p''_1
\delta(E''-2E''_1)
\nonumber\\
&\times&
\frac{m_N^2}{(E''_1)^2}
\Theta(p'_1,p'_2,h_1,h_2) \,.
\end{eqnarray}
Now we change variables $p''_1\rightarrow E''_1$, and integrate over $E''_1$
using $p''_1dp''_1=E''_1dE''_1$,
\begin{equation}
G({\bf h}_1,{\bf h}_2,q,\omega)
=
\frac{m_N^2}{2} \frac{p''_1}{E''_1}
\int d\Omega''_1
\Theta(p'_1,p'_2,h_1,h_2) \,.
\end{equation}
The remaining integral of the step function over the emission angles
is in general non-trivial and has to be performed numerically.
If there is no Pauli blocking, the above integral takes its maximum value:
\begin{equation}
G({\bf h}_1,{\bf h}_2,q,\omega)_{n.p.b}
=
4\pi \frac{m_N^2}{2} \frac{p''_1}{E''_1} \,.
\end{equation}
What remains to be performed is the integral over ${\bf h}_1,{\bf h}_2$, that in
general should be evaluated numerically. However, in
the frozen nucleon approximation one assumes that the integrand
depends very mildly on ${\bf h}_1,{\bf h}_2$, and therefore one can employ this fact to fix the kinematics to the frozen nucleon value, $h_1=h_2=0$.
The phase-space integral in this case is trivial, and takes on the value
\begin{equation}
F(q,\omega)_{n.p.b}
=
4\pi
\left( \frac{4}{3}\pi k_F^3 \right)^2
\frac{m_N^2}{2} \frac{p''_1}{E''_1} \,,
\label{analytical}
\end{equation}
where the ratio $p''_1/E''_1$ in the frozen nucleon approximation is given by
\begin{equation}
\frac{p''_1}{E''_1}
= \sqrt{1-\frac{4m_N^2}{(2m_N+\omega)^2-q^2}} \,.
\end{equation}
Note that in the asymptotic limit $\omega\rightarrow\infty$,
a constant value is obtained,
\begin{equation}
F(q,\infty)
=
4\pi
\left( \frac{4}{3}\pi k_F^3 \right)^2
\frac{m_N^2}{2} \, .
\end{equation}
This asymptotic limit is in agreement with the one obtained in~\cite{Rui14}
by integration in the Lab system.
\begin{figure}
\includegraphics[width=8cm, bb=130 590 430 780]{fig3.ps}
\caption{
\label{fig3}
(Color online) Phase-space function in the frozen nucleon approximation for
$q=3$ GeV/c, computed in the CM using the analytic formula without
Pauli blocking (n.p.b.), and computed numerically in the Lab system
including Pauli blocking (p.b.). }
\end{figure}
As an example, we show in Fig.~\ref{fig3} the phase-space function
$F(q,\omega)$ for $q=3$ GeV/c,
computed using the analytic formula without Pauli
blocking, Eq.~(\ref{analytical}), and by numerical integration in the
Lab frame using the method of \cite{Rui14} with Pauli blocking. Both
results agree except in the small region around the quasielastic peak,
where Pauli blocking produces the very small difference seen between the two results; there the Pauli-blocked function $F(q,\omega)$ is slightly below the analytic result.
\section{Conclusions and perspectives}
In this work we have analyzed the angular distribution of 2p-2h
final states in the relativistic Fermi gas, finding the connections
between the CM and Lab systems. Theoretical calculations of
many-particle emission in neutrino and electron scattering usually
rely on the Lab frame to be the most appropriate to perform the
calculations, since the Fermi gas state description is simpler,
mainly because Pauli blocking necessarily has to be checked in the
Lab system where the initial nucleons are below the Fermi surface.
However the description of the 2p angular distribution is simpler in
the CM frame, where the angular dependence is isotropic, if no Pauli
blocking is assumed.
On the contrary, the phase-space integral in the Lab system has the
difficulty that the angular distribution has a singularity at the
maximum allowed angle. The integration of this singularity in the Lab
system was made in our previous work~\cite{Rui14}. Here we have
studied the alternative method of performing the angular integral in
the CM frame, where the angular dependence is trivial. We show that
such an integral can be solved analytically in the absence of Pauli
blocking.
Of interest for the neutrino scattering data analysis, we have shown
that the algorithms used in Monte Carlo event generators produce 2p
angular distributions that are in agreement with the theoretical
calculations in the Lab system if the nuclear current is disregarded.
We have considered the angular distribution coming from phase
space alone. In a complete calculation one is involved with the
interaction between the two nucleons and the lepton that introduces an
additional angular dependence which needs to be evaluated to correctly
describe the events. A proper model of 2p-2h emission requires at
least the introduction of meson-exchange currents, or nuclear
correlations~\cite{DePace03,Ama10a}. Work along these lines is in
progress.
Finally, the integration method proposed here could also be used to
compute the 2p-2h hadronic tensor in Eq.~(\ref{hadronic2}) as an
alternative procedure to the common Lab frame calculations. Comparisons of
the two methods would be of interest because neither of them presents clear numerical
advantages. Although angular integration in
the CM frame allows one to avoid the divergence arising in the Lab frame, it
introduces the difficulty of having to perform a different boost
inside the integral for each pair of holes $({\bf h}_1,{\bf h}_2)$.
\section*{Acknowledgments}
This work was supported by DGI (Spain): FIS2011-24149 and
FIS2011-28738-C02-01, by the Junta de Andaluc\'{\i}a (FQM-225 and
FQM-160), by the Spanish Consolider-Ingenio 2010 programmed CPAN,
by U.S. Department of Energy under cooperative agreement
DE-FC02-94ER40818 (TWD) and by INFN under project MANYBODY (MBB).
C.A. is supported by a CPAN postdoctoral contract.
|
\section{Introduction}
Nonclassical states with non-Gaussian Wigner function (WF) have brought
great interest in quantum optics and quantum information science \cite{1}.
For example, a single-photon state with non-Gaussian behavior in phase space
has been found many applications in quantum information processing. In
particular, any single-mode nonclassical state has become a sufficient
resource to generate a two-mode entanglement via a beam-splitter \cite{1a}.
Recently, the non-Gaussian states have attracted more attention of both
experimentalists and theoreticians \cite{2,3,4,5,6}. It is possible to
generate and manipulate various non-Gaussian states through subtracting or
adding photon operation or photon subtraction-addition coherent
superposition operation on traditional quantum states or Gaussian states
\cite{7}. For example, the photon subtraction transforms a Gaussian
entangled state (two-mode squeezed state) to a non-Gaussian entangled state
for a nonlocality test \cite{8} and entanglement distillation \cite{9}. The
photon addition can also transform a classical state to a nonclassical state
\cite{10}. In laboratory the operation of photon subtraction or addition is
now realized practically \cite{11,12}. In Ref. \cite{13}, Lee \textit{et al.}
consider a coherent superposition of photon subtraction and addition,
ta+ra^{\dag }$, acting on a coherent state and a thermal state to form
non-Gaussian states, and propose the experimental scheme to implement this
elementary coherent operation. Furthermore, other non-Gaussian state is
obtained theoretically through $m$ times coherent superposition of photon
subtraction and addition, $\left( ta+ra^{\dag }\right) ^{m}$, acting on
thermal state \cite{13a} and coherent state \cite{13b}, respectively.
On the other hand, as a kind of nonclassical state, the so-called Schr\H{o
dinger cat states (SCS, quantum superpositions of coherent states \cite{14
), play an important role in fundamental tests of quantum theory \cite{15,16}
and in many quantum information processing tasks, including quantum
computation \cite{17}, quantum teleportation \cite{18} and precision
measurements \cite{19,20}. There have been a great deal of theoretical and
experimental attempts to generate a Schr\H{o}dinger-cat-type state and
considerable experimental progresses have been achieved in recent years \cit
{21,22,23,24,25}. Such as in Ref. \cite{25} a Schr\H{o}dinger-cat-like state
is generated via a coherent superposition of photonic operations.
Thus an interesting question is naturally raised: can we operate the
coherent superposition operator $\left( a\cos \theta +a^{\dagger
}e^{i\varphi }\sin \theta \right) ^{m}$ on odd-Schr\H{o}dinger-cat state
(OSCS) to construct a new\ non-Gaussian quantum state? The answer is
definite. Considering the above reasons, we shall construct a new
nonclassical state (MCSO-OSCS) which is supposed to be realized in
experiment. In this paper, We focus on studying its nonclassical properties
of this state by deriving analytically some expressions, such as normalized
constant, sub-Poissonian statistics, photocount distribution and Wigner
function. In fact, systems are usually surrounded by a thermal reservoir,
and\ decoherence becomes an important topic in the fields of quantum optics.
Enlightened by these ideas, we shall also discuss its decoherence property
in a thermal environment in this paper.
The paper is organized as follows. In Sec. 2, the MCSO-OSCS is constructed
and its normalized constant turns out to be related with the Hermite
polynomial. In Sec. 3, the fidelity between MCSO-OSCS and its original state
(OSCS) shall be obtain. In Sec. 4, the nonclassical properties of the
MCSO-OSCS, such as sub-Poissonian statistics, quadrature squeezing
properties and photocount distribution are calculated analytically and then
discussed in details. In Sec.5, the explicitly analytical expression of WF
for the MCSO-OSCS is derived. According to the negativity of WF, the
nonclassical properties are also discussed in details. In Sec. 6, the
decoherence of the MCSO-OSCS in a thermal environment\ is investigated. In
Sec. 7, we end our work with main conclusions.
\section{Normalization of the MCSO-OSCS}
Theoretically, the MCSO-OSCS can be introduced by repeated application of
coherent superposition operator $\Omega $ to the OSCS ($\left \vert \alpha
_{0}\right \rangle -\left \vert -\alpha _{0}\right \rangle $) for $m$ times,
i.e.
\begin{equation}
\left \vert \psi _{m}\right \rangle =\Omega ^{m}\left( \left \vert \alpha
_{0}\right \rangle -\left \vert -\alpha _{0}\right \rangle \right) ,
\label{1}
\end{equation
\ \ where $\Omega =a\cos \theta +a^{\dagger }e^{i\varphi }\sin \theta $ with
$\left[ a,a^{\dagger }\right] =1$ and $\theta \in \left( 0,\pi /2\right) $,
m$ is the order of coherent superposition operator (a non-negative integer),
$\left \vert \alpha _{0}\right \rangle $ is a coherent state of amplitude
\left \vert \alpha _{0}\right \vert $. The density operator of the MCSO-OSCS
is $\rho _{m}=N_{m}^{-1}\left \vert \psi _{m}\right \rangle \left \langle
\psi _{m}\right \vert $, where $N_{m}$ is a normalized constant of the
MCSO-OSCS to be determined by $\mathtt{Tr}\rho _{m}=1$. If $\Omega ^{m}$
operates on the even SCS$\ \left( \left \vert \alpha _{0}\right \rangle
+\left \vert -\alpha _{0}\right \rangle \right) $, then we obtain the
MCSO-ESCS. We only discuss the properties of MCSO-OSCS in this paper, for an
odd SCS in general show stronger nonclassical properties than an even SCS
\cite{25a}.
In order to obtain the normalized constant $N_{m}$, and note that the
operator $\Omega $ is not always Hermitian due to $\Omega \neq \Omega
^{\dagger }$ when $\cos \theta \neq e^{i\varphi }\sin \theta $, we shall
derive the normal ordering form of $\Omega ^{m}$ firstly. Recalling the
generating function of the Hermite polynomial $H_{m}(x)$ \cite{26}, i.e.
\sum \limits_{m=0}^{\infty }\frac{t^{m}}{m!}H_{m}(x)=\exp \left(
2xt-t^{2}\right) ,$ with
\begin{equation}
H_{m}(x)=\sum \limits_{m=0}^{\left[ m/2\right] }\frac{\left( -1\right)
^{l}m!\left( 2x\right) ^{m-2l}}{l!\left( m-2l\right) !}=\frac{\partial ^{m}}
\partial t^{m}}\left. \exp \left( 2xt-t^{2}\right) \right \vert _{t=0},
\label{2}
\end{equation
using the Baker-Hausdorff formula $e^{A+B}=e^{A}e^{B}e^{-\frac{1}{2
[A,B]}=e^{B}e^{A}e^{-\frac{1}{2}[B,A]}$ \cite{27} and the technique of
integration within an ordered product (IWOP) of operators \cite{28}, we hav
\begin{eqnarray}
e^{\lambda \Omega } &=&\colon e^{\lambda \Omega +\frac{1}{2}\lambda
^{2}e^{i\varphi }\sin \theta \cos \theta }\colon \notag \\
&=&\sum \limits_{m=0}^{\infty }\frac{\lambda ^{m}\left( -i\sqrt{\frac{1}{2
e^{i\varphi }\sin \theta \cos \theta }\right) ^{m}}{m!}\colon H_{m}(\frac
i\Omega }{\sqrt{2e^{i\varphi }\sin \theta \cos \theta }})\colon , \label{3}
\end{eqnarray
where the symbol $\colon \colon $stands for the normally ordering. Comparing
Eq.(\ref{3}) with the expansion of $e^{\lambda \Omega }$, i.e. $e^{\lambda
\Omega }=\sum \limits_{m=0}^{\infty }\frac{\lambda ^{m}}{m!}\Omega ^{m}$, we
can easily obtain the normal ordering form of $\Omega ^{m}$:
\begin{equation}
\Omega ^{m}=\left( -i\sqrt{\frac{1}{2}e^{i\varphi }\sin \theta \cos \theta
\right) ^{m}\colon H_{m}(\frac{i\Omega }{\sqrt{2e^{i\varphi }\sin \theta
\cos \theta }})\colon . \label{4}
\end{equation}
Similarly, $\Omega ^{\dagger m}=\left( a^{\dagger }\cos \theta
+ae^{-i\varphi }\sin \theta \right) ^{m}$ has the normal ordering form as
follows:
\begin{equation}
\Omega ^{\dagger m}=\left( -i\sqrt{\frac{1}{2}e^{-i\varphi }\sin \theta \cos
\theta }\right) ^{m}\colon H_{m}(\frac{i\Omega ^{\dagger }}{\sqrt
2e^{-i\varphi }\sin \theta \cos \theta }})\colon . \label{5}
\end{equation
From Eq. (\ref{4}) and (\ref{5}), we also give the following relations
\begin{equation}
\left \langle \beta \right \vert \Omega ^{m}\left \vert \alpha \right
\rangle =\left( -i\sqrt{\frac{1}{2}e^{i\varphi }\sin \theta \cos \theta
\right) ^{m}H_{m}(\frac{i\left( \alpha \cos \theta +\beta ^{\ast
}e^{i\varphi }\sin \theta \right) }{\sqrt{2e^{i\varphi }\sin \theta \cos
\theta }})\left \langle \beta \right \vert \left. \alpha \right \rangle ,
\label{6}
\end{equation
and
\begin{equation}
\left \langle \beta \right \vert \Omega ^{\dagger m}\left \vert \alpha
\right \rangle =\left( -i\sqrt{\frac{1}{2}e^{-i\varphi }\sin \theta \cos
\theta }\right) ^{m}H_{m}(\frac{i\left( \beta ^{\ast }\cos \theta +\alpha
e^{-i\varphi }\sin \theta \right) }{\sqrt{2e^{-i\varphi }\sin \theta \cos
\theta }})\left \langle \beta \right \vert \left. \alpha \right \rangle ,
\label{7}
\end{equation
where $\left \vert \alpha \right \rangle $ and $\left \vert \beta
\right
\rangle $ are coherent states and $\left \langle \beta \right \vert
\left. \alpha \right \rangle =\exp \left[ -\frac{1}{2}\left( \left \vert
\alpha \right \vert ^{2}+\left \vert \beta \right \vert ^{2}\right) +\beta
^{\ast }\alpha \right] $ \cite{29,30}. Eq. (\ref{6}) and (\ref{7}) are very
useful in the following calculations.
Next, according to Tr$\rho _{m}=1$, we obtai
\begin{equation}
N_{m}=\mathtt{Tr}\left[ \left \vert \psi _{m}\right \rangle \left \langle
\psi _{m}\right \vert \right] =\left \langle \psi _{m}\right \vert \left.
\psi _{m}\right \rangle . \label{8}
\end{equation
Substituting Eq. (\ref{1}), (\ref{4})-(\ref{5}) into (\ref{8}), and
inserting the completeness\ relation of the coherent state $\int \frac{d^{2}
}{\pi }\left \vert z\right \rangle \left \langle z\right \vert =1,$
furthermore, with the help of Eq. (\ref{2}) and the following integral
formula
\begin{eqnarray}
&&\int \frac{d^{2}z}{\pi }\exp \left( \zeta \left \vert z\right \vert
^{2}+\xi z+\eta z^{\ast }+fz^{2}+gz^{\ast 2}\right) \notag \\
&=&\frac{1}{\sqrt{\zeta ^{2}-4fg}}\exp \left( \frac{-\zeta \xi \eta +\xi
^{2}g+\eta ^{2}f}{\zeta ^{2}-4fg}\right) , \label{9}
\end{eqnarray
whose convergent condition is Re$(\zeta \pm f\pm g)<0$, Re$[(\zeta
^{2}-4fg)/(\zeta \pm f\pm g)]<0$, we obtai
\begin{equation}
N_{m}=2\chi ^{m}\left[ \sum \limits_{k=0}^{m}\left( -1\right) ^{m}A\left
\vert H_{m-k}\left( B\right) \right \vert ^{2}-\sum \limits_{k=0}^{m}\left(
-1\right) ^{k}A\left \vert H_{m-k}\left( C\right) \right \vert ^{2}\exp
\left( -2\left \vert \alpha _{0}\right \vert ^{2}\right) \right] ,
\label{10}
\end{equation
\ where we have set
\begin{eqnarray}
A &=&\left( 2\frac{\sin \theta }{\cos \theta }\right) ^{k}\frac{1}{k!}\left(
\frac{m!}{\left( m-k\right) !}\right) ^{2}, \notag \\
B &=&i\sqrt{\frac{e^{-i\varphi }\sin \theta }{2\cos \theta }}\alpha _{0}
\frac{i\sqrt{e^{i\varphi }\cos \theta }}{\sqrt{2\sin \theta }}\alpha
_{0}^{\ast }, \notag \\
C &=&\frac{i\sqrt{e^{-i\varphi }\cos \theta }}{\sqrt{2\sin \theta }}\alpha
_{0}-\frac{i\sqrt{e^{i\varphi }\sin \theta }}{\sqrt{2\cos \theta }}\alpha
_{0}^{\ast }, \notag \\
\chi &=&-\frac{1}{2}\sin \theta \cos \theta , \label{11}
\end{eqnarray
and we have used the recurrence relation of $H_{m}\left( x\right) $
\begin{equation}
\frac{\partial }{\partial x^{l}}H_{m}\left( x\right) =\frac{2^{l}m!}{\left(
m-l\right) !}H_{m-l}\left( x\right) . \label{12}
\end{equation
Eq.(\ref{10}) indicates that the normalization factor $N_{m}$ is just
related to a Hermite polynomial. Obviously, when $m=0$, the MCSO-OSCS just
reduces to the odd SCS. The analytical expression of $N_{m}$ is important
for further investigating the properties of MCSO-OSCS. For MCSO-ESCS, we can
change the negative sign "$-$" before the second sign of sum in Eq.((\ref{10
) to the positive sign "$+$" and obtain its normalized constant.
\section{Fidelity between MCSO-OSCS and OSCS}
In quantum teleportation, the fidelity $F$, which measures how close the
teleported state is to the original state, is the projection of the original
pure state $\left \vert \Psi _{in}\right \rangle $ of the density operator
\rho _{in}=\left \vert \Psi _{in}\right \rangle \left \langle \Psi
_{in}\right \vert $ onto the teleported state $\left \vert \Psi
_{out}\right
\rangle $ of the density operator $\rho _{out}:F=$ Tr$\left(
\rho _{out}\rho _{in}\right) $ \cite{31,32}. Here the fidelity measures how
close the new state (MCSO-OSCS) is to the original state (OSCS). The
fidelity between MCSO-OSCS (density matrix is $\rho _{m}$) and its original
OSCS ($\rho _{o}$) is defined as \cite{7}
\begin{equation}
F_{\theta ,\varphi ,m}=\frac{\text{Tr}\left( \rho _{m}\rho _{o}\right) }
\text{Tr}\left( \rho _{o}^{2}\right) }. \label{13}
\end{equation
In general, $0\leq F\leq 1$, and $F=1$ shows that the two states are same,
while $F=0$ shows that the two states are anamorphic absolutely. Employing
the similar procedure of deriving the normalization constant, the fidelity
for MCSO-OSCS can be calculated out a
\begin{equation}
F_{\theta ,\varphi ,m}=\frac{N_{m}^{-1}\chi ^{m}}{4\left( 1-e^{-2\left \vert
\alpha _{0}\right \vert ^{2}}\right) ^{2}}\left \vert \left[ \left(
-1\right) ^{m}+1\right] \left[ H_{m}(B^{\ast })-e^{-2\left \vert \alpha
_{0}\right \vert ^{2}}H_{m}(C)\right] \right \vert ^{2}. \label{14}
\end{equation
\begin{figure}[tbp]
\label{Fig1} \centering \includegraphics[width=0.8\columnwidth]{F1.eps}
\caption{Fidelity between MCSO-OSCS and OSCS as a function of $\protec
\alpha _{0}$ (here $\protect \alpha _{0}$ is setted as real number) (a)\
\protect \theta =\frac{\protect \pi }{3},$ $\protect \varphi =0,$ with
different $m$ values; (b) $m=2,$ $\protect \varphi =0,$ with different
\protect \theta $ values; (c) $m=2$, $\protect \theta =\frac{\protect \pi }{3},$
with different $\protect \varphi $ values. }
\end{figure}
In particular, when $m=0$, leading to $\rho _{m}=\rho _{o}$, then Eq. (\re
{13}) reduces to $F_{\theta ,\varphi ,m}=1$ (see Fig.1(a)), which indicates
that the MCSO-OSCS is reduced to the OSCS, as expected. In Eq. (\ref{14})
the fidelity $F_{\theta ,\varphi ,m}$ is equal to $0$ as $m$ is an odd
number because of the term $\left( \left( -1\right) ^{m}+1\right) $. In
Fig.1(a),\ we plot the fidelity $F_{\theta ,\varphi ,m}$ as the function of
\alpha _{0}$ for some different $m$ values with some given $\theta ,\varphi $
values, here $\alpha _{0}$ is setted as a real number, the same as in Figs.
2, 3, 8. It is obvious to note that $F_{\theta ,\varphi ,m}=0$ when $m=1,3$,
as expected, and $F_{\theta ,\varphi ,m}\neq 0$\ when $m=2,4$. When $m$ is
an even number, the fidelity increases monotonously with the increment of
\alpha _{0}$ and tends to $1$ finally, which indicates that the coherent
superposition operation has no influence on the filed when the field is
strong enough. Comparing with the curves of $m=0,2,4$, we find that the
smaller the value of $m$ is, the bigger the fidelity is. In order to see the
effect of different $\theta $ values on the fidelity, we plot the fidelity
as the function of $\alpha _{0}$ for some different $\theta $ values and
given $m,\varphi $\ values, see Fig.1(b). It is shown that the fidelity
decreases as $\theta $ increases. In addition, we study the relation of the
fidelity and parameter $\varphi $ through the plot of Fig.1(c). It is shown
that the values of parameter $\varphi $ have little influence on the
fidelity. It is also shown that the fidelity increases as the amplitude
\alpha _{0}$ increases from Fig.1(b) and 1(c).
\section{Nonclassical properties of MCSO-OSCS}
In this section, we shall discuss the nonclassical properties of the
MCSO-OSCS in terms of sub-Posissonian statistics, quadrature squeezing
properties and the negativity of its Wigner function.
\subsection{Mandel's Q-parameter}
The Mandel's Q-parameter measures the deviation of the variance of the
photon number distribution of the field state under consideration from the
Poissonian distribution of the coherent state, which has been defined as
\cite{34
\begin{equation}
Q=\frac{\left \langle a^{\dagger 2}a^{2}\right \rangle }{\left \langle
a^{\dagger }a\right \rangle }-\left \langle a^{\dagger }a\right \rangle .
\label{15}
\end{equation
The quantum states has the Poissonian, sub-Poissonian and super-Poissonian
statistics for $Q=0,Q<0$ and $Q>0$, respectively. It is well known that the
negativity of $Q$-parameter refers to the nonclassical character of the
state, but a state may be nonclassical even though $Q$-parameter is positive
as pointed out in \cite{35}.
Using Eqs.(\ref{4}), (\ref{5}), $\rho _{m}=N_{m}^{-1}\left \vert \psi
_{m}\right \rangle \left \langle \psi _{m}\right \vert $ and IWOP technique
of operators, one can calculate $\left \langle a^{\dagger }a\right \rangle $
as
\begin{eqnarray}
\left \langle a^{\dagger }a\right \rangle &=&\text{Tr}\left( \rho
_{m}a^{\dagger }a\right) \notag \\
&=&N_{m}^{-1}\chi ^{m}\sum \limits_{k=0}^{m}I\left[
\begin{array}{c}
+\left( -1\right) ^{m-k}\left( k+1+\left \vert \alpha _{0}\right \vert
^{2}\right) \left \vert H_{m-k}\left( B\right) \right \vert ^{2} \\
-\left( k+1-\left \vert \alpha _{0}\right \vert ^{2}\right) \left \vert
H_{m-k}\left( C\right) \right \vert ^{2}e^{-2\left \vert \alpha _{0}\right
\vert ^{2}} \\
+2\func{Re}\left[ R^{\ast }\alpha _{0}^{\ast }\left( m-k\right)
H_{m-k}\left( -B\right) H_{m-k-1}\left( B^{\ast }\right) \right] \\
-2\func{Re}\left[ R^{\ast }\alpha _{0}^{\ast }\left( m-k\right)
H_{m-k}\left( C^{\ast }\right) H_{m-k-1}\left( C\right) e^{-2\left \vert
\alpha _{0}\right \vert ^{2}}\right
\end{array
\right] -1, \label{16}
\end{eqnarray
wher
\begin{eqnarray}
R &=&\frac{i\sqrt{2e^{-i\varphi }\sin \theta }}{\sqrt{\cos \theta }}, \notag
\\
I &=&\frac{2}{k!}\left( \frac{m!}{\left( m-k\right) !}\right) ^{2}\left(
-\left \vert R\right \vert ^{2}\right) ^{k}, \label{17}
\end{eqnarray
and get the value of $\left \langle a^{2}a^{\dagger 2}\right \rangle $ a
\begin{equation}
\left \langle a^{2}a^{\dagger 2}\right \rangle =f_{1}\left( \alpha
_{0}\right) +f_{1}\left( -\alpha _{0}\right) -f_{2}\left( \alpha _{0}\right)
-f_{2}\left( -\alpha _{0}\right) , \label{18}
\end{equation
where
\begin{eqnarray}
f_{1}\left( \alpha _{0}\right) &=&N_{m}^{-1}\chi ^{m}\left. \frac{\partial
^{2m+2}}{\partial t^{m}\partial s^{m}\partial \lambda \partial \eta }\frac
e^{-\left \vert \alpha _{0}\right \vert ^{2}-K^{\ast }t+Ks-s^{2}-t^{2}}}
\sqrt{1-4\lambda \eta }}e^{\left( H+H_{0}+\left \vert \alpha _{0}\right
\vert ^{2}\right) /\left( 1-4\lambda \eta \right) }\right \vert
_{s=t=\lambda =\eta =0}, \notag \\
f_{2}\left( \alpha _{0}\right) &=&N_{m}^{-1}\chi ^{m}\left. \frac{\partial
^{2m+2}}{\partial t^{m}\partial s^{m}\partial \lambda \partial \eta }\frac
e^{-\left \vert \alpha _{0}\right \vert ^{2}-K^{\ast }t-Ks-s^{2}-t^{2}}}
\sqrt{1-4\lambda \eta }}e^{\left( H-L_{0}-\left \vert \alpha _{0}\right
\vert ^{2}\right) /\left( 1-4\lambda \eta \right) }\right \vert
_{s=t=\lambda =\eta =0}, \label{19}
\end{eqnarray
an
\begin{eqnarray}
K &=&i\frac{\sqrt{2e^{-i\varphi }\cos \theta }}{\sqrt{\sin \theta }}\alpha
_{0}, \notag \\
H &=&-\left \vert R\right \vert ^{2}ts+R^{2}t^{2}\eta +\alpha _{0}^{\ast
2}\eta +\alpha _{0}^{2}\lambda +R^{\ast 2}s^{2}\lambda , \notag \\
H_{0} &=&R\alpha _{0}t-R^{\ast }\alpha _{0}^{\ast }s+2R\alpha _{0}^{\ast
}t\eta -2R^{\ast }\alpha _{0}\lambda s, \notag \\
L_{0} &=&R\alpha _{0}t+R^{\ast }\alpha _{0}^{\ast }s-2R\alpha _{0}^{\ast
}t\eta -2R^{\ast }\alpha _{0}\lambda s. \label{20}
\end{eqnarray
Here $s,t,\lambda ,\eta $ are parameters introduced into the calculation
process and will be eliminated after finishing the calculation by setting
them to zero. Furthermore, one can use the relation $\left[ a,a^{\dagger
\right] =1$ to obtain
\begin{equation}
\left \langle a^{\dagger 2}a^{2}\right \rangle =\left \langle
a^{2}a^{\dagger 2}\right \rangle -4\left \langle a^{\dagger }a\right \rangle
-2. \label{21}
\end{equation
\begin{figure}[tbp]
\label{Fig2} \centering \includegraphics[width=0.8\columnwidth]{F2.eps}
\caption{Mandel's Q-parameter of MCSO-OSCS as a function of $\protect \alpha
_{0}$ (a)\ $\protect \theta =\frac{\protect \pi }{4}$, $\protect \varphi =0,$
with different $m$ values; (b) $\protect \theta =\frac{\protect \pi }{8}
\protect \varphi =0,$ with different $m$ values; (c) $m=1,$ $\protect \varphi
=0,$with different $\protect \theta $ values; (d) $m=1,$ $\protect \theta
\frac{\protect \pi }{4},$with $\protect \varphi =\frac{\protect \pi }{3},\frac
\protect \pi }{4},\frac{\protect \pi }{6},0$ (from upper to lower curves). }
\end{figure}
Substituting Eqs. (\ref{16}), and (\ref{21}) into (\ref{15}), and using the
method of numerical calculation we can study the property of Mandel's
Q-parameter for the MCSO-OSCS. The Q-parameters of MCSO-OSCS as the function
of $\alpha _{0}$ are depicted in Fig. 2 for several different values of $m$,
$\theta $ and $\varphi $. It is interesting to note that the values of
Q-parameter are always smaller than zero for different $m$ values under the
given $\theta =\frac{\pi }{4}$ and $\varphi =0$ (see Fig. 2(a)), which
indicates sub-Poissonian statistics. In addition, the absolute value of
Q-parameter decreases with the increment of $\alpha _{0}$ till tends to
zero, which indicates that all states under different $m$ values will tend
to the Poissonian statistics (the distribution of a coherent state) when the
value of $\alpha _{0}$ is big enough. However, we can see that the range of
the Q-parameter is $\left[ -1,0.5\right] $ in Fig. 2(b)\ with different $m$
values and for given $\theta =\frac{\pi }{8}$ and $\varphi =0$. That
indicates the\ MCSO-OSCS with small value of $\theta $ may do not exhibit
sub-Poissonian statistics but exhibit super-Poissonian statistics.
In particular, when $m=1$ and $\varphi =0$, the MCSO-OSCS deduces to the
COSCS ($\Omega \left( \left \vert \alpha _{0}\right \rangle -\left \vert
-\alpha _{0}\right \rangle \right) $) \cite{25a}. From Fig. 2(c), We can see
that the range of the Q-parameter is $\left[ -1,0.7\right] $. From the
criteria of Q-parameter, one finds that the MCSO-OSCS exhibits the
sub-Poissonian statistics for $\theta =\frac{\pi }{4}$, $\frac{\pi }{3}$ and
$\frac{\pi }{2.1}$. In the area of $\theta \in (0,\frac{\pi }{2})$, the
bigger the value of $\theta $ is, the more chance the state exhibits the
sub-Poissonian statistics. The similar conclusion can also be seen in Ref.
\cite{25a}. In addition, from Fig. 2(d) we can see that the absolute value
of Q-parameter decreases as $\varphi $\ increases, but this difference is
not obvious.
\subsection{Quadrature squeezing properties of MCSO-OSCS}
One observes nonclassical effects not only through sub-Poissonian statistics
but also through squeezing effects, which do not allow classical
interpretation of photoelectric counting events. Here, we consider an
appropriate quadrature operator $X_{\theta }=ae^{-i\theta }+a^{\dagger
}e^{i\theta }$, and the squeezing can be characterized by $\left \langle
\left( \Delta X_{\theta }\right) ^{2}\right \rangle _{\min }<1$ with respect
to angle $\theta $, or by the normal ordering form $\left \langle \colon
\left( \Delta X_{\theta }\right) ^{2}\colon \right \rangle _{\min }<0$ \cit
{39}. Upon expanding the terms in $\left \langle \colon \left( \Delta
X_{\theta }\right) ^{2}\colon \right \rangle _{\min }$, one can minimize its
value over the whole angle $\theta $, which is given by \cite{40}
\begin{equation}
S=\left \langle \colon \left( \Delta X_{\theta }\right) ^{2}\colon \right
\rangle _{\min }=-2\left \vert \left \langle a^{\dagger 2}\right \rangle
-\left \langle a^{\dagger }\right \rangle ^{2}\right \vert +2\left \langle
a^{\dagger }a\right \rangle -2\left \vert \left \langle a^{\dagger }\right
\rangle \right \vert ^{2}. \label{23}
\end{equation
Then its negative value in the range $\left[ -1,0\right) $ indicates
squeezing (or nonclassicality). Similarly, using the integration formula
\ref{9}), we obtain
\begin{equation}
\left \langle a^{\dagger }\right \rangle =0, \label{24}
\end{equation
an
\begin{equation}
\left \langle a^{\dagger 2}\right \rangle =f_{3}\left( \alpha _{0}\right)
+f_{3}\left( -\alpha _{0}\right) -f_{4}\left( \alpha _{0}\right)
-f_{4}\left( -\alpha _{0}\right) , \label{25}
\end{equation
wher
\begin{eqnarray}
f_{3}\left( \alpha _{0}\right) &=&N_{m}^{-1}\chi ^{m}\frac{\partial ^{2m}}
\partial t^{m}\partial s^{m}}\left. \left( Rt+\alpha _{0}^{\ast }\right)
^{2}e^{-t^{2}-s^{2}-\left \vert R\right \vert ^{2}st+2Bt-2B^{\ast }s}\right
\vert _{s=t=0}, \notag \\
f_{4}\left( \alpha _{0}\right) &=&N_{m}^{-1}\chi ^{m}\frac{\partial ^{2m}}
\partial t^{m}\partial s^{m}}\left. \left( Rt+\alpha _{0}^{\ast }\right)
^{2}e^{-t^{2}-s^{2}-\left \vert R\right \vert ^{2}st-2C^{\ast }t-2Cs-2\left
\vert \alpha _{0}\right \vert ^{2}}\right \vert _{s=t=0}. \label{26}
\end{eqnarray}
Using Eqs.(\ref{16}), (\ref{24}) , (\ref{25}) and (\ref{23}), one can obtain
the expression of the quadrature squeezing $S$ of MCSO-OSCS.
\begin{figure}[tbp]
\label{Fig3} \centering \includegraphics[width=0.8\columnwidth]{F3.eps}
\caption{Quadrature squeezing of MCSO-OSCS as a function of $\protect \theta $
(a)\ $\protect \alpha _{0}=0.1$, $\protect \varphi =0$, with different $m$
values; (b) $\protect \alpha _{0}=0.1$, $\protect \varphi =0$, with different
m$ values; (c) $m=1$, $\protect \varphi =0,$with different $\protect \alpha
_{0}$ values; (d) $m=1$, $\protect \alpha _{0}=0.1$, with different $\protec
\varphi $ values. }
\end{figure}
We plot the graph of quadrature squeezing $S$ as a function of $\theta $ for
some different $m$ values and for given $\varphi $ and $\alpha _{0}$ values,
(say $\varphi =0$ and $\alpha _{0}=0.1$),\ see Fig. 3(a) and (b). It is
interesting to find that the MCSO-OSCS can exhibit squeezing when the
parameter $m$ is odd ($m=1,3,5,7$) and the angle $\theta $ is smaller than a
threshold, while can't exhibit squeezing when the parameter $m$ is even (
m=0,2,4,10$) for any angle $\theta $. Furthermore, we find that the original
state $\left( m=0\right) $ can't exhibit squeezing, which implies that the
odd times coherent superposition operation ($\Omega ^{m},m$ is odd.) can
achieve squeezing. Small angle $\theta $ corresponds to the case that the
subtracting photon operation is in the ascendant, which indicates that
subtracting photon operation is benefit to squeezing under the case of odd
m $.
In Fig. 3(c), we plot the graph of $S$ as a function of $\theta $ for some
different $\alpha _{0}$ values and for given $\varphi $ and $m$ values, (say
$\varphi =0$ and $m=1$). We find that small value of $\alpha _{0}$ is
helpful to squeezing on condition that the angle $\theta $ is smaller than a
threshold. From Fig. 3(d), We can see that different$\ \varphi $ values have
no effect on the squeezing of MCSO-OSCS.
\subsection{Photocount Distribution of MCSO-OSCS}
For the case of a single radiation mode of registering $n$ photoelectrons in
the time interval $T$, the photon counting distribution $P\left( n\right) $
is given by \cite{41},
\begin{equation}
P\left( n\right) =Tr\left[ \rho \colon \frac{\left( \xi a^{\dagger }a\right)
^{n}}{n!}e^{-\xi a^{\dagger }a}\colon \right] , \label{27}
\end{equation
where $\xi \propto T$ is called the quantum efficiency (a measure) of the
detector, $\rho $ is a single-mode density operator of the light field
concerned. When $\xi =1$, $P\left( n\right) $ becomes the photon number
distribution (PND)\ for a given state. By virtue of the technique of IWOP of
operators, Fan and Hu deduce a reformed formula as showed in reference \cit
{42},
\begin{equation}
P\left( n\right) =\frac{\xi ^{n}}{\left( \xi -1\right) ^{n}}\int \frac{d^{2}
}{\pi }e^{-\xi \left \vert z\right \vert ^{2}}L_{n}\left( \left \vert
z\right \vert ^{2}\right) Q\left( \sqrt{1-\xi }z\right) , \label{28}
\end{equation
where $Q\left( \beta \right) =\left \langle \beta \right \vert \rho
\left
\vert \beta \right \rangle $ is the Q-function, $\left \vert \beta
\right
\rangle $ is the coherent state, and $L_{n}\left( x\right) $ is the
Laguerre polynomials. Once the Q-function of $\rho $ is known, it is easy to
calculate the photocount distribution of MCSO-OSCS from Eq.(\ref{28}).
The Q-function of MCSO-OSCS is given by
\begin{eqnarray}
Q\left( \beta \right) &=&\left \langle \beta \right \vert \rho _{m}\left
\vert \beta \right \rangle \notag \\
&=&N_{m}^{-1}\chi m\left( \left \langle \beta \right \vert \colon
H_{m}(a,a^{\dag })\colon \left( \left \vert \alpha _{0}\right \rangle -\left
\vert -\alpha _{0}\right \rangle \right) \left( \left \langle \alpha
_{0}\right \vert -\left \langle -\alpha _{0}\right \vert \right) \colon
H_{m}^{\ast }(a,a^{\dag })\colon \left \vert \beta \right \rangle \right) ,
\label{29}
\end{eqnarray
where $H_{m}(a,a^{\dag })=H_{m}(i\Omega /\sqrt{2e^{i\varphi }\sin \theta
\cos \theta })$. Then substituting Eq.(\ref{29}) into Eq.(\ref{28}) and
using Eq.(\ref{9}) and the two-variable Hermite polynomials expression of
Laguerre polynomials \cite{26
\begin{equation}
L_{n}\left( zz^{\ast }\right) =\frac{\left( -1\right) ^{n}}{n!}H_{n,n}\left(
z,z^{\ast }\right) =\frac{\left( -1\right) ^{n}}{n!}\left. \frac{\partial
^{2n}}{\partial \mu ^{n}\partial \nu ^{n}}e^{-\mu \nu +\mu z+\nu z^{\ast
}}\right \vert _{\mu =\nu =0}, \label{30}
\end{equation
we obtain the final result of $P\left( n\right) $
\begin{equation}
P\left( n\right) =2T_{m,n}\sum \limits_{j=0}^{m}\sum
\limits_{l,k=0}^{n}A_{j,k} \left[
\begin{array}{c}
e^{-\xi \left \vert \alpha _{0}\right \vert ^{2}}H_{m-l-j}\left( \frac
K-J^{\ast }}{2}\right) H_{m-j-k}\left( \frac{J-K^{\ast }}{2}\right) \\
-\left( -1\right) ^{n-k}e^{\left( \xi -2\right) \left \vert \alpha
_{0}\right \vert ^{2}}H_{m-l-j}\left( \frac{K^{\ast }+J}{2}\right)
H_{m-j-k}\left( \frac{K+J^{\ast }}{2}\right
\end{array
\right] , \label{31}
\end{equation
where we have set
\begin{eqnarray}
A_{j,k} &=&\frac{\left( -1\right) ^{j+k}G^{j+l}G^{\ast j+k}F^{n-l}F^{\ast
n-k}}{l!j!k!\left( m-l-j\right) !\left( n-l\right) !\left( n-k\right)
!\left( m-j-k\right) !}, \notag \\
F &=&\sqrt{1-\xi }\alpha _{0},G=-\sqrt{1-\xi }R^{\ast }, \notag \\
J &=&\left( 1-\xi \right) R\alpha _{0},T_{m,n}=N_{m}^{-1}\chi ^{m}\frac
n!\left( m!\right) ^{2}\xi ^{n}}{\left( 1-\xi \right) ^{n}}. \label{32}
\end{eqnarray
\begin{figure}[tbp]
\label{Fig4} \centering \includegraphics[width=0.7\columnwidth]{F4.eps}
\caption{Photocount distribution of MCSO-OSCS as a function of $n$ (a
non-negative integer) for $\protect \alpha _{0}=0.5+i0.5$ (a) $m=4$, $\protec
\theta =\frac{\protect \pi }{4}$, $\protect \varphi =0$, $\protect \xi =0.2$;
(b) $m=4$, $\protect \theta =\frac{\protect \pi }{4}$, $\protect \varphi =0$,
\protect \xi =0.9$; (c) $m=1$, $\protect \theta =\frac{\protect \pi }{4}$,
\protect \varphi =0$, $\protect \xi =0.2$.}
\end{figure}
\begin{figure}[tbp]
\label{Fig5} \centering \includegraphics[width=0.7\columnwidth]{F5.eps}
\caption{Photocount distribution of MCSO-OSCS as a function of $n$ (a
non-negative integer) for $\protect \alpha _{0}=0.5+i0.5$ (a) $m=4$, $\protec
\theta =\frac{\protect \pi }{8}$, $\protect \varphi =0$, $\protect \xi =0.9$;
(b)$m=4$, $\protect \theta =\frac{\protect \pi }{8}$, $\protect \varphi =\frac
\protect \pi }{2}$, $\protect \xi =0.9$.}
\end{figure}
In order to discuss the photocount distribution of MCSO-OSCS, we plot the
graph of $P\left( n\right) $ for several given parameters $\varphi ,$
\theta $, $\alpha _{0}$, $m$, or $\xi $ in Figs. 4 and 5. Comparing with
Figs. 4(a) and (b), we find that for some given values of $m,$ $\alpha _{0},$
$\varphi ,$ and $\theta $, the corresponding probability-peak of photocount
distribution moves from $n=1$ to $n=4$ as $\xi =0.2$ increases to $0.9$,
which means that the probability of registering big photon-numbers is
increasing gradually while the probability of registering small
photon-numbers is decreasing when we increase the time interval $T$.
Meanwhile, the larger the $\xi $ is, the wider tail of photocount
distribution of MCSO-OSCS has. We can also see that the probability of
finding big photon-numbers increases with the increment of the parameter $m$
(see Figs. 4(a) and 4(c)) or $\theta $ (see Figs. 4(b) and 5(a)). Similarly,
the probability of finding big photon-numbers increases with the increment
of parameter $\varphi $ (see Figs. 5(a) and 5(b)).
\section{Wigner Function of the MCSO-OSCS}
The WF is a quasi-probability distribution, which fully describes the state
of a quantum system in phase space. The partial negativity of the WF is
indeed a good indication of the highly nonclassical character of the state
\cite{43}. Therefore it is worth obtaining the WF for any states and using
the negative region to check whether a state has nonclassicality. For a
single-mode system, the WF $W\left( \alpha ,\alpha ^{\ast }\right) $
associated with a quantum state density matrix $\rho $ can be expressed as
\cite{44}:
\begin{equation}
W\left( \alpha \right) =\frac{1}{\pi }e^{2\left \vert \alpha \right \vert
^{2}}\int \frac{d^{2}z}{\pi }\left \langle -z\right \vert \rho \left \vert
z\right \rangle e^{-2\left( \alpha ^{\ast }z-\alpha z^{\ast }\right) },
\label{33}
\end{equation
where $\left \vert z\right \rangle $ is the coherent state. Substituting
\rho _{m}=N_{m}^{-1}\left \vert \psi _{m}\right \rangle \left \langle \psi
_{m}\right \vert $ into Eq.(\ref{33}), we can finally obtain the WF of
MCSO-OSCS:
\begin{equation}
W\left( \alpha \right) =W_{\alpha _{0}}\left( \alpha \right) +W_{-\alpha
_{0}}\left( \alpha \right) -W_{\alpha _{0}}^{\prime }\left( \alpha \right)
-W_{-\alpha _{0}}^{\prime }\left( \alpha \right) , \label{34}
\end{equation
where we have set
\begin{eqnarray}
W_{\alpha _{0}}\left( \alpha \right) &=&\sum \limits_{k=0}^{m}D_{m}\left(
-1\right) ^{k}e^{-2\left \vert \alpha -\alpha _{0}\right \vert ^{2}}\left
\vert H_{m-k}\left( -C^{\ast }+R\alpha \right) \right \vert ^{2}, \notag \\
W_{\alpha _{0}}^{\prime }\left( \alpha \right) &=&\sum
\limits_{k=0}^{m}D_{m}\left( -1\right) ^{m}e^{2\alpha _{0}^{\ast }\alpha
-2\alpha ^{\ast }\alpha _{0}-2\left \vert \alpha \right \vert
^{2}}H_{m-k}\left( B^{\ast }-R^{\ast }\alpha ^{\ast }\right) H_{m-k}\left(
B+R\alpha \right) , \label{35}
\end{eqnarray
and
\begin{equation}
D_{m}=\frac{1}{\pi }N_{m}^{-1}\left( \frac{1}{2}\sin \theta \cos \theta
\right) ^{m}\left( 2\frac{\sin \theta }{\cos \theta }\right) ^{k}\frac{1}{k!
\left( \frac{m!}{\left( m-k\right) !}\right) ^{2}. \label{36}
\end{equation}
It is found that the sum of $W_{\alpha _{0}}^{\prime }\left( \alpha \right) $
and $W_{-\alpha _{0}}^{\prime }\left( \alpha \right) $ is a real function
due to $W_{-\alpha _{0}}^{\prime }\left( \alpha \right) =\left[ W_{\alpha
_{0}}^{\prime }\left( \alpha \right) \right] ^{\ast }$. By using Eq. (\re
{34}), the WFs as a function of real and imaginary parts of $\alpha $ for
several different values of $m$, $\alpha _{0}$ and $\theta $ are depicted in
Figs. 6 and 7.
\begin{figure}[tbp]
\label{Fig6} \centering \includegraphics[width=0.8\columnwidth]{F6.eps}
\caption{Wigner function distributions of MCSO-OSCS with $\protect \varphi =0
\protect \theta =\frac{\protect \pi }{3},\protect \alpha _{0}=1+i$ (a) $m=0$;
(b) $m=1$; (c) $m=2$; (d) $m=3$.}
\end{figure}
\begin{figure}[tbp]
\label{Fig7} \centering \includegraphics[width=0.8\columnwidth]{F7.eps}
\caption{Wigner function distributions of MCSO-OSCS with $m=2,\protec
\varphi =0$ (a) $\protect \alpha _{0}=1+i,\protect \theta =\frac{\protect \pi }
8}$; (b) $\protect \alpha _{0}=2(1+i),\protect \theta =\frac{\protect \pi }{3}
. }
\end{figure}
We can see clearly that the figures of WF distribution are non-Gaussian. In
addition, as evidence of the nonclassicality of the state, it is easy to see
that there is a negative region of the WF in each plot. From Fig. 6, We can
see\ that the figures of WF exist odd (even) negative peaks when the values
of $m$ are even (odd) for given $\alpha _{0}$, $\varphi $ and $\theta $, and
exhibit more vibration character as the value of $m$ increasing. Meanwhile,
we can find that the minimum value of the WF occurs at the center of the
figure when $m$ is an even number (see Fig. 6(a) and Fig. 6(c)). But the
case is not true when $m$ is an odd number (see Fig. 6(b) and Fig. 6(d)).
Comparing Fig. 6(c) ($\theta =\frac{\pi }{3}$, $m=2$, $\alpha _{0}=1+i$)
with Fig. 7(a) ($\theta =\frac{\pi }{8}$, $m=2$, $\alpha _{0}=1+i$), we can
see that the width of\ the figure of WF in one direction increases as
increasing the value of $\theta $. Comparing Fig. 6(c) ($\alpha _{0}=1+i$,
m=2$, $\theta =\frac{\pi }{3}$) with Fig. 7(b) ($\alpha _{0}=2+2i,m=2,\theta
=\frac{\pi }{3}$), we can also see that the figure of WF also\ exhibits more
vibration character as increasing the value of amplitude $\left \vert \alpha
_{0}\right \vert $.
The volume of the negative part of the WF were used in \cite{45,46} to
describe the interference effects which determine the departure from
classical behavior. In order to further evaluate how these parameters $m$,
\alpha _{0}$, and $\theta $ affect the negative part of WF distribution for
MCSO-OSCS, we shall consider the negative part volume of WF which may be
written a
\begin{equation}
\delta =\frac{1}{2}\left[ \int d^{2}\alpha \left \vert W\left( \alpha
\right) \right \vert -1\right] . \label{37}
\end{equation}
By definition, the quantity $\delta $ is equal to zero for coherent and
squeezed vacuum states, as their WFs are non-negative. Once knowing the
Wigner function of a quantum state, we can obtain the negative part volume
of WF through numerical integration.
\begin{figure}[tbp]
\label{Fig8} \centering \includegraphics[width=0.8\columnwidth]{F8.eps}
\caption{The volume of the negative part of the WF for MCSO-OSCS as the
function of $\protect \theta $ (a) $\protect \alpha _{0}=0.1,\protect \varphi
=0 $; (b) $m=1,\protect \varphi =0$.}
\end{figure}
In Fig. 8, we plot the negative part volume $\delta $ of WF for MCSO-OSCS as
the function of $\theta $. It is shown that the negative part volume $\delta
$ generally increases as $\theta $ increases when $m\neq 0$. In addition, it
is interesting to note that $\delta $ is sensitive to parameter $m,$ and
\delta $ increases as $m$ increases when parameter $\theta $ is bigger than
a threshold (see Fig. 8(a)). In other words, the MCSO-OSCS may exhibit more
nonclassicality by increasing the value of$\ m$. Meanwhile, $\delta $
increases as the value of $\alpha _{0}$ increases when parameter $\theta $
is smaller than a threshold (see Fig. 8(b)).
\section{The decoherence of the MCSO-OSCS in a thermal environment}
When the MCSO-OSCS evolves in the thermal channel, the evolution of the
density matrix in the Born-Markov approximation and the interaction picture
can be described by the master equation \cite{47}
\begin{equation}
\frac{d\rho }{dt}=\kappa \left( \bar{n}+1\right) \left( 2a\rho a^{\dagger
}-a^{\dagger }a\rho -\rho a^{\dagger }a\right) +\kappa \bar{n}\left(
2a^{\dagger }\rho a-aa^{\dagger }\rho -\rho aa^{\dagger }\right) ,
\label{38}
\end{equation
where $\kappa $ represents the dissipative coefficient and $\bar{n}$ ($\bar{
}=\frac{1}{e^{\hbar \omega /(k_{B}T)}-1}$, $T$ is temperature.) denotes the
average thermal photon number of the environment \cite{48}. Using the
thermal entangled state representation \cite{49}, the time evolution of
distribution functions in the dissipative channels are derived \cite{50,51}.
The evolutions of the WF is governed by the following integration equatio
\begin{equation}
W\left( \gamma ,\gamma ^{\ast },t\right) =\frac{2}{\left( 2\bar{n}+1\right)
\Gamma }\int \frac{d^{2}\alpha }{\pi }W\left( \alpha ,\alpha ^{\ast
},0\right) \exp \left[ -2\frac{\left \vert \gamma -\alpha e^{-\kappa
t}\right \vert ^{2}}{\left( 2\bar{n}+1\right) \Gamma }\right] , \label{39}
\end{equation
where $\Gamma =1-e^{-2\kappa t}$ and $W\left( \alpha ,\alpha ^{\ast
},0\right) $ is the WF of the initial state. Thus the WF at any time can be
obtained by performing the integration when the initial WF is known.
Substituting Eq. (\ref{34}) into Eq. (\ref{39}), we have
\begin{equation}
W\left( \gamma ,\gamma ^{\ast },t\right) =W_{\alpha _{0}}\left( \gamma
,\gamma ^{\ast },t\right) +W_{-\alpha _{0}}\left( \gamma ,\gamma ^{\ast
},t\right) -\left( W_{\alpha _{0}}^{\prime }\left( \gamma ,\gamma ^{\ast
},t\right) +c.c.\right) , \label{40}
\end{equation
where
\begin{equation}
W_{\alpha _{0}}\left( \gamma ,\gamma ^{\ast },t\right) =\sum
\limits_{k=0}^{m}\sum_{l=0}^{m-k}MVU^{l}e^{-2V\left \vert \gamma -\alpha
_{0}e^{-\kappa t}\right \vert ^{2}}\left \vert H_{m-k-l}\left( -C^{\ast
}+R\alpha _{0}U+R\gamma e^{-\kappa t}V\right) \right \vert ^{2}, \label{41}
\end{equation
\begin{eqnarray}
W_{\alpha _{0}}^{\prime }\left( \gamma ,\gamma ^{\ast },t\right) &=&\sum
\limits_{k=0}^{m}\sum_{l=0}^{m-k}MVU^{l}e^{\left( -2\left \vert \gamma
\right \vert ^{2}V-2\left \vert \alpha _{0}\right \vert ^{2}U+2\gamma
e^{-\kappa t}\alpha _{0}^{\ast }V-2\gamma ^{\ast }e^{-\kappa t}\alpha
_{0}V\right) } \notag \\
&&\times H_{m-k-l}\left( -B^{\ast }+R^{\ast }\alpha _{0}^{\ast }U+R^{\ast
}\gamma ^{\ast }e^{-\kappa t}V\right) H_{m-k-l}\left( B-R\alpha
_{0}U+R\gamma e^{-\kappa t}V\right) , \label{42}
\end{eqnarray
and
\begin{eqnarray}
V &=&\frac{1}{2\bar{n}\Gamma +1},U=1-e^{-2\kappa t}V, \notag \\
M &=&\frac{N_{m}^{-1}\left( -1\right) ^{k}2^{2l+k-m}\left( m!\right) ^{2}}
\pi k!l!\left( \left( m-k-l\right) !\right) ^{2}}\frac{\sin ^{k+l+m}\theta }
\cos ^{k+l-m}\theta }. \label{43}
\end{eqnarray
Further, when $t=0,\Gamma =0$, Eq.(\ref{40}) just reduces to (\ref{34}), as
expected.
\begin{figure}[tbp]
\label{Fig9} \centering \includegraphics[width=0.7\columnwidth]{F9.eps}
\caption{The time evolution of Wigner function for MCSO-OSCS in the thermal
environment with $\protect \varphi =0,\protect \theta =\frac{\protect \pi }{3}
\protect \alpha _{0}=1+i$, $m=1,\bar{n}=0.2$ (a) $\protect \kappa t=0.001$;
(b) $\protect \kappa t=0.05$; (c) $\protect \kappa t=0.1$ ; (d) $\protec
\kappa t=3$. }
\end{figure}
\begin{figure}[tbp]
\label{Fig10} \centering \includegraphics[width=0.7\columnwidth]{F10.eps}
\caption{Wigner function distributions of MCSO-OSCS in the thermal
environment for $\protect \varphi =0,\protect \theta =\frac{\protect \pi }{3}
\protect \alpha _{0}=1+i,m=1,\protect \kappa t=0.05$ with different parameter
\bar{n}$ (a) $\bar{n}=0$; (b) $\bar{n}=0.5$; (c) $\bar{n}=2;$ (d) $\bar{n}=8
.}
\end{figure}
In order to see the decoherence of the MCSO-OSCS in the thermal
environment,\ we plot the time\ evolution of WF $W\left( \gamma ,\gamma
^{\ast },t\right) $ as a function of real and imaginary parts of $\gamma $
for some different $t$ values and for a given $m$ value (say, $m=1$) in
Figs. 9. It is shown that as time proceeds the negative part of WF and
muti-peaks vibration structure of the plot disappear gradually, and finally
the plot evolves to a wave packet structure, the figure of a Gaussian
distribution (see Fig. 9(d)), which means that\ the MCSO-OSCS has reduced to
the thermal state. In Fig. 10, we plot the picture of $W\left( \gamma
,\gamma ^{\ast },t\right) $ for some different $\bar{n}$ values and for a
given $m$ value (say, $m=1$) at the given time (say, $\kappa t=0.05$). It is
interesting to note that the negative part of WF decreases as the average
photon number $\bar{n}$ increases, i.e., the lager $\bar{n}$ the more
rapidly the nonclassicality is lost, which means that the higher the
temperature of thermal field, the more rapidly the nonclassicality of the
MCSO-OSCS is lost. This result is same as Ref. \cite{52}. \
\section{Conclusions}
In summary, we investigate the nonclassicality of MCSO-OSCS which is
obtained through $m$ times coherent superposition operator $a\cos \theta
+a^{\dagger }e^{i\varphi }\sin \theta $ operating on an odd-Schr\H{o
dinger-cat state. For arbitrary $m$ value, through IWOP technique we have
obtained an analytical expression of the normalization constant, which turns
out to be ralated with the Hermite polynomial. Then the fidelity between
MCSO-OSCS and its original OSCS is discussed. By numerical plot, it is
obvious to note that the fidelity is equal to $0$ when $m$ is odd and not
equal to $0$ when $m$ is even. The nonclassical properties of the state,
such as sub-Poissonian statistics, quadrature squeezing properties, and
photocount distribution are also discussed in details. We find that
MCSO-OSCS has more chance to exhibit sub-Poissonian statistics with bigger
value of $\theta $ in the area of $\theta \in (0,\frac{\pi }{2})$. We also
find that MCSO-OSCS can exhibit squeezing when the parameter $m$ is odd and
the angle $\theta $ is smaller, which indicates that the subtracting photon
operation is benifit to squeezing for odd $m$. Furthermore, the
nonclassicality of MCSO-OSCS is investigated in terms of WF and the negative
part volume of WF after deriving the analytical expression of WF. It is
shown that the WF of the MCSO-OSCS always has negative values which implies
the highly nonclassical properties of quantum states. The negative part
volume of WF increases as $m$ increases when $m\neq 0$ and parameter $\theta
$ is bigger than a threshold, and increases as the value of $\alpha _{0}$
increases when parameter $\theta $ is smaller than a threshold. Especially,
the negative part volume of WF increases with the increment of parameter
\theta $ except the case of $m=0$.
We also investigate the decoherence of the MCSO-OSCS in terms of the
fadeaway of the negativity of WF in a thermal environment. It is shown that
nonclassicality of the MCSO-OSCS decreases as time proceeds and the
MCSO-OSCS reduces to the thermal state finally. It is also shown that
nonclassicality is influenced by the temperature of environment, the higher
the temperature is, the more rapidly the nonclassicality of the MCSO-OSCS is
lost. We wish that our results will benefit for instructing experiments, for
example, the new state would become a sufficient resource to generate a
two-mode entanglement via a beam-splitter.
\begin{acknowledgments}
This project was supported by the National Natural Science Foundation of
China (Nos.11264016, 11364022) and the Natural Science Foundation of Jiangxi
Province of China (No.20142BAB202004) as well as the Research Foundation of
the Education Department of Jiangxi Province of China (Nos.GJJ12171,
GJJ12172).
\end{acknowledgments}
|
\section{Introduction}
Experimental evidence of critical or quasi-critical behavior in brain
networks was gathered over the past decade
\cite{BP03,BP04,Plenz07,Beggs08,Peterman09,Haimovici,Chialvo10}. The
discovery of scale-invariant avalanches of neural activity led to the
conjecture that the brain might operate close to a critical point
\cite{BP03,BP04}. It was argued that critical behavior might bear
functional advantages; for instance, the divergence at criticality of
quantities such as susceptibilities and correlation lengths could
entail the ability of brain networks to coordinate system-wide
activities and efficiently respond to a broad range of stimuli. A
vast number of studies have since flowered, focusing on the numerical
simulation of simple dynamical models that could recover
phenomenologically the hallmarks of criticality observed in
experiments \cite{Kaiser07,Kaiser10,Rubinov,Zhou11,Zhou12}. In
particular, it was noted that effective highly-simplified models of
activity propagation --such as the contact process and the
quiescent-excited-refractory-quiescent model-- could provide valuable
information on large-scale brain properties
\cite{Grinstein-Linsker}. In these ideal models, an active ``unit'' or
node --be it a neuron at a microscopic scale or a coarse-grained
active region at a larger mesoscopic scale-- can propagate its
activity to neighboring units and/or become deactivated. Such simple
dynamics --where activity propagation involves a single active node--
lead generically to continuous phase transitions, with a critical
point separating an active from a quiescent phase
\cite{Marro-Dickman,Liggett,Odor-book,Hinrichsen-book}. Moreover,
relatively simple modifications of these models implementing standard
mechanisms of self-organized criticality lead to robust critical or
quasi-critical behavior without the need of parameter fine tuning
\cite{Levina07,Mill10,JaboJSTAT}.
However, a closer look at real neural dynamics suggests that neural
activity propagation may follow more complicated rules. In particular,
individual neurons usually require to integrate up to hundreds of
post-synaptic potentials before spiking themselves, as typically
captured by integrate-and-fire models
\cite{Abbott,Burkitt-I,Burkitt-II}. At mesoscopic scales, such
requirement may be less stringent; however it is reasonable to
consider that a few neighboring active units might be required to
generate further activity: i.e. the dynamics follow a schematic rule of
the type: $n$A$\to (n+m)$A, with $n>1$ and $m$ of the order of a few
units, where each A stands for an active location or site
\cite{Marro-Dickman,Liggett,Odor-book,Hinrichsen-book}.
Such types of $(n,m)$-processes are well known in reaction-diffusion
systems \cite{Marro-Dickman,Liggett,Odor-book,Hinrichsen-book}, and
they are often used in the modeling of neural dynamics. In particular,
it is known that they lead to broad phases of sustained activity
\cite{Kaiser07} and enhanced dynamic ranges
\cite{Kinouchi-Copelli,Gollo12}. However, these processes are also
well-known to lead to pattern formation (Turing patterns) and to
discontinuous phase transitions between active and quiescent
phases, with associated phase coexistence \cite{Windus-Jensen07} and
lacking critical points, in seeming contradiction with the observation
of scale-invariant behavior in brain dynamics. In other words, the
requirement of more than one source of activity to generate further
activity leads to discontinuous phase transitions, separating two
highly different active and quiescent phases, with no sign of
criticality nor scale-invariance in between
\cite{Marro-Dickman,Liggett,Odor-book,Hinrichsen-book}.
Our goal here is to reconcile the need for signal integration at the
neuron scale, supposedly leading to discontinuous transitions, with
the empirical observation of critical-like features, characteristic of
continuous phase transitions. As we hope to convincingly argue, the
key to this puzzling ambiguity lies in the topology of the underlying
network of neural connections, which we will prove responsible for the
generic rounding of discontinuous transitions. In a nutshell, and in
analogy with what happens in problems of thermodynamic equilibrium at
low dimensions, the existence of some form of structural disorder
implies that potentially discontinuous transitions are rounded-off, thus making
the system critical.
\section{The underlying network of neural connections}
In spite of the huge complexity that would be required to represent
the detailed structure of the brain down to the single-neuron level,
an effective coarse-grained description of neural contact patterns can
be provided by a network --the connectome-- whose nodes represent
groups of neurons, such as cortical columns, and whose links represent
the groups of fibers connecting them \cite{Kaiser10}.
Studies employing different neuroimaging techniques have revealed that
the Human Connectome, the current mapping of human brain connections,
is organized in a hierarchical and modular fashion, in which local
regions are clustered into large-scale moduli, which in turn form
higher level structures and so on
\cite{Sporns05,Meunier,Kaiser07b,Sporns07,Hagmann,Zamora-Lopez}. The
resulting hierarchical modular network (HMN) can be visualized as
built-up from moduli of large internal neural connectivity, enclosed
into higher-level sparser moduli, in a nested hierarchical fashion.
HMNs have been recently found to play a crucial role in neural
dynamics. In particular, simple models of activity propagation were
recently found to display Griffiths phases when running on top of HMNs
\cite{NatCommun}, corroborating the experimental observation of
extended critical regions in the human at its resting state
\cite{Taglia}. Similarly, they were argued to extend the region of
apparent criticality in self-organized models of neural activity and
they were used to explain the ability of the brain to sustain activity
over extended time windows
\cite{Kaiser07,Kaiser10,Rubinov,Zhou11,Zhou12,NatCommun}.
Here we shall use a simple structural model to build-up synthetic HMNs
as follows: local densely connected moduli are used as building blocks;
they are recursively grouped by establishing additional inter-moduli links in a
level-dependent way, as exemplified in Figure \ref{fig:network}.
\begin{figure}
\begin{center}
\vspace{-0cm}
\includegraphics[width=12cm]{network.eps}
\caption{Sketch of the HMN construction method. Given a positive
integer $s$, consider $2^s$ basal fully connected moduli of size
$M$. At the lowest hierarchical level, moduli are linked pairwise
into super-moduli by establishing a fixed number $\alpha$ of random
unweighted and undirected links between the elements of each modulus
($\alpha = 2$ in figure). Newly formed blocks are then iteratively
linked pairwise with the same $\alpha$ for a total of $s$
iterations, until the network becomes connected. The resulting
network has size $N=2^s M$.}
\label{fig:network}
\end{center}
\end{figure}
Further details of the construction methods can be found in Reference
\cite{NatCommun}. A crucial feature of HMNs is represented by their
finite topological dimension $D$. The topological dimension of a
network can be defined as follows: starting from a single node, the
number of neighbors $N_z$ reachable after $z$ steps is computed for
increasing $z$ until the entire network is covered \cite{MAM-GP}. The
network is finite dimensional with dimension $D$ if $\langle
N_z\rangle\sim z^D$, generalizing the familiar behavior of regular
lattices. The topological dimension of a HMN can be tuned easily, by
changing the average number $\alpha$ of links between pairs of modules
at each hierarchical level (see Fig. \ref{fig:network} and
\cite{NatCommun}). Although brain moduli and columns may be densely
connected, at larger mesoscopic and macroscopic scales the
hierarchical contact patterns become very sparse. At such scales, the
effective network becomes finite dimensional \cite{Gallos,NatCommun}.
\section{Continuous versus discontinuous transitions in the presence of
disorder}
It was recently conjectured that dynamical models of activity
propagation characterized by discontinuous phase transitions at the
mean-field level exhibit a rounding phenomenon in finite dimensional
disordered systems, eventually leading to continuous phase transitions
at dimension $D \le 2$ \cite{Villa14}. Such behavior has been
envisaged as the non-equilibrium analogue of the well-know Imry-Ma
criterion, which states that --in the presence of quenched disorder--
spontaneous symmetry breaking as well as first-order phase transitions
are prohibited in equilibrium systems at $D \le 2$
\cite{Imry-Ma,Aizenman,Berker}. Analogously to quenched disorder in
lattices, structural disorder is integral to HMNs as defined above and
may thus be responsible for the rounding of discontinuous phase
transitions in such systems \cite{MAM-GP}. In this light, we
investigate how this form of topological disorder can potentially
alter the order of phase transitions exhibited by simple
$(n,m)$-models of neural activity, which would normally exhibit first-order
phase transitions in networks well described by the mean-field
approximation. In analogy with the Imry-Ma criterion, {\it a priori},
this effect should be expected to occur in networks with topological
dimension $D$ less than $2$ \cite{Villa14}.
\section{Results}
In what follows, we provide extensive numerical tests of the above
conjecture, showing how the topological dimension of a disordered
network can tune the nature of the dynamical phase transition,
ultimately forcing $n$A$\to (n+m)$A dynamics to exhibit continuous
transitions for $D \le 2$. To this end, we consider a prototypical
model, in which we choose $n=2$ and $m=1$ (a $(2,1)$-process in our
notation), whose Monte Carlo implementation is as follows: each of the
$N$ nodes of the network is endowed with a binary state variable
$\sigma=0,1$, inactive or active, and $\rho(t)$ is the density of
active nodes at time $t$; i) at each time step an active node
$\mathrm{R}_1$ is selected and time is increased by $[N\rho(t)]^{-1}$;
ii) with {\it death} probability $p_{\mathrm{d}}$, $\mathrm{R}_1$ is
deactivated, while with complementary probability $1-p_{\mathrm{d}}$,
a neighboring node $\mathrm{R}_2$ is considered and one of the
following actions is taken; iiia) if $\mathrm{R}_2$ is inactive,
activity diffuses to $\mathrm{R}_2$, leaving $\mathrm{R}_1$; iiib) if
$\mathrm{R}_2$ is active, a new neighbor $\mathrm{R}_3$ of
$\mathrm{R}_1$ is considered and, if inactive, it is activated with
{\it birth} probability $p_{\mathrm{b}}$. From a neurophysiological
perspective, $p_{\mathrm{d}}$ encodes the exhaustion mechanism that
accounts for spontaneous deactivation of neurons and neural regions,
which proves essential in maintaining sustained activity bounded
\cite{Kaiser07}. The integrated activation is tuned by
$p_{\mathrm{b}}$. A simple mean-field equation for this type of dynamics is
\begin{equation}
\dot{\rho}(t) =
-p_{\mathrm{d}}\rho(t)
+(1-p_{\mathrm{d}})p_{\mathrm{b}} \, \rho^2(t) \, [1-\rho(t)]
\label{mf}
\end{equation}
which exhibits a discontinuous phase transition (fold bifurcation) at
$p_{\mathrm{d}} = p_{\mathrm{b}}/(4+p_{\mathrm{b}})$
\cite{Windus-Jensen07}. Notice that different choices of $n,m>1$,
which might account for enhanced realism in the physiological
description of brain networks at the mesoscale, would not affect such
behavior. On the other hand, a more detailed theoretical description
of this system --taking explicitly into account the underlying network
topology-- would be provided by a quenched-mean-field approach
\cite{Castellano10,Mata13}.
\begin{figure}
\begin{center}
\vspace{-0cm}
\includegraphics[width=12cm]{phasediagrams.eps}
\caption{ (Color online) Phase diagrams for topological dimension
$D$ above and below $D=2$ respectively. In high dimension, the
phenomenology of a discontinuous phase transition is recovered,
in agreement with the mean-field prediction for the
$(2,1)$-process. Below $D=2$, the transition becomes continuous.
A feeble appearance of hysteretic behavior is recorded in the
$D>2$ case, where different colors correspond to different
initial conditions and spinodal points marking the transition are
located at $p_{\mathrm{d}_{\mathrm{thr}}}\approx 0.0732(1)$ and
$p_{\mathrm{d}_{\mathrm{thr}}}\approx 0.0734(1)$ for $\rho_0=0$
and $\rho_0=1$ respectively (not distinguishable in figure).
Such dependence disappears for $D<2$, in accordance with the the
hypothesis of a continuous phase transition, the critical point
being located at $p_{\mathrm{d}_{\mathrm{thr}}}\approx 0.0402(1)$
Simulations are run on HMNs of size $N=2^{17}=131072$,
partitioned into $s=13$ hierarchical levels.}
\label{fig:phasediagrams}
\end{center}
\end{figure}
As a substrate on which the above dynamics run, we considered
different HMNs, characterized by different topological dimensions $D$.
In the rest of the paper we will show results for HMN extracted from
two ensembles $\mathcal{N}_-$ and $\mathcal{N}_+$, each with a fixed
average dimension $D_{-,+}$ below and above the threshold value
$D=2$. In particular, we show results for $D_{-}\approx 1.6$ and
$D_{+}\approx 2.8$ (networks with such properties are obtained by
choosing $\alpha=1$ and $\alpha=4$ respectively, in the HMN
building-up process; see Figure \ref{fig:network}).
We ran Monte Carlo simulations of the above $(2,1)$-process on such
networks. Figure \ref{fig:phasediagrams} shows the steady state value
of the average activity density, as a function of the control
parameter $p_{\mathrm{d}}$, respectively below and above dimension
$D=2$, in spreading simulations starting both from localized active
seeds ($\rho_0\approx 0$) and from the homogeneously active state
($\rho_0=1$). While the discontinuity encountered above $D=2$ is in
agreement with the mean-field behavior for this type of dynamics,
below $D=2$ the dynamical phase transition is evidently continuous,
confirming the conjecture of a low-dimensional rounding.
\begin{figure}
\begin{center}
\vspace{-0cm}
\includegraphics[width=12cm]{bistability.eps}
\caption{(Color online) Time evolution of the activity density
$\rho$ for different initial values $\rho_0$ (different colors),
for dimension above and below $D=2$ respectively. In both cases, the
control parameter $p_{\mathrm{d}}$ is chosen at the threshold value. For
high dimension, bistable behavior is recovered, as in standard
first-order transitions, whereas no sign of bistability is
encountered below $D=2$. Notice however that both configuration
converge very slowly to their expected behavior. In particular,
in the $D>2$ (discontinuous) case, large enough initial
conditions lead to very long transients, which could be
misinterpreted as continuous behavior for short simulation
times. Such traits of quasi-critical states become stronger as
$D=2$ is approached from above, and corroborate the picture of a
rounding phenomenon.}
\label{fig:bistability}
\end{center}
\end{figure}
To provide further evidence of the radical difference in the
transition nature, Figure \ref{fig:bistability} shows the time
evolution of the average activity density $\rho$ upon changing the
initial activity $\rho_0$, for both cases in Figure
\ref{fig:phasediagrams}, each at the estimated threshold
$p_{\mathrm{d}}$. Above $D=2$ clear signs of bistability emerge,
signaling coexistence phenomena, which typically characterize
discontinuous phase transitions. Below $D=2$, however, the steady
state does not depend on the initial condition anymore, as expected
for a continuous transition, in which correlations become system-wide
and coexistence is prohibited. In order to gain a deeper understanding
of the rounding phenomenon, we can analyze the nature of the inactive
(absorbing) phase in both cases.
\begin{figure}
\begin{center}
\vspace{-0cm}
\includegraphics[width=12cm]{decay.eps}
\caption{(Color online) Time evolution of the activity density
$\rho$ for different values of the control parameter $p_{\mathrm{d}}$
(different colors) below the dynamic threshold, for dimension above and
below $D=2$ respectively. For high dimension, the transition
between the inactive and the active state is abrupt, with no
signs of criticality. Below $D=2$, a Griffiths phase emerges
for $p_{\mathrm{d}}>p_{\mathrm{d}_{\mathrm{thr}}}$, characterized by
generic
power-law relaxation and critical-like behavior.}
\label{fig:decay}
\end{center}
\end{figure}
To this end, let us consider simulations starting from a homogeneous
$\rho_0=1$ state.
Time evolution of $\rho$ is shown in Figure \ref{fig:decay}, for
different values of $p_{\mathrm{d}}$ in the inactive phase.
As usual, above $D=2$ results
for $p_{\mathrm{d}}$ point to an abrupt change in behavior at the dynamic
threshold $p_{\mathrm{d}_{\mathrm{thr}}}$, below which activity dies off
exponentially fast as soon as a large enough fluctuation breaks the
coexistence of active and inactive islands. At
$p_{\mathrm{d}_{\mathrm{thr}}}$, such coexistence becomes stable in the
large-$N$ limit and the phenomenology of a discontinuous phase
transition is recovered. For dimensions below $D=2$, instead, the
system displays a Griffiths phase\cite{Vojta-review,Vojta-Lee-prl}: the average
activity density decays as power laws with continuously varying exponents as a
function of the control parameter $p_\mathrm{d}$. A critical point,
characterized by activated scaling logarithmic time decay
($p_{\mathrm{d}_\mathrm{thr}}\approx 0.0402(1)$ in Fig.\ref{fig:decay})
separates the Griffiths phase from the active phase, marking the recovery of
critical behavior at low dimension.
Griffiths phases are a manifestation of rare-region effects:
islands of localized activity are able to remain active for long times. Their
relevance for both complex networks \cite{MAM-GP} and brain
networks \cite{NatCommun} has been recently discussed in the literature.
Activity propagation models without signal integration
yield such behavior in HMNs as they naturally lead to
continuous phase transitions regardless of dimensionality
constraints \cite{NatCommun}.
Remarkably, the $(2,1)$-model dynamics at low dimensions recovers here those
fingerprints of criticality, in spite of being typically associated with
discontinuous transitions at mean field.
Interestingly, upon approaching the threshold dimension from above, $D\to 2^+$,
the discontinuous nature of the transition is rounded:
although coexistence is genuinely recovered at very large times, activity is
able to self-sustain even in the absorbing state for times potentially longer
than any observation window.
\section{Discussion and Conclusions}
A recent study showed that the $(2,1)$-process adopted here as a
paradigm for first-order phase transitions may show tricritical
behavior in certain families of ordered fractal lattices of dimension
$1<D<2$ \cite{Windus-Jensen09}. Such a finding implies that for each
family of ordered fractals, there exist a ``critical'' dimension
$1<D_\mathrm{c}<2$, below which the transition is continuous,
recovering the known behavior of one dimensional chains, and above
which the transition becomes discontinuous, anticipating the behavior
of pure two-dimensional lattices. In our study we have introduced
disorder in the topology and shown that finite-dimensional disordered
hierarchical modular networks of relevance in neuroscience {\it
always} display continuous phase transitions for $D<2$. In fact even
square lattices ($D=2$) exhibit this behavior provided that disorder
is introduced, specifically in the form of quenched impurities
\cite{Villa14}. Such results corroborate the conjecture that, due to
disorder, non-equilibrium systems with absorbing states do not sustain
first-order dynamical phase transitions for {\it any} $D\le 2$. We
provided further evidence for this claim, proving its validity for
HMNs, and we focused on the relevance that such result may have for
neuroscience. Brain activity is known to exhibit critical-like
behavior, which would suggest its ability to sit constantly in the
vicinity of a continuous transition. We have shown that even if
realistic dynamic models lead to first-order phase transitions in the
mean-field approximation, in low-dimensional disordered systems such
transitions are rounded. A natural question arises whether the brain
actually is a low-dimensional network, provided that each single
neuron may have up to thousands of neighbors. The solution to this
apparent contradiction comes from the hierarchical organization of
brain connections. At the lowest scales, neurons are grouped in well
connected moduli which act as small worlds of diverging topological
dimension. At such scales, integrate-and-fire dynamics naturally
trigger coexistence and local discontinuous activations of moduli. At
the largest scales, however, inter-moduli connectivity is very sparse
in order to maintain the volume of white-fiber matter bounded
\cite{Kaiser10}, allowing only for weak small-world effects \cite{Gallos}.
Such connectivity patterns become finite dimensional, and discontinuous phase
transitions are prohibited.
Notice that the $D=2$ bound should not be read strictly
in real systems. We found that significant traits of quasi-critical
behavior appear even above $D=2$, suggesting a gradual rounding
phenomenon. Such systems will theoretically show discontinuous
transitions for large enough times, yet they are able to sustain
anomalous activity for typical time windows of experimental
observations. More detailed and realistic models of
neural dynamics could be provided. While the behavior presented here
is conserved for different choices of the parameters $(n,m)$,
realistic models supposedly include ingredients such as refractory
times, explicit time integration, inhibition and dependence on synapse
directedness. While such details are of primary importance in
correctly describing physiological aspects of brain activity, we
believe that our simple approach has the advantage of focusing on the
large-scale topology of the Human Connectome
\cite{Sporns05,Hagmann}, in order to provide
insight about its large-scale behavior. An understanding of low-level
synaptic activity requires realistic neuron models and remains a
formidable task.
In conclusion, we have studied the properties of a family of dynamic
models of relevance in the description of neural activity in the
presence of signal integration. Although signal integration may be
responsible for the emergence of first-order phase transitions in
generic networks, we have shown that phase transitions are rounded in
finite dimensional hierarchical networks, eventually turning
continuous for $D\le 2$. Such finding is relevant in explaining the
observation of critical behavior in the brain at large scales, in
spite of the high degree of signal integration required to fire neuron
activity at small scale.
\section*{Acknowledgements}
We acknowledge support from the J. de Andaluc\'{i}a project of
Excellence P09-FQM-4682 and from the Spanish MEC project
FIS2009--08451.
\section*{References}
\providecommand{\newblock}{}
|
\section{Introduction}
The extended binomial coefficients, occasionally called polynomial coefficients (e.g., \cite[p.\ 77]{Comtet}), are defined as the coefficients in the expansion
\begin{equation}\label{ext}\sum_{k=0}^{\infty}\binom{n}{k}^{(q)} x^k = \left(1+x+x^2+\cdots+x^{q}\right)^n,\quad n,q \in \mathbb{N}=\{1,2,\ldots\}.
\end{equation}
In written form, they presumably appeared for the first time in works by De Moivre \cite[p.\ 41]{DeMoivre} and later they also were addressed by Euler \cite{Euler}. Since then, the extended binomial coefficients played a role mainly in the theory of compositions of integers as the number \(c(k,n,q)\) of compositions of \(k\) with \(n\) parts not exceeding \(q\) is given by
\[c(k,n,q)=\binom{n}{k-n}^{(q-1)}. \]
Thus, the extended binomial coefficients and their modifications have been studied in various papers and from different perspectives \cite{Andrews, Balakrishnan, Banderier, Caiado, Eger, Eger2, Fahssi, Heubach, Knopfmacher, Star}, and among the properties their distribution is of particular interest. Recently, Eger \cite{Eger2} showed (using a slightly different notation) that
\[\binom{n}{nq/2}^{(q)}\sim \frac{(q+1)^n}{\sqrt{2\pi n \frac{q(q+2)}{12}}},\]
as \(n\rightarrow\infty\), meaning that the quotient of both sides tends to unity. Moreover, based upon numerical simulations \cite{Eger2} the question arises how well those coefficients can be approximated by ``normal approximations'' in general. It is the aim of this note to give a precise and comprehensive answer to this question by establishing a complete asymptotic expansion for the extended binomial coefficients with error terms holding uniformly with respect to all integer \(k\). More precisely, we show the following.
\begin{theorem} For all integers \(N\geq 2\) we have
\[\sqrt{\frac{q(q+2) n}{12}}\frac{1}{(1+q)^n} \binom{n}{k}^{(q)}=\frac{1}{\sqrt{2 \pi}} e^{-x^2/2}+\sum_{\nu=1}^{\left[(N-2)/2\right]} \frac{q_{2\nu}(x)}{n^{\nu}} +o\left(\frac{1}{n^{(N-2)/2}}\right),\]
as \(n\rightarrow \infty\), uniformly with respect to all \(k\in\mathbb{Z}\), with
\[x=\frac{\sqrt{12}}{\sqrt{q(q+2) n}}\left(k-\frac{q}{2}n\right),\]
\end{theorem}
where the functions \(q_{2\nu}(x)\) are given explicitly as sums of Hermite polynomials and Bernoulli numbers (see Theorem \ref{Main} below for the exact formulae).
Although we only deal with the very basic situation of the extended binomial coefficients in (\ref{ext}) here, the presented approach is a general one, which admits the derivation of (complete) asymptotic expansions in many applications. However, it is not always possible to obtain the involved quantities in a very explicit form, which is an instance making the case of the extended binomial coefficients further interesting and worth to be presented.
\section{Proof of the main result}
First of all we fix some notations following Petrov \cite{Petrov}. For a (real) random variable \(X\) we denote its characteristic function by
\[\varphi_X (t)=Ee^{itX}, \quad t\in \mathbb{R},\]
where, as usual, \(E\) means the mathematical expectation with respect to the underlying probability distribution. If \(X\) has finite moments up to \(k\)-th order, then \(\varphi_X\) is \(k\) times continuously differentiable on \(\mathbb{R}\) and we have
\[\frac{d^k}{dt^k} \varphi_X (t) \Big\vert_{t=0} = \frac{1}{i^k} EX^k.\]
Moreover, in this case we define the cumulants of order \(k\) by
\[\gamma_k =\frac{1}{i^k} \frac{d^k}{dt^k} \log \varphi_X (t)\Big\vert_{t=0},\]
where the logarithm takes its principal branch. Now, let \(\left(X_n\right)\) be a sequence of independent integer-valued random variables having a common distribution and suppose that for all positive integer values of \(k\) we have
\[E\vert X_1\vert^k < \infty\]
and
\[EX_1=\mu,\quad Var X_1 =\sigma^2 >0. \]
Thus, for the sum given by
\[S_n=\sum_{\nu=1}^n X_{\nu}\]
we obtain
\[ES_n=n \mu,\quad Var S_n = n\sigma^2,\]
and for integer \(k\) we define the probabilities
\[p_n(k)=P\left(S_n =k\right).\]
Furthermore, we introduce the Hermite polynomials (in the probabilist's version)
\[H_m (x)=(-1)^m e^{x^2/2} \frac{d^m}{dx^m} e^{-x^2/2},\]
and for positive integers \(\nu\) we define the functions
\begin{equation}\label{qq}q_{\nu} (x) = \frac{1}{\sqrt{2 \pi}} e^{-x^2 \slash 2}
\sum_{k_1, \ldots, k_{\nu} \geq 0 \atop k_1 + 2 k_2 + \cdots + \nu k_{\nu} = \nu} ~
H_{\nu + 2s} (x) ~ \prod\limits^{\nu}_{m=1} ~ \frac{1}{k_m!}
\left( \frac{\gamma_{m+2}}{(m+2)! \sigma^{m+2}} \right)^{k_m},
\end{equation}
where \(s = k_1 + \cdots + k_\nu\) and \(\gamma_{m+2}\) denotes the cumulant of order \(m+2\) of \(X_1\).
Finally, we demand (for convenience) that the maximal span of the distribution of \(X_1\) is equal to one. This means that there are no numbers \(a\) and \(h>1\) such that the values taken on by \(X_1\) with probability one can be expressed in the form \(a+hk\) (\(k\in\mathbb{Z}\)). Under all these assumptions we have the following complete asymptotic expansion in the sense of a local central limit theorem \cite[p.\ 205]{Petrov}.
\begin{theorem}\label{EE} For all integers \(N\geq 2\) we have
\begin{equation}\label{EE1}\sigma \sqrt{n}p_n(k)=\frac{1}{\sqrt{2 \pi}} e^{-x^2/2}+\sum_{\nu=1}^{N-2} \frac{q_{\nu}(x)}{n^{\nu/2}} +o\left(\frac{1}{n^{(N-2)/2}}\right),
\end{equation}
as \(n\rightarrow \infty\), uniformly with respect to all \(k\in\mathbb{Z}\), where we have
\[x=\frac{k-n\mu}{\sigma \sqrt{n}}.\]
\end{theorem}
In the following we choose \(X_1\) to take the integer values \(\{0,\ldots,q\}\) with
\[P(X_1=k)=\frac{1}{q+1}, \quad k\in\{0,\ldots,q\}.\]
Hence, we obtain
\begin{equation}\label{P}p_n(k)=P(S_n=k)=\frac{1}{(1+q)^n} \binom{n}{k}^{(q)}, \quad k\in\mathbb{Z}.
\end{equation}
It is our aim to apply Theorem \ref{EE} in full generality and we want compute all cumulants as explicit as possible.
\begin{lemma} \label{CU}For the \(k\)-th order cumulant \(\gamma_k\) of \(X_1\) we have
\begin{equation}\label{GAMMA}\gamma_k= \begin{cases} \frac{q}{2}, &\text{if}~ k=1; \\
0, & \text{if}~ k ~\text{odd}~\text{and}~ k>1;\\
\frac{\mathcal{B}_{2l}}{2l} \left((q+1)^{2l}-1\right), & \text{if}~ k=2l, l\geq 1,\end{cases}
\end{equation}
where $\mathcal{B}_{\nu},\, \nu \geq 0$, denote the Bernoulli numbers, e.g., \cite[p.\ 22]{Grad}.
\end{lemma}
\begin{proof} First, we observe that the characteristic function of \(X_1\) is given by
\[\varphi_{X_1}(t)=\frac{1+e^{it}+\cdots+e^{qit}}{1+q}.\]
According to the definition of the cumulants we obtain for a positive integer \(k\)
\begin{align*}\gamma_k &=\frac{1}{i^k} \frac{d^k}{dt^k} \log \varphi_{X_1} (t)\Big\vert_{t=0}\\
&=\frac{1}{i^k} \frac{d^k}{dt^k} \left\{\log\left(1+e^{it}+\cdots+e^{qit}\right) -\log(1+q)\right\}\Big\vert_{t=0}\\
&=\frac{1}{i^k} \frac{d^k}{dt^k}\log\left(\frac{e^{(q+1)it}-1}{e^{it}-1}\right)\Big\vert_{t=0}\\
&=\frac{1}{i^k} \frac{d^k}{dt^k}\left\{\frac{q}{2}it+\log \left(\frac{\sin \frac{q+1}{2}t}{\sin \frac{t}{2}}\right)\right\} \Big\vert_{t=0}\\
&=\frac{q}{2}\delta_{k,1}+\frac{1}{i^k} \frac{d^k}{dt^k}\left\{\log \left(\frac{\sin \frac{q+1}{2}t}{ \frac{q+1}{2} t}\right)-\log \left(\frac{\sin \frac{t}{2}}{\frac{t}{2}}\right)\right\}\Big\vert_{t=0},
\end{align*}
where \(\delta_{k,1}\) denotes the Kronecker delta.
Using
\[\frac{d}{dz} \log \left(\frac{\sin z}{z}\right)=\cotan z -\frac{1}{z}\]
yields
\[\gamma_k=\frac{q}{2}\delta_{k,1}+\frac{1}{i^k} \frac{d^{k-1}}{dt^{k-1}}\left\{\frac{q+1}{2}\left(\cotan \frac{q+1}{2} t -\frac{2}{(q+1)t}\right)-\frac{1}{2}\left(\cotan \frac{t}{2} -\frac{2}{t}\right)\right\}\Big\vert_{t=0}.\]
Now, making use of the following expansion (see, e.g., \cite[p.\ 35]{Grad})
\[\cotan z - \frac{1}{z} = \sum\limits^{\infty}_{m = 1} (-1)^{m} \frac{4^{m}}{(2 m)!} \mathcal{B}_{2 m} z^{2 m - 1}
~~,~~ 0 < |z| < \pi ,\]
after some algebra we obtain
\[\gamma_k=\frac{q}{2}\delta_{k,1}+\frac{1}{i^k} \frac{d^{k-1}}{dt^{k-1}}\sum\limits^{\infty}_{m = 1} (-1)^{m} \frac{\mathcal{B}_{2 m}}{(2 m)!} \left((q+1)^{2m}-1\right)t^{2 m - 1}\Big\vert_{t=0}.\]
Carrying out the differentiation under the summation sign immediately gives us (\ref{GAMMA}).
\end{proof}
\begin{remark} As an immediate consequence of Lemma \ref{CU} we obtain
\[EX_1 =\mu=\gamma_1 =\frac{q}{2}\]
and, as we know \(\mathcal{B}_2=\frac{1}{6}\),
\[Var X_1 = \sigma^2=\gamma_2 =\frac{\mathcal{B}_2}{2}\left((q+1)^2-1\right)=\frac{q(q+2)}{12}. \]
\end{remark}
We now are ready to state the main theorem in form of a complete asymptotic expansion with explicit coefficients for the extended binomial coefficients \(\binom{n}{k}^{(q)}\).
\begin{theorem}\label{Main}For all integers \(N\geq 2\) we have
\[\sqrt{\frac{q(q+2) n}{12}}\frac{1}{(1+q)^n} \binom{n}{k}^{(q)}=\frac{1}{\sqrt{2 \pi}} e^{-x^2/2}+\sum_{\nu=1}^{\left[(N-2)/2\right]} \frac{q_{2\nu}(x)}{n^{\nu}} +o\left(\frac{1}{n^{(N-2)/2}}\right),\]
as \(n\rightarrow \infty\), uniformly with respect to all \(k\in\mathbb{Z}\), with
\[x=\frac{\sqrt{12}}{\sqrt{q(q+2) n}}\left(k-\frac{q}{2}n\right),\]
and
\begin{align}\label{q}&q_{2\nu} (x) = \frac{1}{\sqrt{2 \pi}} \left(\frac{12}{q(q+2)}\right)^{\nu} e^{-x^2 \slash 2}\\ \nonumber
&\times \sum_{k_2, k_4, \ldots, k_{2\nu} \geq 0 \atop k_2 + 2 k_4 + \cdots + \nu k_{2\nu} = \nu} ~
H_{2(\nu + s)} (x) \left(\frac{6}{q(q+2)}\right)^s \prod\limits^{\nu}_{m=1} ~ \frac{1}{k_{2m}!}
\left( \frac{\mathcal{B}_{2(m+1)} \left((q+1)^{2m+2}-1\right)}{(2m+2)! (m+1) } \right)^{k_{2m}},
\end{align}
where \(s=k_2+k_4+\cdots+k_{2\nu}\).
\end{theorem}
\begin{proof} The proof is based on an application of Theorem \ref{EE} to the probabilities defined in (\ref{P}). First we observe that in our situation the functions given in (\ref{qq}) vanish identically for odd indices, which turns out to be a consequence of (\ref{GAMMA}). Indeed, if \(\nu=2l+1\) for an integer \(l \geq 0\), then in every solution \(k_1,\ldots, k_{2l+1} \geq 0\) of the equation
\[k_1 +2k_2+\cdots+(2l+1)k_{2l+1}=2l+1\]
there is at least one odd index \(i\) with \(k_i >0\). Consequently, using (\ref{GAMMA}) we have
\[\prod\limits^{2l+1}_{m=1} ~ \frac{1}{k_m!}
\left( \frac{\gamma_{m+2}}{(m+2)! \sigma^{m+2}} \right)^{k_m}=0,\]
from which follows that \(q_{2l+1}(x)\) vanishes identically. Thus, only the functions \(q_{2\nu}(x)\) appear in (\ref{EE1}) and here we have
\[q_{2\nu} (x) = \frac{1}{\sqrt{2 \pi}} e^{-x^2 \slash 2}
\sum_{k_1, \ldots, k_{2\nu} \geq 0 \atop k_1 + 2 k_2 + \cdots + 2\nu k_{2\nu} = 2\nu} ~
H_{2(\nu + s)} (x) ~ \prod\limits^{2\nu}_{m=1} ~ \frac{1}{k_m!}
\left( \frac{\gamma_{m+2}}{(m+2)! \sigma^{m+2}} \right)^{k_m},
\]
where \(s = k_1 + \cdots + k_{2\nu}\). An analogous argument as in the odd case above shows that a solution \(k_1,\ldots,k_{2\nu}\) of the equation
\[k_1 +2k_2+\cdots+2\nu k_{2\nu}=2\nu\]
with a positive entry at an odd index does not give any contribution to the whole sum, so that we can write
\[q_{2\nu} (x) = \frac{1}{\sqrt{2 \pi}} e^{-x^2 \slash 2}
\sum_{k_2, k_4, \ldots, k_{2\nu} \geq 0 \atop k_2 + 2 k_4+ \cdots + \nu k_{2\nu} = \nu} ~
H_{2(\nu + s)} (x) ~ \prod\limits^{\nu}_{m=1} ~ \frac{1}{k_{2m}!}
\left( \frac{\gamma_{2m+2}}{(2m+2)! \sigma^{2m+2}} \right)^{k_{2m}},
\]
where \(s = k_2 + k_4+ \cdots + k_{2\nu}\). Now, taking the explicit form of the cumulants in (\ref{GAMMA}) into account, after some elementary computation we obtain (\ref{q}).
\end{proof}
\begin{remark} As a concluding remark we state the meaning of Theorem \ref{Main} for \(N=5\) explicitly. Using the known facts
\[H_4 (x)=x^4 -6x^2 +3,\quad \mathcal{B}_4=-\frac{1}{30},\]
we obtain
\[\sqrt{\frac{q(q+2) n}{12}}\frac{1}{(1+q)^n} \binom{n}{k}^{(q)}=\frac{1}{\sqrt{2 \pi}} e^{-x^2/2}\left\{1-\frac{\left((q+1)^4-1\right)\left(x^4 -6x^2 +3\right)}{20nq^2 (q+2)^2}\right\} +o\left(\frac{1}{n^{3/2}}\right),\]
as \(n\rightarrow \infty\), uniformly with respect to all \(k\in\mathbb{Z}\), where we have
\[x=\frac{\sqrt{12}}{\sqrt{q(q+2) n}}\left(k-\frac{q}{2}n\right).\]
\end{remark}
\section{Acknowledgments} This work is supported by KU Leuven research grant OT\slash12\slash073 and the Belgian Interuniversity Attraction Pole P07/18.
|
\section{Introduction}
Wireless Sensor Networks (WSNs) have attracted much interest in recent years \cite{akyildiz_commag02}. Detection, classification, or estimation of certain events, targets, or phenomena, in a region of interest, is an important application of sensor networks. Different aspects of this problem have been investigated by the research community over the last few decades~\cite{Viswanathan,veer} mostly in the context of the parallel network topology. In such a framework, due to power and bandwidth constraints, each node, instead of sending its raw data, sends quantized data to a central observer or fusion center (FC). The FC combines these local nodes' data to make a global inference~\cite{Varshney:book}.
Given a parallel topology, the objective is to find efficient quantization rules for the nodes and efficient inference rule for the FC, which maximize the global performance at the FC.
Note that, in general, the problem of designing optimal inference rules is computationally expensive (NP-hard)~\cite{JNT}.
For example, in a distributed detection framework, under the assumption of conditional independence, the optimal decision rule for each node takes the form of a likelihood ratio test, with a suitably chosen threshold. However, finding the optimal thresholds requires the solution of a system of non-linear equations and, therefore, the problem is difficult to solve, even for the network of moderate size. The analysis of optimal detection system performance is tractable only in asymptotic regime.
It has been shown that the use of identical thresholds is asymptotically optimal~\cite{tsit}.
Under the assumption of identical thresholds, several authors have considered the problem of designing optimal decision rules in the past~\cite{Zhang,Shi,Kailkhura}.
In contrast to the distributed detection problem, in a classification problem, each decision is usually represented by ${\log}_2 M$ information
bits, where $M$ is the number of classes to be distinguished.
The problem of classification using ${\log}_2 M$ information bits has been studied for parallel topology \cite{BaekB95}. Due to bandwidth constraints, it is desirable that the local node decisions are sent to the FC with as few bits as possible. To overcome this problem, distributed classification has been proposed in which the local nodes make 1-bit (rather than ${\log}_2 M$ bit) local decisions and send them to the FC~\cite{Zhangm,Zhu2004,Wang_jsac05}. The FC then uses the local decisions collectively and makes a global inference about the underlying phenomenon.
In~\cite{DLI,xiaohong}, the authors consider the problem of parameter estimation in a parallel topology.
Received signal strength based methods have been proposed which employ least-squares or maximum likelihood (ML) based parameter estimation techniques. These techniques are not suitable for power and bandwidth constrained networks. To overcome these drawbacks, distributed parameter estimation using quantized measurements has been addressed in \cite{RibeiroG2006a,RibeiroG2006b,NiuV2006}. Similar to the problem of distributed detection, the system design issues of distributed estimation have also been addressed only in certain scenarios, such as in \cite{VempatyCV2013}, where it has been shown that identical quantizers are optimal under certain conditions. To simplify things, in~\cite{adierror}, coding theory based iterative schemes were proposed for target localization using parallel topology where at every iteration, the FC solves an $M$-ary hypothesis testing problem and decides the region of interest for the next iteration.
Even though the parallel topology has received significant attention, there are many practical situations where parallel topology cannot be implemented due to several factors, such as, the FC being outside the communication range of the nodes and limited energy budget of the nodes~\cite{Lin}.
In such cases, a multi-hop network is employed, where nodes are organized hierarchically into multiple levels (tree networks).
Some examples of tree networks include wireless sensor and military communication networks.
For instance, the IEEE 802.15.4 (Zigbee) specifications \cite{Aliance} and IEEE 802.22b \cite{802_22bDraft}
can support tree based topologies.
There have been limited attempts to address the distributed inference problems in tree networks~\cite{TayTW:IT08,ZhangCPMH13,Kailkhura2013,bhavyaj}. In all but the simplest cases, optimal strategies in tree based networks are difficult to derive. Most of the work on tree networks focuses on person-by-person optimal (PBPO) strategies~\cite{TayTW:IT08,ZhangCPMH13,Kailkhura2013,bhavyaj}. Also, the above works address the problem of distributed detection in tree networks while, to the best of our knowledge, the problem of distributed estimation in tree networks has not received any attention. Due to the complexity of classification and estimation in tree networks as compared to detection, these problems have been left unexplored by researchers.
In this paper, we take a first step to address the distributed inference (classification and estimation) problems in tree networks by developing an analytically tractable framework.
We first consider the distributed classification problem and propose to use coding theory based techniques to solve the problem. We analyze the asymptotic classification performance of our
scheme for two different classes of tree networks: fixed height tree networks, and fixed degree tree networks. We show that the proposed scheme is asymptotically optimal under certain conditions. Building on these results, we extend our scheme to consider the distributed estimation problem and analyze the asymptotic estimation performance of our
scheme. The remainder of the paper is organized as follows. In Section~\ref{sec:prel}, we describe the system architecture and present a brief overview of Distributed Classification Fusion using Error Correcting Codes (DCFECC) scheme~\cite{Wang_jsac05} which serves as a foundation for the schemes presented in this paper. We propose our basic coding scheme for distributed classification in tree networks in Section~\ref{sec:approachclass}. The performance of the proposed scheme in the asymptotic regime is also analyzed. We present some numerical results to gain insights into the solution. We extend this scheme for distributed estimation in tree networks in Section~\ref{sec:approachest} by formulating the estimation problem as a sequence of $M$-ary classification problems. The performance of the proposed scheme in the asymptotic regime is analyzed and some numerical results are presented. We also provide a technique for optimal region splitting for distributed estimation.
Finally, we conclude our paper in Section~\ref{sec:disc} with some discussion on possible future work.
\section{Preliminaries}
\label{sec:prel}
\subsection{General Network Architecture}
\label{sec:syst}
Consider a perfect tree, $T(K,\;N)$, rooted at the FC (Please see Figure~\ref{syst}). Nodes at level $k$, for $1\leq k\leq K-1$, are referred to as intermediate nodes and nodes at the last level of the tree, i.e., $k=K$, are called the leaf nodes. In a perfect tree, all the intermediate nodes have an equal number of immediate successors and the number of such successors $N$ is referred to as the degree of the tree.
\begin{figure}[t]
\centering
\includegraphics[height=2in, width=!]{systree}
\vspace*{-0.3in}
\caption{A distributed inference system organized as a perfect binary tree; $T(3,\;2)$ is shown as an example.}\label{syst}
\vspace*{-0.1in}
\end{figure}
We assume that the network is designed to infer about a particular phenomenon.
Each node $j$ at level $k$ performs two basic operations:
\begin{itemize}
\item Depending on the task, sense data regarding the phenomenon and/or collect data from its successors at level $k+1$, denoted by $S^{k+1}(j)$.
\item Compress the data available at node $j$ about the phenomenon and transmit a 1-bit version to its predecessor at level $k-1$, denoted by $P^{k-1}(j)$.
\end{itemize}
Local observation of node $j$ at level $k$ is denoted as $y_j^k$. Received data vector at node $j$ of level $k$ from its successors $S^{k+1}(j)$ at level $k+1$ is denoted as $\mathbf{v_j^k}\in\{0,1\}^N$. After processing the data at the node according to a processing model (Please see Figure~\ref{syst1}),
every node $j$ at level $k$ sends its one-bit local decision $u_j^k\in\{0,1\}$ to its immediate predecessor. This processing model is designed based on the inference problem considered, i.e., Figure~\ref{syst2} for classification or Figure~\ref{syst3} for estimation. Finally, the FC receives the inference vector $\mathbf{u^1}=(u_{1}^{1},\cdots,u_{N}^{1})\in\{0,1\}^N$ and fuses this data to infer about the underlying phenomenon. In our analysis, we consider error-free links in the network. However, we do provide some simulation results for the case where there are erroneous links, to examine the robustness of the proposed schemes.
Given a tree network, our objective is to find the appropriate processing scheme for nodes at all levels depending on the inference problem considered. Next, we describe the distributed classification fusion using error-correcting codes scheme (originally proposed for parallel topology in \cite{Wang_jsac05}) which serves as the mathematical basis for the ideas proposed in this paper.
\begin{figure}[t]
\centering
\includegraphics[height=2in, width=!]{sys1}
\caption{Data processing for distributed inference at node $j$ at level $k$. Here $y_j^k$ and $\mathbf{v_j^k}$ are the inputs and $u_j^k\in\{0,1\}$ is the output of the process at node $j$.}\label{syst1}
\end{figure}
\subsection{Distributed Classification Fusion using Error-Correcting Codes (DCFECC)}
\label{sec:DCFECC}
In~\cite{Wang_jsac05}, the authors proposed the DCFECC scheme for $M$-ary distributed classification using binary quantized local data for a parallel topology network. The idea behind the DCFECC scheme is to select a binary code matrix ${C}$ to determine the local decision rules at the nodes, and to perform fault-tolerant fusion at the FC. For a network with $N$ nodes trying to distinguish among $M$ hypotheses, the code matrix ${C}$ is an $M\times N$ binary matrix. Each row of ${C}$ corresponds to one of the $M$ possible hypotheses $H_1,\cdots,H_M$ and each column represents the binary decision rule of the corresponding node. Given this code matrix, the node $j$ sends its binary decision $u_j\in\{0,1\}$ to the FC. After receiving the binary decisions $\mathbf{u}=(u_1,\cdots,u_N)$ from local nodes, the final classification decision is made at the FC using minimum Hamming distance based fusion given by:
Decide $H_m$ where
\begin{equation}
\label{eq:hamm}
m=\argmin_{1 \leq l\leq M}{ d_H(\mathbf{u},\mathbf{r}_l)},
\end{equation}
where $d_H(\mathbf{x},\mathbf{y})$ is the Hamming distance between $\mathbf{x}$ and $\mathbf{y}$, and $\mathbf{r}_l=(c_{l1},\cdots,c_{lN})$ is the $l$th row of ${C}$ which corresponds to hypothesis $H_l$.
The tie-break rule is to randomly pick a row of the code matrix ${C}$ from those with the smallest Hamming distance to the received vector $\mathbf{u}$. The performance of the scheme depends on the code matrix ${C}$ since it is used for designing the local decision rules as well as for the fusion rule at the FC. Several approaches to design the matrix $C$, e.g., based on simulated annealing and cyclic column replacement, were presented in~\cite{Wang_jsac05}.
For example, consider the code matrix used by a parallel network of $N=7$ nodes performing an $(M=4)$-ary classification problem
\[ C= \left[ \begin{array}{ccccccc}
1 & 0 & 0 & 0 & 1 & 0 & 1 \\
0 & 0 & 1 & 0 & 0 & 0 & 0 \\
1 & 0 & 1 & 1 & 0 & 1 & 0 \\
0 & 1 & 1 & 1 & 1 & 1 & 1 \end{array} \right].\]
When the true hypothesis is $H_1$ corresponding to the first row, all the nodes are supposed to send the first element of their column. However, due to imperfect observations at the nodes, consider the case when the FC receives the vector $[1 1 1 0 1 0 1]$. The FC evaluates the Hamming distance between this received vector and each of the rows resulting in the Hamming distance values $(2,4,5,3)$. Therefore, it decides the hypothesis corresponding to the first row, $H_1$, as the true hypothesis.
\section{Distributed Classification in Tree Networks}
\label{sec:approachclass}
In this section, we consider the problem of distributed classification in tree networks.
We model the classification problem as an $M$-ary hypotheses testing problem. Let $H_l$, where $l=1,\cdots,M$ and $M\geq 2$, denote the $M$ hypotheses\footnote{In order to distinguish among the $M$ hypotheses using binary decisions, we assume that $N\geq{\log}_2 M$.}. The \textit{a priori} probabilities of these $M$ hypotheses are denoted by $Pr(H_l)=P_l$, for $l=1,\cdots,M$.
\subsection{Proposed Scheme}
\label{ap:class}
We assume that under each hypothesis $H_l$, every leaf node $j$ acts as a source and makes an independent and identically distributed (i.i.d.) observation $y_j^K$. After processing the observations locally, every leaf node $j$ sends its local decision\footnote{In this context, ``decision" is a binary quantized value determined by the processing model.} $u_{j}^{K}\in \{0,1\}$ according to a transmission mapping $\tau_j^K(\cdot)$ to its immediate predecessor $P^{K-1}(j)$.
Each intermediate node $j$ at level $k$ receives the decision vector consisting of local decisions made by its immediate successors $S^{k+1}(j)$ at level $k+1$, which can be expressed as $\mathbf{v_j^k}=\mathbf{u^{k+1}}=(u_{1}^{k+1},\cdots,u_{N}^{k+1})$. After fusing this data using fusion rule $f_j^{k}(\cdot)$, this intermediate node $j$ at level $k$ makes a classification decision $y_j^k\in\{1,\cdots,M\}$. Then, it sends a 1-bit version of this decision, $u_{j}^{k}\in \{0,1\}$, according to its transmission mapping $\tau_j^{k}(\cdot)$ to its immediate predecessor $P^{k-1}(j)$. Finally, the FC receives the decision vector $\mathbf{u^1}=(u_{1}^{1},\cdots,u_{N}^{1})$ and fuses this data to decide the underlying hypothesis.
The proposed scheme builds on the DCFECC scheme (Section~\ref{sec:DCFECC}). To summarize, each node $j$ at level $k$, for $1\leq k\leq K-1$, performs two basic operations (Please see Figure~\ref{syst2}):
\begin{itemize}
\item Collect data from its successors $S^{k+1}(j)$ and fuse their data using fusion rule $f_j^k(\cdot)$ to locally decide the hypothesis, denoted by $y_j^k$.
\item Compress the decision $y_j^k$ at node $j$ about the hypothesis and transmit a 1-bit version $u_j^k$ to its predecessor $P^{k-1}(j)$ using the transmission mapping $\tau_j^k(\cdot)$.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[height=1.5in, width=!]{sys2}
\vspace{-0.2in}
\caption{Data processing for distributed classification at node $j$ at level $1\leq k\leq K-1$. Here $\mathbf{v_j^k} \in \{0,1\}^N$, $y_j^k\in\{1,\cdots,M\}$, and $u_j^k\in\{0,1\}$. Therefore, the mappings are $f_j^k: \{0,1\}^N \to \{1,\cdots,M\}$ and $\tau_j^k: \{1,\cdots,M\} \to \{0,1\}$}\label{syst2}
\end{figure}
For the leaf nodes (level $K$), there are no successors and, therefore, only the second operation needs to be performed. And, for the FC (level `0'), only the first operation needs to be performed. Each of the functions $f_j^k(\cdot)$ and $\tau_j^k(\cdot)$ depend on the code matrix used at level $k$. By appealing to symmetry, we assume that each node at the same level $k$, uses an identical code matrix $C^k$ for transmission to its predecessor and $C^{k+1}$ for fusion of data from its successors. We start with the design of transmission mapping $\tau^K_j(\cdot)$ of the leaf nodes. When the leaf nodes use code matrix $C^K$ for transmission, the probability of misclassification at level $K-1$ is given by \cite{Wang_jsac05}
\begin{equation}
P_e^{K-1}=
\sum_{\mathbf{i},l}\int_{\mathbf{y^K}}P_lP(u_1^K=i_1|y_1^K)\times\cdots\times P(u_N^K=i_N|y_N^K)p(\mathbf{y^K}|H_l) \psi^{K}_{\mathbf{i},l},\label{eq:leaf}
\end{equation}
where $\mathbf{i}=[i_1,\cdots,i_N]\in\{0,1\}^N$ is a realization of the received codeword $\mathbf{u^K}$, $\mathbf{y^K}=[y^K_1,\cdots,y^K_N]$ are the local observations of leaf nodes, and $\psi^{K}_{\mathbf{i},l}$ is the cost associated with a global decision $H_l$ at level $K-1$ when the received vector from level $K$ is $\mathbf{i}$. This cost is:
\begin{equation}
\label{eq:cost}
\psi_{\mathbf{i},l}^k=
\begin{cases}
1-\frac{1}{\varrho} &\mbox{if } \mathbf{i} \mbox{ is in the decision region of } H_l\\
1 & \mbox{otherwise.}
\end{cases}
\end{equation}
for $k=K$, where $\varrho$ is the number of decision regions corresponding to a received codeword $\mathbf{i}$. In other words, it is the number of rows of code matrix $C^K$ which have the same minimum Hamming distance with the received codeword $\mathbf{i}$. Usually this value is 1, however $\varrho$ can be greater than one when there is a tie at the node at level $K-1$ and in those cases, the tie-breaking rule is to choose one of them randomly.
Employing a person-by-person optimization approach, we can find the local transmission mapping of the leaf nodes as follows \cite{Wang_jsac05}:
\begin{equation}
u_j^K=\tau_j^K(y_j^K)=\\
\begin{cases}
0, &\text{if $\sum_{l}p(y_j^K|H_l)A_{jl}<0$}\\
1, &\text{otherwise}
\end{cases},
\end{equation}
where $A=\{A_{jl}\}$ is the weight matrix whose values\footnote{We refer the reader to \cite{Wang_jsac05} for further details.} are given by,
\begin{multline}
A_{jl}=\sum_{i_1,\cdots,i_{j-1},i_{j+1},\cdots,i_N}P_lP(u_1^K=i_1|H_l)
\times\cdots\times P(u^K_{j-1}=i_{j-1}|H_l)P(u^K_{j+1}=i_{j+1}|H_l)\\
\times\cdots\times P(u_{N}^K=i_{N}|H_l)
\times [\psi^K_{i_1,\cdots,i_{j-1},0,i_{j+1},\cdots,i_N,l}-\psi^K_{i_1,\cdots,i_{j-1},1,i_{j+1},\cdots,i_N,l}].
\end{multline}
For $1\leq k\leq K-1$, the local classification decision $y_j^k \in \{1,\cdots,M\}$ made using the data from the successors $S^{k+1}(j)$ is discrete and, therefore, the transmission mapping $\tau^k(\cdot)$ is straight-forward and is given as follows:
\begin{equation}
u_j^k=\tau^k_j(y_j^k)=c^{k}_{y_j^kj}, \quad\text{if $1\leq k\leq K-1$.}
\end{equation}
In other words, the one-bit decision $u_j^k$ is the element of $C^{k}$ corresponding to $y_j^k$th row and $j$th column. For every intermediate node, the fusion rule $f_j^k(\cdot)$ is the minimum Hamming distance fusion rule as given in \eqref{eq:hamm}. Therefore, the performance of the scheme depends on the minimum Hamming distance of the code matrices. Let $d_{min}^k$ be the minimum Hamming distance of the code matrix $C^k$. In the remainder of this section, we derive the error expressions at intermediate nodes which will later be used for the design of code matrices at every level.
\begin{proposition}
\label{prop1}
The probability of misclassification $P_e^{k-1}$ at level $k-1$ due to the data received from level $k$ and using code matrix $C^k=\{c_{mj}^k\}$ ($1\leq k\leq K-1$, $1\leq m\leq M$, $1\leq j\leq N$) is:
\begin{equation}
\label{eq:pe_inter}
P_e^{k-1}=\sum_{\mathbf{i},l}P_l\prod_{j=1}^N \left[(2i_j-1)\sum_{m=1}^Mc^k_{mj}P^k_{ml} +(1-i_j)\right]\psi^k_{\mathbf{i},l} \mbox{,}
\end{equation}
where $\mathbf{i}=[i_1, \cdots, i_N] \in \{0, 1\}^N$ is the realization of the received codeword $\mathbf{u^k}$, matrix $P^k=\{P_{ml}^k\}$ is the confusion matrix of the local decisions at level $k$, and $\psi^k_{\mathbf{i},l}$ is the cost associated with a global decision $H_l$ at level $k-1$ when the received vector from level $k$ is $\mathbf{i}$. This cost is given by \eqref{eq:cost}.
\end{proposition}
\begin{IEEEproof}
If $u_j^k$ denotes the bit sent by the node $j$ at level $k$ and the global decision is made using the Hamming distance criterion:
\begin{equation}
P_{e}^{k-1}=\sum_{\mathbf{i},l}P_lP(\mathbf{u^k}=\mathbf{i}|H_l)\psi_{\mathbf{i},l}^k\mbox{.}
\end{equation}
Since local decisions are conditionally independent, $P(\mathbf{u^k}=\mathbf{i}|H_l)=\prod_{j=1}^N P(u_j^k=i_j|H_l)$.
Further,
\begin{align*}
P(u_j^k=i_j|H_l) &= i_jP(u_j^k=1|H_l)+(1-i_j)P(u_j^k=0|H_l)\\
&=(1-i_j)+(2i_j-1)P(u_j^k=1|H_l)\\
&=(1-i_j)+(2i_j-1)\sum_{m=1}^Mc^k_{mj}P(y_j^k=m|H_l)\\
&=(1-i_j)+(2i_j-1)\sum_{m=1}^Mc^k_{mj}P^k_{ml}
\end{align*}
where $y_j^k$ is the local classification decision made by node $j$ after collecting data from its successors $S^{k+1}(j)$ at level $k+1$. The desired result follows.
\end{IEEEproof}
Note that this suggests that the probability of misclassification at level $k-1$ is dependent on the confusion matrix at level $k$. These can be derived easily as follows:
\begin{equation}
P^k_{ml}{\stackrel{\Delta}{=}} P(\text{decide $H_m$ at level $k$}|\text{$H_l$ is true})=
1-\sum_{\mathbf{i}}p(\mathbf{u^k}=\mathbf{i}|H_l)\psi_{\mathbf{i},m}^{k+1}\label{eq:conf}
\end{equation}
From these expressions, we can observe that there is a recursive structure, where the probability of misclassification at level $k$ is dependent on the confusion matrix of level $k+1$. Therefore, the performance at the FC depends on all the code matrices in a recursive manner. As mentioned before, we propose a simpler approach by assuming that each node of the same level uses the same code matrix which is designed by optimizing on a person-by-person sequential basis. We start with the code design at level $K-1$ to fuse data from level $K$. This is designed by optimizing the expression in \eqref{eq:leaf}. Once we have designed the optimal code matrix at this level, we derive the corresponding confusion matrix from \eqref{eq:conf}, which is used to design the code matrix at the next level by optimizing the expression in \eqref{eq:pe_inter}. Following this method, we can design all the code matrices. Note that each of these optimizations can be performed offline using approaches such as simulated annealing or cyclic-column replacement \cite{Wang_jsac05}.
In the following subsection, we analyze our scheme in the asymptotic regime and show that the scheme is asymptotically optimal.
\subsection{Asymptotic Optimality}
\label{sec:a-optimalclass}
We study the asymptotic classification performance of our scheme for two different classes
of tree networks. The first one is the class of \textit{fixed height trees} in which the
height of the tree, $K$, is assumed to be fixed while the second is the class of \textit{fixed degree trees} in which the degree of the tree, $N$, is assumed to be fixed.
More specifically, we study the classification performance of minimum Hamming distance fusion in fixed height tree networks when the number of nodes tends to infinity and in fixed degree tree networks when the height of the tree tends to infinity.
We first provide the following bound on the misclassification probability at the FC which will be used to prove the asymptotic optimality. Let $Q^k_{m}$ be the probability of misclassifying hypothesis $H_m$ at level $k$ and define $q_{max}^k {\stackrel{\Delta}{=}} \max_{1\leq m\leq M} Q_m^k$. Note that for levels $0\leq k \leq K-1$, we have $Q^k_{m}=1-P_{mm}^k$ where $P_{ml}^k$ are the elements of the confusion matrix. For $k=K$, $Q_m^K=1-Pr(\text{decide $H_m$ at level $K | H_m$ is true})$.
\begin{proposition}
\label{asy}
In a perfect tree structure $T(K,N)$ employing the proposed scheme, the misclassification probability at the FC, $P_{e}^{0}$, is bounded as follows
\begin{equation}
P_{e}^{0}\leq \left[q_{max}^K\right]^{\displaystyle\prod_{k=1}^{K}\frac{d_{min}^k}{a_k}},\label{eq:bound}
\end{equation}
if $q_{max}^K< \frac{1}{2}$ and
\begin{equation}
\label{dcon}
d_{min}^k\geq \dfrac{2(M-2)}{[1-4q_{max}^k(1-q_{max}^k)]-(1/a_k)[(2/q_{max}^k)-2]},\;\forall k,
\end{equation}
where $a_k$ is a parameter which satisfies the following condition
\begin{equation}
\label{acon}
a_k > \dfrac{2(1-q_{max}^k)}{q_{max}^k-4(q_{max}^k)^2(1-q_{max}^k)},\forall k.
\end{equation}
\end{proposition}
\begin{proof}
To prove the proposition, we start with the inequality \eqref{dcon}
\begin{eqnarray}
d_{min}^k\geq \dfrac{2(M-2)}{[1-4q_{max}^k(1-q_{max}^k)]-(1/a_k)[(2/q_{max}^k)-2]}\\
\implies\frac{d_{min}^k}{2}\left(1-4q_{max}^k(1-q_{max}^k)\right)\geq (M-2)+\frac{d_{min}^k}{a_k}\left(\frac{1}{q_{max}^k}-1\right)\label{eq11}\\
\implies\frac{d_{min}^k}{2}\log\left(\frac{1}{4q_{max}^k(1-q_{max}^k}\right)\geq \log(M-1)+\frac{d_{min}^k}{a_k}\log\frac{1}{q_{max}^k}\label{thirdineq}\\
\implies\left[q_{max}^k\right]^{\frac{d_{min}^k}{a_k}}\geq (M-1) \left[\sqrt{4q_{max}^{k}(1-q_{max}^{k})}\right]^{d_{min}^k}
\end{eqnarray}
where \eqref{eq11} is true because $a_k$ satisfies \eqref{acon}, and \eqref{thirdineq} can be proved by applying the logarithm inequality: $(x-1)\geq\log x \geq \dfrac{x-1}{x}$, for $x>0$. Now, for $k=1,\cdots,K$
\begin{eqnarray}
Q_{m}^{k-1}= Pr(\text{decision at level $k-1$} \neq H_m\;|\; H_m)&\leq& Pr(d^k(\mathbf{u^k},\mathbf{c_m^k})\geq \underset{1 \leq l \leq M,\;l\neq m}{\min}d^k(\mathbf{u^k},\mathbf{c_l^k})\;|\;H_m)\nonumber\\
&\leq& \sum\limits_{\underset{l\neq m}{l=1}}^{M} Pr(d^k(\mathbf{u^k},\mathbf{c_m^k})\geq d^k(\mathbf{u^k},\mathbf{c_l^k})\;|\;H_m)\nonumber\\
&\leq& \sum\limits_{\underset{l\neq m}{l=1}}^{M} \left[\sqrt{4Q_{mm}^{k}(1-Q_{mm}^{k})}\right]^{d^k(\mathbf{c_l^k},\mathbf{c_m^k})}\label{17}\\
&\leq& (M-1) \left[\sqrt{4q_{max}^{k}(1-q_{max}^{k})}\right]^{d_{min}^k}\\
&\leq& \left[q_{max}^k\right]^{\frac{d_{min}^k}{a_k}}\label{19}.
\end{eqnarray}
Note that \eqref{17} is true when $Q_{m}^{k}<\frac{1}{2}$ as shown in \cite{ChenWHKY05}, which holds when \eqref{dcon} and \eqref{acon} are true. Using the above results, the average probability of error can be bounded as follows.
\begin{eqnarray*}
P_e^0&=& \sum\limits_{m=1}^{M}P_m Pr(\text{decision at the FC}\neq H_m\;|\; H_m)\\
&\leq&\sum\limits_{m=1}^{M}P_m \left[q_{max}^1\right]^{\frac{d_{min}^1}{a_1}}\\
&=& \left[q_{max}^1\right]^{\frac{d_{min}^1}{a_1}}
\end{eqnarray*}
Now, since $Q_{m}^{1}\leq\left[q_{max}^2\right]^{\frac{d_{min}^2}{a_2}}$ $\forall m$, we have $q_{max}^{1}\leq\left[q_{max}^2\right]^{\frac{d_{min}^2}{a_2}}$. Continuing in this manner, we get
\begin{eqnarray*}
P_e^0&\leq& \left[q_{max}^K\right]^{\prod_{k=1}^{K}\frac{d_{min}^k}{a_k}}.
\end{eqnarray*}
\end{proof}
The results obtained in Proposition~\ref{asy} show that the misclassification probability for minimum Hamming distance fusion can be upper-bounded by a quantity determined by the minimum Hamming distance of the code matrices ($d_{min}^k,\;\forall k$), and the largest local classification error among all hypotheses ($q_{max}^k,\;\forall k$). Also, note that the parameter $a_k$ in \eqref{eq:bound} can be chosen appropriately to make the bound tighter. For example, if $a_k$ is chosen such that
\begin{equation*}
\dfrac{q_{max}^k(2M-2)+2(1-q_{max}^k)}{q_{max}^k-4(q_{max}^k)^2(1-q_{max}^k)}>a_k > \dfrac{2(1-q_{max}^k)}{q_{max}^k-4(q_{max}^k)^2(1-q_{max}^k)},\forall k,
\end{equation*}
then, $(d_k/a_k)>1,\forall k$, and we have $$P_{e}^{0}\leq \left[q_{max}^K\right]^{\prod_{k=1}^{K}(d_{min}^k/a_k)}\leq \left[q_{max}^K\right]^{(d_{min}^K/a_K)}.$$
As a consequence, for fixed height trees, the decoding error of the proposed scheme vanishes as $d_{min}^K$ approaches infinity which happens when $N\to\infty$.
Also, for fixed degree trees,
the decoding error of the proposed scheme vanishes as $K$ approaches infinity.
These results can be summarized in the following theorem:
\begin{theorem}
Under conditions \eqref{dcon} and \eqref{acon}, the proposed coding theory based distributed classification scheme is asymptotically optimal, for both classes of tree networks: fixed
height tree networks and fixed degree tree networks, as long as the probabilities of correct local
classification for all hypotheses of the leaf nodes are greater than one half.
\end{theorem}
The conditions required for the above theorem depend on the minimum Hamming distance $d_{min}^k$ of the code matrices used at each level and can be interpreted as follows: the proposed scheme is optimal when the minimum Hamming distance of the code matrices is ``large enough" to ensure that perfect classification is made at every level of the tree. When the rows of the code matrices are well separated due to
large minimum Hamming distance, the proposed scheme can handle more errors and have good performance.
Also, observe that these results imply that when \eqref{dcon} and \eqref{acon} are satisfied, there is no loss in asymptotic performance when all the nodes at level $k$, for $k=1,\cdots,K$, use identical transmission mapping and identical fusion rules.
\subsection{Simulation Results}
\label{sec:simclass}
In this section, we evaluate the performance of the proposed scheme using simulations. Consider a tree network $T(3,7)$ consisting of a total $N_{total}=400$ nodes, including the FC. The leaf nodes sense the environment to identify among four ($M=4$) equally likely hypotheses. As discussed before, we assume that all the leaf node measurements are independent and identically distributed. Under each hypothesis, the probability density function is assumed to
be Gaussian distribution with the same variance ($\sigma^2=1$) but with different means $0,s,2s$, and $3s$ respectively. The signal-to-noise power ratio (SNR) of observations at each local node is given by $20\log_2{s}$. The code matrices are designed using the scheme described in Section~\ref{ap:class} and simulated annealing for optimization. The designed code matrices used at different levels of the tree are found to be
\begin{equation}
C^1=[11,8,9,9,3,9,12]
\end{equation}
\begin{equation}
C^2=[7,6,3,12,12,9,14]
\end{equation}
\begin{equation}
C^3=[3,8,14,12,9,12,9]
\end{equation}
where the code matrix is represented by a vector of $M$-bit integers. Each integer $m_j$ represents a
column of any arbitrary code matrix $C$ and can be expressed as $m_j=\sum_{l=1}^{M}c_{lj}$. For example, the integer $9$ in column 5 of $C^3$ represents $c^3_{15}=1,c^3_{25}=0,c^3_{35}=0$, and $c^3_{45}=1$.
\begin{figure}[t]
\centering
\includegraphics[width=3in,height=!]{simulation}\vspace{-0.25in}
\caption{Probability of misclassification versus SNR}
\label{fig:sim}
\end{figure}
In Figure~\ref{fig:sim}, we plot the final probability of misclassification at the FC with varying SNR values. Note that this probability of misclassification is empirically found by performing $N_{mc}=5000$ Monte-Carlo runs. As we can observe, the performance of the scheme improves with increasing SNR and approaches 0 as early as 5dB. Since the proposed scheme is based on error-correcting codes, it can also tolerate some errors in data. These errors could be due to various reasons: presence of a faulty node \cite{Wang_jsac05}, presence of imperfect links between levels \cite{varshney_spmag06}, or presence of a malicious node sending falsified data \cite{vempaty_spm13}. In order to check the fault-tolerance capability of the scheme, we have simulated the case when the links between the levels are binary symmetric channels with crossover probabilities $\beta=0.05$ and $\beta=0.1$. As shown in Figure~\ref{fig:sim}, the proposed scheme still performs reasonably well even in the presence of imperfect data due to non-ideal channels modeled as binary symmetric channels.
Building on these results, in the following section, we address the parameter estimation problem in tree based networks. More specifically, we break the parameter estimation problem into a sequence of $M$-ary decision making problems, and each of these $M$-ary decision making problems is solved using a technique similar to the distributed classification scheme of the previous section.
\section{Distributed Parameter Estimation in Tree Networks using Iterative Classification}
\label{sec:approachest}
Consider a distributed parameter estimation problem where the goal is to estimate a random scalar parameter $\theta$ at the FC. The parameter $\theta$ has a prior probability density function (pdf) $p_\theta(\theta)$ where $\theta\in \Theta$. We propose a scheme to estimate the parameter $\theta$ using iterative classification. By doing
so, we break the parameter estimation problem into a sequence of $M$-ary decision making problems. This is essentially a process of iterative rejection of unlikely objects where the most undesirable options are discarded and the scope of options is progressively narrowed down until exactly one option is left.
\subsection{Proposed Scheme}
\label{sec:systestm}
\begin{figure}[t]
\centering
\includegraphics[width=3in,height=!]{normal}\vspace{-0.25in}
\caption{An example of splitting of parameter space.}
\label{fig:normal}
\end{figure}
We consider a distributed estimation system with the topology of a perfect tree, $T(K,\;N)$, rooted at the FC.
We model the parameter estimation problem as an $M$-ary hypotheses testing problem. Our scheme is iterative in which at every iteration $1\leq s \leq K$, the parameter space is split into $M$ regions and an $M$-ary hypothesis test is performed at the level $(K+1-s)$ of the tree to determine the parameter space for the next level in the tree. The optimal splitting of the parameter space at every iteration can be determined offline (which will be explained later in the paper in Section~\ref{sec:split}). For now, we assume that the $M^K$ final regions and their corresponding representation points are known. Let $H_l^k$, where $l=1,\cdots,M$ and $M\geq 2$, denote the $M$ hypotheses\footnote{As before, we assume that $N\geq{\log}_2 M$.} being tested at level $k$. Figure~\ref{fig:normal} shows an example of parameter space splitting when $p_\theta(\cdot)$ is standard Gaussian, and $M=K=2$. Every node at level $k=2$ first performs a classification task to determine if the parameter $\theta$ is positive or negative (differentiate between hypotheses $H_1^2$ and $H_2^2$). After a decision is made, the nodes at level $k=1$, `zoom' into the decided hypothesis, say $H_1^2$, and perform a classification task to determine if $\theta$ belongs to hypothesis $H_1^2$ or $H_2^1$. In this manner, the FC at level `0' eventually decides the true hypothesis, among the $M^K$ hypotheses, where $\theta$ belongs.
The \textit{a priori} probabilities of the $M^K$ hypotheses are denoted by $Pr(H_l^k)=P_l^k$, for $l=1,\cdots,M$ and $k=1,\cdots,K$. $P_l^k$ depends on $p_\theta(\cdot)$ and the region corresponding to $H_l^k$.
We assume that every node $j'$ at level $k+1$ acts as a source and makes a conditionally independent and identically distributed (i.i.d.) observation $y_{j'}^{k+1}$, conditioned under each hypothesis $H_l^k$. After processing the observations locally, every node sends its local decision $u_{j'}^{k+1}\in \{0,1\}$ according to a transmission mapping $\tau^{k+1}(\cdot)$ to its immediate predecessor $P^k(j')$.
Each intermediate node $j$ at level $k$ receives the decision vector $\mathbf{v_{j}^k}$ consisting of local decisions made by its immediate successors $S^{k+1}(j)$ at level $k+1$. Intermediate nodes at level $k$, through collaboration\footnote{In collaboration phase, node $j$ at level $k$ shares $\mathbf{v_j^k}$ (the data collected from its successors $S^{k+1}(j)$) with other nodes at level $k$. In this paper, we assume that nodes do not compress $\mathbf{v_j^k}$ for collaboration and, therefore, after collaboration phase receive the data $\mathbf{v^k}=\left[\mathbf{v_1^k},\cdots,\mathbf{v_N^{k}}\right]\in\{0,1\}^{N^k}$.} and fusion, decide on the result of the $M$-ary hypotheses test as the new parameter space for them.
\begin{figure}[t]
\centering
\includegraphics[height=.95in, width=!]{sys3}
\vspace{-0.2in}
\caption{Data processing for distributed estimation at node $j$ at level $1\leq k\leq K-1$. Here, $\mathbf{v_j^k}\in\{0,1\}^N$, $\mathbf{v^k}\in\{0,1\}^{N^k}$, $y_j^k\in\mathbb{R}$, and $u_j^k\in\{0,1\}$. Therefore, the mappings are $f_j^k:\{0,1\}^{N^k}\to\{1,\cdots,M\}$ and $\tau_j^k:\mathbb{R}\to\{0,1\}$.}\label{syst3}
\end{figure}
The scheme builds on the DCFECC scheme proposed for distributed classification. Each node $j$ at level $k$, for $1\leq k\leq K-1$ performs four basic operations (Please see Figure~\ref{syst3}):
\begin{itemize}
\item Collect data from its successors $S^{k+1}(j)$ and collaborate with other nodes at level $k$.
\item Decide the new parameter space (hypotheses to test) by fusing data using fusion rule $f_j^k(\cdot)$.
\item Acquire observation $y_j^k$ and perform hypothesis testing to determine the new parameter space.
\item Compress the observation at node $j$ about the hypothesis (new parameter space) and transmit a 1-bit version to the predecessor $P^{k-1}(j)$ using the transmission mapping $\tau_j^k(\cdot)$.
\end{itemize}
For the leaf nodes (level $K$), there are no successors and, therefore, only the third and fourth operations need to be performed. The FC (level `0') collects data from its successors and makes the final decision regarding the region where $\theta$ belongs. Given a tree network, our objective is to find efficient transmission
mappings and fusion rules for nodes at all levels, to
maximize the estimation performance at the FC.
\textit{Remark:} There are three major differences between the scheme proposed here for distributed estimation in tree networks and the scheme proposed in Section~\ref{sec:approachclass} for distributed classification in tree networks:
\begin{itemize}
\item In the scheme proposed here, every node acts as a source node and senses the phenomenon while for the classification problem in Section~\ref{sec:approachclass}, only the leaf nodes act as source nodes and intermediate nodes act only as relay nodes.
\item In Section~\ref{sec:approachclass}, each node performs the same classification task, or in other words, the set of classes are the same. On the other hand, in the scheme proposed here, the parameter space is `zoomed' at every level, which changes the corresponding classes to be tested.
\item An important step in the scheme proposed in this section is the collaboration step which is not required for the classification problem of Section~\ref{sec:approachclass}.
\end{itemize}
By appealing to symmetry, we assume that each node at the same level $k$, uses an identical code matrix $C^k$ for transmission to its predecessor and $C^{k+1}$ for fusion of data from its successors. Each of the functions $f_j^k(\cdot)$ and $\tau_j^k(\cdot)$ depend on the code matrix used at level $k$. Although the performance metric in this
framework is the Mean Square Error (MSE), it is difficult
to obtain a closed form representation for MSE. Therefore,
typically, one uses the bounds on MSE to characterize the
performance of the estimator. Here, we use an analytically tractable metric to analyze the performance of the proposed scheme which is the probability of misclassification of the parameter region. It is an important metric when the final goal of the parameter estimation task is to find the approximate region or neighborhood where the parameter lies rather than the true value of the parameter itself. Since the final region could be one of the $M^K$ regions, a metric of interest is the probability of `zooming' into the correct region. In other words, it is the probability that the true value of the parameter and the estimated value of the parameter lie in the same region.
Now, we design the transmission mapping $\tau_j^k(\cdot)$ of nodes at level $k$. Notice that the final region of the estimated value of the parameter is the same as the true value of the parameter, if and only if we `zoom' into the
correct region at every iteration of the proposed scheme. Thus, when the nodes at level $k$ use code matrix $C^k$ for transmission, the probability of misclassification at level $k-1$ is given by the following proposition.
\begin{proposition}
\label{prop2}
The probability of misclassification $P_e^{k-1}$ at level $k-1$ due to the data received from level $k$ and using code matrix $C^k=\{c_{mj}^k\}$ ($1\leq k\leq K$, $1\leq m\leq M$, $1\leq j\leq N^k$) is
\begin{equation}
P_e^{k-1}= 1-\prod\limits_{t=k}^{K}
\Bigg[1-\sum_{\mathbf{i},l}\int_{\mathbf{y^t}}P_l^t P(u_1^t=i_1|y_1^t)\times\cdots
\times P(u_{N^t}^t=i_{N^t}|y_{N^t}^t)p(\mathbf{y^t}|H_l^t) \psi^{t}_{\mathbf{i},l}\Bigg],\label{eq:nk}
\end{equation}
where $\mathbf{i}=[i_1, \cdots, i_{N^t}] \in \{0, 1\}^{N^t}$ is the realization of the received codeword $\mathbf{u^t}$, $\mathbf{y^t}=[y^t_1,\cdots,y^t_{N^t}]$ are the local observations of nodes at level $t$, matrix $P^t=\{P_{ml}^t\}$ is the confusion matrix of the local decisions at level $t$, and $\psi^{t-1}_{\mathbf{i},l}$ is the cost associated with a global decision $H_l^{t-1}$ at level $t-1$ when the received vector from level $t$ is $\mathbf{i}$. This cost is:
\begin{equation}
\label{eq:cost1}
\psi_{\mathbf{i},l}^t=
\begin{cases}
1-\frac{1}{\varrho} &\mbox{if } \mathbf{i} \mbox{ is in decision region of } H_l^t\\
1 & \mbox{otherwise.}
\end{cases}
\end{equation}
where as before $\varrho$ is the number of decision regions corresponding to a received codeword $\mathbf{i}$. In other words, it is the number of rows of code matrix $C^t$ which have the same minimum Hamming distance with the received codeword $\mathbf{i}$.
$P_l^t$ is the prior probability of hypothesis $H_l^t$ at level $t$.
\end{proposition}
\begin{proof}
Note that a correct decision is made at level $k-1$ if and only if the decision at all levels from $t=k$ to $t=K$ are correct. Therefore, using \eqref{eq:leaf} in a recursive manner at every level of the tree, we get the desired result.
\end{proof}
From \eqref{eq:nk}, we can observe that the performance at the FC depends on all the code matrices in a recursive manner. In this paper, we employ a simpler approach by assuming that code matrices are designed by optimizing on a person-by-person basis. The code matrix at each level is designed using an approach similar to that of a parallel topology. Note that each of these optimizations can be performed offline using approaches such as simulated annealing or cyclic-column replacement \cite{Wang_jsac05}.
Employing a person-by-person optimization approach, we can find the local transmission mapping of the nodes at level $k$ as follows:
\begin{equation}
u_j^k=\tau_j^k(y_j^k)=\\
\begin{cases}
0, &\text{if $\sum_{l}p(y_j^k|H_l^k)A_{jl}^k<0$}\\
1, &\text{otherwise}
\end{cases},
\end{equation}
where $A^k=\{A^k_{jl}\}$ is a weight matrix whose values are given by,
\begin{multline}
A_{jl}^k=\sum_{i_1,\cdots,i_{j-1},i_{j+1},\cdots,i_{N^k}}P_l^k P(u_1^k=i_1|H_l^k)
\times\cdots\times P(u^k_{j-1}=i_{j-1}|H_l^k)P(u^k_{j+1}=i_{j+1}|H_l^k)\\
\times\cdots\times P(u_{N^k}^K=i_{N^k}|H_l^k)
\times [\psi^k_{i_1,\cdots,i_{j-1},0,i_{j+1},\cdots,i_{N^k},l}-\psi^k_{i_1,\cdots,i_{j-1},1,i_{j+1},\cdots,i_{N^k},l}].
\end{multline}
For every intermediate node, the fusion rule $f_j^k(\cdot)$ is the minimum Hamming distance fusion rule as given in \eqref{eq:hamm}. Therefore, the performance of the scheme depends on the minimum Hamming distance of the code matrices. Let $d_{min}^k$ be the minimum Hamming distance of the code matrix $C^k$. In the remainder of this section, we analyze our scheme in the asymptotic regime and show that the scheme is asymptotically optimal.
\subsection{Asymptotic Optimality}
\label{sec:a-optimalest}
As before, we study the asymptotic performance of our scheme for two different classes
of tree networks, \textit{fixed height trees} and \textit{fixed degree trees}. We also analyze the scenarios where both the number of nodes and the height of the tree tend to infinity. We first provide the following bound on the misclassification probability at the FC which will be used to prove the asymptotic optimality. Let $Q^k_{m}$ be the probability of misclassifying hypothesis $H_m^k$ at level $k$ and define $q_{max}^k {\stackrel{\Delta}{=}} \max_{1\leq m\leq M} Q_m^k$. For $k=1,\cdots,K$, $Q_m^K=1-Pr(\text{decide $H_m^k$ at level $K | H_m^k$ is true})$.
\begin{proposition}
\label{asy1}
In a perfect tree structure $T(K,N)$ employing the proposed scheme, if $q_{max}^k<\frac{1}{2}$, the misclassification probability at the FC, $P_{e}^{0}$, is bounded as follows
\begin{equation}
\label{asym}
P_e^0\leq 1-\prod\limits_{k=1}^{K}\left[1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right].
\end{equation}
\end{proposition}
\begin{proof}
To prove the proposition, we first establish the following set of inequalities (Please see \eqref{17}-\eqref{19})
\begin{eqnarray*}
\sum\limits_{m=1}^{M} P_m^k Pr(\text{decision at level $k-1$}\neq H_m^k\;|\; H_m^k)&\leq& \sum\limits_{m=1}^{M} P_m^kPr(d^k(\mathbf{u^k},\mathbf{c_m^k})\geq \underset{1 \leq l \leq M,\;l\neq m}{\min}d^k(\mathbf{u^k},\mathbf{c_l^k})\;|\;H_m^k)\nonumber\\
&\leq&\sum\limits_{m=1}^{M} P_m^k \sum\limits_{\underset{l\neq m}{l=1}}^{M} Pr(d^k(\mathbf{u^k},\mathbf{c_m^k})\geq d^k(\mathbf{u^k},\mathbf{c_l^k})\;|\;H_m^k)\nonumber\\
&\leq& \sum\limits_{m=1}^{M}P_m^k \sum\limits_{\underset{l\neq m}{l=1}}^{M} \left[\sqrt{4Q_{mm}^{k}(1-Q_{mm}^{k})}\right]^{d^k(\mathbf{c_l^k},\mathbf{c_m^k})}\\
&\leq& (M-1) \left[\sqrt{4q_{max}^{k}(1-q_{max}^{k})}\right]^{d_{min}^k}.
\end{eqnarray*}
Therefore,
\begin{eqnarray*}
\sum\limits_{l=1}^{M} P_m^k Pr(\text{decision at level $k-1$}\neq H_m^k\;|\; H_m^k)&\leq& (M-1) \left[\sqrt{4q_{max}^{k}(1-q_{max}^{k})}\right]^{d_{min}^k}\\
\Leftrightarrow
\sum\limits_{l=1}^{M} P_m^k Pr(\text{decision at level $k-1$}= H_m^k\;|\; H_m^k)&\geq &1-(M-1) \left[\sqrt{4q_{max}^{k}(1-q_{max}^{k})}\right]^{d_{min}^k}.
\end{eqnarray*}
Now,
\begin{eqnarray*}
P_e^0&=&1-\prod\limits_{k=1}^{K}\sum\limits_{l=1}^{M} P_m^k Pr(\text{decision at level $k-1$}= H_m^k\;|\; H_m^k)\\
&\leq&1-\prod\limits_{k=1}^{K}\left[1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right].
\end{eqnarray*}
\end{proof}
As a consequence of Proposition~\eqref{asy1}, for fixed height trees, the probability of `zooming' into the incorrect region of the proposed scheme vanishes as $d_{min}^k$ approaches infinity which happens when $N\to\infty$.
\begin{small}
\begin{eqnarray*}
\lim\limits_{N\rightarrow\infty} P_e^0&\leq& \lim\limits_{N\rightarrow\infty} \left[ 1-\prod\limits_{k=1}^{K}\left(1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right)\right]\\
&=& \left[ 1-\prod\limits_{k=1}^{K}\lim\limits_{N\rightarrow\infty}\left(1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right)\right]\\
&=& 1-\prod\limits_{k=1}^{K}\left[1-(M-1)\lim\limits_{N\rightarrow\infty}\left((4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right)\right]\\
&=& 1-\prod\limits_{k=1}^{K}\left[1-(M-1)0\right]\\
&=& 0.
\end{eqnarray*}
\end{small}
Hence, the overall detection probability becomes `1' as the
degree of the tree $N$ goes to infinity. This shows that
the proposed scheme asymptotically attains perfect region
detection probability for bounded height tree networks if $q_{max}^k<1/2$ $\forall k=1,\cdots,K$. Notice that perfect region detection probability does not imply that the estimation error will vanish. It just provides a coarse estimate of the parameter. For estimation error to vanish, $M^K\rightarrow\infty$, which can be achieved by letting $K$ approach infinity.
However, for fixed degree trees,
misclassification error of the proposed scheme need not vanish as $K$ approaches infinity.
\begin{eqnarray*}
\lim\limits_{K\rightarrow\infty} P_e^0&\leq& \lim\limits_{K\rightarrow\infty} \left[ 1-\prod\limits_{k=1}^{K}\left(1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right)\right]\\
&=& \left[ 1-\lim\limits_{K\rightarrow\infty}\prod\limits_{k=1}^{K}\left(1-(M-1)(4q_{max}^k(1-q_{max}^k))^{\frac{d_{min}^k}{2}}\right)\right]
\end{eqnarray*}
For misclassification error to vanish, every term in the product should vanish, which obviously is not true for the above equation.
These results can be summarized as the following theorem:
\begin{theorem}
The proposed iterative classification scheme for distributed parameter estimation in tree based networks is asymptotically optimal (when both the degree $N$ and number of levels $K$ simultaneously approach infinity), as long as the probabilities of correct local classification for all hypotheses at each node is greater than one half.
\end{theorem}
\textit{Remark:} Note that, while for distributed classification, we have shown that the proposed scheme is asymptotically optimal if either $N$ or $K$ tend to infinity, for the distributed estimation case, we have proved that the scheme is asymptotically optimal when both $N$ and $K$ tend to infinity.
Next, we address the remaining aspect of the scheme which is the discretization of the continuous parameter space to perform estimation as iterative classification.
\subsection{Optimal Splitting of the Parameter Space}
\label{sec:split}
As mentioned before, the scheme splits the parameter space $\Theta$ into $M^K$ regions. Therefore, the MSE between the true parameter value $\theta$ and the FC's estimate $\hat{\theta}$ is affected by two factors: the quantization of the continuous region $\Theta$ into $M^K$ discrete points and the probability of misclassifying the region where the true parameter belongs. In Section~\ref{sec:a-optimalest}, we showed that the probability of misclassification can be made to tend to zero by using a large sensor network. Therefore, in order to minimize the MSE, we need to minimize the error due to the quantization of $\Theta$ into $M^K$ points. This optimal splitting depends on the prior pdf $p_\theta(\cdot)$ and can be determined by using ideas from rate distortion theory \cite{cover_inftheory}. As mentioned in \cite{cover_inftheory}, the optimal regions for quantization are given by Voronoi regions and the reconstruction points should minimize the conditional expected distortion over their respective assignments. One of the most popular algorithms used to determine these regions is the Llyod-Max algorithm \cite{lloyd,max}. This algorithm is iterative where we start with an initial set of reconstruction points which are typically chosen at random. It then repeatedly executes the following steps until convergence:
\begin{itemize}
\item Compute the optimal set of reconstruction regions (Voronoi regions) and
\item Find the set of optimal reconstruction points for these regions (centroid of the Voronoi regions).
\end{itemize}
In this paper, we use this algorithm which is performed offline and, therefore, is not a computational issue.
\subsection{Simulation Results}
\label{sec:simest}
In this section, we provide simulation results to evaluate the performance of the proposed scheme. As before, consider a tree network $T (3, 7)$ consisting of a total $N_{total} = 400$ nodes, including the FC. The observation at each node is Gaussian distributed with unknown mean $\theta$ and variance $\sigma^2 = 1$. This unknown parameter $\theta$ is uniformly distributed in $(0, \theta_{max} )$ where the region size is varied by varying the maximum value $\theta_{max}$. At each level, the nodes perform an $M$-ary classification where $M = 4$. Therefore, there are a total of $M^K = 4^3 = 64$ possible estimates of $\theta$. Since the parameter is uniformly distributed, the optimal splitting is uniform quantization into $M^K$ regions with the mid-points of the regions as the corresponding representation. Due to the complexity in designing the optimal matrix of size $4\times343$ for transmission at level $3$ (due to collaboration, each node at level $2$ has data of all nodes at level $3$), we employ a sub-optimal approach by concatenating the optimal code matrix of size $4\times7$. For level $3$, it is concatenated $49$ times, and for transmission at level $2$, it is concatenated $7$ times. The smaller code matrix of size $4\times7$ is designed using the simulated annealing approach.
\begin{figure}[t]
\centering
\includegraphics[width=3in,height=!]{MSE}\vspace{-0.25in}
\caption{MSE as a function of the range of $\theta$}
\label{fig:MSE}
\end{figure}
In Figure \ref{fig:MSE}, we plot the mean square error (MSE) between the true value of $\theta$ and its estimate $\hat{\theta}$ at the FC\footnote{As discussed before, this estimate is one of the $M^K$ discrete points representing the quantized regions (centroids of the Voronoi regions, please see Section~\ref{sec:split}).} as a function of $\theta_{max}$. This value of MSE is empirically found by performing $N_{mc} = 5000$ Monte-Carlo runs. As we can observe, the performance of the scheme gets worse with increasing region size. This is because, when the range of $\theta$ is increased while the total number of possible estimates remains fixed, the error due to quantization increases. Since the proposed scheme is based on error-correcting codes, it can tolerate some errors in data. As mentioned before, these errors could be due to various reasons \cite{VempatyHVV14}. We have also simulated the case when the links between the levels are modeled as binary symmetric channels with crossover probability $\beta = 0.1$. As shown in Figure~\ref{fig:MSE}, the proposed scheme is quite robust to the presence of imperfect data arising due to non-ideal channels modeled as binary symmetric channels. As alluded to before, this robustness in performance is due to the use of error-correcting codes.
\section{Conclusion}
\label{sec:disc}
In this paper, we considered the general framework of distributed inference problem in tree networks. We proposed an analytically tractable scheme to solve these problems and proved the asymptotic optimality of the proposed schemes. For the classification problem, when the number of hypotheses is $M=2$, the proposed scheme is a majority-vote scheme for distributed detection in tree networks. Also, note that since the proposed scheme uses error-correcting codes, it works well even in scenarios with unreliable data \cite{VempatyHVV14}. It should be pointed out that the proposed scheme is not limited to wireless sensor networks, although the application of wireless sensor networks has been considered in this paper. The DCFECC scheme has been found to be applicable to a number of other applications including the paradigm of crowdsourcing \cite{VempatyVV2014}. We believe that one can use these results to address several other applications involving tree structures.
\bibliographystyle{IEEEtran}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.